Close
About
FAQ
Home
Collections
Login
USC Login
Register
0
Selected
Invert selection
Deselect all
Deselect all
Click here to refresh results
Click here to refresh results
USC
/
Digital Library
/
University of Southern California Dissertations and Theses
/
Smart buildings: employing modern technology to create an integrated, data-driven, intelligent, self-optimizing, human-centered, building automation system
(USC Thesis Other)
Smart buildings: employing modern technology to create an integrated, data-driven, intelligent, self-optimizing, human-centered, building automation system
PDF
Download
Share
Open document
Flip pages
Contact Us
Contact Us
Copy asset link
Request this asset
Transcript (if available)
Content
1
Smart Buildings: employing modern technology to create an Integrated, Data-Driven, Intelligent,
Self-Optimizing, Human-Centered, Building Automation System
Sean Robert Birnbaum
Green Technologies
Master of Science
University of Southern California
Degree conferral: August 2019
2
Abstract: 3
Introduction: 4
Literature Review: 5
The Big Picture 20
Climate Zones 24
Where does the energy go and how much is used? 26
Contemporary Building Automation Systems 27
The Importance of Data 28
Recurrent Neural Networks- Machine Learning 29
The Human Element 32
Developed Concepts Revisited and/or Improved: 36
Workplace Applications: 36
Occupancy Forecasting: 38
Smart Lighting: 40
Plug Load Forecasting: 41
Model Predictive Control of HVAC and HVAC Forecasting: 42
HVAC System Optimization: 45
Use of External Data Inputs (Weather and Utility): 47
Renewable Energy Forecasting: 48
New Ideas for Next Generation BAS: 49
Distributed Generation Dispatching and CHP/ Alternative Energy Forecasting: 49
Electrical Energy Storage (EES) Forecasting/Dispatching: 51
Thermal Energy Storage (TES) Forecasting/Dispatching: 52
Building Level Energy Forecasting: 54
Building Operational Cost Forecasting: 56
Integrated BAS Architecture: 57
Future Development: 58
Conclusion: 59
Bibliography: 61
3
Abstract:
Advanced technology is becoming increasingly more accessible, affordable, and applicable with respect
to building energy systems which once existed and operated in an isolated manner and were managed by
unsophisticated systems. These systems are becoming increasingly more complex, integrated, and
carefully controlled largely due to rising energy costs and expectations for higher levels of indoor
environmental quality. In the past, and in same cases even today, lighting, HVAC, energy storage, plug
loads, and distributed generation systems were viewed as independent systems that enabled limited
human interaction, system monitoring, or control, but today there is growing expectation for optimized
services from these systems requiring a more dynamic, integrated, and intelligent approach to managing
these systems with the added expectation of a provision some sort of human feedback interface to make
buildings more accommodating to their occupants. For too long occupants have felt that buildings are not
functioning in a manner to accommodate their needs, and rather operating to fulfill programmed
parameters. For decades building controls reduced to following relatively simple sequences of operation
developed by engineers to meet basic criteria that fit discrete operational scenarios. The problem with this
approach is that, to some extent, changes in weather and occupancy patterns are not accounted for in the
control scheme, resulting in the unnecessary expenditure of large quantities of energy to operate systems
in specific places at times where they are not really needed. Not only that, but now many facilities are
adding on-site power generation, energy storage, and advanced lighting systems in addition to modern
HVAC systems. As a result of these advancements there is now a growing expectation that buildings
become virtually sentient beings capable of both maintaining healthy and comfortable environments while
also minimizing the use of energy to keep utility costs low. It doesn’t take much research to realize that
costs of energy are on the rise, and are expected to continue for the foreseeable future. Despite that, with
the growing integration of Internet of Things (IoT) in devices and the development and utilization of cloud
computing for offsite computational capacity and data management along with advancements in Artificial
Intelligence, Machine Learning has entered the lexicon of the building automation industry. Building
Automation was once delegated to merely managing HVAC operations with other systems being
managed by other control systems; the expectation now is complete integration of all energy systems
(and in some cases other systems as well) wherein these various systems can communicate and share
data with one another to fully optimize their services. In this thesis it will be shown how Artificial
Intelligence (AI), namely Recurrent Neural Network machine learning (RNN)(Machine Learning), stands to
make vast improvements in the building automation realm, and how Machine Learning can tie supply side
and demand side energy phenomena into an integrated system. By empowering occupants with
workplace applications, and generating forecasts of lighting, plug loads, and HVAC energy usage based
on occupancy forecasts, while also strategically utilizing on-site energy generation and storage, building
owners can track operational costs whilst also providing optimum indoor environmental conditions that
keep tenants comfortable and maximize the operational lifespan of building equipment. This thesis
provides a framework and methodology by which such an integrated system could arranged and
operated.
4
Introduction:
Machine learning is not a new concept, as a matter of fact it has been in use by other industries for quite
some time. Only recently, due to advancements in computing, networking, and sensing, has the topic
become of interest in the facilities management realm. Driving this new found interest in the industry are
ever increasing costs of energy and evolving utility rate schedules. Many corporations are also now
operating with the expectation of demonstrating an internal commitment to combating climate change
through energy efficiency and conservation while also delivering human centric environmental comfort
since studies have shown that comfortable employees are productive employees.
Although it is certainly true that improvements in energy efficiency both in terms of equipment
developments and also in terms of system design have made significant improvements in buildings’
performance potentials, efficiency is only half the solution. A high efficiency system obviously saves
energy compared to lesser efficient systems when both are operating over the same duration of time, but
a lesser efficient system that is well controlled can actually save more energy than a poorly controlled
high efficiency system. Knowing when, for how long, and at what level to run a system is also equally
important in addition to the efficiency of the system. For this functionality we must look to building
automation. Building automation currently exists, like all things, as a spectrum of offerings from the very
basic to the highly sophisticated. Despite this spectrum of offerings by far and large building automation
systems are still static systems in their setup largely relying on building operators to tell them what to do.
Unfortunately, optimizing programming and scheduling often does not make up the bulk of these
individuals time which results in subpar system performance, uncomfortable occupants, unnecessary or
over-conditioning of indoor spaces, and higher utility bills than what could otherwise be achieved with
effective management. Enter the next generation of energy management system operator.
Machine Learning, a form of Artificial Intelligence, takes reliance on an operator out of the picture, instead
relying on aggregated historical data of building operations to produce a model that seeks to predict future
scenarios. Machine Learning can analyze datasets and identify patterns far too complex for humans to do
in any reasonable time frame and learn from that data as well react to it in turn with a favorable control
response. Thanks to communication interconnections that buildings can establish across integral
equipment, sensors, and external data inputs, facilities have the capacity to use weather forecasts, utility
forecasts, historical occupancy pattern data, seasonal weather pattern information, solar path and
intensity forecasts, lighting sensor inputs, HVAC system inputs, building energy meters data streams, and
occupant inputs to best determine both how to operate a building’s many systems and how to best react
to real-time changes induced both internally and externally. Some systems can react more independently
(i.e. lighting) while others need computational intelligence telling them how to run (i.e. HVAC systems).
Every building is it’s own unique animal, so to speak, and is capable, with modern technology, of being
sentient of its surroundings and of its occupants’ behaviors. We have reached a point where the
independently developed intelligence of numerous subsystems are at a confluence of integration where
emergent optimums can be determined in real-time. This thesis presents a controls framework dependent
on AI and optimization intelligence to forecast, control, and optimize the vast array of energy reliant
systems in facilities of all sizes.
5
Literature Review:
In the past ten years the world has seen massive developments in Artificial Intelligence, with a particular
focus on Machine Learning methods. The applications of this technology, however, have been limited to
e-commerce advertising, stock-trading, robotics, web browser search engines, facial and speech
recognition, numerous scientific pursuits, and used internally by high-tech firms to develop their products.
Meanwhile, however, the built environment industry has taking notice of this developing technology and
it’s possible relevance to the industry. The built environment industry is notorious for being a late adopter
of new technology largely due to the high cost to implement, the liability of failure, and the dependence on
building systems. This is an industry where failure to perform is not very acceptable, and the desire to
spend as little as possible to achieve a desired function reigns king. That said, in the wake of ever
increasing energy costs and evolving utility rate schemes, more than ever, building owners and tenant
organizations are looking for ways to reduce energy consumption since those savings go directly to a
business’s bottom line. For years these parties have had to accept control systems that rely on static
programming, and offer little tenant engagement. These parties and the industry that provides building
controls realize that occupancy, weather patterns, and utility prices are not explicitly regular which proves
problematic when relying on a control system that effectively views these as static unchanging
parameters. As a result of this, and the recognition of Machine Learning nearing a state of maturity, the
industry is exploring the application of Machine Learning to building controls. The following review looks
at the current literature on this pursuit as well as industry developments and identifies where work
remains to be done to facilitate full utilization of this technology to provide dynamic and evolving control
schemes for buildings. It should also be noted that for the purposes of this review building automation
systems shall be viewed as managing the following: renewable energy, distributed generation, energy
storage, HVAC, lighting, demand response, and load shedding. Also it should also be pointed out that
existing literature focuses on a few discrete categories such as the application of Machine Learning to:
Sequence of Operation and Equipment Scheduling Optimization, Occupied Zone Setpoints and
Scheduling, Occupancy Prediction, Thermal Comfort Prediction, Energy Storage Scheduling, and and
Renewable Energy Forecasting. A review of the current literature in the aforementioned categories, in no
particular order, is presented in the following text.
A fundamental problem that has plagued facilities for decades is the general lack of feedback by
occupants to a facility’s building automation system. In effect the BAS relies solely on temperature
setpoints that exist within a zone thermostat [9]. The issue with this is that it ignores the thermal comfort
perception of the occupant. Instead systems are designed to provide an environment that is statistically
supposed to yield an environment with an expected degree of favorability [9]. This is problematic in that
occupant thermal comfort is dictated by far more than dry bulb temperature, which is typically the only
actively monitored zone parameter, with some exceptions bringing CO2 or relative humidity
measurements into the control scheme [9]. To make matters worse, zone thermostats typically are often
inconveniently placed discouraging many occupants from bothering to make adjustments despite
perceived discomfort [9]. The mere act of having to get up from one’s workstation to provide input to a
thermostat is disruptive to workplace productivity, and thermostats often provide a very limited range of
adjustment, if not a placebo range, which further discourages occupants from feeling like their input
results in any meaningful change. In many cases zone thermostats are even covered by plastic
enclosures to discourage occupants from tampering with zone setpoints which certainly does not go over
well with occupants; this has historically been done by office managers to keep operating costs down [9].
Enter a new breed of solution, the workplace thermal comfort application. Instead of operating HVAC
6
systems based on the predicted mean vote (PMV) methodology, or even the more modern Machine
Learning enhanced versions of PMV (such as the one that used Bagging to produce highly accurate
voting predictions) [34], researchers and developers are turning to occupants for direct feedback instead
of trying to guess what their perception should be based on HVAC operational schemes. Even though
PMV can accurately predict an occupant’s perception, very few of the parameters associated with PMV
are actually monitored, and as a result virtually nothing is known about the comfort of the occupants. Thus
a system designed around the PMV concept cannot reliably deliver a comfortable environment with a high
degree of certainty since the status of the key parameters that dictate PMV are not fed back into the
control system resulting in an effectively open control system. The PMV method as a control scheme can
work, but it is prohibitively expensive and intrusive to do since it would require the tracking of many
variables both environmentally and of each particular occupant. Fortunately there is another better way.
Returning to the workplace application concept, much like the residential NEST device that enables
homeowners to control their HVAC via their mobile devices, workplace applications work in a similar
manner albeit with some more sophisticated features to interface with the more complex systems that
exist in buildings. Workplace applications are very new to the market with few competitors in the field [5].
The capabilities of these tools are currently limited, but they are powerful in what they can do. Presently
these systems are only compatible with VAV based HVAC systems, which is a little bit of a setback since
VAV systems are now a 40 year old technology with the only meaningful improvement being the addition
of variable speed fans [5]. Although that improvement yielded a significant energy efficiency improvement
it is by no means the modern system of choice. It is worth noting, though, that many buildings do have
this kind of system in place. With that said, workplace applications still need development to be
compatible with more energy efficient systems such as: radiant heating/cooling, chilled beams, variable
refrigerant flow, displacement ventilation, natural ventilation, demand control ventilation, and even older
systems that although less efficient may be well maintained and not justified to be replaced. An upshot of
the current workplace applications are that they utilize Machine Learning to reprogram zone setpoint
schedules [5]. Not only do they change setpoints but they also have the capability of narrowing or
widening setpoint deadbands [5]. It is, unfortunately not stated in the literature, what kind of Machine
Learning model (i.e. ANN, SVM, etc.) is used by the application to define future setpoints [5]. This is likely
left out to protect intellectual property. Despite that, the key benefit of the workplace application is that the
tool enables occupants to provide a direct form of feedback to the building automation system and that
feedback also results in an immediate response from the HVAC system, which is what every
uncomfortable occupant expects [5]. It’s hard to say explicitly, though, if the short-term (~15 minute) [5]
response provided by the HVAC system to the occupants request, as a result of the direction provided by
the currently existing application, is sufficient; occupants may actually prefer a longer term response, or
perhaps the duration of that response is too long. Regardless of the perception of the duration of the
response, in using this system, if an occupant’s comfort degrades after the short-term response expires
the occupant is forced to respond to the system again [5]. Fortunately, as one would expect, the
aggregation of occupants’ feedback through the integrated artificial intelligence of the application results
in permanent adjustments to zone setpoints with the intent of refining those setpoints to reduce the
amount of occupant input [5]. Although the permission of occupant feedback to trigger immediate HVAC
responses is great for occupant comfort it casts a limitation in how the data provided can be used. The
current systems collect votes anonymously and establishes setpoints on a zone by zone basis. Naturally,
when a zone serves multiple occupants those occupants may have different preferences which will
ultimately result in some occupants being more comfortable than others. This factor inherently
complicates how to make occupants feel empowered to be able change their indoor environment while
not discrediting the perception of other occupants in a space; this is something that workplace application
developers and university researchers are exploring [5]. First, the fact that occupant inputs are not
7
distinguishable from one another, in the current application, could be an unfavorable limitation. It could be
beneficial to track individual occupant inputs as they move about a facility, effectively allowing an HVAC
system to develop occupant comfort profiles and adjust zone setpoints in real-time to match occupancy
trends. Developers are already part of the way there since by the very nature of using a mobile device to
trigger a response from a building HVAC system the application must know where an occupant is to know
what space to condition; this is achievable by any of the following; selecting a space on an interactive
map, through a drop down menu of rooms, active or passive RFID tags, or via GPS. Current applications
appear to depend an interactive floor plan. That has a caveat though, location of occupants is only
identified when a request for a temperature change is requested, otherwise the system is unaware of
where occupants are and how many are present. By taking into account location, time of day, time of
year, temperature, airspeed, humidity, and/or CO2 concentration, the application (or BAS) could develop
individual occupant profiles that could be used to influence, in aggregate, zone setpoints in real-time. In
recognizing many of the setbacks in existing thermal comfort (workplace) applications, one group of
researchers set about trying to solve some of these challenges with existing workplace applications by
using a wholly different detection scheme. Instead of relying on any of the well established occupancy
sensing equipment available on the market to determine where and how many occupants are present in a
given space a study was conducted using passive RFID tags and occupant feedback via web-based
applications on PCs or mobile devices [8]. The application the occupants were equipped with enabled
them to provide comfort votes which triggered a response from the BAS [8]. The unique thing about this
strategy is that not only is every occupant uniquely recognized, but the logic of the system accounts for all
occupants in the room regardless of whether or not they provide votes or not [8]. The control logic
associated with this study only permits temperature changes when a majority of the occupants place
votes for a change [8]. This of course being a provision to provide comfort for as many people as possible
in a given space. This is somewhat counter to the currently marketed workplace application solution that
yields a HVAC system response when just a single vote is submitted as well as several. Although majority
rule is a seemingly superior method, it does have a flaw, at least as it is presented in the Lisbon
University study. In the event that there are many occupants are in a space and only a few provide inputs
of discomfort, the lack of inputs from the rest of the occupants would overrule the requesting parties
preferred change resulting in no change [8]. What this study does not address is the idea of an affirmation
by the non-voters agreeing to a change to increase overall occupant comfort. That is to say perhaps the
other occupants could comfortably tolerate a change that satisfies the uncomfortable occupants. If the
system gave non voters a chance to approve, deny, or ignore (effectively approving) requested changes
this strategy could potentially work even better.
Secondly, it’s worth pointing out that workplace applications currently do not influence how equipment
operates, they merely trigger a predetermined sequence of operations for a set period of time. This may
be adequate, but perhaps there’s an untapped opportunity here to refine terminal unit sequences of
operation to facilitate an optimum of energy conservation and expedient delivery of a thermally
comfortable environment.
Thirdly, there are two perceived reasons for implementing occupant driven comfort in a building. The first,
which the existing literature focuses greatly on, is delivering comfortable spaces to occupants. As was
previously stated there are numerous historical studies that identify that workplace productivity and well
being increases when employees or just occupants in general are comfortable. The second, which
appears to be an additional growing argument for this approach, is energy savings. To prove out the
energy savings that could be yielded by adopting an occupant centered environmental control strategy
some researchers performed an experiment by creating 400 occupant comfort profiles using existing data
in the RP884 database to be evaluated across four different climate zones and over two seasons
8
(summer and winter) [3]. The result of the study was a significant improvement in perceived thermal
comfort, upwards of 40%, and upwards of 5% energy savings [3]. When the inclusion of wider setpoint
deadbands were considered energy savings jumped to upwards of 20%[3]! There is something key to
recognize that this research points out; although Predicted Mean Vote (PMV) isn’t of much value when it
comes to operating a real building with occupants it is a very useful tool for designing new systems and
for performing energy modeling of systems [3]. In the absence of occupant data, or profiles, modern
adaptations of the PMV method can yield HVAC systems that have a lot of potential to provide
comfortable environments, hence why ASHRAE 55 uses it, and why it is codified in most municipalities
[3]. Unlike the all inclusive nature of the PMV model it should be noted that workplace applications know
nothing about about an occupant other than their degree of comfort (unless of course the occupant is
enabled with providing additional information that an HVAC system can leverage). Workplace applications
currently ignore factors such as age, sex, clothing level, and general state of health and fitness which
could cause system confusion since these are all significant factors that play into an occupant’s perceived
comfort level. As was previously stated, although PMV is an all inclusive approach to occupant comfort
that does account for those variables just mentioned, it doesn’t account for real people in actual spaces.
Were it that all of the parameters that dictate PMV were actively fed into a PMV based algorithm derived
from occupants feedback such a system could actually work. Bridging this disconnect could prove
beneficial, but it appears undecided if that is necessary given what workplace application are currently
able to do.
Although workplace applications are certainly an exciting and valuable development, another concept that
is actively being researched, and is of equally great importance, is occupancy forecasting. Occupancy
forecasting is an integral component of Model Predictive Control (MPC) which is a technique for
controlling building systems based on occupancy predictions, an HVAC system model, internal load
predictions, weather predictions, and other relevant predictions effectively forecasting future states of the
building system and identifying optimal control actions to accommodate those predictions and conserve
energy. In essence, this thesis presents a modern evolution of an MPC control scheme. One particular
pair of researchers dug into the specifics of MPC and identified some recents developments. The
researchers compared physics based MPC against data-driven MPC and also explore the specifics of
control theory, and optimization techniques [20]. It was later identified that physics based MPC is best at
providing generalizations while data-driven MPC yield greater accuracy and control performance [20]. It
was also identified that MPC seeks to find an optimal sequence of operations by evaluating a predictive
model of HVAC, internal loads, occupancy, and weather [20]. As part of the description of MPC the
researchers also identified the significance of occupancy prediction on MPC performance; meager
performance of 5-10% savings were seen without it and upwards of 40% savings were seen with it [20].
Part of their discussion about occupancy presented a novel idea of using mobile phone WiFi signals as
occupancy sensors, that could be paired with GPS information also derived from those devices [20]. It
was, however, recommended that this approach we combined with traditional PIR sensors for occupancy
sensing [20]. The researchers also identified that the popular optimization methods for HVAC control are:
Sequential Quadratic Programming, Direct Search, Lagrange method, Univariate Search, Conjugate
Gradient Method, Branch and Bound, & Evolutionary Methods [20]. They go on to identify that the Genetic
Algorithm, a meta-heuristic evolutionary algorithm, is capable of solving an MPC optimization problem
[20], but the researchers did not reference anyone that had tried it yet. Ultimately it was determined that
research opportunities existed in finding methods to produce reliable baseline building models,
developing more comprehensive and accurate occupancy behavior models, and in finding an acceptable
trade-off between complexity and performance in MPC control system [20]. The authors unfortunately did
not identify the benefits of IoT sensing equipment or that building control systems can exist and operate in
the cloud. Restricting control system computational capabilities to on-site computing equipment does
9
inherently make using MPC to control and optimize facilities expensive which has been one of main
detractors from mainstream adoption of MPC. When cloud computing is taken into consideration,
however, MPC should start to look more attractive and viable since the computing resource doesn’t serve
a single client and is seemingly limitless in its capacity. Although several isolated references to using
machine learning as parts of MPC, it was never recognized that machine learning could be the primary
modeling platform for MPC. Since the goal of MPC is fundamentally to learn and predict future states of
the building, and RNN machine learning models are designed to perform such functions, it only seems
like a natural fit to combine the two to yield a new form of MPC.
The driving factor for developing tools to facilitate occupancy prediction is the tendency of facilities to
condition spaces in which there are no people present; this is a surprisingly common occurrence. This
inherently results in the expenditure of unnecessary energy to maintain environmental conditions where
people might be without determining if they are actually present. Increasingly, companies and their facility
managers are requesting for a solution to help predict occupancy patterns in facilities so that all energy
systems can be used more efficiently. Not only does this present the opportunity for energy conservation,
but it also extends the service life of those environmental systems since they wouldn’t need to be active
as often. There is presently no clear best solution to facilitate this. Proposed solutions are far ranging from
the reliance on conventional Passive Infrared (PIR) and hypersonic sensors, to conventional CCTV
cameras and depth cameras tied into sophisticated analytical software, to Bluetooth and Wifi wearables,
use of mobile devices, and most recently to Wifi IoT signal interference detection [38]. Some researchers
have explored methods that minimize the amount of hardware required, reducing upfront system costs,
and instead relying heavily on Machine Learning to extract occupancy trends[37]. For example, a group of
researchers went so far as to claim upper 90 percentile forecasting accuracy by monitoring ambient room
temperature, outside air temperature, solar angle, time of day, and light energy. They extracted
occupancy detection, initially by comparing measurements against Energy Model simulated results and
later relying on Machine Learning models (SVM and RNN) to continue to detect occupants. In essence
they detected occupancy by identifying fine temperature fluctuations and correlated the size of those
fluctuations with number of occupants present in a space [37]. This technique does appear to have some
merit, but the reliance on Energy Modeling does present a problem; models are only as good as the
operators that produce them, and they are constrained by the algorithms that run them. In this instance,
which evaluated a very simple office space, occupancy prediction accuracy was quite high [37]. It’s
difficult to guarantee that would still be the case, however, in a more complex multi-room, multi-zone
commercial space. Considering how many variables that are at play that impact room temperature in the
real world it seems doubtful that an occupancy forecasting method that relies solely on temperature
measurements could truly be accurate in real world application. Perhaps this technique is viable, but I am
inclined to suspect that other academics have refuted this technique; there does not appear to have been
any further developments of this methodology which makes me think the technique has been disproven.
Something to be aware of, which is identified in much of the literature, is that the established occupancy
detection technologies work quite well at detecting the presence of a person or of people, the challenge
lies in not only detecting presence but also in determining how many are present at any given time in any
given location. Fundamentally, this breaks down to tracking a dynamic thermal load for which a HVAC
system would be expected to respond in turn to by ensuring comfortable indoor environmental conditions
are provided since that is what the information is intended to be used for. In light of this, one pair of
researchers went about developing a occupancy prediction model by using motion sensors and cameras
and then developed a modified version of a Markov Machine-Learning model that utilizes what is referred
to as “change point analysis” and “moving window training” as well as a modified random sampling
hierarchical approach and compared their performance to a different type of Markov model, the Reinhart
10
Light Switch Model, a Support Vector Machine (SVM) model and an Artificial Neural Network (ANN)
model’s predictive performance to evaluate accuracy of occupancy prediction; that’s two new methods
compared with 4 pre-existing methods [14]. The researchers evaluated data collected at 5 minute
intervals and verified occupancy with video footage captured by the cameras [14]. The researchers also
evaluated 4 different offices to ensure their results were not anomalous [14]. In addition to comparing the
various models the researchers all vetted out different prediction windows, such as 15 minute, 30 minute,
1 hour, and 24 hours [14]. The results of this experimentation were something of a mixed bag. The
modified Markov model outperformed all of the other models with ANN and SVM being the second and
third highest performers in most cases [14]. The machine learning models, however, did outperform all
others in the 24hr prediction [14]. Interestingly, all the various models saw some impact in their predictive
performance depending on the prediction time scale, with shorter-term windows yielding higher
performance [14]. The degree of decay in performance with increasing time scale, however, was different
for each model [14]. Unfortunately this experiment evaluated the older and generally less accurate Feed
Forward ANN instead of the higher performing and more contemporary versions of RNN or other modern
ANN variants. This shortfall may cause the results of this experiment to favor the modified Markov model
that was used more than is justified. Another thing this study failed to identify is what the value even is of
predicting occupancy on time scales less than 24 hours; at best the researchers suggest they did this
because that is what past MPC studies did or suggested be done in future work. The researchers do
redeem themselves a little bit on this though by identifying that future research should focus on improving
the performance of 24hr predictions [14]. One key takeaway from this study is that the machine learning
depended on the detection of occupants via motion sensors and by cameras. The cameras had their
feeds run through analytical software designed to count occupants [14]. The former technology is
inconsequential, non-invasive and prolific, the latter technology pairing presents a breach of privacy/trust
issue. For cameras to be used as an occupancy detection means, analytical software would need to be
integral to the camera with video feeds not being permitted to be transmitted outside of the device to
ensure the privacy and security of occupants; effectively disabling the ability to use the video feeds for
other purposes. The alternative, and likely less popular approach, is informing occupants that such a
device is used both for occupancy detection, and for security monitoring with the intended use of
capturing inappropriate or criminal activity for future review. In the scenario of a workplace, occupants
would likely feel like their every move was being watched by upper management all the time which could
have corrosive workplace repercussions. For these reasons the camera approach does not appear to be
winning much popularity despite its effectiveness. It would have to be assured that video feeds were not
accessible outside of the occupancy device.
Cameras aren’t the only devices presenting their share of headaches in the occupancy
detection/prediction realm. A big problem that crops up in many studies is the challenge of classifying
information from IoT sensing devices that can be used to ascertain occupancy. A group of researchers
tackled this directly by training Machine-Learning models with batch data delivery, and then dispatched
those models to accurately detect occupancy using an array of IoT sensors as inputs; the goal being
accurate determination of occupancy in real-time [11]. Although the application of Machine-Learning to
interpret the readings of sensors not conventionally viewed as occupancy detection sensors (i.e.
temperature, humidity, CO2, and light) for that purpose is novel, it may be unnecessary since established
technologies exist precisely for that purpose. The authors poignantly identify that their system’s lack of the
ability to identify number of occupants in a space, let alone their activity level, which would arguably be a
greater takeaway from this study had they achieved it, is a shortcoming of their study [11]. In light of
existing sensing equipment on the market and at the time of the study, the sensing equipment used by
these researchers was overly basic as equipment exists that can detect presence and number of
occupants with high levels of accuracy. Although multiple sources of data can validate occupancy
11
detection, considering the capabilities of more advanced IoT sensors, doing so is probably unnecessary.
It could be argued, though, that although sensing equipment exists for the purpose of detection and
occupant counting, reliance on more basic sensing equipment and use of more advanced software and
data analytics to extract occupancy information could be less expensive. Ultimately, the application of
Machine-Learning in this case, although related to Occupancy detection, could be perceived as a waste
of effort. What is valuable about what was accomplished is the identification of the capability of using
Machine-Learning to evaluate data streams from IoT devices in real-time to extract an unknown variable
which could then be used to influence building control systems. It also enables buildings to get more use
out of fewer pieces of equipment using some for multiple functions.
Transitioning now from occupancy prediction methods and MPC, a topic of growing interest in industry
and academia is the concept of energy consumption prediction. By default, knowing how much energy a
facility is going to use lends itself to being able to predict, or forecast if you will, energy costs. With a
desire to keep energy costs down and to reduce GHG emissions across the country the ability to forecast
energy consumption is of great value. Several groups of researchers that are evaluated here have
explored many facets of this development. In recognizing the importance of indoor air temperature and its
impact on energy performance of a building one group of researchers evaluated numerous techniques for
forecasting indoor environmental temperature. These techniques consisted of two linear regression
methods and two machine learning methods [19]. The result was that the machine learning methods were
superior at forecasting and despite efforts to further boost their performance with data clustering, no gains
were made [19]. The high performer in this case was a Multilayer Perceptron (MLP) ANN augmented with
a Non-Linear autoregressive technique (NARX) [19]. The data-set used to predict the indoor air
temperature included: calendar data, humidity, outdoor air temperature, zone setpoints, thermal loads of
zones, and measured indoor temperature [19]. This study unfortunately left out parameters such as: air
flow, precipitation, cloud cover, and solar radiation, amongst others, which also impact building
performance and to some extent indoor air temperature. Despite those shortcoming, taking a more direct
approach to energy forecasting, another group went about performing a study comparing a Feed Forward
ANN Machine-Learning method against the decision tree ensemble based method called Random Forest
(RF) [2]. The study relied on a dataset of 5 minute increment energy consumption data over a time frame
extending beyond 1 year [2]. Included in addition to this dataset was trending of outdoor weather
conditions which encompassed outdoor air temperature, dew point temperature, windspeed, and relative
humidity [2]. It should be noted that this study only forecasted the energy consumption of HVAC systems.
The result of the study was that marginally better prediction was generated by the ANN model compared
to RF, but the authors do note that the RF model is more simple than the ANN model and has greater
tunability [2]. Despite not being an all encompassing energy prediction solution, these researchers did
succeed in producing data-driven machine learning models that predicted energy consumption with a high
degree of accuracy (~96%) [2]. For comparison, another group approached energy prediction through a
different method relying instead on a Least-Squares Support Vector Machine (LSSVM) [12]. SVM models
have recently faded in interest due to lack of ability to compete or outperform other methods like ANN.
What this group did is swapped the conventional real-code Genetic Algorithm (RCGA) used to determine
the two defining free parameters of LSSVM, that is notoriously computationally time-consuming to use,
with a hybrid pair of a Direct Search Optimization (DSO) algorithm and RCGA; the goal was faster
computation and higher accuracy of prediction [12]. To test their hypothesis, energy and weather data
associated with a particular building were provided as the learning set [12]. Simulations were then
performed comparing three LSSVM methods against real energy consumption data [12]. The researchers
concluded their research demonstrating that the proposed hybridized SVM model had higher rates of
convergence and accuracy than the legacy LSSVM methods [12]. Despite referencing a source about
ANNs, that source failed to explicitly identify the shortcomings of ANN compared to LSSVM, or provide
12
comparable metrics to those used in the study [12]. As result of this claim, the researchers did not
evaluate an ANN as part of their study despite it being a notably competing methodology for prediction.
Regardless, the model presented could be a viable prediction tool, or it could be inadequate compared to
other models, it’s hard to say without more data. To validate the findings presented, each of the machine
learning strategies would need to be evaluated using comparable metrics. In contrast to that study,
another research venture took a more comprehensive strategy using weather data, schedule variables,
and machine learning to predict year-ahead energy consumption at 1 hr time steps [23]. The results were
compared against energy modeling simulations to establish overall performance against a baseline [23].
This inherently has a fundamental flaw; weather is not typically accurately forecastable outside of a two
week window, sometimes not even outside of a few days. That said, the net effect of weather over a
years time in an area is usually an aggregated normal (hence climate zones), so perhaps on the whole
this technique isn’t too far off, but viewed in relatively small time steps (i.e hours,days, weeks) large
inaccuracies are likely to be present. For added clarification Energy Modeling relies on the use of
historical weather data files provided by weather stations to estimate the influence of weather induced
dynamics on building energy consumption. For reasons previously mentioned and identified by the
authors of this particular study, reliance on Energy Modeling, as opposed to sheer operational data is
bound to be less accurate simply due to limitations in capturing stochastic behaviors, emergent system
characteristics, expertise of the model constructor, and reliance on physics algorithms that cannot capture
all physical dynamics within a building. Energy Models of buildings are tremendously complex and by
default create innumerable opportunities to miss important parameters that impact building performance,
not to mention the sheer time commitment required to produce an energy model which makes reliance on
Energy Models arguably impractical approach. All said, reliance on Energy Modeling, although convenient
for academics, is largely impractical in real world application for energy prediction. Energy Modeling is
really best suited to evaluating pros/cons of Energy Efficiency Measures and Energy Conservation
Measures; that is to say it’s a great tool for relativistic comparisons but not for accurate predictions. Thus
comparing the accuracy and benefits of prediction using data-driven machine learning vs. energy
modeling has limited validity. Comparison against metered energy usage of a building would have been a
more appropriate comparison for validating the machine learning approach presented by these
researchers. I am inclined presume that this was not done because data was either unavailable, not
pursued, or simply because the researchers were familiar with Energy Modeling and wanted to explore
other possible uses for it. In addition to this, a fundamental concept that was also not considered by the
study was what timeframe of prediction would be of most value in the industry. The answer to this is likely,
daily, weekly, monthly, or possibly quarterly predictions given the way organizational and institutional
budgets typically are formulated. Likewise the researchers failed to recognize what organizations would
conceivably use predicted energy consumption and cost data for. If that had been identified then
establishing meaningful timeframes and time-steps for prediction could have been determined. Despite
those shortfalls the researchers did evaluate a couple of statistical approaches, namely, multi-variable
linear regression, and ridge regression against two variations of ANN ultimately identifying ANN paired
with NARX to be the best performer in their experiment [23]. Bringing MPC back into the picture, this
time with a focus toward energy consumption prediction, it was discovered that a team of researchers
developed an MPC system that used Machine-Learning to model facility energy usage. The result of their
deployed MPC system was a 95% accurate prediction model [33]. A part of the MPC was designed to
determine operational parameters to conserve energy, referred to by the researchers as a “business
engine” to optimize system operations [33]. Demonstration of energy conservation through use of this
feature was not explicitly defined unfortunately with focus of the study instead being on the prediction
engine [33]. The researchers, do however proclaim the “business engine” and online evaluator features of
their MPC system to be effective, yet the plots provided of predicted versus actual performance show
higher actual power demand and energy usage which would seem to conflict with the conclusions drawn
13
by the researchers [33]. The machine learning method used was SVM and no other Machine-Learning
methodologies were tested [33]. Ultimately the value of this study was the demonstration of
Machine-Learning applied to MPC. Transitioning again to other approaches for energy use prediction,
another approach to energy prediction has been the use of the Gradient Boosting Machine Learning
Algorithm, a type of decision-tree Machine-Learning model [28]. This approach was compared against the
,at the time, current industry best practices of Piecewise Linear Regression and Random Forest [28]. With
accuracies only reaching into the mid 80 percentile range [28], though, other methods have proved to be
better options. That said, this method did yield relatively quick convergence on its learning with less than
a thousand iterations required [28]. The value of this study is the demonstration that these approaches
are not optimal and without further modifications demonstrate that these model tactics should not be
used for energy forecasting in light of other methods that have yielded far superior accuracy of prediction.
Despite those underwhelming results, as a demonstration of the capabilities of ANN Machine-Learning
and hybridized variants of it, other researchers compared a Feed Forward Artificial Neural Network with
an Adaptive Neuro Fuzzy Interface which is a blend of ANN and Fuzzy Logic. Surprisingly, both of these
models yielded very nearly the same level of accuracy of about 97% [10]. This would seem to suggest
that hybridization neither helped nor hurt the models accuracy. It should be noted that this study was
over-simplified in that it only predicted energy usage of chillers for three facilities and not the energy use
of all loads in the facilities [10]. Also, the Machine-Learning models only predicted energy use, they did
not provide any control or optimization insights (although this seems true of most studies exploring
applications of machine-learning to energy use prediction) [10]. Although energy use prediction is of great
value, identification of inefficient operation is also valuable in facilities; current systems are only able to
recognize when equipment malfunctions.
Taking a slightly different approach, instead of merely forecasting energy usage, one team of researchers
sought to explore the use of Machine-Learning techniques to identify inefficient operations of building
systems as they occurred by establishing a normalized model through data collection and comparing it
with real-time data streams to identify inefficient operations [26]. Although their results were successful,
the nature of the experiment was somewhat rigged and oversimplified. For one, this study reviewed
inefficient operation of only a single piece of equipment [26]. Buildings, especially those with greater than
10,000 gross square foot of occupiable space often have many pieces of equipment. The expectation in
those cases would be detection of inefficient operation at the zone level, and at the system level. This
inherently means monitoring far more data and distinguishing all of the various pieces of equipment from
one another would be necessary. Furthermore, the researchers in this case, established a canned listed
of criteria to look for as signs of inefficiency; this method requires expert input to establish “rules” or
“signs” of inefficiency [26]. In large facilities, the provision of a significant number of rules is likely
impractical. An alternative approach that would advance this research is use of machine learning to both
identify, or develop, those inefficient “rules” for a given building, and to use machine learning to recognize
future occurrences and then to provide control logic, also driven by machine learning, to correct those
inefficient operations. This could be seen as an optimization algorithm. This brings us to a study
performed on HVAC optimization in facilities that takes into consideration indoor air temperature
expectations and indoor air quality. In the quest for minimized energy usage, indoor air quality (IAQ) is
often the first thing to suffer. The study in question presents a method that effectively provides both at an
optimum in real-time. That is not to say the method provides the lowest energy consumption, but it does
provide the lowest without compromising IAQ (which in this case is determined based on CO2 levels)
[35]. To do this a mathematical framework known by the name Lyapunov Optimization Techniques (LOT)
was used [35]. This method is not dependent on nor does it require knowledge of future values of key
variables, it’s responds only to live data streams [35]. Effectively the researchers found it either infeasible,
beyond of the scope of their study, or not worth the effort to predict future values of key variables; it is
14
also not entirely clear why they chose to pursue LOT amongst other options. On the surface it would
appear that what these researchers proposed is no different than a VAV control scheme that is dictated
by thermostats and CO2 sensors, but the optimization piece enabled by the LOT approach enabled the
ability to dynamically change the zone setpoints in response to changes in the cost of electricity; hence
this scheme is most relevant where real-time pricing is in effect [35]. Although LOT was largely effective
at finding optimums in real-time, the researchers did conclude use of machine-learning techniques as
opposed to LOT would yield further improvement since it could learn the cost of thermal discomfort of
occupants (somehow, the authors did elaborate on how that could be ascertained), provide control of
more complex HVAC systems, and be able to leverage the value of forecasting [35]. Although the
introduction of dynamic zone temperature setpoint is novel, the lack of feedback from occupants presents
some restriction in the effectiveness of this approach. Energy savings made be had, but occupant
productivity (depending on the building use) may suffer without the system being aware of it. With salaries
often being one of the larger costs of an organization, typically larger than energy costs, sacrificing
comfort and by association productivity may be an unwise tradeoff to yield savings on utility bills. With
respect to the VAV operations the authors are right in that the main factors at play are fan speed and
cooling coil usage but the optimization scheme presented seems shortsighted and overly restrictive. A
better optimization scheme would focus on the responsiveness of zone sensors to inputs from the VAV
system and coordinating those inputs with changes in utility costs whilst also still providing some bounds
on zone temperature and CO2 concentration fluctuation. In essence they’ve followed an evolved
thermostat operational scheme rather than a predictive pseudo timing based approach. Furthermore,
manipulation of AHU operations including air mixtures could have been included to yield greater cost
savings, but perhaps this added layer of complexity is what they were referring to in looking to machine
learning for future work.
Returning, once again to the concept of model predictive control, researchers in another study evaluated
the use of a deep reinforcement machine learning by using the Markov Decision Process to replace
traditional rule and model based strategies [31]. Although energy models were used for initial training of
the model, operational data was used to continue to fine-tune the model [31]. This approach has merit
since reliance on real operational data would resolve inaccuracies imposed by a physics driven energy
model that does not account for stochastic phenomena and other variables not accounted in an energy
model. The control strategy proposed maintains room temperatures within acceptable ranges by instituted
two different control strategies. One method provides binary control, where-as the other a multi-speed
airflow approach to control [31]. In light of many systems now in use in facilities that rely on variable
speed capable equipment, this approach seems short-sighted, although relevant to older systems.
Changes to equipment operations were made in association with changes in utility pricing, outdoor
weather, and indoor temperature fluctuations, but did not take into account occupancy trends or indoor air
quality [31]. Although the overarching model used here was the Markov Decision Process, it should be
noted than an integral part of this model was use of an back propagation artificial neural network to
determine “Q values” used for learning and for control decisions [31]. This is a good starting point
towards data-driven HVAC control scheme, but the fact that this is only an HVAC model limits it’s value as
a cost saver or predictor. An integrated system that includes on-site energy storage, renewables,
net-metering, plug loads, lighting, demand-response, load shedding, setpoint adjustments, sequences of
operation, and a more comprehensive approach to maintaining an acceptable level of indoor air quality as
part of its control scheme would be far superior.
Although mentioned previously as a historical, or traditional modeling technique, research continues on
the refinement of regression based machine learning methodologies. A particular research group
explored the use of a Random Forest regression technique to provide predictive energy efficient control of
15
HVAC systems [17]. A nice feature of this technique was the inclusion of weather forecasting data and
energy pricing data [17]. Another benefit of this study was the exercise of the proposed control logic on a
real building. Most of the research in this realm relies on the simulation of performance rather than actual
performance. In this study use of the Random Forest regression technique yielded savings of 30-40%
over the pre-existing control scenario [17]. Unfortunately, it is not possible to know what the results would
be without the machine learning logic for a given period of time without evaluating duplicate buildings in
roughly the same location, or through the use of artificial simulation, or without historical data on days of
similar weather. Although neural networks seem to be the going trend, this study points out that other
machine learning methods can also yield respectable results [17].
With that said, another study evaluated the performance of a hybridized method using an evolutionary
algorithm and a data driven model to improve convergence and accuracy in energy usage prediction [15].
To do this a teaching learning based optimization evolutionary algorithm is used in place of the more
conventional genetic algorithm (GA) in conjunction with an ANN for the data-driven model [15]. The
researchers compared the performance of this arrangement with other common data driven approaches
and demonstrated significantly faster convergence times [15]. They also demonstrated an improvement in
prediction accuracy but this improvement was very minor being mere a fraction of a percent [15]. One
thing these researchers did not include was a control scheme; the method proposed was a merely a tool
to accurately predict electricity consumption. Presumably this could be easily expanded to include other
energy forms but it is unknown if this method could be effectively applied to optimize energy usage in a
building. Interestingly, the researchers identified that for effective building energy forecasting five types of
data need to be gathered: meteorological data, calendar data, occupancy data, historical energy data,
and any other data sets relevant to a particular building [15]. Although they recognized that, one data set
that was used included only meteorological data and energy data and the other relied on occupancy data
deduced from facility scheduling, a local weather station, and building level electrical meter readings [15].
Arguably, both of these datasets do not comply with what had been recommended, but with accuracies of
prediction at 97+% [15] perhaps these shortfalls are acceptable. On that note, I shall transition to
discussing the prediction of performance from renewable energies.
In the built environment the most commonly pursued approach to leverage renewable energy is through
the deployment of photovoltaic arrays. Although wind power does come up, it far less often pursued than
solar. As a result the literature review was focused on methods studied for predicting power generation of
only photovoltaic systems. To some extent, however, those methods presented can be translated in a
manner to forecast wind power generation as well. With that said, one team explored the use of an
Extreme Learning Machine technique that integrated the use of Particle Swarm Optimization (PSO) [7].
The results of this method were then compared with other machine learning methods. Ultimately it
appeared that the ELM method was a higher performer but it was also determined that hybridized
versions of data driven machine learning methods and physical methods yielded the highest performance
[7]. Due to the very technical nature of the article, it was difficult to ascertain any limitations or shortfalls of
the study, and hence none were identified, but that doesn’t mean that there are not any. Despite that,
another research group took the approach of defining weather classification categories and using
machine learning to extract solar irradiance from weather forecast data which could then be used to
forecast power generation from a photovoltaic array [30]. This study focused on on prediction of
generation at 24hr and shorter term prediction time ranges [30]. In this case use of an SVM model was
implemented and prediction accuracy up to 92% was attained [30]. The big take away was that the finer
the weather classification could be done, the more accurate the predictions would be since weather is
constantly in flux and can run through a multitude of classes within a 24hr period [30]. The more discrete
those differences can be made the more accurate the predictions can be. Unfortunately relying on
16
classification to extract solar irradiance has its limitations since weather exists as a continuum of great
variability and does not fit into classification categories easily without compromise. Within this study it is
difficult to say whether trying a method that extracts solar irradiance from weather data directly using
machine learning as opposed to using classification to reduce the noise of the data set would yield better
or worse performance. Not to mention what benefits could be had in exercising a different machine
learning approach since only one was explored. Moving on, another study performed by the Swiss,
although not explicitly a study focused on prediction of power generation from a single PV system,
evaluated the rooftop PV performance potential of structures across the entire country of Switzerland
using machine learning,GIS, weather data, and known physical parameters and requirements for optimal
performance of PV systems (i.e. module tilt) [4]. This appears to have been successful, although given
the scope of the study it cannot be verified for accuracy. That said, since the researchers were successful
at generating production estimates, it is conceivably feasible to use a similar technique restricted to the
area of single facility provided information about an existing or proposed PV system to forecast
performance of a single building. Effectively this is another method of data driven performance
forecasting. The reliance on GIS does, however, does impose a potentially unnecessary computational
burden. In the context of this study it made sense to use, but for a single facility it may actually yield little
benefit; without demonstration of this method applied to a single building it’s hard to say what the result
would be. Similar to the Swiss study, another group of researchers took an approach to solar irradiance
prediction over a broad area only this method relied on an grid of nodes that were tied to particular
latitude and longitudinal locations [18]. Effectively, every node was an equivalent distance from every
other node [18]. Within the quadrant of nodes there may or may not have been weather stations [18]. The
study used SVM machine learning methodologies to predict future solar irradiance of an area using both
data from surrounding weather stations and inter nodal weather stations [18]. The study evaluated two
things, the ability to predict with a high degree of accuracy solar insolation across a broad area, and to
predict solar irradiance for locations where there were not weather stations, effectively relying on the
machine learning to interpolate weather phenomena based on surrounding phenomena [18]. This was
accomplished with great success but was unfortunately performed in place that is largely devoid of
topography and foliage, Oklahoma [18]. It is uncertain if such as method would work in a locale with more
complex geography. Similarly, it is hard to say how dependable the predictions would be for a small area.
Taking a more comprehensive modeling methodology, another study evaluated all of the most commonly
and recently presented machine learning and statistical methodologies used for solar energy prediction
and determined that ANN and its many variants were consistently the top performers [29]. Around the
same time, another study specifically focused on the application of a Hidden Markov machine learning
model compared against SVM [13]. This method was not compared against the Genetic Algorithm or
Neural Network model since the author perceived them as being flawed and unable to address the unique
problems of solving for solar irradiance due to the dependence on physical state and motion of the
atmosphere and a need for determining which variables are the most important contributors to future solar
irradiance [13]. Given the deluge of other studies that do evaluate Neural Networks, and yield high
performance, I’m inclined to disagree and believe ANN should have been included for comparison
purposes. The result of this study was demonstration of relatively high ~92% accuracy prediction on a 5
min time frame with accuracy dropping with increasing time frame to ~62% 30 minutes out [13]. Thus the
proposed method would really only be valuable on a very short-term forecast, which is arguably of little
value. Interestingly, the Hidden Markov model had more consistent performance, but the SVM model had
similar if not slightly better performance, but was also more unreliable [13]. The last study explored for
renewable energy performance prediction takes a more comprehensive approach. In this case both
classification schemes and prediction methods were explored [22]. It was identified that in terms of
publication quantity, recent history has been dominated by research in the areas of Big Data and Machine
Learning and more specifically on Artificial Neural Networks, Support Vector Machines, and Decision
17
Trees [22]. The article goes on to describe that classification methods are defined by the type of learning
used, the type of reasoning used, the type of task executed, the type of learning process, and by other
criteria. From the historical references provided it was apparent the preferred methods for Solar power
prediction are ANN, SVM, ANFIS, and PNN [22]. Unfortunately, the references only account for research
up to 2014, so there are certainly many not listed that occurred in the last few years. The article goes on
to identify classification issues that crop up for the following circumstances: wind speed/power prediction,
fault diagnosis, power quality disturbance detection, and appliance/load monitoring [22]. Unfortunately,
the advice provided is largely tailored towards wind power. There is, however, advice specifically provided
for solar power. The general recommendation was use of ANN or SVM to evaluate meteorological data
and satellite imagery, with a focus on cloud classification and prediction [22]. Both have demonstrated
high levels of prediction accuracy [22]. From here we shall transition to energy storage.
Energy Storage is a concept of growing interest amongst organization across the country. In light of
time-of-use and real-time-pricing (RTP) utility tariffs or simply high penalties for demand ratchets there is
value in developing control schemes that optimize the use of distributed generation options to keep costs
down. There are many flavors of energy storage, but the articles referenced here are specific to electrical
energy storage and ice storage; two of the more common options. With minor adjustments combined heat
and power and fuel cells (other forms of distributed generation) could likely also follow the same
guidelines. The first study we’ll look at hear is not explicitly about energy storage, but does explore
management of thermal loads for a facility run under an RTP scheme [1]. The researchers of this study
defined a simple rule-based algorithm to adjust HVAC operations in response to RTP fluctuations to
minimize cost [1]. Arguably a similar approach could be applied to the use of thermal energy storage. By
making adjustments in how thermal storage is used operational costs can be reduced. That said the
method proposed induces broader fluctuations in indoor temperatures and effectively sacrifices indoor
comfort and air quality to minimize energy costs as a result of a RTP scheme [1]; there are better
methodologies as will be shown. The first of these alternate options explores strictly the use of energy
storage in the presence of a RTP scheme without the added consideration of on-site renewables [27]. In
this scheme a relatively simple algorithm is setup to enable a user to buy as much energy as possible
when it’s inexpensive and later use it when it’s more expensive [27]. Since prices are expected to
fluctuate throughout the day, but are forecasted a day in advance the algorithm plans when and how
much to store or dispatch and when [27]. In the study researchers have also arranged the energy storage
as effectively a third party business able to buy and sell energy/power to numerous clients [27]. All that
feature is not particularly to to this thesis, many features of the overall methodology are applicable to a
single facility or a campus of facilities. The applied methodology is mathematically rigorous and
presumably sound and does result in cost savings, but as will be explained more sophisticated
methodologies exist to really optimize use of energy storage. Another study takes this a step farther by
presenting a method to determine both the optimal size for an energy storage system and determine the
daily, and hourly use of the system; an optimized method of leveraging energy storage in a RTP scheme
[24]. The method presented yields a solution with the lowest net present cost and levelized cost of
electricity [24]. The problems lie in the execution of this approach. The control strategy is appropriate and
effective but the simulation methodology for annual energy flow through the storage system leaves
something to be desired. It also is not explained how future demands are determined nor how system
actions in response to changes in cost impact the simulations. Ultimately the simulation methodology
appears oversimplified. Similarly, the sizing strategy is vaguely described and seems, again,
oversimplified. Otherwise the rest of the methodology is sound. Ultimately the worst part about the
presented method is that its not automatic. Everything is computed in a very manual fashion [24]. For all
these reasons it is doubtful that the proposed method is truly effective or practical. In contrast to this
method another group of researchers explored the use of a finite horizon optimization algorithm to
18
compete against the likes of a Markov decision process and Lyapunov optimization [6]. The result is a
mathematically more simple approach to optimizing the dispatching and storage of energy in an energy
storage system (ESS) following both load projections and RTP projections; 10-13% savings were
estimated [6]. The researchers even went so far as to provide modified algorithms that could account for
contributions from renewable energy generation [6]. One key problem in both this and the previously
mentioned study is the prediction of future loads. Some facilities are, in fact, very regular, while others can
see a fair amount of fluctuation in their day to day loads. Ultimately there are many factors driving what
those loads will be. To really map out future loads and optimize the use of energy storage a more rigorous
load prediction method is required. Taking these shortfalls into consideration researchers have also
explored using a Markov decision process to schedule utilization of energy storage and use a
combination of Approximate Dynamic Programming and a Deep Recurrent Neural Network (RNN) for
forecasting facility loads, and controlling the dispatching of energy to loads and the grid [36]. Predictions
from the Neural Network are used in the Markov model to schedule the energy storage [36]. Optimization
of scheduling is also managed by an RNN [36]. Although the Markov method may or may not be the best
optimization algorithm, as other studies have described, the load prediction method is by far the most
rigorous yielding high accuracy predictions [36]. This study does evaluate a high complexity system with
multiple energy sources, grid connected power, energy storage, and real-time pricing [36]. Because it
looks at the management of a Micro-Grid it may be more extensive in the macro energy management
than this thesis seeks to evaluate, but nonetheless many features of it are relevant [36]. The end result of
this study was the demonstration of upwards of 10% cost savings [36]. Although electrical energy storage
is more flexible in how it can be deployed, many facilities are privy to large cooling loads that depend on
high electrical demand chillers. In some regions of the country electrical energy storage for one reason or
another may not be a cost effective option but thermal energy storage may be. At any rate, chilled water
or ice storage is an indirect form of electrical energy storage, hence there is some relevance of thermal
energy storage studies to those focused largely on batteries. Ice storage is often pursued where high
cooling loads exist and a time of use pricing scheme is in place. One proposed method for effectively
using ice storage has been the use of Sequential Quadratic Programming [16]. This method relied on
machine learning algorithm load predictions in addition to time-of-use rate scheduling [16]. The result was
the ability to reduce costs by ~10% both per day and per month [16]. For comparison the method was
evaluated against traditional heuristic control strategies and proved to be more effective [16]. Taking
complexity up a notch another team of researchers evaluated the optimization of utilization and cost
savings of ice storage in response to a RTP scheme [21]. This was done using air temperature
forecasting using an RNN model and then using dual function load shifting optimization algorithms [21].
One algorithm optimized tank usage while the other optimized cost [21]. In day ahead RTP forecasting
the proposed system yielded up 16% savings [21]. Savings were grossly impacted by what amount of
cooling load could be offset, with partial load offsetting being more cost effective [21]. When RTP was
seen as an unknown, however, all savings virtually disappeared [21]. This is largely due to the fact that
the system had no reliable prediction of cost to rely on [21]. Any proposed optimization scheme would be
subject to failure wherever the real costing misaligned with projections [21]. Fortunately most utilities that
are adopting RTP schemes are providing day ahead forecasting so systems like the one proposed here
are in fact effective to have in place.
In light of the ever increasing expectations of a building control system to not only control a facility’s
systems in an efficient manner but to also forecast usage costs, and effectively deploy alternative energy
sources there is need for defining a framework by which the whole system functions and determining
what services it should provide. A team a researchers went about creating a framework that could
ascertain the overall intelligence of a building management system by taking multiple criteria and features
into consideration [32]. The framework and its rules are very well defined but the framework ultimately
19
falls shorts on the addition of things like energy storage, renewable energy use, strategic use of grid
power and net metering, as well the inclusion of cost forecasting, energy use forecasting, or system
optimization methodologies [32]. References to features that machine-learning can provide were also
limited [32]. Ultimately though, the researchers were not after the establishment of features and defining
the structure of building management systems, what they were after was a quantifiable method to
compare building automation systems and their overall intelligence against one another [32]. Although it
has no impact on this thesis, it should be noted that the weights the authors’ applied to the various
features identified inherently makes the objective tool flawed since the weights are ultimately subjective
valuations of those features biased towards a favored goal or expectation that not all evaluators may
agree with. The convenience of a tool is understandable given the complexity of building automation
systems which make them difficult to compare, but this may not be the best method. Rather identification
of standard features and unique features and comparing them in a manner to how consumers compare
consumer goods for purchase may be more appropriate since the bias is removed.
20
The Big Picture
To fully understand why it’s worth investigating the use high performance computing in building control
systems it’s appropriate to acknowledge macro level energy flows. The American consumption of energy,
and global consumption for that matter, breaks down into a four key categories: Industry, Transportation,
Commercial, and Residential (figure 1). Shockingly, Buildings (which include both Residential and
Commercial) account for a whooping 38% of the pie, 38%!
Figure 1
When you consider that a sizable amount of the Residential sector energy consumption is tied to large
facilities (13%)(figure 2), which have centralized systems, they too are effectively what we conventionally
think of as buildings.
21
Figure 2
As a result, in combining commercial buildings with those Multi-Family Residential buildings, upwards of
33% of macro energy flows are tied up in what we colloquially refer to as buildings. It turns out that,
generally speaking, the gross majority of that 33% is controlled by advanced building automation systems
of some sort. These are systems that have the potential to optimize and finely control major energy end
uses within the buildings. Surprisingly though, they are underutilized and underleveraged simply because
they are not autonomous enough, cumbersome to engage with, poorly programmed, and often
overlooked and forgotten about. Let’s take a step back though, what are we saying when we say that 33%
of all energy used is in buildings? Buildings consume energy predominantly through the forms of
Electricity, Natural Gas, Diesel, and Heat. Two of those are secondary energy sources (electricity and
heat) which are derived from centralized power plants and large, fuel fired boilers. What this breaks down
to is whether it’s happening at the facility or upstream the gross majority of energy used is derived from
combustion whether from fossil fuels, biomass, or municipal waste. The below figure shows the
distribution electrical power across the country by type and size (figure 3). As can be seen, the makeup of
power generation varies significantly across the country, and that inherently impacts the associated
emissions of buildings depending where they are in the country.
22
Figure 3
Combustion is one of the single largest producers of greenhouse gases and air pollutants. Fortunately, in
the case of heat production nearly all of the energy entrained in the fuel can be converted to heat.
Unfortunately, the same cannot said for electricity production, wherein most power plants operate at
around 35% efficiency, with newer state-of-the-art natural gas power plants operating at upwards of 50%
efficiency. In either case a lot of energy is lost, and for all the fuel expelled to produce that electricity
emissions are generated. To capture this visually, the below charts depict macro-level energy flows, and
CO2 emissions in the United States (figure 4) and globally (figure 5).
23
Figure 4
Figure 5
Plainly stated, if reductions in energy use in buildings occurs, greenhouse gas generation (and air
pollution) will decline, significantly! We can reduce our dependence on fossil fuels through energy
24
conservation, efficiency, or use of alternative energy source options and as a result greenhouse gas
production will decline. It’s not even just GHGs, it all the other pollutants associated with combustion as
well that will decline if we use less energy derived from fossil fuels or substitute those for non-polluting
energy sources. Research has suggested that use of artificial intelligence in buildings has the potential to
reduce building energy consumption by 15-30% in the USA. In less developed countries that value is
even higher. So back to big picture thinking. For simplicity, let’s say the USA uses 100 units of energy,
and 33 of that is from buildings. Use of artificial intelligence in buildings, conservatively, would bring that
33 down to 28. As a result, what could be accomplished is an overall reduction in energy consumption
(before even considering alternative energy options) of 5% (10% if 30% reductions were achieved) just by
better managing our facilities. That may not sound like much, but consider how many buildings there are
compared to number of vehicles, or the number of and complexity of our industrial processes (i.e.
manufacturing and agricultural). That 5% is far more accessible than any other gross improvement that
could be had from any of the other categories of major energy consumers.
Climate Zones
No discussion about energy usage in buildings would be complete without recognizing the diverse array
of climate zones in which our buildings exist around the world. These climate zones describe regions with
uniquely identifiable weather patterns that can be anticipated in any given location around the world.
These climate zones also inherently dictate many of the influencing parameters that drive what buildings
have to do to maintain comfortable indoor environments. As is clearly displayed by the charts below
(figures 6 & 7), there are numerous climate zones around the globe and likewise within the United States.
As a result, the technologies needed to maintain comfortable indoor environments vary across the many
different climate zones, and likewise the extremity of the climate zone and the type of systems used by
facilities in those regions will largely drive how much energy buildings need to use to maintain comfortable
indoor environments.
Figure 6. IPCC (Intergovernmental Panel on Climate Change) 2012 Climate Zones
25
Figure 7. IECC (International Energy Conservation Code) Climate Zone Map of the United States
And for relation to the thermal comfort profiles mentioned in a later section, the below chart (figure 8)
depicts where many of the climate zones’ average temperature and humidity profiles’ live on a
psychrometric chart.
Figure 8. ISA (International Society of Automation), “Getting a Grip on Humidity” by James Tennermann
26
Where does the energy go and how much is used?
Buildings, as is shown by the charts below (figures 9 & 10), have a diverse range of uses for energy. It
turns out that the gross majority of these are controllable, such as: space heating, lighting, refrigeration,
ventilation, cooling, and water heating.
Figure 9
27
Figure 10
To effectively manage energy use in a building it is imperative that it is understood where energy in used
within a building and in what quantities. With that information in hand, users can focus their efforts on the
end uses that are of greatest significance to yield the greatest overall improvement. To monitor energy
consumption, energy sources have to either be metered directly, or consumption extrapolated based on
other datastreams and equipment performance assumptions. Ideally, each end use is monitored
separately (through submetering), but oftentimes there are only building level (or tenant level) meters for
electricity, natural gas, hot water (district heating), chilled water (district cooling), domestic water, and/or
sanitary sewer. This notably complicates matters since energy uses are represented in aggregate, and to
break them apart would require third-party energy auditing which is not something that can be done
regularly. As one may expect, submetering of end uses is often only found in large facilities where the
cost to have that added capability is deemed worthwhile. Fortunately, the cost of submeters has come
down significantly in recent years, especially with respect to electricity metering due to “clamp-on” style
inductive meters entering the market. In the past, meters had to be installed integral to electrical system
which made them far more expensive. Natural Gas and Water meters are still integral inline units, and
thus expensive to install.
The technologies presented in the rest of this thesis grants the ability to finely control most of the
end-uses identified here. This will be shown to be achieved through direct control of equipment or through
curtailment of phantom plug-loads based on occupancy patterns.
Contemporary Building Automation Systems
It should be stated up-front that there is a spectrum of building automation systems on the market and
currently in use in buildings. These range from off-the-shelf systems functioning similar to a single-family
residence, to customized, internet based systems capable of providing analytics, diagnostics, trending,
visualizations, and customizable programming. Furthermore, these systems exist as either isolated to to
Heating Ventilation and Air Conditioning (HVAC) control or integrated with other systems such as lighting,
security, and life safety. With regards to HVAC, control systems exist as either pneumatically controlled
systems (a legacy technology that still exists in older well kept buildings) or direct-digital-control (DDC)
systems (which is representative of all modern systems); either form can be managed by any type of
28
building automation system. It is conventional that every major piece of equipment (heating plant, cooling
plant, etc.) or zone terminal (VAV box, Fan Terminal Unit, VRV diffuser, radiant panel, chilled beam, etc.)
has its own controller. All of these controllers contain their own programming on board, but are networked
to a centralized hub which can be interfaced with via a programming interface (keypad, or combination
function thermostat), or a graphical user interface (GUI). Many newer systems also offer internet based
access to the building automation system (BAS) via a GUI, but it should be noted that all computation
efforts regarding building controls resides within the controllers in the building. The problem with BAS
whether new or old systems is three fold. They do not capture feedback from occupants regarding their
comfort, they do not track or respond to changes in occupancy, and they operate based on static
programming that at best functions based on preset scheduling. The result is buildings that try to provide
comfortable indoor environments without any knowledge of the perception of their occupants and as
result oftentimes fail to meet expectations. Similarly, buildings will condition spaces regardless of whether
or not there are people are present since they function based on a schedule and not in response to
occupancy tracking. There is one exception to this, and that is use of demand-control ventilation which
monitors CO2 levels and modulates ventilation in turn (but that approach does not make adjustments with
respect to temperature or humidity). The fundamental flaws of contemporary BAS is the reliance on
reactive, corrective control and its general rigidity to functioning under just one scenario. Programming in
BAS is typically done in accordance with Sequences of Operations written by the engineers of record that
designed the systems being controlled. Oftentimes those sequences are canned concepts that address
extreme scenarios, but otherwise do not provide methods of optimization when between extremes. It is
true that there are parameters in the programming that are inflexible since they impact life safety, but it
turns out that many parameters in controls programming of buildings can actually be flexible. Another
problem with the programming in most buildings is that is set when the system is initially installed and
then handed over to operational staff that either does not understand the programming, or that does not
have the time to comb through the programming to optimize its functionality, and this is largely due to the
fact that those persons are often more focused with keeping with preventative and active maintenance of
equipment in a building. For those professionals, it would arguably be more valuable if the BAS would
learn from and adjust to its equipment to reduce energy use while also providing useful guidance on
where to focus physical efforts (some a BAS will never be able to do) to keep building systems running as
best as possible. In other words the BAS should control the building autonomously, and facility managers
should be enabled to troubleshoot problems and perform repairs based on feedback from the BAS.
The Importance of Data
As the old saying goes, “one cannot manage what it cannot measure”. To manage energy use in
buildings we must rely on a diverse array of datastreams to make intelligent decisions about how to
control our equipment. Data in buildings takes many forms and is not limited to but certainly takes the
forms of: temperatures, airflows, water flow rates, pressures, valve positions, damper positions, CO2
levels, occupancy detection, occupancy counts, lighting levels, power demand, electric metering, gas
metering, water metering, and equipment operational statuses. Typically the datastreams from these
sensors are either voltage or amperage readings which are then coordinated into their respective
measurements within the controllers they are connected too. This means that any evaluation of the data
would be have to done through the controller data feed rather than through the sensor itself. The
invention of Internet-of-Things (IoT), does change this slightly. Sensors that are IoT enabled provide a
datastream that is both uniquely identifiable, directly retrievable on the internet, and directly
representative of the signal that is being monitored (i.e. temperature is shown as temperature). In order
for a BAS to control systems within a building, its needs data; a lot of data. Currently much of this data is
29
processed within the controllers that make up a BAS, which is fine for the static control schemes used
presently. To transcend this control scheme to one that is dynamic, data would need to be transported to
Cloud Based servers where advanced computation could be performed to yield dynamic control
responses from building systems. To to this, data would either need to be streamed/accessed through
existing BAS equipment and/or IoT sensors that circumvent the BAS. All of the data streamed to the
Cloud would need be classified/identifiable to specific equipment/controllers to effectively use it to affect
building control schemes. Similarly, all output (control) signals from the Cloud would need to be classified
appropriately so that they yielded the desired control responses in the building. Thus all the datastreams
need to be matched up with their associated respective equipment in the building. Without this kind of
datastream control/coordination advanced control using cloud-based computing is not possible. To
contrast this with current systems, sensors feed information into controllers where they are used within
the controller to determine how programming should respond, and otherwise the data signals do not go
anywhere else unless specifically programmed to do so. Some newer systems, however, are able to
trends sensors inputs, but again this is physically facilitated at the controller, but is accessible from the
GUI which is networked to all of the controllers in the building. Since there is an communications
backbone that connects all zone and equipment controllers to a centralized point, all of the input sensors
signals’ could also just as easily be transmitted across that communications network. Likewise, that
communications network could be used to transmit output data signals back to specific controls to
facilitate a specific control response.
Recurrent Neural Networks- Machine Learning
As could no doubt be ascertained from the literature review, all of the key topics reviewed revolve around
the concept of Machine Learning. It turns out that Machine Learning takes on many different forms but not
all are created equal and each is best suited to certain applications. Fundamentally, the various models
and associated algorithms all seek to do the same thing; they review a piece or collection of data and
then respond to it in some desired fashion that has been learned. It turns out that there are many
applications for Machine Learning and that some models work better at some applications than others.
Ultimately the goal is to develop Machine Learning models that yield near unity of accuracy for their
intended use whether that’s classification, regression, prediction, or some other function. With that said,
as was made clear in the literature review, there are some recurring trends, and models of focus, that are
proving to be prevalent choices in the Building Industry; Markov, Support Vector Regression (SVR), Feed
Forward Artificial Neural Network (FFANN), and Recurrent Neural Network (RNN) models seem to be
favored for their deep learning capabilities and high levels of performance (accuracy). With that said, RNN
seems to be at the forefront of development since it is both a deep learning methodology and is tailored to
manage sequential phenomena and has yielded high performance in other applications. RNN models
function by taking in new inputs and recycling previous outputs to determine the next output as is shown
by Figure 11 & 12. This is counter to traditional feed forward ANNs which are a simple pass through of
information and do not reuse past outputs to determine new, next step, outputs. The recurrence capability
makes RNN well suited to predicting time step progression events that are dependent on a multitude of
variables. For this reason RNN Machine Learning would seem the ideal tool for forecasting energy usage
and operational costs in buildings. RNNs, and ANN in general, do not have a set number of hidden layers
or nodes, and are rather designed to have as many layers or nodes as is deemed useful for the
application. RNN models alone, however, are plagued with an issue referred to a Vanishing Gradient.
Neural Networks learn by passing inputs through the various hidden layers of the network to produce an
output. That output is then compared with a true value to produce an error function. This error function is
then corrected by an activation function to yield a gradient. The gradient is then applied by
30
backpropagation, through a process known as Gradient Descent, to the connections between the hidden
layer nodes. The problem is that the gradient is exponentially reduced as it steps through the hidden
layers of an RNN. Effectively what this means is that the earlier hidden layers learn less from the applied
gradient than the later hidden layers. As result an uncorrected RNN has a short term memory problem
which encumbers its ability to learn. To mitigate this, RNN models can be equipped with what is referred
to as a Long Short Term Memory (LSTM)(figure 13) feature or a Gated Recurrent Unit (GRU)(figure 14),
which modifies how the neural network manages information. Both of these approaches accomplish very
nearly the same thing with the GRU being slightly more simple to adopt and more efficient to use. As a
result the RNN is better able to learn by training all of the hidden layers more equally and effectively and
by default you end up with a model that can learn faster and be more accurate. As result all references to
Machine Learning from here on out shall be viewed as synonymous with RNN GRU modeling.
Figure 11. Depiction of a Recurrent Neural Network
Figure 12. Simplified Depiction of a Recurrent Neural Network (Wikipedia-Francois Deloche)
This work is licensed under the Creative Commons Attribution-ShareAlike 4.0 International License. To
view a copy of this license, visit http://creativecommons.org/licenses/by-sa/4.0/ or send a letter to Creative
Commons, PO Box 1866, Mountain View, CA 94042, USA.
31
Figure 13. Depiction of the LSTM Unit (Wikipedia-Francois Deloche)
This work is licensed under the Creative Commons Attribution-ShareAlike 4.0 International License. To
view a copy of this license, visit http://creativecommons.org/licenses/by-sa/4.0/ or send a letter to Creative
Commons, PO Box 1866, Mountain View, CA 94042, USA.
Figure 14. Depiction of the GRU Unit (Wikipedia-Francois Deloche)
This work is licensed under the Creative Commons Attribution-ShareAlike 4.0 International License. To
view a copy of this license, visit http://creativecommons.org/licenses/by-sa/4.0/ or send a letter to Creative
Commons, PO Box 1866, Mountain View, CA 94042, USA.
32
The Human Element
Fundamental to the whole discussion of this thesis are people. Buildings exist to protect people from the
elements and to maintain a comfortable indoor environment for them. Humans are so dependent on this
concept of shelter that the Environmental Protection Agency (EPA) has published that, “ Americans, on
average, spend approximately 90% of their time indoors.” By that notion, we naturally expect those places
to be comfortable and inviting. Yes it is true that in the case of buildings that serve as a place of
employment the building is also a fundamental piece of infrastructure that enables people to conduct the
services they provide for the greater community. By that notion, buildings also have to be accommodating
to the equipment that enables people to perform work, but with respect to indoor environmental quality
(IEQ), human expectations are more the limiting factor than equipment. Research has even shown that
IEQ has a distinct impact on human emotions (mood) and performance (productivity). In the workplace
this has an important impact since oftentimes the most expensive cost in any business is the payroll of
the staff. Staff that is uncomfortable, for any number of reasons, is inclined to be less productive because
they are distracted by their discomfort. Poor IEQ has a deleterious impact in many different settings, yet it
is often overlooked for the sake of energy conservation. This thesis presents an approach that balances
efficiency, conservation, and optimal IEQ. Before we get to that though it worth pointing out what all is
entailed in IEQ. IEQ is dependent on all of the following: airborne particulates, odors, CO2 levels,
humidity, airflow, mean radiant temperature, dry bulb temperature, illumination, and glare. Human comfort
is dependent all of the those things as well as: state of health, gender, age, metabolism, and clothing
level. Many of the factors of IEQ are controllable by a building, while the additional factors specific to
people are of course not, and there inlies the challenge. Buildings are expected to provide a comfortable
environment to all occupants regardless of those uncontrollable variables. Due to the many variables
involved, for locations where many people are collocated it is virtually impossible to guarantee comfort to
everyone. In a perfect world every occupant would have their own HVAC zone customized to meet their
needs; that is not the world we live in though. Such systems have been deemed too expensive and
complex to justify, so we have to make due with compromises. Despite that, however, it has been
determined that upwards of 80% of collocated people of a diverse group, using the Predicted Mean Vote
(PMV) method for system design, can be satisfied with regards to thermal comfort, and that has been
accepted as good enough industry wide (ASHRAE Standard 55). The problem is, despite good intentions
when designed, many systems fail to provide that level of performance and that is largely because their is
often no means by which occupants can provide feedback about their sensory experience when they are
uncomfortable. It has been found, however, that if the indoor environment is maintained within the thermal
bounds shown below (figure 15) most occupants will be satisfied. The chart below depicts,
psychometrically, the bounds of human thermal comfort in terms of operative temperature temperature,
which is a value representative of both air temperature and mean radiant temperature.
33
Figure 15. ASHRAE 55 (American Society of Heating Refrigeration and Air Conditioning Engineers)
Thermal Comfort Profile
34
But as was mentioned previously, temperature and humidity aren’t the only factors that influence thermal
comfort. The below chart (figure 16) shows the influence of airspeed on thermal comfort.
Figure 16. ASHRAE 55 Comparison of Airspeed and Operative Temperature and Clothing Levels
influence on perceived Thermal Comfort.
And lastly the chart below (figure 17) explains how the occupant specific factors influence the perception
of thermal comfort.
Figure 17. ASHRAE 55 Influence of Clothing Level and Metabolic Rate on Thermal Comfort
35
In light of these realizations, the following chart (figure 18) describes the general response expected from
an HVAC system whenever thermal conditions of an environment are outside of the comfort zone.
Figure 18.
Other notable standards related to IEQ, and Indoor Air Quality (IAQ) (which is a component of IEQ), are
ASHRAE 62.2 (Ventilation and Indoor Air Quality), ASHRAE/IESNA 90.1 (Illuminating Engineering
Society of North America) (which includes requirements for comfortable lighting), ASHRAE 189 (Standard
for High Performance Buildings), LEED V4 (Leadership in Energy and Environmental Design), and WELL
V2.
36
Developed Concepts Revisited and/or Improved:
Workplace Applications:
Addressing existing research and COMFY:
Workplace Applications are a very new technology with only one identified as commercially available
presently, COMFY. As was identified in the literature review researchers have explored numerous
methods to facilitate occupant engagement with a BAS and more specifically the HVAC system. Many
different tactics have been utilized to facilitate an interface for such an application such as, mobile
phones, digital wearable devices, computer workstations, and combinations of devices using RFID tags,
occupancy sensors, and cameras. In addition to the above described function the workplace application is
also intended to provide a BAS with valuable feedback that it can use to modify it’s operational scheme to
reduce energy use. In the case of COMFY, the interface of choice is the mobile device. This application,
with respect to occupant comfort, is limited only interfacing with HVAC systems managed by a VAV air
distribution system and only enable occupants to submit votes for being too hot or too cold. As was
identified earlier, there are many other controllable parameters that influence the environmental
perception of occupants. The remainder of this section will address the other forms of HVAC systems a
workplace application should be able to interface with and present a conceptual framework for a more
advanced workplace application.
The Evolved Workplace Application:
For the purpose of this thesis, I am going to elaborate of the use of mobile phones as an interface for a
workplace application and the use of various types of IoT occupancy sensors to yield optimal functionality
of the system. Below I presented a few new features and adjustments to current concept of the workplace
application.
For starters, similar to the existing product, COMFY, I propose for future development an application that
would give users the ability to proclaim votes such as: “I’m Cold”, “I’m Hot”, “It’s Too Bright”, “It’s Not
Bright Enough”, “It’s Muggy in here”, “It smells bad in here”, and “It’s too drafty in here”. Equipping
occupants with more voting options better enables the HVAC system to provide high levels of IEQ and
also integrates a connection to lighting or shading systems to again facilitate improvements in IEQ. I also
propose the addition of an approval feedback in the form a notification that says something to the effect of
“I’m fine with a small change if others are requesting one”. It doesn’t quite make sense to empower every
occupant with a guaranteed HVAC system response in response to an application input; doing so risks
putting other occupants in the same zone in discomfort which would be self-defeating. By this notion the
better approach to handling occupant feedback votes is one dependent on a majority rule concept, and
one that gives comfortable occupants an option to grant or deny a change in environmental conditions.
Use of majority rule is not a new concept, it was identified in the literature, but the addition of an approval
feedback is a new concept to help facilitate majority rule decision making. The reason for this is simple,
individuals submitting requests are those that are dissatisfied. Satisfied occupants would have no need to
submit a vote. Of course this form of management isn’t always applicable, such as in the case of private
offices, but regardless it has merit in the face of unrestricted instantaneous HVAC responses to every
submitted request in a high occupancy space.
There’s also the issue of in what manner should an HVAC system respond to an occupants input. To
shed light on this it should be noted that the predominant systems in use in most facilities these days fall
into the categories of: VAV, Fan Coil, VRF, Radiant, Active Chilled Beams, Displacement Ventilation, and
Demand Control Ventilation (which is an adaption of VAV). This variety of systems does present some
37
difficulty and limitations in making a one size fits all type of solution. Fundamentally what’s being
controlled are water flow rates, refrigerant flow rates, or air flow rates via valves, dampers, compressors,
and blowers/fans. It gets complicated when you consider that some valves are modulating while others
are two position. Similarly, fans/blowers may have binary , multi-speed, or variable speed control
capabilities. And likewise compressors (or pumps for that matter) may binary, staged, multi-speed or
variable speed. These factors greater complicate coordinating a control response. The most realistic
approach to addressing this challenge is to have a unique control sequence for each of the various types
of systems that the application could select from after detecting, through the building automation system,
what kinds of inputs/outputs are present on a zone terminal unit and coordinating a match. As for the type
of response the HVAC system should provide, in response to occupant(s) input(s), a short term 10-15
minute response with a 1-2 degree temperature change should be provided before resuming operations
per a schedule. This brings us to the real party piece of workplace applications and their tie-in to HVAC
controls. Conventionally, zones are scheduled with an Occupied Time and an Unoccupied Time, which
represent mere preset windows of time at predetermined static zone setpoints that only consider dry-bulb
temperature. The problems with this strategy is that the HVAC system will condition spaces regardless of
whether or not people are present, provide an ambient air temperature that may not be comfortable in the
presence of other variables (such as radiant heat, air flow, or humidity), and apply the same deadbands
regardless of occupancy. It turns out that if a system logs the votes of occupants and compares that
information with weather phenomena, setpoints can systematically float to match circumstances, and
likewise deadbands can narrow and widen which collectively results in optimal indoor environmental
conditions when people are present and energy conservation when they’re not. To accomplish this
functionality Machine-Learning can be used. Machine-Learning has the capability to manage a multitude
of data inputs, provide calculated responses, and forecast future conditions. By monitoring and logging
occupants votes, occupancy trends, temperature changes, and weather changes, setpoints and
deadbands can be automatically “scheduled” and adjusted as time elapses. By having a system learn its
occupant’s preferences it will, with time, need fewer corrections in the form of occupant votes to ensure
an optimal environment is provided.
This is does present a concern that particular occupants that are always hot or cold may cause a system
to respond in kind to their desires at the disdain of other occupants in the same space. To mitigate this, as
was mentioned previously, use of a majority vote system could be utilized. To do this would require two
way communication between the application on each occupants mobile device and the Building
Automation System. Herein lays one of the great challenges. If people are comfortable they’re not going
to be inclined to interface with their device. Also, to institute a majority rule control scheme the system
would need to know which occupants were present, which means that anonymity, in some respect, would
need to be sacrificed. The best way to facilitate this is via GPS and IP-Addresses; the former provides a
location, and the latter a unique identifier. Assuming occupants spend most of their time on a particular
floor of a building (and could provide that basic information upon account setup) this could conceivably
work. For privacy and security reasons the application could also hide all occupant identifiers of a
particular space from those in other spaces. Using this approach if one or several occupants submit votes
for changes, whichever type is most numerous,a time sensitive request (perhaps 1 minute) could be sent
to the other occupants in the space asking for permission to make the requested change. If the request
were ignored by some or all non-requesters the system would default to approval votes and then
determine if the gross majority (say at least 60%) wants a change in which case a short term response
would be provided, and otherwise no change would be made. Repeated accepted votes would result in
machine-learning algorithms making “permanent” changes to setpoints that would stay until repeated
votes opting for a different change were made.
Another component to this, which was mentioned previously, was the use of variable deadbands.
Machine-Learning could manipulate the control deadbands of indoor air temperature by tracking
38
occupancy data trends. For periods of time that a machine-learning algorithm finds a space to be
regularly unoccupied deadbands (as well as setpoints) could be adjusted. Similarly, for periods of time
when spaces are either occupied or highly occupied, control deadbands could be narrowed to ensure
comfortable conditions are maintained tightly. Again this is all a data-driven approach which would
complete negate the need for conventional control schedules since the internal usage patterns of a
building would be learned by the system that controls it. This lends systems to ability to schedule various
zones in completely different ways in a completely dynamic manner changing with occupancy trends as
they occur. The below diagram (figure 19) shows the systematic structure of the workplace application
and the government machine learning model associated with it.
Figure 19. Workplace Application Machine Learning Model diagram
Occupancy Forecasting:
In addition to to the improvements in energy conservation and IEQ provided by Workplace Applications,
directly tracking occupancy in a building has notable benefits. As was mentioned previously, nearly all
buildings currently function based on set schedules with occupied hours and unoccupied hours. This very
basic arrangement, however, fails to account for occupancy patterns. Plainly stated, most buildings have
a multitude of rooms, and many of those rooms are not occupied for considerable portions of any given
day, yet because we rely blanket schedules those spaces are carefully managed to provide comfortable
environmental conditions so that in the event that someone does decide to use the space it is ready. This
approach results in large amounts of wasted energy. Occupancy tracking has threefold benefits: effective
use of artificial lighting, effective management of plug loads, and efficient utilization of HVAC systems. If a
building knew, with a high degree of confidence, what occupancy patterns looked like on any given day
for all of the spaces within it, it could schedule its systems on its own to match occupancy patterns. Not
only that, but as occupancy patterns change over time, the building could adapt on its own. As a result of
this, in addition to the implementation of Workplace Applications in buildings, it is also beneficial to have
39
occupancy sensors in all zones/workstations. The type of sensors to use vary depending on the nature of
occupancy in a space. Individual workstations are good candidates for combination IR/ultrasonic sensors.
For spaces that see multi-occupancy (conference rooms) or high transiency occupancy (hallways) use of
combination detection/counting occupancy sensors is preferred. Although it is true that many facilities do
have occupancy sensors, these sensors are often only used as relay to trigger an action or set of actions,
but do not provide any sort of datastream.
Acknowledgement of Existing Occupancy Forecasting:
Although most buildings equipped with occupancy sensors utilize them in the manner previously
described, it is important to acknowledge that research and case studies have already been performed
that resulted in the successful collection of occupancy data that were then subjected to machine learning
evaluation to yield accurate occupancy forecasts. Occupancy forecasting is not an entirely new concept,
and it has already been used to influence the operations of HVAC systems. Occupancy forecasting is
however, an integral component to the system presented in this thesis so it worth identifying how these
systems work.
Where datastreams of occupancy sensors cannot be established through BAS controllers use of IoT
devices can provide the needed data stream. It is debatable what the best approach is to capturing
occupancy data, but it would appear that infrared sensors with counting abilities and hypersonic (motion)
sensors are the best and least intrusive options for capturing this data. By collecting and, time and
date-stamping occupancy information, and aggregating that data a machine learning model can
characterize the occupancy patterns of the building and produce forecasts of future occupancy. Those
forecasts can be used to influence activation of environmental systems as well as the prediction of plug
load and lighting system usage. The below diagram (figure 20) depicts the system architecture for
occupancy forecasting.
New ways to wield Occupancy Forecasting:
Although occupancy forecasting is not a new concept, what is new is utilizing this forecast to influence
other forecasts that pertain to energy use, electrical demand, energy storage utilization and operational
cost forecasting. Existing literature appears to only identify the use of occupancy forecasting in managing
the control of HVAC systems, but does not appear to apply this data to any of the other energy using
systems in a building that are also related to occupancy trends. It also appears that this form of
forecasting has not before been facilitated via cloud computing through a recurrent neural network, hence
the use of these computational methods also appears to be novel.
40
Figure 20. Occupancy Machine Learning Model diagram
Smart Lighting:
The lighting brilliance that already is...
There are a few notable trends in lighting that are pertinent to the subject matter of this thesis, namely
daylighting, and occupancy driven lighting controls, and network addressable lighting systems. Prior to
these technologies lighting was either controlled via a static schedule through binary control, or manually
through wall switches, and in many cases still is. The new technologies are a means to conserve energy
when adequate natural illumination is available or for when occupants are not present. Network
addressable lighting systems lends the flexibility of redefining lighting “scenes” as floor space utilization
changes over time and also provides a data stream that either directly, or through correlation, can be
used to provide energy consumption information related to the lighting system. The nature of Smart
Lighting means that it functions with respect to weather phenomena and occupancy patterns to conserve
as much energy as possible; Smart Lighting also tends to be LED which has higher efficacy and efficiency
that any other type of lighting. LED lighting is also capable of dimming (which also reduces energy use),
which many legacy technologies were not able to do.
Making Lighting a Forecastable Load
Paired with data associated with occupancy and weather, energy consumption from lighting systems can
conceivably be forecasted using machine-learning and included with and aggregated with forecasts
related to Plug Loads, and HVAC systems for overall building energy usage estimates. It is not suggested
or recommended that lighting systems be operated according to forecasted occupancy though. Due to
their near instantaneous response, LEDs luminaires are well suited to responding in-situ to occupancy
sensor inputs and daylighting sensor inputs, wherein they activate and modulate in the presence of
people and to the intensity of natural illumination from the outdoors. For the purposes of energy use and
cost forecasting, however, forecastable energy use from lighting systems is valuable, and that is only
feasible by tracking lighting system operations, occupancy, and weather patterns. The below diagram
(figure 21) details how forecasting of lighting system energy usage could be facilitated.
41
Figure 21. Lighting System Machine Learning Model diagram
Plug Load Forecasting:
What we already know, and what already exists
It should come as no surprise that a significant contributor to energy consumption in buildings is from
items that are plugged into wall sockets. Plug loads present a unique problem with respect to energy
management in that most equipment when plugged into a wall socket regardless of whether it’s “on” or
not, will use energy. Obviously, when items plugged into wall sockets are actively used they use
considerably more power than when they’re not, but it turns out, much like a car, a considerable amount
of energy is wasted sitting in an idle state. As a result, the preference is to deactivate outlets when
devices that are plugged into them are not in use (obviously the old fashioned approach to this would be
to just unplug the equipment, but oftentimes that is an impractical option). It so turns out, that electrical
outlets exist that can be activated or deactivated in response to occupancy sensors, demand-response
events, or even from Wi-Fi connected devices. Naturally these outlets are referred to as Smart Outlets.
Unfortunately these outlets do not presently provide any usable data stream and instead are just
controllable.
Forecasting the load with a novel approach
Since plug load usage is, however, inherently associated with occupancy, it is conceivable that where
plug loads are submetered within a building that plug load related energy consumption could be
forecasted by comparing metered usage with occupancy information via the use of machine learning. The
42
below diagram (figure 22) described the system framework by which plug load energy usage forecasting
could be accomplished.
Figure 22. Plug Load Machine Learning Model diagram
Model Predictive Control of HVAC and HVAC Forecasting:
Not a new concept, but an important one revisited
No discussion about energy management in facilities would be complete without addressing the elephant
in the room that is HVAC. As was shown earlier HVAC represents a significant portion of energy usage in
facilities. HVAC energy consumption is derived, predominantly, from the operation of cooling plants,
heating plants, fans, and pumps. At a more basic level, this breaks down to the operation of electric
motors used to perform various different functions, electric heating elements, and burners for heating. It
turns out there is a broad spectrum of methodologies by which these devices are controlled; everything
from basic binary control, to dual speed, multi-speed, and variable speed (similar types of control are
feasible with electrical heating elements). Contemporary equipment has largely taken the path of variable
speed control, but virtually all of the different control schemes exist in the diverse array of buildings that
make up the building stock. Before we dig into that farther though, it is appropriate to acknowledge the
other contributing factors that influence the need for HVAC systems. Working our way from the outside in,
the single largest contributor for the need to condition interior spaces is the building envelope and the
buildings orientation. The building envelope not only isolates the natural elements from the indoors, but it
insulates a building from temperature differentials, provides an air and moisture tight seal, and where
windows are present, permits natural illumination to permeate a space. Exterior walls, doors, windows (
and the roof and floor) are all boundaries across which heat will flow depending on the nature of the
differential. Contributors in the interior space are people, lighting, and equipment, which all radiate heat.
The role of the HVAC system is to balance the various heat flows in a space such that an environment is
maintained that is perceived as comfortable by humans. It should be noted that an HVAC system is
tasked with providing clean air to ventilate a space, while also managing humidity and air temperature,
otherwise referred to as latent and sensible loads. In the summer months outside temperatures and
humidity levels are elevated. Buildings are expected to both reduce temperatures and humidity levels
during these scenarios. On the flip side, in the winter months outsides temperatures and humidity levels
are lower requiring both heating and humidification to maintain comfortable indoor conditions. Naturally a
43
major influencer on the need for space conditioning is the status of the weather, both in terms of
temperature, humidity, and solar exposure. The latter here being a element often forgotten about that also
impacts human comfort. Solar gain, as it is referred to in building science, not only heats up materials in
spaces, but also induces a heating sensation in people. This is referred to as Mean Radiant Temperature.
Interestingly, air temperature, relative humidity, and airspeed can all be at favorable levels, but if the
mean radiant temperature is too high people will feel uncomfortable. Control of radiation from the sun is
often managed by either window selection and/or via the use of operable blinds or shades. The problem
with blinds, in most cases, is that they are setup by occupants when either radiation or daylight
illumination are too high and then are not re-adjusted when those scenarios subside. This tendency
results in an unnecessary dependency on artificial lighting and HVAC systems to compensate. In the case
of new buildings, special consideration should be taken into account for how daylighting and solar
radiation are going to be managed. In the case of existing buildings, it would behoove those facilities to
install window tinting or motorized window shades, in lieu of manually operated blinds. For the purposes
of this thesis we’ll assume the building has motorized window shades that are tied into the lighting
sensors. That is to say that when daylighting levels exceed a certain threshold, shades will be drawn to
manage indoor illumination levels. Solar radiation intensity and intensity of natural illumination are related
in that they are just different wavelengths of electromagnetic radiation impinging the surfaces of a
building.
As a result of these insights, it’s plain to see there are many variables impacting the thermal loading of a
building, such that virtually no two buildings will respond the same way. This leads to our next talking
point.
With respect to thermal management of interior spaces, HVAC systems either add or extract heat from a
space using a variety of different means. To effectively control a space to stay within a range of
parameters there are two distinct approaches. The conventional approach is through the use of a
thermostat, wherein there is a deadband in which temperature is allowed to drift between pre-specified
limits before corrective action is made; this is a reactive control scheme. The alternative approach,
although not a new concept, but one that is back in vogue, is predictive control. All spaces are influenced
by a variety of parameters that affect the responsiveness of a space to correction. By monitoring indoor
temperature, occupancy, and outdoor weather phenomena, is hast been demonstrated that HVAC
systems can be dispatched in a more predictive and proactive manner by learning from past responses
which enable them to facilitate more tightly controlled environmental conditions. This control scheme
integrates seamlessly with the control scheme discussed earlier regarding Workplace Applications. Past
attempts have been made at predictive control by relying on physics based energy models, but due to the
inherent limitations of these models true performance always seems to be lacking. More recent research,
however, has demonstrated the improvements that can be had by using a purely data-driven approach or
even a hybrid approach. Research has demonstrated that by using a machine learning model that
analyzes data provided by all of the elements included as part of the HVAC control system a data-driven
model predictive control scheme can be adopted. This means that each building managed by this type of
control system would be intimately aware of how its HVAC system responds to internal loads and its
building envelope by merely relying on data streams provided by integral sensing equipment.
.
Like lighting, HVAC systems are often established on their own electrical circuits making it feasible to
submeter the energy usage of the HVAC system. HVAC systems do present the added challenge that in
many parts of the country, where the infrastructure exists to provide it, buildings rely on natural gas for
heating. In the case of those facilities, gas usage as it relates to HVAC would need to be separately
metered, as it often is for billing purposes. Like lighting systems, by analyzing meter readings along with
data-streams from the many pieces of equipment and sensors in a building along with weather data, a
44
machine learning model is able to forecast the energy consumption of an HVAC system. The below
diagram (figure 23) describes the complex nature by which control, forecasting, notification, and
optimization of an HVAC system could be managed via Machine Learning.
Adding new useful features to the MPC scheme
In monitoring the innumerable inputs connected to a building automation system, not only can an MPC
system determine future control responses and energy usage, but also by monitoring key parameters
over time a machine learning model also has the capacity to recognize anomalous circumstances such as
performance degradation. These automated evaluations can be used to generate notifications provided to
facility management staff to investigate equipment for maintenance or repair before a disruptive problem
presents itself. Over time such a system could also establish its own customized preventative
maintenance schedule based on historical system use and degradation periods. No two buildings operate
in the same manner, and occupancy patterns are not a fixed normal, hence prescribed preventative
maintenance schedules do not inherently make sense. Similarly, relying simply on runtime is not
adequate as equipment wear over many starts and stops is not the same as that induced by continuous
operation over the same aggregated run-time. A more sophisticated approach is needed to schedule
maintenance activities. Facility maintenance staff never have a shortage of tasks to accomplish, it’s time
that the building automation system provides information that better utilizes personnel time.
Figure 23. HVAC Machine Learning Model
45
HVAC System Optimization:
A new approach to using less
Building on the knowledge gathered by a predictive model, a separate machine learning model could be
developed to optimize the functionality of a HVAC system. Forecasting control responses of an HVAC
system is one thing, improving those control responses is wholly separate task and is actually a
multi-faceted concept. There is optimization of functionality of individual pieces of equipment, and
optimization of combined system response, all with the intent of minimizing energy usage while yielding a
comfortable indoor environment. As it turns out, all buildings are commissioned with a set sequence of
operations applied to system components. These sequences are best guesses at how to control the
indoor environment. The parameters, particularly the ones pertaining to thermal control, really ought not
to be static parameters, and instead should be variables that can be modified by the BAS. By monitoring
parameters associated with the equipment as well as the space the equipment serves, the BAS can
makes adjustments to those governing parameters of the sequence of operations to optimize the use of
that equipment. Similarly, upstream equipment operational parameters can be optimized by correlating
their influence on downstream performance. A machine learning model that operates on this scheme
would learn how to precisely time events, and adjust operational parameters, to yield a desired end
response while using as little energy as needed to do so. Such changes and improvements would be
reflected in the data provided by meters, and possibly even from extrapolation of data provided by
equipment sensors (i.e. voltage and current sensors). To perform such optimization processes, use of the
GRU RNN alone would likely not be sufficient. To supplement that model, use of a Genetic Algorithm
(GA) would enable this system to find those optimal arrangements that minimize energy usage while
maintaining acceptable levels of IEQ. Genetic Algorithms function by implementing a process similar to
that of biological Natural Selection. The algorithm evaluates fitness levels for each member of a
population level (i.e. ability to achieve certain end goals; in biology those goals are survival and
reproduction). In this case fitness level would be based on lowest is best for energy usage and the ability
to maintain high levels of IEQ or better stated maintained IEQ variables within certain bounds that ensure
comfort. There are three main steps to a GA; selection, crossover, and mutation. What a GA effectively
does is manipulates the weights of connections between the nodes in a Neural Network to produce new
results which are evaluated for “fitness”. It’s an iterative process that eventually converges on a network
that will produce the desired end result consistently. GRU RNNs are suited for prediction and not
optimization, for that added functionality a GA (or some other similar type of algorithm) must be applied.
Because GA’s often require many (often hundreds if not thousands) of iterations to converge on a
optimum, it is recommended that optimization be performed on historical data offline, in the cloud,
periodically and that the results be applied to active MPC neural network model(s). How frequently
optimization should be performed is uncertain, but a good starting point would likely be monthly model
optimization. Doing so would yield regularly optimized system performance. The below diagrams (figure
24 & 25) depict the integration of HVAC optimization into the Machine Learning based control scheme as
well as the functional process of a GA algorithm.
46
Figure 24. Diagram of HVAC Optimization Model with relation to the HVAC Machine Learning Model
Figure 25. Interface of a Genetic Algorithm with a Neural Network
47
Use of External Data Inputs (Weather and Utility):
It has now been alluded to numerous times, that use of externally sourced data is imperative to fully
optimize energy consumption and cost savings in buildings. With respect to weather, it turns out that the
National Weather Service (NWS) provides a continuous datastream of multiple parameters that define the
state of the weather in a given location at any given time. It also turns out that weather forecasts are
produced both for large regions (i.e. a metropolitan area) and for smaller regions (i.e. towns or
communities) and are completed according to a grid mosaic. Effectively what this means is that there is a
weather forecast for any building’s particular part of town with a resolution of just a few square miles,
which is extremely useful since that permits it to not only account for what’s happening in the sky, but also
how everything on the ground is influencing what is occuring in the sky. It should come as no surprise that
weather forecasts themselves are the result of machine learning computations. Thus weather data is both
as a real-time feed and a forecast. Together this information can be used by machine-learning models to
influence decisions and forecasts related to energy systems associated with a building. In many cases.
Buildings themselves, will having outdoor environmental sensors that can be used to characterize the
micro-climate around the building. This information can be used in addition to weather forecasts, and
nearby weather station readings.
Similarly, with respect to Real-Time-Pricing, utilities provide a datastream to users in the form of a 24hr
forecast often with a resolution of 15 minutes intervals. This datastream can be used to influence
decisions regarding energy system usage or curtailment. In the case of Time-of-Use pricing, a datastream
isn’t necessarily provided, but the parameters (i.e. costs) at various time intervals are provided. These
parameters can be programmed into a machine learning model to influence decisions regarding energy
system usage or curtailment. The below diagrams (Figure 26) described the many data streams that
make-up weather forecasts and define the current state of the weather.
Figure 26. Diagram of the parameters of weather
48
Renewable Energy Forecasting:
For several decades it was the great quandary of the energy sector, how to estimate the generation of a
renewable energy system. Renewables are notoriously intermittent for a number of reasons. The largest
influencer on performance being the nature of the weather in a region. The weather of the world is a
complex system influenced by the hydrologic cycle, solar radiation, geography, concentration of sun
absorbing gases or sun blocking particulate in the atmosphere, as well as celestial movement. Despite all
of this complexity, researchers have established that electricity generation from renewable energy
systems can be accurately forecasted a handful of days into the future. Weather forecasts, unfortunately,
outside of a few days become progressively unreliable, which inherently inhibits renewable energy
generation forecasting beyond that same timeframe. The National Weather Service (NWS) is
continuously looking to improve its weather forecasts, and is itself utilizing machine learning methods in
it’s weather modeling to improve long term accuracy. As a result, as time progresses and NWS advances
in its methodologies, it is reasonable to assume the time horizon for accurate weather forecasts will
elongate. All said, by evaluating past performance against past weather parameters machine learning
models can be trained to correlate two and those models when equipped with forecasted weather
phenomena are able to forecast renewable energy system generation. This forecasting, although not a
wholly new concept, is an integral component to the greater system being presented later in this thesis.
Figure 27, shown below, depicts the system dynamics of renewable energy forecasting.
Figure 27. Renewable Energy Machine Learning Model diagram
49
New Ideas for Next Generation BAS:
Distributed Generation Dispatching and CHP/ Alternative Energy Forecasting:
With the growing trend to install distributed generation at facilities, it behooves those that own those
systems to ensure they are utilized in a manner that yields optimal cost savings. Although often touted as
a way to reduce reliance on the Power Grid and indirectly reduce GHG emissions, distributed generation,
in practice is more so viewed as a cost savings measure. Distributed generation in most cases resolves to
photovoltaic installations, small scale wind turbines, biomass generators, natural gas fuel cells, or natural
gas cogeneration systems. It should be noted that biomass and cogeneration systems do not help with
reductions in GHG emissions, they only save money and/or reduce dependency on fossil fuels. With
respect to cost savings, some key dynamics to be aware of are the presence of Time-Of-Use electrical
billing, Real-Time-Pricing electrical billing, variable Natural Gas pricing, and weather influences on system
performance. Most cities now have net-metering policies in place enabling producers to sell electricity to
the wholesale electric marketplace when they overproduce. Distributed generation, ultimately, provides
the opportunity to use on-site generated inexpensive power when prices are high and rely on grid
supplied power when prices are low. Additionally, since facility loads follow a recurring profile, generally,
there are often times where the generation of a DG system may be larger than the load it is serving,
which presents the opportunity of being able to sell or store excess power.
For biomass, fuel cells, and co-gen systems, finding and leveraging optimal times to use these systems is
easy enough to accomplish since its just of matter of comparing cost to produce electricity vs. the utility
price of electricity. For renewables, however, this is a more difficult task since availability of generation
and the quantity of generation are variable. So as not to rehash a previous section, I will just reiterate that
methods have already been developed to forecast renewable energy generation. With regards to all forms
of distributed generation, including renewable energy, the new concept being presented here is the
intelligent dispatching of distributed generation for the purposes of selling to the wholesale market, storing
energy, or directly utilizing energy.
It turns out the use of machining learning methods can not only forecast when the best times are to use
these resources but also when to sell the power generated from those assets, and to forecast the cost
savings to be had. This is achievable by training a machine learning model provided with system
performance, building demands, and utility pricing. To effectively use this model, the model would need to
be provided with a levelized cost of electricity associated with the DG system. The model would then
instruct use, sale, or storage of energy based on the forecasted comparison of LCOE, utility cost, and
anticipated level of generation. In the case of real-time-pricing, which typically functions on 15 minute time
frames, systems could be strategically dispatched to coordinate with those time windows to yield
maximum cost savings via the use of DG systems. The below diagram (figure 28) depicts these system
dynamics and shows how machine learning can be leveraged to effectively dispatch and utilize distributed
generation. Similarly figure 29 (same as figure 27) shows how renewable energy generation could
dispatched, which is a new idea to make the most of that resource.
50
Figure 28. Distributed Generation Machine Learning Model diagram
Figure 29. Renewable Energy Dispatching
51
Electrical Energy Storage (EES) Forecasting/Dispatching:
In light of many recent developments in battery technologies, EES is growing in popularity as it is being
implemented in a diverse array of applications. An EES system provides the convenient flexibility of a
stored state of energy that can be used in a diverse number of ways unlike thermal energy which has
limited uses. EES systems have been viewed as both a new form of emergency backup power, a way to
strategically buy and store utility provided power, and as a way to firm the capacity of renewable
generation. For the purposes of this thesis we will focus on the latter two uses. EES is already being used
to store inexpensive power in Time-Of-Use schemes, and some systems are even starting to be used in
Real-Time-Pricing scenarios, but none of these systems provide cost forecasting, capacity forecasting,
and are not tied into a whole building energy management system. This thesis presents a way to take a
more inclusive approach to integrated EES with other building systems.
Since EES functions as a intermediary between the utility, renewable energy systems, and building loads
it is incumbent upon the user that this resource be effectively utilized. To effectively utilize EES, multiple
datastreams would need be analyzed to forecast a control strategy. Obviously, EES is a finite resource
with limited capacity, which inherently constrains how it can be used. Because of the complex
interdynamics between utility pricing, variable generation capacity from renewables, and variable loads
from a building, it not necessarily straightforward how to best utilize EES. To resolve this complex issue,
machine learning, once again, can save the day. Machine learning could conceivably determine it’s
control strategy by evaluating utility cost, versus renewable system LCOE, along with forecasted
renewable generation, and forecasted building loads. Effectively this boils down to when to store utility
power, when to store power from renewables, when to use stored energy in the building, and when to sell
stored energy to the wholesale electric market. For example, it is not uncommon for prices of electricity to
be low in the evening, and for building loads to ramp up early in the morning. For this scenario, low cost
energy could be stored during the evening, and then partially used in the morning when prices may be
higher. As the sun rises energy generation from a photovoltaic system would start to rise and would then
be used to directly power systems within a building. By midday utility prices may change due to an overall
increased demand from air conditioning systems. Meanwhile, photovoltaic systems would be running at
maximum capacity. In this case we’ll assume the photovoltaic system is overproducing to meet the needs
of the facility. As result of overproduction from the renewables the excess power would be sold to the
market. We’ll even assume the building is using whatever energy may still have stored as well. Move
forward a little ways in time and cost of electricity changes again to a lower price. Renewable generation
is still high, but not at its peak. Meanwhile, we also know that tomorrow is supposed to be a warm cloudy
day and that peak pricing will be effect tomorrow. As result of this insight the control system decides to
store the remaining energy from the renewables for use tomorrow and relies on grid power for the rest of
day. This whole scheme is not possible with current technologies. The inclusion of a machine model to
control an EES system would enable this kind of foresight and finely timed control. The below diagram
(figure 30) describes the structure by which EES capacity can be forecasted, and its stored energy
strategically dispatched for optimal cost savings.
52
Figure 30. Electrical Energy Storage Machine Learning Model diagram
Thermal Energy Storage (TES) Forecasting/Dispatching:
Similar to Electrical Energy Storage systems, Thermal Energy Storage is merely a means of storing low
cost energy for future use when energy prices are high. Thermal Energy Storage most commonly takes
the form of cooling capacity in the form of either ice, chilled water, or through the use of some other phase
change substance. Thermal Energy Storage for heating typically only occurs where fuel demand charges
are in effect, or where Solar Thermal Hot Water panels are in use. The latter case is typically just a
means to reduce dependence on fossil fuels for heating, and instead relying on the radiant energy of the
sun to heat a working fluid to transfer heat into a building. What’s more, most facilities use either
electricity in the form of heat-pumps, or natural gas via furnaces or boilers to heat facilities. In the case of
the former, it is virtually unheard of to storage hot air for future use, and in the case of the latter, with
exception of minimizing fuel demand charges, inter-day price fluctuations are a non-factor unlike
electricity, leaving little incentive for thermal energy storage.
For Time-of-Use (TOU) electricity pricing use of TES is fairly straightforward and requires rather basic
control schemes. Effectively equipment is scheduled to store thermal energy at Off-Peak times and then
switch to using the stored energy at On-Peak times. It should be noted, that although the control scheme
for TOU pricing is well established those systems do not provide cost projections. Those control scheme
also do not optimize how much energy to store, instead they usually just fully fill tanks and then deplete
them as much as possible during On-Peak hours. Use of machine-learning can help to optimize the use
of TES by using forecasted loads to determine how much energy to store. In the case of
53
Real-Time-Pricing schemes, machine learning would really come into its own by determining how to much
to store and when based on day ahead pricing schedules and load forecasts. Similar strategies could also
be used in the case of dependence on Solar Thermal Hot Water by estimating heating capacities based
on weather forecasts. Machine learning would have little issue using data to draw correlation between
heat production and weather phenomena. It is conceivably possible that heat energy forecasting for Solar
Thermal heating could be forecasted several days out, and using that information it could be determined
how much a building would need to rely on other heating sources.
Presently it is uncommon to find both TES and EES serving the same facility. For one, due to the
versatility of EES, there is little benefit to also having TES. It is conceivable, though, that some
installations that already have a TES system may adopt an EES system for the sole purpose of firming a
renewable energy system, or to serve as an emergency backup power source, or to expand the energy
storage capacity that is has in the form of a TES; these are all functions a TES is not suited to serve. It is
also conceivable that the cost of EES may be prohibitive to cover both renewable energy generation and
HVAC systems, in which case it might be conceivable that a TES system would also be installed. The last
possibility is that an organization heats their facility with a hydronic system and would like to take
advantage of the sun’s radiation to supplement another system and would also like to leverage the sun’s
radiation to produce electricity via a photovoltaic array. The thing about the solar thermal heating
arrangements is that these systems are more often than not only large enough to enable the primary
heating system to dial back its input. Rarely are they large enough to produce more heat than a facility
needs, let alone all of what it needs. When the systems would conceivably be producing more heat than
is needed is when there would be no need for heat (i.e. summer months) and as a result there would be
no need to store the excess heat (with possible exception of being used for domestic hot water use). In
light of the growing trend towards green buildings, we have to assume that at some point, if it hasn’t
already happened, a facility will be designed such that it can rely solely on solar thermal heating, and for
the sake of argument let’s say it can produce more heat that it needs at peak loads. Due to the fact that
the sun rises and falls every day and that numerous variables impact the intensity of the sunlight
received, use of thermal energy storage to condition spaces during periods of low solar intensity (i.e.
mornings, evenings, inclimate weather days, etc.) would be necessary. In this circumstance
Machine-Learning would likely be helpful in determining how to size the storage system and to a lesser
extent how to utilize it. Unlike current arrangements where TES is used as an alternative source based on
cost, this would be the sole source with no driving economic factor determining when to use it. Machine
Learning, could however, learn the responsiveness of a facility due to internal and external variables and
dispatch heating TES to provide steady indoor environmental conditions (this is true for other systems as
well). It is not likely that a TES system would ever be designed to only provide heating during dimly lit
periods which is contrary to EES (gathering solar energy for use during dimly lit periods) or the inverse of
cooling oriented TES (storing energy when there’s no demand for cooling and using it when there is).
Regardless, to best utilize TES, utilization of a machine-learning based control scheme is favorable. The
below diagram (figure 31) details how a TES machine-learning control scheme would function.
54
Figure 31. Thermal Energy Storage Machine Learning Model diagram
Building Level Energy Forecasting:
At this point we have addressed all of the individual elements of the greater building system that may or
may not be present. To forecast the overall energy usage of a building, all that would need to be
performed is a an aggregation of the all the contributing factors. Thus a simple algorithm could sum the
forecasts associated with HVAC, lighting, and plug loads, to determine how much energy the building is
going to need to perform its anticipating functions to accommodate the anticipated occupants and their
activities in the building. How much of a forecast is the subject of a separate discussion, but at minimum
research has suggested that a minimum of a 24hr time horizon is preferable. It is arguable that forecasts
as large as weekly, monthly, or quarterly could be desirable, but with the limiting factors such weather
variability and utility pricing uncertainty it is hard to see how forecasts beyond a few days would yield any
meaningful information since their accuracy will likely quickly diminish with longer timeframes. It is also
debatable whether or not building level energy forecasting has an real intrinsic value to the end user.
From a machine-learning point of view this aggregated value could be of value for future forecasting
especially as it is applied to cost forecasting. The below diagram (figure 32) showcases the overall
structure of energy management in a building that includes all of the aforementioned system. Figure 33
describes a Machine-Learning that would forecast the overall building energy consumption based on the
forecasts of the individual submetered systems.
55
Figure 32. Overall Integrated Building Systems Model
Figure 33. Total Building Energy Usage Machine Learning Model diagram
56
Building Operational Cost Forecasting:
Since operational costs of a facility are a major influencer on budgets managed by organizations that
occupy buildings, there is definite value in knowing how much it is going to cost to run a facility before it
has occurred. Unlike energy forecasting, cost forecasting is an actionable tool an organization can wield.
Presently, at best, organizations are able to participate in Demand-Response events, or to trigger their
own load-shedding procedures to cut costs where possible and to help the utility control costs. They also
are able to utilize EES, TES, and renewable energy to cut costs, but without the kind of advanced control
I previously described. The addition of that control would yield additional cost savings. By using the
machine-learning based control schemes presented here, costs can also be forecasted by correlating
energy usage as time elapses against energy costs in affect and by taking into account the forecasted
use of EES, TES, and DG. Collectively this provides an overall cost to operate estimate that can be
compared against actual cost, and broken apart to identify opportunities for improvement that will
continue to keep costs down. In other words, it can be tracked how much money EES, TES, and DG are
saving, and how much money HVAC, Lighting, and Plug Loads are costing. As a result, a system that
functions in this manner is able to provide a service that, to my knowledge, has never been provided
before. That service being real-time cost trending, and forecastable building operational costs. The
diagram below (figure 34) describes how a machine-learning model could used to generate operational
cost forecasts.
Figure 34. Cost of Energy Machine Learning Model diagram
57
Integrated BAS Architecture:
Up this point I have described a multitude of systems to be managed by data-driven Model Predictive
Control operated by recurrent neural networks. Although alluded to, it has not been explicitly identified
how all of the pieces in this integrated control methodology interface with one another. Referring back to
Figure 31, fundamentally, what the below architecture describes, is an approach by which two-way
information sharing amongst all of the individual energy users and suppliers can communicate and inform
one another of their needs, capacities, and plans. Historically many of these systems have operated in a
mutually exclusive fashion, when in reality it’s all one large integrated system. The chart below (figure 35)
provides a high level depiction of what an integrated architecture would look like that leverages cloud
computing, machine-learning control logic, existing BAS controllers and programming, data streams from
all the various energy systems previously described, as well as the inputs provided by occupants via a
workplace application. To my knowledge this integrated framework is the first of its kind.
Figure 35. Data-Driven MPC System Architecture Diagram
58
Future Development:
This thesis presents a framework and methodology by which machine-learning can be leveraged to
optimize the operations of a building. It does not however delve into the specifics of what exactly those
fundamental algorithms would look like and how the tremendous amounts of data provided by a building
would be managed and processed within those machine-learning models. Data scientists would need to
explore and develop each of the machine-learning models identified in this thesis to yield a functional
system that fulfills the capabilities described by this framework. Similarly, data scientists and/or
programmers would need to determine how the cloud based machine-learning models would interface
with existing building automation systems, and determine how those models be able to facilitate
adjustments in programming in an autonomous manner. Despite those uncertainties I do hypothesize,
however, that implementation of all the many integrations defined here reliant on two way communication
and machine learning will not only reduce energy usage in buildings appreciably, but also reduce
operational costs, extend equipment lifetimes, provide facility managers with more actionable information,
and enable higher levels of sustained IEQ than was previously attainable. I also hypothesize that the
added intelligence that this integrated control methodology affords enables buildings to be more agile,
resilient, strategic, and better able to manage the stochastic phenomena that occurs around and within
buildings that affects the consumption of energy. I encourage industry researchers and academics to
evaluate the ideas presented in this thesis with the programming logic that would be required to facilitate
it and determine whether or not the improvements I am projecting can in fact be realised.
Furthermore, it would have to be determined how the machine-learning models could be universally
applicable to work with any type of building automation system. I hypothesize that integration with a
multitude of existing products is feasible, but that some degree of tailoring may be necessary to facilitate
integrations with the various brands’ products and encourage researchers to validate or disprove that
hypothesis.
Although the methods of control defined in this thesis are seemingly conveniently scalable, an economic
study would need to be performed to determine if such a service would be cost beneficial in all cases, or if
there’s a threshold below which smaller facilities would not recoup the costs of the service in savings. I
hypothesize that there may be limit of building simplicity where such a control scheme is not a cost
effective technology to adopt.
In addition to that, this thesis has been tailored towards standalone facilities occupied by one tenant
organization. In reality, especially with regards to large buildings, facilities are occupied by multiple tenant
organizations. I believe there is a way to institute a control scheme like what I have described here for
such facilities, but I would encourage future research to explore how that would be facilitated. Similarly, in
some cases, facilities are grouped together as a campus and share some energy systems. I hypothesize
that the framework described here could be expanded to apply to those scenarios, but also recognize that
the scenario becomes complicated with having to manage separate buildings in a mutually exclusive
fashion while also coordinating shared resources to serve those separate buildings. This scenario starts
to blur the lines between Smart/Micro Grid Operation and Smart Buildings. I encourage researchers to
explore, and even determine how delineation of the systems would be managed, while also retaining the
level of integration proposed in this thesis.
59
Conclusion:
Through full integration of workplace applications, IoT sensors, smart lighting, HVAC systems, plug load
control, connected thermal energy storage, electric energy storage, and dispatchable renewable energy
or alternative energy technologies, energy costs can be minimized, occupant comfort assured, and
energy usage optimized.
To facilitate this weather forecasts can be utilized to forecast HVAC and lighting demands as well as
contributions from renewable energy generation. Real-time pricing from electrical utilities could be used to
indicate when to sell on-site generation to the grid or when to store or utilize electrical and or thermal
energy within the facility to both reduce electric demand and avoid high periods of energy cost. By tying
energy supply systems into a Building Management System (BMS)(aka BAS) and facilitating
communication between the supply side and the demand side, systems can be fully optimized in their
operations. This optimization can be facilitated by utilizing machine learning and IoT devices in buildings
making them fully aware and able to respond to their influencing parameters whether they be the weather,
occupancy patterns, or utility supply dynamics. Lighting, HVAC, Plug Load management, Renewables,
Alternative Energy, and Energy Storage Systems have all seen great advancements independently
leading towards this key integral moment. Development of standardized data protocols (open source) and
data tagging methodologies are also helping bring us closer to this comprehensive state of integration.
And more than anything, the bridging technology that brings these systems together and enables detailed
control is the development of Artificial Intelligence in the form of Recurrent Neural Network
machine-learning models.
No two buildings are completely identical and the full integration of systems, as described previously,
enables buildings to learn about themselves and optimize themselves by leveraging Artificial Intelligence.
The world is not a stagnant nor entirely predictable place, and our buildings need to be flexible and
responsive in a dynamic manner to match that. Although many high level cosmic phenomena are
predictable (such as the daily passage of the sun and seasons) low level local phenomena are not
predictable outside of a short time frame, and patterns of short term phenomena aggregated over long
periods of time (i.e. climate) can be and are changing, and our buildings need to be flexible to those
changes. As a result of this, and with the provision of Artificial Intelligence, buildings can constantly make
improvements to optimize for new scenarios instead of assuming input parameters will always be the
same. In a sense this future-proofs a building’s systems making them more flexible to changes in the
future which is something many legacy or current systems are not setup to handle. That said, all systems
in a building have a life span and as a result their components will wear out, fail, or become obsolete and
will inevitably need to be replaced. Conveniently though, in a building managed by machine-learning, all
underlying systems can be replaced, and the control system will just adjust to accommodate the changes
automatically; something that legacy systems or even current systems are not able to do. For clarity what
I’m suggesting is that the software is somewhat future-proof, and any and all of the controls hardware in a
building has a finite life expectancy, and like the systems they serve will wear out, fail, or become
obsolete and will require replacement eventually. The machine-learning models, on the-other-hand, would
only need periodic updates that could be performed over the internet and will otherwise never wear out.
This inherently instills greater intrinsic value in the control system than a system suited for only a single
designed scenario, or set of scenarios, and that is otherwise incapable of changing operational
parameters, or is too difficult to use making it infeasible or not worthwhile to implement changes to.
60
In conclusion this thesis presents a new breed of building automation system that truly manages all
energy systems within a building through comprehensive integration, and provides the capability to keep
occupants in the loop ensuring that the building operates in a manner to serve its occupants, as it is
intended to do, while also minimizing the amount of energy (and GHG emissions) required to accomodate
them and minimize the cost to the run the building.
61
Bibliography:
[1] Agüero-Rubio, J., Giménez-Fernández, A., Callejón-Ferre, A., & López-Martínez, J. (2014, 09). Simple rule
for management of thermal loads with real-time prices. Journal of Cleaner Production, 78 , 48-53.
doi:10.1016/j.jclepro.2014.05.001
[2] Ahmad, M. W., Mourshed, M., & Rezgui, Y. (2017, 07). Trees vs Neurons: Comparison between random
forest and ANN for high-resolution prediction of building energy consumption. Energy and Buildings,
147 , 77-89. doi:10.1016/j.enbuild.2017.04.038
[3] Aryal, A., & Becerik-Gerber, B. (2018, 10). Energy consequences of Comfort-driven temperature setpoints in
office buildings. Energy and Buildings, 177 , 33-46. doi:10.1016/j.enbuild.2018.08.013
[4] Assouline, D., Mohajeri, N., & Scartezzini, J. (2017, 01). Quantifying rooftop photovoltaic solar energy
potential: A machine learning approach. Solar Energy, 141 , 278-296. doi:10.1016/j.solener.2016.11.045
[5] Baker, L. and Hoyt, T. (2016). Control for the People: How Machine Learning Enables Efficient HVAC Use
Across Diverse Thermal Preferences. ACEEE, [online] (Summer Study on Energy Efficiency in
Buildings). Available at: https://aceee.org/files/proceedings/2016/data/papers/8_57.pdf
[6] Bedi, A. S., P., P. A., Ahmad, M. W., Swapnil, S., Rajawat, K., & Anand, S. (2018, 10). Online algorithms for
storage utilization under real-time pricing in smart grid. International Journal of Electrical Power &
Energy Systems, 101 , 50-59. doi:10.1016/j.ijepes.2018.02.034
[7] Behera, M. K., Majumder, I., & Nayak, N. (2018, 06). Solar photovoltaic power forecasting using optimized
modified extreme learning machine technique. Engineering Science and Technology, an International
Journal, 21 (3), 428-438. doi:10.1016/j.jestch.2018.04.013
[8] Carreira, P., Costa, A. A., Mansur, V., & Arsénio, A. (2018, 08). Can HVAC really learn from users? A
simulation-based study on the effectiveness of voting for comfort and energy use optimization.
Sustainable Cities and Society, 41 , 275-285. doi:10.1016/j.scs.2018.05.043
[9] Castaldo, V. L., Pigliautile, I., Rosso, F., Cotana, F., Giorgio, F. D., & Pisello, A. L. (2018, 11). How
subjective and non-physical parameters affect occupants’ environmental comfort perception. Energy
and Buildings, 178 , 107-129. doi:10.1016/j.enbuild.2018.08.020
[10] Deb, C., Eang, L. S., Yang, J., & Santamouris, M. (2015). Forecasting Energy Consumption of Institutional
62
Buildings in Singapore. Procedia Engineering, 121 , 1734-1740. doi:10.1016/j.proeng.2015.09.144
[11] Elkhoukhi, H., Naitmalek, Y., Berouine, A., Bakhouya, M., Elouadghiri, D., & Essaaidi, M. (2018). Towards
a Real-time Occupancy Detection Approach for Smart Buildings. Procedia Computer Science, 134 ,
114-120. doi:10.1016/j.procs.2018.07.151
[12] Jung, H. C., Kim, J. S., & Heo, H. (2015, 03). Prediction of building energy consumption using an improved
real coded genetic algorithm based least squares support vector machine approach. Energy and
Buildings, 90 , 76-84. doi:10.1016/j.enbuild.2014.12.029
[13] Li, J., Ward, J. K., Tong, J., Collins, L., & Platt, G. (2016, 05). Machine learning for solar irradiance
forecasting of photovoltaic system. Renewable Energy, 90 , 542-553. doi:10.1016/j.renene.2015.12.069
[14] Li, K., Xie, X., Xue, W., Dai, X., Chen, X., & Yang, X. (2018, 09). A hybrid teaching-learning artificial neural
network for building electrical energy consumption prediction. Energy and Buildings, 174 , 323-334.
doi:10.1016/j.enbuild.2018.06.017
[15] Li, Z., & Dong, B. (2018, 01). Short term predictions of occupancy in commercial buildings—Performance
analysis for stochastic models and machine learning approaches. Energy and Buildings, 158 , 268-281.
doi:10.1016/j.enbuild.2017.09.052
[16] Luo, N., Hong, T., Li, H., Jia, R., & Weng, W. (2017, 10). Data analytics and optimization of an ice-based
energy storage system for commercial buildings. Applied Energy, 204 , 459-475.
doi:10.1016/j.apenergy.2017.07.048
[17] Manjarres, D., Mera, A., Perea, E., Lejarazu, A., & Gil-Lopez, S. (2017, 10). An energy-efficient predictive
control for HVAC systems applied to tertiary buildings based on regression techniques. Energy and
Buildings, 152 , 409-417. doi:10.1016/j.enbuild.2017.07.056
[18] Martin, R., Aler, R., Valls, J. M., & Galvan, I. M. (2015, 08). Machine learning techniques for daily solar
energy prediction and interpolation using numerical weather models. Concurrency and Computation:
Practice and Experience, 28 (4), 1261-1274. doi:10.1002/cpe.3631
[19] Mateo, F., Carrasco, J. J., Sellami, A., Millán-Giraldo, M., Domínguez, M., & Soria-Olivas, E. (2013, 03).
Machine learning methods to forecast temperature in buildings. Expert Systems with Applications, 40 (4),
1061-1068. doi:10.1016/j.eswa.2012.08.030
[20] Mirakhorli, A., & Dong, B. (2016, 10). Occupancy behavior based model predictive control for building
63
indoor climate—A critical review. Energy and Buildings, 129 , 499-513.
doi:10.1016/j.enbuild.2016.07.036
[21] Murphy, M., O’Mahony, M., & Upton, J. (2015, 07). Comparison of control systems for the optimisation of
ice storage in a dynamic real time electricity pricing environment. Applied Energy, 149 , 392-403.
doi:10.1016/j.apenergy.2015.03.006
[22] Pérez-Ortiz, M., Jiménez-Fernández, S., Gutiérrez, P., Alexandre, E., Hervás-Martínez, C., &
Salcedo-Sanz, S. (2016, 08). A Review of Classification Problems and Algorithms in Renewable Energy
Applications. Energies, 9 (8), 607. doi:10.3390/en9080607
[23] Rahman, A., & Smith, A. D. (2017, 10). Predicting fuel consumption for commercial buildings with machine
learning algorithms. Energy and Buildings, 152 , 341-358. doi:10.1016/j.enbuild.2017.07.017
[24] Rahman, A., & Smith, A. D. (2018, 10). Predicting heating demand and sizing a stratified thermal storage
tank using deep learning algorithms. Applied Energy, 228 , 108-121.
doi:10.1016/j.apenergy.2018.06.064
[25] Salcedo-Sanz, S., Cornejo-Bueno, L., Prieto, L., Paredes, D., & García-Herrera, R. (2018, 07). Feature
selection in machine learning prediction systems for renewable energy applications. Renewable and
Sustainable Energy Reviews, 90 , 728-741. doi:10.1016/j.rser.2018.04.008
[26] Shioya, M., Yairi, T., Masukawa, Y., Yoshida, K.. (2015). Energy fault detection in office building system by
machine learning methods. ASHRAE Transactions , vol. 121, no. 1, 2015, p. 185+.
[27] Steriotis, K., Tsaousoglou, G., Efthymiopoulos, N., Makris, P., & Varvarigos, E. (2018, 08). Real-time
pricing in environments with shared energy storage systems. Energy Efficiency .
doi:10.1007/s12053-018-9723-8
[28] Touzani, S., Granderson, J., & Fernandes, S. (2018, 01). Gradient boosting machine for modeling the
energy consumption of commercial buildings. Energy and Buildings, 158 , 1533-1543.
doi:10.1016/j.enbuild.2017.11.039
[29] Voyant, C., Notton, G., Kalogirou, S., Nivet, M., Paoli, C., Motte, F., & Fouilloy, A. (2017, 05). Machine
learning methods for solar radiation forecasting: A review. Renewable Energy, 105 , 569-582.
doi:10.1016/j.renene.2016.12.095
[30] Wang, F., Zhen, Z., Mi, Z., Sun, H., Su, S., & Yang, G. (2015, 01). Solar irradiance feature extraction and
64
support vector machines based weather status pattern recognition model for short-term photovoltaic
power forecasting. Energy and Buildings, 86 , 427-438. doi:10.1016/j.enbuild.2014.10.002
[31] Wei, T., Wang, Y., & Zhu, Q. (2017). Deep Reinforcement Learning for Building HVAC Control.
Proceedings of the 54th Annual Design Automation Conference 2017 on - DAC '17 .
doi:10.1145/3061639.3062224
[32] Wong, J. K., & Li, H. (2009, 01). Development of intelligence analytic models for integrated building
management systems (IBMS) in intelligent buildings. Intelligent Buildings International, 1 (1), 5-22.
doi:10.3763/inbi.2009.0011
[33] Wu, L., Kaiser, G., Solomon, D., Winter, R., Boulanger, A., & Anderson, R. (2012, 05). Improving efficiency
and reliability of building systems using machine learning and automated online evaluation. 2012 IEEE
Long Island Systems, Applications and Technology Conference (LISAT) .
doi:10.1109/lisat.2012.6223192
[34] Wu, Z., Li, N., Peng, J., Cui, H., Liu, P., Li, H., & Li, X. (2018, 08). Using an ensemble machine learning
methodology-Bagging to predict occupants’ thermal comfort in buildings. Energy and Buildings, 173 ,
117-127. doi:10.1016/j.enbuild.2018.05.031
[35] Yu, L., Xie, D., Huang, C., Jiang, T., & Zou, Y. (2018). Energy Optimization of HVAC Systems in
Commercial Buildings Considering Indoor Air Quality Management. IEEE Transactions on Smart Grid,
1-1. doi:10.1109/tsg.2018.2875727
[36] Zeng, P., Li, H., He, H., & Li, S. (2018). Dynamic Energy Management of a Microgrid using Approximate
Dynamic Programming and Deep Recurrent Neural Network Learning. IEEE Transactions on Smart
Grid, 1-1. doi:10.1109/tsg.2018.2859821
[37] Zhao, H., Hua, Q., Chen, H., Ye, Y., Wang, H., Tan, S. X., & Tlelo-Cuautle, E. (2018, 06).
Thermal-Sensor-Based Occupancy Detection for Smart Buildings Using Machine-Learning Methods.
ACM Transactions on Design Automation of Electronic Systems, 23 (4), 1-21. doi:10.1145/3200904
[38] Zou, H., Zhou, Y., Yang, J., & Spanos, C. J. (2018, 10). Towards occupant activity driven smart buildings
via WiFi-enabled IoT devices and deep learning. Energy and Buildings, 177 , 12-22.
doi:10.1016/j.enbuild.2018.08.010
Abstract (if available)
Abstract
Advanced technology is becoming increasingly more accessible, affordable, and applicable with respect to building energy systems which once existed and operated in an isolated manner and were managed by unsophisticated systems. These systems are becoming increasingly more complex, integrated, and carefully controlled largely due to rising energy costs and expectations for higher levels of indoor environmental quality. In the past, and in same cases even today, lighting, HVAC, energy storage, plug loads, and distributed generation systems were viewed as independent systems that enabled limited human interaction, system monitoring, or control, but today there is growing expectation for optimized services from these systems requiring a more dynamic, integrated, and intelligent approach to managing these systems with the added expectation of a provision some sort of human feedback interface to make buildings more accommodating to their occupants. For too long occupants have felt that buildings are not functioning in a manner to accommodate their needs, and rather operating to fulfill programmed parameters. For decades building controls reduced to following relatively simple sequences of operation developed by engineers to meet basic criteria that fit discrete operational scenarios. The problem with this approach is that, to some extent, changes in weather and occupancy patterns are not accounted for in the control scheme, resulting in the unnecessary expenditure of large quantities of energy to operate systems in specific places at times where they are not really needed. Not only that, but now many facilities are adding on-site power generation, energy storage, and advanced lighting systems in addition to modern HVAC systems. As a result of these advancements there is now a growing expectation that buildings become virtually sentient beings capable of both maintaining healthy and comfortable environments while also minimizing the use of energy to keep utility costs low. It doesn’t take much research to realize that costs of energy are on the rise, and are expected to continue for the foreseeable future. Despite that, with the growing integration of Internet of Things (IoT) in devices and the development and utilization of cloud computing for offsite computational capacity and data management along with advancements in Artificial Intelligence, Machine Learning has entered the lexicon of the building automation industry. Building Automation was once delegated to merely managing HVAC operations with other systems being managed by other control systems
Linked assets
University of Southern California Dissertations and Theses
Conceptually similar
PDF
Learning personal thermal comfort and integrating personal comfort requirements into HVAC system control loop
PDF
Intelligent adaptive automation: activity-driven and user-centered building automation
PDF
Smart buildings: synergy in structural control, structural health monitoring and environmental systems
PDF
Developing environmental controls using a data-driven approach for enhancing environmental comfort and energy performance
PDF
Economic model predictive control for building energy systems
PDF
Integration of energy-efficient infrastructures and policies in smart grid
PDF
Distributed adaptive control with application to heating, ventilation and air-conditioning systems
PDF
A joint framework of design, control, and applications of energy generation and energy storage systems
PDF
"Watts per person" paradigm to design net zero energy buildings: examining technology interventions and integrating occupant feedback to reduce plug loads in a commercial building
PDF
The next generation of power-system operations: modeling and optimization innovations to mitigate renewable uncertainty
PDF
User-centric smart sensing for non-intrusive electricity consumption disaggregation in buildings
PDF
Active state tracking in heterogeneous sensor networks
PDF
Defending industrial control systems: an end-to-end approach for managing cyber-physical risk
PDF
Charge-mode analog IC design: a scalable, energy-efficient approach for designing analog circuits in ultra-deep sub-µm all-digital CMOS technologies
PDF
Towards health-conscious spaces: building for human well-being and performance
PDF
Architectural innovations for mitigating data movement cost on graphics processing units and storage systems
Asset Metadata
Creator
Birnbaum, Sean Robert
(author)
Core Title
Smart buildings: employing modern technology to create an integrated, data-driven, intelligent, self-optimizing, human-centered, building automation system
School
Viterbi School of Engineering
Degree
Master of Science
Degree Program
Green Technologies
Publication Date
07/26/2019
Defense Date
04/17/2019
Publisher
University of Southern California
(original),
University of Southern California. Libraries
(digital)
Tag
alternative energy,BAS,big data,BMS,building automation systems,building control systems,building load forecasting,building management system,climate change,cloud computing,CO2 emissions,cogeneration,distributed generation,dynamic control programming,dynamic setpoints,electrical energy storage,EMS,energy conservation,energy efficiency,energy management system,energy storage,energy system integration,GHGs,global energy usage,Green Technologies,greenhouse gas emissions,high efficiency HVAC,IAQ,IEQ,indoor air quality,indoor environmental quality,Internet of Things,IoT,load shedding,machine learning,mobile applications,modern buildings,OAI-PMH Harvest,real-time pricing,recurrent neural networks,renewable energy,renewables generation forecasting,roof-top solar PV,SaaS,sequences of operation,smart buildings,smart HVAC,smart lighting,smart programming,system integration,thermal comfort,thermal energy storage,utility cost forecasting,workplace applications
Format
application/pdf
(imt)
Language
English
Contributor
Electronically uploaded by the author
(provenance)
Advisor
Maby, Edward (
committee chair
), Choi, Joon-Ho (
committee member
), Krishnamachari, Bhaskar (
committee member
), Prasanna, Viktor (
committee member
)
Creator Email
sbirnbau@usc.edu,sean.robert.birnbaum@gmail.com
Permanent Link (DOI)
https://doi.org/10.25549/usctheses-c89-193100
Unique identifier
UC11662486
Identifier
etd-BirnbaumSe-7637.pdf (filename),usctheses-c89-193100 (legacy record id)
Legacy Identifier
etd-BirnbaumSe-7637.pdf
Dmrecord
193100
Document Type
Thesis
Format
application/pdf (imt)
Rights
Birnbaum, Sean Robert
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
Tags
alternative energy
BAS
big data
BMS
building automation systems
building control systems
building load forecasting
building management system
climate change
cloud computing
CO2 emissions
cogeneration
distributed generation
dynamic control programming
dynamic setpoints
electrical energy storage
EMS
energy conservation
energy efficiency
energy management system
energy storage
energy system integration
GHGs
global energy usage
greenhouse gas emissions
high efficiency HVAC
IAQ
IEQ
indoor air quality
indoor environmental quality
Internet of Things
IoT
load shedding
machine learning
mobile applications
modern buildings
real-time pricing
recurrent neural networks
renewable energy
renewables generation forecasting
roof-top solar PV
SaaS
sequences of operation
smart buildings
smart HVAC
smart lighting
smart programming
system integration
thermal comfort
thermal energy storage
utility cost forecasting
workplace applications