University of Southern California Dissertations and Theses

Bridging performance gaps by occupancy and weather data-driven energy prediction modeling using neural networks

(USC Thesis Other)

Bridging performance gaps by occupancy and weather data-driven energy prediction modeling using neural networks

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content

Bridging Performance Gaps
By Occupancy and Weather Data-Driven Energy Prediction Modeling using Neural
Networks

By
Aditya Sudhir Dharane

Presented to the
FACULTY OF THE
SCHOOL OF ARCHITECTURE
UNIVERSITY OF SOUTHERN CALIFORNIA
In partial fulfillment of the
Requirements of degree
MASTER OF BUILDING SCIENCE
MAY 2016
2

COMMITTEE
Prof. Joon-Ho Choi (Chair)
Assistant Professor
USC School of Architecture
joonhoch@usc.edu

Prof. Karen Kensek
Assistant Professor
USC School of Architecture
kensek@usc.edu

Prof. Kyle Konis
Associate Professor
USC School of Architecture
kkonis@usc.edu

3

ABSTRACT
Studies have shown that the prediction of energy usage by simulation program does not always
match the actual energy usage. Assumptions about occupancy and use of weather data different
from site conditions are few of the factors affecting the accuracy of prediction. Hence real time
occupancy, sub- metering and weather data for an office building in downtown Los Angeles was
measured. Using the measured data, two methods of prediction, calibration simulation and
artificial neural network, were studied to predict end use energy usage. Measured sub-metered and
occupancy data was used to generate lighting, equipment, and occupancy profile that were used
for calibrating the model. Results showed that when real time occupancy was used the prediction
were less accurate compared to when lighting and equipment profiles were used. This indicates
that occupancy was not modeled accurately in the simulation program. Different neural network
configuration and time series data with different time lag were studied to best predict end use
energy usage. The fan energy was best predicted with 60 minute time lag considered. Finally,
comparison of neural network with simulation program showed that neural network predictions
were more accurate.

HYPOTHESIS
Building operation energy use can be predicted more accurately by using the artificial neural
network and measured occupancy, weather and sub-metering data.

4

ACKNOWLEDGMENTS
The author is grateful to M/S BuroHappold Engineering for allowing the access to the office and
sharing the data and knowledge. Special thanks to Mr. Gideon Susman from M/S BuroHappold
for all his inputs and guidance throughout the past year.

5

TABLE OF CONTENTS
Contents
ABSTRACT .................................................................................................................................... 3
ACKNOWLEDGMENTS .............................................................................................................. 4
TABLE OF CONTENTS ................................................................................................................ 5
LIST OF FIGURES ...................................................................................................................... 10
LIST OF TABLES ........................................................................................................................ 14
LIST OF ABBREVIATIONS ....................................................................................................... 16
Chapter 1: Introduction ................................................................................................................. 17
Overview: .................................................................................................................................. 17
List of Terms ............................................................................................................................. 17
1.1 Importance of prediction of building energy in design and operational stage ............... 17
1.2 Problem with current prediction method – performance gaps ....................................... 18
1.3 Reasons for Performance Gaps ...................................................................................... 18
1.4 Significance of occupancy (behavior) in energy gap: .................................................... 20
1.5 Significance of weather in performance gaps ................................................................ 23
1.6 Goals and objectives....................................................................................................... 23
1.7 Scope of work................................................................................................................. 24
1.8 Hypothesis Statement ..................................................................................................... 24
1.9 Chapter Summary ........................................................................................................... 24
Chapter 2: Background and Literature Review ............................................................................ 26
Overview: .................................................................................................................................. 26
List of terms .............................................................................................................................. 26
2.1 State of the Art: .............................................................................................................. 27
6

2.2 Simulation Calibration ................................................................................................... 27
2.3 Use of Calibration Simulation: ....................................................................................... 28
2.4 Current mathematical regression models ....................................................................... 30
2.5 Potential use of Artificial Neural Network: ................................................................... 31
2.6 Occupancy Measurements.............................................................................................. 32
2.6.1 What parameters need to be measured? .................................................................. 32
2.6.2 Different methods of detecting occupancy ............................................................. 32
2.7 Summary: ....................................................................................................................... 35
Chapter 3: Methodology ............................................................................................................... 36
3.1 Overview of Methodology ............................................................................................. 36
3.2 Data Collection:.............................................................................................................. 37
3.2.1 Installing data collection tools ................................................................................ 39
3.2.2 Weather ................................................................................................................... 39
3.2.3 Indoor conditions .................................................................................................... 41
3.2.4 Sub-metering for energy consumption.................................................................... 42
3.2.5 Occupancy Sensors: ................................................................................................ 43
3.2.6 Data monitoring & Collection ................................................................................ 46
3.3 Methodology for simulation calibration ......................................................................... 46
3.3.1 Baseline Energy Model ........................................................................................... 47
3.3.2 Creation of Weather File ......................................................................................... 48
3.3.3 Sub-metering Data to Energy Model ...................................................................... 49
3.3.4 Occupancy Profile from actual measured occupancy ............................................. 50
3.4 Data Driven Modeling .................................................................................................... 50
3.4.1 Model building – artificial neural network: ............................................................ 51
7

3.4.2 Validation of model: ............................................................................................... 51
3.5 Summary ........................................................................................................................ 52
Chapter 4: Results - Occupancy Information and measurements data ......................................... 53
Overview: .................................................................................................................................. 53
4.1 Climate data.................................................................................................................... 53
4.1.1 Outdoor Temperature in degree Celsius: ................................................................ 53
4.1.2 Outdoor solar radiation (Watt/m
2
) .......................................................................... 57
4.1.3 Outdoor Wind direction: ......................................................................................... 57
4.2 Indoor Condition – Building Management System........................................................ 59
4.2.1 Central Chiller Chilled Water Temperature ............................................................ 59
4.2.2 Temperature of Air Handling Unit ......................................................................... 60
4.2.3 Air handling unit outside air volume (cms) and static pressure (Pa) ...................... 60
4.3 Indoor Conditions – Data Logger................................................................................... 61
4.3.1 Indoor spot measurement of temperature and relative humidity ............................ 61
4.3.2 Indoor spot measurements for lighting ................................................................... 62
4.3.3 Indoor spot measurement of carbon dioxide levels (ppm): .................................... 63
4.4 Sub-metered Data ........................................................................................................... 64
4.4.1 Pie-chart for distribution of total energy use in study period by end use: .............. 64
4.4.2 Sub-Metered Data – Fans........................................................................................ 65
4.4.3 Sub-metered Data – Water heating ......................................................................... 67
4.4.4 Sub-metered data- Lighting .................................................................................... 69
4.4.5 Sub-metering data – Kitchen energy use ................................................................ 70
4.4.6 Sub-metering – Refrigeration ................................................................................. 70
4.4.7 Sub-metering Data – Computer (W) ....................................................................... 72
8

4.4.8 Sub metering Total energy use: .............................................................................. 73
4.5 Occupancy ...................................................................................................................... 74
4.5.1 Occupancy data from the administration: ............................................................... 74
4.5.2 Occupancy data from sensor: .................................................................................. 75
4.6 Summary ........................................................................................................................ 76
Chapter 5: Analysis and Discussion ............................................................................................. 77
Overview: .................................................................................................................................. 77
5.1 Base case simulation results ........................................................................................... 77
5.1.1 Base case - measured versus simulated lighting energy ......................................... 80
5.1.2 Base case - measured versus simulated computer load .......................................... 81
5.1.3 Base case - measured versus simulated plug load .................................................. 82
5.1.4 Base case - measured versus simulated fan power ................................................. 84
5.2 Changing weather data ................................................................................................... 85
5.2.1 Changing weather file - measured versus simulated lighting, computer and plug
load 85
5.2.2 Changing weather file - measured versus simulated fan energy usage .................. 87
5.3 Lighting, computer and plug load profiles changed ....................................................... 88
5.3.1 Changing profile - measured versus simulated lighting, computer and plug load.. 88
5.3.2 Changing profiles - measured versus simulated fan energy usage ......................... 91
5.4 Occupancy profile change .............................................................................................. 92
5.4.1 Changing occupancy profiles - measured versus simulated lighting, computer and
plug load ................................................................................................................................ 92
5.4.2 Changing occupancy profiles - measured versus simulated fan energy usage ....... 94
5.5 Summary of simulation results:...................................................................................... 96
5.6 Artificial Neural Network Model ................................................................................... 97
9

5.6.1 Lighting energy usage prediction............................................................................ 99
5.6.2 Fan energy usage prediction: ................................................................................ 106
5.7 Comparisons of prediction by energy simulation model and artificial neural network
model 109
5.7.1 Comparison of different lighting energy prediction models ................................. 110
5.7.2 Comparison of different fan energy prediction models ........................................ 110
5.8 Summary ...................................................................................................................... 111
Chapter 6: Conclusion................................................................................................................. 112
Overview: ................................................................................................................................ 112
6.1 Lessons learnt ............................................................................................................... 112
6.2 Limitation of study: ...................................................................................................... 113
6.3 Future work .................................................................................................................. 114
6.4 Summary ...................................................................................................................... 114
Bibliography: .............................................................................................................................. 115

10

LIST OF FIGURES
Figure 1 Factor showing energy gap (“Maintenance 2016"; “People” 2016; PSDgraphics.com
2016; “Energyplus” 2016) ............................................................................................................ 20
Figure 2:
Comparison of measured and predicted energy use for 121 LEED New Construction buildings
(Cathy and Mark 2008) ................................................................................................................. 21
Figure 3: Measured air conditioning electricity consumption per unit floor area in summer of a
residential building in Beijing (Zhaojian Li, Yi Jiang, and QP Wei 2007) .................................. 22
Figure 4 Components consuming energy in typical office building that is affected by human
behavior (“Officescene” 2016; “Thermostat” 2016; “Houseandesign.top” 2016; “LG ” 2016;
“Espresso-Machine” 2016; “Rainshower” 2016) ......................................................................... 23
Figure 5 Showing the overview of the methodology (“AIA2012-2” 2016) ................................. 37
Figure 6 Open Office Layout chosen for as study space (“Buro-Happold-DT-LA-Office” 2016)
....................................................................................................................................................... 38
Figure 7 Office floor plan (Courtesy of M/s BuroHappold engineers) ........................................ 38
Figure 8 Typical layout of central conditioning system in a commercial building (“Typical
Layout of Central Conditioning System in a Commercial Building,” 2003.) .............................. 38
Figure 9 Location of sensors (1.5 m above floor level) installed in study space .......................... 39
Figure 10 Weather station (Integrated Sensor Suite) used to collect weather data (Davis
Instruments Vantage Pro2 Weather Station, 2004.) ..................................................................... 40
Figure 11 Screenshot of Building Management System............................................................... 41
Figure 12 HOBO Temperature and Humidity Data Logger by Onset Corp and Telaire Carbon
Dioxide Sensor .............................................................................................................................. 42
Figure 13 Wireless Circuit Level Clip on sensors for collecting sub-metered data ..................... 42
Figure 14 USB Based Infra-red People Counter (All-Tag People Counter with USB Capability,
Software Included, Wireless Battery Operated, n.d.) ................................................................... 44
Figure 15 Picture of installed occupancy sensor on site – marked in red ..................................... 45
11

Figure 16 Close up of the installed occupancy sensor on site ...................................................... 45
Figure 17 Showing methodology for calibration .......................................................................... 46
Figure 18 Process of creating weather file as per measured weather data .................................... 49
Figure 19 Methodology for Data Driven Modeling (Images source: “Icon Experience,” 2012) 50
Figure 20 Recorded Outdoor Temperature ................................................................................... 54
Figure 21 Day profile for outdoor temperature and relative humidity recorded for hottest day
(Monday October 12
th
) and average TMY2 values for October ................................................... 55
Figure 22 Day profile for outdoor temperature and relative humidity recorded for recorded
coldest day (Tuesday November 17
th
) and average TMY2 values for November ....................... 56
Figure 23 Outdoor Solar Radiation ............................................................................................... 57
Figure 24 Wind rose diagram from Climate consultant software ................................................. 58
Figure 25 Cumulative wind direction for entire study period ....................................................... 58
Figure 26 Chilled water inlet and outlet temprature plot .............................................................. 59
Figure 27 Air handling unit temperature ...................................................................................... 60
Figure 28 Outside air volume and static pressure ......................................................................... 61
Figure 29 Indoor spot measurement ............................................................................................. 62
Figure 30 Measured light intensity ............................................................................................... 63
Figure 31 Measured Carbo dioxide level ...................................................................................... 64
Figure 32 Energy consumption ..................................................................................................... 65
Figure 33 Measured fan energy usage .......................................................................................... 66
Figure 34 Average and Peak Daily fan energy use ....................................................................... 67
Figure 35 Water heating energy usage.......................................................................................... 68
Figure 36 Average water heating energy use ................................................................................ 68
Figure 37 Sub-metering data – lighting ........................................................................................ 69
Figure 38 Sub-metering data – Kitchen ........................................................................................ 70
12

Figure 39 Average daily refrigeration energy use ........................................................................ 71
Figure 40 sub-metering data- computer ........................................................................................ 72
Figure 41 Average daily computer energy use ............................................................................. 73
Figure 42 Sub metering total energy use ...................................................................................... 74
Figure 43 Measured occupancy .................................................................................................... 75
Figure 44 Comparison of ASHRAE 90.1 profile versus actual profile derived from measured
occupancy ..................................................................................................................................... 76
Figure 45 Base case model in IES VE .......................................................................................... 78
Figure 46 HVAC model in APACHE HVAC module in IES VE ................................................ 78
Figure 47 Symbols used in Apache HVAC .................................................................................. 79
Figure 48 Measure vs Simulated lighting power for week 1 ........................................................ 80
Figure 49 Measure vs simulated computer for week 1 ................................................................. 81
Figure 50 Measure vs Simulated plug load for week 1 ................................................................ 82
Figure 51 Measured versus simulated fan power for week 1 ....................................................... 84
Figure 52 Lighting energy usage .................................................................................................. 85
Figure 53 Computer energy usage ................................................................................................ 86
Figure 54 Plug load energy usage ................................................................................................. 86
Figure 55 Fan energy usage after changing weather file .............................................................. 87
Figure 56 Lighting energy after changing equipment profile ....................................................... 89
Figure 57 Plug load after changes in profile ................................................................................. 89
Figure 58 Computer energy after changes in profile .................................................................... 90
Figure 59 Fan energy usage after changes in profile .................................................................... 91
Figure 60 Lighting energy after changing occupancy profiles ..................................................... 92
Figure 61 Computer energy usage after changing occupancy profile .......................................... 93
13

Figure 62 Plug load energy usage after changing occupancy profile ........................................... 93
Figure 63 Fan energy usage after change in occupancy data ........................................................ 95
Figure 64 Typical Process in Rapid Miner software .................................................................... 97
Figure 65 Typical validation process ............................................................................................ 98
Figure 66 Neural network with one and two hidden layer with six neurons ................................ 99
Figure 67 Neural network with two hidden layer with three neurons and six three neurons ..... 100
Figure 68 Lighting energy usage preidction by neural network ................................................. 101
Figure 69 Lighting energy usage ................................................................................................ 103
Figure 70 Lighting energy usage comparison ............................................................................ 104
Figure 71 Lighting energy prediction with time lag ................................................................... 105
Figure 72 Fan energy prediction with neural network ................................................................ 107
Figure 73 Fan energy preidction with time lag ........................................................................... 108
Figure 74 Comparison of simulation and neural network for lighting ....................................... 110
Figure 75 Comparison of simulation and neural network fan energy ........................................ 111

14

LIST OF TABLES
Table 1: Advantage and disadvantage of calibration simulation .................................................. 29
Table 2 Advantages and disadvantages of regression model: ...................................................... 29
Table 3 Comparison of different various occupancy detection methods systems ........................ 33
Table 4 Description of parameters measured for climate conditions: (“Wireless Vantage Pro2
TM

& Vantage Pro2
TM
Plus Stations,” 2003.) ..................................................................................... 40
Table 5 Indoor Parameters and Tools Used .................................................................................. 41
Table 6 Location of Wireless Clip-Ons and what they represent ................................................. 43
Table 7 Different cases studies ..................................................................................................... 47
Table 8: Material Properties as defined in basis of design ........................................................... 47
Table 9 Number of people employed in office as per HR manager ............................................. 75
Table 10 Baseline lighting MBE and CVRSME values ............................................................... 81
Table 11 Baseline computer MBE and CVRSME values ............................................................ 82
Table 12 MBE and CVRSME values for base case plug load ...................................................... 83
Table 13 MBE and CVRSME values for base case fan energy .................................................... 84
Table 14 MBE and CVRSME after changing weather file........................................................... 87
Table 15 MBE and CVRSME values for fan energy usage after changing weather file.............. 88
Table 16 MBE and CVRSME values after changing profile ........................................................ 90
Table 17 MBE and CVRSME values for fan energy after changing lighting, computer and plug
load profiles .................................................................................................................................. 91
Table 18 MBE and CVRSME value after changing occupancy profile ....................................... 94
Table 19 MBE and CVRSME values for fan energy after changing profiles .............................. 96
Table 20 MBE and CVRSME values for different simulation cases ............................................ 96
Table 21 MBE and CVRSME values for lighting energy prediction ......................................... 101
Table 22 MBE and CVRSME values ......................................................................................... 103
15

Table 23 MBE and CVRSME values for lighting prediction with time lag ............................... 106
Table 24 MBE and CVRSME values – Fan prediction .............................................................. 107
Table 25 MBE and CVRSME values for fan energy prediction with timelag ........................... 108
Table 26 MBE and CVRSME values of lighting energy prediction .......................................... 110
Table 27 MBE and CVRSME values for fan energy prediction ................................................ 111

16

LIST OF ABBREVIATIONS
ASHRAE American Society of Heating, Refrigerating and Air Conditioning Engineers
BAS Building Automation System
BMS Building Management System
BTU British thermal unit
CDD Cooling Degree Days
DOE Department of Energy
EPA Environmental Protection Agency
EUI Energy Use Intensity
GJ Gigajoule
HDD Heating Degree Days
HVAC Heating, Ventilation and Air-Conditioning
IEQ Indoor Environmental Quality
MLR Multiple Linear Regression
POE Post Occupancy Evaluation
Quad Quadrillion Btu (10^15 Btu)
THSW Temperature/Humidity/Sun/Wind
TMY2 Typical Meteorological Year 2 (TMY2)
UV Ultra Violet
WWR Window Wall Ratio
ANN Artificial neural network
IES Integrated environment solution
VE Virtual environment

17

Chapter 1: Introduction
Overview:
In this chapter, the importance of predicting of building energy, problems associated with it are
discussed. The significance of weather data and occupancy behavior on the prediction of building
energy is discussed. At the end of the chapter, the goal and objective and hypothesis statement are
stated.
List of Terms
Sensitivity analysis:
It is a technique used to determine how different values of an independent variable will
impact a particular dependent variable under a given set of assumptions (root 2003). In
building energy simulation, sensitivity analysis describes the energy impact of a range of
building physical features and operational practices, representing the energy use
characteristics of the building (Jonathan Heller and Morgan Heater 2011).
1.1 Importance of prediction of building energy in design and operational stage
40% of the global energy is consumed by buildings (Initiative 2009). Thus, there is great potential
to conserve energy if building energy is optimized. For energy conservation, building simulation
applications have emerged as a cost-effective method for supporting energy efficient design and
operation of buildings. Energy simulation programs are mostly used for code and energy standard
compliance (Zhao 2015). The building simulation programs are also used to evaluate the
architectural concepts, compare different HVAC systems, or different energy conservation
approaches and technologies. In practice, decisions are made based on these simulation results. As
per ASHRAE Research Report 1051- procedure for reconciling computer-calculation results with
measured energy data, published in 2006, 220 billion dollars of energy is consumed by buildings,
and by 2016, 90% of building stock will comprise of existing buildings (Reddy and Maor 2006).
Thus, energy conservation in existing building is important, if not more important, than new
construction. Studies have shown that predicting building operational energy consumption can
save energy in existing buildings (Wang, Mathew, and Pang 2012). Energy saving is possible as
the building management team can optimize daily operational and implement better control
strategies with an accurate prediction of building energy consumption. With better controls and
optimization demand side management (DSM) is more effective and adds value to client / property
18

owner. Further, if one can predict building energy accurately in operational stage, then sensitivity
analysis will be more reliable. Sensitivity analysis has proven to be very helpful in effective
retrofitting for a building to save energy (Gustafsson 1998).
1.2 Problem with current prediction method – performance gaps
Many decisions are made based on the results of the simulation program. Thus, the accuracy of
these building simulation programs is critical. However, in some cases, it is seen that the predicted
building energy use by simulation software does not always match the measured energy once the
building is in operation. In some instances, the operational energy is seen to be 1.5 to 2 times the
predicted energy (Pieter 2009). The difference between measured and predicted energy is called
performance gap or energy gap. There are various reports, like Carbon Hub, LEED report on new
buildings 2008, Carbon Trust, and scientific papers, that support the evidence that energy gap
exists in new and existing building (“Performance Gap Evidence Review Report Released” 2016),
(Cathy and Mark 2008), (“Annual-Report-2014-2015.pdf,” 2015), (Newsham, Mancini, and Birt
2009; Cathy and Mark 2008; Pieter de Wilde 2009; Blight and Coley, n.d.; Sunikka-Blank and
Galvin 2012).

1.3 Reasons for Performance Gaps
Various factors lead to energy gap (Figure 1). They span throughout building life cycle starting
from conceptual stage and lasting till building’s operation and maintenance stage.
Miscommunication between client and design team or between different actors of the team
regarding the performance of the building is the cause of performance issue (Newsham, Mancini,
and Birt 2009). There might be an initial issue in the design, incorporate inefficient systems, wrong
or missing construction details, or lack simplicity or buildability (Pieter de Wilde 2009). During
the design stage, many energy efficient technologies are utilized which do not perform as per
intended. The prediction of building energy use at design stage are based on assumptions,
modeling techniques, software tools made by the simulation engineers. Obviously, this requires
that the simulation engineers should be well trained and have adequate experience so that they use
appropriate tools, assumption, and modeling techniques. This is not true in most instances, and
19

human error is introduced which lead to energy gap. Testing validation and verification in the field
of building energy modeling are emerging areas that still need further development.
The construction process introduces energy gaps (Pieter de Wilde 2009). The required insulation
and air tightness levels are challenging. Construction is typically layered and hence the errors and
defects in construction might not be visible to naked eye. Value engineering and change orders are
also seen to the cause for energy gaps. While value engineering will reduce the construction cost
but in doing so, the performance parameter might be overlooked, introducing energy gap (Blank
and Galvin 2012). Due to budget and time constraints, full performance evaluation and testing may
not be done during building commissioning and hand-over.
Once the building is in operation, the weather data used for performance estimation of a building
is most cases have seen not to reveal the actual weather conditions on site (Wang, Mathew, and
Pang 2012). Control settings of thermostats might not be represented or programmed as intended
within the Building Energy Management System (BEMS) (Guo, et al. 2010). Furthermore, one
has to accept that, metering itself comes with issues and uncertainties; this is especially true when
it comes to capturing contextual factors such as weather data and occupant behavior. Measurement
techniques can by itself have accuracy, missing, or incomplete data, as well as incredible cleaning
of metering data, is therefore essential, but can introduce further threats to the validity of the
results (Pieter de Wilde 2009).
Daylighting strategies are used to improve the performance of the building. With the advance in
technologies and sophistication of simulation software, these are simulated in building energy
performance software. These simulated strategies often in reality not seen to perform as intended
because of glare, values engineering, and constructability problem. Energy gaps are introduced
because of this. The figure below summarizes the factors that introduce performance or energy
gap throughout the building life cycle.
20

Figure 1 Factor showing energy gap (“Maintenance 2016"; “People” 2016; PSDgraphics.com
2016; “Energyplus” 2016)
1.4 Significance of occupancy (behavior) in energy gap:
International Energy Agency (IEA), Energy in the Buildings and Communities program (EBC),
Annex 53: identified climate,
building envelope, building energy and services systems, indoor design criteria, building
operation nad maintenance, and occupant behaviour, as the driving factor in energy use
in buildings (“EBC Annex 53 Total Energy Use in Buildings: Analysis & Evaluation Methods”
2016).
Significant progress has been made in the first five focus areas but due to the lack of scientific and
robust methods to model occupant behavior in building simulation models this area is still lagging
behind others (Yan et al. 2015). Occupant behavior is now widely recognized as an important
contributing factor to the uncertainty of building performance (Yan et al. 2015).
Further, at the 74
th
Executive Committee Meeting of the IEA Energy in Building and Communities
Program, held in Dublin, Ireland, Annex 66 project was approved unanimously. The aim of the
annex is to set up a standard occupant behavior definition platform, establish a quantitative
21

simulation technology to model occupant behavior in building, and understand the influence of
occupant behavior on building energy use and the indoor environment (“World: [Energy] Balances
for 2012,”2013).
The behaviors include occupants’ interaction with operable windows, light, blinds, thermostats,
and plug load in appliances. The importance of “human factor” is evident (Zhang et al. 2014; Attia
et al. 2013) in building a simulation. In future, as the building becomes more energy efficient the
role of occupancy behavior place an important role.
Thus, occupant behavior is a one of the most important factor affecting the predictions of building
simulation software.

Figure 2:
Comparison of measured and predicted energy use for 121 LEED New Construction buildings
(Cathy and Mark 2008)
A study conducted in a large residential building in Beijing, which consisted of 25 apartments
illustrates the impact of occupancy behavior (Li et al., 2007). In the study, electricity consumption
of the 25 apartments was measured and studied. Even though the apartments were identical, the
consumption varied largely, as per resident’s use of the split air-conditioning unit (Figure 3).

22

Figure 3: Measured air conditioning electricity consumption per unit floor area in summer of a
residential building in Beijing (Zhaojian Li, Yi Jiang, and QP Wei 2007)
An apartment where the occupant kept the air conditioning on for longer durations or in larger sp
aces consumed more energy than an apartment using the AC for a shorter period and/or in smaller
space.
Thus, the occupant is the driver of the energy consumption, rather than the design of the apartme
nts. Uncertainty in occupant behavior was studied in building energy models, using various
occupancy schedules and environmental preferences and found that the energy consumption
differed150% (or more) if the occupant-related inputs were maximized and minimized (even
for typical occupant behavior patterns). Studies show that the wat occupants used the heating
system makes a significant impact on the heating energy consumption. Thus, it will be wise to say
that even if the weather conditions, the building envelope, and the equipment are well‐defined,
occupants’ presence and interaction with various building components significantly affect the
energy consumption predictions made by energy simulation (Yan et al. 2015).

23

Figure 4 Components consuming energy in typical office building that is affected by human
behavior (“Officescene” 2016; “Thermostat” 2016; “Houseandesign.top” 2016; “LG ” 2016;
“Espresso-Machine” 2016; “Rainshower” 2016)
1.5 Significance of weather in performance gaps
To predict the annual energy usage by using simulation programs, a weather file is required. These
weather files are typical for a city, and it often is seen that the site weather data does not match the
site conditions. Energy gaps are introduced because of it. Further, a variety of weather files is
available when an energy simulation engineer does simulations. Some of them are locally
recorded, measured weather data to preselected “typical” year. Typical weather file includes –
WYEC2, TMY2, CWEC, and CTZ2. (Huang and Crawley 1996a), compared the variations caused
by the use of different weather files. It was seen that the use of different files introduced a variation
of 11% in total annual energy usage while the variation in annual peak electric demand was
reported 10%.
1.6 Goals and objectives
Discussion till now was about importance of predicting building energy and how the accuracy of
this prediction is still a significant problem. The goal is to discover ways/ technique to improve
the accuracy of prediction of building energy modeling techniques for existing buildings. Though
predicting building energy at design and operation is essential due to available resources and time,
• Computers
• Lights
• Equipments
• Coffee
Machine
• Water Heater
for Shower
• Heating and
Cooling
24

the focus was to predict building energy usage in operational stage. Also, the discussion still now
advocate that amongst the various factors affecting the prediction of building energy, human
behavior (occupancy) and weather are major contributors for inaccuracy in predicting building
energy. Thus, following are the objectives of this research:
 To investigate the impact of measured real-time weather data on simulation accuracy in
energy modeling program.
 To analyze the significance of actual occupancy data in building operational energy by
comparing prediction results with measured energy consumption.
1.7 Scope of work
An office building in downtown Los Angeles is selected as a case study to improve the predictions
of end-use energy usage at the operational stage and analyze the impact of a human factor in energy
consumption. To improve the accuracy, a time series mathematical model with the artificial neural
network was developed. For the model, measured occupancy, sub-metering, and weather data were
used. Further, a calibration of an energy model for the office building was performed. For this
same measured occupancy, sub-metering, and weather information was used. Then results of
calibrated model were compared with the new proposed mathematical model. For this type of
study multiple offices in multiple locations and over a longer duration would be considered ideal
to arrive at a reasonable conclusion. Due to the limitation of time and available resources, only
one office for six week period was studied. However, high-resolution data at 15-minute time
interval was investigated to arrive at a reasonable conclusion.
1.8 Hypothesis Statement
Building operation energy use for an existing building can be predicted more accurately by using
the artificial neural network and measured occupancy, weather, and sub-metering data.
1.9 Chapter Summary
Emphasis of predicting building energy was explained and how it can be useful to saving energy.
Factors affecting building energy prediction and role of the human factor and weather on building
energy prediction was discussed in detail. The goals, objective, and scope of work were defined.
25

The next chapter focuses on the literature review. In Chapter 3 methodology for data collection,
energy modeling and mathematical modeling is discussed. In chapter 4 collected data is presented.
In chapter 5 the results are documented and in chapter 6 conclusion, limitation, and future work
are conferred
26

Chapter 2: Background and Literature Review
Overview:
In the previous chapter, practical application of prediction of building energy use and the factor
affecting the accuracy of predictions were discussed. The dialogue also suggests that human factor
is one of the major factors leading to inaccuracies in predicting energy. In this chapter, the literature
related to the current building energy predictions methods (state of the art), their application are
discussed, various method of predictions are compared. In addition, various methods of occupancy
measurement are discussed.
List of terms
1. Autocorrelation:
A mathematical representation of the degree of similarity between a given time series and
a lagged of itself over a successive time interval is called auto-correlation (root 2006).
2. Multi-collinearity:
When in a regression model two or more predictors are highly or moderately correlated,
the accuracy of the model decrease and this problem is called multicollinearity (“Lesson
12: Multi-collinearity and Other Regression Pitfalls | STAT 501” 2016).
3. Predictive mean vote (PMV):
The average thermal sensation response of a large number of subjects, using ASHRAE
thermal sensation scale is called the predictive mean vote (Brandem 2016) . PMV is the
most widely used thermal comfort index used (“Human Thermal Comfort | Sustainability
Workshop” 2016).
4. Time series
Sequence of measurement of the same variable collected over time is called time series
(“1.1 Overview of Time Series Characteristics | STAT 510” 2016).
5. Autoregressive model (AR):
An autoregressive model is when a value from a time series is regressed on previous
values from that same time series (“14.1 - Autoregressive Models | STAT 501” 2016).
6. Autoregressive integrated moving average (ARIMA)
For time series mean values calculated over a shorter duration of time which are moving
are called moving average. A range of models can be constructed suing the moving
27

average and they are called MA models. When MA models are combined with AR
models, they are called ARMA or ARIMA (I stands for integrated) (“Statistical Analysis
Handbook” 2016).
7. Autoregressive moving average with exogenous input model (ARMAX)
When an independent variable is used in the ARMA model, the resulting model is
ARMAX model (Rob J Hyndman 2010).
8. Fourier series:
A Fourier series is a sum of sine and cosine function that describes a periodic signal
(“Fourier Series - MATLAB & Simulink” 2016).
9. Un-stationary changes
When mean and variance is not constant over time in a time series data, the series is said
to have un-stationary changes (“Regression - Why Does a Time Series Have to Be
Stationary? - Cross Validated” 2016)
2.1 State of the Art:
Predicting of energy in operational stage of building is often referred to inverse modeling. Inverse
modeling uses measured data to predict future energy consumption in the building. Inverse
modeling can be further divided into two categories based on the method used. There are two
common methods: calibration simulation and mathematical regression model.
2.2 Simulation Calibration
Calibration simulation is the process of using an existing building simulation computer program
and “tuning” or calibrating the various inputs to the program so that predictions match closely with
observed energy consumption (Reddy and Maor 2006). By using statistical methods, more reliable
and insightful prediction can be made once this is done (Tso and Yau 2007). The first attempt was
made in 1980 to calibrate the energy model by using utility bills (Oh 2013). The calibration process
involves detail modeling and the use of many input parameters, which restricts the accuracy and
reliability of the calibration process (Royapoor and Roskilly 2015). After this researcher started
using short-term end, use data to improve the accuracy. Finally, researchers started using
instantaneous data on an hourly basis for calibrating the model. The researcher concluded that that
use of hourly data for a month, season or year can effectively represent the building system
dynamics (Gunay, O’Brien, and Beausoleil-Morrison 2013).
28

2.3 Use of Calibration Simulation:
Calibrated simulation can be used for the following purposes (Reddy and Maor 2006):
 To improve the models used in simulation program (Clarke, Strachan, and Pernot 1993);
 To educate the owner about thermal and electrical loads usage in their building (Soebarto
and Williamson 2001);
 To predict impact of different energy conservation and load control measures on the
aggregated electrical load (“DrCEUS: Energy and Demand Usage from Commercial On-
Site Survey Data (Abstract Only) | Energy Efficiency Program Library” 2016);
 To support investment-grade recommendations made by an energy auditor who has to
identify cost-effective ECMs unique to the individual building and determine their payback
(Reddy and Maor 2006);
 For monitoring and verification (M&V) under the following circumstances:
o if ECM are implemented calibration simulation can be used to find baseline
energy use against which saving can be measured,
o if unanticipated changes like due to creep in load, changes in operational
hours or changes in occupancy conditions occur the correction in the
contractual baseline energy can be made,
o if M&V requires whole building monitoring data to be used for verifying
the effect of end-use retrofit,
o if the effect of individual retrofit needs to be isolated when retrofits are
complex and interactive,
o if no data before or after retrofit is available,
o if the time required for post-retrofit verification needs to be reduced (Reddy
and Mazor 2006).
 The facility/building management service can provide the owner the capability of
implementing:
o continuous commissioning or fault detection (FD) measures by identifying
equipment malfunction and take appropriate action like as
tuning/optimizing HVAC and primary equipment controls,
29

o Optimal supervisory control, equipment scheduling and operation of the
building and its systems, either under normal operation or active load
control in response to real-time pricing (RTP) price signals (Reddy and
Maor 2006).
Advantages and disadvantages of calibration simulation are tabulated below (Table 1):
Table 1: Advantage and disadvantage of calibration simulation
Advantage Disadvantage
Calibration
Simulation
1. Knowledge of data science/data
mining techniques is not required
(Clarke, Strachan, and Pernot
1993).
2. Prediction accuracy of 2% is
demonstrated in few journal and
conference papers(Lam et al. 2014).
3. No need to develop software or
algorithm. Existing simulation
software’s can be used (Clarke,
Strachan, and Pernot 1993).
4. Can represent hourly or sub-hourly
demand (Tso and Yau 2007)
5. It is not data driven but based on
fundamental principles of
thermodynamics and building
physics (Oh 2013).
1. No specific method/ standard is
available(Royapoor and Roskilly
2015).
2. The accuracy of results depends on
energy simulator’s experience and
skills (Reddy and Maor 2006).
3. Building physics cannot always
represent reality (Pieter de Wilde
2009).

Advantages and disadvantages of regression model are tabulated below:
Table 2 Advantages and disadvantages of regression model:
Advantage Disadvantage
Regression
Models
1. Accuracy is not dependent on the
individual; it depends on the sound
use of data mining/ data science
techniques (Tso and Yau 2007).
2. As per ASHRAE Guideline 14,
more accurate prediction are
possible using these models
(Heating et al. 2004).
3. Suited for average energy
consumption for a day or year
(Neto and Fiorelli 2008)
1. The exact relationship between the
parameter cannot be defined
(Chang and Hong 2013).
2. Based on the data, the relationship
can be estimated, but that is never
going to be exact (Kalogirou and
Bojic 2000).
3. Hourly or sub-hourly energy
demand is not reflected (Tso and
Yau 2007).

30

4. Time-consuming method, as one
need to find the best-fit model
(Ekici and Aksoy 2009).

2.4 Current mathematical regression models
A mathematical regression model, in the statistic, is a modeling approach to predict the
relationship between scalar dependent variable y and one or more explanatory variables
(“Introduction to Linear Regression Analysis” 2016). A mathematical regression model is used
in trend estimation of stock, oil price, etc. in the medical field to prove that tobacco smoking is
injurious to health, in finance field to analyze and quantify the systematic risk of investment
(“Linear Regression” 2016). Building energy prediction with the help of regression model has
been demonstrated effectively in many journal papers (Fels 1986; Dhar, Reddy, and Claridge
1999; Ruch and Claridge 1992). However, regression models are unable to reflect hourly or sub-
hourly energy demand. They are best suited to predict average energy consumption for longer
duration like day or month (Yang, Rivard, and Zmeureanu 2005). Selecting appropriate time
scale and regresses to fit best the model, for different building with different environment and
weather conditions takes hard work. Auto-correlation or multi-collinearity problems lead to
model uncertainty and must be considered carefully while evaluating the performance of
prediction.
Energy use is a function of time as well, so it makes more sense to use time series analysis
techniques for forecasting energy. Experiment with the autoregressive integrated moving average
(ARIMA) model and found the performance of ARIMA to be better than a two-dimensional
autoregressive (AR) model. (Yang, Rivard, and Zmeureanu 2005)
Autoregressive moving average with exogenous input model (ARMAX) were used to derive many
models and applications. For a given set of time series data, ARIMA and ARMAX model can
capture the relationship between the hourly energy consumption and time variation. ARMA and
AR assume that the present value is a linear combination of previous ones, which is incorrect.
ARMIA and ARMAX model handle the un-stationary process changes, but they require many
other parameter estimations. The accuracy of the prediction is strongly affected by the correlation
between different variables. To estimate the energy demand in an institutional building Fourier
31

series model was tested and it was found that even though Fourier series model performs better
than the time series model, they assume that energy use in the building is periodic which is not
always true (Dhar, Reddy, and Claridge 1998). Fourier series models are unable to cope up with
dramatic changes (Dhar, Reddy, and Claridge 1999). To take care of the dramatic changes, high-
frequency Fourier components must be included, which increases the computational time
significantly.
2.5 Potential use of Artificial Neural Network:
Artificial neural network is families of models, mathematical usually, which mimics the behavior
of human brain by using artificial intelligence technique. For a complicated system, it can
approximate a nonlinear relationship between input variable to generate an output or outputs. Self-
learning capabilities are one of the major advantages of ANN. ANN are faster at estimating
parameters because they learn from examples automatically. Generally, for other models noise in
the data makes it hard to distinguish structure but ANN memorizes noise. Many researchers before
have made use of ANN to predict the building energy consumption (Yang, Rivard, and Zmeureanu
2005). Though all of them share some similarities, they differ from each other as each one predicts
a particular type of energy for specific built environment.
Some of them tell us about static predictions while others talk about the use of dynamic predictions
(Yang, Rivard, and Zmeureanu 2005). In static prediction, historical data is used to set up a model
in advance which cannot be changed later. Because of this, the model can become invalid with
new patterns or trends in recent data. Extensive historical data is required to set up this kind of
models. On the other hand, dynamic model can change itself to such pattern changes. Classic
mathematical theories are used to develop the regression and time-series model. Hence, the
behavior and parameters of these models are well understood. However, these models do not
perform well under non-linear conditions. Thus, for predicting building energy system with these
models is not accurate as the behavior of building energy system is mostly non-linear. On the other
hand, ANN performs better on a non-linear system. However, the performance of ANN depends
on the choice of input and output parameters, the structure of ANN, the number of hidden layer,
the number of neurons used in each layer, and the training algorithm. There are few studies which
compare the performance of different models. Comparison of ANN with Fourier series model
(Dhar, Reddy, and Claridge 1999). While using ANN, one need not define the relationship between
32

input and output parameter as the relationship is identified through self-learning process. This
might be perhaps the biggest advantage of using ANN over other mathematical models. Because
of this, use of ANN, while establishing a proper mathematical model, results in significant time
and effort reduction as compared to other conventional prediction methodology. The ANN-based
models performed better compared to other models in the two Great Energy Shootout competitions
organized by ASHRAE. In the literature review, separate cases of use of Fourier series or time
series and ANN were encountered. However, none of them has focused on time series or Fourier
series coupled with ANN. This maybe because of the limited or available technology/ tools and
computational capabilities in past. However, now with the recent development and availability of
tools / software in computer science domain, this is possible.
2.6 Occupancy Measurements
There are several parameters that effect occupancy measurements.
2.6.1 What parameters need to be measured?
To have a better understanding of the occupancy, six spatial- temporal properties to describe
occupancy information in a typical office building were defined in (Labeodan et al. 2015).
 Presence – Is the person present? Simulation programs use diversity factor to define
occupancy presence.
 Location – Which zone is the person in? As office building are divided into a different
zones, this information is necessary.
 Count- How many number of people are present?
 Activity – What is the person doing? The activity level will directly affect the metabolic
rate and CO2 production, and is used to determine the mean predictive vote (PMV) value.
This is commonly used the model to determine how much people are comfortable.
 Identity – Who is the person?
 Track: Where the person is coming from? This data can be difficult to obtain.
2.6.2 Different methods of detecting occupancy
There are various occupancy detection methods, techniques, and technology available. Various
studies show the application of these different technologies and techniques. Detecting occupancy
33

using CO2 sensors can help in demand control ventilation strategies to save energy (Nassif 2012).
Passive infrared detection system (PIR) sensors can be effective in lighting control (Guo et
al.2010). Ultrasonic detection system can provide occupancy information about presence and
location and to control lighting (Haq et al. 2014). Use of image detection system can control
environmental systems (Benezeth et al. 2011). The sound detection system can determine the
number of people of in space (Uziel, S, et al., 2009). Other authors have used successfully
demonstrated the use of electromagnetic signal (EM), energy management based, computer
activity, a sensor fusion detection system to collect occupancy information (Li et al. 2012;
Milenkovic and Amft 2013; Khoury and Kamat 2009; Christensen et al. 2014; Dong et al. 2010).
The various occupancy detection techniques can be grouped based on method, function, and
infrastructure. The method of occupancy measurement can be categorized into two types– terminal
and non-terminal. When occupancy is detected by use of devices like mobile or radio frequency
tag which are carried by the occupants, then that technique of detection is called terminal method.
On the other hand, if sensors such as infra-red (PIR), carbon dioxide sensors are used, which do
not require occupancy to carry, then that technique can be called non-terminal. Based on the
function, occupancy detection techniques can be individualized or non-individualized. Occupancy
detection techniques that have the capabilities to detect, identify and track individual occupants
are termed as an individualized system. While the detection techniques that do not that capability
to detect, identify or track individual occupant are termed as a non-individualized system. The
occupancy measurement techniques based on infrastructure can be grouped in two ways- implicit
and explicit. The occupancy measurement techniques that give direct information about occupancy
are termed as implicit. The examples of these are CO 2 and PIR sensors. While the measurement
techniques that use information to infer occupancy measurements are termed as explicit. The
examples of these are occupancy measurement by information about the use of a pattern of
building appliances such as computers, printers, and other similar appliance.
Various occupancy detection techniques and their advantage and disadvantage based on method,
function, infrastructure, location, presence, count, activity, identity, and track are summarized
(Table 3) (Labeodan et al. 2015):
Table 3 Comparison of different various occupancy detection methods systems
34

Sensors Method Function
Infra-
structure

Termi
nal
Non-
terminal
Individual
ized
Non-
Individualized Implicit
Expli
cit
CO2 sensor × √ × √ × √
PIR sensors × √ × √ × √
Ultrasonic
sensors × √ × √ × √
Image
sensors × √ × √ √ √
Sound
Sensors × √ × √ × √
EM Signal √ × × √ √ √
Power
Meters × √ × √ × √
Computer
App. × √ × √ √ ×
Sensor
fusion √ × √ √ √ √

35

Locatio
n Presence Count Activity Identity Track
CO2 sensor √ √ √ √ × ×
PIR sensors √ √ × × × ×
Ultrasonic
sensors √ √ × × × ×
Image
sensors √ √ √ √ √ √
Sound
Sensors √ √ × × × ×
EM Signal √ √ √ √ √ √
Power
Meters √ √ × √ × ×
Computer
App. √ √ √ √ √ ×
Sensor
fusion √ √ √ √ √ √

It is clear that sensor fusion detection system, a system developed by using multiple sensors, can
give a comprehensive information and better insight regarding the occupancy data.
2.7 Summary:
In this chapter, a review of different inverse energy models and their practical applications was
made. From the literature review, it was found that time series regression models with artificial
neural network could prove to give more accurate results in all of the regression models. Hence,
an artificial neural network model was selected to be tested to predict end-use energy usage, and
it is compared with calibrated energy model. Spatial-temporal properties that are needed to be
captured to get a comprehensive insight of occupancy were discussed and comparison of different
occupancy detection techniques was made. In the next chapter methodology for data collection,
calibration and development of ANN model are discussed in detail.

36

Chapter 3: Methodology
Overview
The intention was to find the impact of occupancy, weather and sub-metering on building energy
predictions when the building is in operation by using real-time occupancy, weather, and sub-
metering data. In this chapter methodology to achieve this is discussed.
3.1 Overview of Methodology
For an office building located in downtown Los Angeles, actual occupancy, weather and sub-
metering data was measured. Sensors are placed for sub-metering energy use at the circuit level,
weather data was collected from a weather station located on the site, and indoor conditions are
collected from the building BMS system and data loggers, occupants, are counted using a bi-
directional occupancy. The measured data was used to perform calibration of energy model. To
find the impact of occupancy, weather and sub-metering data, the measured data was inputted in
the energy model individually and all together. Further, a time series regression model with
artificial neural network (ANN) with different neural structure are tested and compared. The
generic process consists of comparing two results (
Figure 5). Detailed methodology for calibration and ANN process in a chapter later discussed
later.

37

Figure 5 Showing the overview of the methodology (“AIA2012-2” 2016)
3.2 Data Collection:
An office space is downtown Los Angeles; California was chosen for the study mainly because
access was granted and has a typical office setting in a city (
Figure 6 Open Office Layout chosen for as study space). The office is leased by an engineering
consulting firm and people working in the office are mainly engineers with few office
administrators. The office has a total of 12,336 square-foot, and it is located on the 16
th
floor at the
corner of Wilshire Boulevard and Flower Street. The commercial building was built in 1971. All
floors are occupied as office space. The building has a window to wall ratio of 65%, with single
pane clear glazing. The building has no external shading devices, and indoor roll-down opaque
shading is provided to the east and south sides. The last major renovation was done in 2007 with
a new chiller being installed. Chiller runs on fixed schedules for working hours; 6 am to 6 pm on
weekdays and 9 am to 1 pm on Saturday. On each floor, a central recirculating air handling unit
with filter, cooling coil and fan section. The office being studied has an underfloor air distribution
system for the open space and variable air volume with terminal reheat for breakout area and
meeting rooms. The AHU supplies cool and dehumidified air to the underfloor air distribution
plenum. Since the studied space used here has an open office layout, occupied single rooms are
not covered in the study.

38

Figure 6 Open Office Layout chosen for as study space (“Buro-Happold-DT-LA-Office” 2016)

Figure 7 Office floor plan (Courtesy of M/s BuroHappold engineers)
The office building has a central chilled water system. The typical central chilled water system
layout in shown in figure below.

Figure 8 Typical layout of central conditioning system in a commercial building (“Typical
Layout of Central Conditioning System in a Commercial Building,” 2003.)
In the chilled water system, a centralized cooling is located in the single room, in this case, the
basement plant room. This plant supplies chilled water to an air-handling unit situated on each
floor.
39

3.2.1 Installing data collection tools
Sensors were placed for collecting outdoor weather parameters using a weather station; indoor
parameters are collected from the building management system, and data loggers, sub-metering
measuring energy consumption at the circuit level, and occupants are counted using a bi-
directional occupancy counter (Figure 9). All sensor are place at working height that was
approximately 1.5 m from the floor level. Sensors function and capabilities are explained below.

Figure 9 Location of sensors (1.5 m above floor level) installed in study space
3.2.2 Weather
For measuring the weather parameters wireless vantage pro2 plus including UV & Solar Radiation
sensors was used (Figure 10) (“Wireless Vantage Pro2
TM
& Vantage Pro2
TM
Plus Stations,”
2004.).
40

Figure 10 Weather station (Integrated Sensor Suite) used to collect weather data (Davis
Instruments Vantage Pro2 Weather Station, 2004.)
This weather station consists of six sensors: rain collector, temperature and humidity sensors,
anemometer, solar radiation, and UV sensor (Table 4).
Table 4 Description of parameters measured for climate conditions: (“Wireless Vantage Pro2
TM

& Vantage Pro2
TM
Plus Stations,” 2003.)
No. Parameter Description
1 Wind The anemometer measures wind speed/run (nautical miles) and direction.
The console calculates a 10-minute average wind speed and 10-minute
dominant wind direction.
2 Temperature The temperature sensor is vented and shielded to minimize the solar
radiation induced temperature error.
3 Humidity Humidity refers to the amount of water vapor in the air. Relative humidity
reflects the quantity of water vapor in the air as a percentage of the amount
the air is capable of holding.
4 Dew Point Dew point is the temperature to which air must be cooled for saturation
(100% relative humidity) to occur, providing there be no change in water
vapor content.
5 Rain No rain was recorded during the study period.
6 Solar
Radiation
A measure of the intensity of the sun’s radiation reaching the horizontal
surface at any given time. This irradiance includes both the direct
component from the sun and the reflected component from the rest of the
sky.

41

3.2.3 Indoor conditions
Indoor conditions are measured using two different tools. The first is the buildings inbuilt building
management system (BMS) that provides zonal information through a software developed by
Schneider electric. BMS is a monitoring and control system that allows integrating access control,
video surveillance, HVAC and lighting control (“Andover Continuum - Schneider Electric” 2016).
HOBO Data Loggers are placed to for spot measurements (“HOBO,”2003) (Table 5).
Table 5 Indoor Parameters and Tools Used
No Parameter Tool Description
1 Zonal Temperature Building Management System, Andover Continuum (Figure 11
Screenshot of Building Management System)
2 Spot Temperature HOBO Data Logger (Figure 12 HOBO Temperature and Humidity
Data Logger by Onset Corp and Telaire Carbon Dioxide Sensor)
3 Relative Humidity HOBO Data Logger (Figure 12 HOBO Temperature and Humidity
Data Logger by Onset Corp and Telaire Carbon Dioxide Sensor)

Figure 11 Screenshot of Building Management System

42

Figure 12 HOBO Temperature and Humidity Data Logger by Onset Corp and Telaire Carbon
Dioxide Sensor
3.2.4 Sub-metering for energy consumption
The building energy consumption was broken down into eight major end-use loads: (1) HVAC,
hot water supply (HWS), kitchen, lighting, refrigerator, amusement and information, housework
and sanitary, and others. Wireless sensors developed by panoramic power are used on a circuit
level deployed across two main power boards. Data was collected by transmitting information
every 10 seconds and stored in the cloud and can be monitored online (“Panoramic Power,”
2007). By installing sensors for real-time circuit-level energy measurement that gives visibility
to energy usage by real-time dashboards and reports.

Figure 13 Wireless Circuit Level Clip on sensors for collecting sub-metered data
43

Hardware Used for Sub-metering: PAN-14 high-current wireless sensor; powered by magnetic
fields (does not require maintenance, service or battery replacement) Size: 1.33 x 1.14 x 1.67
inch Manufacturer: Panoramic Power. A total of 73 clip-on were used to measure energy
consumption at the circuit level.Energy-using items in chosen offices are categorize by type
(Table 6):
Table 6 Location of Wireless Clip-Ons and what they represent
No. Parameter
Category
Description
1 Boiler Water heater 7, water heater 9, water heater 11, Instant hot at men’s
shower 10
2 Servers Server Rcpt 10, Server Recept 12, Server Rcpt 14, Server Rcpt 16.
3 Computer Furniture Partition 1,2,3,4,5,6,7,8
4 Display
Screen
TV Rcpt 18, TV Receptacle 33
5 Fan EF1 & EF2 6, Ceiling Fan 13
6 HVAC AHU 19, 21, 23, RA-1 HVAC, VAV 4, HVAC Controls 25
7 Kitchen
Equipment
Ice Machine 36, Dish Washer #1 38, Dish Washer #2 40, Microwave 23,
Kitchen Rcpt 20, Kitchen Rcpt 22, Garbage Disposal 25
8 Lighting Electrical and machine lights 42, PE1, PE7, PE10, Exit signs 21, Track
Lighting 27, EM Egress Lights 15, EM Egress Lights 18, EM Egress
Lights 30, Shower Lights 14
9 Mains Work Room Rcpt 11, Work Room Receptacles 31, Conv. Outlets 27,
Work Room Fixtures 19, Phone Boot Hallway 29, Conv and phone room,
recept 30, Keyless Fixtures 15, Keyless Fixtures 17, Conv. Outlet 29,
Floor Receptacles Reception 10, Work Room receptacles 9, Phone room
receptacles 34.
10 Office
Appliances
Copier Outlet 16, Copier Outlet 18, Copier Rcpt 21, Copier 32, Plotter
and IT Storage 15, Plotter and storage rcpt 13, Existing WiFi 29
11 Misc CUI 35, 16L Panel 12, 16L Panel 17, CUI 37, GFI recept in bathrooms
17, CSFD 23, Spare 13
12 Refrigeration Refrigeration #1 24, Refrigeration #2 26
13 Sub Panel
Mains
16A Panel Sub Main (W), 16B Sub-main (W)
The number accompanied to each element like water heater, furniture partition are circuit number
in the electrical panel making them easy to trace the circuit in the field.
3.2.5 Occupancy Sensors:
USB based infra-red people counter was the used occupancy detection system for the study. The
main objective was to detect the presence and count the number of people and both these objectives
are achieved by using infra-red sensor. Also, available resources like time and money are of the
reasons for choosing USB based infra-red people counter.
44

The technical specification of occupancy sensor used are as below:
Hardware – USB Based Infra-red People Counter with PC Data Viewing (PC-002U)
Dimensions - 116.4 x 68.6 x 22.3 mm
Material - ABS Plastic, Power Supply - 2 x 1.5v AA approximately 120uA
Battery life - Approximately four years, Maximum counting distance: 33 feet (10 m)
Maximum Data Memory: >1 month at 1min. interval
Data – USB
The transmitter unit (PTX20-1) transmits two infrared beams to the receiver unit (PRX20U1).
When someone passes between the left beam first (seen as facing transmitter) and breaks the
infrared beam, the internal counter increases by one for counter A. Vice versa on the right side,
counter B. Therefore the total count for the space occupants is counter A minus counter B. The
data is stored in the counter, and the data can be downloaded to a PC via the USB drive provided.
The picture of installed occupancy sensor on site (Figure 15 Picture of installed occupancy sensor
on siteand Figure 16 Close up of the installed occupancy sensor on site

Figure 14 USB Based Infra-red People Counter (All-Tag People Counter with USB Capability,
Software Included, Wireless Battery Operated, n.d.)

45

Figure 15 Picture of installed occupancy sensor on site – marked in red

Figure 16 Close up of the installed occupancy sensor on site
46

3.2.6 Data monitoring & Collection
All sensors ran for six weeks starting from October 10 to November 22, 2015. After that, all
building-related data through weather station, building automation systems, data loggers,
occupancy sensors, sub-metering data was extracted and gathered, and then the database was
constructed. Since each tool has a specific software for data extraction and does not share the same
extension for data files, all data were converted to an excel spreadsheet file. All data was collected
on a minute resolution; the database will consist of data broken down to 1 minute, 15 minutes and
60 minutes resolution. Also, hourly and daily data as a whole was compiled.
3.3 Methodology for simulation calibration
Simulation calibration was done in four steps (Figure 17).

Figure 17 Showing methodology for calibration
The baseline model was created in Integrate Environmental Solution (IES) software. Then the
weather file was edited as per the actual measured weather data. Based on the sub-metering data,
lighting, plug load and computer profiles are derived and finally from the measured occupancy
data; occupancy profiles are derived. To derive the diversity profile for each load type and
occupancy – the measured value at each time step was divided by the maximum measured value
47

of that type during the study period. Results are compared against one another. The three
different cases studied are explained in table below
Table 7 Different cases studies
Measured Weather Data Actual Sub-Metering Data Actual Occupancy
Base Case × × ×
Case 1 √ × ×
Case 2 √ √ ×
Case 3 √ √ √
3.3.1 Baseline Energy Model
Based on the architectural drawing the building geometry and all the architectural elements like
glazing and shading devices are modeled in IES energy simulation program. The office is
divided into open office space, meeting room, breakout area, etc. (Figure 9 Location of sensors
(1.5 m above floor level) installed in study space) showing the office floor plan. However, while
modeling the geometry, these space are not merged into one space, and the office is modeled as
one space because zones occupancy data is not measured. Though the office is located on sixteen
floors, only one floor was modeled and the adjacent building ignored because of the time
constraint. Material properties are used based on the basis of design (BOD) report available with
from the office. The table below shows the values for BOD prepared by GAIA. Occupancy,
lighting, and equipment schedule as defined by ASHRAE 90.1 Appendix G as used for the base
case. The occupancy for the space was designed at maximum capacity of 60 people. Lighting is
designed at 0.8W/ft2. Ventialtion rate is ASHRAE 62.1 plus 30%. All these values are taken
from record document - basis of design (BOD) and as built drawing – from office.
Table 8: Material Properties as defined in basis of design
Exposed Elements U-Values SHGC Construction
Type
Comments
Roof 0.214 - Light weight
concrete
Heat capacity 4.46
Btu/deg. F-sq. ft.
Walls, above grade 0.438 - Frame Wall Heat capacity 3.18
Btu/deg. F-so. ft.
Windows (No film) 1.19 0.68 - 30-40% WWR
Windows (Low E film)
Skylight
1.04 0.35 - 30-40% WWR
The default HVAC system with equipment efficiency meeting the minimum criteria as specified
in California’s Title 24 and typical meteorological year (TMY3) weather file downloaded from the
department of Energy (DOE) website for Los Angeles was used for simulation. TMY3 are data
sets of hourly values of solar radiation and meteorological elements for a (Huang and Crawley
48

1996b) one year period. Their intended use was for computer simulation of solar energy conversion
system to facilitate performance comparisons of different system types, configuration, and
locations. Simulation was run only for six week period starting from October 10 to November 22
in 2015 for which the data was collect. Following setting are used while running the simulation:
1. Simulation time step – 10 min
2. Reporting interval – 30 min
3. Preconditioning period – 10 days
Comparison with sub-metering data:
Collected submetering data has different end-use load. These end use loads are compared with
simulation results.Mean biased error (MBE) and Coefficient of Variation of the Root mean square
error (CVRSME) are used as evaluation indices. As per American Society of Heating Refrigeration
and Air Conditioning Engineering (ASHRAE) Guideline 14-2002 the computer model shall have
an MBE & CVRSME value less than or equal to 5% and 15% respectively for monthly calibration
for hourly it shall be less than or equal to 10% and 30% respectively.
MBE and CVRSME are evaluated as follows:
𝑀𝐵𝐸 =
∑ (𝑦 𝑖 − 𝑦 𝑖 ̂)
𝑁𝑠
𝑖 =1
∑ 𝑦 𝑖 𝑁𝑠
𝑖 =1

𝑌 𝑠 ̅
=
∑ 𝑦 𝑖 𝑁 𝑠 𝑖 =1
𝑁 𝑠
𝐶𝑉𝑅𝑀𝑆𝐸 (𝑠 )
=
√∑ ((𝑦 𝑖 − 𝑦 𝑖 ̂ )
2
𝑁 𝑠 ⁄ )
𝑁 𝑠 𝑖 =1
𝑁 𝑠
where y i is the measured data; 𝑦 𝑖 ̂ is the simulated data; N s is the sample size; and 𝑌 𝑠 ̅
is the sample
mean of measured data (Lam et al. 2014).
3.3.2 Creation of Weather File
EnergyPlus has developed an application which can convert various forms of weather file in .epw
format. It is called EnergyPlus weather converter. This application was used for converting Los
Angeles (LAX) typical meteorological year (TMY3) weather file into .csv format. The .csv format
can be opened in Microsoft excel. The parameters from the weather file are replaced by actual
49

measured weather parameters like a dry bulb, wet bulb, dew point temperature, relative humidity,
solar radiation, the wind, etc. Then this .csv file was converted again to .epw format and was used
for simulation (Figure 16). Simulation was run using the same setting are explained in the section
above. MBE and CVRSME are calculated to evaluate how close the simulation results are
compared with the actual measured energy data.

Figure 18 Process of creating weather file as per measured weather data
3.3.3 Sub-metering Data to Energy Model
From the sub-metered data lighting, computer and plug loads profiles are derived by dividing the
measured value at a given time step by maximum value during the study period. IES Ergon a
cloud-based application used to create the file format of the profiles which can be read by the
software. ERGON allows an individual or energy modeler to import, manage and interrogate real
building profile or schedule data (down to 1-minute time steps) for use within your VE
simulations. One can utilize measured data from the actual building they are investigating to create
profiles that enhance model calibration. Alternatively, one can use normalized benchmark data
from other buildings of its type. Such profiles can be used to improve operational models or help
close the performance gap by bringing energy models closer to reality (“IES Ergon,” 2013). The
created profiles are used in simulation software.
50

3.3.4 Occupancy Profile from actual measured occupancy
In section 43, the measurement technique and information about sensors was explained in details.
Occupancy profile was created from the measured occupancy data.
3.4 Data Driven Modeling
The schematic layout for methodology for data driven modelling (Figure 19):
`

Figure 19 Methodology for Data Driven Modeling (Images source: “Icon Experience,” 2012)
Methodology for using data driven modeling can be explained stepwise as below:
a) Data collection:
Various parameters are measured for an office building for six weeks. The process of
measuring and collecting data was explained in detail in section 3.2.
b) Parameter identification:
Data Set
Identify key
parameter
Artificial
Neural
Network
Predict End
Use Energy
Usage
51

Measured parameters are mainly of two types – independent and dependent parameters.
The end use energy usage are dependent parameters while occupancy and weather
parameters like temperature, relative humidity, and other are independent parameters.
c) Model building- artificial neural network
A time series artificial neural network model with independent variables trying to predict
dependent variables. The model was trained on six week data. Rapid Miner, an open source
predictive analytical platform, was used for building neural network model (“Rapid
Miner,” 2013). The model was validated using time series validation.
d) For each model tested predictions are made for 20
th
Nov for fan and lighting energy usage.
The MBE and CVRSME values are calculated and results are compared.
e) Different parameter set and network configurations are tested to find best fit model.
f) To study the effect of thermal mass on energy usage, occupancy measured at time interval
t was used to predict energy usage at time interval t+15 minutes, t+30 minutes, t+45
minutes and t+60 minutes, from the best fit model.
3.4.1 Model building – artificial neural network:
Rapid Miner, an open source predictive analytical tool was used to build neural network model.
“Rapid Miner is free open source visual environment for predictive analysis and data mining.
Rapid Miner is based on an XML internal process structure, it has an intuitive graphical user
interface and no programming is required. For these reasons, it is one of the best open source
data mining tools both in terms of technology and applicability” (D’Oca and Hong 2015). The
model was built on 3874 data points.
3.4.2 Validation of model:
Rapid Miner has a built in time series validation operator called sliding window validation. This
operator has three setting – training window width, training window step size and testing
window width. Training window width – number of examples in window which is used for
training. Training window step size – number of examples the window is moved after each
iteration. Test window width- number of examples which are used for testing. For model
building following setting were used. Training window width was set to 390, testing window
width was set to 100 and training window test size was set to 390. So the tool will do the
52

following calculation- it will train the model based on first 390 data points, then it will test the
model based on next 100 data points that is 391- 491 and then built the model based on next 390
data points that is 391 to 780 and so on till it reaches 3874 data points (“Question Abut: Sliding
Window Validation...” 2016). The purpose of selecting 390 is that the data set should be divided
into at least 10 sets so that the validation is valid (Rob J Hyndmanby 2010).
3.5 Summary
The methodology for data collection, simulation calibration and predicting the energy
consumption by using the artificial neural network was discussed. Data was collecting from
weather station, data logger, sub metering system, building management system and occupancy
sensor. Calibration simulation was done in four steps, with first step – creating base, second step
– editing weather data, third step – editing equipment and lighting profiles, fourth step – use real
time occupancy data. For data driven modeling data is collected, parameters are identified, model
was built and end use energy usage was predicted. In the next chapter, the collected data is
documented.

53

Chapter 4: Results - Occupancy Information and measurements data
Overview:
In chapter 3, the methodology for data collection using various sensor and tools was discussed. In
this chapter the collected data for six week is presented.
Week Period – All in year 2015
1 11 Oct to 17 Oct
2 18 Oct to 24 Oct
3 25 Oct to 31 Oct
4 1 Nov to 7 Nov
5 8 Nov to 14 Nov
6 15 Nov to 21 Nov
Although the data was collected at 1 minute resolution the occupancy sensor collects the data at
15 minute interval. Hence taking this in consideration, the data was compiled at 15 minute
resolution. Example for one hour is - 1:00, 1:15, 1:30 and 1:45.
4.1 Climate data
All recorded climate data - outdoor dry bulb temperature, solar radiation and wind directions
were documented in graphical format.
4.1.1 Outdoor Temperature in degree Celsius:
Variation in temperature change was seen in the recorded temperature for the measured outdoor
dry bulb temperature (Figure 20 Recorded Outdoor Temperature).
54

Figure 20 Recorded Outdoor Temperature
The hottest recorded dry bulb temperature was on Monday October 12
th
at 3:00 pm, followed by
Saturday October 24
th
at 1:45 pm.
Hourly comparison of recorded dry bulb temperature and relative humidity for the hottest day and
coldest versus average TMY2 values for October and November respectively (Figure 21 and
Figure 22).
0.00
5.00
10.00
15.00
20.00
25.00
30.00
35.00
40.00
DB (
0
C)
Week
Outdoor Dry Bulb Temprature (
0
C)
Dry Bulb
55

Figure 21 Day profile for outdoor temperature and relative humidity recorded for hottest day
(Monday October 12
th
) and average TMY2 values for October

0
10
20
30
40
50
60
70
80
90
100
0
5
10
15
20
25
30
35
40
3:30 AM
4:15 AM
5:00 AM
5:45 AM
6:30 AM
7:15 AM
8:00 AM
8:45 AM
9:30 AM
10:15 AM
11:00 AM
11:45 AM
12:30 PM
1:15 PM
2:00 PM
2:45 PM
3:30 PM
4:15 PM
5:00 PM
5:45 PM
6:30 PM
7:15 PM
8:00 PM
8:45 PM
9:31 PM
10:15 PM
11:00 PM
11:45 PM
Relative Humidity (%)
Temperature ⁰C
Hourly comparison of recorded climate and historical average
Temperature Temperature TMY2 Relative Humidity (%) RH TMY2
56

Figure 22 Day profile for outdoor temperature and relative humidity recorded for recorded
coldest day (Tuesday November 17
th
) and average TMY2 values for November
Based on the weather data collected for Los Angeles, and comparison with historical averages
(TMY2 weather file) the recorded data shows a typical pattern. However, it shows a significant
difference, about 8 to 10 degrees Celsius. This supports the studies suggesting that urban areas,
in this case Downtown Los Angeles, has higher temperature due to heat island effect (CalEPA)
2016).

0
10
20
30
40
50
60
70
80
90
100
0
5
10
15
20
25
30
35
40
12:00 AM
12:45 AM
1:30 AM
2:15 AM
3:00 AM
3:45 AM
4:30 AM
5:15 AM
6:00 AM
6:45 AM
7:30 AM
8:15 AM
9:00 AM
9:45 AM
10:30 AM
11:15 AM
12:00 PM
12:45 PM
1:30 PM
2:15 PM
4:00 PM
4:45 PM
5:30 PM
6:15 PM
7:00 PM
7:45 PM
8:30 PM
9:15 PM
10:00 PM
10:45 PM
11:30 PM
Relative HUmidity (%)
Temperature ⁰C
Hourly comparison of recorded climate and historical average
Low Temperature Temperature TMY2 Relative Humidity RH TMY2
57

4.1.2 Outdoor solar radiation (Watt/m
2
)
The solar radiation was recorded for six weeks (Figure 23). These values were used to update
weather file.

Figure 23 Outdoor Solar Radiation
The maximum recorded solar radiation was 888 W/m
2
on October 25
th
at 12: 00pm. Solar
Radiation levels seem to be lower in Week 5 and 6 compared to levels in Week 2 and 3.
4.1.3 Outdoor Wind direction:
The wind rose diagram from climate consultant for Los Angeles weather file TMY2 and the
cumulative wind direction for the entire study showed similar trends (Figures 26 and 27).
0
100
200
300
400
500
600
700
800
900
1000
0
100
200
300
400
500
600
700
800
900
1000
Solar Radiation (W/m
2
)
Solar Radiation (W/m2)
Outdoor Solar Radiaton (W/m
2
)
Rad. (W/m2) Daily Peak Average Peak Value
58

Figure 24 Wind rose diagram from Climate consultant software

Figure 25 Cumulative wind direction for entire study period
NORTH
5%
NNE
10%
NE
6%
ENE
4%
EAST
2%
ESE
1%
SE
2%
SSE
2%
SOUTH
4%
SSW
9%
SW
21%
WSW
16%
WEST
4%
WNW
5% NW
6%
NNW
4%
Wind Direction
59

The wind was predominant from that SW side with 21% of the reading wind recorded from SW
direction. WSW as the next most frequent direction from which wind was blowing, followed by
NNE and NE.
4.2 Indoor Condition – Building Management System
For the indoor condition, the central chiller chilled water temperature, temperature of air
handling unit, and air handling unit air pressure and volume, were measured.
4.2.1 Central Chiller Chilled Water Temperature
The difference in inlet and outlet temperature can be noticed which indicates the time for which
the system was working in cooling mode (Figure 26 Chilled water inlet and outlet temprature
plot).

Figure 26 Chilled water inlet and outlet temprature plot
Whenever outlet temperature was more than the inlet temperature means that the system was
working in cooling mode. Considering this, out of the recorded 1008 hours, it was notices that
the cooling was on for 895 hours. The maximum recorded temperature difference between
chilled water outlet and inlet temperature was 7.33
0
C.
0.00
5.00
10.00
15.00
20.00
25.00
Chilled Water Inlet and Outlet Temperature (
0
C)
CHW Inlet CHW Outlet
60

4.2.2 Temperature of Air Handling Unit
The difference in recorded supply and return air temperature (SA and RA respectively)
temperature which indicates the system operating in cooling or heating mode was observed
(Figure 27 Air handling unit temperature).

Figure 27 Air handling unit temperature
The minimum recorded supply air temperature was 14.04
0
C. As the basics of design report, the
system was designed for 15
0
C supply air temperature. The temperature goes 1
0
C below at peak
conditions 15
0
C, indicating that system was optimally sized as per studies (Engdahl and
Johansson 2004).
4.2.3 Air handling unit outside air volume (cms) and static pressure (Pa)
Typical weekly pattern of measured air volume flow rate in cubic meter per second (cms) and
static pressure in Pascal (Pa) of the fan (Figure 28 Outside air volume and static pressure.

10.00
15.00
20.00
25.00
30.00
35.00
Air Handling Unit Temperature
SA Temp RA Temp
61

Figure 28 Outside air volume and static pressure
The maximum recorded outdoor air volume rate was 7.73 m/s (16372 ft
3
/min) on November 3
rd

2015 at 12:00 pm. The recorded occupancy at this time was 31. Thus, outdoor airflow rate per
person was 0.24 m/s (528 ft
3
/min), which was more than the required outdoor airflow rate per
person as per ASHRAE standard 62.1.

4.3 Indoor Conditions – Data Logger
For the indoor condition, the air temperature, relative humidity, lighting and carbon dioxide were
measured. Temperature, RH, lighting and carbon dioxide sensors with data logging capabilities
were ran for six weeks starting from October 10 to November 22 and the location of sensors
(Figure 9 Location of sensors (1.5 m above floor level) installed in study space).
4.3.1 Indoor spot measurement of temperature and relative humidity
The recorded temperature and relative humidity for the six weeks shows the temperature was
maintained with the comfort range 24+2
0
C during working hours (Figure 29).
0
100
200
300
400
500
600
0.0000
1.0000
2.0000
3.0000
4.0000
5.0000
6.0000
7.0000
8.0000
9.0000
Static pressure (Pa)
Air vlume (cms)
Axis Title
Outside Air Volume and Static
Outside Air CFM Static Pa
62

Figure 29 Indoor spot measurement

The space was designed for comfort range of 24 + 2
0
C and relative humidity not greater 60% as
per the basis of design document. The working hours are 8 am to 5 pm for the office. Thus for 6
weeks considering weekday only there are 270 hours. For these 270 hours the space was
maintained within the specified conditions for 98% of time.
4.3.2 Indoor spot measurements for lighting
The measure lux level in the office for six weeks shows the lighting level was above 400 lux
during working hours (Figure 30).
0
10
20
30
40
50
60
70
0
5
10
15
20
25
30
35
Relative Humidity (%)
DB (
0
C)
Weeks
Indoor Spot Measurement
Temp (0C) Relative Humidity
63

Figure 30 Measured light intensity
The recommended Lux level for offices is between 300 - 500 Lux as per standard EN 12464-
1:2011(“Philips Lighting Questions Proper Light-Level Standards for Office Workers” 2016).
The maximum recorded lux level was 563.82 lux on 12
th
November at 12:00 pm. Considering
the total 270 working hours the lighting level was found to be at or above 500 lux for 94% of
these hours. During the working hours the lighting level was never recorded below 450 lux.
4.3.3 Indoor spot measurement of carbon dioxide levels (ppm):
The recorded carbon dioxide levels in parts per million (ppm) for six week (Figure 31).
0
100
200
300
400
500
600
Light intensity (lux)
Weeks
Light Intensity (Lux)
Light Intensity (Lux)
64

Figure 31 Measured Carbo dioxide level
As per ASHRAE Standard 62.1: 2007 and occupational safety and health act (OSHA) the
recommended CO 2 level is below 1000 parts per million (ppm) (“Carbon Dioxide Concentration
- Comfort Levels” 2016). The CO2 level was never recorded above 1000 ppm during the 270
working hours except for few hours during 28
th
and 29
th
October. CO 2 level above does not
impose health risk, however is used as an indicator of occupant acceptance of odors.
4.4 Sub-metered Data
4.4.1 Pie-chart for distribution of total energy use in study period by end use:
The summation of all the sub-metered categories was done and the data was rearranged to create
the summary of cumulative power distribution recorded during the six week study period. The
percentage breakdown of energy consumed by each type of end use can be noticed (Figure 32).
400
500
600
700
800
900
1000
1100
1200
1300
CO
2
(PPM)
Axis Title
Carbon Dioxide (CO
2
)
Carbon Dioxide (CO2)
65

Figure 32 Energy consumption
Computer consume most of the energy at 33%, followed by lighting at 24% and then HVAC,
which consist of AHU + Ventilation Fans, at 23%.
4.4.2 Sub-Metered Data – Fans
The recorded fan energy usage and typical trend of energy usage by fan are noted (Figure 33).
HVAC
23%
Water Heating
1%
Lighting
24%
Cooking
6%
Refrigeration
1%
Office Eq.
5%
Computers
33%
Others
7%
Energy Consumption by End-Use
66

Figure 33 Measured fan energy usage
The maximum measured fan energy usage was 11658.24 W on 13
th
October at 12 pm.
From the average daily use and peak daily energy usage difference in average and peak can be
noticed (Figure 34).
0
2000
4000
6000
8000
10000
12000
14000
Fan energy (W)
Weeks
Fan (W)
Fans (W)
67

Figure 34 Average and Peak Daily fan energy use
The difference in average and peak energy indicates that variable air volume system was
working efficiently. The system was responding to the part load demand and was reducing the
fan air volume rates accordingly and thus reducing the fan power.
4.4.3 Sub-metered Data – Water heating
The energy consumed by water heater during the six week study period and the daily average
energy consumption pattern were recorded (Figure 35).
6,000
7,000
8,000
9,000
10,000
11,000
12,000
Energy (Wh)
AHU Fan Energy Use (Wh)
Peak Day (Wh) Average (Wh)
68

Figure 35 Water heating energy usage

Figure 36 Average water heating energy use
Water Heating measured here refers to instant electric hot water in the male and female shower
and also the dishwasher/ kitchen sink. The peak at 8:30am was mainly because some employees
freshen up in the morning as they come to work on bicycle.
0
1000
2000
3000
4000
5000
6000
7000
Water heating (W)
Weeks
Water Heating (W)
Water Heating (W)
0
200
400
600
800
1,000
1,200
1,400
Energy (Wh)
Average Water Heating Energy Use
69

4.4.4 Sub-metered data- Lighting
From the recorded energy usage by the lighting over the studied six week, it was noticed that
lighting energy use throughout the day does not fluctuate too much (Figure 37).

Figure 37 Sub-metering data – lighting
It was observed that the energy usage by lighting peaks about 7:30 am and doesn’t flucuate much
througout the day. When compared against the average occupant schedule, it does not co-vary.
This is explainable as the fixtures are motion-sensitive and does not take into account number of
occupants. If the first person enters the office, lights will be turned on. This was also seen at the
end of the day, where occupants start to leave, lighting use remain the same. It was also noted
that lighting starts at 2.8 kWh. This could mean some lights are switched on all hours such as
egress or corridor lights.
1500
2000
2500
3000
3500
4000
Lighting (W)
Weeks
Lighting (W)
Lighting (W)
70

4.4.5 Sub-metering data – Kitchen energy use
Kitchen equipment here refers to a coffee maker, coffee grinder, microwave, toaster, dishwashers
and sink incinerator. The recorded energy use by kitchen equipment for the studied six weeks
showed that energy use never drops to zero, even during weekends and unoccupied hours (Figure
38).

Figure 38 Sub-metering data – Kitchen
4.4.6 Sub-metering – Refrigeration
The kitchen was also equipped with two commercial size refrigerators. It was not included in the
kitchen equipment category because it operates all day 24 hours. The average daily use by
refrigerator showed peak in morning(Figure 39).

0
500
1000
1500
2000
2500
3000
3500
4000
Kitchen equipment (W)
Axis Title
Kitchen equipment (W)
Kitchen eq. (W)
71

Figure 39 Average daily refrigeration energy use
The daily profile pattern is typical and can be related to how the kitchen space was used, peaks in
the morning, constant all day.
80
100
120
140
160
180
200
Energy (Wh)
Average Daily Refrigerator Energy Use (Wh)
Average
72

4.4.7 Sub-metering Data – Computer (W)
The energy consumed by computer over the studied six week period shows that the computer
base load has decrease after the third week because eight employee left the company and nine
management level employees were working remotely or were on business tour after third week.
(Figure 40)

Figure 40 sub-metering data- computer
The average daily energy consumed by computer and average daily occupancy showed similar
pattern (Figure 41).
0
1000
2000
3000
4000
5000
6000
7000
8000
9000
Computer energy use (W)
Weeks
Computer Energy Use(W)
Computers(W)
73

Figure 41 Average daily computer energy use

It is evident that the computer energy usage and occupancy are correlated.
4.4.8 Sub metering Total energy use:
The cumulative energy consumed by all equipment showed change in based load after third week
noticed (Figure 42).
0
5
10
15
20
25
30
35
40
45
3,000
3,500
4,000
4,500
5,000
5,500
6,000
6:00 AM
6:30 AM
7:00 AM
7:30 AM
8:00 AM
8:30 AM
9:00 AM
9:30 AM
10:00 AM
10:30 AM
11:00 AM
11:30 AM
12:00 PM
12:30 PM
1:00 PM
1:30 PM
2:00 PM
2:30 PM
3:00 PM
3:30 PM
4:00 PM
4:30 PM
5:00 PM
5:30 PM
6:00 PM
Energy (Wh)
Average Daily Computer Energy Use Profile (Wh)
Average Computer Average Occupancy
74

Figure 42 Sub metering total energy use
In the later part of the third week, the base load has dropped from 9000 W to 5000 W. This was
mainly because there were only 36 people working against 53 people working in first three
weeks.
4.5 Occupancy
The main objective was to detect the presence and count the number of people and both these
objectives are achieved by using infra-red sensor. Hence, a USB based infra-red people counter
was the used for occupancy detection. The details of the sensor used, its location and the method
of collecting occupancy data was explained in detail in section 3.2.5.
4.5.1 Occupancy data from the administration:
The office human resource (HR) manager was asked how many people were expected to be in the
office in the week. Although most weeks were similar, the data from the table stood out as they
were significantly different from each other (Table 9 Number of people employed in office as per
. In week 1, when the study started there were 53 people in the office. From August to mid-October,
eight people left the company, hence there are 45 people left in the office. In November, the
management level (9 people) in the company were away, leaving 36 people in the office (week 3).
5000
10000
15000
20000
25000
30000
35000
Total energy (W)
Week
Total - LA Office (W)
Total - LA Office (W)
75

Table 9 Number of people employed in office as per HR manager
Week Week Period People
1 02 August to 08 August 2016 53
2 18 October to 24 October 2016 45
3 1 November to 07 November 2016 36
4.5.2 Occupancy data from sensor:
The occupancy data measured by occupancy sensor for six (Figure 43).

Figure 43 Measured occupancy
The maximum measured occupancy was 66 on 5
th
November 2015. The effect on fewer people
working in office from week 4 was evident (Figure 43). The comparison of recommended
occupancy profile in ASHRAE 90.1 appendix G versus the measured occupancy profile were
significantly different form each other (Figure 44)
0
10
20
30
40
50
60
70
Occupancy (number)
Axis Title
Occupancy (Number)
Occupancy
76

Figure 44 Comparison of ASHRAE 90.1 profile versus actual profile derived from measured
occupancy
4.6 Summary
All the climate and building related data was analyzed and graphically represented. The
comparison of the measured climate data with TMY2 file showed that the measured
temperatures are significantly higher (8
0
C to 10
0
C) than TMY2 file, highlighting the evidence of
heat island effect in urban environment. Even more thorough investigation is required, but from
the spot measurement of indoor conditions, it was evident that the indoor air quality of the
studied office space is good. From the sub metering data end use energy distribution was studied.
Computer followed lighting and then HVAC were the top three energy users. The comparison of
recommended occupancy profile in ASHRAE 90.1 appendix G versus the measured occupancy
profile showed significant differences. The next chapter will discuss results from simulation and
neural network model.
0.00
10.00
20.00
30.00
40.00
50.00
60.00
70.00
80.00
90.00
100.00
5:00 AM
5:45 AM
6:30 AM
7:15 AM
8:00 AM
8:45 AM
9:30 AM
10:15 AM
11:00 AM
11:45 AM
12:30 PM
1:15 PM
2:00 PM
2:45 PM
3:30 PM
4:15 PM
5:00 PM
5:45 PM
6:30 PM
7:15 PM
8:00 PM
8:45 PM
9:30 PM
10:15 PM
11:00 PM
11:45 PM
12:30 AM
1:15 AM
2:00 AM
2:45 AM
3:30 AM
4:15 AM
Percentage of Occupants present
Time
Occupancy profile comparison
Monday Occupancy Tuesday Wednesday Thursday
77

Chapter 5: Analysis and Discussion
Overview:
In the previous chapter collected data was analyzed and graphically presented. The results of the
simulation and analysis are presented and discussed. With the data for weather, occupancy, and
energy collected the next step was to create energy model.
5.1 Base case simulation results
Based on the architectural drawing the building geometry and all the architectural elements like
glazing and shading devices are modeled in IES energy simulation program. The office is
divided into open office space, meeting room, breakout area, etc. as seen in the office floor plan
(Figure 9). However, while modeling the geometry, these space were merged into one space, and
the office is modeled as one space because occupancy data for individual zone was not
measured. The detail process and the assumptions made to create the base case were discussed in
section 3.3.1. The simulation model was created in IES VE (Figure 45). The material properties
were described previously (
Table 8).

78

Figure 45 Base case model in IES VE

Data for fan power, computer lighting and plug load was collected. To compare the loads the
simulation results should highlight fan, lighting, computer and plug load separately. Hence, to
separate the computer and plug load from lighting load, both the load as given a miscellaneous
source of energy (this option is available in IES) and lighting was on electricity. In order to estimate
the fan power detailed APACHE HVAC system was modelled in Apache HVAC module in the
software. Apache HVAC is module in IES VE which uses flexible schematic component-based
approach that enable to quickly assemble HVAC plant and control system. The details HVAC
model was created the IES VE (Figure 46).

Figure 46 HVAC model in APACHE HVAC module in IES VE

79

The symbols and their meaning are explained (Figure 47).

Figure 47 Symbols used in Apache HVAC

80

5.1.1 Base case - measured versus simulated lighting energy
The predicted lighting energy usage by baseline energy model shows significant discrepancy
from the measured energy usage (
Figure 48).

Figure 48 Measure vs Simulated lighting power for week 1
The actual total energy consume during the study period was 5989.62 kW against the predicted
3017.36 kW. The simulated lighting was seen to follow the lighting diversity profile inputted.
But, it was observed that the actual diversity profile was significantly different from the
recommended in ASHRAE 90.1 appendix G. The main reason seen was that ASHRAE
recommends minimum load of 5% during evening 8pm to morning 6am but the actual lighting
never reaches the 5% diversity. A field investigation was done and it was find out that the office
has large number of emergency lighting which has to operate 24 x 7 as per fire norms. The office
has an open plan layout with core of the building in center and corridor all along the core area.
0
0.5
1
1.5
2
2.5
3
3.5
4
Energy (kW)
Day
Measure vs Simulated lighting energy for week 1
Light (kW) Measured Light (kW)
81

These corridors are fire exit. Thus the lights in this large corridor are always on, leading to a lot
of energy wastage.
Table 10 Baseline lighting MBE and CVRSME values

Weather File
changed
Actual Plug Load & Lighting
profiles
Occupancy
Lighting
MBE CVRSME
Baseline X X X 49.62 76.85
*ASHRAE guideline 14 recommends MBE less than 10% and CVRSME less than 30% for hourly
simulation
5.1.2 Base case - measured versus simulated computer load
The predicted computer energy usage by baseline energy model showed significant discrepancy
from measured energy usage (Figure 50 and Table 11).

Figure 49 Measure vs simulated computer for week 1
Similar observation to lighting were seen in the computer loads energy usage
0
1
2
3
4
5
6
7
8
Energy (kW)
Days
Measure vs Simulated computer energy for week 1
Computer (kW) Measured Computer (kW)
82

Table 11 Baseline computer MBE and CVRSME values

Weather File
changed
Actual Plug
Load &
Lighting
profiles
Occupancy
Computer
MBE CVRSME
Baseline X X X -0.04 99.42
Even though, MBE in computer energy usage prediction was within recommended ASHRAE
limit, it CVRSME value was way above the recommended limit. This is because MBE is sum of
difference of measure and predicted energy, which many times cancels positive and negative
value.
5.1.3 Base case - measured versus simulated plug load
The predicted plug load energy usage showed significant discrepancy from measured plug load
energy usage (Figure 50 and Table 12 ).

Figure 50 Measure vs Simulated plug load for week 1

0
2
4
6
8
10
Energy (kW)
Days
Measure vs Simulated Plug load energy usage for week 1
Plug Load Measured Plug Load (kW)
83

Table 12 MBE and CVRSME values for base case plug load

Weather File
changed
Actual Plug Load &
Lighting profiles
Actual Occupancy
Data
Plug Load
MBE CVRSME
Baseline X X X -22.01 131.66
Similar observation to lighting are seen in the plug loads energy usage. The plug load consists of
server load kitchen equipment like coffee machine, microwave oven, boiler, server, display
screen, refrigerator, printer and miscellaneous loads which are unknown. Thus the plug load
consists of server and refrigerator which has to operate 24 x 7 but during field investigation it
was noticed that office appliance like microwave, display screens, printer and coffee machine
were also never switched off. The switching of appliances should be the responsibility of the
employees or office administrative staff as all these small appliances cannot be automated or
scheduled. Thus, the non-switching off of appliances shows the evidence of how occupancy
behavior leading to gap in measured versus predicted energy by energy simulation as discussed
in section 1.4.

84

5.1.4 Base case - measured versus simulated fan power
The comparison of measured fan energy versus the predicted energy usage of fan by baseline
energy model showed that the predicted energy usage by baseline energy model was far apart
from the measured energy usage (Figure 51 and
Table 13).

Figure 51 Measured versus simulated fan power for week 1
Table 13 MBE and CVRSME values for base case fan energy

Fans
MBE CVRSME
Baseline 60.17 78.46
The fan energy used during the period of study was 5788.83kW against the energy predicted by
energy model 2305.81 kW.

0
2
4
6
8
10
12
Energy (kW)
Days
Measured versus simulated fan power for
week 1
Fan Measured Baseline
85

5.2 Changing weather data
EnergyPlus weather converter application was used to convert Los Angeles (LAX) typical
meteorological year (TMY3) weather file into .csv format. Then .csv format was edited with
measured weather data in Microsoft excel. Then edited .csv file was converted again to .epw format
and was used for simulation (Figure 16).
5.2.1 Changing weather file - measured versus simulated lighting, computer and plug load
The predicted lighting, computer and plug load energy after changing the weather file was
unchanged. (Figure 52, 52, 53 and Table 14).

Figure 52 Lighting energy usage

0
0.5
1
1.5
2
2.5
3
3.5
4
Energy (kW)
Lighting energy usage
Measured Data Baseline Changing weather file
86

Figure 53 Computer energy usage

Figure 54 Plug load energy usage
0
1
2
3
4
5
6
7
8
Energy (kW)
Date
Computer energy usage
Computer (kW) Measured Baseline Chaniging weather file
0
2
4
6
8
10
12
Energy (kW)
Days
Plug load energy usage
Plug Load Measured Baaseline Changing weather file
87

Table 14 MBE and CVRSME after changing weather file

Weathe
r File
change
d
Actual
Plug
Load &
Lightin
g
profiles
Actual
Occupanc
y Data
Lighting Computer Plug Load
MB
E
CVRSM
E
MB
E
CVRSM
E
MBE
CVRSM
E
Baselin
e
X X X
49.6
2
76.85
-
0.04
99.42
-
22.01
131.66
Case 1
√ X X
49.6
2
76.85
-
0.04
99.42
-
22.01
131.66
Both the lines are overlapping. This is due to the fact that simulation program uses diversity
profile to predict the lighting energy usage which in reality is very different.
5.2.2 Changing weather file - measured versus simulated fan energy usage
The predicted fan energy usage after changing the weather file was closer to the measured
energy usage compared to baseline model (Figure 55 and Table 15)

Figure 55 Fan energy usage after changing weather file

0
2
4
6
8
10
12
Energy Usage (kW)
Date
Fan energy usage
Fan Measured Baseline Weathre file changed
88

Table 15 MBE and CVRSME values for fan energy usage after changing weather file

Fans
MBE CVRSME
Baseline 60.17 78.46
Case 1 32.16 61.86
*ASHRAE guideline 14 recommends MBE less than 10% and CVRSME less than 30% for hourly
simulation
From the graph and the tables, it is observed that replacing the weather file with real weather data
gives more accurate prediction of fan energy usage.
5.3 Lighting, computer and plug load profiles changed
From the sub-metered data lighting, computer and plug loads profiles are derived by dividing the
measured value at a given time step by maximum value during the study period. IES Ergon a
cloud-based application used to create the file format of the profiles which can be read by the
software.
5.3.1 Changing profile - measured versus simulated lighting, computer and plug load
The predicted lighting, computer and plug load energy usage after changing the profiles was
accurate to the measured and calculated MBE and CVRSME values (Figure 56, 57, 58, and
Table 16).
89

Figure 56 Lighting energy after changing equipment profile

Figure 57 Plug load after changes in profile
0
0.5
1
1.5
2
2.5
3
3.5
4
Energy (kW)
Lighting energy after changing equipment profile
Light (kW) Measured Baseline Weather file change Profile changed
0
2
4
6
8
10
12
Power (kW)
Date
Plug Load
Plug Load Measured Baaseline Changing weather file Chanigng profile
90

Figure 58 Computer energy after changes in profile
Table 16 MBE and CVRSME values after changing profile

Weather
File
changed
Actual Plug
Load &
Lighting
profiles
Actual
Occupancy
Data
Lighting Computer Plug Load
MBE CVRSME MBE CVRSME MBE CVRSME
Baseline
X X X 49.62 76.85 -0.04 99.42 -22.01 131.66
Case 1
√ X X 49.62 76.85 -0.04 99.42 -22.01 131.66
Case 2
√ √ X -0.20 15.09 2.30 14.69 0.38 18.24
*ASHRAE guideline 14 recommends MBE less than 10% and CVRSME less than 30% for hourly
simulation
It was observed that prediction of lighting, computer and plug load was depended on the
accuracy of diversity profile used for these type. If the profiles were more accurate than the
prediction were more accurate. But in reality these end use energy usage are dependent on
human behavior (Fabi et al. 2013) and thus cutting off the relationship between end use energy
usage and human factor by simulation program affect the accuracy of the prediction.
0
1
2
3
4
5
6
7
8
Power (kW)
Date
Comparsion of computer energy usage
Computer (kW) Measured Baseline Chaniging weather file Changing profiles
91

5.3.2 Changing profiles - measured versus simulated fan energy usage
The comparison of predicted fan energy usage after changing the lighting, computer and plug
load profiles showed that the prediction after changing profiles were more accurate compared to
other cases studied (Figure 59 and Table 17)

Figure 59 Fan energy usage after changes in profile
Table 17 MBE and CVRSME values for fan energy after changing lighting, computer and plug
load profiles

Weather File
changed
Actual Plug Load &
Lighting profiles
Actual
Occupancy Data
Fans
MBE CVRSME
Baseline X X X 60.17 78.46
Case 1 √ X X 32.16 61.86
Case 2 √ √ X -4.84 39.39
It was observed that prediction were more accurate compared to baseline energy model with use
of measured occupancy, lighting, plug load because this resembles reality more accurately.
0
2
4
6
8
10
12
14
Energy Usage (kW)
Date
Fan energy usage
Fan Measured Baseline Profiles changed
92

5.4 Occupancy profile change
Using the measured occupancy data occupancy profiles at 15 minute interval are generated. And
IES VE cloud tool Ergon, was used to create file that could input these profiles in software for
simulation.
5.4.1 Changing occupancy profiles - measured versus simulated lighting, computer and plug
load
The comparison of predicted lighting, computer and plug load energy usage after changing
occupancy profile and showed no change compared to case 2 studied in last section. (Figure 60,
61, 62, and
Table 18)

Figure 60 Lighting energy after changing occupancy profiles

93

Figure 61 Computer energy usage after changing occupancy profile

Figure 62 Plug load energy usage after changing occupancy profile

0
1
2
3
4
5
6
7
8
Energy (kW)
Day
Comparison of computer load
Computer (kW) Measured Baseline Changing profiles
0
2
4
6
8
10
12
Power (kW)
Date
Plug Load
Plug Load Measured Baaseline Chanigng profile
94

Table 18 MBE and CVRSME value after changing occupancy profile

Weather
File
changed
Actual Plug
Load &
Lighting
profiles
Actual
Occupancy
Data
Lighting Computer Plug Load
MBE CVRSME MBE CVRSME MBE CVRSME
Baseline
X X X 49.62 76.85 -0.04 99.42 -22.01 131.66
Case 1
√ X X 49.62 76.85 -0.04 99.42 -22.01 131.66
Case 2
√ √ X -0.20 15.09 2.30 14.69 0.38 18.24
Case 3
√ √ √
-0.20
15.09
2.30
14.69
0.38
18.24

5.4.2 Changing occupancy profiles - measured versus simulated fan energy usage
The comparison of fan energy prediction after changing the occupancy profile against the
measured fan energy and baseline and calculated MBE and CVRSME (Figure 63 and
95

Table 19)

Figure 63 Fan energy usage after change in occupancy data

0
2
4
6
8
10
12
14
Energy Usage (kW)
Date
Fan energy usage
Fan Measured Baseline Changing Occupancy profile
96

Table 19 MBE and CVRSME values for fan energy after changing profiles

Weather File
changed
Actual Plug Load &
Lighting profiles
Actual
Occupancy Data
Fans
MBE CVRSME
Baseline X X X 60.17 78.46
Case 1 √ X X 32.16 61.86
Case 2 √ √ X -4.84 39.39
Case 3 √ √ √ 22.65 33.14
It was observed that prediction of fan energy usage was deviating further from the case 2 after
changing the occupancy profile. Ideally it was expected that using real occupancy profile would
yield more accurate results but the results suggest otherwise. This supports the discussion in
section 1.4 that occupancy is not modelled accurately and is the leading cause of deviation of
simulated energy from the measured energy.
5.5 Summary of simulation results:
The summary of all the four cases studied in the simulation can be best judges by comparing the
MBE and CVRSME values (Table 20).
Table 20 MBE and CVRSME values for different simulation cases

Weathe
r File
change
d
Actual
Plug
Load &
Lightin
g
profiles
Actual
Occupanc
y Data
Lighting Computer Plug Load Fans
MB
E
CVRSM
E
MB
E
CVRSM
E
MBE
CVRSM
E
MB
E
CVRSM
E
Baselin
e
X X X
49.6
2
76.85
-
0.04
99.42
-
22.01
131.66
60.1
7
78.46
Case 1
√ X X
49.6
2
76.85
-
0.04
99.42
-
22.01
131.66
32.1
6
61.86
Case 2
√ √ X -0.20 15.09 2.30 14.69 0.38 18.24 -4.84 39.39
Case 3
√ √ √
-0.20
15.09
2.30
14.69
0.38
18.24
22.6
5
33.14
*ASHRAE guideline 14 recommends MBE less than 10% and CVRSME less than 30% for
hourly simulation
It was observed that changing the weather file has no effect on lighting, plug load or computer
load. Whereas prediction of fan energy usage after changing the weather file (using real weather
data) was more accurate compared to using standard weather file TMY3 or TMY2 available from
DOE website or other sources. Deriving diversity profile for lighting, computer and plug load from
97

the measured data and using them for simulation yielded much better MBE and CVRSME value,
thus illustrating better accuracy. In case 3, diversity profiles are derived from actual measured data.
Using actual measured profiles of occupancy one expects the simulation result to be more accurate
compared to case 2, but the results suggest otherwise. This further supports the premise that
occupancy behavior was not modelled accurately in the simulation programs.
5.6 Artificial Neural Network Model
Artificial neural network is families of models, mathematical usually, which mimics the behavior
of human brain by using artificial intelligence technique. A time series artificial neural network
model was used to predict lighting and fan energy. The model was trained on six week data. Rapid
Miner, an open source predictive analytical platform, was used for building neural network model
(“Rapid Miner,” 2013). The model was validated using time series validation. Methodology for
predicting end use energy usage was discussed in detail in section 3.4.
A typical process for predicting end use energy usage consists of training, testing data, windowing
operator, validation and applying model and results. (Figure 64).

Figure 64 Typical Process in Rapid Miner software
Training data- It is an Excel file which is imported into Rapid miner software. In this case it’s a
six week data at 15 minute interval.
Training Data

Training Data

Training Data

Training Data

Training Data
Apply model

Training Data

Training Data

Training Data

Training Data
Validation

Training Data

Training Data

Training Data

Training Data
Windowing

Training Data

Training Data

Training Data

Training Data
Windowing

Training Data

Training Data

Training Data

Training Data
Prediction Data

Training Data

Training Data

Training Data

Training Data
Results

Training Data

Training Data

Training Data

Training Data
98

Testing data – It is an excel file for with all predictor variables and response variable.
Windowing operator – It is an operator which creates the time series for the given data.
Validation operator – This is a nested operator meaning there is another operation which runs
before this operation. The nested operation divides the data into training and testing data. The
training side and tested is shown (Figure 65). In the training side the algorithm used for
prediction is defined. In the case neural network.

Figure 65 Typical validation process
Apply model – This operator will find the best fit model based on the training operation.
Performance operator – This operator will tell how well the model is performing by measuring
the prediction trend accuracy.
Hidden layer – The hidden layer in neural network transforms the input into something the
output can use.
Neurons – A neuron in a neural network node in the network that form the network which mimic
a biological network of neurons in human brain.

Training algorithm
Performance
Operator

Training Data

Training Data

Training Data

Training Data
Apply model

Training Data

Training Data

Training Data

Training Data
99

5.6.1 Lighting energy usage prediction
For lighting prediction two sets of parameters were tested to predict the lighting energy usage. The
first set consist of following seven parameters: dry bulb temperature, relative humidity, wind
speed, wind direction, solar radiation, occupancy, and dew point temperature. The second set of
parameters consisted of solar radiation and occupancy.
Studies have shown that solar radiation and occupancy are the parameters that affect the lighting
energy usage (Haq et al. 2014). Hence these two parameters are selected in one set. The effect of
other parameters on lighting usage is not known, hence they are all grouped into the other set of
parameters to predict the lighting energy usage.
The performance of ANN depends on the choice of input and output parameters, the structure of
ANN, the number of hidden layer, the number of neurons used in each layer, and the training
algorithm. Considering the time constrains and limited computational capabilities, to find the
optimal solution, four different network configuration were tested to predict energy usage with the
first seven set of parameters. First – one layer with 6 neurons, second – two layer with 6 neurons
in each layer, third- two layer with 3 neurons each, fourth – two layer with 6 neurons in first layer
followed by 3 layers in next layer (Figure 66, Figure 67).

Figure 66 Neural network with one and two hidden layer with six neurons

100

Figure 67 Neural network with two hidden layer with three neurons and six three neurons

101

The predicted lighting energy usage for different configuration with measured lighting energy
usage showed that network with two hidden layer and six neurons in each layer was most
accurate (Figure 68)

Figure 68 Lighting energy usage preidction by neural network
Table 21 MBE and CVRSME values for lighting energy prediction
Number of
layer
1 2 2 2
Neurons in each
layer
6 6-6 3-3 6-3
MBE CVRSME MBE CVRSME MBE CVRSME MBE CVRSME
Lighting 0.65% 3.89% -
0.32%
2.98% -
1.78%
3.22% 0.14% 3.68%
2800
2900
3000
3100
3200
3300
3400
3500
11/20/15 12:00 AM
11/20/15 12:45 AM
11/20/15 1:30 AM
11/20/15 2:15 AM
11/20/15 3:00 AM
11/20/15 3:45 AM
11/20/15 4:30 AM
11/20/15 5:15 AM
11/20/15 6:00 AM
11/20/15 6:45 AM
11/20/15 7:30 AM
11/20/15 8:15 AM
11/20/15 9:00 AM
11/20/15 9:45 AM
11/20/15 10:30 AM
11/20/15 11:15 AM
11/20/15 12:00 PM
11/20/15 12:45 PM
11/20/15 1:30 PM
11/20/15 2:15 PM
11/20/15 3:00 PM
11/20/15 3:45 PM
11/20/15 4:30 PM
11/20/15 5:15 PM
11/20/15 6:00 PM
11/20/15 6:45 PM
11/20/15 7:30 PM
11/20/15 8:15 PM
11/20/15 9:00 PM
11/20/15 9:45 PM
11/20/15 10:30 PM
11/20/15 11:15 PM
Energy (W)
Comparison of Various Neural Network Configuration
Measured Predicted 1 layer 6-6 neurons
Predicted 2 layer 6-6 neurons Predicted 2 layer 6-3 neurons
102

*ASHRAE guideline 14 recommends MBE less than 10% and CVRSME less than 30% for hourly
simulation.
All the results are with the limit specified in ASHRAE guideline 14. For hourly prediction the
specified limit is 10% for MBE values and 30% for CVRSME. Though, the values of MBE and
CVRSME are not too far apart from each other, but neural network with two layer and 6 neurons
in each layer has least CVRSME value and neural network with two layer - 6 neurons and 3
neurons respectively had the least value for MBE. Based on the graph and CVRSME value the
second configuration with two layer and six neuron each gives the best prediction of lighting
energy amongst the four configuration tested.
For the second set of parameters, five network configuration were tested. Following were the
configuration tested:
1. One layer – 3 neurons
2. Two layer – 3 -3 neurons
3. Two layers – 2 – 2 neurons
4. Two layers – 2 – 1 neurons
5. Two layers – 3 – 2 neurons
103

The results of the tested five network showed that network configuration with one hidden layer
and three neurons in each layer yielded the most accurate results (Figure 69).

Figure 69 Lighting energy usage
Table 22 MBE and CVRSME values
Number of layer 1 2 2 2 2
Neurons in each layer 3 3-3 2-2 2-1 3-2
MBE CVRSME MBE CVRSME MBE CVRSME MBE CVRSME MBE CVRSME
Lighting 0.57% 2.94% -1.78% 3.28% 0.98% 3.09% -1.51% 3.04% -3.56% 4.47%
*ASHRAE guideline 14 recommends MBE less than 10% and CVRSME less than 30% for hourly
simulation
2800
2900
3000
3100
3200
3300
3400
3500
3600
11/20/15 12:00 AM
11/20/15 12:45 AM
11/20/15 1:30 AM
11/20/15 2:15 AM
11/20/15 3:00 AM
11/20/15 3:45 AM
11/20/15 4:30 AM
11/20/15 5:15 AM
11/20/15 6:00 AM
11/20/15 6:45 AM
11/20/15 7:30 AM
11/20/15 8:15 AM
11/20/15 9:00 AM
11/20/15 9:45 AM
11/20/15 10:30 AM
11/20/15 11:15 AM
11/20/15 12:00 PM
11/20/15 12:45 PM
11/20/15 1:30 PM
11/20/15 2:15 PM
11/20/15 3:00 PM
11/20/15 3:45 PM
11/20/15 4:30 PM
11/20/15 5:15 PM
11/20/15 6:00 PM
11/20/15 6:45 PM
11/20/15 7:30 PM
11/20/15 8:15 PM
11/20/15 9:00 PM
11/20/15 9:45 PM
11/20/15 10:30 PM
11/20/15 11:15 PM
Energy (W)
Comparison of Various Neural Network Configuration
Measured
Predicted 1 layer 3 neurons - Solar and occupancy only
Predicted 2 layer 3-3 neurons - Solar and occupancy only
Predicted 2 layer 2-2 neurons - Solar and occupancy only
Predicted 2 layer 2-1 neurons - Solar and occupancy only
Predicted 2 layer 3-2 neurons - Solar and occupancy only
104

All the results are with the limit specified in ASHRAE guideline 14. The first configuration with
one hidden layer and three neurons give the least CVRSME value of 2.94% and MBE value of
0.57%.
Comparison of two set of parameters:
The comparison of best network configuration of two set of parameters tested showed that the
parameter set with solar radiation and occupancy gave more accurate results (Figure 70).

Figure 70 Lighting energy usage comparison
The comparison of two set of parameters shows that the second set with solar radiation and
occupancy selected as predictor to predict lighting energy usage give better prediction results with
2800
2900
3000
3100
3200
3300
3400
3500
11/20/15 12:00 AM
11/20/15 12:45 AM
11/20/15 1:30 AM
11/20/15 2:15 AM
11/20/15 3:00 AM
11/20/15 3:45 AM
11/20/15 4:30 AM
11/20/15 5:15 AM
11/20/15 6:00 AM
11/20/15 6:45 AM
11/20/15 7:30 AM
11/20/15 8:15 AM
11/20/15 9:00 AM
11/20/15 9:45 AM
11/20/15 10:30 AM
11/20/15 11:15 AM
11/20/15 12:00 PM
11/20/15 12:45 PM
11/20/15 1:30 PM
11/20/15 2:15 PM
11/20/15 3:00 PM
11/20/15 3:45 PM
11/20/15 4:30 PM
11/20/15 5:15 PM
11/20/15 6:00 PM
11/20/15 6:45 PM
11/20/15 7:30 PM
11/20/15 8:15 PM
11/20/15 9:00 PM
11/20/15 9:45 PM
11/20/15 10:30 PM
11/20/15 11:15 PM
Energy (W)
Lighting energy comparison
Measured
Predicted 2 layer 6-6 neurons
Predicted 1 layer 3 neurons - Solar and occupancy only
105

MBE and CVRSME values of 0.57% and 2.94% respectively compared to -0.32% and 2.98%
respectively. This result further justifies studies that have shown that solar radiation and occupancy
effect lighting energy usage.
Lighting energy usage with time lag:
Lighting energy prediction with time lag taken into consideration were tested. The time delay in
change of temperature due to the thermal mass of the material of the building is called time lag
(“Time Lag and Decrement Factor” 2016). To study the effect of thermal mass on energy usage,
occupancy measured at time interval t is used to predict energy usage at time interval t+15 minutes,
t+30 minutes, t+45 minutes and t+60 minutes and MBE and CVRSME values are calculate (Figure
71 and Table 23).

Figure 71 Lighting energy prediction with time lag

2800
2900
3000
3100
3200
3300
3400
3500
11/20/15 12:00 AM
11/20/15 12:45 AM
11/20/15 1:30 AM
11/20/15 2:15 AM
11/20/15 3:00 AM
11/20/15 3:45 AM
11/20/15 4:30 AM
11/20/15 5:15 AM
11/20/15 6:00 AM
11/20/15 6:45 AM
11/20/15 7:30 AM
11/20/15 8:15 AM
11/20/15 9:00 AM
11/20/15 9:45 AM
11/20/15 10:30 AM
11/20/15 11:15 AM
11/20/15 12:00 PM
11/20/15 12:45 PM
11/20/15 1:30 PM
11/20/15 2:15 PM
11/20/15 3:00 PM
11/20/15 3:45 PM
11/20/15 4:30 PM
11/20/15 5:15 PM
11/20/15 6:00 PM
11/20/15 6:45 PM
11/20/15 7:30 PM
11/20/15 8:15 PM
11/20/15 9:00 PM
11/20/15 9:45 PM
11/20/15 10:30 PM
11/20/15 11:15 PM
Lighting Energy Usage (W)
Lighting Energy Use - Time Lag
Lighting Lighting 15mins Lighting 30mins Lighting 45mins Lighting 60mins
106

Table 23 MBE and CVRSME values for lighting prediction with time lag
15 minutes 30 minutes 45 minutes 60 minutes
MBE CVRSME MBE CVRSME MBE CVRSME MBE CVRSME
Lighting -0.59% 2.99% -2.26% 4.64% 2.29% 3.89% 1.76% 3.89%
*ASHRAE guideline 14 recommends MBE less than 10% and CVRSME less than 30% for
hourly simulation
The lighting energy usage prediction with time lag does not give better prediction results
compared to lighting energy prediction without time lag. Thus it can be conclude that occupancy
affect the lighting energy usage without any time lag for the studied data set. The data set is not
large enough to conclude that this is applicable in cases. But observation in real building will
indicate that this is true. If lighting is insufficient in the built environment people inside will
switch on the light and the lighting energy usage will be affected instantaneously.
5.6.2 Fan energy usage prediction:
To predict the fan energy usage, seven independent parameters were selected. 1. Dry bulb
temperature, 2. Relative humidity 3. Wind speed 4. Wind direction 5. Solar radiation 6. Occupancy
7. Dew point temperature. Since the system is operating from 6am to 7pm, the prediction are only
made for this time period at 15 minute interval. Initially, three network configuration were tested
to predict fan energy usage. In these first three configuration one hidden layer was added in each
iteration. It was noted that configuration with two layer was the most accurate of the three. Hence
few alteration, change in number of neurons were made in the network with two hidden layer were
made. This a standard practice to optimize the neural network for better prediction accuracy. Thus,
total five network configuration as below were tested to predict fan energy usage:
1. One layer with six neuron
2. Two layer with six neurons reach
3. Three layers with 6 neurons each
4. Two layer with six and three neurons in first and second layer respectively
5. Two layer with three and three neurons in first and second layer.
The results of five tested network configuration and calculated MBE and CVRSME values showed
second configuration was most accurate (Figure 72 and Table 24).
107

Figure 72 Fan energy prediction with neural network
Table 24 MBE and CVRSME values – Fan prediction
Number of layer 1 2 3 2 2
Neurons in each
layer
6 6-6 6-6-6 6-3 3-3
MBE CVRSME MBE CVRSME MBE CVRSME MBE CVRSME MBE CVRSME
Lighting -
1.66%
32.94% -
1.87%
24.29% 25.49% 44.72% 15.34% 40.70% -
15.17%
35.85%
*ASHRAE guideline 14 recommends MBE less than 10% and CVRSME less than 30% for hourly
simulation
From the graph and the table the configuration with two hidden layer and six neurons in each layer
give the closest prediction of fan energy usage. The MBE and CVRSME values are -1.87% and
24.29%
0
2000
4000
6000
8000
10000
12000
Energy (W)
Fan energy usage
Actual Predicted 1 layer Predicted 2 layer (6-6)
Predicted 3 layer Predicted 2 layer (6-3) Predicted 2 layer (3-3)
108

Effect of time lag on fan energy usage:
Four cases of time lag were tested for the best network of the five network tested in previous
section. The time delay in change of temperature due to the thermal mass of the material of the
building is called time lag (“Time Lag and Decrement Factor” 2016). To study the effect of
thermal mass on energy usage, occupancy measured at time interval t is used to predict energy
usage at time interval t+15 minutes, t+30 minutes, t+45 minutes and t+60 minutes and MBE and
CVRSME values are calculate (Figure 73 and Table 25).

Figure 73 Fan energy preidction with time lag
Table 25 MBE and CVRSME values for fan energy prediction with timelag
15 minutes 30 minutes 45 minutes 60 minutes
MBE CVRSME MBE CVRSME MBE CVRSME MBE CVRSME
Fan 19.74 31.05 5.48 27.92 5.98 25.76 -0.84 25.09
*ASHRAE guideline 14 recommends MBE less than 10% and CVRSME less than 30% for hourly
simulation
0
2000
4000
6000
8000
10000
Energy (W)
Fan Power Prediction - TimeLag
Actual Predicted 2 layer(6-6) - 15 mins
Predicted 2 layer(6-6) - 30 mins Predicted 2 layer(6-6) - 45 mins
Predicted 2 layer(6-6) - 60 mins
109

From the calculated MBE and CVRSME values the time lag of 60 minutes predicts the energy
usage more accurately compared to other time lag duration studied. Thus, it can be concluded that
the fan power is affect approximately 60 minutes after there is change in the occupancy for the
studied data set. This can be explained by as follows:
In reality whenever a person enters the built environment, he/ she will dissipate the heat and this
will affect the heating and cooling energy use but not immediately because each space has its own
thermal mass. Further, the person might use lighting and other equipment in office and that will
affect the fan energy use but that to not immediately. So it can be concluded that for the studied
built environment fan energy will be affected 60 minutes after the occupant enter the office space
for the studied data set.
5.7 Comparisons of prediction by energy simulation model and artificial neural network model
Using measured data and artificial neural network, lighting and fan energy usage was predicted
for 20
th
November 2015. The results of the prediction are discussed in above section. In this
section, results from section 5.5 and 5.6 are compared. The calibration simulation results are at
30 minutes interval while neural network results are at 15 minutes interval. Hence results at same
time at 30 minute interval are compared.
110

5.7.1 Comparison of different lighting energy prediction models
The comparison of best cases of lighting energy prediction by simulation and neural network
showed neural network model was more accuracy (Figure 74 and Table 26).

Figure 74 Comparison of simulation and neural network for lighting
Table 26 MBE and CVRSME values of lighting energy prediction
MBE CVRSME
Calibrated Model -0.20% 15.09%
Neural Network All Parameter -0.32% 2.99%
Neural Network Two Parameters 0.57% 2.94%
*ASHRAE guideline 14 recommends MBE less than 10% and CVRSME less than 30% for hourly
simulation
The evaluated MBE and CVRSME values and the graph, indicates that artificial neural network
with two parameters gives better prediction compared to other models for the studied data set.
5.7.2 Comparison of different fan energy prediction models
The comparison of best cases of fan energy prediction by simulation against the neural network
method showed that neural network model was more accurate (Figure 75 and Table 27).
2.8
2.9
3
3.1
3.2
3.3
3.4
3.5
0:15 1:45 3:15 4:45 6:15 7:45 9:15 10:45 12:15 13:45 15:15 16:45 18:15 19:45 21:15 22:45
Energy (kW)
Comparison of simulation and neural network for lighting
Light (kW) Measured Neural Network All Parameters
All Profile Neural Netwrok - Two Parameter
111

Figure 75 Comparison of simulation and neural network fan energy
Table 27 MBE and CVRSME values for fan energy prediction
MBE CVRSME
Calibrated Model 22.65 33.14
Neural Network 1.83 24.29
*ASHRAE guideline 14 recommends MBE less than 10% and CVRSME less than 30% for hourly
simulation.
The evaluated MBE and CVRSME values and the graph, indicates that artificial neural network
predicts fan energy usage more accurately compared to simulation calibration for the studied data
set.
5.8 Summary
Both energy simulation and neural network model were used to predict the end use energy usage.
The neural network showed better prediction accuracy of end use energy usage. Further, the
results showed that occupancy (behavior) is not modelling accurately in energy simulation
program. In next chapter conclusions, limitations, and future work are discussed.

0
2
4
6
8
10
12
Energy (kW)
Comparison of simulation and neural network fan energy
Fan Measured All Profile Neural Network
112

Chapter 6: Conclusion
Overview:
In the previous chapter results of the analysis are discussed. In this chapter lessons learnt,
limitations, and further work are discussed. The goal as discussed in section 1.6 was to discover a
technique to improve the accuracy of building energy prediction. Thus, two methods were studied
for predicting energy usage – first energy modeling simulation programs and second data driven
artificial neural network model.
6.1 Lessons learnt
Four different building energy modeling simulation are studied and their results are compared. For
the first case, a building energy model was built in IES VE energy modeling software based on as
built conditions and data from basis of design document. This was called the baseline energy
model. The total measured energy consumption of lighting, plug load, computer and fan was 24.73
MW for the studied time period against the simulation results of 19.81 MW. This was due to the
fact the lighting and other equipment did not follow the diversity profile used in simulation which
are recommended in ASHRAE Standard 90.1 appendix G. Thus, highlighting that diversity
profiles in reality are very different from the one recommended in standards and major factor
contributing to performance gap.
In the second iteration the typical meteorological year weather file (TMY3) was edited and
replaced by measured real time weather data. Changing the weather file did not affect any of the
equipment energy usage while the smaller values of calculated MBE and CVRSME indicated that
predicted fan energy usage in second iteration was closer to measured fan energy usage compared
to baseline case.
In the third iteration, from the measured sub-metering data diversity profiles for computer, plug
load and lighting were derived, and the IES cloud tool Ergon was used to create files that can input
the derived profiles in the simulation program. This iteration matched the measured energy very
close to measured data, with the values of MBE and CVRSME being lower than the recommended
values in ASHRAE guideline 14 of less than 10 and 30 for both respectively. But still the prediction
of fan energy usage was not close to measured fan energy.
113

In fourth iteration, diversity profiles were derived from real time measured occupancy data and
were used in simulation. It was expected that results of this iteration would yield better fan energy
predictions compared to the third iteration. But the values of MBE and CVRSME suggest
otherwise. This highlights the fact discussed in section 1.4 that human occupancy (behavior) is not
modeled accurately and is the leading cause of performance gap in energy simulation programs.
After testing the simulation programs, a time series model with neural network algorithm was used
to predict lighting and fan energy usage. For predicting lighting two set of parameters with
different network configuration were tested. Of all the different configuration and parameter sets
it was found that model with measured occupancy and solar radiation as parameter set and one
hidden layer and three neurons yielded the closest prediction with measured data based on the
lowest MBE and CVRSME values. Lighting energy prediction with time lag taken into
consideration were also tested. But the results show that lighting energy was affect instantaneously
with change in occupancy.
Different network configurations were tested for predicting fan energy usage with weather and
occupancy as predictors. The results indicated that out of the five network configuration tested the
network with two hidden layer and six neurons in each layer predicted fan energy usage closet to
the measured data. Effect of time lag on fan energy usage was also tested, and it was found that
considering a time lag of 60 minutes, resulted in more accurate prediction of fan energy usage
compared to not considering time lag.
Comparison of simulation and neural network model showed that neural network model gave
better prediction of lighting and fan energy usage compared to simulation results from energy
modeling software.
6.2 Limitation of study:
The study and conclusions made are based on one month data only. So all the observations and
conclusions made in above section and chapters hold true for studied month and for the studied
office. This was due to the available resource and time limitation. But high resolution data at 15
minutes interval were studied to arrive at reasonable conclusion. Further, the conclusions cannot
be generalized and apply to all types of office building or are true for all around the year. The
114

conclusion that fan power was affect 60 minutes after change in occupancy cannot be
generalized and applied for all around the year. The occupancy was measured at one entry point
only, but the office has another exit which was rarely used. This might have introduced an error
in measurement of occupancy. The sub-metered data has an element called miscellaneous load in
the measurement for which type of load was unknown. This was added in plug load but this
could be light or computer which can lead to further errors in percentage of calculated MBE and
CVRSME. The time series neural network model was trained and tested based on six week of
data. As per the literature review in section 2.5, though this data was sufficient but data sets for
longer duration have resulted in better prediction accuracy. As occupancy at zonal level or reach
room was not available, the office was modeled as single space and thus this would have affect
the prediction of energy model. The comparison was based on MBE and CVRSME values and
graph only but other parameters like time required, cost and computational power may give a
better insight in which method would be more suitable for use in practice.
6.3 Future work
An alternative and more accurate prediction method to predict fan and lighting energy usage in
office building was demonstrated. The same study could be done on larger duration of data set
and on different type of building. Other factor like time, cost and computational power would be
measured to decide which method can be used in practice.
6.4 Summary
Comparison of simulation results with neural network showed that neural network model
predicted end use energy usage more accurately. Using the current sensor technology and the big
data available from building management, the neural network can be used for more reliable
sensitivity analysis. Sensitivity analysis has proved very helpful is selecting retrofit options to
save energy. Better prediction can help building management team to optimize daily operation
and implement better control strategies. With better controls and optimization demand side
management is more effective and adds value to client or property owner.

115

Bibliography:
“1.1 Overview of Time Series Characteristics | STAT 510.” 2016. Accessed March 17.
https://onlinecourses.science.psu.edu/stat510/node/47.
“14.1 - Autoregressive Models | STAT 501.” 2016. Accessed March 17.
https://onlinecourses.science.psu.edu/stat501/node/358.
A. Dhar, T.A.Reddy, and D E Claridge. 1999. “A Fourier Series Model to Predict Hourly
Heating and Cooling Energy Use in Commercial Buildings With Outdoor Temperature as
the Only Weather Variable.” Journal of Solar Energy Engineering 121 (1): 7.
“AIA2012-2.jpg (432×283).” 2016. Accessed March 24.
http://www.cadalyst.com/files/cadalyst/nodes/2012/14616/AIA2012-2.jpg.
All-Tag People Counter with USB Capability, Software Included, Wireless Battery Operated.
n.d.
“Andover Continuum - Schneider Electric.” 2016. Accessed February 18. http://www.schneider-
electric.com/en/product-range/6823-andover-continuum/.
“Annual-Report-2014-2015.pdf.” n.d.
Attia, Shady, Andre De Herde, Elisabeth Gratia, and Jan L. M. Hensen. 2013. “Achieving
Informed Decision-Making for Net Zero Energy Buildings Design Using Building
Performance Simulation Tools.” Building Simulation 6 (1): 3–21. doi:10.1007/s12273-
013-0105-z.
Benezeth, Y., H. Laurent, B. Emile, and C. Rosenberger. 2011. “Towards a Sensor for Detecting
Human Presence and Characterizing Activity.” Energy and Buildings 43 (2-3): 305–14.
doi:10.1016/j.enbuild.2010.09.014.
Blight, T., and David A. Coley. n.d. “Modeling Occupant Behaviors in Passivhaus Buildings:
Bridging the Energy Gap.” In CIBSE Technical Symposium, DeMonfort University,
Leicester, UK, 9:1–13.
Brandem. 2016. “ThermalComfort.pdf.” Predicting Thermal Comfort. Accessed March 17.
http://ceae.colorado.edu/~brandem/aren3050/docs/ThermalComfort.pdf.
“Buro-Happold-DT-LA-Office_cropped-Copy.jpg (901×452).” 2016. Accessed March 13.
http://blog.archpaper.com/wp-content/uploads/2014/01/Buro-Happold-DT-LA-
Office_cropped-copy.jpg.
(CalEPA), California Environmental Protection Agency. 2016. “Urban Heat Island Index for
California.” Accessed March 20. http://www.calepa.ca.gov/UrbanHeat/Index.htm.
“Carbon Dioxide Concentration - Comfort Levels.” 2016. Accessed March 16.
http://www.engineeringtoolbox.com/co2-comfort-level-d_1024.html.
Cathy, Turner, and Frankel Mark. 2008. “Energy Performance of LEED for New Construction
Building.” New Building Instituiton.
Chang, Wen-Kuei, and Tianzhen Hong. 2013. “Statistical Analysis and Modeling of Occupancy
Patterns in Open-Plan Offices Using Measured Lighting-Switch Data.” Building
Simulation 6 (1): 23–32. doi:10.1007/s12273-013-0106-y.
116

Christensen, K., R. Melfi, B. Nordman, B. Rosenblum, and R. Viera. 2014. “Using Existing
Network Infrastructure to Estimate Building Occupancy and Control Plugged-in Devices
in User Workspaces.” International Journal of Communication Networks and Distributed
Systems 12 (1): 4–29. doi:10.1504/IJCNDS.2014.057985.
Clarke, J. A., P. A. Strachan, and C. Pernot. 1993. “An Approach to the Calibration of Building
Energy Simulation Models.” TRANSACTIONS-AMERICAN SOCIETY OF HEATING
REFRIGERATING AND AIR CONDITIONING ENGINEERS 99: 917–917.
David Ruch, and David Claridge. 1992. “A Four-Parameter Change-Point Model for Predicting
Energy Consumption in Commercial Buildings.” Journal of Solar Energy Engineering
114 (2): 7.
Davis Instruments Vantage Pro2 Weather Station. n.d.
Dhar,A, Reddy T.A, and Claridge D.E. 1998. “Modeling Hourly Energy Use in Commercial
Buildings with Fourier Series Functional Forms.” Journal of Solar Energy Engineering,
Transactions of the ASME 120 (3): 217–23.
D’Oca, Simona, and Tianzhen Hong. 2015. “Occupancy Schedules Learning Process through a
Data Mining Framework.” Energy and Buildings 88 (February): 395–408.
doi:10.1016/j.enbuild.2014.11.065.
Dong, Bing, Burton Andrews, Khee Poh Lam, Michael Höynck, Rui Zhang, Yun-Shang Chiou,
and Diego Benitez. 2010. “An Information Technology Enabled Sustainability Test-Bed
(ITEST) for Occupancy Detection through an Environmental Sensing Network.” Energy
and Buildings 42 (7): 1038–46. doi:10.1016/j.enbuild.2010.01.016.
“DrCEUS: Energy and Demand Usage from Commercial On-Site Survey Data (Abstract Only) |
Energy Efficiency Program Library.” 2016. Accessed March 17.
http://library.cee1.org/content/drceus-energy-and-demand-usage-commercial-site-survey-
data.
“EBC Annex 53 Total Energy Use in Buildings: Analysis & Evaluation Methods.” 2016.
Accessed March 13. http://www.iea-ebc.org/index.php?id=141.
Ekici, Betul Bektas, and U. Teoman Aksoy. 2009. “Prediction of Building Energy Consumption
by Using Artificial Neural Networks.” Advances in Engineering Software 40 (5): 356–62.
doi:10.1016/j.advengsoft.2008.05.003.
“Energyplus.jpg (182×121).” 2016. Accessed March 14.
https://energyplus.net/sites/all/themes/eplus_bootstrap/images/energyplus.jpg.
Engdahl, Fredrik, and Dennis Johansson. 2004. “Optimal Supply Air Temperature with Respect
to Energy Use in a Variable Air Volume System.” Energy and Buildings 36 (3): 205–18.
doi:10.1016/j.enbuild.2003.09.007.
“Espresso-Machine.jpg (350×318).” 2016. Accessed March 18.
http://coffeefancier.com/blog/wp-content/uploads/2013/10/Espresso-Machine.jpg.
Fabi, Valentina, Rune Vinther Andersen, Stefano P. Corgnati, and Bjarne W. Olesen. 2013. “A
Methodology for Modelling Energy-Related Human Behaviour: Application to Window
Opening Behaviour in Residential Buildings.” Building Simulation 6 (4): 415–27.
doi:10.1007/s12273-013-0119-6.
117

Fels, Margaret F. 1986. “PRISM: An Introduction.” Energy and Buildings 9 (1-2): 5–18.
“Fourier Series - MATLAB & Simulink.” 2016. Accessed March 18.
http://www.mathworks.com/help/curvefit/fourier.html.
Gunay, H. Burak, William O’Brien, and Ian Beausoleil-Morrison. 2013. “A Critical Review of
Observation Studies, Modeling, and Simulation of Adaptive Occupant Behaviors in
Offices.” Building and Environment 70 (December): 31–47.
doi:10.1016/j.buildenv.2013.07.020.
Gustafsson, S.-I. 1998. “Sensitivity Analysis of Building Energy Retrofits.” Applied Energy 61
(1): 13–23.
Haq, Mohammad Asif ul, Mohammad Yusri Hassan, Hayati Abdullah, Hasimah Abdul Rahman,
Md Pauzi Abdullah, Faridah Hussin, and Dalila Mat Said. 2014. “A Review on Lighting
Control Technologies in Commercial Buildings, Their Performance and Affecting
Factors.” Renewable and Sustainable Energy Reviews 33 (May): 268–79.
doi:10.1016/j.rser.2014.01.090.
Heating, American Society of, Refrigerating, Air-Conditioning Engineers, and American
National Standards Institute. 2004. Thermal Environmental Conditions for Human
Occupancy. Vol. 55. 2004. American Society of Heating, Refrigerating and Air-
Conditioning Engineers.
“HOBO.” n.d. http://www.onsetcomp.com/products/data-loggers-sensors/temperature.
Huang, Y. Joe, and D. B. Crawley. 1996a. “Does It Matter Which Weather Data You Use in
Energy Simulations?” Proceedings of American Council for an Energy Efficient Summer
Study, Pacific Grove, CA, USA.
https://publications.lbl.gov/islandora/object/ir%3A110538/datastream/PDF/download/cita
tion.pdf.
———. 1996b. “Does It Matter Which Weather Data You Use in Energy Simulations?”
Proceedings of American Council for an Energy Efficient Summer Study, Pacific Grove,
CA, USA.
https://publications.lbl.gov/islandora/object/ir%3A110538/datastream/PDF/download/cita
tion.pdf.
“Human Thermal Comfort | Sustainability Workshop.” 2016. Accessed March 17.
http://sustainabilityworkshop.autodesk.com/buildings/human-thermal-comfort.
“Icon Experience.” n.d. https://www.iconexperience.com/g_collection.
“IES Ergon.” n.d. IESVE. https://www.iesve.com/software/cloud-solutions/ergon.
Initiative, Climate. 2009. “Buildings and Climate Change.”
http://admin.indiaenvironmentportal.org.in/files/SBCI-BCCSummary.pdf.
“Introduction to Linear Regression Analysis.” 2016. Accessed March 17.
http://people.duke.edu/~rnau/regintro.htm.
Jonathan Heller, and Morgan Heater. 2011. “Sensitivity Analysis: Comparing the Impact of
Design, Operation, and Tenant Behavior on Building Energy Performance.”
http://newbuildings.org/wp-content/uploads/2015/11/SensitivityAnalysisReport1.pdf.
118

Kalogirou, Soteris A., and Milorad Bojic. 2000. “Artificial Neural Networks for the Prediction of
the Energy Consumption of a Passive Solar Building.” Energy 25 (5): 479–91.
Karatasou, S., M. Santamouris, and V. Geros. 2006. “Modeling and Predicting Building’s
Energy Use with Artificial Neural Networks: Methods and Results.” Energy and
Buildings 38 (8): 949–58. doi:10.1016/j.enbuild.2005.11.005.
Khoury, Hiam M., and Vineet R. Kamat. 2009. “Evaluation of Position Tracking Technologies
for User Localization in Indoor Construction Environments.” Automation in Construction
18 (4): 444–57. doi:10.1016/j.autcon.2008.10.011.
Labeodan, Timilehin, Wim Zeiler, Gert Boxem, and Yang Zhao. 2015. “Occupancy
Measurement in Commercial Office Buildings for Demand-Driven Control
applications—A Survey and Detection System Evaluation.” Energy and Buildings 93
(April): 303–14. doi:10.1016/j.enbuild.2015.02.028.
Lam, Khee Poh, Jie Zhao, Erik B. Ydstie, Jason Wirick, Meiwei Qi, and Jihyun Park. 2014. “An
EnergyPlus Whole Building Energy Model Calibration Method for Office Buildings
Using Occupant Behavior Data Mining and Empirical Data.” In Building Simulation
Conference, 160–67. https://www.ashrae.net/File%20Library/docLib/Events/ASHRAE-
IPBSA-USA/Presentations/21_Lam.pdf.
“Lesson 12: Multicollinearity and Other Regression Pitfalls | STAT 501.” 2016. Accessed March
17. https://onlinecourses.science.psu.edu/stat501/node/343.
“LG_refrigerator_interior.jpg (2448×3264).” 2016. Accessed March 18.
https://upload.wikimedia.org/wikipedia/commons/8/85/LG_refrigerator_interior.jpg.
Li, Nan, Gulben Calis, and Burcin Becerik-Gerber. 2012. “Measuring and Monitoring
Occupancy with an RFID Based System for Demand-Driven HVAC Operations.”
Automation in Construction 24 (July): 89–99. doi:10.1016/j.autcon.2012.02.013.
“Linear Regression.” 2016. Wikipedia, the Free Encyclopedia.
https://en.wikipedia.org/w/index.php?title=Linear_regression&oldid=710339263.
“Maintenance_man.jpg (2357×2121).” 2016. Accessed March 18. http://www.sa4u.net/wp-
content/uploads/2015/09/maintenance_man.jpg.
Milenkovic, Marija, and Oliver Amft. 2013. “Recognizing Energy-Related Activities Using
Sensors Commonly Installed in Office Buildings.” Procedia Computer Science 19: 669–
77. doi:10.1016/j.procs.2013.06.089.
Nassif, Nabil. 2012. “A Robust CO2-Based Demand-Controlled Ventilation Control Strategy for
Multi-Zone HVAC Systems.” Energy and Buildings 45 (February): 72–81.
doi:10.1016/j.enbuild.2011.10.018.
Neto, Alberto Hernandez, and Flávio Augusto Sanzovo Fiorelli. 2008. “Comparison between
Detailed Model Simulation and Artificial Neural Network for Forecasting Building
Energy Consumption.” Energy and Buildings 40 (12): 2169–76.
doi:10.1016/j.enbuild.2008.06.013.
Newsham, Guy R., Sandra Mancini, and Benjamin J. Birt. 2009. “Do LEED-Certified Buildings
Save Energy? Yes, But….” Energy and Buildings 41 (8): 897–905.
doi:10.1016/j.enbuild.2009.03.014.
119

“Office Light Fixtures Office Lighting Fixtures(ACM3209) – China Acmelite,Office Lighting …
| Houseandesign.top.” 2016. Accessed March 18. http://houseandesign.top/office-light-
fixtures/office-light-fixtures-office-lighting-fixturesacm3209-china-acmeliteoffice-
lighting/.
“Officescene.jpg (1024×768).” 2016. Accessed March 14. https://gretchenrubin.com/wp-
content/uploads/2015/01/officescene.jpg.
Oh, Sukjoon. 2013. “Origins of Analysis Methods in Energy Simulation Programs Used for High
Performance Commercial Buildings.” Texas A&M University.
https://oaktrust.library.tamu.edu/handle/1969.1/151151.
“Panaromic Power.” n.d. http://www.panpwr.com/technology.
“People-220284_960_720.jpg (838×720).” 2016. Accessed March 18.
https://pixabay.com/static/uploads/photo/2013/11/28/11/31/people-220284_960_720.jpg.
“Performance Gap Evidence Review Report Released.” 2016. Accessed February 23.
http://www.zerocarbonhub.org/news/performance-gap-evidence-review-report-released.
“Philips Lighting Questions Proper Light-Level Standards for Office Workers.” 2016. Accessed
March 16. http://www.ledsmagazine.com/articles/iif/2015/03/philips-lighting-questions-
proper-light-level-standards-for-office-workers.html.
Pieter de Wilde. 2009. “The Building Energy Performnace Gap: Up Close and Personal.” In .
PSDgraphics.com. 2016. “Weather Icon (PSD) | PSDGraphics.” Accessed March 18.
http://www.psdgraphics.com/psd/weather-icon-psd/.
“Question Abut: Sliding Window Validation...” 2016. Accessed March 14. https://rapid-
i.com/rapidforum/index.php?topic=7109.0.
“Rainshower-System-Shower-System-for-Wall-Mounting.jpg (470×364).” 2016. Accessed
March 18. http://cdn12.grohe.com/~mi/483/953/rainshower-system-shower-system-for-
wall-mounting.jpg.
“RapidMiner.” n.d. https://rapidminer.com/.
“Regression - Why Does a Time Series Have to Be Stationary? - Cross Validated.” 2016.
Accessed March 18. http://stats.stackexchange.com/questions/19715/why-does-a-time-
series-have-to-be-stationary.
Rob J Hyndman. 2010. “The ARIMAX Model Muddle.” Hyndsight. October 4.
http://robjhyndman.com/hyndsight/arimax/.
Rob J Hyndmanby. 2010. “Why Every Statistician Should Know about Cross-Validation.”
Hyndsight. October 4. http://robjhyndman.com/hyndsight/crossvalidation/.
root. 2003. “Sensitivity Analysis Definition.” Investopedia. November 26.
http://www.investopedia.com/terms/s/sensitivityanalysis.asp.
———. 2006. “Autocorrelation Definition.” Investopedia. March 12.
http://www.investopedia.com/terms/a/autocorrelation.asp.
120

Royapoor, Mohammad, and Tony Roskilly. 2015. “Building Model Calibration Using Energy
and Environmental Data.” Energy and Buildings 94 (May): 109–20.
doi:10.1016/j.enbuild.2015.02.050.
Soebarto V.I, and T.J. Williamson. 2001. “Deriving Loadshapes from Utility Bills through
Scaled Simulation.” Seminar Slides presented at the ASHRAE Seminar Slides, Kansas
City, MI.
“Statistical Analysis Handbook.” 2016. Accessed March 17.
http://www.statsref.com/HTML/index.html?arima.html.
Sunikka-Blank, Minna, and Ray Galvin. 2012. “Introducing the Prebound Effect: The Gap
between Performance and Actual Energy Consumption.” Building Research &
Information 40 (3): 260–73. doi:10.1080/09613218.2012.690952.
T.A.Reddy, and Dr. Itzhak Maor. 2006. “Procedure for Reconciling Computer - Calculated
Results with Measured Energy Data.” Research 1051. ASHRAE.
“Thermostat.jpg (847×567).” 2016. Accessed March 18. http://www.bbb.org/blog/wp-
content/uploads/2011/07/thermostat.jpg.
“Time Lag and Decrement Factor.” 2016. Accessed March 21. http://www.new-
learn.info/packages/clear/thermal/buildings/building_fabric/properties/time_lag.html.
Tso, Geoffrey K.F., and Kelvin K.W. Yau. 2007. “Predicting Electricity Energy Consumption: A
Comparison of Regression Analysis, Decision Tree and Neural Networks.” Energy 32
(9): 1761–68. doi:10.1016/j.energy.2006.11.010.
“Typical Layout of Central Conditioning System in a Commercial Building.” n.d.
http://bea.touchstoneenergy.com/resourcelibrary/article/1825.
Uziel, S, Elste, T, Kattanek, W, Goetze, S, and Gerlach. n.d. “Networked Embedded Acoustic
Processing System for Smart Building Applications.” In . Cagliari, Italy.
Wang, Liping, Paul Mathew, and Xiufeng Pang. 2012. “Uncertainties in Energy Consumption
Introduced by Building Operations and Weather for a Medium-Size Office Building.”
Energy and Buildings 53 (October): 152–58. doi:10.1016/j.enbuild.2012.06.017.
“Wireless Vantage Pro2
TM
& Vantage Pro2
TM
Plus Stations.” n.d.
http://www.davisnet.com/weather/products/wx_product_docs.asp?pnum=06162.
“World: [Energy] Balances for 2012.” n.d. International Energy Agency (IEA).
X. Guo, PhD, DK Tiller, DPhil, and GP Henze, PhD PE. 2010. “The Performance of Occupancy-
Based Lighting Control Systems: A Review.” Lighting Research and Technology 42
(December).
Yan, Da, William O’Brien, Tianzhen Hong, Xiaohang Feng, H. Burak Gunay, Farhang
Tahmasebi, and Ardeshir Mahdavi. 2015. “Occupant Behavior Modeling for Building
Performance Simulation: Current State and Future Challenges.” Energy and Buildings
107 (November): 264–78. doi:10.1016/j.enbuild.2015.08.032.
Yang, Jin, Hugues Rivard, and Radu Zmeureanu. 2005. “On-Line Building Energy Prediction
Using Adaptive Artificial Neural Networks.” Energy and Buildings 37 (12): 1250–59.
doi:10.1016/j.enbuild.2005.02.005.
121

Zhang, Weiqian, Shen Tan, Yizhong Lei, and Shoubing Wang. 2014. “Life Cycle Assessment of
a Single-Family Residential Building in Canada: A Case Study.” Building Simulation 7
(4): 429–38. doi:10.1007/s12273-013-0159-y.
Zhaojian Li, Yi Jiang, and QP Wei. 2007. “Survey on Energy Consumption of Air Conditioning
in Summer in a Residential Building in Beijing.” Journal of Heating Ventilation and Air
Conditioning 37 (4): 46–51.
Zhao, Jie. 2015. “Design-Build-Operate Energy Information Modeling for Occupant-Oriented
Predictive Building Control.” Carnegie Mellon University.
http://repository.cmu.edu/cgi/viewcontent.cgi?article=1472&context=dissertations.

Abstract (if available)

Abstract Studies have shown that the prediction of energy usage by simulation program does not always match the actual energy usage. Assumptions about occupancy and use of weather data different from site conditions are few of the factors affecting the accuracy of prediction. Hence, real-time occupancy, sub-metering and weather data for an office building in downtown Los Angeles was measured. Using the measured data, two methods of prediction, calibration simulation and artificial neural network, were studied to predict end-use energy usage. Measured sub-metered and occupancy data was used to generate lighting, equipment, and occupancy profile that were used for calibrating the model. Results showed that when real-time occupancy was used the prediction were less accurate compared to when lighting and equipment profiles were used. This indicates that occupancy was not modeled accurately in the simulation program. Different neural network configuration and time series data with different time lag were studied to best predict end-use energy usage. The fan energy was best predicted with 60-minutes time lag considered. Finally, comparison of neural network with simulation program showed that neural network predictions were more accurate.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

A simplified building energy simulation tool: material and environmental properties effects on HVAC performance

PDF

Impact of occupants in building performance: extracting information from building data

PDF

Building energy performance estimation approach: facade visual information-driven benchmark performance model

PDF

Developing environmental controls using a data-driven approach for enhancing environmental comfort and energy performance

PDF

A parametric study of the thermal performance of green roofs in different climates through energy modeling

PDF

Energy simulation in existing buildings: calibrating the model for retrofit studies

PDF

Bridging the gap: a tool to support bim data transparency for interoperability with building energy performance software

PDF

Microclimate and building energy performance

PDF

Energy efficient buildings: a method of probabilistic risk assessment using building energy simulation

PDF

Impacts of building performance on occupants' work productivity: a post occupancy evaluation study

PDF

Nursing comfort study: are hospitals performing for nurses? IEQ analysis of working conditions inside hospitals

PDF

Post occupancy energy evaluation of Ronald Tutor Hall using eQUEST; computer based simulation of existing building and comparison of data

PDF

Visualizing thermal data in a building information model

PDF

Enhancing thermal comfort: data-driven approach to control air temperature based on facial skin temperature

PDF

Energy savings by using dynamic environmental controls in the cavity of double skin facades

PDF

Energy use intensity estimation method based on building façade features by using regression models

PDF

Multi-occupancy environmental control for smart connected communities

PDF

Exploration for the prediction of thermal comfort & sensation with application of building HVAC automation

PDF

Smart buildings: employing modern technology to create an integrated, data-driven, intelligent, self-optimizing, human-centered, building automation system

PDF

Multi-domain assessment of a kinetic facade: determining the control strategy of a kinetic façade using BIM based on energy performance, daylighting, and occupants’ preferences; Multi-domain asse...

Asset Metadata

Creator Dharane, Aditya Sudhir (author)

Core Title Bridging performance gaps by occupancy and weather data-driven energy prediction modeling using neural networks

School School of Architecture

Degree Master of Building Science

Degree Program Building Science

Publication Date 04/20/2018

Defense Date 03/21/2016

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

Tag building energy calibration modeling,data driven building energy modeling,end-use energy usage,energy prediction,IES VE,neural network,OAI-PMH Harvest,occupancy behaviour,occupancy measurement,occupancy modeling,performance gap,submetering,thermal mass,time lag,time series,weather data

Format application/pdf (imt)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Choi, Joon-Ho (committee chair), Kensek, Karen (committee member), Konis, Kyle (committee member)

Creator Email adityadharane@gmail.com,dharane@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c40-237382

Unique identifier UC11277240

Identifier etd-DharaneAdi-4329.pdf (filename),usctheses-c40-237382 (legacy record id)

Legacy Identifier etd-DharaneAdi-4329-0.pdf

Dmrecord 237382

Document Type Thesis

Format application/pdf (imt)

Rights Dharane, Aditya Sudhir

Type texts

Source 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 a...

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