University of Southern California Dissertations and Theses

Using demand-side management for decarbonization: developing methods to quantify the impact of altering electricity consumption patterns

(USC Thesis Other)

Using demand-side management for decarbonization: developing methods to quantify the impact of altering electricity consumption patterns

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content Using demand-side management for decarbonization: developing methods to quantify the impact of altering electricity consumption patterns Stepp Mayes A Dissertation Presented to the FACULTY OF THE GRADUATE SCHOOL UNIVERSITY OF SOUTHERN CALIFORNIA In Partial Fulfillment o f the Requirements for the Degree DOCTOR OF PHILOSOPHY (ENVIRONMENTAL ENGINEERING) December 2023 Acknowledgements I would like to first thank my advisor, Dr. Kelly Sanders, for being a mentor in every sense of the word, but especially for the guidance, the support, and the trust you have shown in me throughout the last few years. I would like to also thank my qualifying and defense committee members for helping to shape my research projects and challenge my understanding of engineering. Similarly, I would like to thank my professors and mentors at USC for their efforts. I am extremely grateful for my friends who have supported me in this process; thank you for visiting me, providing welcome distractions, and making LA feel like home. I would also like thank my friends at USC for making my PhD experience an enjoyable and social one, and my lab group for doing all of that as well as making time for many conversations about research projects and coding problems. A special thanks to McKenna and Andrew, for being with me every step of the way. Finally, a big thank you to my family members who have always been there for me and believed in me. Thank you to my sister, mom, and dad for years of advice and support. I have learned so much from you all. i Table of Contents Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . i List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiii Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xv Chapter 1: Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.2 Summary of research gaps and the contributions of this work . . . . . . . . . 2 1.2.1 Quantifying the impacts of residential precooling . . . . . . . . . . . 2 1.2.2 Improving the estimation of marginal emissions factors . . . . . . . . 3 1.3 Research Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 1.4 Structure of this document and resulting publications to date . . . . . . . . . 4 Chapter 2: Quantifying the electricity, CO2 emissions, and economic tradeoffs of precooling strategies for a single-family home in Southern California 6 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 2.2 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 2.2.1 Simulation Program . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 2.2.2 Simulation Details . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 2.2.3 Thermal Comfort . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 2.2.4 Output Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 2.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 2.3.1 Understanding the hourly impact of precooling . . . . . . . . . . . . . 19 2.3.2 Cumulative impact of precooling . . . . . . . . . . . . . . . . . . . . 23 2.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 ii 2.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 Chapter 3: A data-driven framework for quantifying consumption-based monthly and hourly marginal emissions factors . . . . . . . . . . . . . . . 33 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33 3.2 Contributions of This Study . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 3.3 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38 3.3.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38 3.3.2 The System Boundaries . . . . . . . . . . . . . . . . . . . . . . . . . 39 3.3.3 Data Sources and Data Processing Steps . . . . . . . . . . . . . . . . 40 3.3.4 Consumption-Based Hourly CO2 Emissions Estimates . . . . . . . . . 42 3.3.5 The Averaging Method for Estimating AEFs . . . . . . . . . . . . . . 42 3.3.6 The Regression Model for Estimating MEFs . . . . . . . . . . . . . . 43 3.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 3.4.1 Correlation Between Emissions and Demand . . . . . . . . . . . . . . 46 3.4.2 MEFs by Demand Level in Each Month . . . . . . . . . . . . . . . . 48 3.4.3 Month-Hour MEFs . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49 3.4.4 Month-Hour AEFs . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 3.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51 3.5.1 Consumption-Based Versus Generation-Based MEFs . . . . . . . . . 51 3.5.2 Year-to-Year Variation in MEFs . . . . . . . . . . . . . . . . . . . . . 52 3.5.3 The Influence of Hydropower and Imports on AEFs and MEFs . . . . 54 3.5.4 Comparing MEFs Versus AEFs . . . . . . . . . . . . . . . . . . . . . 56 3.5.5 The Influence of Other Interactions . . . . . . . . . . . . . . . . . . . 57 3.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 Chapter 4: Residential precooling on a high-solar grid: Impacts on CO2 emissions, peak period demand, and electricity costs across California . 59 iii 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59 4.2 Literature review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61 4.3 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64 4.3.1 Single-Family Home Selection . . . . . . . . . . . . . . . . . . . . . . 64 4.3.2 Weather Files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65 4.3.3 Precooling Schedules . . . . . . . . . . . . . . . . . . . . . . . . . . . 66 4.3.4 EnergyPlus Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . 67 4.3.5 Thermal Comfort . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68 4.3.6 Output Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69 4.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72 4.4.1 The impact of precooling schedule parameters . . . . . . . . . . . . . 73 4.4.2 The impact of building properties and Climate Zone . . . . . . . . . . 75 4.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78 4.5.1 Impact of Precooling on Thermal Comfort . . . . . . . . . . . . . . . 78 4.5.2 Influence of TOU Rates . . . . . . . . . . . . . . . . . . . . . . . . . 80 4.5.3 Precooling for CAISO’s Flex Alert Program . . . . . . . . . . . . . . 81 4.5.4 Impact of Precooling on CO2 Emissions . . . . . . . . . . . . . . . . . 82 4.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83 Chapter 5: Using Neural Networks to Forecast Marginal Emissions Factors: A CAISO Case Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86 5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86 5.2 Material and methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91 5.2.1 Preparation of Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . 91 5.2.2 Composite Model Design . . . . . . . . . . . . . . . . . . . . . . . . . 93 5.2.3 Model Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . 95 5.2.4 Model Outputs for Historical Data . . . . . . . . . . . . . . . . . . . 95 5.3 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97 iv 5.3.1 Hourly MEFs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97 5.3.2 Forecasting Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99 5.3.3 Drivers of MEFs and Feature Importance . . . . . . . . . . . . . . . . 100 5.3.4 Application to DSM . . . . . . . . . . . . . . . . . . . . . . . . . . . 102 5.3.5 Limitations and Future Improvements . . . . . . . . . . . . . . . . . 104 5.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106 Chapter 6: Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109 Appendices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133 A. Supplemental information for Chapter 1 . . . . . . . . . . . . . . . . . . . 133 A.1 Building Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133 A.2 Climate Zone Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133 A.3 Hourly Average Emissions Factors . . . . . . . . . . . . . . . . . . . . . . . . 134 A.4 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136 B. Supplemental information for Chapter 2 . . . . . . . . . . . . . . . . . . . 147 B.1 Summary of relevant literature . . . . . . . . . . . . . . . . . . . . . . . . . . 147 B.2 Data Repository . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147 B.3 CAISO Electricity Trades . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148 B.4 Outlier Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148 B.5 Monthly MEFs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150 B.6 Coefficient of Determination for MEFs . . . . . . . . . . . . . . . . . . . . . 150 B.7 Changes in MEFs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151 B.8 MEF-AEF Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151 B.9 Comparison of Generation-Based vs Consumption-Based MEFs . . . . . . . 152 v C. Supplemental information for Chapter 3 . . . . . . . . . . . . . . . . . . . 153 C.1 Building Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153 C.2 MEF and AEF Calculations . . . . . . . . . . . . . . . . . . . . . . . . . . . 154 C.3 Additional Results: The impact of building properties and Climate Zone . . 154 vi List of Figures 1 Net demand for electricity in the CAISO region on April 24th, 2023. Net demand is equal to demand minus the supply of wind and solar power. Data from CAISO. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 2 Total CAISO electricity supply mix (top row), net demand (middle row), and electricity CO2 emissions intensity (bottom row) for June 2019 (left) and 20 June 2019 (right). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 3 Overview of our modeling framework. . . . . . . . . . . . . . . . . . . . . . . 15 4 A sample thermostat setpoint schedule illustrating the simulation periods defined in Table 1, with corresponding hourly electricity usage dedicated to air conditioning. For this schedule, the length of the precool is 3 h, the depth is 3 °F, and the offset is 3 °F. The AC setpoint (blue line) is lower than the baseline temperature in the middle of the day and higher than the baseline temperature during the peak demand period. The corresponding hourly AC electricity consumption (orange line) peaks in the middle of the afternoon (coinciding hours that have a relatively low average emissions factor, as shown in Figure 2), and is lower during the evening hours (when the average emissions factor is relatively high during the diurnal period). . . . . . . . . . . . . . . 17 5 Top: the changes in average hourly electricity consumption for a shallow precool (4 h length, 1 °F depth, 3 °F offset) scenario, as compared to the 75 °F baseline simulation, are shown in blue for an average day in the month of September. Bottom: the same is shown for a deeper precool (3 h length, 3 °F depth, 3 °F offset). Hourly PMV is shown in green, which stays between the +0.5 and −0.5 limits for both schedules. . . . . . . . . . . . . . . . . . . 20 vii 6 Top: the changes in average hourly CO2 emissions between a shallow precool (4 h length, 1 °F depth) and the 75 °F baseline simulation are shown in blue for the month of September. Bottom: the same is shown for a deeper precool (3 h length, 3 °F depth). Hourly PMV is shown in green. . . . . . . . . . . . 22 7 Top: the change in average hourly residential electricity costs for a household for a shallow precool (4 h length, 1 °F depth) compared to the 75 °F baseline simulation is shown in blue for the month of September. Bottom: the same is shown for a deeper precool (3 h length, 3 °F depth). Hourly PMV is shown in green. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 8 Each dot represents one of the 70 distinct precooling schedules that passed the PMV comfort test. The x-axis shows percent change in aggregate peak demand period consumption and the y-axis shows the percent change in CO2 emissions between the cooling schedule and the baseline simulation over the five-month period. The color of each dot represents the percent change in residential electricity cost over this same period. Cooling schedules are illustrated by the black bars and groupings, with the width of the bar proportional to the length of the precool, the height of the bar proportional to the depth of the precool, and the offset shown by each grouping. . . . . . . . . . . . . . . 25 9 Includes subset of 3 °F offset scenarios from Figure 8 The schedule that reduced peak period cooling electricity most is shown with purple bars, and the schedule that reduced peak period cooling electricity consumption and CO2 emissions most is shown with orange bars. The color scale for residential cost is maintained from Figure 8. . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 10 Visual representation of the proposed framework to evaluate MEF values . . 45 viii 11 Scatter plots of hourly CO2 emissions and hourly demand for electricity in 2019 (Subplot A) and 2020 (Subplot B), as well scatter plots of changes between consecutive hours in hourly emissions and hourly demand between in 2019 (Subplot C) and 2020 (Subplot D). The line for best fit is shown in yellow, with R2=0.35 (Subplot A) and 0.50 (Subplot B) for hourly emissions and demand, and R2=0.40 (Subplot C) and 0.39 (Subplot D) for hourly changes in emissions and demand. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47 12 MEFs for each demand level and month combination in 2019 and 2020 with line of best fit. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49 13 Month-hour distribution of AEFs (top) and MEFs (bottom) in 2019 and 2020. Colors represent the magnitude of emissions factor in kgCO2/MWh. . . . . . 50 14 Average hourly generation for each resource serving CAISO’s electricity demand in March and July of 2019 and 2020. Data from CAISO. . . . . . . . . 54 15 Average changes in hourly generation between two consecutive hours for each resource serving CAISO’s electricity demand in March and July of 2019 and 2020. Data from CAISO. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 16 Theoretical precooling schedule with illustration of key terms from Table 5. . 67 17 Month-hourly Average (left) and Marginal (right) Emissions Factors for the CAISO grid in 2021. Factors for the months of July, August, and September were included in this study. . . . . . . . . . . . . . . . . . . . . . . . . . . . 71 18 Changes in daily cooling related CO2 emissions, daily kWh of electricity consumed for cooling, and monthly cooling costs over the 3-month analysis period for multiple single-family home designs in California Climate Zone 9. Each point represents the precooling schedule that achieved maximum CO2 emissions reductions for a given precooling length-depth combination, with the length and depth of the schedule shown as the width and height of the crosshairs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74 ix 19 The emissions, peak period electricity consumption, and residential electricity costs for the baseline cooling schedule as well as the precooling schedule that maximized CO2 reductions and the one that maximized peak period electricity reductions for each of the 4 single-family home designs in various California building climate zones. The length and depth of the schedule are shown as the width and height of the cross-hairs. . . . . . . . . . . . . . . . . . . . . . 76 20 The emissions, peak period electricity consumption, and residential electricity costs for the baseline cooling schedule as well as the precooling schedule that maximized CO2 reductions and the one that maximized peak period electricity reductions for Buildings 1 and 4 in five different California building climate zones. The length and depth of the schedule are shown as the width and height of the cross-hairs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 21 Mean PMV during the offset period as a function of the length and depth of the precool for Buildings 1 and 4. PMV values outside the range of -0.5 to +0.5 are considered uncomfortable; all values for Building 1 fall above +0.5 and all for Building 4 fall between 0 and +0.5. . . . . . . . . . . . . . . . . . 79 22 Minimum on-peak to off-peak electricity price ratio needed to make the the schedule that maximizes peak period electricity reductions also cost-reducing for each home and climate zone combination. Climate zones are organized by number of CDD from low to high. . . . . . . . . . . . . . . . . . . . . . . . . 81 23 Conceptualization of the data sources used, the MLP-linear composite model, and the learning process. The model predicts hourly changes in the emissions associated with demand in the CAISO region. . . . . . . . . . . . . . . . . . 94 24 Hourly marginal emissions factors calculated with the composite model for CAISO for 2019-2021. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98 x 25 Hourly percent difference between the MEFs predicted by the composite model using actual versus forecasted CAISO demand and variable renewable energy for 2021. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100 26 Mean Shapley value by month of 2021. . . . . . . . . . . . . . . . . . . . . . 102 27 Difference in MEFs between an initial hour and a new hour on April 17th, 2021 that an activity could occur in. Shades of blue represent switches in the timing of an activity that would reduce emissions. . . . . . . . . . . . . . . . 103 A.1 Top of each sub-figure: the changes in average hourly electricity consumption for a shallow precool (4 h length, 1 °F depth, 3 °F offset) scenario, as compared to the 75 °F baseline simulation, are shown in blue for an average day in each month June through October. Bottom of each sub-figure: the same is shown for a deeper precool (3 h length, 3 °F depth, 3 °F offset). Hourly PMV is shown in green, which stays between the +0.5 and −0.5 limits for both schedules in each month. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144 B.1 California Independent System Operator (CAISO) has many direct exchange of electricity with many other balancing authorities in Western Electricity Coordinating Council mainly in the Southwest and Northwest regions. . . . . 148 B.2 Hourly Emissions for March 23, 2019. The emissions data reported for hour 15 are an outlier. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149 B.3 MEF values across all demand levels in each month of the years 2019 and 2020.150 B.4 MEF values for each demand level occurring in each month of the year in 2019 and 2020 and corresponding R2 values. . . . . . . . . . . . . . . . . . . . . . 151 B.5 Percent relative change in 2020 month-hour MEFs compared to 2019. . . . . 152 B.6 Percent difference between MEF and AEF in 2019 and 2020. . . . . . . . . . 152 B.7 omparison of this study’s month-hour MEFs with CEDM’s estimated MEFs. 153 xi C.8 The emissions, peak period electricity consumption, and residential electricity costs for the baseline cooling schedule as well as the precooling schedule that maximized CO2 reductions and the one that maximized peak period electricity reductions for Buildings 2 and 3 in five different California building climate zones. The length and depth of the schedule are shown as the width and height of the cross-hairs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155 C.9 The emissions, peak period electricity consumption, and residential electricity costs for the baseline cooling schedule as well as the precooling schedule that maximized CO2 reductions and the one that maximized peak period electricity reductions for each of the four single-family home designs in various California building climate zones. The length and depth of the schedule are shown as the width and height of the cross-hairs. . . . . . . . . . . . . . . . . . . . . . 157 xii List of Tables 1 Description of precooling schedule periods and definition of the key terms used. 13 2 The ten schedules most effective at reducing CO2 emissions and the associated changes in CO2 emissions, cost, and peak demand period consumption. . . . 27 3 Electricity supply sources in the CAISO region (total electricity supplied was 219.5 TWh in 2019 and 218.5 TWh in 2020) . . . . . . . . . . . . . . . . . . 39 4 Summary of California’s Climate Zones . . . . . . . . . . . . . . . . . . . . . 65 5 Definitions of precooling-related terms used in this study . . . . . . . . . . . 66 6 Accuracy of different MEF models when using actual historical data and historical forecasted data as inputs. Only the MLP composite model is structured to use forecasted data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99 7 Correlation between MEFs and measures of grid load (i.e., hourly demand and net load), hourly changes in measures of grid load, and hour of the day. Correlations are shown for each year of the study period as well as for the entire period. Correlation is measured by Pearson’s rho for continuous variables, and pseudo-rho (square root of goodness-of-fit R-squared) for categorical variables. 101 A1 Building properties of the residential building prototype used in the precooling simulations. The prototype was designed by the Pacific Northwest National Laboratory. This table, and information about the development of the prototypes can be found at https://www.energycodes.gov/development/residential/iecc models133 A2 The full set of hourly average emissions factors (AEFs) for CO2 for the California Independent System Operator (CAISO). The AEFs were calculated using 2019 data available for public download on the CAISO website with a regression analysis. Recreated from Zohrabian, Sanders. . . . . . . . . . . . . 134 xiii A3 The R2 values corresponding to each AEF calculated above, which correspond to a coefficient in the binned linear regression model. These values are consistently close to 1.00, with a minimum value of 0.86, which implies a strong correlation between emissions and demand for a given hour h of month m (or equivalently, after grouping by an hour of the day and month of the year pairing, the variations in emissions are highly explainable by variations in demand). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135 A4 Percent change in cumulative peak period cooling electricity consumption, residential electricity cost, and cooling associated CO2 emissions over the fivemonth simulation period between each of the 70 precooling schedules and the baseline schedule of 75 oF. These 70 schedules represent all of the precooling schedules out of the 504 that were simulated that maintain thermal comfort (PMV between -0.5 and +0.5) for occupants during the peak period. The precooling hours refers to the number of hours before 5pm that precooling occurred, the depth refers to the number of degrees Fahrenheit below 75 that AC was set at during precooling, and the offset refers to the number of degrees Fahrenheit above the baseline temperature that the AC was set at during the peak demand period (5-8pm). . . . . . . . . . . . . . . . . . . . . . . . . . . 145 B1 Summary of major regression-based studies evaluating MEFs for different electric grids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147 B2 Key Properties of Selected ResStock Buildings . . . . . . . . . . . . . . . . . 153 xiv Abstract As grids across the U.S. transition to higher penetrations of variable renewable energy (VRE), there is a need for new strategies designed to change when end-users consume electricity. Strategies that shift the timing of electricity demand fall in the category of demand-side management (DSM) and have primarily been used to reduce peak electricity demand, thus reducing costs, and providing grid reliability benefits. Increasing penetrations of VRE have also created the potential to use DSM to reduce emissions, because the generation resources that are used to meet demand, and their associated emissions, vary throughout the day and year. Researchers often rely on marginal emissions factors (MEFs) to determine the impact of changes in demand on emissions, but existing methods of calculating MEFs fail to properly account for modern, high-renewable grids and lack the properties that would make them most useful for DSM. This body of work introduces a new methodology for calculating MEFs that incorporates both renewable generation sources and electricity trades between regions, and results in estimates that have higher temporal resolution, are more accurate, and can be forecasted, improving significantly upon existing statistical methods. These emissions factors quantify the effect that implementing a DSM strategy would have on emissions and can therefore inform DSM program design. This dissertation also applies MEFs by examining the impact of precooling (a DSM strategy that shifts AC load) on singlefamily homes across California to evaluate precooling as an emissions-reduction strategy. The results show that precooling can effectively reduce emissions, in addition to providing traditional DSM management benefits, highlighting the potential of MEFs to be used to create lower-emissions consumption patterns. xv Chapter 1: Introduction 1.1 Motivation 00:00 03:00 06:00 09:00 12:00 15:00 18:00 21:00 Hour of The Day 0 5000 10000 15000 20000 Net Demand (MW) April 24th, 2023 Figure 1: Net demand for electricity in the CAISO region on April 24th, 2023. Net demand is equal to demand minus the supply of wind and solar power. Data from CAISO (1). Altering the timing of electricity consumption is known as load-shifting. This strategy is part of the field of demand-side management (DSM), which focuses on the way end-users consume electricity (2). DSM programs have traditionally focused on reducing electricity consumption during periods of high demand to ensure grid reliability, as well as reducing demand when the cost to generate electricity is high, thus saving utilities and other electricity providers money (3). These benefits can be passed on to consumers to incentivize participation in the programs via dynamic pricing for electricity, such as time-of-use (TOU) rates or real-time pricing (4). A less explored benefit of DSM, and load-shifting specifically, 1 is the potential to reduce CO2 emissions. Using DSM to reduce emissions requires accurate estimations of the effect on emissions that a specific strategy will have. The impact of altering the demand for electricity on emissions depends on the generation resources being used to meet demand at a certain time, and, more specifically, on the generators that respond to changes in demand. This work introduces a framework for quantifying the impact that a change in demand will have on emissions; this framework can be used to develop new DSM strategies and programs that focus on decarbonization. 1.2 Summary of research gaps and the contributions of this work This body of work calculates the emissions impact of precooling, a strategy that shifts airconditioning consumption from one time of day to another. To date, the potential emissions benefits of precooling have been largely unexplored, but the results of this work demonstrate that precooling, and load-shifting strategies more broadly, can reduce emissions, and that this impact can be quantified with emissions factors. The analysis of precooling also informs the sections of this dissertation devoted to methods of calculating marginal emissions factors (MEFs). MEFs describe the relationship between changes in demand and changes in emissions, but current methods used to calculate these factors have many limitations that reduce their applicability to DSM. This dissertation introduces new methods of calculating MEFs that address these gaps. The contributions of this collection of studies to precooling and MEFs are summarized in the following sections. 1.2.1 Quantifying the impacts of residential precooling Precooling is a cooling strategy that features intense use of building air-conditioning (AC) for a period of time in order to reduce the need for AC at a later time, effectively shifting some of the electricity required for cooling by a few hours. Precooling has typically been evaluated as a method to reduce peak-demand for electricity and as a cost-saving mechanism 2 under TOU electricity pricing, with minimal attention given to precooling as an emissionsreduction strategy. This gap in the literature can be addressed through the application of emissions factors. My work determines the efficacy of precooling at reducing the CO2 emissions associated with AC use and evaluates if these benefits can be delivered in congruence with the more traditional DSM benefits of reduced peak demand and cost savings for utility customers. I use EnergyPlus, a building simulation software, to simulate residential homes under different precooling schedules and analyze the resulting electricity consumption to determine the impact on CO2 emissions, residential electricity costs, and peak period electricity consumption. This dissertation begins in Chapter 2 with a proof-of-concept analysis of precooling that uses average emissions factors and features one single-family home design, and then expands the scope and applicability of precooling in Chapter 4 by using MEFs and evaluating multiple single-family home designs across all 16 California climate zones. These analyses also motivate improvements to the methods of calculating MEFs because implementing precooling relies on accurate, high-temporal-resolution MEFs that are known in advance. 1.2.2 Improving the estimation of marginal emissions factors Common methods of calculating MEFs are not tailored to DSM applications. These methods describe the relationship between emissions and electricity generation (not demand), fail to account for electricity trades between regions, and create inaccurate estimates on grids with high renewable energy penetrations. My work in Chapter 3 creates a novel, data-driven framework for calculating MEFs that incorporates both trades and renewable generation, and better isolates the impact of demand on emissions. Additionally, existing methods calculate historical MEFs and have limited temporal resolution, making it difficult to apply the resulting MEFs to DSM planning that aims to take advantage of future variations in emissions intensity. My research in Chapter 5 also addresses this gap by introducing a novel composite model made up of a multi-layer perceptron model that feeds into a linear model 3 to predict changes in emissions. The MEFs calculated with this new method have hourly granularity, increased accuracy, and can be forecasted 24 hours in advance, making them ideal for demand-side management applications. 1.3 Research Questions The central research questions answered in this dissertation are: 1. Can precooling offer emissions reductions benefits in addition to reducing residential electricity costs and peak demand? 2. Can we develop new frameworks to calculate demand-based MEFs that are appropriate for DSM applications within high variable renewable energy grids? 3. How do building and climate zone characteristics influence the CO2 emissions, peak energy, and economic tradeoffs of precooling? 4. Can we forecast MEFs with high granularity and accuracy, increasing their utility for DSM applications? 1.4 Structure of this document and resulting publications to date This document is organized into 6 Chapters, with Chapters 2, 3, 4, and 5 corresponding to the peer-reviewed articles listed below. Chapters 2 through 5 also answer research questions 1 through 4 respectively. Chapter 6 summarizes and concludes this body of work. • Chapter 2: Mayes, Stepp, and Kelly Sanders. “Quantifying the electricity, CO2 emissions, and economic tradeoffs of precooling strategies for a single-family home in Southern California.” Environmental Research: Infrastructure and Sustainability 2.2 (2022): 025001. 4 • Chapter 3: Zohrabian, Angineh, Stepp Mayes, and Kelly T. Sanders. “A data-driven framework for quantifying consumption-based monthly and hourly marginal emissions factors.” Journal of Cleaner Production 396 (2023): 136296. • Chapter 4: Mayes, Stepp, Tong Zhang, and Kelly T. Sanders. “Residential precooling on a high-solar grid: impacts on CO2 emissions, peak period demand, and electricity costs across California.” Environmental Research: Energy 1.1 (2023): 015001. • Chapter 5: Mayes, Stepp, Nicholas Klein, and Kelly T. Sanders. “Using Neural Networks to Forecast Marginal Emissions Factors: A CAISO Case Study.” Journal of Cleaner Production (in review) 5 Chapter 2: Quantifying the electricity, CO2 emissions, and economic tradeoffs of precooling strategies for a single-family home in Southern California The content included in this Chapter is published in: Mayes, S., & Sanders, K. (2022). Quantifying the electricity, CO2 emissions, and economic tradeoffs of precooling strategies for a single-family home in Southern California. Environmental Research: Infrastructure and Sustainability, 2(2), 025001. 2.1 Introduction Increasing penetrations of variable renewable power generators can challenge reliable grid management, particularly in electricity grids where there is a temporal mismatch between peak renewable energy generation and peak electricity demand. Managing the large ramping requirements needed to quickly bring generation resources online to transition from periods of low net demand (which often align with high renewable energy penetration) to periods of high net demand can be difficult (5). (Note: here we define net demand as total demand minus variable renewable energy generation (6).) This phenomenon is already evident in California’s Independent System Operator (CAISO) territory, where high penetrations of solar power have created challenging grid conditions commonly referred to as the ’duck curve’. The duck curve refers to CAISO’s diurnal net load curve, which is characterized by (1) a deep daytime net load dip (i.e., the “belly” of the duck) caused by high solar photovoltaic penetration and relatively low demand, and (2) a steep increase in net load that occurs as the sun sets in the early evening and CAISO simultaneously enters its evening peak load period (7). To accommodate the large generation ramping requirements needed to meet rising demand at the time of diminishing solar resources, fast-reacting (but expensive, inefficient, and 6 polluting) natural gas generators are brought online to replace solar generators. Additionally, CAISO increases the amount of electricity imported from neighboring states, which in general have a dirtier grid mix than CAISO, and therefore contribute significantly to total emissions. Thus, during the middle of the day, electricity within CAISO tends to be cheap with a low greenhouse gas emissions intensity on average, whereas in the evening, when electricity demand tends to peak, the average grid mix is much more polluting and expensive (8). The rapid increase in net demand seen in the late afternoon (see Figure 2) and early evening also creates a risk of demand exceeding supply, and CAISO has reacted by creating its “Flex Alert” electricity conservation program (9). Flex alert programs attempt to reduce the temporal mismatches between electricity supply and demand by calling on electricity consumers to voluntarily reduce or shift the timing of their electricity consumption to reduce demand during periods when there is a risk of undergeneration or excessive stress on the electric grid. During a Flex Alert, CAISO recommends that users increase their thermostat setpoints, avoid using appliances, and minimize the use of lighting. In the window of time preceding a Flex Alert, CAISO recommends running necessary appliances, closing window coverings, and increasing AC usage in order to “precool” their homes so that less energy is needed during the Flex Alert (10). As illustrated with CAISO’s Flex Alert guidance, the concept of precooling a home has been proposed in previous literature as a strategy to shift the timing of electricity consumption (11; 12). Precooling involves aggressively cooling a building for a period of time in order to reduce the cooling needed during a later time when electricity might be more expensive. In grids with large variable renewable energy penetrations, it can also have a greenhouse gas reduction benefit by moving demand away from periods of high fossil fuel generation and towards periods of high renewable energy generation. However, there are physical factors that can limit precooling as an effective demand response or load-levelling strategy. Precooling relies on a building’s ’thermal mass’, or ability to maintain its temperature, a property that 7 Figure 2: Total CAISO electricity supply mix (top row), net demand (middle row), and electricity CO2 emissions intensity (bottom row) for June 2019 (left) and 20 June 2019 (right). 8 varies according to the materials used in the envelope of the building, the volume of air in the building, and the objects inside (13). Thus, the efficacy of precooling a building depends largely on the magnitude of its thermal mass (14; 15). Adjusting the timing of residential cooling to mitigate the challenges posed by grids with high penetrations of daytime solar and an evening peak demand is attractive for a variety of reasons. First, residential cooling represents a significant and growing fraction of overall load. In 2019, the residential sector was responsible for 41% of US electricity consumption (16), with air conditioning representing roughly 16% of total US residential sector electricity usage (17). Second, unlike loads like lighting, the timing of home cooling (or heating) can be slightly decoupled from resident occupancy due to the structure’s thermal mass, which allows it to maintain the temperature of cooled (or warmed) air over time. Finally, growing AC usage presents its own challenges for peak load management, maintaining grid reliability, and mitigating greenhouse gas emissions from the grid, making net load flattening across the day an important priority to reduce the need for expensive generation investments (18; 19; 20). Precooling studies in the literature can be classified by their focus on either residential or commercial buildings, and by the analytical approach used to determine the effectiveness of precooling, including experimental studies with physical buildings, direct thermal models, or building simulation engines that incorporate thermal models. Many residential sector studies have examined the impacts of increasing cooling in the middle of the day to reduce cooling needs in the late afternoon and evening, and they have consistently found that this type of precooling is effective for shifting the timing of electricity consumption. These studies often use building simulation programs, such as Department of Energy’s (DOE) EnergyPlus software (12; 21; 22; 23; 24; 25). German et al (21) used a coupled methodology of field testing and EnergyPlus simulation of a high-performance house (i.e., improved wall and window insulation, reduced infiltration). They found that aggressive precooling strategies can eliminate almost all peak demand period cooling needs in the high-performance house and significantly reduce peak demand period cooling in a more standard house, although 9 this decrease in peak electricity consumption was accompanied by a small increase in total electricity consumption. Cole et al (23) similarly simulated a typical home in EnergyPlus and found that an optimal precooling strategy could reduce peak demand period consumption by over 50%, but at the expense of an 11% increase in total daily electricity consumption. Turner et al (26) used REGCAP, a building simulation tool similar to EnergyPlus, to study a typical residential building and concluded that precooling in the residential sector could eliminate more than half of peak demand period cooling electricity consumption at the cost of increasing total electricity consumption. Their results indicate that peak demand period cooling could be nearly eliminated with a penalty of a 67% increase in total cooling load electricity consumption. Other precooling studies in the residential sector have used direct thermal models (27) or experimental studies to analyze existing buildings (28; 29). Herter Energy Research Solutions examined multiple precooling schedules in 152 occupied homes on eight distinct days and found that a six-hour precool at a temperature 3 °F lower than a baseline temperature could reduce peak demand period cooling electricity consumption by 43%, with a slight increase in total energy consumption (30). Furthermore, this precooling schedule was rated as the most comfortable by the building occupants, even when compared to a business-asusual cooling schedule. These results (13; 21; 23; 26; 27; 28) suggest that precooling can significantly reduce peak demand period electricity consumption in the residential sector, but there is typically a slight to moderate increase in total electricity consumption caused by the increased cooling load in the afternoon (typically a warm part of the day), with the magnitude of these changes in consumption depending on the specifics of the precooling schedule, building location, and building thermal mass. Strengers (31) observes in a study of dynamic peak pricing in Australia that utilities can benefit twice from precooling in the form of reduced peak demand and increased total demand. In the commercial sector, studies focus on overnight precooling with the goal of reducing cooling during the day in order to reduce total cooling electricity consumption, total cool10 ing electricity costs, or both (32; 33; 34; 35; 36; 37). Overnight precooling shifts a portion of daytime cooling consumption to nighttime, when lower outdoor temperatures increase cooling efficiency and utilities often offer a reduced electricity price per kWh. The results of these studies consistently show that daytime cooling can be effectively reduced with an overnight precool, and a larger fraction of commercial (compared to residential) precooling studies found that reductions in both the target period cooling electricity consumption (daytime) and total electricity consumption is possible (32; 34; 35). These results conflict with residential sector studies because commercial buildings typically have larger thermal mass, and overnight precooling strategies leverage cooler outdoor temperatures, making it possible to reduce both total and peak demand period cooling consumption with a single precooling strategy. Xu et al (34) experimentally studied two large office buildings located in a hot California climate on specific days in August and September and found that overnight precooling can reduce the total cooling load by 20%–30% on hot days without a significant reduction in occupant comfort. Morgan and Krarti (33) simulated a prototypical office building with three distinct levels of thermal mass in several climate zones and found that in specific zones a reduction in daytime cooling energy consumption of 31% was possible with a penalty in total energy consumption, and reductions in daytime cooling of up to 20% could be achieved without a total consumption penalty. While previous studies have consistently found that residential precooling can reduce peak period electricity consumption, this reduction usually comes at the expense of increasing total electricity consumption. In grids with high fossil fuel generation, there is strong coupling between total power sector emissions and total electricity usage, but as grids transition to high penetrations of carbon free generators, this relationship becomes much weaker. In other words, grid fleets (i.e., the fleet of generators operating at any given time to meet electricity demand) vary over hours, days, seasons, and years, so the emissions intensity of consuming a unit of electricity also can vary a great deal across a given time period (see Figure 2). Even though increasing renewable energy generation will accelerate the decoupling trend 11 between total electricity consumption and total greenhouse gas emissions, little attention has been directed to how precooling could be leveraged to simultaneously reduce peak period electricity consumption and total daily CO2 emissions. The high penetration of renewable energy in CAISO provides an ideal environment to explore this concept since there are times when grid emissions are already somewhat decoupled from electricity consumption. In this study, we evaluate the environmental and economic tradeoffs of aggressively cooling a residential building during times of high solar penetration to relieve strain on CAISO during peak demand hours when solar penetration is low. Specifically, we quantify changes in peak electricity consumption and CO2 emissions for a set of 504 unique precooling schedules for a generic home located in California Climate Zone 9, an area surrounding Los Angeles and characterized by warm to hot summers and generally mild winters (38). We also calculate changes in residential electricity costs under a residential time-of-use pricing scheme offered by the electricity supply company Southern California Edison (SCE) operating within CAISO (39). Thus, this study explores the CO2 emissions tradeoffs of traditional precooling strategies, filling a knowledge gap and providing insight as to how precooling can be used as a load-levelling strategy while also incurring a reduction in CO2 emissions in high renewable energy penetration environments. 2.2 Methodology This study analyzes the potential efficacy of residential precooling strategies in reducing the carbon footprint of electricity consumed for cooling within a single-family home located in Southern California. The home was simulated in EnergyPlus, with 505 distinct cooling schedules, including one ‘baseline simulation’ that featured a constant temperature AC setpoint of 75 °F. The remainder of the cooling schedules were precooling strategies, each varying the temporal ‘length’ and temperature ‘depth’ of precooling prior to the defined peak demand period beginning at 5 pm (see Table 1). 12 Table 1: Description of precooling schedule periods and definition of the key terms used. Cooling schedule period Time of day Setpoint characteristics Modeling definitions used to describe precooling scenarios Precooling Defined time interval occurring before peak demand period; varies by scenario AC setpoint < 75 °F ’Length’ of precool: varied from 1 to 6 h in duration ’Depth’ of precool: number of degrees that the thermostat setpoint is set below the baseline temperature; varied from 1 to 12 °F below 75 °F Peak demand Defined as 5 to 8 pm for all simulations AC setpoint > 75 °F Temperature ’offset’: number of degrees the thermostat setpoint is set above 75 °F during the peak demand period; varied from 0 to 6 °F (i.e., operates as a maximum temperature constraint) Normal operation Defined as any hour outside of precooling and peak demand periods AC setpoint = 75 °F for all scenarios Setpoint of 75 °F, the baseline temperature assumed in this analysis 2.2.1 Simulation Program Using building simulation programs makes it possible to examine a large variety of building characteristics and precooling schedules across different climate zones and seasonal time frames, which would not be feasible in experimental studies. EnergyPlus is a free building simulation tool funded by the DOE used here to model the energy consumption of each air conditioning schedule. EnergyPlus has been validated and verified in DOE funded studies as well as by independent research teams with both comparisons to other building simulation programs as well as to empirical data (40; 41; 42; 43; 44). It uses algorithms based on heat and mass transfer between zones to perform a detailed heat balance at a sub-hourly level based on user inputs (45). Its ability to meet a wide variety of building structure and system specifications (46) make it an ideal choice for this study. EnergyPlus requires two inputs: a building data file describing the building’s physical characteristics and loads, and a weather data file. In this study, a building prototype representing an average U.S. single family home was used. This residential home prototype was developed by the Pacific Northwest National Laboratory (PNNL) based on the 2018 version 13 of the International Energy Conservation Code (47). The prototype is a two-story, detached building that is 2376 square feet with 15% window area (measured relative to conditioned floor area) and includes four options for heating technology (electric resistance, gas furnace, oil furnace, heat pump) and four options for foundation (slab, crawlspace, heated basement, unheated basement). We selected this prototype because it was originally developed to quantify the energy consumption impacts of changes in codes and standards, and its options regarding heating, ventilation, and air conditioning (HVAC) technology and scheduling can be easily modified (47; 48; 49), making it an appropriate choice for determining the energetic and thermal comfort tradeoffs of different AC operation schedules. We modeled a house with a slab foundation due to its prevalence in the Pacific region (in 2013, 55% of new singlefamily homes built in this region were built on a slab, down from over 65% in 2005 (50; 51)). We selected the natural gas furnace option for space heating, which represented over half of all homes that used heating equipment in the Pacific region in 2015 (52). (Although there is currently a push towards space heating electrification in California, this analysis focuses on warm months when space heating is unlikely to impact results, regardless of technology choice.) All residential prototypes developed by PNNL feature central AC, and all other thermal, physical, and operational house properties were left unchanged from PNNL’s default assumptions, apart from the home’s thermostat setpoint schedule, which governs hourly AC and heating operations. These schedules were specified manually as described in more detail below. A full list of building properties can be found in Table A1. The weather file used (53) describes typical conditions for California climate zone 9, which includes Los Angeles as its reference city (54). This weather file represents the average weather in the Los Angeles area which was found by averaging weather conditions over a time-span of multiple years, and was created by the California Energy Commission for the purpose of determining compliance with California Building Energy Efficiency Standards (title 24) (53). While EnergyPlus can estimate many output variables, those utilized in this analysis include hourly electricity consumption, cooling-specific electricity consumption, indoor air 14 temperature, and indoor relative humidity. 2.2.2 Simulation Details The EnergyPlus model described in Figure 3 was run for one reference case and 504 unique precooling schedule simulations, with each simulation spanning five months in duration. The five warmest months, as measured by average daily high temperature (Jun–Oct), for a typical year in Los Angeles were selected for analysis, as this period is when the majority of cooling is expected to occur (55). The baseline simulation assumed a constant thermostat setpoint temperature of 75 °F (the baseline temperature assumed throughout this study) for the entire five month simulation period and was used as the reference case to compare all precooling schedule simulation results. EnergyPlus PNNL Residential Prototype CA CZ-9 Weather File Thermostat Schedules (1 Baseline, 504 Precooling) Outputs Does not meet PMV comfort constraints (Discarded) PMV Check Hourly CO2 emissions Hourly Residential Electricity Costs EnergyPlus Simulation Hourly Average Emissions Factors Time-of-use Rates Filtering Post-Processing Hourly electricity Meets PMV comfort constraints Figure 3: Overview of our modeling framework. Each precooling schedule simulation partitioned each 24 h day into three unique periods designated by a specific thermostat temperature setpoint (applied during the precooling and normal hour periods) or maximum temperature constraint (applied during the peak demand period). The three periods defining each daily precooling simulation schedule are described in Table 1. The 5 to 8 pm peak demand period was defined to reflect the window of time that California’s Investor-Owned Utilities typically assign the most expensive time-of-use rates in efforts to dissuade electricity consumption during CAISO’s peak demand period. 15 Identical scheduling assumptions were applied for each day across the total period of a respective simulation, regardless of month or season. By varying the AC setpoints during the precooling and peak demand periods and varying the length of the precooling period, we create a search space that encompasses a large diversity of precooling schedules. It should be underscored that the scheduling periods selected in this study were selected primarily to reduce peak demand electricity consumption and secondarily to align precooling periods with renewable energy availability. Thus, the purpose of the study is not to minimize greenhouse gas emissions for cooling explicitly, but rather to evaluate the tradeoffs among several variables related to the magnitude and timing of residential electricity consumption (Figure 4). 2.2.3 Thermal Comfort Thermal comfort was quantified according to predicted mean vote (PMV), as defined by the American Society of Heating, Refrigerating, and Air-Conditioning Engineers (ASHRAE). PMV is a non-linear function of dry bulb temperature, mean radiant temperature, air speed, relative humidity, metabolic rate, and clothing insulation, and describes the relative comfort of building occupants. A PMV of +0.5 (or −0.5) corresponds to a prediction that 10% of occupants will report themselves as uncomfortable (note that at a PMV of 0.0, 5% of occupants are still expected to report themselves as uncomfortable) (56). Calculations of PMV were done in Python with the pythermal comfort model (57), with the variables air speed, metabolic rate, and clothing insulation held at constant values (0.1 m s−1 , 1.2, and 0.5 respectively), and other measured hourly inputs provided by the EnergyPlus simulation. We used ASHRAE’s defined comfort range of PMV values, which span −0.5 to +0.5 (56). (Note: the AC setpoint in the baseline simulation was selected to give occupants a close to neutral thermal comfort level during afternoon and evening hours, with the best integer value found to be 75 °F.) All 504 precooling schedules were simulated and any schedules that created uncomfortable 16 0 0.2 0.4 0.6 0.8 1 1.2 1.4 69 70 71 72 73 74 75 76 77 78 79 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 Cooling Electricity [kWh] Temperature [°F] Cooling Setpoint [F] Cooling E- [kWh] Normal Operation Period (constant setpoint) Precooling Peak Demand Offset Length Depth Figure 4: A sample thermostat setpoint schedule illustrating the simulation periods defined in Table 1, with corresponding hourly electricity usage dedicated to air conditioning. For this schedule, the length of the precool is 3 h, the depth is 3 °F, and the offset is 3 °F. The AC setpoint (blue line) is lower than the baseline temperature in the middle of the day and higher than the baseline temperature during the peak demand period. The corresponding hourly AC electricity consumption (orange line) peaks in the middle of the afternoon (coinciding hours that have a relatively low average emissions factor, as shown in Figure 2), and is lower during the evening hours (when the average emissions factor is relatively high during the diurnal period). conditions (per the PMV indicator) during the 5 to 8 pm peak demand period on any day in the simulation period were removed, leaving 70 distinct schedules that met the comfort conditions. While maintaining a PMV between −0.5 and 0.5 during the peak demand period was set as a criterion for the inclusion of a specific precooling schedule, PMV was allowed to go below this range during the precooling period. Comfortable conditions were generally maintained throughout the entire day for shallow precooling schedules, but deep precools created uncomfortable conditions for occupants before 5 pm during the precooling period. We assume that occupants are either out of their home prior to this peak demand period or, as active participants in precooling, can adjust accordingly (e.g., wearing warmer clothes). 17 Thermal comfort is also not considered overnight. (While the normal operation AC setpoint should ensure that occupants are not too warm, occupant comfort could also be affected by the natural gas heating setpoint, which is beyond the scope of this study.) 2.2.4 Output Analysis The key EnergyPlus output analyzed was hourly AC electricity consumption, which was used to quantify changes in peak demand period electricity consumption, total daily electricity consumption, and residential electricity cost. Hourly AC electricity consumption was also multiplied by average hourly CO2 average emissions factors for the CAISO grid to determine hourly CO2 emissions. Hourly emissions factors were calculated with grid level electricity load and emissions data (including both electricity imports and exports) from CAISO in 2019 (8) using the linear regression model described in equation 1. The model first bins hourly electricity load by month and hour of the day (ex. the bin ’June, hour 1’, has 30 data points) and then regresses the hourly total CO2 emissions (E, in kg of CO2) on hourly total electricity load (D, in MWh) within each bin. The resulting coefficient D is defined as the average CO2 emissions factor (AEF, in kgCO2/MWh); a total of 288 coefficients were calculated with each coefficient representing a specific month and hour of the day. Eh,m = AEFh,m · Dh,m ; h = 1, 2, ..., 24 ; m = 1, 2, ..., 12 (1) Here, the subscripts h and m refer to the hour of the day and month of the year respectively, since the average emissions factor varies throughout the day due to changes in the grid mix. More information about the regression model can be found in Section A.3. The hourly electricity consumption across all precooling schedules was analyzed from a residential electricity cost perspective assuming a time-of-use residential rate plan designed to reduce peak demand period consumption and offered by SCE, a utility within CAISO, servicing about 15 million people in Southern and Central California (39). The SCE rate scheme applied for this analysis has a basic structure $0.52 per kWh from 5 to 8 pm, and 18 $0.26 per kWh at all other times (58). Total homeowner costs were calculated by multiplying the hourly electricity consumption by cost of electricity at that time. 2.3 Results Relative to a constant setpoint scenario, precooling scenarios shift the timing and magnitude of both electricity consumption and the CO2 emissions associated with that electricity consumption. These scenarios also change electricity costs to the consumer. Post-filtering, the 70 remaining precooling schedules were evaluated relative to the baseline simulation by examining changes in electricity consumption, CO2 emissions, and residential electricity costs at the hourly level as well as cumulatively across the five month period. 2.3.1 Understanding the hourly impact of precooling Using the simulation outputs for the month of September, we calculate hourly changes in electricity consumption, CO2 emissions, and residential electricity cost in order to analyze results at sub-daily timescales. (We highlight the month of September due to its high temperatures in the selected climate zone, which provided good conditions for understanding the potential impact of precooling on days when AC loads tend to be highest.) The measured hourly electricity consumption associated with each precooling schedule (or baseline schedule) was averaged for each hour of the day across the 31 days of September to create an average or representative September day for each unique schedule. This consumption was then used to determine the homeowner’s associated electric utility costs and CO2 emissions. We then calculate the difference between a precooling schedule and the baseline schedule to get insight as to thermal behavior of the house over the course of the day and the resulting benefits and consequences. The shallow and deeper precooling scenarios illustrated in Figure 5 both shift peak demand period electricity consumption to other hours of the day, first increasing consumption during the precooling period, then reducing consumption during the peak demand period, 19 Precooling Period Peak Demand Period Shallow Precool Deeper Precool Figure 5: Top: the changes in average hourly electricity consumption for a shallow precool (4 h length, 1 °F depth, 3 °F offset) scenario, as compared to the 75 °F baseline simulation, are shown in blue for an average day in the month of September. Bottom: the same is shown for a deeper precool (3 h length, 3 °F depth, 3 °F offset). Hourly PMV is shown in green, which stays between the +0.5 and −0.5 limits for both schedules. and then again increasing consumption after the peak demand period. During the precooling period, the lower temperature setpoint causes an increase in AC output and a resulting increase in electricity consumption, with a larger increase the deeper the precool. During the peak demand period, when offset is occurring, the higher temperature setpoint reduces AC output and thus decreases electricity consumption. Deeper and longer precools will reduce AC consumption more during the peak demand period for a specific level of offset, but progressive increases in precooling depth or length offer progressively less reduction. In other words, there are diminishing returns for precooling when the goal is reduced peak period electricity consumption. After the peak demand period, returning the house to the baseline temperature causes a spike in electricity consumption as the AC turns on to reduce the indoor temperature back to the baseline setpoint. This increase could be mitigated by improving natural airflow, such as by opening windows, on nights when the outdoor tem20 perature is below the post-8 pm setpoint of 75 °F. However, this study focuses solely on air-conditioning operation, and maintains the same window positions throughout the day for both the baseline cooling schedule and the precooling schedules. The magnitude of this spike is commensurate with the magnitude of the offset temperature reached at the end of the peak demand period (i.e., a 3 °F offset requires the AC to reduce the temperature by 3 °F starting at the end of the peak demand period). Regarding thermal comfort, occupants might experience a slightly cool sensation during precooling as the indoor air temperature is below the neutral-sensation temperature. For shallower precools, this sensation remains in the comfortable range, with PMVs above ASHRAE’s lower bound of −0.5. However, deeper precools (3 °F or more) create indoor air temperatures significantly below the baseline temperature of 75 °F during the precooling, leading to PMVs below −0.5. Many of these deeper precools also created uncomfortable conditions during the peak demand period, and thus were eliminated in the PMV filtering stage, but a small number of precools with a depth of 3 °F created uncomfortable conditions during the precooling period but not during the peak demand period, and thus are included in this analysis. During the peak demand period, occupants would experience a slightly warm sensation due to the elevated indoor temperature that is proportional on the amount of offset, but all schedules included in these results maintain PMV levels between 0.0 and +0.5, the maximum permitted level. After the peak demand period, the PMV level approaches 0.0, or a neutral sensation, as the house returns to the baseline temperature. 21 Precooling Period Peak Demand Period Shallow Precool Deeper Precool Figure 6: Top: the changes in average hourly CO2 emissions between a shallow precool (4 h length, 1 °F depth) and the 75 °F baseline simulation are shown in blue for the month of September. Bottom: the same is shown for a deeper precool (3 h length, 3 °F depth). Hourly PMV is shown in green. These precooling scenarios shift the magnitude and timing of when CO2 emissions (Figure 6) and residential electricity costs (Figure 7) occur by moving them from the peak demand period to normal operation hours, following the same general temporal trends as changes in electricity consumption. These changes are scaled by the hourly AEFs for CO2 emissions, and by the hourly price of a kWh of electricity for residential electricity costs. For example, the large difference between peak demand period electricity price and off period electricity price (factor of 2x) creates larger reductions (relative to the baseline) in residential electricity cost during the peak demand period than increases in cost during the precooling period. It is important to note that the AEFs and the price structure serve as hourly multipliers, making it possible for reductions in cumulative CO2 emissions or residential electricity costs, even if total electricity consumption is unchanged or even increased. With a better understanding of the hourly impacts of precooling, we move next to an analysis of the aggregate results 22 Precooling Period Peak Demand Period Shallow Precool Deeper Precool Figure 7: Top: the change in average hourly residential electricity costs for a household for a shallow precool (4 h length, 1 °F depth) compared to the 75 °F baseline simulation is shown in blue for the month of September. Bottom: the same is shown for a deeper precool (3 h length, 3 °F depth). Hourly PMV is shown in green. over a cooling season. 2.3.2 Cumulative impact of precooling The precooling schedules were also analyzed over the entire five-month period from May through October by summing the hourly outputs for each hour in this period to get aggregate results for the AC electricity consumption, CO2 emissions, and residential electricity cost, respectively. The data points lying below the dashed horizontal axis of Figure 8 represent precooling strategies that are effective in simultaneously reducing CO2 emissions, peak demand period electricity consumption, and residential electricity costs. While all the precooling schedules simulated reduce peak demand period electricity consumption, the schedules that achieved the most reduction feature large amounts of offset, deep precools, and, to a lesser extent, 23 long precools. Increasing the depth of the precool or the amount of offset increases the gap between the initial temperature of the house during the peak demand period and the temperature at which the AC cycles back on. This leads to a longer period during which the AC is inactive as the house warms from the precooling period temperature to the offset period temperature. Increasing the length of the precool reduces peak demand period consumption due to the increased removal of heat from portions of the building that contribute to thermal mass beyond the air itself. In other words, the longer the precool, the more heat is removed from the internal walls and furniture, thus slowing the rate of indoor air temperature increase when the precooling period ends. However, increasing in the length of a precool provides a much smaller reduction in peak demand period consumption than increasing the depth of the precool or the amount of offset. In summary, schedules with deep precools and large offsets create conditions in which the AC operates for only a small portion of the peak demand period, effectively reducing peak demand period consumption, and this reduction in AC operation can be reduced slightly more with longer precools that help to keep the house cool during the peak demand period. The precools that reduce peak demand period consumption the most achieve a reduction of around 70%. Over half of the simulated schedules increase CO2 emissions, but some schedules achieve small reductions in emissions relative to the baseline. Schedules that reduce CO2 emissions generally feature 2 °F or less of depth, with the most effective schedules being those that have large amounts of offset, and only 1 °F of depth. Precools deeper than 1 °F often have the effect of increasing total electricity consumption, despite decreasing peak demand period consumption. CO2 emissions reductions are possible for these schedules that increase total electricity consumption since consumption is shifted from the peak demand period, when there are higher AEFs, to the precooling period, when there are lower AEFs, but these results suggest that for most deeper precools, the increase in electricity consumption is significant enough that the difference between precooling period AEFs and peak demand period AEFs cannot compensate. Instead, the shallow precooling schedules that are most effective in 24 PC Depth 3F 2F 1F 6H PC Length 1H 0°F Offset 1°F Offset 2°F Offset 3°F Offset +2% -13% Change in Resident Cost Figure 8: Each dot represents one of the 70 distinct precooling schedules that passed the PMV comfort test. The x-axis shows percent change in aggregate peak demand period consumption and the y-axis shows the percent change in CO2 emissions between the cooling schedule and the baseline simulation over the five-month period. The color of each dot represents the percent change in residential electricity cost over this same period. Cooling schedules are illustrated by the black bars and groupings, with the width of the bar proportional to the length of the precool, the height of the bar proportional to the depth of the precool, and the offset shown by each grouping. reducing emissions feature a large offset that reduces the peak demand period consumption when AEFs are high but do not significantly increase total electricity consumption. The most effective schedules at decreasing CO2 emissions achieve reductions of approximately 3%–3.5%. The schedules that achieve residential cost savings tend to overlap with the schedules that accomplish significant CO2 emissions reductions. A higher fraction of precooling schedules reduce residential electricity costs than reduce emissions because of the large difference between the on- and off-peak pricing (factor of 2x) schedule considered in this analysis. This makes it possible for a precooling schedule to significantly reduce residential costs despite increasing total electricity consumption. This result implies that deeper precools would need 25 precooling period emissions factors to be less than 50% of peak demand period emissions factors in order for deeper precools to successfully reduce emissions, which is not the case for the AEFs used in this study. The effect of this large price multiplier is also seen in the magnitude of cost reductions, with the optimal precooling schedule reducing residential electricity bills by nearly 13%, outpacing the CO2 emissions reductions. -9% -13% Change in Resident Cost 3F Offset PC Depth PC Length 6H 1H 1F 3F Figure 9: Includes subset of 3 °F offset scenarios from Figure 8 The schedule that reduced peak period cooling electricity most is shown with purple bars, and the schedule that reduced peak period cooling electricity consumption and CO2 emissions most is shown with orange bars. The color scale for residential cost is maintained from Figure 8. The precooling schedules that simultaneously reduced peak demand period consumption, CO2 emissions, and residential cost feature large offsets of 2 to 3 °F, and, in general, use shallower precools of varying length; the most successful schedules feature 1 °F of precooling for 1–5 h. This type of schedule has enough offset to significantly reduce peak demand period consumption but does not feature the deep precools that significantly increase total electricity consumption and make reductions in CO2 emissions impossible. Figure 9 shows the subset of 3 °F offset schedules from Figure 8, with the schedule that reduced peak period 26 electricity consumption most and the schedule that simultaneously reduced both cost and CO2 emissions most highlighted in purple and orange, respectively. The ten schedules that most reduced CO2 emissions are included in the table below, with full results of all the schedules included in Table A4. Table 2: The ten schedules most effective at reducing CO2 emissions and the associated changes in CO2 emissions, cost, and peak demand period consumption. Offset PC hours PC Depth Change in peak period electricity electricity % Change in CO2 emissions (%) Change in residential cost (%) 3 2 1 -57 -3.5 -13 3 1 1 -55 -3.0 -12 3 4 1 -57 -2.9 -12 3 5 1 -58 -2.6 -12 3 3 1 -57 -2.5 -12 2 2 1 -42 -2.5 -9.0 2 0 0 -33 -2.5 -8.1 3 6 1 -58 -2,3 -11 2 1 1 -40 -2.0 -8.5 3 1 2 -61 -2.0 -12 Table 2, row 7, shows that a significant amount of CO2 and cost reductions can be achieved with the use of an offset in the absence of any precooling (i.e. constant setpoint of 75 °F in all hours before the peak demand period) due to the reduced AC operation during 27 the peak demand period. Introducing a shallow precooling schedule, in combination with this offset, significantly increases the reduction in peak demand period electricity consumption, and the presence of the hourly AEFs and cost profile make it possible for this to simultaneously provide slightly bigger reductions in both CO2 emissions and cost. Therefore, under the selected SCE utility pricing schedule and with emissions calculated using hourly AEFs, shallow-precools outperform the strategy of simply using a higher AC setpoint during peak demand period hours. 28 2.4 Discussion Several simplifications or assumptions were necessary to complete the modeling phase of this study. For example, the building prototype used for this model, though well vetted by PNNL to represent an average building at the national level, may not be as representative of homes in California, and in Southern California specifically, even with the most common foundation and heating/cooling technology selected. California homes are smaller than the prototype used here on average. This simulation assumes active AC setpoints for all hours of the day. In reality, residents who are not home in the middle of the day may turn off their ACs entirely while out of the house. This study does not attempt to compare precooling results to other intermittent AC usage schedules, only to a constant setpoint schedule. Some deep precooling schedules also reduced indoor temperatures below the comfortable range during the precooling period, which limits the time that occupant comfort can be ensured. Residents would need to either accept this risk of discomfort and adapt accordingly (e.g., with warmer clothing), or be absent during the precooling period. In reality, individuals will have unique tolerances for feeling ‘too warm’ versus ‘too cold’; the PMV metric used here only expresses average populationlevel comfort preferences. While the specific implementation of precooling is beyond the scope of this paper, the increasing penetration of smart homes, appliances, and thermostats could make precooling easier for homeowners via preset thermostat schedules. Lastly, this study relied on hourly average emissions factors for the CAISO. Average emissions factors represent the current mix of sources supplying power to the grid in a time period, but a more accurate estimation of the emissions associated with shifting the timing of electricity consumption may come from the use of marginal emissions factors. In this case, CO2 emissions would be calculated by multiplying the hourly change in electricity consumption with hourly marginal emissions factors. While there is not currently an established set of marginal emissions factors for CAISO, these factors could be estimated with a model similar to the regression model used in this study, but instead of regressing total hourly CO2 29 emissions on total hourly changes in demand, one would regress changes in hourly CO2 emissions on changes in hourly demand. Using marginal, instead of average, emissions factors may increase the amount of calculated CO2 emissions reductions, potentially making deeper precools more beneficial, as marginal emissions factors are particularly high during the grid ramping that occurs in early evening since natural gas combustion generators are typically at the margin (59). 2.5 Conclusion This study used EnergyPlus to evaluate the efficacy of residential precooling as a strategy to reduce CO2 emissions in grids that have periods of high variable renewable energy penetrations in addition to providing the demand response and load-levelling benefits that are traditionally associated with precooling. Our results suggest that in CAISO, precooling can achieve multi-faceted benefits in terms of flattening CAISO’s net demand curve as well as reducing CO2 emissions, residential utility costs, and peak demand electricity usage, even at the expense of an overall increase in total residential electricity use. Relative to a constant AC setpoint, simultaneous reductions of 57% of peak demand period consumption, 3.5% of CO2 emissions, and 13% of residential cost is possible to achieve when precooling a typical home in California CZ9 over the five-month period studied in this analysis. Key insights from this study include: • CO2 emissions benefits were realized in many precooling schedules despite increases in total electricity demand due to the daytime decoupling between electricity consumption and greenhouse gas emissions within CAISO, a grid with high penetrations of daytime variable renewable energy resources. • Achieving larger reductions in peak demand period electricity consumption typically required larger total electricity consumption penalties. • While deep precools reduced peak demand period consumption more than shallow 30 precooling schedules, they often caused an increase in total CO2 emissions due to large increases in total electricity demand. • Shallow precools could significantly reduce peak demand period consumption while also decreasing total CO2 emissions due to more mild increases in total electricity consumption than deep precools. • Residential precooling strategies can also provide economic benefits to the consumers when utility rate structures use higher costs to penalize electricity usage during peak hours The results of this study confirm that the timing of electricity consumption (as opposed to simply the magnitude of electricity consumption) is becoming increasingly important as grids transition to higher penetrations of variable renewable power sources, and can inform the implementation of precooling as a demand response strategy within CAISO. For example, shallow precools of 1–2 °F below the baseline temperature with precooling periods of 1 to 4 h in length can achieve demand response benefits without increasing total CO2 emissions or causing large periods of discomfort. Thus, the real-time variations in grid mix create an opportunity for optimizing precooling as an emissions reduction strategy. This conclusion offers important policy insights, particularly for grids that have high penetrations of daytime variable renewable energy resources. Given the pace of energy transitions around the world, we believe that this study is a significant contribution to the literature for researchers, industry members, and policy makers interested in designing strategies that meet climate mitigation and demand response priorities simultaneously. Whereas our goal was to evaluate multifaceted tradeoffs of precooling for peak electricity demand, the environment, and occupants, future studies could design precooling schedules that have the explicit goal of minimizing cooling-associated CO2 emissions. Minimizing emissions may involve changing the timing of the precooling period, which currently occurs directly before the peak demand period and partially overlaps with the rapid ramping of 31 net demand that occurs in the late afternoon (see Figure 2). Implementing precooling prior to this ramping period, instead of overlapping with it, may offer larger reductions in CO2 emissions due to the rapid increases in average and marginal emissions factors that occur during the ramping period due to the rapid deployment of fast-reacting natural gas generators needed to meet rising demand as solar resources diminish. 32 Chapter 3: A data-driven framework for quantifying consumption-based monthly and hourly marginal emissions factors The content included in this Chapter is published in: Zohrabian, A., Mayes, S., & Sanders, K. T. (2023). A data-driven framework for quantifying consumption-based monthly and hourly marginal emissions factors. Journal of Cleaner Production, 396, 136296. 3.1 Introduction Understanding the interactions between electricity demand and electricity supply is an important first step in quantifying the emissions impacts of load modifying interventions and leveraging demand-side resources for deep decarbonization. In an electric grid, power supply must be balanced with electricity demand at any given time of day. This means that each time there is an increase in electricity demand, there is a commensurate increase in the supply of electricity (and conversely, electricity supply must be reduced if electricity demand decreases). Marginal generators are the power generation units that respond to these changes in demand; altering their output to match the new level of demand. Tracking the operation of marginal generators in the context of the electric grid operation provides key information in greenhouse gas accounting and developing emissions mitigation measures. Traditionally, quantifying emissions of the electric grid focuses on total grid emissions (i.e., the sum of all emissions associated with the electricity supply over a given period of time) or the average emissions factor (i.e., the total grid emissions for a period of time divided by the total demand over the same period). These quantities are often easy to calculate, pending adequate data availability, but estimating marginal emissions at a point in time is not as straightforward because it is typically difficult to accurately identify marginal 33 generators and isolate their emissions. Theoretically, electricity generators are dispatched according to the lowest marginal cost of electricity generation given operational and transmission constraints (60). When compared to renewable electricity generators, fossil fuel-based generators have relatively high operational costs, and therefore, are often dispatched after renewables. Several studies have used comprehensive electric grid simulation models for assessing long-term (implying that future capacity expansion changes the generation capacity mix) marginal emissions associated with electricity generation systems (61) and short-term (implying that the generation capacity mix does not change significantly), and have applied their estimates to different cases such as electric vehicle charging (62; 63; 64), renewable energy integration (65; 66), energy efficiency (65) and energy storage (67). For instance, one study developed a reduced-order power plant dispatch model for the North American Electric Reliability Corporation regions for years 2014 to 2017 and showed that clusters of low- or high-emitting power plants of similar production cost could create large changes in marginal emissions factors (MEFs) as ascending the order of dispatching generators (68). Another study builds upon cost-based dispatch by incorporating ramping limitations and applies this methodology to Great Britain’s power system, finding significant variations in MEF within a day and over the course of a year (69) One limitation of simulation modeling approaches is that actual grid operation is more complicated than most simulations can capture. For example, more than one resource or generation unit can contribute to supply the last unit of electricity, or cheaper generators with limited generation capacity, such as hydropower, may choose to produce more electricity during more profitable times of the day. Khan (70) found that oil was the only consistent marginal resource in Bangladesh, but in New Zealand, a study found that peak demands were primarily met by hydropower (71). Li and colleagues (72) found that for the Midcontinent Independent System Operator (MISO) in the years spanning 2015 through 2018, fossil-fuel burning generators were generally the marginal resource, but it also found that at high demand times, hydropower also contributed to marginal generation in the MISO Central subregional grid, while wind energy contributed 34 to marginal generation in the MISO North subregional grid (which had a high penetration of wind energy, about 27 % of total electricity generation). Electricity imports and exports dynamics are among other factors that further complicate electric grid operation beyond what models can capture. Alternatively, regression models are used to track short-term marginal emissions (typically on the scale of hour-to-hour) and are generally more accurate than simulated values obtained from electricity system models (68). Regression-based models rely primarily on historical granular electricity generation, consumption, and greenhouse gas emissions data to estimate marginal emissions. One advantage of using observed historical data to estimate marginal emissions is that the data capture grid operation constraints as they occurred, in contrast to some simplifying assumptions in electric grid models that might ignore generator outages or transmission congestion constraints. Different imports and electricity purchase structures and load serving obligations can also be hard to accurately simulate in electric grid models. Other constraints, such as maintaining grid stability and reliability, might result in deviations from loading orders and/or lifting or enforcing certain environmental rules, which are hard to predict and capture in electric grid simulations that are suited to mostly capture the grid’s normal operations (73). Another advantage is the low complexity of regression models (as compared to electricity system models), which makes it easier to interpret and validate results. Different regression models have been previously used to analyze marginal emissions of Great Britain’s electric grid by (74), India’s electric grid by (75), Ontario’s electricity system in Canada by (76), Pennsylvania, Jersey, Maryland Power Pool (PJM) in the US by (77), MISO by (78), and North Electric Reliability Council (NERC) regions in the US by (79). These studies utilize various statistical models with a range of different independent variables. Some studies used a simple linear regression model with hourly changes in generation as the independent variable (79; 64; 74). In one of the earliest MEF quantifying studies, Siler-Evans and colleagues (79) assumed that changes in electricity demand are only balanced by changes in the output of fossil fuel-based generators, ignoring the role of 35 changes in production from non-fossil fuel powered generators in meeting marginal demand. Seckinger and Radgen (80) similarly assume that all marginal demand is met with fossil fuel technologies when calculating MEFs for the German grid, citing the prioritizing of renewables as justification for this structure. Although this might be an adequate assumption for grids with low penetrations of renewable energy, or with strict resource queuing rules, many regional grids incorporate significant levels of alternative sources, and emissions free generators are increasingly on the margin. In fact, specific loading orders can also facilitate the marginal operation of renewables, regardless of their low production cost advantages (e.g., California’s 2003 Energy Action Plan that prioritizes renewables over fossil fuel generation (81). Some studies have included electricity generation from variable energy resources, as well as levels of electricity demand as predicting variables that marginal emissions depend on (67; 82; 76). A number of these studies are summarized in Supplemental information Table B.1. Growing shares of battery storage deployments will also create new grid operation dynamics. It has yet to be explored how different charging and discharging patterns can impact marginal generators and emissions. Regardless of the statistical model and predicting variables used, all previous studies reported that MEF trends are distinct among seasons and hours of the day, and can also be significantly different from the concurrent average emissions factors (AEFs) due to factors such as electricity demand patterns, electricity generation fleet, the legacy of technology mix, fuel type, operational cost, dispatchability and grid interconnectedness (65; 74; 60). Geographic boundaries are critical for accurate emissions accounting (83). Previous studies have typically taken a generation-based approach for calculating changes in emissions and generation by considering generators located within their region of focus (79; 74). This approach assumes that changes in electricity demand are equal to changes in electricity supply in that same geographic area, ignoring the influence of electricity trades with external areas. Electricity trades were modeled by Tranberg et al. (84), who assessed real-time grid-wide average emissions in European electricity markets, and by De Chalendar et al. (85) , who 36 analyzed annual and median daily emissions (not marginal emissions) for 66 electricity balancing authorities across the US. The latter study showed that exchanges between regions play an especially large role in the Western Interconnection (where California Independent System Operator or CAISO is located), since, as an example, 2016 net imports accounted for 29 % of annual consumption by net importing regions and 2016 net exports accounted for 37 % of annual generation in net exporting regions (85). While electricity trades have been modeled in a few average grid emissions studied, they are widely absent in regression-based MEF assessment studies. The large variation in these factors among different electric grids highlight the need for evaluating regional specific MEFs. Moreover, rapid shifts in electricity supply mix make it essential to frequently re-assess MEF estimates. In this study, MEFs of CAISO , a grid that generated 20 % of its annual electricity consumption from solar PV and wind turbines in 2019 (and 21 % in 2020), are evaluated. The analysis is done independently in each year and examine changes that occur between the two years. This comparison provides insight as to the impact of supply factors, such as the use of hydropower, as well as the effect of changes in electricity demand patterns on MEFs. 3.2 Contributions of This Study Our framework refines three major aspects of the regression models previously used in MEF assessment literature through its approach in considering electricity trades and variable renewable energy generation. These improvements increase the ability to methodically quantify the efficacy of demand-side management (DSM) strategies for reducing emissions, since DSM affects the subset of generators at the margin as opposed to the whole set of generators across the grid. 1. This framework takes a consumption-based approach, rather than a generation-based approach (i.e., emissions associated with electricity consumption in CAISO are accounted for regardless of whether electricity is produced with in-region resources or 37 imported from out-of-region). In addition, electricity generated with in-region resources and exported out-of-region is not considered for in-region electricity demand. The emissions associated with these exports are removed from the total emissions before emissions factor calculations. 2. The regression model used in this study mimics a net load curve to explicitly account for generation from emissions-free non-dispatchable solar PV and wind turbines, which is particularly important for analyzing grids with high levels of renewable energy penetration. Specifically, the model includes a designated term for variable renewable energy in the regression model, an inclusion lacking in previous US-based studies. 3. Lastly, this framework also makes methodological improvements that account for MEFdemand dependency and enables the analysis of more granular emissions dynamics and careful quantification of hourly and monthly MEFs. 3.3 Methods This section describes the methodology developed to use historical data for estimating hourly average emissions factors and marginal emissions factors (defined in the next subsection) for CAISO independently for two years of 2019 and 2020. 3.3.1 Definitions Here, the following definitions are used for quantifying emissions: Average Emissions Factor (AEF): This metric quantifies the emissions associated with the average unit of electricity consumed by accounting for emissions from all electricity generating units in the CAISO region and net electricity imports. In this study, AEFs are calculated using periods of length one-hour. Marginal Emissions Factor (MEF): This metric shows the change in emissions due to one unit change in electricity demand, as changes in demand impact marginal generators 38 Table 3: Electricity supply sources in the CAISO region (total electricity supplied was 219.5 TWh in 2019 and 218.5 TWh in 2020). Electricity supply data from CAISO (86), and the breakdown of net imports from EIA (87). 2019 2020 Total Supply (Sum to 100 %) Fossil fuel (mostly natural gas) 29 % 33 % Nuclear 7 % 7 % Hydro 12 % 6 % Solar PV and Wind turbines 20 % 21 % Other Renewables 7 % 7 % Net Imports 24 % 26 % Breakdown of Net Imports from BAs to CAISO (Sum to 100 %) Los Angeles Dep. Water & Power (LDWP) 35 % 31 % Bonneville Power Administration (BPAT) 13 % 26 % Salt River Project (SRP) 25 % 19 % Arizona Public Service Company (AZPS) 19 % 11 % Balancing Authority of Northern California (BANC) 6 % 6 % Nevada Power Company (NEVP) -1 % 5 % Imperial Irrigation District (IID) 5 % 4 % Western Area Power Administration - Desert Southwest Region (WALC) 2 % 2 % PacifiCorp West (PACW) <1 % <1 % Turlock Irrigation District (TIDC) -2 % -3 % Centro Nacional de Control de Energia (CEN), and Comision Federal de Electricidad (CFE) (in Mexico) -1 % -1 % rather than all generators. In this study, MEFs are calculated using hour-to-hour changes in demand, variable renewable generation, and CO2 emissions. 3.3.2 The System Boundaries This study is bounded around the CAISO region located within the Western Electricity Coordinating Council. Based on CAISO data (86), in 2019, about 76 % of total annual electricity (219.5 TWh) was generated within the region (74 % in 2020); the remainder was imported from multiple balancing authorities (BAs) in the Southwest and Northwest regions which are listed in Table 3. This electricity was generated from a mix of technologies that use natural gas, nuclear, hydro, and renewables, which included large fractions of solar PV and wind technologies (20-21 % of total supply). 39 3.3.3 Data Sources and Data Processing Steps The utilized data sources and their corresponding main processing steps are as follows: 1. The list of power plants operating in each BA in each year were identified based on the US Energy Information Administration (EIA) Form EIA-860 reports (88) by filtering power plant identification numbers for the two sectors of electric utility and independent power producers (IPP) for both CHP (combined heat and power) and non-CHP plants. 2. Hourly emissions data were extracted from the US Environmental Protection Agency’s (EPA) Air Markets Program Data for each state, and were rearranged for each BA based on the list of power plants identified within each BA in the previous step (89). (Note: EPA’s Air Markets Program Data reports emissions associated for fossil-fueled power plants with capacities greater than 25 MW.) 3. Hourly data for electricity generation at BA level as well as electricity exchanges between BAs were collected from EIA’s Electric System Operating Data (87). Note that because these data are bidirectional between BAs, each electricity exchange is reported in two locations. Total electricity generation data were collected from each corresponding BA file (for example, total electricity generation of LDWP was extracted from LDWP file). The only exception was BPAT where historical data reported on the BPAT website (90) were used instead of EIA’s data (87) due to the large discrepancies in the reported EIA values. 4. Data for CAISO’s electricity demand, total electricity generation and solar PV and wind turbine generation, as well as total imported electricity were sourced from the CAISO website (86). These data are reported in five-minute increments, but this study uses averaged values to represent each hour. Several considerations and adjustments were made when cleaning and processing data. 40 First, the timestamps for the data associated with the three BAs located within the Arizona time zone (i.e., AZPS, SRP and WALC) were shifted for one hour for the affected data points between November and March of each year, so that the all timestamps are aligned with the Pacific Time zone. Second, electricity trades between CAISO and the two BAs located in Mexico (CEN and CFE) were ignored in this analysis due to lack of associated emissions data. Instead, the magnitude of these electricity trades were distributed among the other BAs within the US in proportion to their concurrent electricity trades with CAISO. (The electricity traded between CAISO and the two BAs in Mexico was only about 57 GWh, or 1 %, of total net imports to CAISO in 2020, as shown in Table 3.) Third, CAISO’s reported total net imports were cross-checked with the sum of electricity trades between all of the individual BAs and CAISO to ensure that the two values were equal. When the sum of hourly electricity trades between CAISO and other BAs did not match CAISO’s reported total net electricity imports in a given hour, CAISO’s total net import value was used to normalize the interchange electricity amounts with each BA (scaling the magnitude of the exchanges with each BA to ensure that the two totals matched). Fourth, for the bidirectional exchanges reported by the EIA, the values reported from CAISO’s perspective were primarily used, and data from other BAs were only used to verify or fill in missing or misreported electricity interchanges. If hourly data were missing from both sources, five-day averaged values for the same hour were used. Fifth, it was ensured that total electricity demand for CAISO was balanced with the sum of total electricity generation and total net imports in each hour. Finally, before beginning the regression process, an outlier analysis was performed to identify data points that were likely the result of incorrectly reported emissions values by BAs, or exacerbated by limitations in the data processing methodology (e.g., a small error in emissions reporting could be magnified by the scaling process that matches BA-reported imports with CAISO-reported imports). More details on the analysis can be found in Supplemental Information. 41 3.3.4 Consumption-Based Hourly CO2 Emissions Estimates Hourly emissions associated with CAISO’s electricity demand were tracked using Equation 2, which explicitly accounts for the emissions associated with electricity imports and exports. In this equation, E C i,j , E X i,j , and E I i,j are emissions associated with in-CAISO power generation, exports (from CAISO, C, to other BAs, M) and imports (from other BAs, M, to CAISO, C), and i and j are indexes for the month and the hour of day. Additionally, in this equation, XC→M GC calculates the fraction of emissions associated with the electricity produced in CAISO (GC) but consumed in other BAs; while IM→C GM calculates the fraction of emissions associated with the electricity produced outside CAISO (GM) but consumed in the CAISO region. It is notable that this accounting method for CO2 emissions is an improvement over the methodology used in reporting CO2 emissions in five-minute increments on the CAISO Today’s Outlook website (91) where emissions are approximated using resourcespecific CO2 emissions rate for in-CAISO generators and a fixed unspecified emissions rate (i.e., 0.428 mTCO2/MWh as established by California Air Resources Board) for all imported and exported electricity (92). Ei,j = E C i,j − E X i,j + E I i,j = E C i,j − ( X M E C · XC→M GC )i,j + (X M E M · I M→C GM )i,j (2) 3.3.5 The Averaging Method for Estimating AEFs For AEF calculations, the emissions data were first grouped by month j and hour of the day i (e.g., “January, hour 1” has 31 data points), and then derived a regression model based on Equation 3 to calculate an average CO2 emissions factor (AEFi,j ) using electricity demand (Di,j ) as a predicting variable for each hour (i) and month (j) pairing. Twenty-four hourly AEFs were calculated in kg CO2/MWh for each month (e.g., 24 AEFs in January and 288 AEFs in 2019), which represent an average day in that month. 42 Ei,j = AEFi,j · Di,j (3) 3.3.6 The Regression Model for Estimating MEFs Following the steps illustrated in Figure 10, the CAISO data were first re-ordered from lowest to highest hourly electricity demand to form the load duration curve for each year for both years of analysis. Then, electricity demand was partitioned into 10 equal bins, where each bin represented 10 % of the range between the lowest and the highest demand (shown in the load duration subplot in Figure 10). This binning method allowed quantifying MEFs for hours with similar demand level (in contrast to AEFs, where binning was done by time of day, so similar hours of different days in a month were grouped together). This binning method results in a variable number of data points in each of these 10 bins. For instance, Bins 7-10 contained a smaller number of hours compared to Bin 2-4, as these bins covered the highest demand hours that occur infrequently throughout the year (see the steep slope area in the load duration curve in Figure 10), but they might have more significance to the regression slope due to larger supply needs during these hours. Considering the very small number of hours in Bin 10, (e.g., only 33 hours in year 2020), Bins 9 and 10 were combined prior to the regression step of the analysis. Further combination of bins was performed ad hoc during the regression process to ensure a minimum number of data points for each regression (described in more details later in this section). It is notable that because the bins are defined at the annual level, some months might not have certain demand levels; for example, demand levels that fall in Bin 9 and Bin 10 only occur during summer months. Secondly, after assigning bins to each hour, the differences between consecutive hours were calculated for emissions (∆Ek,j ) in kg, electricity demand (∆Dk,j ) in MW h, and the sum of solar PV and wind generation values (∆Rk,j ) in MW h. These variables describe the changes between the j th and j − 1 th hour and have a label of k that specifies the bin their demand level falls in. This step is illustrated in the table presented in the center of Figure 43 10. Thirdly, the multiple linear regression model shown in Equation 4 was applied for each demand level and month combination to estimate MEF values. In the event that a demand level and month combination only occurred in a small number of hours, demand level bins were combined with an adjacent bin until a minimum threshold of 25 hours was reached. In this regression model, ak,j estimates MEF and bk,j estimates the impact of solar PV and wind generation (both in kgCO2/MW h) for each bin k and month j. ∆Ek,j = ak,j · ∆Dk,j + bk,j · ∆Rk,j + ck,j (4) Detailed regression results for each month and demand level combination can be found in the data repository mentioned at the end of this document. Finally, MEF values were converted from a specific month/bin pairing to month/hour pairing by multiplying the month and bin-specific MEF values by weight factors wi,k that are associated with the histogram of bins in each hour of each month (see Equation 5). In other words, hourly MEFs are found with a weighted sum, where the weight is dependent on the bin distribution for a specific hour in a specific month, and these weights are multiplied by the MEFs for those month/bin pairings. The resulting hourly MEFs are obtained in the format of ”month-hour” in which 24 MEFs represent an average day in each month. For illustration, weight factors that had hourly electricity demand in Bins 2-4 are displayed at the bottom-right of Figure 10 for January 2020. As the plot suggests, the majority of hours had demand levels concentrated within a single demand bin (e.g., 48 % of electricity demand in hour 5 of January 2020 was within the demand range of Bin 2 and 52 % within Bin 3; whereas, electricity demands in hours 2-4 were entirely within the demand range of Bin 2). MEFi,j = X k (wi,k · ak,j ) (5) 44 1) Partitioning electricity demand to 10 bins using the annual load duration curve, and assigning a bin number to each hourly record (from 1 to 10) 2) Calculating hourly changes in electricity demand (D), solar + wind generation (R), and CO2 emissions (E) 3) Estimating regressionbased MEF values for similar demand bins of each month (Equation 2) 4) Calculating hourly MEF values by multiplying monthly specific MEF values of step (3) by weight factors associated with the histogram of demand bins in each hour of that month (Equations 3) Bin # Demand (GW) Bin # Demand (GW) 1 15.6 <D < 18.7 6 31.3 <D < 34.4 2 18.7 <D < 21.9 7 34.4 <D < 37.6 3 21.9 <D < 25.0 8 37.6 <D < 40.7 4 25.0 <D < 28.1 9 40.7 <D < 43.8 5 28.1 <D < 31.3 10 43.4 <D < 47.0 Month () Hour of month D Bin () R E Regression for each demand bin: ∆, = ,Δ, + ,∆, + , ∆, ∆, ∆, 1 1 d1 2 r 1 e1 1 2 d2 1 r 2 e2 e2 -e1 d2 -d1 r 2 -r 1 1 3 d3 1 r 3 e3 e3 -e2 d3 -d2 r 3 -r 2 1 4 d4 1 r 4 e4 e4 -e3 d4 -d3 r 4 -r 3 1 5 d5 2 r 5 e5 e5 -e4 d5 -d4 r 5 -r 4 1 6 d6 2 r 6 e6 e6 -e5 d6 -d5 r 6 -r 5 1 7 d7 3 r 7 e7 e7 -e6 d7 -d6 r 7 -r 6 1 … … … … … … … … 1 744 d744 2 r 744 e744 e744-e743 d744-d743 r 744-r 743 Duration in hours Bin 10 Bin 9 Bin 8 Bin 7 Bin 6 Bin 5 Bin 4 Bin 3 Bin 2 Bin 1 45.5 35.5 25.5 15.5 Hour of day 0 2000 4000 6000 8000 0 0.2 0.4 0.6 0.8 1 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 Hour of Day Bin 4 Bin 3 Bin 2 Example distribution of demand bins for each hour of the day in January 2020 Figure 10: Visual representation of the proposed framework to evaluate MEF values 3.4 Results In this section, the regression results for MEF values are presented for each month and demand level in 2019 and 2020, as well as the estimated month and hour level MEF and AEF values. Additional evidence from the operation of the electric grid was provided to support the responsiveness of different grid resources to changes in electricity demand which help validating the estimated MEF trends. In general, the results are consistent with the range of AEFs and MEFs estimated in other studies for CAISO. The coefficient of determination (R2 ) for the regression results lies between 0.40 and 0.98 for the distinct regressions, with better predictability in higher demand levels. From a practical standpoint, having higher accuracy of MEF estimation in higher demand levels is most important since periods with high MEFs (normally coincident with 45 high demands) have the greatest implications for emissions reductions and increases due to changes in demand (see Supplemental Information for more details). It is worth noting that the model shows a better performance in 2019 as compared to 2020, likely due to data quality issues and the larger spread of the data that occurred in 2020. As other studies reported, electricity demand patterns were impacted by COVID-19 pandemic-related lockdowns, such as lower consumption levels than previous years especially in March and April months (93), and sector-specific changes in loads, such as increased residential sector consumption in the middle of the day (94). 3.4.1 Correlation Between Emissions and Demand The hourly data distribution in both 2019 and 2020 suggests a strong linear correlation between hourly demand and hourly CO2 emissions, as shown in Figure 11 subplots a and b. As demand goes up, less spread is seen in the magnitude of hourly emissions and data points than for lower demand, implying a tighter correlation between emissions and demand for higher demand values. At higher levels of demand, the emissions produced also tend to lie above the line of best fit, which suggests that as demand increases, the emissions per unit of demand also increase. This is consistent with the common operational practice of bringing fast-reacting and dirty electricity natural gas combustion units online to meet the highest levels of demand when the range of the typical resources used to meet demand is exceeded. In 2019, there was an average of 266 kgCO2 generated per MWh, and in 2020 this number rose to 310 kgCO2 per MWh, with these values equal to the slopes of the lines of best fit for the top two graphs of Figure 11. 46 15000 20000 25000 30000 35000 40000 45000 50000 Demand (MWh) 0 2000 4000 6000 8000 10000 12000 14000 Emissions (thousands of kgCO2) Subplot A (2019) 15000 20000 25000 30000 35000 40000 45000 50000 Demand (MWh) 0 2000 4000 6000 8000 10000 12000 14000 Emissions (thousands of kgCO2) Subplot B (2020) 6000 4000 2000 0 2000 4000 6000 Change in Demand (MWh) 2000 1000 0 1000 2000 Change in Emissions (thousands of kgCO2) Q2 Q1 Q3 Q4 Subplot C (2019) 6000 4000 2000 0 2000 4000 6000 Change in Demand (MWh) 2000 1000 0 1000 2000 Change in Emissions (thousands of kgCO2) Q2 Q1 Q3 Q4 Subplot D (2020) Figure 11: Scatter plots of hourly CO2 emissions and hourly demand for electricity in 2019 (Subplot A) and 2020 (Subplot B), as well scatter plots of changes between consecutive hours in hourly emissions and hourly demand between in 2019 (Subplot C) and 2020 (Subplot D). The line for best fit is shown in yellow, with R2=0.35 (Subplot A) and 0.50 (Subplot B) for hourly emissions and demand, and R2=0.40 (Subplot C) and 0.39 (Subplot D) for hourly changes in emissions and demand. Figure 11 also examines the correlation between changes in emissions and changes in demand calculated as the difference between two consecutive hours for each variable (subplots C and D). In 2019 there was an average emission of 322 kgCO2 per MWh of marginal electricity generated and in 2020 this number fell to 308 kgCO2, with these values equal to the slopes of the lines of best fit for subplots C and D. The magnitude of changes in demand in 2019 cover a wider range compared to 2020, highlighting higher hour to hour variations in load, while changes in emissions remained approximately in the same range in both years. Overall, there is a wider variation in the range of changes in emissions when the magnitude of demand changes are positive and large (Q1) than when the magnitude of demand changes 47 are large but negative (Q3) for subplots C and D. Additionally, a significant number of hours in Q4 of subplots C and D have an increase in demand on the scale of GWh and decrease in emissions on the scale of thousands of metric tons. These events, and more generally, the large deviations from the lines of best fit, emphasize that changes in demand alone cannot accurately predict changes in emissions and that a multiple linear regression model capable of capturing the influence of the varying supply of renewable energy is necessary. 3.4.2 MEFs by Demand Level in Each Month Our data binning and regression model resulted in a total of 67, for 2019, and 61, for 2020, independent MEF values ( one for each distinct demand bin and month pairing in that year). These values are shown in Figure 12, and the plotted trend lines show a positive correlation between MEF and demand level. The higher MEFs for high demand levels are caused by emissions-intensive marginal generators. The relationship between demand level and marginal generation also results in seasonal variations in MEFs. Summer months, which typically have higher demand levels due to increased electricity usage for air conditioning show relatively high MEFs when compared with cooler months. While high demand levels are centralized around summer months, lower demand levels occur across many months and have a wider range of corresponding MEFs than higher demand levels. The slope of the line of best fit was steeper in 2020 than in 2019, meaning that MEFs were more responsive to demand level in 2020. The causes of this shift are explored further in the discussion section. Figures displaying MEFs organized by month instead of demand level can be found in Supplemental Information. 48 1 2 3 4 5 6 7 8 9 Demand Level 0 100 200 300 400 500 600 MEF in kgCO2/MWh 2019 Month Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec 1 2 3 4 5 6 7 8 9 Demand Level 0 100 200 300 400 500 600 MEF in kgCO2/MWh 2020 Month Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Figure 12: MEFs for each demand level and month combination in 2019 and 2020 with line of best fit. 3.4.3 Month-Hour MEFs Demand-level based MEFs, like those shown in Figure 12, can be directly applied in situations in which demand level can be estimated; however, they do not provide information about diurnal MEF patterns, which are important for predicting the changes in emissions associated with changes in electricity consumption that occur at specific times of the day. To address this need, month-hourly MEFs were derived using the relationship between time-of-day and demand. The results are shown in the form of heat maps in Figure 13. The MEF values in 2019 and 2020 are consistent and similar in terms of hours when the highest and lowest MEF values are concentrated (i.e., highest in the evening of summer months and lowest in the morning of spring months). However, there are two significant differences between the MEFs in 2019 and 2020 that are worth highlighting. First, the highest MEF values reached in 2020 are substantially higher than those in 2019; i.e., nearly 500 kg CO2/MWh in evening hours in July 2020 compared to roughly 370 kgCO2/MWh in August 2019. Second, lower MEFs during the evening hours of spring months in 2020 (dropping as low as 100 kgCO2/MWh in the evening in March) show a contrast to 2019, where MEFs are around 330 kgCO2/MWh at similar periods. In fact, the MEFs calculated for the evening of March 2020 are significantly lower than the lowest MEFs at any point in 2019. These differences in MEFs imply that different resources were operating at the margin in the two studied years. 49 These differences were investigated in further detail in the discussion section by examining the electricity generation fleet in March and July of each respective year. Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Month 0 2 4 6 8 10 12 14 16 18 20 22 Hour 2019 AEFs 100 200 300 400 500 AEF in kgCO2/MWh Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Month 0 2 4 6 8 10 12 14 16 18 20 22 Hour 2020 AEFs 100 200 300 400 500 AEF in kgCO2/MWh Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Month 0 2 4 6 8 10 12 14 16 18 20 22 Hour 2019 MEFs 100 200 300 400 500 MEF in kgCO2/MWh Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Month 0 2 4 6 8 10 12 14 16 18 20 22 Hour 2020 MEFs 100 200 300 400 500 MEF in kgCO2/MWh Figure 13: Month-hour distribution of AEFs (top) and MEFs (bottom) in 2019 and 2020. Colors represent the magnitude of emissions factor in kgCO2/MWh. 3.4.4 Month-Hour AEFs The temporally based AEFs show strong diurnal and seasonal trends in Figure 13. From a seasonal perspective, the AEFs, which are representative of the average grid mix, reach their lowest values in the spring when there is a high availability of clean power sources like solar, wind, and hydropower, as well as moderate demand levels across CAISO. The AEFs are significantly higher in the late summer/fall months when electricity consumption is much higher, as well as in winter months when supplies from solar resources are limited. AEFs tend to be lower in the middle of the day due to the availability of solar power and the relatively low level of demand. In most months, AEFs increase in evening hours as demand 50 increases and solar PV comes offline causing a larger fraction of the load to be met with natural gas generators and imports. 3.5 Discussion In this section, some important aspects of the observed trends in MEFs and AEFs are discussed. 3.5.1 Consumption-Based Versus Generation-Based MEFs Regression-based hourly-level MEF estimates are rare in literature; however, hourly MEF values have been periodically reported by the Center for Climate, Energy, and Environmental Decision Making (95) for various regional aggregations using a generation-based method. The consumption-based month-hour MEFs in 2019 were compared to MEF estimates reported by CEDM for CAISO in year 2018, the most recent year that MEFs were reported (95). (Note that CEDM used a methodology similar to (79) for estimating MEFs that, as discussed earlier, relied on hourly changes in fossil fuel generation as the single variable for predicting changes in emissions.) The comparison shows that differences between the estimated MEFs and CEDM’s estimated MEFs in the same month and hour ranged from -15 % to 126 % (42 % average difference and 39 % median difference), and the estimates were lower in value in 99 % of hours. While some year-to-year variation could explain these differences, lower MEF estimates across the majority of hours in the study are expected given the differences in methodology in terms of marginal generators. As shown in Figure 13, the hourly changes in demand were typically larger than the hourly changes in natural gas generation, requiring other supply resources such as hydropower, imports, and, in some cases, renewables to respond to changes in demand (CEDM’s MEF methodology assumes all marginal generation is met by in-region fossil fuel plants). While many renewable sources are first-to-take, Figure 13 shows that emissions-free hydropower generation can exhibit strong load-following behaviour in evening hours. Additional demand in the evening is met by a 51 mix of imports, natural gas, and hydropower, resulting in MEFs that can be significantly different and often lower than those calculated using in-region fossil fuel generation changes alone. Additionally, in terms of applications, the use of demand change as a marginal emissions estimator is advantageous over fossil-fuel generation change for a couple of reasons. First, having knowledge of the changes at a specific time in the electric grid’s fossil fuel generation is much more data-intensive and complicated than knowing about the changes in electricity demand for a region or balancing authority. Secondly, the MEF values that simply represent CO2 emissions change per unit of demand change are more ideal to quantify marginal emissions changes associated with end-use demand changes than MEFs derived from fossil fuel generation change. (Fossil fuel generation change would be indirectly correlated with end-use demand changes, which may be highly uncertain in many hours of the year.) Practically, MEFs developed in this analysis can directly be multiplied by measured changes in electricity consumption for any end-use, allowing for more precise and simple monitoring of emissions displacement by demand-side management measures such as load shifting or load shedding. While the MEF estimation exercise relied on historical annual data, applying this methodology with real-time emissions and demand data would enable a close-to-real-time MEF estimation that could be a useful tool to quantify the emissions associated with end-use electricity consumption patterns. 3.5.2 Year-to-Year Variation in MEFs Since regression-derived MEFs rely on historical data, the predictability of MEFs for future years may be limited by how sensitive MEF values are to resource dispatch and fuel mix changes that occur from year to year. Comparing MEF values in 2019 and 2020 helps us answer this question. In fact, a wide range of differences were found in hourly MEFs between 2019 and 2020. (Note: The relative differences between month-hour MEFs in 2019 and 2020 are plotted in Figure B.5 of Supplemental Information.) While summer and winter months 52 in 2020 typically show smaller differences in MEF values between the two years for the same month-hour combination, greater differences are observed in spring months. Despite the fact that both years had similar fossil-fuel and renewable generation shares (shown in Table 3), the role of hydropower seems to be significant in explaining these differences, as hydropower’s share of generation decreased from 12 % of total supplies in 2019 to only 6 % in 2020 (more discussion of this change is provided in Section 3.5.3. The wide range of differences in hourly MEF values suggests that hourly-level MEF values should be evaluated often and that historical-based MEFs require a great caution if used for future years. It is important to understand whether a less granular temporal MEF value would sufficiently represent changes in marginal emissions. To test this, the regression equation was applied to all hour-to-hour changes, regardless of demand level, to calculate a single annual MEF. The result was an annual MEF of 302 kgCO2/MWh in 2019 and 285 kg CO2/MWh in 2020. However, the month-hour MEFs calculated in this study ranged from 169 to 372 kgCO2/MWh in 2019 and from 89 to 503 kgCO2/MWh in 2020. The large range of MEFs that occurred in both years indicate that analyses that use an annual MEF (for example in (96)) for emissions calculations could significantly misrepresent emissions for activities with dynamic temporal patterns, and that being able to capture the diurnal and seasonal trends in MEF values is essential. Regarding average emissions, month-hour AEFs fall in a narrower range of values closer to the annual AEF. For an application in which AEFs are appropriate, using an annual AEF in place of month-hour AEFs for emissions calculations would be less erroneous than making the same simplification for MEFs. 53 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 Hour 0 5000 10000 15000 20000 25000 30000 35000 Load and Resource Generation (MW) Average Day March 2019 Load NG Hydro VRE Imports 2000 3000 4000 5000 6000 7000 8000 Tons of CO2 CO2 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 Hour 0 5000 10000 15000 20000 25000 30000 35000 Load and Resource Generation (MW) Average Day July 2019 2000 3000 4000 5000 6000 7000 8000 Tons of CO2 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 Hour 0 5000 10000 15000 20000 25000 30000 35000 Load and Resource Generation (MW) Average Day March 2020 2000 3000 4000 5000 6000 7000 8000 Tons of CO2 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 Hour 0 5000 10000 15000 20000 25000 30000 35000 Load and Resource Generation (MW) Average Day July 2020 2000 3000 4000 5000 6000 7000 8000 Tons of CO2 Figure 14: Average hourly generation for each resource serving CAISO’s electricity demand in March and July of 2019 and 2020. Data from CAISO (86). 3.5.3 The Influence of Hydropower and Imports on AEFs and MEFs In Figure 15, the changes in generation were compared between consecutive hours for each fuel serving the electricity demand in the months of March and July of 2019 and 2020. This figure identifies which fuels respond more to changes in electricity load as well as diurnal trends in solar and wind availability. It appears that hydropower was more responsive to increases in demand in evening hours of March 2020 than in March 2019 (MEFs of 300-400 kgCO2/MW h in evening hours of 2019 compared to 100-200 kgCO2/MW h in 2020) and was able to reduce CAISO’s reliance on imports and natural gas for marginal generation, despite demand changes being similar across the two years. As a result, hydropower effectively reduced marginal emissions and MEFs during the evening hours of 2020 when compared to 2019 (Figure 13). Despite the limitation of hydropower generation in the year 2020 compared to 2019 (see Table 3), the hydropower dispatch ramp-up in 2020 was complimentary with 54 renewable energy availability and successfully replaced fossil-fuel based marginal generation. This example provides further evidence that energy-limited resources like hydro, if dispatched strategically to offset the need for the dirtiest marginal generators, can help reduce emissions, even in a dry year. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 Hour 4000 3000 2000 1000 0 1000 2000 3000 4000 Change in Load and Resources (MWh/h) Average Day March 2019 Load NG Hydro VRE Imports 1000 500 0 500 1000 Change in Tons of CO2 CO2 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 Hour 4000 3000 2000 1000 0 1000 2000 3000 4000 Change in Load and Resources (MWh/h) Average Day March 2020 1000 500 0 500 1000 Change in Tons of CO2 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 Hour 4000 3000 2000 1000 0 1000 2000 3000 4000 Change in Load and Resources (MWh/h) Average Day July 2019 1000 500 0 500 1000 Change in Tons of CO2 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 Hour 4000 3000 2000 1000 0 1000 2000 3000 4000 Change in Load and Resources (MW/h) Average Day July 2020 1000 500 0 500 1000 Change in Tons of CO2 Figure 15: Average changes in hourly generation between two consecutive hours for each resource serving CAISO’s electricity demand in March and July of 2019 and 2020. Data from CAISO (86). Comparing the hourly generation changes between July 2019 and July 2020, it appears that the increased reliance on natural gas generation (in place of imports) in the early evening hours of 2020 compared to 2019 could have driven higher MEF values in 2020. When high temperatures in July spur increases in electricity consumption, it is often the case that neighboring BAs’ electric energy consumption values are high as well. As a result, imports from neighboring BAs might often be limited when CAISO’s demand reaches its highest levels. While the breakdown of hourly electricity generated by technology (namely, natural gas combined cycles, gas turbines, steam turbines, and internal combustion engines) is not 55 available for generation within CAISO or imports, the observed MEF values of roughly 500 kgCO2/MWh in the early evening hours of July 2020 suggest that generators at the margin could have been natural gas combustion turbines (known as ”peaker” plants), which on average emit about 550 kgCO2/MWh (97). Considering the additional natural gas generation in July 2020 as compared to 2019 and the higher MEFs in 2020, it is likely that natural gas combustion units were more active during the peak hours of July 2020. In terms of the magnitude of the demand changes, electricity consumption increased more aggressively during the afternoon hours in 2020 (an increase of about 1,800 MW in 2020 versus 1,400 MW in 2019 between 3 and 4pm). 3.5.4 Comparing MEFs Versus AEFs Consistent with other studies, the results show that MEFs are significantly different from AEFs in most hours of the year (see Figure 13 and Figure B.6 ). In fact, in late spring and early summer months, the MEF can be nearly three times the magnitude of the AEF. This occurs when there is a high fraction of renewable energy on the grid, which results in lower AEFs, while the last unit of demand is still often met by fossil-fuel generation. While the MEF for a given hour for CAISO is typically higher than the AEF, this is not always the case given the complex dynamics of hourly changes in the fleet mix, which can be met in part by clean resources such as hydropower and or clean imports. For example, as it was explained in the previous section, operational changes such as hydropower generation timing could result in lower MEFs (100-200 kgCO2/MW h) compared to AEFs ( 200-300 kgCO2/MW h) during evening hours of March in 2020. However, hourly AEFs in general were lower in March 2019, in part due to the abundance of hydropower resources thanks to the wet conditions in the state in that year (98). 56 3.5.5 The Influence of Other Interactions Relying on historical data provides holistic context and evidence to understand various trends and driving factors in marginal emissions. However, some interactions can still be refined to provide better representation of regional complexities and dynamics. Although the presented method captures the electricity trades and emissions associated with imports and exports, it is limited in answering questions related to inter-regional influences on MEFs values (e.g., how long-distance renewable exports from CAISO to other regions can effectively reduce CO2 emissions elsewhere. A larger scale regional regression model (for example, WECC-wide) with sub-regional representation is needed to be able to answer these types of questions. Additionally, with rising penetration of grid-scale battery storage technologies, the role of storage for displacing fossil-fuel-based marginal generators should be investigated in future studies. It is worth noting that although the proposed method is well suited to capture historical dynamics of the electric grid, it is less insightful to provide information about future dynamics. Given the fast pace of structural and operational changes in electric grids due to renewable energy adoption, electric power grid modeling may be a more effective option if long-term MEFs are of interest (such as in Gagnon et al. (61)). However, historical data and regression-based MEFs are useful to validate the MEFs calculated through modeling exercises. Over the short-term analysis, although the use of AEFs for DSM emissions quantification is still widespread (99; 100; 101), developing regression-based MEFs is much insightful and necessary, especially for grids with growing penetration of renewables. 3.6 Conclusion In this paper, a proposed multiple linear regression model is used to quantify MEFs at the hourly level, relying on historical hourly emissions, electricity generation and consumption data. This model was applied for CAISO using historical data for 2019 and 2020. This paper’s methodology improves previous MEF estimates by taking a consumption-based ap57 proach that accounts for electricity trades with neighboring regions, as well as including a specific term to account for generation from variable renewable sources (i.e., solar PV and wind). This study shows that capturing these factors is important in grids like CAISO that have high levels of renewable energy penetration and meet considerable fractions of their demand with imports. The proposed method will become increasingly applicable as electric grids across the country incorporate more renewable technologies and aim for around the clock net-zero emissions targets. These methodological changes allow for better isolation of the impact of electricity demand on CO2 emissions and explore the temporal variations in the emissions intensity of marginal demand. The MEFs calculated through the proposed methodology can also be used for evaluating the effectiveness of energy management measures and different grid-connected technologies for reducing emissions. For example, policymakers could use these granular MEFs to facilitate programs that can strategically utilize flexible loads (e.g., electric vehicle charging, heating and cooling, etc.) to reduce demand during the most emissions-intensive hours of the day. Accurate, up-to-date MEFs are an essential step in monitoring emissions and leveraging the timing of electricity consumption to effectively manage and reduce emissions. 58 Chapter 4: Residential precooling on a high-solar grid: Impacts on CO2 emissions, peak period demand, and electricity costs across California The content included in this Chapter is published in: Mayes, S., Zhang, T., & Sanders, K. T. (2023). Residential precooling on a high-solar grid: impacts on CO2 emissions, peak period demand, and electricity costs across California. Environmental Research: Energy, 1(1), 015001. 4.1 Introduction In the building sector, precooling refers to an air-conditioning (AC) operation strategy where the AC is used to overcool a space for a period of time in efforts to reduce cooling needs at a later time. Because precooling can effectively shift some electricity demand from one time period to another by using the thermal storage properties of the building itself, it has gained interest as an attractive application for demand response (DR), a subgroup of demand-side management (DSM) strategies designed to shift the magnitude or timing of electricity demand on the end-user side of electricity meters (102). A key component of DR, compared to other DSM strategies, is its focus on incentivizing end-users to temporarily adjust the timing of their daily electricity consumption from one period to another rather than simply reducing the magnitude of their total consumption (103). Precooling is a particularly attractive DR application in grids with high solar generation, since AC can be used more aggressively when solar resources and daytime temperatures are highest, and then relaxed when more emissions-intensive generation (e.g., natural gas and imports) is needed to be quickly dispatched to replace lost solar generation. As the mix of resources servicing the grid changes over the course of a day, so do the per-unit emissions associated with consuming electricity (104; 105; 106; 107). Solar availability varies 59 diurnally, seasonally, and climatically, and, in the absence of grid-scale storage, electricity from variable resources must be consumed instantaneously or curtailed. Therefore, strategies like precooling that can shift the timing of electricity consumption from fossil-fuel dominated hours to renewable energy dominated hours could theoretically achieve reductions in CO2 emissions by reducing solar curtailment and/or lowering the average emissions intensity of the electricity, even if the total amount of electricity consumed does not change or increases by a moderate amount. As load shifting and demand response strategies become increasingly important for grid operators attempting to manage peak demand, it is important to evaluate the effectiveness of precooling as a peak demand- and cost-reducing strategy for a wide variety of home types and climatic conditions. Growing penetrations of renewable energy in regional grids, like the one overseen by the California Independent System Operator (CAISO, which covers 80% of California’s (108) bulk power transmission), also present an emerging opportunity to compound these benefits through CO2 emissions reductions. CAISO reached a peak 5- minute solar penetration of 72% of total load on a day in 2022 (109) and in 2021 over a third of annual in-state generation in California was sourced from renewable power, with 17.1% coming from solar power (110). Here we use EnergyPlus to simulate the cooling-related energy demand, CO2 emissions, and utility bill costs associated with 480 unique residential precooling schedules for four distinct single-family home designs in each of California’s 16 climate zones (spanning the area covered by CAISO). These four buildings have distinct thermal mass and envelope characteristics and span a wide range of performance levels so that we can evaluate the ability of precooling to deliver simultaneous reductions in residential electricity costs, peak period electricity demand in California, and cooling-associated CO2 emissions, and examine the trade-offs when prioritizing reductions in one of the variables of interest for a variety of conditions. We chose CAISO as a case study because of its high mid-day solar penetration, which makes it an interesting case study to evaluate if precooling is an effective tool for 60 mitigating growing challenges related to peak demand management as high quantities of solar generation are replaced by quick-ramping natural gas generators in the early evening as the sun goes down. This study provides novel analysis on precooling’s ability to reduce emissions on a high solar power grid using marginal emissions factors, which are better suited for demand-side interventions than average emissions factors. 4.2 Literature review The load-shifting potential of precooling varies according to a building’s thermal mass, which is defined as its ability to store energy (111). In the event of a difference between the indoor and outdoor temperature, the thermal inertia of buildings slows the rate at which the indoor temperature approaches the outdoor temperature. Precooling has been studied extensively in the commercial sector for buildings of high thermal mass, and more recently has been explored for lower-mass residential buildings. In the commercial sector, Keeney, Braun (112) performed one of the first precooling studies by precooling a large office building overnight, and achieved reductions in max demand for cooling electricity and operational cost while maintaining thermal comfort. Xu et al. (113) simulated a building in Santa Rosa, CA and found that over 80% of on-peak load could be shifted to an off-peak period with precooling. Xu et al. (114) extended this study using the building simulation software EnergyPlus to explore a larger number of scenarios and confirmed precooling’s ability to shift load for large commercial buildings. Because homes have a limited amount of thermal mass and ability to retain energy compared to larger buildings, precooling must occur in close proximity to the period in which reduced cooling demand is desired, limiting the temporal flexibility of residential precooling. Hence, most residential sector studies explore precooling in the middle of the day or early afternoon to reduce late-afternoon and evening electricity demand, which shifts AC demand away from periods of high grid-level demand and aligns well with typical time-of-use (TOU) rate plans that charge more for electricity during peak demand periods. Numerous 61 studies in the residential sector have confirmed precooling’s ability to reduce electricity bills for customers exposed to TOU electricity rates (115; 116; 117; 118; 119; 120; 121; 122; 123; 30). Wang, Tang, and Song (115) developed a linear programming method capable of determining the optimal precooling schedule for a specific building and determined that cost-optimized schedules could reduce costs by as much as 56% when compared to rulebased cooling schedules. Nelson et al. (119) used EnergyPlus to simulate precooling a building in multiple climate zones and found cost reductions as high as 23.5% were possible in Phoenix, AZ, although the savings in Los Angeles, CA, were found to be just 1.7%. The reduced benefits in Los Angeles may be the result of a more temperate annual climate or a consequence of less aggressive local utility TOU rates. The ability of precooling to reduce demand during a predefined peak period has also consistently been confirmed in the residential sector (115; 116; 117; 118; 119; 122; 123; 30; 124; 125). Cole et al. (122) used a building simulation software to explore the efficacy of precooling a home in Austin, Texas during the months of July and August, during which precooling successfully reduced peak energy consumption by an average of 70%. Turner, Walker, and Roux (125) explored precooling across twelve distinct US climate zones, and found that at least 50% of annual cooling consumption can be shifted away from a peak period in every zone. While precooling has been shown to reduce both residential electricity costs and peak period electricity consumption, a number of factors have been shown to influence the potential reductions, including the building’s thermal mass and envelope characteristics (118; 120; 123; 30; 125; 126; 127; 128; 129), which influence its ability to maintain its temperature, and the location or climate zone the building is located in (116; 118; 119; 123). For example, in (125) all locations realized peak load reductions of greater than 50%, but some climate zones were able to reduce peak load by as much as 99% by precooling, and the increase for total cooling electricity consumption needed to achieve these peak period reductions varied greatly, though building envelope properties were not held constant between locations. German and Hoeschele (118) used EnergyPlus to simulate a building in seven different U.S. cities and 62 found that reductions in peak period electricity consumption varied from near zero to near 50%. Herter Energy Research Solutions (30) tested precooling using occupied residential buildings in the Sacramento, California area and recommended a minimum ceiling R-value of 38 in order to effectively shed evening load while maintaining thermal comfort and minimizing any penalties in total electricity consumption. Several studies have explored modern wall materials and insulation techniques to determine their impact on precooling, including Wijesuriya, Brandt, and Tabares-Velasco (127), who studied phase change material (PCM) enhanced drywall; Kishore et al. (128), who also studied PCM-integrated walls; and Dehwah, Krarti (129) who explored precooling in a home with switchable insulation systems. The results of these residential precooling studies show that precooling’s effectiveness is difficult to generalize because the specific building properties and location need to be considered. While the impact of precooling on peak period electricity consumption and residential electricity bills has been well-researched, the influence of precooling on CO2 emissions has received significantly less attention. Mayes and Sanders (130) simulated a single-family residential building prototype in the Climate Zone that includes Los Angeles, CA, and calculated the resulting CO2 emissions using average emissions factors (AEFs) for the CAISO region. They found that precooling can reduce CO2 by 3 to 4% when compared to a constant setpoint schedule via leveraging the large diurnal variations that exist in CAISO grid’s emissions intensity in daytime versus evening hours. Stopps and Touchie (131) examined precooling for high-rise residential buildings and calculated the impact on greenhouse gas emissions using marginal emissions factors (MEFs) for the grid in Toronto, Canada. Their study showed mixed results for precooling’s ability to reduce peak demand depending on the specific unit in the high-rise, but even for units where peak demand was reduced, they did not find significant or consistent greenhouse gas emissions reductions despite the large diurnal variations in MEFs for this region caused by the varying mix of hydropower and natural gas on the margin. Unlike many residential precooling studies, this study focused on overnight and morning precooling, the times when the local grid’s marginal resources were 63 the cleanest. 4.3 Methodology 4.3.1 Single-Family Home Selection The single-family residential homes used in this study come from the National Renewable Energy Lab’s (NREL) ResStock model (132). NREL developed a large set of residential prototypes designed to represent the spread of buildings present in the residential sector across the country. These prototypes are created via permutations of a large number of building properties, and are designed to be region-specific. In total, the ResStock model has 549,871 residential building designs available for download (133). The four single-family selected from this database contain many common values in order to increase the comparability of results (finished area, no pool, central air conditioning, window types, slab foundation, natural gas heating, etc.) but have differences in key insulation and AC properties. The single-family homes span a range of wall (none to R-19) and ceiling (R-13 to R-49) insulation levels as well as a range of AC efficiency levels (SEER 10 to SEER 13). Key building properties can be found in Table B2 and the full list of building properties can be found in the online data repository (https://data.mendeley.com/datasets/333jfmpvb8/1). The variation in these parameters provides a spread of results that show the impacts of building envelope and AC efficiency on the variables of interest. At opposite ends of the spectrum, Building 1 (B1) represents a poorly insulated single-family home with low AC efficiency and Building 4 (B4) represents a well-insulated single-family home with high AC efficiency. Considering that these prototypes were developed to represent actual homes, we treat Buildings 1 and 4 as representing a reasonable low- and high-end of performance levels for the existing single-family California building stock (with Buildings 2 and 3 falling in-between these extremes). We label these homes B1 through B4 in figures when relevant. 64 4.3.2 Weather Files The weather files used in this study were downloaded from the EnergyPlus website (134) and represent a typical meteorological year (over a 30-year period) for each of the 16 California Building Climate Zones, as specified by the Climate Design Data 2009 ASHRAE Handbook. The climate zones were defined as part of California’s Title 24 Building Energy Efficiency Standards (135). Each zone describes a unique geographic area in California and has its own set of Title 24 requirements for ceiling and wall insulation and window u-values (136). The reference city for each zone and a brief description of its characteristics are included in Table 4. Table 4: Summary of California’s Climate Zones (136) Climate Zone Reference City CDDs Warmest Month (Max, Mean, Min T in ◦F) CZ1 Eureka 15 September (62, 57, 51) CZ2 Napa 500 August (82, 67, 51) CZ3 Oakland 183 September (73, 65, 57) CZ4 San Jose 666 July (82, 70, 58) CZ5 Santa Maria 464 September (74, 63, 51) CZ6 Los Angeles (LAX) 742 August (76, 70, 63) CZ7 San Diego 865 September (62, 58, 51) CZ8 Long Beach 1072 August (83, 73, 62) CZ9 Los Angeles (Downtown) 1456 August (83, 74, 65) CZ10 Riverside 1620 August (94, 78, 60) CZ11 Red Bluff 1354 July (98, 82, 67) CZ12 Stockton 1226 July (95, 77, 59) CZ13 Fresno 1599 July (98, 81, 63) CZ14 Barstow 3056 July (102, 84, 67) CZ15 Brawley 4760 July (107, 91, 76) CZ16 Bishop 596 July (97, 77, 56) These climate zones span a large range of cooling degree-days (CDDs), and hence, a large range of potential cooling needs (CDDs range from climate zones that typically have fewer than 50 annual CDDs to zones that have 4,000 to 6,000 annual CDDs). 65 4.3.3 Precooling Schedules Precooling schedules typically feature a time period where the AC setpoint is below a baseline temperature, closely followed by a period where the AC setpoint is above the baseline temperature. In this study we define and refer to specific precooling schedules with the terms specified in Table 5 and illustrated in Figure 16. Table 5: Definitions of precooling-related terms used in this study Term Significance Units Peak Period Period during which CAISO experiences high demand, defined as 4-9pm (also common hours for Flex Alerts (137)) Twelve-hour clock range Precooling Period The portion of the day during which precooling occurs (AC setpoint below baseline temperature) Twelve-hour clock range Offset Period Three-hour period immediately after precooling, always falls within the Peak Period (AC setpoint above baseline temperature) Twelve-hour clock range Start Time Time of day at which transition from precooling period to offset period occurs Twelve-hour clock time Baseline Temperature Temperature at which house is kept outside of the precooling and offset period ◦F Length Duration of the precooling period Hours (h) Depth Number of degrees below the baseline temperature the AC is set at during the precooling period ◦F Offset Temperature Number of degrees above the baseline temperature the AC is set at during the offset period ◦F Reset Method Method of returning to baseline temperature at the end of the offset period 1: Setpoint returns instantly (sudden) 2: Setpoint returns linearly over two hours (gradual) To define the full set of precooling schedules, we vary the precooling length from 1 to 5 hours, the depth from 1 to 4◦F (0.56 to 2.22◦C), the offset temperature from 1 to 4◦F (0.56 to 2.22◦C), the start time from 4 to 6pm, and the reset method through both of its options. We restrict the ranges for precooling depth and offset temperature to 4◦F (2.22◦C) to avoid creating exceedingly uncomfortable conditions, as discussed further in Sections 4.3.5 and 4.5.1, and due to computational limitations. Possible values are restricted to integer values, and these permutations create a total of 480 distinct precooling schedules to test for each single-family home design/climate zone combination. Precooling length, depth, and offset temperature, are restricted to integer values to represent schedules that can be easily employed by homeowners, given that most thermostats by default take integers for their setpoints, and sub-hourly adjustments in temperature are impractical for those without 66 Range of precooling lengths (1 to 5 hours) Range of precooling depths (1 to 4 oF) Range of offset temperatures (1 to 4 oF) Precooling Period Offset Period Reset Method (1, 2) Variable start time (4pm, 5pm, 6pm) Baseline Temperature (75 oF) Figure 16: Theoretical precooling schedule with illustration of key terms from Table 5. high-tech smart thermostats. The large number of tested schedules provides a broad domain from which to select a schedule that reduces a variable of interest. The baseline cooling schedule uses a constant setpoint of 75◦F (23.89◦C) at all hours of the day and serves as a reference point with which the precooling schedules are compared. 4.3.4 EnergyPlus Simulation Simulations of each precooling schedule are done using EnergyPlus, a building energy simulation software developed by the Department of Energy (138). EnergyPlus has been tested and verified both internally and externally, with relevant studies listed on the EnergyPlus website (139). EnergyPlus has been used both directly and indirectly (through programs that make use of the EnergyPlus engine or results) for a number of precooling studies in both the residential and commercial sector due to its ability to explore a large number of 67 alternatives in a time-efficient manner (117; 118; 119; 122; 124; 127; 129; 140; 141; 142). For each EnergyPlus simulation, we input a modified data file from the ResStock project that specifies the home’s properties and AC setpoint schedule, as well as a weather file that describes one typical year of weather in the climate zone in question. We simulate each single-family home using all of the precooling schedules and the baseline schedule in each of California’s building climate zones (a total of 30,784 simulations). EnergyPlus returns information at the hourly level, including indoor and outdoor climatic variables and electricity consumption—the main output used for the analysis of the precooling schedules. 4.3.5 Thermal Comfort An occupant’s thermal comfort is an important consideration when analyzing the effect of a precooling schedule. Thermal comfort describes a person’s perceived thermal sensation and can be estimated with a large set of methods and metrics (143). Several previous precooling studies (112; 124; 125; 130; 144; 145) have made use of Predicted Mean Vote (PMV), a thermal comfort metric developed by Fanger (146). The PMV metric is designed to estimate the self-described mean sensation of a large group of people based on variables such as current temperature, humidity, wind speed, activity level, and clothing level. In the PMV model, thermal comfort is measured on a seven point scale (centered at a neutral sensation, PMV=0), and The American Society of Heating, Refrigerating, and Air-Conditioning Engineers (ASHRAE) defines an acceptable range for PMV as -0.5 to +0.5 (147). In this study, standard values were used for wind speed (0.1 m/s), activity level (1.1: seated, typing), and clothing level (0.5, typical summer indoor clothing), while temperature and relative humidity were reported hourly by EnergyPlus. The baseline temperature of 75◦F (23.89◦C) was selected because it provides a close to neutral mean PMV for the test period for building 4 (the best insulated home). For the homes with better insulation and higher AC efficiency, there is a close correlation between AC setpoint and actual indoor temperature, but for the more poorly insulated homes with less efficient AC the indoor 68 temperature can significantly exceed the AC setpoint at a given time. In response, we use PMV to define a range of permissible AC setpoints by only allowing schedules that would fall in an acceptable comfort range assuming the indoor temperature matched the AC setpoint. Using constant values for the other variables, a PMV range of -0.5 to +0.5 translates to roughly a 6◦F (3.33◦C) range (148), reducing our precooling schedule space to schedules that use a precooling depth of no more than 3◦F (1.67◦C), and a maximum temperature of 3 ◦F (1.67◦C). This filter reduces our simulations from 480 per building-climate zone to 270, and our total simulations from 30,784 to 17,316. We discuss the impact of precooling on actual thermal comfort levels in the absence of this filter in Section 4.5.1. 4.3.6 Output Processing Using the hourly electricity consumption values generated by EnergyPlus, we calculate the CO2 emissions, residential electricity costs, and peak period electricity consumption for each precooling schedule as well as the baseline schedule, and repeat this process for each home design and climate zone. We restrict the analysis of these results to the months of July, August, and September due to the increased cooling needs during this summer period. Precooling schedules generally cause higher total daily electricity consumption than a constant setpoint schedule, with consumption increasing for longer and deeper precools, but we focus on the peak period when analyzing electricity consumption in this analysis and consider the total impacts on CO2 emissions and residential electricity costs. To determine the emissions associated with a precooling schedule, we use grid-level emissions factors calculated from historical data with linear regression models. Emissions factors generally quantify the amount of emissions associated with a unit of activity, in this case, the amount of CO2 emissions associated with consuming a unit of electricity. Researchers typically draw a distinction between two types of emissions factors: average emissions factors (AEFs) and marginal emissions factors (MEFs) (149; 150). Average emissions factors describe the relationship between the total amount of emissions being produced by the grid 69 and total supply of (or demand for) electricity; hence, they are useful for evaluating the emissions impacts of existing and regular electricity loads. MEFs, by contrast, describe the relationship between changes in emissions and changes in generation (or demand); hence, they are used for evaluating the impact of changes in the magnitude or timing of existing loads on emissions. MEFs are more applicable for load-shifting strategies like precooling, since these types of DR activities impose temporary changes in electricity-consuming behavior, prompting a temporary increase or decrease of marginal generation that would not occur if the DR strategy was not employed. To calculate the emissions associated with a cooling schedule, we first use AEFs calculated at the month-hour resolution (AEF is specific to the month of the year and hour of the day) to determine the emissions from the baseline schedule, and then use month-hour MEFs to adjust these emissions for each precooling schedule depending on the difference in hourly electricity consumption between the precooling and baseline schedule. Equation 6 shows this calculation at the hourly level, and results for the full 3-month simulation period are reported by summing the hourly results. This method treats the baseline cooling consumption as an existing/established load, and assumes that the changes in consumption caused by precooling are met by marginal generation. This distinction is consistent with recommended advice for using AEFs and MEFs in existing literature (106; 151). CO2precooling,h = AEFh × e − baseline,h + MEFh × (e − precooling,h − e − baseline,h) (6) Here e − denotes the hourly consumption of electricity for cooling in kWh (with subscripts specifying the baseline or precooling schedule), AEF and MEF refer to the average and marginal emissions factor in kgCO2 kW h at a point in time, and the subscript h denotes a specific hour in the study period. The MEFs and AEFs used are specific to CAISO’s grid, and are calculated following the methodology developed by Zohrabian, Mayes, and Sanders (105) for the year 2021 and are shown in Figure 17. This method was designed to develop MEFs that isolate the impact of demand on emissions, which is ideal for demand-side management 70 applications such as precooling. The data used to calculate these AEFs and MEFs can be found in the online data repository (https://data.mendeley.com/datasets/333jfmpvb8/1) and additional notes on the AEF and MEF calculations can be found in Section C.2. Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Month 0 2 4 6 8 10 12 14 16 18 20 22 Hour 2021 AEFs 100 150 200 250 300 350 AEF in kgCO2/MWh Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Month 0 2 4 6 8 10 12 14 16 18 20 22 Hour 2021 MEFs 150 200 250 300 350 MEF in kgCO2/MWh Figure 17: Month-hourly Average (left) and Marginal (right) Emissions Factors for the CAISO grid in 2021. Factors for the months of July, August, and September were included in this study. Reductions in peak period electricity consumption are found by summing the daily electricity consumption between 4 and 9pm for the baseline schedule and all of the precooling schedules. This time period includes the hours during which CAISO’s grid typically reaches peak net load (1) (net load is defined as total demand minus VRE production). For peak period consumption and CO2 emissions we report average daily values for the three-month study period. The hourly residential electricity cost associated with a precooling schedule (Ch in $), as well as the baseline schedule, is calculated based on a selected hourly TOU rate schedule as shown in Equation 7. Here rh ( $ kW h) refers to the hourly price for residential electricity during a specific hour under the TOU plan. Ch = rh × e − h (7) In Sections 4.4.1 and 4.4.2, we use a TOU program offered by Southern California Edison (152), an investor owned utility serving 15 million people in Southern California to calculate 71 costs. This program has an off-peak price that applies to hours outside of 5-8pm, and an onpeak price during 5-8pm that is twice as high. This ratio is used for illustrative purposes in these sections, but Section 4.5.1 explores the impact of different on-peak to off-peak ratios to better reflect the various TOU programs offered by different California electricity providers. Hourly residential electricity cost for cooling is reported in terms of a monthly value by dividing the total summed hourly cost by the number of months in the simulation period to be more easily interpretable for individuals who pay monthly electricity bills. By comparing the results for each precooling schedule, we can find the best schedule for reducing a specific variable of interest for each home design and climate zone combination, and examine trade-offs when maximizing reductions in different variables. In the remainder of this paper, we analyze the potential for reductions in peak period electricity use and reductions in CO2 emissions while using the residential electricity cost metric to confirm that these benefits can be delivered without economic penalties to residents. 4.4 Results For all four single-family homes and 16 California climate zones, the results of our simulations show that precooling with an offset temperature can achieve significant reductions in both peak period electricity demand and total CO2 emissions relative to a constant setpoint cooling schedule. Additionally, precooling can also substantially reduce residential electricity costs given typical TOU rate structures offered in the state of California (153; 152; 154; 155). Our results suggest that precooling is capable of eliminating 8-61% of peak period electricity demand for Building 1 (low AC efficiency and insulation) and 18-82% for Building 4 (high AC efficiency and insulation) depending on the climate zone. Reductions in CO2 emissions fell in the range of 1-22% for the Building 1 and 2-19% for the Building 4. While the home designs evaluated demonstrated a similar range of percent reductions for both variables across precooling schedules, the homes with inferior insulation and lower AC efficiency had much larger initial electricity consumption, and hence, much larger net absolute reductions 72 in CO2 and peak period electricity consumption. 4.4.1 The impact of precooling schedule parameters We first examine the impact of a precooling schedule on CO2 emissions and residential electricity costs for a specific climate zone. In Figure 18 we plot the changes in CO2 emissions, peak period consumption, and residential electricity cost (assuming SCE’s TOU rate plan) associated with a specific precooling length-depth combination for all four homes using Climate Zone 9 as an example. Climate Zone 9 is a region that includes the highly populated downtown Los Angeles and is characterized by having moderately hot summers. Figure 18 includes one point for the simulation that most reduces CO2 emissions for each precooling length-depth combination. Thus, each point might vary in terms of offset temperature, start-time, and reset method (see Table 5 for definitions) to achieve maximum emissions reductions for a given length-depth pair. For each selected schedule, the length and depth of the precooling schedule is shown as the length and height of the cross-hairs on the associated point. 73 4.0 3.5 3.0 2.5 2.0 1.5 Change in Peak Period Electricity Consumption (kWh/day) 0.3 0.2 0.1 0.0 0.1 Change CO2 Emissions (kg/day) Building 1 30 25 20 15 10 Change Cost ($/month) 4.0 3.5 3.0 2.5 2.0 1.5 Change in Peak Period Electricity Consumption (kWh/day) 0.30 0.25 0.20 0.15 0.10 0.05 0.00 0.05 0.10 Change CO2 Emissions (kg/day) 5h 3F Building 2 25 20 15 10 Change Cost ($/month) 2.0 1.8 1.6 1.4 1.2 1.0 Change in Peak Period Electricity Consumption (kWh/day) 0.14 0.12 0.10 0.08 0.06 0.04 0.02 0.00 Change CO2 Emissions (kg/day) Building 3 16 14 12 10 8 6 Change Cost ($/month) 2.2 2.0 1.8 1.6 1.4 Change in Peak Period Electricity Consumption (kWh/day) 0.150 0.125 0.100 0.075 0.050 0.025 0.000 0.025 0.050 Change CO2 Emissions (kg/day) Building 4 10 9 8 7 6 5 Change Cost ($/month) Figure 18: Changes in daily cooling related CO2 emissions, daily kWh of electricity consumed for cooling, and monthly cooling costs over the 3-month analysis period for multiple singlefamily home designs in California Climate Zone 9. Each point represents the precooling schedule that achieved maximum CO2 emissions reductions for a given precooling lengthdepth combination, with the length and depth of the schedule shown as the width and height of the cross-hairs. In Climate Zone 9, all home designs were able to reduce target variables significantly, with an impact of -1 to -4 kWh/day for peak period electricity consumption, and +0.2 to -0.3 kgCO2/day for CO2 emissions, depending on the home and precooling schedule (See Figure 18). For all home designs, it is possible to reduce both peak period consumption and emissions simultaneously, although maximizing reductions for one variable generally reduces the reduction in the other target variable. Deeper precools (i.e., AC setpoint far below baseline during the precooling period), and to a lesser extent, longer precools, result in greater reductions in peak period electricity consumption but significantly reduce reductions in CO2 due to the large increases in electricity consumption needed to reach the precooling 74 temperature in the middle of the day. For these schedules, the additional CO2 emissions during the precooling period are similar in magnitude to the avoided emissions during the offset period, meaning that even deep precools that increase overall daily electricity use for cooling can do so without increasing emissions due to the decoupling between electricity usage and grid-related emissions during hours with high VRE penetrations. On the other hand, shallow precools (AC setpoint slightly below baseline during the precooling period) offer smaller reductions in peak period electricity consumption, but the emissions penalty during this precooling period is smaller than the benefit occurred during the offset period. When selecting the other precooling parameters that most reduced CO2 emissions for a given precooling length-depth combination, the offset temperature selected was always 3 ◦F above the baseline temperature, with higher offsets eliminated by the thermal comfort constraints. Higher offset temperature consistently leads to lower CO2 emissions and peak period demand. The start time that most reduced CO2 emissions varied for the different precooling lengths and depths, as well as the specific home, but the vast majority of selected schedules started at 4 or 5pm, which is the period of the day when the grid transitions from high penetrations of solar generation to large penetrations of natural gas and imports. The reset method that most reduced CO2 emissions also depended on the home selected and precooling length and depth, with a gradual reset being the better strategy more often. The reset method had a relatively small impact on both CO2 emissions and peak period electricity consumption. 4.4.2 The impact of building properties and Climate Zone Next, we examine the trade-offs in precooling schedules that maximize CO2 emissions reductions versus schedules that maximize peak period consumption reductions relative to the baseline schedule for each home within a climate zone. Figure 19 illustrates results for four distinct climate zones: Climate Zone 3 (San Francisco) representing a coastal area with mild weather, Climate Zone 9 (Downtown Los Angeles) representing a dense urban area 75 with periodic heat waves, Climate Zone 13 (Fresno) representing an area with warm, humid summers, and Climate Zone 15 (Brawley) representing a hot, desert region. 0 1 2 3 4 Peak Period Electricity Consumption (kWh/day) 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 CO2 Emissions (kg/day) B1 B2 B3 B4 5h 3F CZ3 0 10 20 30 40 50 60 70 80 Cost ($/month) 0 2 4 6 8 10 12 14 Peak Period Electricity Consumption (kWh/day) 2 3 4 5 6 7 8 CO2 Emissions (kg/day) B1 B2 B3 B4 CZ9 0 50 100 150 200 250 300 Cost ($/month) 0 5 10 15 20 Peak Period Electricity Consumption (kWh/day) 0 2 4 6 8 10 12 14 CO2 Emissions (kg/day) B1 B2 B3 B4 3F5h CZ13 0 100 200 300 400 500 Cost ($/month) 5 10 15 20 25 30 Peak Period Electricity Consumption (kWh/day) 7.5 10.0 12.5 15.0 17.5 20.0 22.5 CO2 Emissions (kg/day) B1 B2 B3 B4 CZ15 0 100 200 300 400 500 600 700 800 Cost ($/month) Figure 19: The emissions, peak period electricity consumption, and residential electricity costs for the baseline cooling schedule as well as the precooling schedule that maximized CO2 reductions and the one that maximized peak period electricity reductions for each of the 4 single-family home designs in various California building climate zones. The length and depth of the schedule are shown as the width and height of the cross-hairs. Figure 19 shows that for all four homes across a variety of climate zones, precooling can significantly reduce peak period electricity consumption with little or no penalty in CO2 emissions. The homes with poor insulation and lower efficiency AC are capable of larger reductions in peak period electricity consumption than those with superior insulation and higher AC efficiency due to the higher initial levels of consumption for the baseline schedule, providing larger potential for reduction. In Climate Zone 13, precooling was able to reduce peak period electricity consumption by 4.2 kWh/day for Building 1 but just 1.2 kWh/day for Building 4. The maximum possible reductions in CO2 emissions are smaller in 76 percentage than the maximum possible reductions in peak period electricity consumption, but the schedules that most reduce CO2 emissions have the co-benefit of reducing peak period consumption significantly, especially in warmer climate zones. For example, for Building 1 in CZ15, the precooling schedule that most reduced CO2 emissions (by 0.44 kgCO2/day) also reduced peak period electricity consumption by 98% of the maximum possible reduction. The potential of precooling to significantly reduce peak period consumption without a CO2 penalty, or even with a small-moderate CO2 benefit, suggests that precooling can be used for a variety of single-family homes on a consistent basis in summer months. We also investigate the impact that climate zone and building properties have on the effectiveness of precooling schedules by comparing the schedule that maximizes CO2 reductions and the schedule that maximizes peak period reductions for each respective home across multiple climate zones. Figure 20 illustrates results for Buildings 1 and 4. 5 0 5 10 15 20 25 30 Peak Period Electricity Consumption (kWh/day) 0 5 10 15 20 25 CO2 Emissions (kg/day) CZ3 CZ7 CZ9 CZ13 CZ15 Building 1 0 100 200 300 400 500 600 700 800 Cost ($/month) 2 0 2 4 6 8 10 Peak Period Electricity Consumption (kWh/day) 0 2 4 6 8 CO2 Emissions (kg/day) CZ3 CZ7 CZ9 CZ13 CZ15 5h 3F Building 4 0 50 100 150 200 250 300 Cost ($/month) Figure 20: The emissions, peak period electricity consumption, and residential electricity costs for the baseline cooling schedule as well as the precooling schedule that maximized CO2 reductions and the one that maximized peak period electricity reductions for Buildings 1 and 4 in five different California building climate zones. The length and depth of the schedule are shown as the width and height of the cross-hairs. Figure 20 shows that reductions in peak period electricity consumption and CO2 emissions from precooling are generally larger in percentage in cooler climate zones, but larger in magnitude in warmer climate zones. Warmer climate zones typically have higher baseline cooling-related electricity consumption across home designs, and thus more potential for 77 reduction, but exact reductions in peak period electricity consumption and CO2 emissions also depend on factors like the diurnal temperature profile (e.g. precooling tends to be more effective in climate zones where evening temperatures remain high). For these climate zones, we observe a spread of potential reductions in peak period electricity consumption ranging from 1.1 kWh/day in Climate Zone 3 to 2.4 kWh/day in Climate Zone 13 for Building 4 and 2.4 kWh/day in Climate Zone 3 to 3.7 kWh/day in Climate Zone 9 for Building 1. Similarly, daily CO2 emissions can be reduced by 0.081 to 0.18 kgCO2/day for Building 1 and 0.26 to 0.40 kgCO2/day for the Building 4, depending on climate zone. 4.5 Discussion 4.5.1 Impact of Precooling on Thermal Comfort More precooling prior to the offset period may offer improved thermal comfort for building occupants by reducing the average temperature during the offset period. Deeper and longer precools remove more heat from the building during precooling, which can cause uncomfortably cool temperatures during precooling, but then help to prevent the home from reaching uncomfortably high temperatures during the offset period. This may be preferred by occupants that find it easier to adapt to cool temperatures, or who would may be gone during the precooling period in the afternoon. We examine the impact of precooling on offset period comfort by finding the mean PMV during the 3-hour offset period when the AC is set above the baseline temperature. Figure 21 shows the mean PMV as a function of the length and depth of the precool in Climate Zone 9 (Downtown Los Angeles) for Buildings 1 and 4 assuming a start time of 5pm, an offset temperature of 3◦F, and reset method 1. In Figure 21, we see that comfort is improved during the offset period when deeper and longer precooling schedules are used, but the relatively small impact on mean PMV for both homes (for even the deepest and longest precooling schedules) suggest that this effect is minimal. For Building 1, the difference in mean PMV between the most and least 78 Figure 21: Mean PMV during the offset period as a function of the length and depth of the precool for Buildings 1 and 4. PMV values outside the range of -0.5 to +0.5 are considered uncomfortable; all values for Building 1 fall above +0.5 and all for Building 4 fall between 0 and +0.5. precooling is 0.12, and the mean PMV exceeds ASHRAE’s comfort standard (PMV ≤ +0.5) for all precooling schedules. For Building 4, the mean PMV has a range of 0.16, and all of the precooling schedules have a mean PMV of less than +0.5 during the offset period and are therefore within the ASHRAE standard. These results suggest that it may be necessary to limit precooling to homes with appropriate building properties or consider co-adaptations to improve comfort for lower-performing homes. We also note that comfort perception has been shown to depend on a variety of psychological factors, such as control of the thermal environment (156; 157). Occupants may view precooling as reducing this control, negatively impacting their thermal comfort level. Additionally, both individual comfort levels (158) and resident occupancy levels (159) have been shown to vary significantly across users. This may create additional precooling opportunities beyond the scope of this study, such as eliminating more peak period consumption by precooling to deeper offsets for unoccupied homes, or further reducing CO2 emissions by raising the offset for users who are comfortable at higher temperatures. On the other hand, precooling may not be an appropriate strategy for potentially sensitive occupants, like young children, the elderly, or those with serious health conditions. 79 4.5.2 Influence of TOU Rates Precooling can offer utility customers cost benefits when used within TOU pricing structures that discourage electricity usage during the grid’s peak demand hours. Many Californian electricity providers offer a selection of TOU programs, but here we focus on TOU pricing schedules designed to reduce consumption from 5-8pm via an “on-peak” price that is higher than the “off-peak” price of electricity (e.g., see those offered by the Sacramento Municipality District (SMUD) (155), Southern California Edison (SCE) (152), and Pacific Gas and Electric (PG&E) (153)). For each home and climate zone combination, we identify the precooling schedule that maximizes reduction in peak period electricity consumption and use it to calculate a break-even ratio between on-peak price and off-peak price: the smallest ratio between on- and off-peak pricing for which the schedule that most reduces peak period electricity consumption also provides residential electricity cost benefits. In Figure 22, we see that cooler climate zones generally require a larger on-peak to offpeak price ratio to make the precooling schedule that minimized peak period electricity consumption also economically beneficial to residents. Outside of the coolest climate zones, most home and climate zone combinations have a minimum ratio significantly lower than 2:1. Utilities offer TOU programs with a variety of ratios, such as 1.37:1 for PG&E, 2:1 for SCE, 1.75:1 for SMUD, and 1.25:1 for San Diego Gas and Electric. Therefore, even precooling with goal of aggressively reducing peak demand will provide reduced residential electricity costs for many locations and single-family homes. Additionally, the schedules that most reduced CO2 consistently provided reductions in residential electricity cost even when the TOU ratio barely exceeded 1, meaning that shallower, shorter precools can reduce residential electricity costs regardless of home design or location, provided offset temperature is maximized within comfort levels. 80 1 6 7 3 5 16 8 2 9 10 12 11 10 14 13 15 CZ B1 B2 B3 B4 Building 1.0 1.2 1.4 1.6 1.8 2.0 2.2 TOU Ratio Figure 22: Minimum on-peak to off-peak electricity price ratio needed to make the the schedule that maximizes peak period electricity reductions also cost-reducing for each home and climate zone combination. Climate zones are organized by number of CDD from low to high. 4.5.3 Precooling for CAISO’s Flex Alert Program In recent years, CAISO has increasingly relied on the “Flex Alerts” program to avoid rolling blackouts. The Flex Alerts notification system was developed by CAISO as a mechanism to request that electricity users voluntarily reduce their consumption during periods of high demand (160). In September 2022, a Flex Alert sent via text was estimated to have reduced demand by over 1 GW in a span of just five minutes and continued to reduce demand over several successive hours (161). On the Flex Alert website, CAISO recommends a variety of strategies for shifting consumption away from a period of high demand, including aggressively cooling your home beforehand and then avoiding cooling use during the Flex Alert period (precooling) (160). Although CAISO recommends an AC setpoint of 78◦F or higher 81 during the Flex Alert period, it does not offer a recommendation for a specific precooling temperature prior to the Flex Alert period. Our results suggest that precooling should be done for several hours leading up to the Flex Alert period with an AC setpoint 3◦F or more below the residents typical baseline temperature (depending on occupants’ thermal comfort flexibility) to achieve large reductions in demand during the Flex Alert period. Importantly, our results show that precooling can be used to by residents in single-family homes to reduce consumption during the Flex Alert period regardless of location and home insulation level. 4.5.4 Impact of Precooling on CO2 Emissions With CAISO’s large diurnal variations in marginal emissions factors between mid-day and evening hours, precooling can moderately reduce cooling-related CO2 emissions, achieving at most 20% reductions when compared to a baseline schedule, but often significantly less, as shown in Section 4.4.1. Multiple factors limit precooling’s effectiveness at reducing CO2 emissions. First, the smaller thermal mass of the single-family homes considered here limits the amount of time cooling load can be shifted by. This makes it difficult to move AC usage from the highest emissions-intensity hour (often peak demand) to the lowest (often peak solar), which can occur far apart within a day. Secondly, precooling schedules take advantage of clean midday generation, which requires low AC setpoint temperatures during some of the hottest hours of the day, consuming significant additional electricity. Lastly, returning from the offset temperature to the baseline temperature causes a late-night spike in electricity consumption, a period during which the emissions intensity of CAISO is often high due to high natural gas consumption. This effect can be mitigated by a gradual reset to the baseline temperature (reset method 2), although this strategy shows mixed effectiveness depending on climate zone and selected home. Despite the modest reductions in CO2 emissions, our results do suggest that we have the option of using precooling to aggressively reduce electricity demand during the peak period without increasing CO2 emissions, or even with some reductions (see Figures 18, 19, 20). Many schedules can reduce daily demand during the 82 peak period by multiple kWh per single-family home while still decreasing emissions, even when total daily electricity consumption increases. In summer months, CAISO’s diurnal profile of both AEFs and MEFs (see Figure 17) feature lower mid-day emissions factors and higher emissions factors in the evening. The precooling schedules discussed in this paper aim to specifically take advantage of this pattern by aligning the increased electricity caused by precooling with periods of lower emissions factors, and reducing cooling when the emissions factors are higher. For other electricity grids that are less solar-dominant, precooling 1) might not be a viable strategy for emissions reductions in these regions, or 2) might have to be scheduled during a different portion of the day that takes advantage of the local daily variations in emissions-intensity due to VRE availability (which might not be possible given a home’s thermal inertia constraints). For example, for the Midcontinent Independent System Operator, marginal emissions factors in summer months (72) show a trend of being lower in the evening; this pattern would be difficult to take advantage of to reduce emissions because cooling needs in the evening are already lower. Generally, to gain emissions benefits (in addition to the standard peak-shaving benefits of precooling) a regional grid would have to transition from lower MEFs to higher MEFs over the span of a few hours, and for AC usage to be needed during the high MEF period; a pattern that may become more common as more US grids transition to higher solar generation. 4.6 Conclusion For all California climate zones and home designs, the precooling schedules that maximize peak period electricity consumption are deeper and longer; however, these schedules offer only slightly greater reductions in peak period consumption than shallow precool schedules that deliver simultaneous reductions in CO2 emissions. The schedules maximizing peak demand benefits may be preferable for DR events such as CAISO Flex Alerts, but residents can achieve similar benefits with the use of a shallow precool and offset temperature, which 83 might be more beneficial for frequent usage due to the reduced CO2 emissions. Furthermore, for a majority of the combinations of climate zone and homes evaluated, precooling can also reduce residential electricity costs for current TOU plans offered by local utilities, with the largest cost-savings occurring for shallower, shorter precooling schedules like those that most reduce CO2 emissions. When precooling, the transition from the precooling period (low AC setpoint) to the offset period (high AC setpoint) should occur in line with the timing of TOU rate plans if the resident is attempting to save money, or the start of a specified high demand period (such as a broadcasted Flex Alert period) if the goal is to reduce demand during a period of grid stress. The most effective precooling schedule will depend on a homeowner’s goal (i.e., reducing peak demand vs CO2 emissions vs electricity costs), as well as factors such as building envelope, HVAC efficiency, and location, With this in mind, residents of single-family buildings would benefit from utility-specific guidance on precooling or the use of optimization technology via smart thermostats. Full results for all precooling simulations (481 schedules x 16 climate zones x 4 single-family homes) can be found in the data repository online (https://data.mendeley.com/datasets/333jfmpvb8/1). Despite the variety of homes included in this study, many specific building characteristics remain untested, such as the impact of residential building size and type of AC. Large residential homes are not included in this analysis, and while more rare, they may contribute significantly to residential electricity demand, and associated CO2 emissions, due to their large daily consumption. This study also focused on homes with central AC, and the impact of precooling on homes with individual window or wall AC units was beyond the scope of this study. While the homes included in this study were selected to be reflective of many single-family homes in California, apartment and multi-family buildings in the residential sector were not considered. For the prototypes used in these simulations, we did not analyze their relative frequency in the existing building stock, and scaling individual building results to grid-level impacts is beyond the scope of this study. While the results of this study show that precooling can reduce CO2 emissions on a daily 84 basis, a permanent or season-long shift to precooling may reduce the applicability of MEFs, and create a scenario where AEFs should be considered when calculating emissions. The similar diurnal patterns of AEFs may create the same co-benefits of precooling, but this has only been studied for a single home design in California (130). All precooling schedules require some flexibility on thermal comfort, and the comfort sensitivity of occupants may limit precooling’s ability to reduce the target variables. Precooling can sometimes create uncomfortably cool conditions (for precools with large depth), and thus rely on occupants either having programmable thermostats and being away from the home, or briefly adapting to a cooler environment (such as through the use of additional clothing). Lastly, local utilities offer a variety of TOU rate plans, and precooling should be done strategically with a specific TOU plan in mind if the goal is to reduce residential electricity cost; in general, if the goal is cost reduction, precooling is best used leading up to the period during which there is an elevated per kWh price for residential electricity, with offset used to the limit of the occupants comfort during the following higher price period. 85 Chapter 5: Using Neural Networks to Forecast Marginal Emissions Factors: A CAISO Case Study The content included in this Chapter is published in: Mayes, S., Klein, N., & Sanders, K. T. (2023). Using neural networks to forecast marginal emissions factors: A CAISO case study. Journal of Cleaner Production, 139895. 5.1 Introduction As electric grids in the United States achieve increasingly high fractions of wind and solar generation, the relationship between amount of electricity demanded and the CO2 emissions associated with producing that electricity is decoupling and highly time-dependent. While grids with primarily fossil-fuel-based generation typically exhibit a close relationship between the amount of electricity demanded and the associated emissions, grids with high fractions of wind and solar Photovoltaic power (referred to collectively as variable renewable energy sources, or VRE) have an emissions-intensity that varies significantly throughout the day and over the course of a year. For example, wind and solar power accounted for 24.9% of total generation in 2021 in California (110), and as a result the California Independent System Operator (CAISO, the organization responsible for managing the majority of California’s bulk power system (108)) experiences diurnal patterns of solar generation and total demand for electricity that often create a low net load (total load minus VRE generation) in the middle of the day and a high net load in the evening. As a result, the grid tends to be less emissions-intense in the middle of the day when there is an abundance of solar power and more emissions-intense in the afternoon and evening when CAISO relies heavily on natural gas generators to meet demand (162; 163). An implication of this phenomenon is that shifting an electricity load from one time period to another can reduce the amount of emissions associated with the load, even if the magnitude of demand is unaffected. Quantifying the exact change in emissions associated with modifying load at a specific time becomes 86 a challenging but important task. One tool used to quantify the changes in emissions associated with changes in demand are emissions factors (EFs). Researchers typically draw a distinction between Marginal Emissions Factors (MEFs) and Average Emissions Factors (AEFs) (149; 164). AEFs describe the relationship between a region of interest’s total generation for a period of time and the total emissions produced from that generation, which is dependent on all of the generators supplying electricity to that region’s grid. MEFs describe the relationship between changes in generation (or demand) and changes in emissions and depend only on the resource or resources that respond to changes in demand at a specific time, making them useful for evaluating the impact of adding, removing, or altering loads (165; 166). Several studies have used AEFs for evaluating the emissions changes caused by demandside management (DSM) strategies that change the timing of electricity loads, including air-conditioning usage (167), EV charging (168), and electricity usage in the water industry (169). However, using MEFs for DSM applications has become more common in recent years under the justification that additional loads or changes in load impact electricity production by marginal generators (as opposed to changing production across the whole generation fleet). Recent studies have used MEFs to evaluate the impact of EV charging patterns (76; 96; 64; 170), electricity storage policy (72), pimm2021using, mckenna2017short, braeuer2020comparing, air-conditioning timing (131), and societal damage factors (171). One common method for estimating MEFs is through grid modelling. Grid models work by representing a fleet of power plants and creating an order in which they respond to demand, with the order often being cost- or merit-based and less frequently dependent on factors such as the location of the demand and/or the individual power plants (164; 172; 61; 173; 174). A downside to the grid-modelling approach is that these models often require simplifications that prevent them from fully representing grid behavior. For example, Gagnon, Cole (175) point out that their model does not account for maximum ramp rates, or minimum up or down times, which can significantly impact MEF estimates (174). Other 87 models ignore trades of electricity between regions, which can be a significant portion of supply and demand (172). The accuracy of these models can also vary significantly depending on location and data availability (173). Alternatively, regression techniques that estimate MEFs based on historical grid data may do a better job of incorporating grid constraints, although they are also limited by data availability and vary in accuracy. Many regression-based analyses follow a methodology similar to that established by Hawkes (176), where changes in demand or generation are regressed on changes in emissions (single-factor linear regression) (96; 64; 170; 72; 177; 178; 179; 171; 180; 106; 181). While this methodology works well for primarily fossil-fuel-based grids, it is less appropriate for grids that have both carbon and non-carbon emitting resources changing generation throughout the day. For example, Siler-Evans, Azevedo, and Morgan (106) regress changes in emissions on changes in fossil-fuel generation to determine the MEF, but this assumes that renewable energy sources are never the marginal resource, which is not the case for grids with significant generation of hydropower or that experience VRE curtailment (182; 183). Li et al. (72) expand on this approach by regressing hourly changes in emissions on hourly changes in generation from all sources. While this better incorporates renewables as marginal generators, it fails to capture grid dynamics during VRE ramping periods. Under this methodology, rapid increases in VRE generation can artificially give the impression that additional demand is emissions-free, or even create situations where hourly emissions drop despite hourly demand increasing (due to the displacement of fossil fuel generators). This could give the false conclusion that the MEF is zero or even negative, when, in the absence of curtailment, additional demand would actually increase emissions. This failure to identify the true marginal emissions could lead to large inaccuracies when estimating the emissions impacts of potential DSM strategies by incorrectly encouraging shifting load to high-renewable hours when, in reality, the marginal resource for some of those hours would be fossil-fuels. To improve upon these single-factor regression methodologies, more recent studies have 88 introduced models that incorporate both changes in total demand and changes in generation from VRE sources. These two variables can be thought of as having separate and opposite effects on changes in emissions and can be represented through a multiple-linear regression model, which was first introduced by Thomsom, Harrison, and Chick (184) to capture the influence of wind generation in Great Britain. This approach was also utilized by Zohrabian, Mayes, and Sanders (182) for VRE generation in California, where the coefficient on the VRE term was called the marginal displacement factor (MDF). These models better isolate the impact of changes in demand (instead of changes in generation or a subset of generation) on emissions, and thus are more appropriate for DSM applications that seek to determine the impact that shifting, shedding, or adding loads has on emissions. While this approach is an improvement on single-factor regression techniques, the application of the resulting MEFs to DSM is still limited by the temporal properties of these models. Beyond being demand-based, MEFs that are intended for evaluating DSM strategies like demand response and load-shifting should also be highly temporally resolved and capable of being estimated in advance. (We refer to these qualities as ”granularity” and ”forecastability”, respectively, in this manuscript.) Accurate, ahead-of-time MEF estimations would enable the creation of data-driven DSM programs that encourage behavioral changes to reduce emissions. Traditional linear regression calculates the model coefficients for groups of data points, which results in assigning the same MEF to multiple points. This lack of granularity is a concern because MEFs can vary significantly hour-to-hour and day-to-day. Coarseresolution MEFs, such as those calculated at the annual level (171; 96; 180; 184; 106; 181), can be useful for understanding the evolution of the grid and high-level changes in the mix of marginal resources, but are poorly-suited for evaluating the emissions implications of DSM applications, which modify loads on sub-daily time intervals. Other studies group their analyses by hour of the day, calculating MEFs at the year-hour level (24 values for the whole year) (184; 170; 106; 181; 179), or month-hour level (288 values for the whole year) (76; 72; 171; 182). MEFs calculated with these methods are better suited for demand-side 89 applications that focus on time of day, but fail to capture the significant changes in grid and consumer behavior that can occur across months, or even days. Factors such as the demand level and amount of renewable energy generation can depend strongly on weather variables that change significantly day-to-day. To increase granularity, Beltrami et al. (185) developed a novel methodology using an auto-regressive integrated moving average model and found that it compared favorably to traditional linear-regression approaches, achieving higher granularity and more accurate predictions of changes in emissions for historical data, though they did not explore using this method for forecasting. Other analyses group their regressions by load-level (177; 106; 181; 178; 64), which could theoretically improve both granularity (i.e., any time at which load is known can be assigned an MEF) and allow for MEF forecasting by using projections of load that are commonly done for electricity providers. Of the listed studies, this forecasting method was explored only by Huber et al. (64) who used MEF forecasts to determine optimal EV charging times. The main limitation of this approach is that binning by demand prior to regression forces all points with a similar level of demand to have the same MEF; in reality, other factors such as the time of day and amount of renewable generation can strongly impact MEFs. New methodologies are needed for calculating demand-based MEFs given the growing importance of DSM strategies for managing the challenges associated with grids with high penetrations of variable renewable energy and the rapid adoption of new grid resources and grid management strategies. The proposed method must be effective at 1) isolating the impact of demand on emissions, 2) temporally resolving the emissions factors (“granularity”), and 3) predicting these emissions factors ahead of time (“forecastability”). We improve upon previous models by creating a composite model composed of a multilayer perceptron (MLP) and a linear model and then apply it to the grid overseen by CAISO, a grid that has high fractions of renewable generation, a variety of resources that contribute to marginal generation, and good data availability. While (to the authors’ knowledge) MLP 90 models have not previously been used to calculate MEFs, there is precedent for using neural networks in the electric grid research space to forecast electricity demand and prices that achieve high degrees of accuracy and granularity (186; 187; 188; 189). The advantages offered by the composite model developed in this study, including higher accuracy and granularity than traditional regression models, are maintained through model forecasting, resulting in MEFs that are well suited for DSM applications. 5.2 Material and methods 5.2.1 Preparation of Dataset We prepared a dataset of CAISO generation, demand, and emissions for the 2019-2021 period with the following data collection and manipulation process. The key data used in this methodology and their sources include: 1. data on CAISO’s demand, variable renewable energy generation, and total imports from CAISO’s website (162). 2. hourly emissions data for individual power plants from the U.S. Environmental Protection Agency’s Clean Air Markets Program Data (190) (it should be noted that the EPA Air Markets Program Data is only for power plants with with capacity greater than 25MW, so the emissions calculated in this study do not include fossil fuel generators under this size) 3. data identifying which power plants belong to which balancing authorities from US Energy Information Administration (EIA) 860 data (191) 4. hourly data for individual BA generation by source and hourly trades between BAs from the EIA (trades are reported bidirectionally) (192). For all of the above sources, the data used are publicly available and free to download. These datasets required several processing steps to fill in missing values and ensure temporal 91 agreement between sources. First all data from time zones other than PST were shifted to the PST time zone. Second, the CAISO load data (reported in MW) reported at 5-minute intervals were aggregated to the hourly demand level (in MWh) via averaging. Third, for hours where CAISO’s electricity trade with another BA was not reported in the EIA data, we first checked to see if these values were present in the dataset from the other BA’s perspective. If values were missing in both directions we filled in the missing values with the average of the values on the 5 closest days at the same hour of the day. Then, the sum of all hourly electricity trades between CAISO and other regions reported in the EIA exchange data was scaled to match CAISO’s hourly total reported net imports (essentially treating CAISO’s total reported value as the ground truth). This scaling was done by taking the difference between CAISO’s reported value and the summed value and distributing this difference to each BA in proportion to the magnitude of their reported electricity trade. Finally EIA 860 data was used to aggregate the hourly powerplant emissions data to the balancing authority level, creating an estimate of total hourly emissions associated with CAISO generation and the total hourly emissions associated with the generation of each balancing authority with whom CAISO trades electricity. These datasets were then combined to create a final dataset describing the hourly demand, VRE, and CO2 emissions in the CAISO region (specifically, the hourly emissions are those associated with demand that occurs within the CAISO region). Calculating these hourly, demand-based emissions for CAISO requires accounting for trades of electricity between CAISO and neighboring BAs and the emissions associated with these trades. This emissions accounting process generally follows the procedure outlined by Chalendar, Taggart, and Benson (193), which was also utilized by Zohrabian, Mayes, and Sanders (182) from which we adapt Equation 8. Eh = E C h − E X h + E I h = E C h − ( X M E C GC · X C→M)h + (X M E M GM · I M→C )h (8) In Equation 8, the total emissions for CAISO at a specific hour (Eh) are calculated as 92 the total emissions from power plants within the CAISO region (E C h ) plus the emissions associated with imports from neighboring BAs (E I h ) minus the emissions associated with exports to neighboring BAs (E X h ). The emissions associated with exports are calculated as the sum over all neighboring BAs of the average emissions for CAISO (C) at a specific hour ( EC GC ) times the the amount of exports (XC→M) to each BA (M) during that hour. The emissions associated with imports are calculated as the sum over all BAs of the average emissions for each BA (M) at a specific hour ( EM GC ) times the amount of imports from that BA (I M→C) during that hour. This method assumes that the electricity traded between BAs is reflective of the overall grid mix of the BA in question at a that specific time. After aligning these calculated CAISO hourly emissions values with hourly demand and hourly VRE generation, we then differentiate the demand, VRE, and emissions at the hourly level to create three additional features representing the hourly changes in these variables. 5.2.2 Composite Model Design The model developed to calculate MEFs and predict the hourly changes in CAISO emissions is depicted in Figure 23. The composite model consists of a MLP network followed by a multi-variable linear model. The MLP takes input features for a given hour and predicts the coefficients of the linear model. The linear model has a term for changes in demand (∆D) with corresponding coefficient MEF, changes in VRE generation (∆V RE) with corresponding coefficient MDF, and an intercept term c. These predicted terms are then plugged in to the linear model to calculate a predicted change in hourly emissions. Comparing this value to the ground truth change in emissions data, a loss term is calculated and used to train the MLP model. This structure provides a way to learn MEF, MDF, and intercept coefficients despite the lack of ground-truth values for these variables by using the error from the predicted change in emissions to inform the predictions of the latent coefficient space. The input features used for the MLP model are hourly demand in the CAISO region, hourly variable renewable generation in CAISO, hour of the day, day of the year, and time 93 Figure 23: Conceptualization of the data sources used, the MLP-linear composite model, and the learning process. The model predicts hourly changes in the emissions associated with demand in the CAISO region. since 2018 in number of hours. These features were selected to capture as many of the dynamics of the electric grid as possible (without introducing unnecessary noise) and because they can be known or estimated ahead of time for forecasting. The level of demand and amount of VRE generation provide information on the type of marginal generator at a given point in time because many resources are used for specific levels of net load. Hour of the day can also be a good predictor of which resource or powerplant is next in the generation queue, and additionally may capture scheduled grid behaviors such as electricity trade commitments or hydropower operation. The day of the year was included to capture seasonal effects, such as hydropower generation, and time since 2018 was incorporated to capture general trends MEFs and MDFs, such as changes caused by the mix of grid resources getting cleaner over time (California’s in-state generation increased from 21% in 2019 to 23% in 2020 and 25% in 2021 (194)). 94 5.2.3 Model Implementation The three years of hourly data was first randomly split into train, validation, and test sets with a 60/20/20 split. All of the data was then standardized based on the mean and variance of the train set. During training, R-squared was computed on the validation set after each epoch and the model with the highest R-squared seen during this process was chosen as the final model. This model was then evaluated on the test set. The structure and key parameters of the MLP model were determined via a standard grid search and manual tuning. The MLP has two hidden layers, the first of size 512 followed by 256. After each hidden layer there is a batch normalization layer, a Rectified Linear Unit activation, and a dropout layer with a dropout probability of 0.5. The model was trained for 40,000 epochs using a full batch size and the AdamW optimizer with a learning rate of .01 and a weight decay of .003. The loss function used was mean squared error with a regularization term that penalized MEF and MDF values that fell outside of a reasonable range. In the process of developing this model, we explored additional features (such as weather variables) as well as more complex neural networks (long short-term memory models and attention-based models) but found that these alternatives did not significantly increase the accuracy of our model, and often decreased ease-of-use, especially for forecasting. The final version of the model used in this analysis, as well as the data used for model training, testing, and validation, is available in an online data repository (https://github.com/ S3researchUSC/MEF-Regression), with further data available upon request. 5.2.4 Model Outputs for Historical Data Using the final trained model, we determined historical MEFs, MDFs, and intercepts at the hourly level for CAISO for 2019-2021. We assess the accuracy of our model by calculating the R-squared values and mean absolute error (MAE) between our model’s predicted change in emissions and historical actual change in emissions, and compare these results to 95 predictions made with a multi-variable (MV) regression model. The MV regression model used for comparison is based on the work of Zohrabian, Mayes, and Sanders and provides a reference point for the accuracy level of these more granular MEFs. After the prediction step, we perform correlation analyses and a feature importance analysis using Shapley Values calculated by Shap.DeepExplainer (195) and discuss the significance of the results. (Shapley Values have been used for feature importance of neural networks in power systems in previous research (196).) Shapley Values are a way of measuring the relative importance of each feature and are calculated by removing a specific feature, finding all permutations of the remaining features (including the null set), and then evaluating the marginal contribution of the feature to the estimate (i.e., how much does including this feature increase the accuracy of the model across all combinations of features). This process is repeated for each feature and the results are used to assign a value that represents the importance of each feature (197). To assess the accuracy of our model for forecasting tasks, we use historical day-ahead forecasts of demand and VRE available on CAISO’s website (198) to create hypothetical demand and VRE forecasts for our test set. First, we split historical demand and VRE data by quintile and calculate the distribution of forecast errors in that quintile for the concurrent forecasts of demand and VRE. Then, we use the historical demand and VRE data from our test set and randomly sample the error distributions to derive hypothetical forecasts of VRE and demand for the test set hours. By using this sampling method, we can assess the sensitivity of our model to the typical level of inaccuracy present in CAISO’s demand and VRE forecasting data, determining how accurate our model would be if it relied on forecasted data for these features (the remaining features–day of the year, hour of the day, and time since 2018–are exactly defined for all future times and dates). We apply this method to the test set so that the model is being tested on forecasts of data that were not included in the training phase, though we note that true forecasting would expand beyond the range of trained values for the time since 2018 feature. We use the constructed day-ahead 96 forecasts of demand and VRE to predict the hourly change in emissions and compare the accuracy of these predictions to those made using actual demand and VRE generation data. 5.3 Results and Discussion 5.3.1 Hourly MEFs The hourly MEFs estimated by our composite model using historical data for 2019, 2020, and 2021 are shown in Figure 24. In all three years, MEFs vary significantly both throughout the day and over the course of the year. Higher MEFs are consistently seen in the late afternoon/early evening hours, but MEFs can also be high during the middle of the day, especially during fall months. MEFs tend to be low in the early morning throughout the year and during the middle of the day in spring and summer months. In the early morning (6-8am), there is typically a rapid increase in VRE generation and a small or no increase in demand; this may lead to an underestimation of MEFs as it is difficult to attribute changes in emissions to changes in demand. Low MEFs in the middle of the day in spring and summer months are expected when daytime solar generation is high, and particularly low during the times when hydropower, wind output and solar generation are simultaneously high (typically in spring months) (198). These hours of very high daytime solar often experience curtailment, and at times, marginal generation may be effectively emissions free (183). 97 Figure 24: Hourly marginal emissions factors calculated with the composite model for CAISO for 2019-2021. The MEFs calculated with this proposed methodology are notably more accurate for historical data than those calculated with the multi-variable (MV) linear regression model as shown in Table 6. Table 6 includes the accuracy of the forecasting portion of this analysis, which is discussed in Section 5.3.2. Our composite model outperforms the multi-variable regression model at predicting changes in emissions as measured by both R-squared and mean absolute error. Considering the significant increase in granularity achieved with our model, this result suggests that MEFs calculated with this methodology are preferable for DSM applications. The 22% reduction in MAE is significant for DSM applications, where the specific strategies depend on accurate estimates on the emissions impacts. This effect is magnified by the increased granularity, with a year being represented by 8760 distinct MEFs as opposed to the 288 MEFs created when binning by month and hour. With this level of information, 98 Table 6: Accuracy of different MEF models when using actual historical data and historical forecasted data as inputs. Only the MLP composite model is structured to use forecasted data. Metric 2019-2021 Historical Data Day-ahead Forecast of Test Set Binned MV Regression Composite Model Composite Model R-squared .85 .91 .88 MAE 156,000 122,000 133,000 DSM strategies can take advantage of MEF variations that occur on specific days but don’t present themselves as diurnal patterns over longer time periods. A number of factors may contribute to uncertainty in MEFs and explain the remaining inaccuracy in the predictions of changes in emissions. For example, limitations in the quality of the data, which was combined and merged across multiple sources, and data preprocessing, which relied on a hierarchy of data sources and a degree of temporal aggregation, may create inaccuracies when estimating MEFs. Additionally, events such as powerplant maintenance or planned in-operation may impact the queue of generation resources in a way that is difficult for the MLP model to predict or recognize. 5.3.2 Forecasting Results The results of the forecasting analysis show that day-ahead forecasts of demand and VRE generated from our composite model produce reliable estimates of MEFs. The percent difference between MEFs predicted with actual versus forecasted data are shown at the hourly level for 2021 in Figure 25, with a mean absolute difference of 9% for the entire year. The difference between the MEFs based on forecasted data and those based on actual data is a function of inaccuracies in CAISO’s forecasts of electricity demand and VRE generation. While CAISO’s forecast of demand is highly accurate (mean absolute error of less than 1%), its forecast of VRE is more uncertain (mean error of approximately 15%). Figure 25 shows that an underestimation of the MEF is more common in the middle hours of the day when using forecasted data. While there are occasionally large differences in forecasted MEFs versus those calculated with actual data, the difference is less than 10% for 72% of the hours 99 Figure 25: Hourly percent difference between the MEFs predicted by the composite model using actual versus forecasted CAISO demand and variable renewable energy for 2021. of the year, providing large utility to DSM planning. 5.3.3 Drivers of MEFs and Feature Importance The MEFs predicted by the MLP model show many non-linear behaviors, with the correlations between MEFs and variables such as demand, net demand, change in demand, and change in net demand being weak for all portions of the studied period (see Table 7). This non-linear behavior emphasizes the need for sophisticated models that can capture multiple contributing factors and the complex relationship between marginal emissions and grid behavior. Additionally, demand-related variables show reducing correlation with MEFs from 2019 to 2021, while hour of the day exhibits a stronger relationship. To examine the relative importance of each input feature to our MLP model, we calculated Shapley values and show them by month for 2021 in Figure 26. Hour of the day was generally 100 Table 7: Correlation between MEFs and measures of grid load (i.e., hourly demand and net load), hourly changes in measures of grid load, and hour of the day. Correlations are shown for each year of the study period as well as for the entire period. Correlation is measured by Pearson’s rho for continuous variables, and pseudo-rho (square root of goodness-of-fit R-squared) for categorical variables. Hourly Hourly Hourly Change Hourly Change Hour Demand Net Load in Demand in Net Load of Day 2019 MEFs 0.19 0.25 0.23 0.33 0.38 2020 MEFs 0.20 0.11 0.30 0.31 0.47 2021 MEFs 0.05 -0.04 0.18 0.18 0.50 2019-2021 MEFs 0.15 0.10 0.24 0.27 0.41 the most valuable feature for reducing model error (difference between predicted and actual changes in emissions plus regularization term). Day of the year, which was included to capture seasonal variations, was an important feature, especially in months with relatively flat MEF and MDF levels, where simply knowing the time of year is enough to make an accurate estimate. This suggests that DSM planners could make use of just these variables to plan load-shifting and DR strategies that reduce emissions, though our results show this would be less effective than using the granular, forecasted MEFs developed in this study. VRE was most important in April and May, months with significant solar and wind production that frequently reached curtailment levels in the middle of the day (over 500 GWh of VRE production was curtailed throughout April and May of 2021) (183). The amount of demand was of medium importance throughout the year, though generally lower in winter months with flatter levels of demand. The least important feature was found to be the total time elapsed since 2018, which was included to capture small general trends, such as a grid mix that is becoming higher percentage renewable energy over the three-year period. It should be noted that these Shapley values pertain to the importance of each feature in reducing prediction error, and therefore are relevant to MEF, MDF and intercept; the MEFs themselves may have a different relative feature importance. 101 Figure 26: Mean Shapley value by month of 2021. 5.3.4 Application to DSM MEFs that can be forecasted 24 hours in advance offer a range of possibilities for using DSM to reduce CO2 emissions, helping to meet decarbonization goals without relying on new technologies or conservation. The large variations in these granular MEFs for all timescales, and the lack of direct correlation with variables such as demand or hour of the day, make it difficult to devise general strategies for DSM. Instead, forecasted MEFs can be made available to the public and used to create short-term advice for load shifting. For example, flexible loads could be shifted from higher MEF to lower MEF hours, reducing emissions without reducing demand. Figure 27 illustrates the the difference in MEF caused by changing the timing of an electricity consuming behavior from one hour to another, using April 17th, 2021 as an example. On this day, it would be particularly valuable to shift demand away from 5pm, and generally valuable to shift demand from the 4 to 6pm window (i.e., peak demand 102 hours) towards the 9am to 1pm window, when solar resources are highest. Figure 27: Difference in MEFs between an initial hour and a new hour on April 17th, 2021 that an activity could occur in. Shades of blue represent switches in the timing of an activity that would reduce emissions. Beyond load-shifting, researchers, policy-makers, and industry members could use forecasted MEFs to devise lower emissions solutions for adding a flexible load to the grid (e.g., EV charging). Hourly MEFs can be used to minimize the amount of emissions associated with a flexible load, subject to constraints such as a window of time during which it is desirable for the task to be completed. This concept could be implemented by a variety of electricity providers and consumers. Time-of-use plans, where the price per unit of electricity varies by time of day (199), or real-time pricing, where consumers experience a rate based on the price of electricity in the whole-sale market (200), are utility rate structure options designed to inform the timing of electricity usage to save utilities money, reduce peak demand, and increase grid stability. Similarly, electricity providers could incentivize 103 consumption patterns that reduce the total amount of emissions associated with electricity consumption. While these reliability and climate mitigation goals may align during certain times of the year, Figure 24 shows that MEFs can also be high outside of typical grid peak hours. Further, Table 7 shows that MEFs are only slightly correlated with demand, so new plans or incentives could be designed to encourage emissions-reducing consumption patterns in addition to patterns that improve grid reliability. Emphasis could be put on the periods when these benefits align, or there may be situations where prioritizing reducing peak demand is a priority (e.g., during hot summer months) and others where grid reliability is a smaller concern and the focus should be on reducing emissions. Utilities could also create plans that weight both emissions and grid benefits or give customers the opportunity to choose their priority. On the end-user side, smart home technologies (201) could be designed so that consumers have the ability to choose flexible appliance and EV charging schedules that prioritize CO2 emissions reductions, in addition to cost savings. 5.3.5 Limitations and Future Improvements Through this methodology we were able to increase the accuracy and granularity of historical estimates of MEFs as well as accurately forecast day-ahead MEFs. Though these MEFs are well suited for DSM tasks, several potential improvements remain. The temporal resolution of MEFs is limited by the least granular dataset used in the emissions calculation, which is reported at the hourly level. More granular MEFs may be able to better describe the behavior of CAISO’s dynamic grid. This model may also benefit from additional inputs such as spatial information about the location of demand and generation that allows the model to better capture transmission constraints, which are expected to have a significant impact on grid-decarbonization (202). The emissions accounting methodology used in this study does not account for the emissions associated with grid-level storage. Although storage was used minimally in the years covered by this project, CAISO has rapidly expanded battery capacity in recent years from 104 less than 1 GW in mid 2021 to over 6 GW in 2023 (198). The use of storage should be addressed via shifting the emissions associated with charging a battery to the time that the battery is discharging (when the demand “actually” occurs), as was done in (184). Shifting emissions in this manner would likely increase MEFs for hours during which batteries are discharged and decrease MEFs for hours during which batteries are being charged. However, co-located renewable energy and storage plants may need to be treated differently, as this storage would be emissions-free. A remaining question is how to use MEFs for DSM given the influence of flexible, emissions-free resources. These resources are sometimes strategically reserved for highdemand periods (for example, the use of hydropower on CAISO’s grid (182)) leading to a low MEF during a period of high net demand. Under these conditions, shifting demand to this period may not actually decrease emissions and could increase grid stress. As mentioned in the methodology, we replicated the MEF forecasting task via historical forecasts of demand and VRE for our model test set, but future forecasting would feature predictions for hours outside of the temporal range of the training data. We believe this would have minimal impact on the accuracy of the model as this would only impact the ”time since 2018” feature, and the testing range of this feature would only be slightly beyond the training range (24 hours ahead). Lastly, applying this methodology to other regions depends on the amount and quality of data made available by other Independent System Operators (ISOs). This project used data from CAISO for demand, VRE generation, forecasted demand, forecasted VRE generation, and imports. Some ISOs share sufficient information on generation, demand, and trades to replicate this methodology exactly, but others might require modifications, such as using EIA data alone to calculate electricity trades or requiring the researcher to develop their own method of forecasting demand and VRE generation. While this method was designed for grids with high fractions of renewable energy, this method should maintain high accuracy for primarily fossil-fuel grids due to the isolation of the demand impact and renewable energy 105 impact achieved in the multivariate model. 5.4 Conclusions The novel methodology introduced in this study produces accurate and high-granularity estimates of both historical and day-ahead MEFs, filling a gap in the existing literature on statistical MEFs. Accurately forecasted MEFs can be utilized for a variety of DSM applications that aim to quantify the CO2 emissions associated with changing the demand for electricity. Hence, they can be leveraged to reduce CO2 emissions in addition to the traditional aims of DSM such as reducing costs for electricity producers and consumers. This study found significant variations in hourly MEFs both between days, and throughout the hours of the day, suggesting that shifting the timing of a flexible load can significantly change the CO2 emissions associated with that consumption. Our results also show that MEFs are not highly correlated with grid-level demand or net demand, underscoring the need for a flexible, non-linear MEF model, such as a neural network. This methodology will become increasingly applicable as grids across the US integrate more renewables and rely more heavily on DSM. 106 Chapter 6: Conclusion The research summarized in this dissertation fills existing research gaps by quantifying the impact of precooling on CO2 emissions and, more broadly, develops a framework for calculating the emissions impact of various DSM interventions. While precooling had been previously shown to reduce electricity costs and successfully shift load, this body of work reveals that precooling can maintain these benefits while also reducing CO2 emissions in a variety of circumstances, a result quantified with emissions factors. This dissertation contributes to existing research in the field of emissions factors by developing new methods for estimating MEFs. The final model developed in this work calculates demand-based MEFs, incorporates renewable generation and electricity trades between regions, and produces more accurate and temporally resolved MEFs than traditional models. Lastly, this method can forecast MEFs 24 hours in advance, a useful window for DSM. The contributions of each chapter are summarized in the following paragraphs. In Chapter 2, the effect of precooling on peak period electricity consumption, residential electricity costs, and CO2 emissions was explored via EnergyPlus for a single-family home in Southern California. The impact on emissions was calculated using average emissions factors. The results suggest that precooling is capable of delivering the traditional benefits of demand response/load-shifting strategies—reduced costs and peak demand—while also delivering moderate reductions in the CO2 emissions associated with cooling. This study was one of the first and remains one of the few to evaluate the carbon mitigation potential of precooling. In Chapter 3, a framework for calculating the impact of changes in demand on emissions was developed. This new approach to calculating MEFs increased the applicability of MEFs to the field of DSM using CAISO’s grid as a case study. Wind and solar generation were included via the use of a multi-variable linear model, and electricity trades between regions were incorporated into the emissions accounting process. Changes in emissions were also 107 regressed on changes in demand, rather than generation, to better isolate the impact that changing demand would have on emissions in load-shifting or demand response scenarios. In Chapter 4, the analysis of precooling was extended by using the demand-based MEFs from Chapter 3 and expanding the number of buildings and geographic area studied. Five distinct single-family homes spanning a range of thermal properties were simulated in each of California’s 16 climate zones. The outputs of these simulations suggest that CO2 emissions could be reduced by 1-22% by switching from a constant setpoint cooling schedule to a precooling schedule, depending on building type and location. These same schedules were also able to eliminate 8-82% of peak period electricity consumption and reduce residential electricity costs. In Chapter 5, our methods for calculating MEFs were improved with a multi-layer perceptron and linear composite model that increased the granularity, accuracy, and forecastability of MEFs. This model was tested on historical forecasts of demand and VRE generation for CAISO, and successfully predicted changes in emissions with R-squared value of 0.89. These demand-based, forecasted MEFs are well suited for programs that aim to reduce CO2 emissions via short-term alterations of load, a significant contribution to the field of DSM. In aggregate, the new methods of estimating MEFs and the demonstrated application of these MEFs to precooling validate the concept of using DSM to reduce CO2 emissions. The MEFs calculated in this body of work reveal a decoupling between the amount of demand for electricity and the corresponding emissions; load-shifting can become another tool in the efforts to decarbonize the grid, helping to meet climate change mitigation goals without requiring reductions in demand or additional generation resources. This methodology can be applied to many DSM strategies and will become increasingly relevant to grids across the U.S. as they transition from fossil-fuel to renewable generation. 108 References [1] C. I. S. Operator, Today’s outlook: Demand, "http://www.caiso.com/ TodaysOutlook/Pages/default.aspx" (2022). [2] M. Albadi, E. El-Saadany, A summary of demand response in electricity markets, Electric Power Systems Research 78 (11) (2008) 1989–1996, www.sciencedirect.com/ science/article/pii/S0378779608001272. doi:10.1016/J.EPSR.2008.04.002. [3] P. Siano, Demand response and smart grids—A survey, Renew. Sustain. Energy Rev. 30 (2014) 461–478, www.sciencedirect.com/science/article/pii/ S1364032113007211. doi:10.1016/J.RSER.2013.10.022. [4] P. L. Joskow, C. D. Wolfram, Dynamic pricing of electricity, American Economic Review 102 (3) (2012) 381–385. [5] D. Lew, G. Brinkman, N. Kumar, P. Besuner, D. Agan, S. Lefton, Impacts of Wind and Solar on Fossil-Fueled Generators (No. NREL/CP-5500-53504), National Renewable Energy Laboratory (NREL), 2012. [6] G. Bredehoeft, E. Krall, Increased solar and wind electricity generation in california are changing net load shapes, Tech. rep., U.S. Energy Information Administration, (EIA) (2014). [7] B. Jones-Albertus, Confronting the duck curve: how to address over-generation of solar energy, https://energy.gov/eere/articles/ confronting-duck-curve-how-address-over-generation-solar-energy. Energy Information Administration (EIA) (2017). [8] CAISO, Today’s outlook: emissions, http://caiso.com/TodaysOutlook/Pages/emissions.html (retrieved 24 November 2020) (2020). 109 [9] CAISO, Summer readiness and flex alerts—california iso, http://caiso.com/about/ Pages/News/SummerReadiness-FlexAlerts.aspx (2021). [10] CAISO, Flex alerts are voluntary calls for consumers to conserve electricity. flex alert: energy conservation tips, save energy on high demand days in california, https:// flexalert.org/https://flexalert.org/ (2021). [11] S. Verbeke, A. Audenaert, Thermal inertia in buildings: a review of impacts across climate and building use, Renew. Sustain. Energy Rev. 82 (2018) 2300–2318, https: //doi.org/10.1016/j.rser.2017.08.083. doi:10.1016/j.rser.2017.08.083. [12] A. German, M. Hoeschele, Residential mechanical precooling, Tech. rep., Davis Energy Group (2014). [13] A. Reilly, O. Kinnane, The impact of thermal mass on building energy consumption, Appl. Energy 198 (2017) 108–121. doi:10.1016/j.apenergy.2017.04.024. [14] R. Yin, P. Xu, M. A. Piette, S. Kiliccote, Study on auto-dr and pre-cooling of commercial buildings with thermal mass in california, Energy Build. 42 (2010) 967–975. doi:10.1016/j.enbuild.2010.01.008. [15] J. E. Braun, Load control using building thermal mass, J. Sol. Energy Eng. 125 (2003) 292–301. doi:10.1115/1.1592184. [16] U. E. I. A. (EIA), Total energy monthly data, https://eia.gov/totalenergy/data/ monthly/index.php (retrieved 24 November 2020) (2020). [17] M. Woodward, C. Berry, U.s. energy information administration—eia—independent statistics and analysis, https://eia.gov/todayinenergy/detail.php?id=36412 (retrieved 23 November 2020) (2018). [18] M. Chen, G. A. Ban-Weiss, K. T. Sanders, Utilizing smart-meter data to project 110 impacts of urban warming on residential electricity use for vulnerable populations in southern california, Environ. Res. Lett. 15 (2020). doi:10.1088/1748-9326/ab6fbe. [19] J. L. Reyna, M. V. Chester, Energy efficiency to reduce residential electricity and natural gas use under climate change, Nat. Commun. 8 (2017) 1–12. doi:10.1038/ ncomms14916. [20] M. E. H. Dyson, S. D. Borgeson, M. D. Tabone, D. S. Callaway, Using smart meter data to estimate demand response potential, with application to solar energy integration, Energy Policy 73 (2014) 607–619. doi:10.1016/j.enpol.2014.05.053. [21] A. German, M. Hoeschele, D. Springer, Maximizing the benefits of residential precooling, in: ACEEE Summer Study on Energy Efficiency in Buildings, American Council for and Energy-Efficient Economy Washington, DC, USA, 2014, pp. 71–83. [22] R. Arababadi, K. Parrish, Modeling and testing multiple precooling strategies in three residential building types in the phoenix climate, in: ASHRAE Transactions, Vol. vol 122, American Society of Heating, Refrigerating and Air-Conditioning Engineers (ASHRAE), 2016, pp. 202–214. [23] W. J. Cole, K. M. Powell, E. T. Hale, T. F. Edgar, Reduced-order residential home modeling for model predictive control, Energy Build. 74 (2014) 69–77. doi:10.1016/ j.enbuild.2014.01.033. [24] C. Booten, P. C. Tabares-Velasco, Using energyplus to simulate the dynamic response of a residential building to advanced cooling strategies, in: 2nd Int. Conf. on Building Energy and Environment, 2012. [25] J. Nelson, N. G. Johnson, P. T. Chinimilli, W. Zhang, Residential cooling using separated and coupled precooling and thermal energy storage strategies, Appl. Energy 252 (2019). doi:10.1016/j.apenergy.2019.113414. 111 [26] W. J. N. Turner, I. S. Walker, J. Roux, Peak load reductions: electric load shifting with mechanical pre-cooling of residential buildings with low thermal mass, Energy 82 (2015) 1057–1067. doi:10.1016/j.energy.2015.02.011. [27] S. Katipamula, N. Lu, Evaluation of residential hvac control strategies for demand response programs, in: ASHRAE Transactions, Vol. vol 112, 2006, pp. 535–548. [28] H. E. R. Solutions, Smud’s 2012 residential precooling study—load impact evaluation, http://herterenergy.com/pdfs/Publications/2013_Herter_SMUD_ ResPrecooling.pdf (2013). [29] J. Wang, C. Y. Tang, L. Song, Design and analysis of optimal precooling in residential buildings, Energy Build. 216 (2020). doi:10.1016/j.enbuild.2020.109951. [30] K. Herter, Y. Okuneva, Smud’s 2012 residential precooling study – load impact evaluation, Tech. rep., Herter Energy Research Solutions (2013). [31] Y. Strengers, Air-conditioning australian households: the impact of dynamic peak pricing, Energy Policy 38 (2010) 7312–7322. doi:10.1016/j.enpol.2010.08.006. [32] K. Keeney, J. E. Braun, Application of building precooling to reduce peak cooling requirements, ASHRAE Trans. 103 (1997). [33] S. Morgan, M. Krarti, Impact of electricity rate structures on energy cost savings of pre-cooling controls for office buildings, Build. Environ. 42 (2007) 2810–2818. doi: 10.1016/j.buildenv.2005.11.010. [34] P. Xu, R. Yin, C. Brown, D. E. Kim, Demand Shifting with Thermal Mass in Large Commercial Buildings in a California Hot Climate Zone, California Energy Commission, PIER Energy-Related Environmental Research Program, 2009. [35] C. J. Klaassen, J. M. House, I. E. Center, Demonstration of load shifting and peak load reduction with control of building thermal mass, in: Teaming for Efficiency: Com112 mercial Buildings: Technologies, Design, Performance Analysis, and Building Industry Trends, Vol. vol 3, 2002, p. p 55. [36] P. Xu, P. Haves, M. A. Piette, L. Zagreus, Demand shifting with thermal mass in large commercial buildings: field tests, simulation and audits, Tech. rep. (2005). [37] J. Ma, J. Qin, T. Salsbury, P. Xu, Demand reduction in building energy systems based on economic model predictive control, Chem. Eng. Sci. 67 (2012) 92–100. doi: 10.1016/j.ces.2011.07.052. [38] GardenZeus, California climate zones, https://gardenzeus.com/climate-zones/ (2020). [39] E. International, https://www.edison.com/home/about-us.html:~:text=About% 20Edison%20International,and%20technologies%2C%20including%20renewable% 20energy. (2020). [40] U. D. of Energy, testing and Validation https://energyplus.net/testing (2020). [41] J. Neymark, R. Judkoff, G. Knabe, H.-T. Le, M. D¨urig, A. Glass, G. Zweifel, Applying the building energy simulation test (bestest) diagnostic method to verification of space conditioning equipment models used in whole-building energy simulation programs, Energy Build. 34 (2002) 917–931. doi:10.1016/s0378-7788(02)00072-5. [42] H. S. Rallapalli, A comparison of energyplus and equest whole building energy simulation results for a medium sized office building, Ph.D. thesis, Arizona State University (2010). [43] S. Shrestha, G. Maxwell, Empirical validation of building energy simulation software: Energyplus, in: Proc. of Building Simulation, 2011, pp. 2935–2942. [44] P. Haves, B. Ravache, A. Fergadiotti, C. Kohler, M. Yazdanian, Accuracy of hvac load 113 predictions: validation of energyplus and doe-2 using flexlab measurements, Tech. rep. (2020). [45] L. B. N. L. National Renewable Energy Laboratory, Energyplushttps: //www.osti.gov/biblio/1395882-energyplus:~:text=;EnergyPlus%E2%84% A2%20is%20a%20whole,and%20water%20use%20in%20buildings. (1999). [46] D. B. Crawley, J. W. Hand, M. Kummert, B. T. Griffith, Contrasting the capabilities of building energy performance simulation programs, Build. Environ. 43 (2008) 661–673. doi:10.1016/j.buildenv.2006.10.027. [47] B. E. C. Program, Residential prototype building models, https://energycodes. gov/development/iecc_models (2020). [48] B. E. C. Program, Residential energy and cost analysis methodology, https:// energycodes.gov/development/residential/methodology (2020). [49] V. V. Mendon, Z. T. Taylor, Development of residential prototype building models and analysis system for large-scale energy efficiency studies using energyplus, in: ASHRAE/IBPSA-USA Building Simulation Conference, 2014. [50] N. Siniavskaia, What foundations are built across the nation?: eye on housing, eye On Housing — National Association of Home Builders Discusses Economics and Housing Policy. http://eyeonhousing.org/2014/10/ what-foundations-are-built-across-the-nation/:~:text=NAHB%20analysis% 20of%20the%20Survey,percent%20have%20a%20crawl%20space (2014). [51] S. Gupta, Characteristics of new single family homes, nAHB https://nahbclassic. org/generic.aspx?genericContentID=64030&print=true (2006). [52] U. E. I. Administration, https://eia.gov/consumption/residential/data/2015/ hc/php/hc6.8.php (2015). 114 [53] U. DOE, Weather data by region, https://energyplus.net/weather-region/ north_and_central_america_wmo_region_4/USA/CA-Zones (2020). [54] P. E. C. C. Information, Guide to california climate zones, https://pge.com/myhome/ edusafety/workshopstraining/pec/toolbox/arch/climate/index.shtmlhttps: //pge.com/myhome/edusafety/workshopstraining/pec/toolbox/arch/climate/ index.shtml (2020). [55] WeatherSpark.com, Average weather in los angeles, california, united states, year round—weather spark, https://weatherspark.com/y/1705/ Average-Weather-in-Los-Angeles-California-United-States-Year-Round (2020). [56] R. . A.-C. E. American Society of Heating, Standard 55: 2010, thermal environmental conditions for human occupancy, Tech. rep., ASHRAE (2010). [57] F. Tartarini, S. Schiavon, T. Hoyt, C. Mackey, Comfort models. comfort models pythermalcomfort 1.3.3 documentation, https://pythermalcomfort.readthedocs.io/en/ latest/reference/pythermalcomfort.htmlid5 (2020). [58] S. C. Edison, Time-of-use residential rate plans, https://sce.com/residential/ rates/Time-Of-Use-Residential-Rate-Plans(retrieved23November2020) (2020). [59] R. McCarthy, C. Yang, Determining marginal electricity for near-term plug-in and fuel cell vehicle demands in california: impacts on vehicle greenhouse gas emissions, J. Power Sources 195 (2010) 2099–2109. doi:10.1016/j.jpowsour.2009.10.024. [60] A. Jenn, K. Clark-Sutton, M. Gallaher, J. Petrusa, Environmental impacts of extreme fast charging, Environ. Res. Lett. 15 (9) (2020) 094060. doi:10.1088/1748-9326/ ab9870. 115 [61] P. Gagnon, W. Cole, Planning for the evolution of the electric grid with a longrun marginal emission rate, iScience 25 (3) (2022) 103915, www.sciencedirect.com/ science/article/pii/S2589004222001857. doi:doi.org/10.1016/j.isci.2022. 103915. [62] G. Kamiya, J. Axsen, C. Crawford, Modeling the GHG emissions intensity of plug-in electric vehicles using short-term and long-term perspectives, Transp Res D Transp Environ 69 (2019) 209–223. doi:10.1016/j.trd.2019.01.027. [63] V. Raichur, D. S. Callaway, S. J. Skerlos, Estimating Emissions from Electricity Generation Using Electricity Dispatch Models: The Importance of System Operating Constraints, J. Ind. Ecol. 20 (1) (2016) 42–53. doi:10.1111/jiec.12276. [64] J. Huber, K. Lohmann, M. Schmidt, C. Weinhardt, Carbon efficient smart charging using forecasts of marginal emission factors, Journal of Cleaner Production 284 (2021) 124766. [65] J. J. Buonocore, P. Luckow, G. Norris, J. D. Spengler, B. Biewald, J. Fisher, J. I. Levy, Health and climate benefits of different energy-efficiency and renewable energy choices, Nat. Clim. Change 6 (1) (2016) 100–106, www.nature.com/articles/nclimate2771# Sec14. doi:10.1038/nclimate2771. [66] M. Li, Y. Yang, T. M. Smith, E. J. Wilson, Wind can reduce storage-induced emissions at grid scales, Appl. Energy 276 (2020) 115420. doi:10.1016/j.apenergy.2020. 115420. [67] A. J. Pimm, J. Palczewski, E. R. Barbour, T. T. Cockerill, Using electricity storage to reduce greenhouse gas emissions, Appl. Energy 282 (2021) 116199. doi:10.1016/ j.apenergy.2020.116199. [68] T. A. Deetjen, I. L. Azevedo, Reduced-Order Dispatch Model for Simulating Marginal Emissions Factors for the United States Power Sector, Environ. Sci. Technol 53 (17) 116 (2019) 10506–10513, pubs.acs.org/doi/abs/10.1021/acs.est.9b02500. doi:10. 1021/acs.est.9b02500. [69] Z. Zheng, F. Han, F. Li, J. Zhu, Assessment of marginal emissions factor in power systems under ramp-rate constraints, CSEE J. Power Energy Syst. 1 (4) (2015) 37–49, ieeexplore.ieee.org/document/7377170. doi:10.17775/CSEEJPES.2015.00049. [70] I. Khan, Importance of GHG emissions assessment in the electricity grid expansion towards a low-carbon future: A time-varying carbon intensity approach, J. Clean. Prod. 196 (2018) 1587–1599. doi:10.1016/J.JCLEPRO.2018.06.162. [71] I. Khan, M. W. Jack, J. Stephenson, Analysis of greenhouse gas emissions in electricity systems using time-varying carbon intensity, J. Cleaner Prod. 184 (2018) 1091–1101. doi:10.1016/J.JCLEPRO.2018.02.309. [72] M. Li, T. M. Smith, Y. Yang, E. J. Wilson, Marginal Emission Factors Considering Renewables: A Case Study of the U.S. Midcontinent Independent System Operator (MISO) System, Environ. Sci. Technol 51 (19) (2017) 11215–11223. doi:10.1021/ acs.est.7b00034. [73] J. Mccall, J. Macknick, D. Hillman, Water-Related Power Plant Curtailments: An Overview of Incidents and Contributing Factors, National Renewable Energy Laboratorywww.nrel.gov/docs/fy17osti/67084.pdf (2016). [74] A. D. Hawkes, Estimating marginal CO2 emissions rates for national electricity systems, Energy Policy 38 (10) (2010) 5977–5987. doi:10.1016/j.enpol.2010.05.053. [75] S. Sengupta, T. Spencer, N. Rodrigues, R. Pachouri, S. Thakare, P. J. Adams, R. Tongia, I. M. Azevedo, Current and Future Estimates of Marginal Emission Factors for Indian Power Generation, Environ. Sci. Technol 56 (13) (2022) 9237– 9250, pubs.acs.org/doi/full/10.1021/acs.est.1c07500. doi:10.1021/ACS.EST. 1C07500/ASSET/IMAGES/LARGE/ES1C07500_0005.JPEG. 117 [76] Y. Gai, A. Wang, L. Pereira, M. Hatzopoulou, I. D. Posen, Marginal Greenhouse Gas Emissions of Ontario’s Electricity System and the Implications of Electric Vehicle Charging, Environ. Sci. Technol 53 (13) (2019) 7903–7912, pubs.acs.org/doi/abs/ 10.1021/acs.est.9b01519. doi:10.1021/acs.est.9b01519. [77] P. L. Donti, J. Z. Kolter, I. L. Azevedo, How Much Are We Saving after All? Characterizing the Effects of Commonly Varying Assumptions on Emissions and Damage Estimates in PJM, Environ. Sci. Technol 53 (16) (2019) 9905–9914. doi: 10.1021/acs.est.8b06586. URL pubs.acs.org/doi/abs/10.1021/acs.est.8b06586 [78] M. P. Thind, E. J. Wilson, I. L. Azevedo, J. D. Marshall, Marginal Emissions Factors for Electricity Generation in the Midcontinent ISO, Environ. Sci. Technol 51 (24) (2017) 14445–14452, pubs.acs.org/doi/abs/10. 1021/acs.est.7b03047?casa_token=d6OZmhVmsOAAAAAA:r5drz_Ms58ZQNi_ 77uY8iT3XN3nqSjDVmthFkp1JMJEBcvqiIe0QsoGRKPcVSjuav7Yul_oJq7Ud2yC1. doi:10.1021/acs.est.7b03047. [79] K. Siler-Evans, I. Lima Azevedo, M. Granger Morgan, Marginal Emissions Factors for the U.S. Electricity System, Environ. Sci. Technol. 46 (9) (2012) 4742–4748, pdfs. semanticscholar.org/8921/bb8e8ab3c3b6623c63b95e2bea601821f2b4.pdf. doi: 10.1021/es300145v. [80] N. Seckinger, P. Radgen, Dynamic prospective average and marginal GHG emission factors—scenario-based method for the german power system until 2050, Energies 14 (9), www.mdpi.com/1996-1073/14/9/2527 (2021). doi:10.3390/en14092527. [81] S. Bender, D. Doughman, Pam Hungerford, S. Korosec, T. Lieberg, M. Merritt, M. Rawson, H. Raitt, J. Sugar, Implementing California’s Loading Order for Elec118 tricity Resources, Tech. rep., California Energy Commission, http://www.energy. ca.gov/2005publications/CEC-400-2005-043/CEC-400-2005-043.PDF (2005). [82] R. C. Thomson, G. P. Harrison, J. P. Chick, Marginal greenhouse gas emissions displacement of wind power in Great Britain, Energy Policy 101 (2017) 201–210, www. sciencedirect.com/science/article/pii/S0301421516306036. doi:10.1016/j. enpol.2016.11.012. [83] N. A. Ryan, J. X. Johnson, G. A. Keoleian, Comparative Assessment of Models and Methods to Calculate Grid Electricity Emissions, Environ. Sci. Technol 50 (17) (2016) 8937–8953. doi:10.1021/acs.est.5b05216. [84] B. Tranberg, O. Corradi, B. Lajoie, T. Gibon, I. Staffell, G. B. Andresen, Realtime carbon accounting method for the European electricity markets, Energy Strategy Rev. 26 (2019) 100367, www.sciencedirect.com/science/article/pii/ S2211467X19300549. doi:10.1016/j.esr.2019.100367. [85] J. A. de Chalendar, J. Taggart, S. M. Benson, Tracking emissions in the US electricity system, Proc. Natl. Acad. Sci. U.S.A 116 (51) (2019) 25497–25502, www.pnas.org/ cgi/doi/10.1073/pnas.1912950116. doi:10.1073/pnas.1912950116. [86] CAISO, California ISO - Managing Oversupply, http://www.caiso.com/informed/ Pages/ManagingOversupply.aspx, accessed on 28/08/21 (2021). [87] U.S. Energy Information Administration, Real-time Operating Grid, www.eia.gov/ electricity/gridmonitor/dashboard/electric_overview/US48/US48, accessed on 08/20/21 (2021). [88] U.S. Energy Information Administration, Form EIA-860 detailed data with previous form data (EIA-860A/860B), www.eia.gov/electricity/data/eia860/, accessed on 08/23/21 (2020). 119 [89] Air Markets Program Data, ampd.epa.gov/ampd/, accessed on 09/12/21 (2021). [90] BPA, BPA: Balancing Authority Load & Total Wind Generation, transmission.bpa. gov/Business/Operations/Wind/, accessed on 25/05/21 (2021). [91] CAISO, California ISO - Todays Outlook, http://www.caiso.com/TodaysOutlook/ Pages/default.aspx, accessed on 16/12/21 (2021). [92] A. Hundiwale, Greenhouse Gas Emission Tracking Methodology, Tech. rep., California Independent System Operator, Folsom, www.caiso.com/Documents/ GreenhouseGasEmissionsTracking-Methodology.pdf (12 2016). [93] M. Krarti, M. Aldubyan, Review analysis of COVID-19 impact on electricity demand for residential buildings, Renew. Sustain. Energy Rev. 143 (2021) 110888, www. sciencedirect.com/science/article/pii/S1364032121001829. doi:doi.org/10. 1016/j.rser.2021.110888. [94] E. Kawka, K. Cetin, Impacts of COVID-19 on residential building energy use and performance, Build. Environ. 205 (August) (2021) 108200. doi:10.1016/j.buildenv. 2021.108200. [95] I. L. Azevedo, T. A. Deetjen, P. L. Donti, N. C. Horner, G. Schivley, B. Sergi, K. Silerevans, P. T. Vaishnav, Electricity Emissions & Damage Factors, Tech. rep., Center For Climate and Energy Decision Making, Pittsburgh: Carnegie Mellon University, cedm.shinyapps.io/MarginalFactors/ (2020). [96] S. P. Holland, M. J. Kotchen, E. T. Mansur, A. J. Yates, Why marginal co2 emissions are not decreasing for us electricity: Estimates and implications for climate policy, Proceedings of the National Academy of Sciences 119 (8) (2022) e2116632119. [97] M. Steen, Greenhouse Gas Emissions from Fossil Fuel Fired Power Generation Systems, Tech. rep., European Commission Joint Research Centre (DG JRC) Institute 120 for Advanced Materials, publications.jrc.ec.europa.eu/repository/bitstream/ JRC21207/EUR19754EN.pdf (2000). [98] CDWR, Water Year 2020 Summary Information, Tech. Rep. September, California Departement of Water Resources, water.ca.gov/-/media/DWR-Website/Web-Pages/ What-We-Do/Drought-Mitigation/Files/Publications-And-Reports/ Water-Year-2020-Handout{_}Final.pdf (2020). [99] S. Mayes, K. Sanders, Quantifying the electricity, CO 2 emissions, and economic tradeoffs of precooling strategies for a single-family home in Southern California, Environ. Res. Infrastruct. Sustainability 2 (2) (2022) 025001. doi:10.1088/2634-4505/ac5d60. [100] C. Samaras, K. Meisterling, Life cycle assessment of greenhouse gas emissions from plug-in hybrid vehicles: Implications for policy, Environ. Sci. Technol 42 (9) (2008) 3170–3176, pubs.acs.org/doi/pdf/10.1021/es702178s. doi:10.1021/es702178s. [101] N. C. Onat, M. Kucukvar, O. Tatari, Conventional, hybrid, plug-in hybrid or electric vehicles? State-based comparative carbon and energy footprint analysis in the United States, Appl. Energy 150 (2015) 36–49, www.sciencedirect.com/science/article/ pii/S0306261915004407. doi:doi.org/10.1016/j.apenergy.2015.04.001. [102] P. Warren, A review of demand-side management policy in the uk, Renewable and Sustainable Energy Reviews 29 (2014) 941–951. [103] E. Guelpa, V. Verda, Demand response and other demand side management techniques for district heating: A review, Energy 219 (2021) 119440. [104] A. Zohrabian, K. T. Sanders, Emitting less without curbing usage? exploring greenhouse gas mitigation strategies in the water industry through load shifting, Applied Energy 298 (2021) 117194. 121 [105] A. Zohrabian, S. Mayes, K. T. Sanders, A data-driven framework for quantifying consumption-based monthly and hourly marginal emissions factors, Journal of Cleaner Production 396 (2023) 136296. [106] K. Siler-Evans, I. L. Azevedo, M. G. Morgan, Marginal emissions factors for the us electricity system, Environmental science & technology 46 (9) (2012) 4742–4748. [107] California average grid electricity used as a transportation fuel in california and electricity supplied under the smart charging or smart electrolysis provision, Tech. rep., California Air Resources Board (2020). [108] C. I. S. Operator, Understanding the iso, "http://www.caiso.com/about/Pages/ OurBusiness/Default.aspx" (2023). [109] CAISO, Monthly renewables performance report, "http://www.caiso.com/ Documents/MonthlyRenewablesPerformanceReport-Dec2022.html" (2022). [110] C. E. Commission, Total system electric generation, "https://www.energy. ca.gov/data-reports/energy-almanac/california-electricity-data/ 2021-total-system-electric-generation" (2022). [111] J. E. Braun, Load control using building thermal mass, J. Sol. Energy Eng. 125 (3) (2003) 292–301. [112] K. R. Keeney, J. E. Braun, Application of building precooling to reduce peak cooling requirements, ASHRAE transactions 103 (1) (1997) 463–469. [113] J. Braun, M. A. Piette, P. Haves, P. Xu, Peak demand reduction from pre-cooling with zone temperature reset in an office building, Citeseer, 2004. [114] P. Xu, P. Haves, M. A. Piette, L. Zagreus, Demand shifting with thermal mass in large commercial buildings: Field tests, simulation and audits (2005). 122 [115] J. Wang, C. Y. Tang, L. Song, Design and analysis of optimal pre-cooling in residential buildings, Energy and Buildings 216 (2020) 109951. [116] J. Wang, C. Y. Tang, L. Song, Analysis of precooling optimization for residential buildings, Applied Energy 323 (2022) 119574. [117] A. German, M. Hoeschele, D. Springer, et al., Maximizing the benefits of residential pre-cooling, in: ACEEE Summer Study on Energy Efficiency in Buildings, American Council for and Energy-Efficient Economy Washington, DC, USA, 2014, pp. 71–83. [118] A. German, M. Hoeschele, Residential mechanical precooling, Tech. rep., Davis Energy Group, Davis, CA (United States). Alliance for Residential . . . (2014). [119] J. Nelson, N. G. Johnson, P. T. Chinimilli, W. Zhang, Residential cooling using separated and coupled precooling and thermal energy storage strategies, Applied Energy 252 (2019) 113414. [120] S. Naderi, S. Heslop, D. Chen, I. MacGill, G. Pignatta, Cost-saving through precooling: A case study of sydney, Environmental Sciences Proceedings 12 (1) (2021) 2. [121] M. Marwan, M. D. Marwan, Designing a pre-cooling model for air conditioning to avoid an electricity price spike for different building characteristics, Journal of King Saud University-Engineering Sciences (2021). [122] W. J. Cole, K. M. Powell, E. T. Hale, T. F. Edgar, Reduced-order residential home modeling for model predictive control, Energy and Buildings 74 (2014) 69–77. [123] W. J. Turner, J. Roux, I. S. Walker, Reducing residential peak electricity demand with mechanical pre-cooling of building thermal mass, Tech. rep., Lawrence Berkeley National Laboratory (2014). 123 [124] Z. Zeng, W. Zhang, K. Sun, M. Wei, T. Hong, Investigation of pre-cooling as a recommended measure to improve residential buildings’ thermal resilience during heat waves, Building and Environment 210 (2022) 108694. [125] W. Turner, I. Walker, J. Roux, Peak load reductions: Electric load shifting with mechanical pre-cooling of residential buildings with low thermal mass, Energy 82 (2015) 1057–1067. [126] R. Arababadi, K. Parrish, Modeling and testing multiple precooling strategies in three residential building types in the phoenix climate., ASHRAE Transactions 122 (2) (2016). [127] S. Wijesuriya, M. Brandt, P. C. Tabares-Velasco, Parametric analysis of a residential building with phase change material (pcm)-enhanced drywall, precooling, and variable electric rates in a hot and dry climate, Applied energy 222 (2018) 497–514. [128] R. A. Kishore, M. V. Bianchi, C. Booten, J. Vidal, R. Jackson, Modulating thermal load through lightweight residential building walls using thermal energy storage and controlled precooling strategy, Applied Thermal Engineering 180 (2020) 115870. [129] A. H. Dehwah, M. Krarti, Performance of precooling strategies using switchable insulation systems for commercial buildings, Applied Energy 303 (2021) 117631. [130] S. Mayes, K. Sanders, Quantifying the electricity, co2 emissions, and economic tradeoffs of precooling strategies for a single-family home in southern california, Environmental Research: Infrastructure and Sustainability 2 (2) (2022) 025001. [131] H. Stopps, M. F. Touchie, Load shifting and energy conservation using smart thermostats in contemporary high-rise residential buildings: Estimation of runtime changes using field data, Energy and Buildings 255 (2022) 111644. 124 [132] J. Reyna, E. Wilson, A. Satre-Meloy, A. Egerter, C. Bianchi, M. Praprost, A. Speake, L. Liu, A. Parker, R. Horsey, et al., Us building stock characterization study: A national typology for decarbonizing us buildings. part 1: Residential buildings, Tech. rep., National Renewable Energy Lab.(NREL), Golden, CO (United States) (2021). [133] N. R. E. Laboratory, End use load profiles: Building energy models, "https://data. openei.org/submissions/4520" (2021). [134] D. of Energy, Weather data by country, "https://energyplus.net/ weather-region/north_and_central_america_wmo_region_4/USA/CA-Zones" (2022). [135] C. E. Commission, Building energy efficiency standards - title 24, /url”https://www.energy.ca.gov/programs-and-topics/programs/building-energyefficiency-standards” (2022). [136] T. P. E. Center, Guide to california climate zones, "https://www.pge.com/myhome/ edusafety/workshopstraining/pec/toolbox/arch/climate/index.shtml" (2009). [137] CAISO, Grid emergencies history report - 1998 to present, Tech. rep., California Independent System Operator (2022). [138] D. of Energy, Energyplus, "https://energyplus.net/" (2022). [139] D. of Energy, Energyplus documentation, "https://energyplus.net/ documentation" (2022). [140] S. Morgan, M. Krarti, Impact of electricity rate structures on energy cost savings of pre-cooling controls for office buildings, Building and environment 42 (8) (2007) 2810–2818. 125 [141] C. Booten, P. C. Tabares-Velasco, Using energyplus to simulate the dynamic response of a residential building to advanced cooling strategies, Tech. rep., National Renewable Energy Lab.(NREL), Golden, CO (United States) (2012). [142] R. Yin, P. Xu, M. A. Piette, S. Kiliccote, Study on auto-dr and pre-cooling of commercial buildings with thermal mass in california, Energy and Buildings 42 (7) (2010) 967–975. [143] K. Fabbri, Indoor thermal comfort perception, A Questionnaire Approach Focusing on Children; Springer: New York City, NY, USA (2015). [144] Z. Tian, B. Si, Y. Wu, X. Zhou, X. Shi, Multi-objective optimization model predictive dispatch precooling and ceiling fans in office buildings under different summer weather conditions, in: Building Simulation, Vol. 12, Springer, 2019, pp. 999–1012. [145] S. Naderi, G. Pignatta, S. Heslop, I. MacGill, D. Chen, Demand response via precooling and solar pre-cooling: a review, Energy and Buildings (2022) 112340. [146] P. Fanger, Thermal environment—human requirements, Environmentalist 6 (4) (1986) 275–278. [147] ASHRAE, Thermal environmental conditions for human occupancy, Tech. rep., American Societ of Heating, Refrigerating, and Air-Conditioning Engineers (2020). [148] F. Tartarini, S. Schiavon, T. Cheung, T. Hoyt, Cbe thermal comfort tool: Online tool for thermal comfort calculations and visualizations, SoftwareX 12 (2020) 100563. [149] N. A. Ryan, J. X. Johnson, G. A. Keoleian, Comparative assessment of models and methods to calculate grid electricity emissions, Environmental science & technology 50 (17) (2016) 8937–8953. [150] A. D. Hawkes, Long-run marginal co2 emissions factors in national electricity systems, Applied Energy 125 (2014) 197–205. 126 [151] W. Schram, I. Lampropoulos, T. AlSkaif, W. Van Sark, M. Helfert, C. Klein, B. Donnellan, et al., On the use of average versus marginal emission factors, in: SMARTGREENS 2019-Proceedings of the 8th International Conference on Smart Cities and Green ICT Systems, SciTePress, 2019, pp. 187–193. [152] S. C. Edison, Time-of-use (tou) rate plans (2023). URL https://www.sce.com/residential/rates/Time-Of-Use-Residential-Rate-Plans [153] P. Gas, Electric, Time-of-use rate plans, "https://www.pge.com/en_US/ residential/rate-plans/rate-plan-options/time-of-use-base-plan/ time-of-use-plan.page" (2023). [154] S. D. Gas, Electric, Time-of-use plans, "https://www.sdge.com/residential/ pricing-plans/about-our-pricing-plans/whenmatters" (2023). [155] S. M. District, Time-of-day (5-8 p.m.) rate, "https://www.smud.org/ en/Rate-Information/Time-of-Day-rates/Time-of-Day-5-8pm-Rate/ Rate-details" (2023). [156] X. Zhou, Q. Ouyang, Y. Zhu, C. Feng, X. Zhang, Experimental study of the influence of anticipated control on human thermal sensation and thermal comfort, Indoor Air 24 (2) (2014) 171–177. [157] M. Luo, B. Cao, W. Ji, Q. Ouyang, B. Lin, Y. Zhu, The underlying linkage between personal control and thermal comfort: psychological or physical effects?, Energy and Buildings 111 (2016) 56–63. [158] M. Pazhoohesh, C. Zhang, A satisfaction-range approach for achieving thermal comfort level in a shared office, Building and Environment 142 (2018) 312–326. [159] D. Malekpour Koupaei, K. S. Cetin, U. Passe, Stochastic residential occupancy sched127 ules based on the american time-use survey, Science and Technology for the Built Environment 28 (6) (2022) 776–790. [160] C. I. S. Operator, Flex alert news, "https://www.flexalert.org/news" (2022). [161] D. Murtaugh, B. Eckhouse, A text alert may have saved california from power blackouts, "https://www.bloomberg.com/news/articles/2022-09-07/ a-text-alert-may-have-saved-california-from-power-blackouts? leadSource=uverify%20wall" (2022). [162] C. I. S. Operator, Today’s outlook: Emissions, "http://www.caiso.com/ TodaysOutlook/Pages/emissions.html",year="2023". [163] P. Denholm, M. O’Connell, G. Brinkman, J. Jorgenson, Overgeneration from solar energy in california. a field guide to the duck chart, Tech. rep., National Renewable Energy Lab.(NREL), Golden, CO (United States) (2015). [164] A. D. Hawkes, Long-run marginal co2 emissions factors in national electricity systems, Applied Energy 125 (2014) 197–205. [165] A. Regett, F. B¨oing, J. Conrad, S. Fattler, C. Kranner, Emission assessment of electricity: Mix vs. marginal power plant method, in: 2018 15th International Conference on the European Energy Market (EEM), IEEE, 2018, pp. 1–5. [166] C. Samaras, K. Meisterling, Life cycle assessment of greenhouse gas emissions from plug-in hybrid vehicles: implications for policy (2008). [167] S. Mayes, K. Sanders, Quantifying the electricity, co2 emissions, and economic tradeoffs of precooling strategies for a single-family home in southern california, Environmental Research: Infrastructure and Sustainability 2 (2) (2022) 025001. [168] M. Noussan, F. Neirotti, Cross-country comparison of hourly electricity mixes for ev charging profiles, Energies 13 (10) (2020) 2527. 128 [169] A. Zohrabian, K. T. Sanders, Emitting less without curbing usage? exploring greenhouse gas mitigation strategies in the water industry through load shifting, Applied Energy 298 (2021) 117194. [170] G. Kamiya, J. Axsen, C. Crawford, Modeling the ghg emissions intensity of plug-in electric vehicles using short-term and long-term perspectives, Transportation Research Part D: Transport and Environment 69 (2019) 209–223. [171] P. L. Donti, J. Z. Kolter, I. L. Azevedo, How much are we saving after all? characterizing the effects of commonly varying assumptions on emissions and damage estimates in pjm, Environmental science & technology 53 (16) (2019) 9905–9914. [172] T. A. Deetjen, I. L. Azevedo, Reduced-order dispatch model for simulating marginal emissions factors for the united states power sector, Environmental science & technology 53 (17) (2019) 10506–10513. [173] S. Sengupta, T. Spencer, N. Rodrigues, R. Pachouri, S. Thakare, P. J. Adams, R. Tongia, I. M. Azevedo, Current and future estimates of marginal emission factors for indian power generation, Environmental Science & Technology 56 (13) (2022) 9237–9250. [174] Z. Zheng, F. Han, F. Li, J. Zhu, Assessment of marginal emissions factor in power systems under ramp-rate constraints, CSEE Journal of Power and Energy Systems 1 (4) (2015) 37–49. [175] P. Gagnon, W. Cole, Planning for the evolution of the electric grid with a long-run marginal emission rate, Iscience 25 (3) (2022) 103915. [176] A. D. Hawkes, Estimating marginal co2 emissions rates for national electricity systems, Energy Policy 38 (10) (2010) 5977–5987. [177] A. J. Pimm, J. Palczewski, E. R. Barbour, T. T. Cockerill, Using electricity storage to reduce greenhouse gas emissions, Applied Energy 282 (2021) 116199. 129 [178] E. McKenna, J. Barton, M. Thomson, Short-run impact of electricity storage on co2 emissions in power systems with high penetrations of wind power: A case-study of ireland, Proceedings of the Institution of Mechanical Engineers, Part A: Journal of Power and Energy 231 (6) (2017) 590–603. [179] F. Braeuer, R. Finck, R. McKenna, Comparing empirical and model-based approaches for calculating dynamic grid emission factors: An application to co2-minimizing storage dispatch in germany, Journal of cleaner production 266 (2020) 121588. [180] N. Seckinger, P. Radgen, Dynamic prospective average and marginal ghg emission factors—scenario-based method for the german power system until 2050, Energies 14 (9) (2021) 2527. [181] M. P. Thind, E. J. Wilson, I. L. Azevedo, J. D. Marshall, Marginal emissions factors for electricity generation in the midcontinent iso, Environmental science & technology 51 (24) (2017) 14445–14452. [182] A. Zohrabian, S. Mayes, K. T. Sanders, A data-driven framework for quantifying consumption-based monthly and hourly marginal emissions factors, Journal of Cleaner Production 396 (2023) 136296. [183] C. I. S. Operator, Managing oversupply (2023). URL http://www.caiso.com/informed/Pages/ManagingOversupply.aspx [184] R. C. Thomson, G. P. Harrison, J. P. Chick, Marginal greenhouse gas emissions displacement of wind power in great britain, Energy Policy 101 (2017) 201–210. [185] F. Beltrami, A. Burlinson, M. Giulietti, L. Grossi, P. Rowley, G. Wilson, Where did the time (series) go? estimation of marginal emission factors with autoregressive components, Energy Economics 91 (2020) 104905. 130 [186] D. C. Park, M. El-Sharkawi, R. Marks, L. Atlas, M. Damborg, Electric load forecasting using an artificial neural network, IEEE transactions on Power Systems 6 (2) (1991) 442–449. [187] D. Singhal, K. Swarup, Electricity price forecasting using artificial neural networks, International Journal of Electrical Power & Energy Systems 33 (3) (2011) 550–555. [188] J. P. d. S. Catal˜ao, S. J. P. S. Mariano, V. Mendes, L. Ferreira, Short-term electricity prices forecasting in a competitive market: A neural network approach, electric power systems research 77 (10) (2007) 1297–1304. [189] Y. T. Chae, R. Horesh, Y. Hwang, Y. M. Lee, Artificial neural network model for forecasting sub-hourly electricity usage in commercial buildings, Energy and Buildings 111 (2016) 184–194. [190] U. E. P. Agency, Clean air markets program data (2023). [191] U.S. Energy Information Administration (EIA), Annual Electric Generator data - EIA860 data file, www.eia.gov/electricity/data/eia860/ (2019). [192] U. EIA, Electricity data, "https://www.eia.gov/electricity/data.php" (2022). [193] J. A. de Chalendar, J. Taggart, S. M. Benson, Tracking emissions in the us electricity system, Proceedings of the National Academy of Sciences 116 (51) (2019) 25497–25502. [194] C. E. Commission, Total system electric generation, "https://www.energy.ca.gov/ media/7311" (2023). [195] S. Lundberg, Shap.deepexplainer, "https://shap-lrjball.readthedocs.io/en/ latest/generated/shap.DeepExplainer.html" (2023). [196] K. Zhang, P. Xu, J. Zhang, Explainable ai in deep reinforcement learning models: A shap method applied in power system emergency control, in: 2020 IEEE 4th Conference on Energy Internet and Energy System Integration (EI2), IEEE, 2020, pp. 711–716. 131 [197] S. M. Lundberg, S.-I. Lee, A unified approach to interpreting model predictions, Advances in neural information processing systems 30 (2017). [198] CAISO, Open acess same-time information system, "http://oasis.caiso.com/ mrioasis/logon.do",year="2023". [199] Y. Wang, L. Li, Time-of-use electricity pricing for industrial customers: A survey of us utilities, Applied Energy 149 (2015) 89–103. [200] H. Allcott, Real time pricing and electricity markets, Harvard University 7 (2009). [201] D. Marikyan, S. Papagiannidis, E. Alamanos, A systematic review of the smart home literature: A user perspective, Technological Forecasting and Social Change 138 (2019) 139–154. [202] P. R. Brown, A. Botterud, The value of inter-regional coordination and transmission in decarbonizing the us electricity system, Joule 5 (1) (2021) 115–134. [203] S. Seabold, J. Perktold, Statsmodels: Econometric and statistical modeling with python, in: 9th Python in Science Conference, 2010. 132 Appendices A Supplemental information for Chapter 1 A.1 Building Properties Table A1: Building properties of the residential building prototype used in the precooling simulations. The prototype was designed by the Pacific Northwest National Laboratory. This table, and information about the development of the prototypes can be found at https://www.energycodes.gov/development/residential/iecc models Parameter Assumption Conditioned floor area 2,376 sq feet Footprint and height 54-ft-by-22 ft, two-story, 8.5-ft-high ceilings Area above unconditioned space 1,188 sq feet Area below roof/ceilings 1,188 sq feet Perimeter length 152 feet Gross exterior wall area 2,584 sq feet Window area (relative to conditioned floor area) Fifteen percent equally distributed to the four cardinal directions Door area 42 sq feet Internal gains 86,761 Btu/day Heating system Natural gas furnace, heat pump, electric furnace, or oil-fired furnace Cooling system Central electric air conditioning Water heating Same as fuel used for space heating, or as required to evaluate domestic hot water-specific code changes A.2 Climate Zone Information The EnergyPlus simulations performed in this study were done in California Climate Zone 9. Climate Zone 9 has a reference city of Los Angeles, and the area covered by the zone has a significant overlap with Los Angeles County. The Pacific Gas and Electric Company, a utility servicing California describes this climate zone as being impacted by both coastal and interior weather patterns, and states that there are a similar number of heating and cooling degree days. They characterize this zone as having warmer summers and cooler winters than more coastal areas (136). 133 A.3 Hourly Average Emissions Factors Table A2: The full set of hourly average emissions factors (AEFs) for CO2 for the California Independent System Operator (CAISO). The AEFs were calculated using 2019 data available for public download on the CAISO website (1) with a regression analysis. Recreated from Zohrabian, Sanders. (169). Hour Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec 1 318 294 262 225 215 241 279 317 337 346 346 340 2 317 293 258 220 214 235 274 310 333 343 343 340 3 317 294 255 218 213 233 269 305 331 341 341 339 4 316 293 253 216 214 233 266 302 328 340 340 336 5 314 291 252 217 213 236 264 300 324 337 337 332 6 314 290 255 230 222 242 266 300 321 331 331 327 7 316 297 267 249 225 233 262 305 329 335 335 329 8 314 280 259 218 186 188 225 271 310 330 330 326 9 274 228 203 155 148 153 189 221 250 266 266 296 10 242 202 158 123 132 142 177 201 218 214 214 273 11 228 187 137 109 122 139 173 195 211 200 200 265 12 220 186 128 103 114 138 173 195 211 195 195 260 13 221 186 124 100 108 136 174 199 213 194 194 259 14 222 188 126 101 108 139 180 207 220 197 197 263 15 231 194 127 101 109 142 185 216 227 198 198 272 16 255 211 136 108 112 154 194 227 238 203 203 291 17 298 250 159 123 119 163 204 238 253 247 247 315 18 320 293 201 149 137 180 219 257 284 311 311 332 19 325 310 258 212 183 215 251 298 332 354 354 335 20 325 309 287 265 235 259 293 333 351 362 362 335 21 327 308 288 276 251 275 306 339 354 360 360 336 22 330 308 281 263 245 272 306 338 356 358 358 339 23 328 302 274 249 232 260 299 335 355 356 356 341 24 323 298 268 235 221 248 292 326 343 350 350 340 To calculate the AEFs, data was binned by month and hour of the day, after which emissions were regressed on electricity load to determine the hourly AEF for each hour of the day in each month. Our precooling study spanned five month and thus used columns May through October of this table. Note that evening CO2 emissions intensities can be over 1.5x greater than midday intensities. 134 The regression model used was an ordinary least squares model, implemented with the python package statsmodels (203). The coefficients calculated for each month and hour of the year pairing show both high coefficients of determination (high R2 ), as well as high significance levels (low p-value). The R2 values corresponding to each coefficient (and thus to each AEF) are included in the table below. Table A3: The R2 values corresponding to each AEF calculated above, which correspond to a coefficient in the binned linear regression model. These values are consistently close to 1.00, with a minimum value of 0.86, which implies a strong correlation between emissions and demand for a given hour h of month m (or equivalently, after grouping by an hour of the day and month of the year pairing, the variations in emissions are highly explainable by variations in demand). Hour Jan Feb Mar Apr May June July Aug Sep Oct Nov Dec 1 0.996 0.996 0.991 0.986 0.993 0.990 0.983 0.997 0.994 0.994 0.998 0.995 2 0.996 0.996 0.991 0.985 0.992 0.990 0.981 0.997 0.993 0.993 0.999 0.996 3 0.995 0.996 0.990 0.984 0.993 0.991 0.980 0.997 0.992 0.992 0.999 0.996 4 0.996 0.996 0.989 0.984 0.993 0.993 0.979 0.997 0.992 0.992 0.998 0.997 5 0.997 0.996 0.991 0.984 0.991 0.994 0.979 0.997 0.992 0.993 0.998 0.996 6 0.997 0.995 0.993 0.985 0.991 0.994 0.980 0.997 0.993 0.994 0.998 0.996 7 0.996 0.994 0.992 0.983 0.987 0.993 0.976 0.997 0.994 0.996 0.998 0.995 8 0.996 0.993 0.983 0.975 0.974 0.986 0.957 0.993 0.991 0.987 0.997 0.996 9 0.987 0.977 0.964 0.945 0.940 0.974 0.931 0.985 0.980 0.990 0.992 0.989 10 0.966 0.959 0.918 0.912 0.925 0.962 0.919 0.980 0.968 0.986 0.988 0.975 11 0.958 0.959 0.902 0.889 0.922 0.941 0.915 0.979 0.960 0.978 0.985 0.971 12 0.953 0.961 0.890 0.881 0.923 0.931 0.912 0.978 0.966 0.974 0.989 0.972 13 0.953 0.957 0.875 0.871 0.915 0.911 0.906 0.979 0.965 0.972 0.987 0.971 14 0.958 0.952 0.870 0.856 0.903 0.904 0.904 0.980 0.965 0.969 0.987 0.971 15 0.965 0.948 0.865 0.864 0.898 0.900 0.913 0.981 0.965 0.967 0.989 0.975 16 0.981 0.956 0.881 0.868 0.902 0.904 0.924 0.984 0.964 0.974 0.994 0.988 17 0.997 0.982 0.895 0.897 0.918 0.914 0.937 0.988 0.972 0.981 0.995 0.996 18 0.998 0.996 0.926 0.937 0.949 0.931 0.951 0.990 0.983 0.985 0.996 0.996 19 0.998 0.997 0.984 0.977 0.979 0.958 0.968 0.994 0.993 0.984 0.996 0.996 20 0.998 0.998 0.995 0.991 0.991 0.981 0.985 0.997 0.995 0.983 0.997 0.997 21 0.998 0.998 0.995 0.992 0.994 0.985 0.989 0.998 0.995 0.995 0.998 0.997 22 0.997 0.998 0.995 0.993 0.995 0.986 0.988 0.998 0.995 0.996 0.998 0.997 23 0.996 0.998 0.994 0.993 0.995 0.989 0.986 0.998 0.995 0.996 0.998 0.997 24 0.996 0.997 0.993 0.990 0.994 0.990 0.984 0.998 0.995 0.996 0.998 0.997 Note: The p-values for each of these AEFs is far below the 0.05 significance threshold, with a largest p-value of order 10−14 . An alternative way to calculate an AEF for a specific hour of a specific month would be 135 to sum all of the emissions that occurred in that specific hour over the course of a month (e.g. all 31 hour 1s in January), and divide by the sum of the electricity demanded in these same time periods (note that this is mathematically equivalent to calculating the AEF for each hour and each day in a month, and then averaging all of the values for the same hour of the day in that month). This cumulative method produces similar results to the regression model used in this analysis, but the regression method was considered preferable due to its use of a squared error term, which is more responsive to data points that fall farther away from the average value. A.4 Simulation Results As discussed in the main text, the scope of this precooling study was limited to June-October, the five warmest months of the year for the study area of Los Angeles, CA. In general, electricity consumption was more strongly impacted in warmer months than cooler ones. This effect includes both larger increases in electricity consumption during the precooling period and larger decreases during the peak demand period in months with higher temperatures. 136 0.4 0.2 0.0 0.2 0.4 E0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 0.4 0.2 0.0 0.2 0.4 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 Change in E- Consumption (kWh) 0.50 0.25 0.00 0.25 0.50 PMV PMV 0.50 0.25 0.00 0.25 0.50 PMV June 0.10 0.05 0.00 0.05 CO2 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 0.10 0.05 0.00 0.05 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 Change in CO2 Emissions (kg) 0.50 0.25 0.00 0.25 0.50 PMV PMV 0.50 0.25 0.00 0.25 0.50 PMV June 137 0.3 0.2 0.1 0.0 0.1 $ 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 0.3 0.2 0.1 0.0 0.1 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 Change in Cost (USD) 0.50 0.25 0.00 0.25 0.50 PMV PMV 0.50 0.25 0.00 0.25 0.50 PMV June 0.50 0.25 0.00 0.25 0.50 E0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 0.50 0.25 0.00 0.25 0.50 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 Change in E- Consumption (kWh) 0.50 0.25 0.00 0.25 0.50 PMV PMV 0.50 0.25 0.00 0.25 0.50 PMV July 138 0.15 0.10 0.05 0.00 0.05 0.10 CO2 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 0.15 0.10 0.05 0.00 0.05 0.10 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 Change in CO2 Emissions (kg) 0.50 0.25 0.00 0.25 0.50 PMV PMV 0.50 0.25 0.00 0.25 0.50 PMV July 0.3 0.2 0.1 0.0 0.1 $ 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 0.3 0.2 0.1 0.0 0.1 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 Change in Cost (USD) 0.50 0.25 0.00 0.25 0.50 PMV PMV 0.50 0.25 0.00 0.25 0.50 PMV July 139 0.50 0.25 0.00 0.25 0.50 E0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 0.50 0.25 0.00 0.25 0.50 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 Change in E- Consumption (kWh) 0.50 0.25 0.00 0.25 0.50 PMV PMV 0.50 0.25 0.00 0.25 0.50 PMV August 0.1 0.0 0.1 CO2 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 0.1 0.0 0.1 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 Change in CO2 Emissions (kg) 0.50 0.25 0.00 0.25 0.50 PMV PMV 0.50 0.25 0.00 0.25 0.50 PMV August 140 0.3 0.2 0.1 0.0 0.1 $ 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 0.3 0.2 0.1 0.0 0.1 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 Change in Cost (USD) 0.50 0.25 0.00 0.25 0.50 PMV PMV 0.50 0.25 0.00 0.25 0.50 PMV August 0.6 0.4 0.2 0.0 0.2 0.4 E0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 0.6 0.4 0.2 0.0 0.2 0.4 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 Change in E- Consumption (kWh) 0.50 0.25 0.00 0.25 0.50 PMV PMV 0.50 0.25 0.00 0.25 0.50 PMV September 141 0.1 0.0 0.1 CO2 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 0.1 0.0 0.1 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 Change in CO2 Emissions (kg) 0.50 0.25 0.00 0.25 0.50 PMV PMV 0.50 0.25 0.00 0.25 0.50 PMV September 0.3 0.2 0.1 0.0 0.1 $ 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 0.3 0.2 0.1 0.0 0.1 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 Change in Cost (USD) 0.50 0.25 0.00 0.25 0.50 PMV PMV 0.50 0.25 0.00 0.25 0.50 PMV September 142 0.4 0.2 0.0 0.2 0.4 E0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 0.4 0.2 0.0 0.2 0.4 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 Change in E- Consumption (kWh) 0.50 0.25 0.00 0.25 0.50 PMV PMV 0.50 0.25 0.00 0.25 0.50 PMV October 0.10 0.05 0.00 0.05 0.10 CO2 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 0.10 0.05 0.00 0.05 0.10 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 Change in CO2 Emissions (kg) 0.50 0.25 0.00 0.25 0.50 PMV PMV 0.50 0.25 0.00 0.25 0.50 PMV October 143 0.2 0.1 0.0 0.1 $ 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 0.2 0.1 0.0 0.1 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 Change in Cost (USD) 0.50 0.25 0.00 0.25 0.50 PMV PMV 0.50 0.25 0.00 0.25 0.50 PMV October Figure A.1: Top of each sub-figure: the changes in average hourly electricity consumption for a shallow precool (4 h length, 1 °F depth, 3 °F offset) scenario, as compared to the 75 °F baseline simulation, are shown in blue for an average day in each month June through October. Bottom of each sub-figure: the same is shown for a deeper precool (3 h length, 3 °F depth, 3 °F offset). Hourly PMV is shown in green, which stays between the +0.5 and −0.5 limits for both schedules in each month. 144 Table A4: Percent change in cumulative peak period cooling electricity consumption, residential electricity cost, and cooling associated CO2 emissions over the five-month simulation period between each of the 70 precooling schedules and the baseline schedule of 75 oF. These 70 schedules represent all of the precooling schedules out of the 504 that were simulated that maintain thermal comfort (PMV between -0.5 and +0.5) for occupants during the peak period. The precooling hours refers to the number of hours before 5pm that precooling occurred, the depth refers to the number of degrees Fahrenheit below 75 that AC was set at during precooling, and the offset refers to the number of degrees Fahrenheit above the baseline temperature that the AC was set at during the peak demand period (5-8pm). Offset (F) Precooling Hours (h) Precooling Depth (F) Change in Cumulative Peak Period Cooling Electricity Consumption (%) Change in Cumulative Residential Electricity Cost (%) Change in Cumulative Cooling-associated CO2 Emissions (%) 0 1 0 0.00 0.00 0.00 1 1 0 -16.49 -4.04 -1.31 2 1 0 -33.05 -8.05 -2.47 0 1 1 -6.04 -0.02 0.87 1 1 1 -23.20 -4.43 -0.77 2 1 1 -39.65 -8.47 -2.02 3 1 1 -55.36 -12.18 -2.99 0 1 2 -12.50 -0.05 1.83 1 1 2 -29.85 -4.54 0.12 2 1 2 -46.27 -8.65 -1.27 3 1 2 -60.94 -11.98 -1.96 0 1 3 -18.85 0.47 3.44 1 1 3 -35.52 -3.78 1.85 2 1 3 -51.87 -7.92 0.37 3 1 3 -65.45 -10.88 -0.10 0 1 4 -24.82 1.13 5.10 1 1 4 -42.17 -3.35 3.37 2 1 4 -57.04 -7.05 2.07 3 1 4 -69.72 -9.71 1.77 0 2 1 -8.38 -0.69 0.30 1 2 1 -25.39 -5.06 -1.32 2 2 1 -41.61 -9.02 -2.51 3 2 1 -56.85 -12.60 -3.45 0 2 2 -16.12 0.05 2.16 1 2 2 -33.09 -4.31 0.53 2 2 2 -49.48 -8.41 -0.85 3 2 2 -63.42 -11.51 -1.43 0 2 3 -23.96 0.86 4.12 1 2 3 -40.83 -3.47 2.48 2 2 3 -56.10 -7.26 1.19 3 2 3 -68.80 -9.93 0.88 0 3 1 -8.52 0.25 1.23 1 3 1 -25.53 -4.11 -0.37 2 3 1 -41.81 -8.06 -1.54 145 Offset (F) Precooling Hours (h) Precooling Depth (F) Change in Cumulative Peak Period Cooling Electricity Consumption (%) Change in Cumulative Residential Electricity Cost (%) Change in Cumulative Cooling-associated CO2 Emissions (%) 3 3 1 -56.98 -11.67 -2.53 0 3 2 -17.87 0.25 2.30 1 3 2 -34.70 -4.07 0.69 2 3 2 -50.73 -8.04 -0.62 3 3 2 -64.53 -11.16 -1.26 0 3 3 -26.72 0.98 4.13 1 3 3 -43.28 -3.27 2.52 2 3 3 -58.05 -6.89 1.34 3 3 3 -70.42 -9.48 1.04 0 4 1 -9.31 -0.01 0.92 1 4 1 -26.12 -4.31 -0.65 2 4 1 -42.23 -8.22 -1.80 3 4 1 -57.33 -11.90 -2.92 0 4 2 -18.46 1.04 2.99 1 4 2 -35.15 -3.25 1.40 2 4 2 -51.11 -7.19 0.12 3 4 2 -64.80 -10.24 -0.47 1 4 3 -44.93 -2.95 2.66 2 4 3 -59.30 -6.42 1.58 3 4 3 -71.40 -8.96 1.29 0 5 1 -9.66 0.33 1.20 1 5 1 -26.48 -3.97 -0.37 2 5 1 -42.60 -7.89 -1.53 3 5 1 -57.59 -11.52 -2.63 0 5 2 -19.84 0.32 2.09 1 5 2 -36.18 -3.85 0.55 2 5 2 -52.00 -7.74 -0.67 3 5 2 -65.37 -10.74 -1.30 0 6 1 -9.88 0.68 1.50 1 6 1 -26.70 -3.62 -0.07 2 6 1 -42.79 -7.53 -1.23 3 6 1 -57.78 -11.16 -2.32 0 6 2 -20.31 1.07 2.74 1 6 2 -36.66 -3.11 1.19 2 6 2 -52.43 -6.97 -0.01 3 6 2 -65.74 -9.98 -0.66 146 B Supplemental information for Chapter 2 B.1 Summary of relevant literature Complementary to the introduction section, a summary is provided comparing the most relevant regression-based literature in quantifying MEFs. Table B1: Summary of major regression-based studies evaluating MEFs for different electric grids # Study Approach MEF predicting factors Reported MEF granularity Geographic Area Assessment Method 1 Siler-Evans et al., (79) Generation based Changes in fossil fuel-based generation Hourly, monthly, and annual United States Single linear regression 2 Hawkes, (74) Consumption based Changes in system load Hourly, monthly, annual, and load level Great Britain Single linear regression 3 Gai et al., (76) Generation based Changes in total generation Month-hour Greater Toronto and Hamilton Area in Ontario, Canada Single and multiple linear regression 4 Li et al., (72) Generation Based Changes in emitting and non-emitting generation Month-hour Midcontinent Independent System Operator Single linear regression 5 Donti et al., (77) Generation based Changes in fossil fuel and nonemitting generation Annual, monthhour PJM Interconnection and the Reliability First Corporation in the US Single linear regression 6 Thind et al., (78) Generation based Changes in total generation Period of day, dayof-week, monthly, and annual Midcontinent Independent System Operator Single linear regression model 7 Seckinger and Radgen, (80) Consumption based Changes in dispatchable generation (primarily fossil fuels) Annual Germany Single linear regression model 8 Sengupta et al., (75) Generation based Changes in total generation Seasonal, period of day India Single linear regression model 9 Thomson et al., (82) Consumption based Changes in demand and wind generation Annual and annual average hourly Great Britain Multiple linear regression 10 This study Consumption based Changes in electricity demand and renewable generation Month-hour average, and demand level California Independent System Operator Multiple linear regression B.2 Data Repository A collection of data and regression results used to generate the figures in this manuscript can be found at data.mendeley.com/datasets/7w87xy5pwj/2 (DOI: 10.17632/7w87xy5pwj.2). This repository contains the raw data used for regression, as well as a table that summarizes the regression results, including coefficents, R-squared values, P-Values, and Variance Inflation Factors (as a multicollinearity check) for each individual regression result. There are also a series of figures that scatter the model prediction errors against the predicted values as a scedasticity check. 147 B.3 CAISO Electricity Trades The geographic distribution of the balancing authorities that CAISO exchanges electricity with are depicted in Figure B.1. Figure B.1: California Independent System Operator (CAISO) has many direct exchange of electricity with many other balancing authorities in Western Electricity Coordinating Council mainly in the Southwest and Northwest regions. B.4 Outlier Analysis Despite the improvements in the estimation of hourly emissions associated with incorporating imports and exports, a number of outliers remained in the hourly emissions data in both 2019 and 2020. Many of these outliers represent hourly data pulled directly from the EPA 148 CEMS dataset, while others may be caused by the estimating and scaling techniques used when the CEMS dataset was missing data for specific hours. Outliers were identified by comparing the changes in emissions in consecutive hours with the normal range of changes in emissions in the same consecutive hours for the same month. For example, emissions tend to increase significantly in the late afternoon, so a large decrease in emissions between 4 and 5pm is cause for suspicion. These points were identified as outliers when a large change in emissions (greater than 1.5x the inner quartile range) in one direction was quickly followed by a large change in the opposite direction, as shown in Figure B.2. 0 5 10 15 20 25 Hour of Day 500 1000 1500 2000 2500 3000 Tons of CO2 Figure B.2: Hourly Emissions for March 23, 2019. The emissions data reported for hour 15 are an outlier. These outliers are not explained by changes in grid mix or demand and are eliminated for the calculation of marginal and average emissions factors. Note that when working with changes in hourly emissions, each outlier creates at minimum two data points that must be eliminated, the hour of the outlier itself as well as the hour immediately after the outlier. 149 This resulted in eliminating 37 data points from 2019 (0.4% of hours) and 46 from 2020 (0.5% of hours). B.5 Monthly MEFs Figure B.3 shows the MEFs for each demand level organized by month of the year. As discussed in the main text, MEF is typically higher for higher demand levels. In Figure B.3, it is apparent in both the higher MEFs that occur in months with higher demand and higher MEFs within a specific month for the higher demand levels. 0 1 2 3 4 5 6 7 8 9 10 11 12 Month 0 100 200 300 400 500 600 MEF in kgCO2/MWh 2019 Demand Level 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 10 11 12 Month 0 100 200 300 400 500 600 MEF in kgCO2/MWh 2020 Demand Level 1 2 3 4 5 6 7 8 9 Figure B.3: MEF values across all demand levels in each month of the years 2019 and 2020. B.6 Coefficient of Determination for MEFs In general, this study found stronger regression results for MEFs at higher demand levels, an important result given that these are the most polluting times (highest MEFs). The coefficient of determination is lowest for high-hydro, low-demand months, when it is difficult to predict if marginal generation will be met by hydropower or traditional resources. See Figure B.4 150 0 1 2 3 4 5 6 7 8 9 Demand Level 0 100 200 300 400 500 600 MEF in kgCO2/MWh 2019 0.4 0.5 0.6 0.7 0.8 0.9 R^2 0 1 2 3 4 5 6 7 8 9 Demand Level 0 100 200 300 400 500 600 MEF in kgCO2/MWh 2020 0.5 0.6 0.7 0.8 0.9 R^2 Figure B.4: MEF values for each demand level occurring in each month of the year in 2019 and 2020 and corresponding R2 values. B.7 Changes in MEFs MEFs at the month-hourly level were often lower in 2020 than in 2019 (shown in Figure B.5), which evidently is a result of the increased use of hydropower as a marginal generator in 2020. This hypothesis is supported by the fact that many of the month-hour combinations that saw a decrease in MEF between 2019 and 2020 were in the high hydropower production season. B.8 MEF-AEF Comparison In 2019, the MEFs were consistently higher than the AEFs for a given month-hour pair (displayed in Figure B.6). In 2020 the MEFs were usually higher, but more often than in 2019 they were at a similar level or lower than the AEFs. This again can be explained by hydropower being primarily a baseload resource (2019) or a load-following resource (2020), as well as a reflection of the low hydropower production levels in 2020 which increased the annual AEF. 151 Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Month 0 2 4 6 8 10 12 14 16 18 20 22 Hour 100 75 50 25 0 25 50 75 100 % Change Figure B.5: Percent relative change in 2020 month-hour MEFs compared to 2019. Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Month 0 2 4 6 8 10 12 14 16 18 20 22 Hour 100 50 0 50 100 150 200 % Change Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Month 0 2 4 6 8 10 12 14 16 18 20 22 Hour 100 50 0 50 100 150 200 % Change Figure B.6: Percent difference between MEF and AEF in 2019 and 2020. B.9 Comparison of Generation-Based vs Consumption-Based MEFs More detailed comparison of month-hour MEFs estimated in this study using a consumptionbased approach is compared with generation-based MEFs reported by CEDM (95). Across 152 vast majority of hours, MEFs for the year 2018 from CEDM were higher than this study’s estimates for 2019. 0 100 200 300 400 500 600 0 24 48 72 96 120 144 168 192 216 240 264 288 Month-Hour MEFs for CAISO (kgCO2/MWh) 2019 (This Study) 2018 (CEDM) Figure B.7: Comparison of this study’s month-hour MEFs with CEDM’s estimated MEFs. Data for CEDM’s MEFs are from (95) C Supplemental information for Chapter 3 C.1 Building Properties The four single-family homes simulated feature a range of window and wall insulation levels as well as AC efficiencies. Higher R-values indicate better insulation, and higher SEER rating indicates a more efficient AC system. Table B2: Key Properties of Selected ResStock Buildings Building Ceiling R-Value Wall R-Value AC Efficiency Building 1 (B1) R-13 None SEER 10 Building 2 (B2) R-19 R-11 SEER 13 Building 3 (B3) R-30 R-15 SEER 10 Building 4 (B4) R-49 R-19 SEER 13 153 C.2 MEF and AEF Calculations As mentioned in Section 4.3.6, the calculation of Average and Marginal Emissions Factors is performed following the methodology of Zohrabian, Mayes, and Sanders (105) and is described in brief here. AEFs were calculated by grouping the 2021 data by month and year (288 groups), and then regressing hourly emissions on hourly demand for electricity (with an intercept term). This resulted in 288 unique AEFs, with all hours of the same month sharing the same AEF. MEFs were calculated by by grouping the 2021 data by month and demand level (10 unique demand levels), and regressing hour-to-hour changes in emissions on changes in demand and supply of variable renewable energy (and an intercept term). The coefficient of the demand term is taken as the MEF for that hour and demand level combination, and the results were then converted to the month-hourly level (to match the AEFs) by taking a probability-weighted sum; the MEF for a month-hour depended on the likelihood that an hour would fall in a demand bin and the MEF associated with that demand bin. Refer to the cited paper for more extensive explanation of the AEF and MEF calculation process, as well as to the data repository (https://data.mendeley.com/datasets/333jfmpvb8/1) for the hourly data used to calculate these AEFs and MEFs. C.3 Additional Results: The impact of building properties and Climate Zone We expand on the results from Section 4.4.2 by including results for additional climate zones and building types. 154 0 5 10 15 20 Peak Period Electricity Consumption (kWh/day) 0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5 CO2 Emissions (kg/day) CZ3 CZ7 CZ9 CZ13 CZ15 Building 2 0 100 200 300 400 500 600 Cost ($/month) 0 2 4 6 8 10 Peak Period Electricity Consumption (kWh/day) 0 2 4 6 8 CO2 Emissions (kg/day) CZ3 CZ7 CZ9 CZ13 CZ15 5h 3F Building 3 0 50 100 150 200 250 300 Cost ($/month) Figure C.8: The emissions, peak period electricity consumption, and residential electricity costs for the baseline cooling schedule as well as the precooling schedule that maximized CO2 reductions and the one that maximized peak period electricity reductions for Buildings 2 and 3 in five different California building climate zones. The length and depth of the schedule are shown as the width and height of the cross-hairs. 0.00 0.25 0.50 0.75 1.00 1.25 1.50 1.75 Peak Period Electricity Consumption (kWh/day) 0.1 0.2 0.3 0.4 0.5 0.6 0.7 CO2 Emissions (kg/day) B1 B2 B3 B4 CZ1 0 5 10 15 20 25 30 Cost ($/month) 0 2 4 6 8 10 Peak Period Electricity Consumption (kWh/day) 1 2 3 4 5 6 CO2 Emissions (kg/day) B1 B2 B3 B4 CZ2 0 50 100 150 200 Cost ($/month) 0 2 4 6 8 10 Peak Period Electricity Consumption (kWh/day) 2 3 4 5 6 CO2 Emissions (kg/day) B1 B2 B3 B4 CZ4 0 50 100 150 200 Cost ($/month) 0 1 2 3 4 Peak Period Electricity Consumption (kWh/day) 0.50 0.75 1.00 1.25 1.50 1.75 2.00 2.25 2.50 CO2 Emissions (kg/day) B1 B2 B3 B4 CZ5 0 20 40 60 80 Cost ($/month) 155 0 1 2 3 4 5 6 Peak Period Electricity Consumption (kWh/day) 1.5 2.0 2.5 3.0 3.5 4.0 CO2 Emissions (kg/day) B1 B2 B3 B4 CZ6 0 20 40 60 80 100 120 140 160 Cost ($/month) 0 2 4 6 8 Peak Period Electricity Consumption (kWh/day) 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 CO2 Emissions (kg/day) B1 B2 B3 B4 CZ7 0 25 50 75 100 125 150 175 200 Cost ($/month) 0 2 4 6 8 10 Peak Period Electricity Consumption (kWh/day) 2 3 4 5 6 7 CO2 Emissions (kg/day) B1 B2 B3 B4 CZ8 0 50 100 150 200 250 Cost ($/month) 0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5 Peak Period Electricity Consumption (kWh/day) 4 6 8 10 CO2 Emissions (kg/day) B1 B2 B3 B4 CZ10 0 50 100 150 200 250 300 350 400 Cost ($/month) 0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5 20.0 Peak Period Electricity Consumption (kWh/day) 2 4 6 8 10 CO2 Emissions (kg/day) B1 B2 B3 B4 CZ11 0 50 100 150 200 250 300 350 400 Cost ($/month) 0 2 4 6 8 10 12 14 16 Peak Period Electricity Consumption (kWh/day) 2 3 4 5 6 7 8 CO2 Emissions (kg/day) B1 B2 B3 B4 CZ12 0 50 100 150 200 250 300 Cost ($/month) 156 2.5 5.0 7.5 10.0 12.5 15.0 17.5 20.0 Peak Period Electricity Consumption (kWh/day) 4 6 8 10 12 14 CO2 Emissions (kg/day) B1 B2 B3 B4 CZ14 0 100 200 300 400 500 Cost ($/month) Figure C.9: The emissions, peak period electricity consumption, and residential electricity costs for the baseline cooling schedule as well as the precooling schedule that maximized CO2 reductions and the one that maximized peak period electricity reductions for each of the four single-family home designs in various California building climate zones. The length and depth of the schedule are shown as the width and height of the cross-hairs. 157

Abstract (if available)

Abstract As grids across the U.S. transition to higher penetrations of variable renewable energy (VRE), there is a need for new strategies designed to change when end-users consume electricity. Strategies that shift the timing of electricity demand fall in the category of demand-side management (DSM) and have primarily been used to reduce peak electricity demand, thus reducing costs, and providing grid reliability benefits. Increasing penetrations of VRE have also created the potential to use DSM to reduce emissions, because the generation resources that are used to meet demand, and their associated emissions, vary throughout the day and year. Researchers often rely on marginal emissions factors (MEFs) to determine the impact of changes in demand on emissions, but existing methods of calculating MEFs fail to properly account for modern, high-renewable grids and lack the properties that would make them most useful for DSM. This body of work introduces a new methodology for calculating MEFs that incorporates both renewable generation sources and electricity trades between regions, and results in estimates that have higher temporal resolution, are more accurate, and can be forecasted, improving significantly upon existing statistical methods. These emissions factors quantify the effect that implementing a DSM strategy would have on emissions and can therefore inform DSM program design. This dissertation also applies MEFs by examining the impact of precooling (a DSM strategy that shifts AC load) on single-family homes across California to evaluate precooling as an emissions-reduction strategy. The results show that precooling can effectively reduce emissions, in addition to providing traditional DSM management benefits, highlighting the potential of MEFs to be used to create lower-emissions consumption patterns.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Residential electricity demand in the context of urban warming: leveraging high resolution smart meter data to quantify spatial and temporal patterns…

PDF

Evaluating the role of energy system decarbonization and land cover properties on urban air quality in southern California

PDF

Developing frameworks to quantify the operational and environmental performance of energy systems within the context of climate change

PDF

Developing high-resolution spatiotemporal methods to model and quantify water use for energy

PDF

Investigating the role of urban emission sources on redox-active PM compounds and the chemical analysis of the standardized diesel exhaust particles

PDF

Investigating the role of climate in affecting residential electricity consumption through high spatiotemporal resolution observations

PDF

From tugboats to trees: investigating the coupled systems of urban air pollution and meteorology

PDF

Beyond greenhouse gases and towards urban-scale climate mitigation: understanding the roles of black carbon aerosols and the urban heat island effect as local to regional radiative forcing agents

PDF

Chemical and toxicological characteristics of particulate matter in urban environments with a focus on its sources, associated health impacts and mitigation policies

PDF

Integrating systems of desalination and potable reuse: reduced energy consumption for increased water supply

PDF

Environmental effects from a large-scale adoption of electric vehicle technology in the City of Los Angeles

PDF

The smart grid network: pricing, markets and incentives

PDF

The next generation of power-system operations: modeling and optimization innovations to mitigate renewable uncertainty

PDF

User-centric smart sensing for non-intrusive electricity consumption disaggregation in buildings

PDF

Demand response management in smart grid from distributed optimization perspective

PDF

Utility – customer interface strategy and user application for electric vehicles and distributed generation management

PDF

Mechanics of abrasion emissions of particulate matter and microplastics

PDF

Semiconductor devices for vacuum electronics, electrochemical reactions, and ultra-low power in-sensor computing

PDF

Emissions markets, power markets and market power: a study of the interactions between contemporary emissions markets and deregulated electricity markets

$Hydraulic fracturing and the environment: risk assessment for groundwater contamination from well casing failure$

PDF

Hydraulic fracturing and the environment: risk assessment for groundwater contamination from well casing failure

Asset Metadata

Creator Mayes, Stepp A. (author)

Core Title Using demand-side management for decarbonization: developing methods to quantify the impact of altering electricity consumption patterns

School Viterbi School of Engineering

Degree Doctor of Philosophy

Degree Program Environmental Engineering

Degree Conferral Date 2023-12

Publication Date 12/21/2023

Defense Date 12/21/2023

Publisher Los Angeles, California (original), University of Southern California (original), University of Southern California. Libraries (digital)

Tag air conditioning,demand-side management,marginal emissions factors,OAI-PMH Harvest,precooling,renewable energy

Format theses (aat)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Sanders, Kelly (committee chair), Olivier, Audrey (committee member), Silva, Sam (committee member)

Creator Email steppmay@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-oUC113800517

Unique identifier UC113800517

Identifier etd-MayesStepp-12583.pdf (filename)

Legacy Identifier etd-MayesStepp-12583

Document Type Dissertation

Format theses (aat)

Rights Mayes, Stepp A.

Internet Media Type application/pdf

Type texts

Source 20240105-usctheses-batch-1118 (batch), University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the author, as the original true and official version of the work, but does not grant the reader permission to use the work if the desired use is covered by copyright. It is the author, as rights holder, who must provide use permission if such use is covered by copyright.

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA

Repository Email cisadmin@lib.usc.edu