FTN EDA

Foundations and Trends
R
in Electronic Design
Automation
Vol. XX, No. XX (2015) 1–150

c 2015 M. Maasoumy and A. Sangiovanni-Vincentelli
DOI: 10.1561/XXXXXXXXXX
Smart Connected Buildings Design Automation:

Foundations and Trends
Mehdi Maasoumy Alberto Sangiovanni-Vincentelli

UC Berkeley UC Berkeley
maasoumy@eecs.berkeley.edu alberto@eecs.berkeley.edu
Contents
1 Introduction 3
1.1 Why Buildings? . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Why Smart Buildings? . . . . . . . . . . . . . . . . . . . 5
1.3 Areas of research . . . . . . . . . . . . . . . . . . . . . . 8
1.4 Organization . . . . . . . . . . . . . . . . . . . . . . . . . 8
2 Simulation Tools 12
2.1 Building Simulation Tools . . . . . . . . . . . . . . . . . . 12
2.2 Building-to-Grid Simulation Tools . . . . . . . . . . . . . . 19
2.3 Comparisons . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.4 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . 21
3 Building Models 23
3.1 Resistor-Capacitor Models . . . . . . . . . . . . . . . . . . 24
3.2 Parameter-Adaptive Building (PAB) Model . . . . . . . . . 31
4 Building Control Design 41

4.1 Rule-Based Control (RBC) and Reinforcement Learning . . 44
4.2 Model Predictive Control (MPC) . . . . . . . . . . . . . . 46
4.3 Randomized Model Predictive Control . . . . . . . . . . . 49
4.4 Robust Model Predictive Control (RMPC) . . . . . . . . . 49
ii
iii
4.5 Stochastic Model Predictive Control (SMPC) . . . . . . . 51

4.6 Exergy-based Model Predictive Control (XMPC) . . . . . . 52
4.7 Comparisons . . . . . . . . . . . . . . . . . . . . . . . . . 57
5 Test-beds and Real-scale Experiments 77

5.1 Review of Experimental Building Studies . . . . . . . . . . 77
5.2 Review of Large-scale Test-beds for Building Studies . . . 79
6 Designing Building Control Systems as Cyber-Physical Sys-

tems 95
6.1 The Co-design Problem . . . . . . . . . . . . . . . . . . . 98
6.2 Sensing and Prediction Accuracy Modeling . . . . . . . . . 100
6.3 Sensing System Design and Accuracy . . . . . . . . . . . . 104
6.4 Design Space Exploration . . . . . . . . . . . . . . . . . . 106
7 Dynamic Contracts for Building-Grid Interaction 110

7.1 A Supply-Following Scenario for Smart Buildings . . . . . 111
7.2 Ancillary Service from Buildings . . . . . . . . . . . . . . . 114
7.3 Dynamic Contracts . . . . . . . . . . . . . . . . . . . . . 120
7.4 Flexibility-aware Contractual Framework . . . . . . . . . . 123
7.5 Computational Results . . . . . . . . . . . . . . . . . . . 130
8 Conclusion 133
8.1 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . 135
Acknowledgements 137
References 138
Abstract
Buildings are the result of a complex integration of multi-physics sub-

systems. Besides the obvious civil engineering infrastructure, thermal,
electrical, mechanical, control, communication and computing subsys-
tems must co-exist and be operated so that the overall operation is
smooth and efficient. This is particularly important for commercial
buildings but is also very relevant for residential buildings especially
apartment buildings. Unfortunately, the design and deployment of these
subsystems is rarely synchronized: lighting, security, heating, ventila-
tion and air conditioning systems are often designed independently.
However, simply putting together a collection of sub-systems, albeit
optimized, has led to inefficient buildings of today. Worldwide, build-
ings consume 42% of all electrical power – more than any other asset
– and it can be proven that much of this can be reduced if a holistic
approach to design, deployment and operation is taken.
Government agencies, academic institutions, building contractors
and owners have realized the significant impact of buildings on the
global environment, the electrical grid, and the mission of their orga-
nizations. However, the economic impact for all constituencies is still
difficult to assess. Government regulations can play a fundamental role,
as it has been the case for the transportation industry where regulations
on emission and fuel consumption have been the single most important
factor of innovation in automotive design.
We are convinced that by leveraging technology and utilizing a
system-level approach to buildings, they will provide comfort, safety
and functionality while minimizing energy cost, supporting a robust
electric grid and mitigating environmental impact. Realizing this vision
requires adding intelligence from the beginning of the design phase, to
deployment, from commissioning to operation, all the way to the end
of the building’s life cycle.
In this issue, we attempt to provide as complete as possible of an
overview of the activities in the field of smart connected building design
automation that attempts to make the vision a reality. The overarching
range of such activities includes developing simulation tools for mod-
eling and designing of buildings, and consequently control algorithms
2
proposed to make buildings smarter and more efficient. Further, we

will overview real-world and large-scale implementation of such control
strategies on physical buildings. We then present a formal co-design
methodology to design buildings taking the view that buildings are
prime examples of cyber-physical systems where the virtual and phys-
ical worlds meet, as more traditional products such as thermostats are
able to connect online and perform complicated computational tasks
to control building temperature effectively. We complete the presenta-
tion describing the growing role of buildings in the operation of the
smart grid where buildings are not only consumers of energy, but also
providers of services and energy themselves to the grid.
The audiences for this article are industry professionals and re-
searchers who work in the area of smart buildings, smart cities and
smart grid, with emphasis on energy efficiency, simulation tools, opti-
mal control, and cyber-physical systems for the emerging power mar-
kets.
M. Maasoumy and A. Sangiovanni-Vincentelli. Smart Connected Buildings Design

Automation:
Foundations and Trends. Foundations and Trends
R
in Electronic Design
Automation, vol. XX, no. XX, pp. 1–150, 2015.
DOI: 10.1561/XXXXXXXXXX.
1
Introduction
The term intelligent or smart building refers to the next generation of

buildings that provide new levels of comfort to the occupants with min-
imum possible energy consumption. They not only follow commands,
but also proactively learn from occupants’ behavior and adapt their
operation based on the indoor and outdoor conditions. These buildings
are no longer solely consumers of energy, but also significant players in
the ecosystem of smart grid, in that they provide regulation services to
the grid as well as energy if equipped with solar panels or other green
sources. Intelligent buildings not only are safe by design, but also re-
act in the case of a fault, system malfunction, or cyber-attack to steer
the system into a safe operating region. There has been much research
in academia and industry towards this goal. Companies such as Nest
(https://nest.com/), recently acquired by Google, have been formed
over the last few years to bring new technologies in this space to the
public. In this paper, we present an overview of the work done in this
domain over the last two decades.
3
4 Introduction
Figure 1.1: US Primary Energy Overview (Quadrillion Btu) EIA.
1.1 Why Buildings?
But why should we care about buildings? Americans, on average, spend

approximately 90% of their time indoors based on an EPA [2009] sur-
vey. Commercial Buildings Energy Consumption Survey (CBECS) es-
timates that there were 5.6 million commercial buildings in the United
States in 2012, comprising 87 billion square feet of floor-space. This
level represents a 14% increase in the number of buildings and a 21%
increase in floor-space since 2003, the last year for which CBECS results
are available. Between the first CBECS (conducted in 1979) and the
latest 2012 CBECS, the number of commercial buildings in the United
States has increased from 3.8 million to 5.6 million, and the amount of
commercial floor-space has increased from 51 billion to 87 billion square
feet. On the residential side, nearly 130 million residential housing units
existed in the U.S. in 2010. Approximately 7.188 million new housing
units were built between 2005 and 2009, according to AHS [September
2008]. The total primary energy consumption in the United States in-
creased from 35 quads1 in 1950 to 78.3 quads in 1980 to over 98.5 quads
in 2014 as shown in Figure 1.1. In 2014 the building sector accounts for
39.87% of this total consumption according to Figure 1.2. The indus-
trial and transportation sectors represent the remaining 31.33% and
27.12%. Electrical energy consumption of buildings doubled in the last
1
A quad is a unit of energy equal to 1.055 × 1018 joules.
1.2. Why Smart Buildings? 5
18 years, and another 25% growth is projected through 2030. Residen-

tial buildings accounted for 54.6% of the total energy consumption in
the building sector, while commercial buildings accounted for the other
45.4%. The building sector is also responsible for almost 40% of green-
house gas emissions and 70% of electricity use. The energy consumption
by Heating Ventilation and Air Conditioning (HVAC) systems is 50%
of the total energy usage in buildings and 20% of the total national en-
ergy usage in European and American countries Pérez-Lombard et al.
[2008]. HVAC energy consumption can exceed 50% of the total energy
usage of a building in tropical climate Chua et al. [2013].
The industrial sector has always been optimizing its processes to
reduce cost and increase profit. In the transportation sector, in the last
30 years a great amount of work has gone into emission and fuel con-
sumption reduction via better engine control, and efforts are already
well under way to find suitable alternatives to oil. Bio-fuels are one
possibility. Alternative types of vehicles – hybrids, electric vehicles, and
vehicles powered by hydrogen fuel cells, for example – all have the goal
of reducing our dependence on oil. The Corporate Average Fuel Econ-
omy (CAFE) standards, initially adopted in 1975, made more stringent
in 2007, and strengthened again in pending legislation, require auto-
mobile manufacturers to build cars with higher average fuel economy.
On the other hand, historically, not much has been done to improve
the energy efficiency of buildings.
Growth in population, increasing demand for building services and
comfort levels, together with the rise in time spent inside buildings,
assure that the upward trend in energy demand will continue in the fu-
ture. For this reason, energy efficiency in buildings is a prime objective
today for energy policy at regional, national and international levels.
1.2 Why Smart Buildings?
Given that we spend on average more than 90% of our time in buildings
and the fact that 40% of total energy consumption is being consumed
in buildings, it is crucial that these systems are safe and comfortable
while consuming the minimum amount possible of energy. In order to
6 Introduction
U.S. Energy Flow, 2014

(Quadrillion Btu)
1 10
Includes lease condensate. Includes 0.16 quadrillion Btu of electricity net imports.
2 11
Natural gas plant liquids. Total energy consumption, which is the sum of primary energy consumption, electricity retail
3
Conventional hydroelectric power, biomass, geothermal, solar/photovoltaic, and wind. sales, and electrical system energy losses. Losses are allocated to the end-use sectors in
4
Crude oil and petroleum products. Includes imports into the Strategic Petroleum Reserve. proportion to each sector’s share of total electricity retail sales. See Note 1, “Electrical Systems
5
Natural gas, coal, coal coke, biofuels, and electricity. Energy Losses,” at the end of U.S. Energy Information Administration, Monthly Energy Review
6 (March 2015), Section 2.
Adjustments, losses, and unaccounted for.
7 Notes: • Data are preliminary. • Values are derived from source data prior to rounding for
Natural gas only; excludes supplemental gaseous fuels. publication. • Totals may not equal sum of components due to independent rounding.
8
Petroleum products, including natural gas plant liquids, and crude oil burned as fuel. Sources: U.S. Energy Information Administration, Monthly Energy Review (March 2015),
9
Includes -0.02 quadrillion Btu of coal coke net imports. Tables 1.1, 1.2, 1.3, 1.4a, 1.4b, and 2.1.
Figure 1.2: US Energy Flow in 2014 (Quadrillion Btu) EIA.
achieve these objectives, we need to make buildings smart about the

way they operate. Studies such as the American Housing Survey for
the United States by EnergySTAR, have shown that 30% of energy
consumption of commercial buildings is wasted and could be saved by
continuously monitoring and adjusting operations of these buildings.
Achieving safety, energy efficiency and comfort is only feasible if all
subsystems in the building continuously sense the environment, com-
municate between different parts of the system and make the right
decision both individually and collectively.
Buildings of the future are perceived as entities for real-time en-
ergy trading, as opposed to passive energy consumers. In this scenario,
buildings not only need to be aware of and responsive to the internal
conditions, but also need to be able to operate their subsystems (e.g.
HVAC, and lighting) in coordination with the grid. Real-time pricing
1.2. Why Smart Buildings? 7
Residential Sector Energy Consumption
Computers
Wet Cleaning 1.7%
3.4% Cooking
3.1% Other
Electronics
10.3%
2.9%
Refrigeration Space Heating
4.0% 44.2%
Space Cooling
8.9%
Lighting Water Heating
4.6% 16.9%
2015 Residential Energy End‐Use Splits
Figure 1.3: Breakdown of energy consumption in a typical building. Over 50% of

energy consumption is related to HVAC systems DOE [2013].
combined with the intelligence of the Building Energy Management

System (BEMS) to operate the building in a cost-effective way, is an
example of such scenarios. More sophisticated scenarios would involve
buildings operating in a cost-effective way given not only the real-time
energy prices, but also rewards that a utility or system operator may
offer buildings to provide flexibility in their energy consumption. The
latter scenario would require a fundamentally different building control
design; operating a system in the most cost-effective manner does not
typically lead to much flexibility around the operating trajectory. In a
scenario where the objective is defined not only by the goal of reducing
energy cost, but also by the reward for operating in certain regions, the
optimization problem becomes multi-objective and nontrivial.
8 Introduction
1.3 Areas of research
About 50% of the energy consumed in buildings is directly related to

space heating, cooling and ventilation (Figure 1.3). As such, reducing
building energy consumption by designing smart control systems to op-
erate the HVAC system in a more efficient way is critically important
to address the worldwide energy and environmental concerns. With
the advent of smart, easily-controllable thermostats, smart meters, and
two-way communication infrastructure between occupants and build-
ings via smart devices such as phones and next generation connected
electric cars, as well as between the buildings as consumers of energy
and utility companies as providers of energy, the role of buildings in
the operation of the smart grid will be even more significant compared
to the current state-of-the-art.
In the last decade a significant amount of work has been done in ar-
eas that, directly or indirectly, have contributed to achieving improved
performance, reliability and efficiency of buildings. We categorize this
work into the following areas:
• Simulation tools;
• Building models;
• Building control design;
• Test-beds and real-scale experiments;
• Buildings as cyber-physical systems;
• Smart buildings in the smart grid ecosystem.
In this article, we provide an overview of what has been achieved

in each of these areas, and we highlight emerging or existing areas for
research.
1.4 Organization
The remaining chapters of this article are organized as follows:

1.4. Organization 9
1.4.1 Chapter 2: Simulation Tools
We start by reviewing the simulation tools that have been developed

over the years in Chapter 2. We cover EnergyPlus and Modelica Li-
braries among other building simulation tools.
1.4.2 Chapter 3: Building Models
We then present our work in modeling buildings. In particular, we

first present Resistor-Capacitor (RC) models, which are the building
blocks of the majority of building simulation tools. Next, we show how
we use related available information from additional sensors such as
CO2 sensors, outside air temperature, and Global Horizontal Irradiance
(GHI) to infer quantities that are not measured, such as internal and
external heat gains and un-modeled dynamics.
Furthermore, we show how the proposed modeling framework can
be enhanced by introducing a Parameter-Adaptive Building (PAB)
model. The proposed PAB model leverages a Kalman filter-based
state estimation algorithm to simultaneously estimate the states and
parameters of the system, resulting in a parameter-varying model.
1.4.3 Chapter 4: Building Control Design
We first provide an overview of classical building HVAC controllers. We

then present a hierarchical control scheme in which the high-level con-
troller optimizes a cost function and sends the optimal set-point to the
local low-level PID controllers. The majority of Chapter 4 is devoted
to obtaining and studying Model Predictive Control (MPC), Robust
Model Predictive Control (RMPC), Stochastic Model Predictive Con-
trol (SMPC), and Exergy-based Model Predictive Control (XMPC),
and studying the performance of each in the presence of model uncer-
tainty. At the end of this chapter we provide a guideline for selecting
the most appropriate control strategy based on the accuracy of the
building model.
10 Introduction
1.4.4 Chapter 5: Test-beds and Real-scale Experiments
In Chapter 5 we review some of the outstanding efforts in this domain,

and present some findings on how effective new control techniques are
when implemented on real buildings. We focus on real-scale implemen-
tation of novel control algorithms on buildings and classify the studies
according to the system that was controlled (e.g. whole building, test
cell), the actuators, the total experiment time, and the MPC model.
1.4.5 Chapter 6: Designing Building Control Systems as Cyber-

Physical Systems
After presenting various control strategies in Chapter 4, and reviewing

real-scale implementation of such algorithms on real, physical buildings,
we present a framework to co-design the control algorithm and the em-
bedded platform for building HVAC systems in Chapter 6, thus treat-
ing a building as a cyber-physical system. As complex cyber-physical
systems, HVAC systems involve three closely related subsystems – the
control algorithm, the physical environment and the embedded imple-
mentation platform. In this chapter, we propose a co-design approach
that analyzes the interaction between the control algorithm and the
embedded platform through a set of interface variables, in particular
the sensing accuracy. Based on the proposed models, we explore the
design space of the control algorithm and the embedded platform to
optimize a system with respect to energy cost and monetary cost while
satisfying the constraints for user comfort level.
1.4.6 Chapter 7: Dynamic Contracts for Building-Grid Interaction
In Chapter 7, we address the future role of smart buildings in the con-

text of the smart grid. We first propose a means to define and quantify
the flexibility of a commercial building. We then propose a contractual
framework that could be used by building operators and utility com-
panies to declare flexibility on one side and reward structure on the
other. Subsequently, we design a control mechanism for the building to
decide its flexibility for the next contractual period to maximize the
reward, given the contractual framework. Finally, we perform at-scale
1.4. Organization 11
experiments to demonstrate the feasibility of the proposed algorithm.
1.4.7 Chapter 8: Conclusion

Finally, Chapter 8 draws the conclusions of the article with a discussion
on the possible directions for future work.
2
Simulation Tools
A variety of simulation tools have been developed for building energy

analysis, simulation, control design and building design automation.
Here we review some of the outstanding works in this domain. We
present building simulation tools, and building-to-grid simulation tools.
2.1 Building Simulation Tools
2.1.1 EnergyPlus
EnergyPlus is a whole building energy simulation software supported

by the U.S. Department Of Energy (DOE) for modeling building heat-
ing, cooling, lighting, ventilation, other energy flows, and water use.
Modeling the performance of a building with EnergyPlus enables build-
ing professionals to optimize the building design for energy and water
usage. EnergyPlus is typically used by engineers, architects, and re-
searchers to perform building energy analysis. It is a good tool to an-
alyze energy usage of a building with a particular design or feature, or
to compare the performance of two different designs or two different
set of features for a given building.
Initially, the U.S. government was developing two different software
12
2.1. Building Simulation Tools 13
tools, BLAST and DOE-2, which were both abandoned after many dis-
cussions but represented a first step and the working basis for Energy-
Plus. EnergyPlus ultimately has the features and capabilities of BLAST
and DOE-2, however it is an entirely new software tool that combines
the heat balance of BLAST with a generic HVAC system Sousa [2012].
EnergyPlus includes many simulation capabilities such as fine time-
steps (i.e. less than an hour), modular systems and plant integrated
with heat balance-based zone simulation, multi-zone air flow, thermal
comfort, water use, natural ventilation, and photovoltaic systems.
EnergyPlus handles integrated, simultaneous solution where the
building response and the primary and secondary systems are tightly
coupled, and it performs iteration when necessary. It also supports
sub-hourly, user-definable time steps for the interaction between the
thermal zones and the environment, and variable time steps for inter-
actions between the thermal zones and the HVAC systems. Energy-
Plus supports ASCII text-based weather, input, and output files in-
cluding hourly or sub-hourly environmental conditions, standard and
user definable reports. It leverages heat balance-based techniques for
building thermal loads that allow for simultaneous calculation of radi-
ant and convective effects at both interior and exterior surface during
each time step. Transient heat conduction through building elements
such as walls, roofs, floors, etc. using conduction transfer functions
is also supported. With the improved ground heat transfer modeling
through links to three-dimensional finite difference ground models and
simplified analytical techniques, EnergyPlus provides a more accurate
model of the heat loss or gain through the building floors. EneryPlus
also features combined heat and mass transfer model that accounts for
moisture absorption/desorption. Thermal comfort models based on ac-
tivity, inside dry bulb, and humidity as well as anisotropic sky model
for improved calculation of diffuse solar on tilted surfaces are available.
Detailed and advanced fenestration calculations in EnergyPlus includes
controllable window blinds, electrochromic glazing, layer-by-layer heat
balances that allow proper assignment of solar energy absorbed by
window panes, and a performance library for numerous commercially
available windows. Daylighting controls including interior luminance
14 Simulation Tools
calculations, glare simulation and control, luminaire controls, and the

effect of reduced artificial lighting on heating and cooling are the other
features supported by EnergyPlus. Atmospheric pollution calculations
that predict CO2, SOx, NOx, CO, particulate matter, and hydrocar-
bon production for both on site and remote energy conversion are also
available for users.
EnergyPlus is regarded as a very detailed whole building simulation
tool. This makes EnergyPlus a good choice if the objective of the study
is to analyze and compare the effects of some building features. On the
other hand, being a very detailed modeling tool, EnergyPlus is not a
program of choice if the task is model-based control design. However,
sometimes to account for that problem a detailed EnergyPlus model is
linearized at some operating point and the linearized model is used for
control purposes as proposed by Ahuja et al. [2011], and Maasoumy
and Vincentelli [2014].
2.1.2 DOE-2
The DOE-2 software was developed by James J. Hirsch & Associates

(JJH) in collaboration with Lawrence Berkeley National Laboratory
(LBNL), with LBNL work performed mostly under funding from the
United States Department of Energy and other work performed mostly
under funding from a wide range of industry organizations.
DOE-2 in its standard form is a command line, batch-oriented pro-
gram, for which one creates input files, using a text editor, with the
building description in DOE-2 building description language (BDL).
The learning curve on raw DOE-2 is 6-12 months and it is very difficult
to master. DOE-2.1E comes in two versions: the standard DOE-2.1E
(pre-110 version numbers) developed in collaboration with LBNL’s
Simulation Research Group, and the enhanced DOE-2.1E (post-110
version numbers) which incorporates many bug fixes and new features
that formed the early initial implementation of a subset of features now
found in DOE-2.2. DOE-2.2 is much changed and upgraded from 2.1E.
Most of the extensive changes are in the HVAC area – particularly in
the central plant descriptions.
2.1.3 eQUEST
eQUEST is a freeware, complete, fully interactive, professional DOE-

2.2 simulation environment based upon technology first developed for
PowerDOE then enhanced for use in the current generation of Win-
dows operating system versions. eQUEST, the QUick Energy Simula-
tion Tool, is a comprehensive building energy simulation tool that can
be used by all design team members at all design phases, including
schematic design. eQUEST is designed to allow the user to perform
detailed analysis of today’s state-of-the-art building design using most
sophisticated building energy use simulation techniques. This is accom-
plished by combining a building creation wizard, an energy efficiency
measure (EEM) wizard, and graphical reporting with a simulation “en-
gine” derived from the latest version of DOE-2.
eQUEST features include support for complex geometries, sun-
spaces and trombe walls, custom glazing, daylighting with optional sun
control, various methods for air infiltration, more than 50 HVAC sys-
tem configurations available in wizards, and some more detailed modes
such as flexible HVAC equipment control strategies, ground-source heat
pumps, and dual-fuel cooling plants. On the other hand, its limitations
include daylighting limitations such as unreliable calculations for light
shelves, skylights with deep wells, and rooms with internal obstructions,
and support of daylighting only calculated through exterior windows.
Natural ventilation modeling is supported only for residential and single
zone type systems. Inter-zonal air flow and calculation of air stratifi-
cation within zones are not supported, and at maximum one HVAC
system per thermal zone is allowed. eQUEST accepts BIN weather file
format. It also can convert from other formats (e.g. from EnergyPlus
EPW format or from a text file)
eQUEST is supported as a part of the Energy Design Resources pro-
gram which is funded by California utility customers and administered
by Pacific Gas and Electric Company, San Diego Gas and Electric, and
Southern California Edison, under the auspices of the California Public
Utilities Commission.
16 Simulation Tools
Figure 2.1: Graphical view of the Modelica model that computes the change in
temperature for two simple room models, as shown in Modelica.
2.1.4 Modelica
Modelica is an equation-based modeling language that is well

positioned to become the de-facto open standard for modeling
multi-physics dynamical systems. It already has a successful track
record in many industrial-scale multi-physics applications (see also
http://www.modelica.org/publications). Dymola is utilized as a
compiler for the Modelica language, which allows the model develop-
ment to focus on the mathematical description of the physical phenom-
ena rather than the solution of simultaneous equations. This approach
provides increased flexibility in integrating models for various com-
ponents of the system. For control purposes, the Modelica models are
either recoded in MATLAB and C or sampled to provide look up tables
for the rapid calculations required by the MPC algorithm.
Modelica buildings library by Wetter et al. [2014], is a free open-
source library that is implemented in Modelica. The library developed
and maintained by LBNL, supports rapid prototyping as well as design
and operation of building energy and control systems. The scope of the
library covers heating, ventilation and air-conditioning systems, multi-
zone heat transfer and multi-zone airflow and contaminant transport.

The class hierarchy used in the library allows implementing component
models by extending partial implementations of base models of heat
and mass exchangers, and by instantiating basic models for conserva-
tion equations and flow resistances. LBNL has also built associated
tools for pre- and post-processing, regression tests, co-simulation and
real-time data exchange with building automation systems. An exam-
ple of such Modelica models is shown in Figure 2.1.
2.1.5 TRNSYS
TRNSYS is a TRaNsient SYstems Simulation Program with a modular

structure. It recognizes a system description language in which the user
specifies the components that constitute the system and the manner in
which they are connected. The TRNSYS library includes many of the
components commonly found in thermal and electrical energy systems,
as well as component routines to handle input of weather data or other
time-dependent forcing functions and output of simulation results. The
modular nature of TRNSYS gives the program tremendous flexibility,
and facilitates the addition to the program of mathematical models
not included in the standard TRNSYS library. TRNSYS is well suited
to detailed analyses of any system whose behavior is dependent on the
passage of time. Main applications include: solar systems (solar thermal
and photovoltaic systems), low energy buildings and HVAC systems,
renewable energy systems, cogeneration, and fuel cells.
2.1.6 ESPr
ESPr is a general-purpose simulation environment, which supports an

in-depth appraisal of the factors which influence the energy and en-
vironmental performance of buildings. The ESPr system has been the
subject of sustained developments since 1974 and in 2002 converted to
the GNU Public License.
ESPr simulates building performance in a realistic manner that
adheres closely to actual physical systems. It also supports early and
detailed design stage appraisals, and enables integrated performance
18 Simulation Tools
assessments in which no single issue is unduly prominent. By address-

ing all aspects simultaneously, ESPr allows the designer to explore the
complex relationships between a building’s form, fabric, air flow, plant
and control. ESPr is based on a finite volume, conservation approach in
which a problem (specified in terms of geometry, construction, opera-
tion, leakage distribution, etc.) is transformed into a set of conservation
equations (for energy, mass, momentum, etc.) which are then integrated
at successive time-steps in response to climate, occupant and control
system influences. In addition to state of the art standard simulation
features, ESPr has powerful capability to simulate scenarios including
daylight utilization, natural ventilation, contaminant distribution, com-
bined heat and electrical power generation and photovoltaic facades,
adaptive 3D transient CFD, multi-gridding (2D and 3D conduction),
and control systems.
ESPr comprises a central project manager around which are ar-
ranged support databases, a simulator, various performance assessment
tools and a variety of third party applications for CAD, visualization
and report generation.
Users can export an ESPr model with materials, constructions, sur-
faces, all three and four sided surfaces as well as those including one
window or one door, to EnergyPlus. Boundary condition attributes
are translated and the parent/child relationship between opaque and
transparent surfaces are established.
However, specialist features require knowledge of the particular sub-
ject. Although robust and increasingly used for consulting, ESPr retains
much of the look and feel of a research tool and lacks the extensive
databases associated with commercial tools. The current Windows im-
plementation does not conform to the standard look and feel of most
Windows applications and lacks a few features available on other plat-
forms.
2.1.7 BRCM
Building Resistance-Capacitance Modeling (BRCM) is a Matlab Tool-

box that facilitates the physical modeling of buildings for model-
based control design. The toolbox provides a means for the fast gen-
2.2. Building-to-Grid Simulation Tools 19
eration of (bi-)linear resistance-capacitance type models from basic

building geometry, construction and systems data. Moreover, it sup-
ports the generation of costs and constraints. The toolbox is based
on previously validated modeling principles. The toolbox itself, the
details of the modeling and the documentation can be found at
www.brcm.ethz.ch Sturzenegger et al. [2014].
2.1.8 MLE+
MLE+ by Bernal et al. [2012] is a tool for energy-efficient building
automation design, co-simulation and analysis. The tool leverages the
high-fidelity building simulation capabilities of EnergyPlus and the sci-
entific computation and design capabilities of Matlab for controller de-
sign. MLE+ facilitates integrated building simulation and controller
design with integrated support for system identification, control de-
sign, optimization, simulation analysis and communication between
software applications and building equipment. It provides streamlined
workflows, a graphical front-end, and debugging support to help con-
trol engineers eliminate design and programming errors and take in-
formed decisions early in the design stage, leading to fewer iterations
in the building automation development cycle. Bernal et al. [2012] show
through an example and two case studies how MLE+ can be used for
designing energy-efficient control algorithms for both simulated build-
ings in EnergyPlus and real building equipment via BACnet.
2.2 Building-to-Grid Simulation Tools
2.2.1 HomeSim / S2 Sim

HomeSim, developed in University of California, San Diego (UCSD)
by Venkatesh et al. [2013] is a residential electrical energy simulation
platform that enables investigating the impact of technologies such
as renewable energy and different battery types. Additionally, Home-
Sim allows simulating different scenarios including centralized vs. dis-
tributed in-home energy storage, intelligent appliance rescheduling, and
outage management. Using measured residential data, HomeSim quan-
tifies different benefits for different technologies and scenarios, includ-
20 Simulation Tools
ing up to 50% reduction in grid energy through a combination of dis-

tributed batteries and reschedulable appliances.
Smart Grid Swarm Simulator (S2Sim), another tool from UCSD
by Aksanli et al. [2014] was developed to study the grid dynamics and
demonstrate how distributed smart building controllers can threaten
the stability of the grid. S2Sim is built to study the effects of smart
building controllers on the grid. It is based on the UCSD Microgrid
circuit structure as the baseline, and hosts distributed smart buildings
and energy generation.
2.2.2 VirGIL
Co-simulation platforms are necessary to study the interactions of com-

plex systems integrated in future smart grids. The Virtual Grid Inte-
gration Laboratory (VirGIL) proposed by Chatzivasileiadis et al. [2015]
is a modular co-simulation platform designed to study interactions be-
tween demand response strategies, building comfort, communication
networks, and power system operation. Chatzivasileiadis et al. [2015]
present the coupling of power systems, buildings, communications and
control under a master algorithm. First, they propose a modular archi-
tecture for VirGIL, based on the Functional Mock-up Interface (FMI),
where several different modules can be added, exchanged, and tested.
Second, they propose a commercial power system simulation platform,
familiar to power system operators. VirGIL further introduces the in-
tegration of the Quantized State System (QSS) methods for simulation
in this co-simulation platform.
2.3 Comparisons
Each of the building simulation software tools mentioned above has

certain characteristics, and suits best to specific applications. In order
to better understand specific features of each one, Table 2.1 presents a
summary table of the features of each of the simulation tools mentioned
above, where we summarize the building simulation tools introduced
in this section using metrics such as modeling approach, interior sur-
face convection, outside surface convection, solar gain, data exchange,
2.4. Concluding Remarks 21
libraries, control capability, and additional capabilities and defects. Ref-

erences for the following comparisons are from Nouidui [2014], Crawley
et al. [2008], Sousa [2012].
2.4 Concluding Remarks
Energy simulation software tools for buildings have been developed

over the years. Currently there are several energy simulation software
tools with different levels of complexity and response to different sim-
ulation scenarios. Among the most mature simulation software tools
are the EnergyPlus, eQUEST, ESPr, TRNSYS, and Modelica. Being
the most complete software tools, these are also the most complex and
therefore require greater expertise. From the analyzed energy simula-
tion software tools, TRNSYS is the most complete, but depending on
the user perspective and final purpose the other software tools could be
more appropriate. The major limitation of TRNSYS is not being able
to connect with the AutoCad software tool for importation and expor-
tation of files. In this aspect EnergyPlus, and ESPr are more appropri-
ate. Modelica, on the other side, is a non-proprietary, object-oriented,
equation based language to model complex physical systems contain-
ing, e.g., mechanical, electrical, electronic, hydraulic, thermal, control,
electric power or process-oriented subcomponents, which makes Mod-
elica unique in the way that it supports a wide variety of systems and
is not just limited to building simulations.
Simulation Modeling Interior Outside Solar gain Data Libraries Control Additional
Tools approach surface surface exchange capability capabilities /
convection convection defects
EnergyPlus heat-balance dependent on DOE-2, inside radiation input extensive via building whole build-
equations temperature, ASHRAE, view factors, solar weather HVAC controls virtual ing analysis,
air flow, surface BLAST, gain from environ- files libraries test bed water usage
heat coefficient, MoWiTT, ment, daylighting, analysis
or user-defined user-defined radiation to air
DOE-2 command user-defined DOE-2 - text input extensive responsive to long learning
line, batch- files HVAC climate, and curve
oriented libraries occupant
program
eQUEST dependent NA DOE-2 radiation to air building walls and responsive to schematic
on DEO2 geometry, windows climate, and design
weather materials occupant
TRNSYS transient dependent on user-defined solar gain from handle low-energy responsive to support for
simulation temperature, environment, day- input of building climate, and renewable,
air flow, or lighting, radiation weather and HVAC occupant cogeneration,
user-defined to air data systems and fuel cells
ESPr finite vol- dependent on MoWiTT, inside radiation building HVAC and responsive to -
ume, energy, temperature, Ito, Kimura, view factors, solar geometry, building climate, oc-
mass, etc. air flow, surface and Oka gain, daylighting, models component cupant and
conservation heat coefficient, correlation, radiation to air export to libraries control system
or user-defined user-defined EnergyPlus
Modelica equation- constant or constant solar radiation, real-time building simple control -
based temperature- coefficient, infrared radiation data ex- and HVAC algorithms
object- dependent dependent from ambient change with
oriented coefficient on wind- environment BAS
modeling speed, wind
language direction and
temperature
BRCM RC modeling constant coeffi- constant coef- - built on - control design -
cient ficient MATLAB capability in
MATLAB
MLE+ RC modeling constant coeffi- constant coef- - MATLAB, - control design system identi-
cient ficient EnergyPlus capability in fication
MATLAB
Table 2.1: Comparison of simulation tools.

3
Building Models
For the simulation and control of buildings, models of appropriate ac-

curacy and complexity are essential. In this chapter, we present our
work in this domain. In particular, we first present RC models, that
are the building blocks of the majority of building simulation tools.
Unmodelled dynamics such as internal and external heat gains are
random, time-varying and hard to accurately model and estimate. We
use an affine mapping from measurable quantities such as CO2 con-
centration level and outside air temperature, respectively, to obtain
an estimate of such unmodelled dynamics. Using historical data, we
calibrate the model, and show the effectiveness of this methodology
through modeling a real building and comparing the results with mea-
surements.
This modeling framework works well for most buildings. However,
the model development process is not easily scalable. From one building
to another, one needs to repeat all the training and calibration. Fur-
thermore, the modeling framework of Section 3.1 of this chapter does
not address the time-varying behavior of the building and its indoor
and outdoor environment. We improve this modeling framework by in-
troducing a PAB model. The architecture of the PAB model and its
23
24 Building Models
components including its Kalman filter-based estimation algorithm are

presented, and the effectiveness of the proposed approach is validated
via simulations.
3.1 Resistor-Capacitor Models
The American Society of Heating, Refrigerating, and Air-Conditioning

Engineers (ASHRAE) has established extensive methodologies for cal-
culation of heating and cooling loads in its Handbook [2009]. Such
methodologies involve heat balance analysis. It is a straight-forward
physics-based and comprehensive methodology that involves calculat-
ing a node-to-node heat balance of a room through consideration of con-
ductive, convective, and radiative heat transfer mechanisms, a node can
either be wall or air in the room. Such cooling and heating load calcu-
lations are greatly facilitated by Resistance-Capacitance (RC) models.
RC modeling allows for better understanding of problem physics and
makes it possible to easily evaluate modeling hypotheses and sensitivity
to different system parameters Bueno et al. [2012].
We present high-fidelity models of building thermal dynamics,
which are obtained using fundamental laws of Thermodynamics and
Heat Transfer. We model the conductive and convective heat transfer
phenomena, which take place in the walls and between the walls and
the environment, respectively. We obtain an RC model of the whole
building. Each wall and room in the building is modeled as a node
with a designated capacitor, and these nodes are connected using resis-
tances. Historical data is then used to calibrate the model and find the
optimum set of parameters for the model. In order to integrate internal
and external disturbances to the model, such as heat generated from
occupancy and heat from sun radiation, CO2 concentration level and
outside temperature from sensors are used, respectively. We call these
two quantities unmodelled dynamics.
3.1.1 Mathematical Modeling
Figure 3.1 depicts the schematic of a typical room. We use lumped

model analysis as proposed in Bergman et al. [John Wiley & Sons In-
3.1. Resistor-Capacitor Models 25
Figure 3.1: Schematic of a typical room with a window. Temperature sensors are
denoted by “S” in this figure.
corporated, 2011] to reduce the complexity of the model, and obtain

a low order model, suitable for control purposes. As a simplifying as-
sumption, temperature is considered uniform inside the room. We use
RC model from Maasoumy et al. [2011] in which the building is consid-
ered as a network of interconnected nodes. We account for time-varying
parameters, such as convective heat transfer coefficient of outside air,
by updating and learning model parameters on the fly.
26 Building Models
Heat Transfer
There are two types of nodes in the building network: walls and rooms.
Consider in total n nodes, m of which represent rooms and the remain-
ing n − m nodes represent walls. We denote the temperature of room
i with Tri . The wall node and temperature of the wall between room i
and j are denoted by (i, j) and Twi,j , respectively. Thermal dynamics
of nodes is governed by the following equation:
Wall Temperature Dynamics
w dTwi,j X Trk − Twi,j
Ci,j = + ri,j αi,j Awi,j Qradi,j (3.1)
dt Ri,jk
k∈Nwi,j
where Ci,jw, α
i,j and Awi,j are heat capacity, radiative heat absorption
coefficient and area of wall between room i and j, respectively. Ri,jk
is the total thermal resistance between the centerline of wall (i, j) and
the side of the wall where node k is located. Qradi,j is the radiative heat
flux density on wall (i, j). Nwi,j is the set of all of neighboring nodes to
node wi,j . ri,j is wall identifier which is equal to 0 for internal walls, and
equal to 1 for peripheral walls (i.e. either i or j is an outside node). In
Equation (3.1) the left term denotes the rate of change of stored heat
in the wall between room i and room j. The first term of the right hand
side of this equation represents the flow of heat between room k and
wall (i, j) due to temperature difference and the second term shows the
heat flow to the wall, due to solar radiation. Temperature dynamics of
the ith room is modeled by the following equation:
Air Temperature Dynamics
dTri X Tk − Tr
Cir = i
+ ṁri ca (Tsi − Tri ) + wi τwi Awini Qradi + Q̇inti (3.2)
dt Ri,ki
k∈Nri
where Tri , Cir and ṁri are the temperature, heat capacity and air mass
flow into room i, respectively. ca is the specific heat capacity of air,
and Tsi is the temperature of the supply air into room i. πi is window
identifier which is equal to 0 if none of the walls surrounding room i

have a window, and is equal to 1 if at least one of them has a window.
τwi is the transmissivity of glass of window i, Awini is the total area
of windows on walls surrounding room i, Qradi is the radiative heat
flux density per unit area radiated to room i, and Q̇inti is the internal
heat generation in room i. Nri is the set of all of the neighboring room
nodes to room i. In Equation (3.2) the left term denotes the rate of
change of stored heat in the air in room i. The first term of the right
hand side of this equation represents the flow of heat between node k
and room i due to temperature difference, the second term shows the
heat flow delivered by the heating system, the third term represents
the total radiative heat passing through the windows and the fourth
term is the internal heat generation inside room i. More details of
building thermal modeling and estimation of the unmodelled dynamics
is available in Maasoumy [2011], Maasoumy et al. [2011], Maasoumy
and Sangiovanni-Vincentelli [2012].
We approximate the values of Qradi (t) and Q̇int (t) based on the
following equations:
Qradi (t) = τ Tout (t) + ζ (3.3)
Q̇int (t) = µΨ(t) + ν (3.4)
where Tout and Ψ are the outside air temperature and CO2 concentra-
tion in the room, respectively Maasoumy Haghighi [2013]. Air venti-
lation is considered constant as a simplifying assumption. A more so-
phisticated model for gas transport process in buildings can be found
in Federspiel [vol. 5, no. 5, pp. 480-489, 1997]. Parameters τ , ζ, µ and
ν are obtained by the parameter estimation algorithm detailed in Sec-
tion 3.2.2.
We model the radiative heat transfer between building and ambient
environment as proposed in Goforth et al. [vol. 4710, no. 1, pp. 203-213,
Orlando, FL, 2002]. The amount of heat transferred from the building
to the environment is given by the Stefan-Boltzmann law:
4
Qbldg = σTbldg (3.5)
where Tbldg is the average temperature of the building. We also con-
sider solar radiation heat transfer, Qsolar absorbed by the walls, and
28 Building Models
the room through the windows. The data used here is based on the
past 30 years monthly average of solar radiation for flat-plate collec-
tors facing south (resembling the south facing flat vertical walls of the
building), and is obtained from NREL (National Renewable Energy
Laboratory) NRE [May 2015] database for Houghton, MI in January.
Furthermore, we take into account the radiation cooling at night (i.e.
sky thermal radiation to the building) based on the proposed relation
in Goforth et al. [vol. 4710, no. 1, pp. 203-213, Orlando, FL, 2002]:
Qsky = (1 + KC 2 )8.78 × 10−13 Tout
5.852
RH 0.07195 (3.6)
where K is the coefficient related to the cloud height and C is a function
of cloud coverage. We use K = 0.34 and C = 0.8 for simulations, based
on the results in Goforth et al. [vol. 4710, no. 1, pp. 203-213, Orlando,
FL, 2002]. Tout is the outside air temperature, and RH is the air relative
humidity percentage. The total radiation exchange between building
and ambient environment is then given by:
Qrad = Qsky + Qsolar − Qbldg (3.7)
Note that Qsky and Qsolar are heat flow into the building, and Qbldg ,
is the heat flow from the building to the environment.
System Dynamics
Heat transfer equations for walls and rooms yield the following system
dynamics:
ẋt = f (xt , ut , dt , t)
yt = Cxt (3.8)
where xt ∈ Rn is the state vector representing the temperature of the
nodes in the thermal network, ut ∈ Rl×m is the input vector represent-
ing the air mass flow rate and discharge air temperature of conditioned
air into each thermal zone, and yt ∈ Rm is the output vector of the
system which represents the temperature of the thermal zones. l is the
number of inputs to each thermal zone (e.g., two for air mass flow and
supply air temperature). C is a matrix of proper dimension and the
disturbance vector is given by dt = g(Qradi (t), Q̇int (t), Tout (t)).
Disturbance
Following the intuitive linear relation between outside temperature
Tout , internal heat generation Q̇int , and solar radiation Qrad , with the
building internal temperature rise we approximate g with an affine
function of these quantities, leading to:
dt = aQradi (t) + bQ̇int (t) + cTout (t) + e (3.9)
where a, b, c, e are constants to be estimated. By substituting Equa-

tions (3.3) and (3.4) into Equation (3.9) and rearranging the terms, we
get:
dt = (aτ + c)Tout (t) + bµΨ(t) + aζ + bν + e

(3.10)
= āTout (t) + b̄Ψ(t) + ē
where ā = aτ + c, b̄ = bµ, and ē = aζ + bν + e. Therefore, only

measurements of outside air temperature and CO2 concentration levels
are needed to determine the disturbance to the model. The values of
ā, b̄, and ē are estimated along with other parameters of the model.
In what follows we use linearized system dynamics for control de-
sign. However, the original non-linear model is used for state estima-
tion, filtering and as the plant to compute the actual temperature evo-
lution. System dynamics is linearized around the nearest equilibrium
point by starting from an initial point and searching, using a Sequen-
tial Quadratic Programming (SQP) algorithm, until it finds the nearest
equilibrium point to the specified operating point of the system (de-
tails in Maasoumy [2011]). Discretizing the state space realization using
zero-order hold leads to the following discrete time LTI system:
xk+1 = Axk + Buk + E dˆk

yk = Cxk (3.11)
3.1.2 Model Uncertainty

Modeling and system identification are the most challenging and
time-consuming parts of building predictive control Privara et al.
[2013]. Model uncertainty, specifically, is a significant challenge to
30 Building Models
more widespread use of MPC for optimizing building energy consump-

tion. Incorporating unmodelled dynamics improves the accuracy of the
mathematical model of the building. To address this challenge, over
the last few years numerous mathematical models of building thermal
dynamics have been proposed in the literature. Resistor-capacitor (RC)
models with disturbances to capture unmodelled dynamics have been
proposed by Maasoumy [2011], Maasoumy and Sangiovanni-Vincentelli
[2012], Maasoumy et al. [2011]. Agbi et al. [pp. 6951-6956, Grand
Wailea Maui, HI, 2012] discovered that time-varying properties such
as occupancy can significantly change the dynamic thermal model and
influence how building models are identified. While modeling a multi-
zone building, Agbi et al. [pp. 6951-6956, Grand Wailea Maui, HI,
2012] observed that the experimental data often did not have sufficient
quality for system identification and therefore proposed a closed-loop
architecture for active system identification using prediction-error iden-
tification method (PEM). Other modeling techniques with application
in building predictive control include: subspace methods, MPC relevant
identification (MRI), deterministic semi-physical modeling (DSPM),
and probabilistic semi-physical modeling (PSPM). In a recent work
by Radecki and Hencey [2012], Unscented Kalman Filter (UKF) has
been used for online estimation of building thermal parameter estima-
tion.
However, the building is exposed to extremely time-varying phe-
nomena in the environment such as occupancy from inside, and solar
radiation and wind from outside. In order to account for such time-
varying phenomena, Maasoumy et al. [2014b] propose a PAB model
framework for online parameter adaptation and state estimation for
building models.
Buildings are dynamical systems with uncertain and time-varying
plant and occupant characteristics. The heat transfer characteristics
of a building are highly dependent on the ambient conditions. For in-
stance, heat transfer properties such as convective heat transfer coeffi-
cient h of peripheral walls is dependent on outside temperature, wind
speed and direction. Also, unmodelled dynamics of a building according
to Maasoumy and Sangiovanni-Vincentelli [2012] is function of
3.2. Parameter-Adaptive Building (PAB) Model 31
1. external factors: ambient weather conditions such as radiative

heat flux into the walls and windows, and cloudiness of the sky,
and
2. internal factors: such as occupancy level, internal heat genera-

tion from lighting, and electric or electronic devices.
These quantities are highly time-varying and therefore the dynam-

ics of the building and, consequently, parameters of the mathematical
model describing the dynamics of the buildings are constantly chang-
ing with time. Accordingly, the estimation algorithms used to identify
these parameters should take the time-varying aspect of buildings into
account and be adaptive in this respect.
Additive uncertainty
We linearize the original nonlinear dynamic system and use Euler’s
discretization method to obtain a linear discrete-time system. We also
add an additive uncertainty to the state update equation to account
for model uncertainties, leading to:
xk+1 = Axk + Buk + E(dk + wk ) (3.12)
where the uncertainty wk ∈ Rr is a stochastic additive disturbance.

t ∈ R refers to time in continuous-time domain and k ∈ Z refers to time
in discrete-time domain. The set of possible disturbance uncertainties
is denoted by W and wk ∈ W ∀k = 0, 1, ..., N − 1. For this study, we
consider box-constrained disturbance uncertainties given by
Wλ = {w : ||w||∞ ≤ λ} (3.13)
3.2 Parameter-Adaptive Building (PAB) Model
Building models proposed in the literature typically depend on many

parameters. This is due to the fact that buildings are composed of
many sub-systems and a variety of thermal mechanisms take place in
the building such as heat conduction through walls, forced convection
32 Building Models
Figure 3.2: Architecture of the building control based on the proposed PAB model
with its components.
due to air conditioning systems, and thermal radiation from outside.

A mathematical model that is descriptive enough to accurately ex-
plain these phenomena will end up with many time-varying parame-
ters. Finding the best parameters at each time step is shown to be
cumbersome as suggested by Agbi et al. [pp. 6951-6956, Grand Wailea
Maui, HI, 2012]. Maasoumy et al. [2014b] proposed a novel parameter
adaptive building (PAB) model that facilitates this parameter tuning
process in an online and automatic fashion. The architecture of the
PAB model is shown in Figure 3.2. Measurement data from various
sensors such as temperature and airflow are stored in a data reposi-
tory. The PAB model has a parameter update module which learns the
parameter on the fly. We present the algorithm in details later in this
section.
The PAB model works as follows: Historical data is used to per-
form off-line, one-step model calibration. The obtained parameters from
model calibration are used in the parameter update module (exploit-

ing Kalman filtering algorithm) as an initial set of parameters. Kalman
filter updates the parameters of the building model as the new mea-
surements arrive. The control module then uses the new updated set
of parameters for the next time step. The PAB model is developed in
the rest of this section.
3.2.1 State-Parameter Estimation

Using Equation (3.1) for each wall and Equation (3.2) for each room
node in the building network, system dynamics is given by:

1 1 1 1 1 1 x2
ẋ1 = · − − − − − ṁr1 ca x1 +
C1r R121 R131 R141 R151 win
R15 R121

x3 x4 x5 T5
+ + + + ca Ts1 ṁr1 + win + Awin τ Qrad + Q̇int1 (3.14a)
R131 R141 R151 R15

1 x1 1 1 T2
ẋ2 = w
. − + x2 + (3.14b)
C21 R211 R211 R212 R212

1 x1 1 1 T3
ẋ3 = w
. − + x3 + (3.14c)
C31 R311 R311 R313 R313

1 x1 1 1 T4
ẋ4 = w
. − + x4 + (3.14d)
C41 R411 R411 R414 R414

1 x1 1 1 T5
ẋ5 = w
. − + x5 + + Aw51 αQrad (3.14e)
C51 R511 R511 R515 R515
where x1 is the room temperature (i.e. Tr1 ), and x2 , x3 , x4 , x5 are

the peripheral walls’ temperatures (i.e. Tw12 , Tw13 , Tw14 , Tw15 ). T2 ,
T3 , T4 , T5 are the temperatures of the surrounding zones, as shown in
Figure 3.1 and Figure 3.3. These temperatures act as disturbance to
the system dynamics for a single zone thermal model, and x is the state
vector:
h iT
x = Tr1 , Tw12 , Tw13 , Tw14 , Tw15 (3.15)
One way to adapt the model to account for time varying parameters
is to assume that all the parameters of the model are independent, and
hence define a state corresponding to each parameter. However, this
would lead to an excessive number of states (e.g. 18 states for a room
34 Building Models
shown in Figure 3.1). To overcome this problem, we take a different ap-

proach. We reduce the number of states by exploiting the redundancies
in the resulting model. For instance, thermal properties of wall mate-
rial (e.g. specific heat capacity and conductive heat transfer coefficient)
are the same across the building, as these are functions of the mate-
rials used as the building walls. In addition, the thickness of internal
walls and thickness of peripheral walls are the same throughout the
building. Following this approach, we are able to reduce the number of
independent parameters from 18 to 10. Hence, we re-write the thermal
equations of the walls, i.e. Equations (3.14b)-(3.14e) as follows:
1 2 T2
ẋ2 = x1 − w w x2 + w w (3.16)
C w Rw C R C R
1 2 T3
ẋ3 = w w x1 − w w x3 + w w (3.17)
C R C R C R
1 2 T4
ẋ4 = w w x1 − w w x4 + w w (3.18)
C R C R C R
1 1 1

ẋ5 = w x1 − wR + w x5
C51 R511 C51 511 C51 R515
T5 Aw51 αQrad
+ w + w (3.19)
C51 R515 C51
As shown in Equations (3.20), Cw Rw is not a function of the wall

area:
Lw /2 1 cw L2w cw Lw
C w Rw = (cw Aw Lw )( + )= + (3.20)
k w Aw hin Aw 2kw hin
where cw , kw , Aw and Lw are the specific heat capacity, conductive

heat transfer coefficient of wall material, area and thickness of wall,
respectively, and hin is the indoor convective heat transfer coefficient.
Hence, we can use one common term to express thermal capacitance-
resistance between centerline of each wall and the node on each side of
the wall for the equations of walls in the building. We designate a state
variable to all the independent time-varying parameters of the system
as follows:
1 1
x6 = x7 = (3.21)
C1r R121 C1r R131
1 1
x8 = x9 = (3.22)
C1r R141 C1r R151
1 1
x10 = x11 = (3.23)
C1r Cw Rw
1 1
x12 = wR x13 = w (3.24)
C51 511 C51 R515
α 1
x14 = w x15 = win (3.25)
C51 R15
Rate of change of these states is equal to zero, as shown in the corre-
sponding state update Equation (3.31). We then add a low-magnitude
fictitious noise to the dynamics of parameters to allow slow changes in
their values over time.
State and Parameter Dynamics Model
ẋ1 = (x6 − x7 − x8 − x9 − x10 x15 − x10 u2 ca )x1 + x6 x2 + x7 x3

+ x8 x4 + x9 x5 + (ca u1 u2 + T5 x15 + Awin τ Qrad + Q̇int ).x10 (3.26)
ẋ2 = (x1 − 2x2 + T2 ).x11 (3.27)
ẋ3 = (x1 − 2x3 + T3 ).x11 (3.28)
ẋ4 = (x1 − 2x4 + T4 ).x11 (3.29)
w51
ẋ5 = x1 x12 − (x12 + x13 )x5 + T5 x13 + A x14 Qrad (3.30)
ẋi = 0 ∀i = 6, 7, ...15. (3.31)
u is the input vector given by:

" #
Ts1
u= (3.32)
ṁr1
In summary, we express the dynamics of the system using following
state update model:
xk = f (xk−1 , uk−1 , dk−1 , wk−1 )

zk = h(xk ) + vk (3.33)
36 Building Models
where wk and vk are the process and measurement noise and are as-
sumed to be zero-mean multivariate Gaussian process with variance
Wk and Vk , (i.e. wk ∼ N (0, Wk ) and vk ∼ N (0, Vk )), respectively.
30 30
25 25
20 20
Outdoor temperature (°C)
Indoor temperature (°C)

15 15
10 Tout 10
T2in
5 5
T3in
0 0
T4in
−5 −5
−10 −10
−15 −15
−20 −20
−25 −25
0 1 2 3 4 5 6 7 8 9 10 11 12 13
Time (Day)
Figure 3.3: Temperatures of neighboring zones acting as disturbance to the PAB

model.
3.2.2 Estimation Algorithm

In order to estimate the unknown parameters of the system we augment
the states of the system with a vector pk which stores the parameters
of the system, with a time evolution dynamics of pk+1 = pk . Due to the
multiplication of states and parameters the resulting dynamic system is
nonlinear. Nonlinear estimation algorithms such as Extended Kalman
Filtering (EKF) or Unscented Kalman Filtering (UKF) can then be
exploited to simultaneously estimate the states and the parameters of
the system.
An alternative to using a Kalman filter would be a simple observer.
However, given the random variations, inaccuracies and uncertainties
Figure 3.4: Inputs to the PAB model.
2.5
x6
x7
2
x8
Parameter value
x9
1.5
x10
x11
1
x12
x13
0.5 x14
x15
0
−0.5
0 1 2 3 4 5 6 7 8 9 10 11 12 13
Time (Day)
Figure 3.5: Estimated parameters of the PAB model using the designed UKF.
in the system dynamics, as described earlier, using a Kalman filter

is suggested in order to get a statistically optimal estimate of system
38 Building Models
states, as suggested by Julier and Uhlmann [pp. 182-193, Seattle, WA,

1997], Julier et al. [vol. 3, pp. 1628-1632, 1995].
In a previous work Maasoumy et al. [Stanford, CA, May 2013]
showed that UKF outperforms EKF for building parameter estimation.
The test-bed that will be introduced later in Section 5.2.7 was used
to collect measurements from January 11 to January 24, 2013. To re-
move noise from the temperature measurements, a second order But-
terworth lowpass filter with cutoff frequency of 0.001 (Hz) was used.
Figure 3.3 shows the temperatures of the neighboring zones and the
outside temperature, which act as disturbance to the PAB model. Fig-
ure 3.4 depicts the model inputs including the air mass flow rate and the
supply air temperature. In order to obtain the best initial parameter
values for the Kalman filter algorithm, we first perform a static param-
eter identification on the historical data. We consider the first part of
the data as training set, and obtain the best parameters that minimize
the least square error between the simulation and the measurement
data. The result of this step is used to simulate the temperature evo-
lution of the room air for the next three days. Due to time-varying
parameters and disturbance to the model, it is difficult to find a set of
parameters for the model, which results in good temperature tracking
for all days.
The obtained initial parameters from the off-line calibration step is
used as initial value for the UKF algorithm. For the off-line parameter
calibration practice, we used the historical data of two weeks where
the first 60% of the data was used for training (calibration) and the
remaining 40% of data was used for testing. The model learns from
new measurements and adjusts the parameters over time as shown in
Figure 3.5. The parameters evolve over time and the steady state values
are not necessarily close to the initial points as expected, due to the
changing ambient environment. More details about the PAB model can
be found in Maasoumy et al. [2014b].
As explained before, one issue with RC models of large buildings is that

such models often lead to very large number of states. Deng et al. [pp.
5118-5123, Baltimore, MD, 2010] described a model reduction method-
ology to obtain a simpler representation of the RC network. The origi-
nal RC network of a building application consists of a large number of
coupled linear differential equations. The proposed technique by Deng
et al. retains the physical intuition of the original model, but is a simpler
RC network. RC modeling is also used for studies on specific parame-
ters related to HVAC load calculations.
Another challenge regarding using RC models is the parameter esti-
mation and model fitting. Platt et al. [2010] proposed real-time HVAC
zone model fitting and prediction techniques based on physical prin-
ciples, as well as the use of genetic algorithms for optimization. The
proposed approach is validated by comparing real-time HVAC zone
model fitting and prediction against the corresponding experimental
measurements. In addition, Platt et al. [2010] used an algorithm based
on feedback-delayed Kalman filters that demonstrated the superiority
of the proposed approach in terms of prediction accuracy.
Fayazbakhsh et al. [2015] summarizes the recent RC sub-circuit
models used for the room and its envelope used in HVAC applications.
Many of the mentioned methods depend on experimental measurements
for acquiring the model parameters, while, other approaches suggest
estimating the RC parameters instead of measuring them. Such models
range from using 1C configuration for room and 1C configuration for
building envelope, to 4R1C and 3R2C configuration for envelope and
2R2C configuration for room model. Fayazbakhsh et al. [2015] also
summarize the advantages of the present RC modeling approach as
follows:
• Real-time simulation of HVAC thermal behavior
• Enabling proactive control of HVAC systems
• Direct measurement of thermal parameters instead of sophisti-

cated mathematical estimations
40 Building Models
• One RC circuit developed for the entire thermal system
• Results validated with experimental data
• Facilitating retrofit of existing HVAC systems through parametric

studies
The PAB model presented in this section provides all the advan-
tages mentioned above, as it is a physics-based RC model. On top of
these advantages, the PAB model learns the changes in the system and
environment and adapts the system parameters accordingly.
4
Building Control Design
Approximately 40% of the global energy consumption occurs in build-

ings, of which, roughly half is used for HVAC in the US and other in-
dustrialized countries. This level of consumption makes HVAC energy
reduction practices very attractive. Such HVAC energy conservation
measures can be realized by improving a building’s HVAC systems and
construction, enhancing its operation, or preferably some combination
of both. However, the majority of the building stock is already in place
and refurbishments of buildings are expensive. Quite differently, control
systems can be upgraded and their operation optimized at much lower
cost. However, safety requirements, the interaction with building users,
increasing comfort requirements, and the complexity of many modern
buildings make the design of energy efficient, economic, robust, and
easy-to-implement building control systems far from trivial.
Advanced control algorithms are considered critical enablers to
achieve comfort and energy efficiency in buildings. Entire sections of the
ASHRAE 90.1 standard Handbook [2009] are dedicated to the specifi-
cation of control requirements. In particular, ASHRAE’s Standard 55
Standard [2004], Thermal Environmental Conditions for Human Oc-
cupancy, suggests the condition, which is acceptable to at least 80% of
41
42 Building Control Design
occupants. According to this standard, the ideal temperature in typical

clothing in summer (0.35-0.6 clo) is in the range of 22.5 ◦ C to 26 ◦ C.
The operative temperature for occupants in normal clothing insulation
in winter which is between 0.8 to 1.2 clo should be in range of 20 ◦ C
to 23.5 ◦ C. Clothing is defined in terms of clo units. Clo is a unit used
to express the thermal insulation provided by garments and clothing
ensembles, where 1 clo = 0.155 (m2 ∗ K/W ) Standard [2004]. This
temperature range is based on a metabolic rate of 1.2 met (70 W/m2 )
and 60% RH. More details can be found in de Dear and Brager [vol.
34, no. 6, pp. 549 - 561, 2002,v], Bradshaw [2010]. ASHRAE’s Stan-
dard 62.1 ASHRAE [2004], Ventilation for Acceptable Indoor Air Qual-
ity, explains outdoor air ventilation requirements for different types of
indoor spaces. When the major contamination source is proportional
to number of occupants, the minimum ventilation rate is enforced in
CF M (L/s) and when other factors play the main role in contamina-
tion, the minimum ventilation rate is enforced in CF M/f t2 (L/s.m2 )
Bradshaw [2010]. We use this standard as a guideline for the control
design analysis presented in this section.
Although the optimal control of an HVAC system is a complex
multi-variable problem, it is standard practice to rely on simple control
strategies that include bang-bang controllers with hysteresis, and PID
controllers, as shown in Figure 4.1. In most cases, standard sequences
of operations for typical installations are used by control contractors.
Each sequence controls the HVAC equipment during an operation phase
such as optimal start, safety shutdown and normal operation. After in-
stallation and tuning, the building is inspected by a commissioning
agent that mainly verifies that the building satisfies the owner’s ex-
pectations. The commissioning agent does not only verify the expected
performance right after installation, but also after the building has
started its operations.
This short snapshot of design and validation practices in the build-
ing industry shows the importance of a model-based design flow for
building controls. To attain energy efficiency, control algorithms need
to be tailored to the physical properties of the building at hand rather
than being an adaptation of a standard sequence designed for a typ-
43
PID
On-Off
On-Off
On-Off
Figure 4.1: Control logic of a typical commercial building. The interaction and
interrelation of building sub-systems is not addressed in the state-of-the-art control
logic of such systems. This may cause inefficient use of energy, such as simultaneous
heating and cooling of air in two different part of the system.
ical building. Thus, a thermal model of the building is needed that

is also suitable for optimal control design. Once such model is made
available, it can be used to design an optimal controller that balances
comfort and energy usage. To achieve building-level energy-optimality,
the model should be able to capture the interaction between physically
connected spaces in the building, occupancy schedules, and state and
input constraints.
The model-based approach enables design methods that are rig-
orous and verifiable. Further, the same models can be used during
building operation for diagnostic purposes. Henze et al. [2004], inves-
tigated the potential of building thermal storage inventory, in par-
ticular the combined utilization of active and passive inventory, for
the reduction of electrical utility cost using common time-of-use rate
differentials. Mendes et al. [2001] elaborated a dynamic multi-nodal

lumped-capacitance non-linear model to describe a building, consid-
ering conduction heat fluxes, envelope thermal capacity, lighting and
people loads, infiltration, fenestration and thermal inertia of heating
systems. Arguello-Serrano and Velez-Reyes [2002] presented a nonlin-
ear disturbance rejection state feedback controller for an HVAC system.
A model-based approach to the design of control algorithms for HVAC
systems was proposed in Maasoumy et al. [2011]. For this purpose a
nonlinear model was developed that takes into account heat transmis-
sion and storage in building elements such as walls and room air, and
external and internal heat gains from different sources such as the sun
and building occupants. The model has then been calibrated against
real data of an existing building.
In this chapter we review different model-based approaches to build-
ing control: Rule-Based Control (RBC) and Reinforcement Learning,
Model Predictive Control (MPC), Robust Model Predictive Control
(RMPC), Stochastic Model Predictive Control (SMPC) and Exergy-
based Model Predictive Control (XMPC). We then compare the most
relevant approaches using our own implementations with respect to a
set of performance indices.
4.1 Rule-Based Control (RBC) and Reinforcement Learning
Current building control algorithms are mainly rule-based (if-then-else

based rules) and thus the performance of a large number of build-
ings heavily depends on the experience of building managers. In recent
years, various advanced control techniques have been proposed to im-
prove upon the performance of the rule-based controllers for HVAC sys-
tems. Dounis and Manolakis [2001] describe the design of a living space
comfort regulator using fuzzy logic. They define comfort as a fuzzy con-
cept, different for different people and depending on the work done in
the space. Singh et al. [2006] present role of fuzzy modeling in HVAC
and control models. A reinforcement learning controller is developed
by Dalamagkidis et al. [2007], where the reinforcement learning signal
used is a function of the thermal comfort of the building occupants, the
4.1. Rule-Based Control (RBC) and Reinforcement Learning 45
indoor air quality and the energy consumption. A direct neural network
(NN) control is designed by Liang and Du [2005] with the control ob-
jective of Predicted Mean Vote (PMV) for indoor thermal environment
regulation. Braun [2003] provides an overview of research related to use
of building thermal mass for shifting and reducing peak cooling loads
in commercial buildings. Studies have demonstrated significant savings
potential for use of building thermal mass in commercial buildings.
However, the savings are sensitive to many factors, including utility
rates, type of equipment, occupancy schedule, building construction,
climate conditions, and control strategy. The problem of computing
optimal control strategies for time-scheduled operation of heating, ven-
tilating and air conditioning systems is explored by Zaheer-Uddin and
Zheng [2000]. A two-zone variable air volume heating system consisting
of a 4 ton heat pump, a storage tank, water and air flow networks and
two environmental zones is considered. The optimization problem that
takes into consideration the building operation schedules consisting of
night-setback, start-up, occupied modes and energy price discounts is
formulated and solved for a given predicted weather profile. Liu and
Henze [2006a] in a two-part investigation of a novel approach present
a framework for optimal control of commercial building passive and
active thermal storage inventory. The proposed building control ap-
proach is based on simulated reinforcement learning, which is a hybrid
control scheme that combines features of model-based optimal control
and model-free learning control. An experimental study was carried
out to analyze the performance of a hybrid controller installed in a
full-scale laboratory facility. The first paper by Liu and Henze [2006a]
introduced the theoretical foundation of this investigation including the
fundamental theory of reinforcement learning control. The second pa-
per by Liu and Henze [2006b] presents a discussion and analysis of the
experiment results. Operating cost savings were attained with the pro-
posed control approach compared with conventional building control;
however, the savings are lower than for the case of model-based predic-
tive optimal control (to be presented in the next section). As for the
case of model-based predictive control, the performance of the hybrid
controller is largely affected by the quality of the training model. Never-
theless, compared with standard reinforcement learning, the proposed

hybrid controller is much more readily implementable in a commercial
building.
Problems of peak power reduction and scheduling of Thermostati-
cally Controlled Loads (TCLs) have also been studied in the literature.
Karmakar et al. [2013] address the problem of maintaining thermal
comfort-bands associated with background loads under peak energy
consumption constraint. They present a feasibility criterion for main-
taining the thermal comfort-band of a given set of background loads
under constraints on the peak available power. They then propose Ther-
mal Comfort-Band Maintenance (TCBM), a new algorithm for schedul-
ing such background loads under peak power constraint. In addition to
limiting peak power demand, the TCBM algorithm avoids undesirable
switching (ON and OFF) of electrical appliances to improve efficiency of
the equipment and reduce failures. Nghiem et al. [2011] present an ap-
proach to fine-grained coordination of energy demand by scheduling the
control systems within a constrained peak while ensuring custom cli-
mate environments are facilitated. The peak constraint is minimized for
energy efficiency, while they provide feasibility conditions for the con-
straint to be realizable by a scheduling policy for the control systems.
Physical systems are then coordinated by the scheduling controller to
satisfy both the peak constraint and the climate/safety constraint.
4.2 Model Predictive Control (MPC)
MPC is based on iterative, finite horizon optimization of a model of

the system to control. At any given time t, the current plant state is
observed, and an optimal control strategy is computed for some finite
time horizon [t, t + H] in the future. An online calculation is used to
explore possible future state trajectories originating from the current
state, finding an optimal control strategy until time t + H . To ensure
robustness with respect to model errors, only the first step of the com-
puted optimal control strategy is implemented; then the plant state is
sampled again, and new calculations are performed on a horizon of H
starting from the new current state. While the global optimality of such
4.2. Model Predictive Control (MPC) 47
receding horizon approach is not ensured, it tends to do well in practice.

In addition to reducing computational complexity, it improves the sys-
tem robustness with respect to exogenous disturbances and modeling
uncertainties.
MPC provides a very attractive framework for optimizing a cost
function subject to constraints over a prediction horizon. This frame-
work is of utmost advantage to control of building HVAC systems.
Optimal control of building HVAC system includes minimizing energy
consumption or monetary cost of operating HVAC systems, subject to
constraints on the room temperature and on the inputs to the system
such as air mass flow input.
In recent years, many papers have analyzed the energy savings po-
tential of MPC in simulations including Kummert and André [2005],
Castilla et al. [2011]. Other outstanding publications in the field of
MPC for building climate control include recent papers by Ma et al.
[2012], Freire et al. [2008], Ferreira et al. [2012], Oldewurtel et al. [pp.
4731-4736, 2008], Zhang et al. [2013].
In particular, model predictive control has gained increasing popu-
larity in utilizing passive or active thermal storage to save energy. With
weather predictions and occupancy schedules, free cooling at night was
used in cooling applications and night setback was adopted for heating
applications by Ma et al. [2012], Kummert and André [2005]. Peak-load
shifting was implemented in MPC to save electricity cost by taking ad-
vantage of the time-of-day rate of electricity price as proposed by Liu
and Henze [2006b]. MPC was also applied to determine optimal tem-
perature set-points at the top level of a hierarchical control, for which
simple PID or on-off controllers were used for lower-level components
such as fans, heating/cooling coils and thermal storage tanks Ma et al.
[2012]. Alternatively, low level components could be directly managed
by an MPC to achieve a higher level of efficiency Castilla et al. [2011].
A hierarchical control architecture for balancing comfort and energy
consumption in buildings was proposed by Maasoumy et al. [2011].
The control architecture comprises a first level that regulates low level
quantities such as air flow, and a second level that balances comfort
and energy consumption.
While on one hand energy saving is important, on the other hand,

occupant thermal comfort also plays a key role in the control of HVAC
systems for low-energy buildings. A large number of the existing con-
trol algorithms were designed such that the neutral temperature was
achieved based on the Fanger’s thermal comfort model. Alternatively,
the so-called effective temperature, which is a combination of the in-
door temperature and relative humidity, could be used as the index
for thermal comfort Freire et al. [2008]. Though the existing HVAC
control algorithms seldom directly optimize a PMV index (or use it as
a constraint), a numerical study demonstrated that using the PMV in
defining the thermal comfort constraint in an MPC could reduce en-
ergy consumption and improve thermal comfort, compared to utilizing
a comfort zone from a psychrometric chart Freire et al. [2008].
Active and passive building thermal storage has been studied by
several authors. Henze et al. [2004] investigated the potential of build-
ing thermal storage inventory, in particular the combined utilization of
active and passive inventory, for the reduction of electrical utility cost
using common time-of-use rate differentials. In particular, they investi-
gate a three-story office building equipped with two chillers with con-
stant coefficient of performance and a thermal energy storage system.
An optimal controller is designed in order to compute the cooling ca-
pacity produced by the two chillers without accounting for pump power.
Experimental results are presented in Liu and Henze [2006a] where the
optimal controller is implemented in a receding horizon fashion on an
unoccupied test bed. Henze et al. [2007], investigate building thermal
mass control of commercial buildings to reduce utility costs with a par-
ticular emphasis on the individual impacts of both adaptive comfort
criteria and of heat waves. Henze et al. [2003] describe simulation-based
results of an investigation of a commercial cooling plant with an ice
storage system. Various ice storage systems, chiller compressors, and
building types were analyzed under four different control strategies.
Oldewurtel et al. [Baltimore, MD, 2010] and Gyalistras and Gwerder
[Zug, Switzerland, 2010] used a model of the building which is bilin-
ear between inputs, states and weather parameters. The approach that
they take is a form of Sequential Quadratic Programming (SQP) for
4.3. Randomized Model Predictive Control 49
solving non-linear problems in which they iteratively linearize the non-

convex constraints around the current solution, solve the optimization
problem and repeat until a convergence condition is met. Larsen et al.
[2009] apply common available methods for wall thermal analysis.
4.3 Randomized Model Predictive Control
Direct incorporation of the PMV in an MPC design for HVAC sys-

tems could pose practical implementation challenges. The calculation
of PMV involves iteration, which could raise computation concerns,
especially for MPC which is known to be computation intensive. Past
work tried to approximate the PMV with a neural network model Dala-
magkidis et al. [2007], Ferreira et al. [2012], Oldewurtel et al. [pp. 4731-
4736, 2008], Zhang et al. [2013] or with a linearized parametrization
model Federspiel and Asada [1994]. In addition, most buildings typi-
cally do not have sensors to continually measure humidity, air velocity
and mean radiant temperature. Even though for laboratory facilities
where the aforementioned environmental sensing data are available,
occupant clothing insulation and activity levels, which could vary with
respect to time and vary among occupants, are seldom monitored con-
tinually and individually. Assuming a uniform and constant clothing
level for occupants could cause errors in predicting occupant thermal
sensations Castilla et al. [2011]. Zhang et al. [2013] proposed a novel
technique called Randomized MPC to improve the control of existing
HVAC systems. Randomized MPC uses weather and occupancy pre-
dictions to minimize the building’s energy consumption. It accounts
for the prediction uncertainties by basing its control actions on a given
number of sampled uncertainty scenarios. The main advantage of Ran-
domized MPC over existing methods is the absence of a probabilistic
disturbance model. This makes the handling of uncertainties straight-
forward, even if they are non-Gaussian or non-additive.
4.4 Robust Model Predictive Control (RMPC)
Optimal control of HVAC components using model-based control tech-

niques has shown promising results for achieving energy efficiency in
buildings Maasoumy and Sangiovanni-Vincentelli [2012], Oldewurtel et

al. [pp. 4731-4736, 2008], Ma et al. [pp. 1-65, 2011], Ma et al. [2010],
Oldewurtel et al. [Baltimore, MD, 2010]. However, these control tech-
niques rely heavily on a perfect (or almost perfect) mathematical model
of the building or a perfect estimation of the unmodelled dynamics of
the system according to Maasoumy and Sangiovanni-Vincentelli [2012].
Although a great deal of progress has been made in modeling the
thermal behavior of building envelope and HVAC system as mentioned
in previous sections, the random nature of some components of build-
ings makes it very hard to predict, with high fidelity, the tempera-
ture evolution of the room using mathematical models. These random
events and phenomena include building occupancy by people, which
along with other internal loads such as the heat emitted from electrical
devices and lighting, account for the total internal heat generation of
the building. The outside environment of the building is also subject
to many random and hard-to-accurately-predict phenomena such as
the wind speed, solar radiation, cloudiness of the sky and outside air
temperature. The aggregate effect of all these factors constitutes the
total external heat gain of the building. We call these two heat gains
of the building, the “unmodelled dynamics”. As mentioned earlier, it
is difficult to obtain a perfect prediction of the loads in future times.
On the other hand, model-based optimal controllers such as MPC are
highly dependent on accurate predictions of these disturbances. In or-
der to account for these modeling deficiencies, it is usually a reasonable
assumption to consider an additive norm-bounded uncertainty to the
model. The question here is how to integrate this uncertainty infor-
mation in the control design to achieve the desired comfort level while
consuming minimum energy.
A framework for handling model uncertainty and selecting best con-
troller based on model uncertainty was proposed by Maasoumy et al.
[2014b], where a nominal MPC and an RMPC were defined, and their
performances analyzed for various model uncertainty levels. MPC as-
sumes that the model is perfect (no uncertainty), and the RMPC as-
sumes that the model is uncertain and designs a robust control policy
for a specific class of uncertainty. The results from MPC and RMPC
4.5. Stochastic Model Predictive Control (SMPC) 51
Figure 4.2: Schematic of the robust model predictive control implementation.
are compared to a conventional Rule-Based Control (RBC) for a typ-

ical building. Novel performance indices are proposed to compare the
performance of these controllers. Also a methodology to select the best
controller among the ones studied in this paper for any given model un-
certainty is proposed, which leads to optimum trade-off between energy
consumption and comfort level.
4.5 Stochastic Model Predictive Control (SMPC)
A bilinear version of an RC model is presented in Oldewurtel et al.

[Baltimore, MD, 2010] that takes into account weather predictions to
increase building energy efficiency. The authors also develop and ana-
lyze a SMPC strategy for building climate control that takes into ac-
count weather predictions to increase energy efficiency while respecting
constraints resulting from desired occupant comfort. They investigate
a bilinear model under stochastic uncertainty with probabilistic, time
varying constraints, and show that the proposed control strategy sig-
nificantly outperform current control practice. Oldewurtel et al. [2012]
presents an investigation of how MPC and weather predictions can
increase the energy efficiency in Integrated Room Automation (IRA)
while respecting occupant comfort. IRA deals with the simultaneous

control of HVAC as well as blind positioning and electric lighting of
a building zone such that the room temperature as well as CO2 and
luminance levels stay within given comfort ranges.
4.6 Exergy-based Model Predictive Control (XMPC)
Exergy is defined as the maximum theoretically available energy that

can do work with respect to a given state via a reversible process Robi-
nett III and Wilson [2011]. As a system moves away from its thermo-
dynamic equilibrium (e.g. a higher temperature gradient with respect
to the environment Dincer and Rosen [2012]), its potential to do work
increases and when a system is at the thermodynamic equilibrium with
its environment, there is no potential to do work, therefore the exergy
of the system in that condition is zero. Exergy is based on the First
and Second Laws of Thermodynamics and is not conserved. The First
Law is related to energy conservation and the Second Law analyzes
entropy generation and irreversibilities, which cause deficiency and en-
ergy waste in a system. Hence, exergy is a more appropriate metric to
analyze power systems as compared to energy.
In addition to the calculation of energy balances, the concept of
exergy is used to evaluate the quality of energy sources, resulting in a
higher flexibility of measures to optimize a building design. A proto-
typical tool integrated into a building information modeling software
is described in Schlueter and Thesseling [2009], enabling instantaneous
energy and exergy calculations and the graphical visualization of the
resulting performance indices.
As mentioned earlier more than 50% of energy demand in build-
ings is related to HVAC, where this heat transfer occurs close to the
ambient temperature. However, these low quality energy demands are
mainly met by high quality (high exergy) sources such as electricity
from grid which itself is obtained from high exergy sources such as fos-
sil fuels. Thus, it is of a great interest to satisfy low exergy demand
with low exergy sources such as renewable energy sources produced by
photo-voltaic panels and wind turbines. By applying exergy aware con-
4.6. Exergy-based Model Predictive Control (XMPC) 53
trol algorithm, HVAC systems can be operated in low exergy fashion.

As a result, irreversibilities will be reduced causing less exergy destruc-
tion. In other words, systems can be operated with less irreversibility
and consequently system operation will be more sustainable and more
energy efficient.
A great deal of study has been done in exergy analysis for build-
ings’ heating and cooling systems and a wide literature on exergy anal-
ysis of heat pumps for both building cooling and heating system exists
in Sakulpipatsin et al. [2010], Yildiz and Güngör [2009], Chengqin et al.
[2002], Bi et al. [2009], Meggers et al. [2012]. Applications of low ex-
ergy systems for building HVAC have been studied before in Schmidt
[2009], Kilkis [2006]. For instance, in Kilkis [2006], it is shown that
HVAC systems are more exergy-efficient if low exergy energy sources
are used. Since conventional HVAC systems use high exergy energy
sources, therefore, they have a low exergy efficiency. This drawback
has not yet been addressed properly in the literature. Given the energy
crisis and the unprecedented focus and research on energy efficiency of
built environment over the last decade, and at the same time increasing
adoption of renewable energy resources, design of controller algorithms
for building HVAC systems with exergy considerations is crucial. Op-
timal control algorithms enable us to reduce greenhouse gas emissions,
exergy destruction and energy consumption of buildings.
In the context of building HVAC control, energy is always used
as an index to evaluate efficiency of HVAC systems. Razmara et al.
[2015] introduce exergy of the system as a more appropriate metric to
evaluate performance of HVAC systems. Also, an exergy model of a
test-bed building is formulated. An MPC strategy is formulated based
on the minimization of exergy destruction. The optimal control prob-
lem reduces irreversible entropy generation of an HVAC system by
minimizing exergy destruction, and addresses not only energy saving
but also environmental concerns by saving exergy. This work is the
first study undertaken to develop an HVAC model predictive control
system based on minimizing exergy destruction. The results are very
promising as the new MPC saves both energy and exergy destruction.
Exergy is also a measure of sustainability. An exergy model of a
building is presented in Razmara et al. [2015] and exergy destruction

is formulated as a function of the physical parameters of the building
model and environment. The exergy model is then used to develop MPC
strategy based on the minimization of exergy destruction. Compared
to a traditional on-off controller, the proposed XMPC is reported to
achieve 22% decrease in exergy destruction and 36% reduction in energy
consumption for the building. Simulation results also demonstrate the
benefit of XMPC compared to energy-based MPC. The results indicate
XMPC leads to 4% less exergy destruction compared to MPC while
providing 12% more energy saving.
Low exergy (LowEx) building systems create more flexibility and
generate new possibilities for the design of high-performance buildings.
Instead of maximizing the barrier between buildings and the environ-
ment using thick insulation, low exergy systems maximize the connec-
tion to the freely available dispersed energy in the environment. LowEx
systems provide many heating and cooling methods for buildings us-
ing moderate supply temperatures and heat pumps that exploit more
valuable energy sources. Meggers et al. [2012] propose an implementa-
tion of integrated LowEx systems that maintains low temperature-lifts,
which can drastically increase heat pump performance from the typical
Coefficient of Performance (COP) range of 3–6 to values ranging from
6 to 13.
The Exergy Model To use exergy as a criterion for optimizing the

behavior of a building controller, a mathematical model of exergy is
needed. In particular, the definition of the reference environment in
terms of reference temperature, pressure, and chemical composition are
crucial. Ambient environment is used as a reference condition for exergy
analysis in buildings for HVAC applications Meggers and Leibundgut
[2012]. In our study the same definition is used while each room in
a building is considered as a control volume. The following equation
governs the exergy balance for a control volume with thermodynamic
4.6. Exergy-based Model Predictive Control (XMPC) 55
exergy: destruction Cengel et al. [2015]:

ẊiH.T.,r
z }| {
r
X T0 H.T.,k X X dXir
Ẋdest i
= (1 − r )Q̇i −Ẇir + ṁri ψ − ṁri ψ −
k∈N r
Ti in out
dt
i
(4.1)
where the rate of exergy destruction in ith room is shown by Ẋdest r

i
.
This term denotes the loss in work potential due to irreversibility such
as air mixing and heat transfer. T0 is the reference air temperature and
H.T.,k
Q̇i is the rate of heat transfer to room i. Rate of exergy transfer by
work is shown by Ẇir and is equal to zero for our system since there is
no associated work in room i. ψ shows the amount of exergy associated
dX r
with flow. dti represents the rate of change of exergy of the room i.
Whenever there is a temperature change, exergy destruction is in-
evitable. Unlike energy which is never destroyed during a process, ac-
cording to Dincer and Rosen [2012], exergy is not a conserved prop-
erty. Rate of exergy destruction (Ẋdest ) is proportional to irreversible
entropy production inside the control volume. Total exergy transfer by
the heat transfer is represented by ẊiH.T.,r in Equation (4.1). This term
is a function of different factors including the building type, insulation
level and temperature difference between room i and surrounding zones.
The third and the fourth terms of right hand side of Equation (4.1)
present flow exergy transferred in and out of the room, respectively.
Exergy change due to heat transfer can be rewritten in the following
terms:
T0 T k − Tir
ẊiH.T.,r =
X
(1 − r )( k
) (4.2)
k∈N r
Ti Ri,j
i
The total exergy of a flowing fluid in a control volume is the sum of

the exergies of its kinetic energy, potential energy and enthalpy. The
following equation shows the unit-mass form of this equality:
V2
ψ = (h − h0 ) − T0 (s − s0 ) +
+ gz (4.3)
2
in which h and h0 indicate enthalpy and dead-state enthalpy of the
fluid. s and s0 show fluid entropy and dead-state entropy. Specific
kinetic energy of fluid and specific gravitational potential energy are

2
represented by V2 and gz, respectively. Changes in kinetic and grav-
itational energy of the supply air are neglected in this study due to
insignificant values. Rate of change of exergy inside room i based on
change in enthalpy and entropy is demonstrated in the following equa-
tion:
Xir =mroom
i [(h − h0 ) − T0 (s − s0 )] ⇒
dXi r dh ds dmroom (4.4)
=mroom
i ( − T0 ) + i
[(h − h0 ) − T0 (s − s0 )]
dt dt dt dt
where mroom
i is mass of the air inside room i. Ideal gas assumption is
considered for change in enthalpy and entropy due to the compressibil-
ity factor being close to one (very low pressure of the air). We neglect
change of mass of the air inside room in Equation (4.4), thus we as-
dmroom
sume i
dt is equal to zero. Hence, the only remaining term of the
Equation (4.4) is the first term. The following equation expresses the
entropy change (∆s) and the enthalpy change (∆h) of air:
Z 2
∆h = cpavg dT ⇒ h2 − h1 = cpavg (T2 − T1 ) (4.5)
1
Z 2
dT v2 T2 v2
∆s = cvavg + R · ln ⇒ s2 − s1 = cvavg · ln + R · ln
1 T v1 T1 v1
(4.6)
where cvavg is average specific heat capacity at constant volume and

cpavg is average specific heat capacity at constant pressure, respectively.
In Equations (4.5) and (4.6), cpavg and cvavg values are found for the
average of air temperature range studied. R and v show gas constant
and specific volume, respectively. Tis is considered entrance tempera-
ture of supply air into room i, while Tir is the exit temperature in the
control volume. By plugging in Equations (4.2), (4.3), (4.4), (4.5),
and (4.6) into Equation (4.1) and discretizing the consequent equation
4.7. Comparisons 57
with sampling time of Tsample the following equation is obtained:
r
X T0 [k] Tkr [k] − Tir [k]
Ẋdest [k] = (1 − )( )+
i
Tkr [k] k
Ri,j
k∈Nir
Tis [k]
ṁri [k]{cpavg (Tis [k] − Tir [k]) − T0 [k]cvavg ln( )}
Tir [k]
mroom
i T r [k]
+ {cpavg (Tir [k] − Tir [k − 1]) − T0 [k]cvavg ln r i }
Tsample Ti [k − 1]
(4.7)
where [k] indicates index of time step. Equation (4.7) expresses the
exergy destruction at each time step based on the outside air temper-
ature, mass flow rate, supply air temperature and zone temperature.
Figure 4.3: Structure of XMPC and Energy-based MPC.
4.7 Comparisons
In this section, we compare whenever possible the approaches that we

presented in previous sections. To do so, we must formulate the op-
timization problems so that a meaningful analysis can be performed.
MPC and RMPC will be compared first also with respect to the base-
line represented by RBC. Then XMPC will be compared with MPC
again using RBC as baseline.
4.7.1 MPC Formulation
A model predictive control problem is formulated with the objective

of minimizing a linear combination of the total and the peak airflow.
Here we consider air mass flow as the control input and keep supply
air temperature constant in order to replicate the control algorithm
of the test-bed. We implement the control inputs obtained from the
MPC with the linearized system dynamics of the model on the original
nonlinear model for forward simulation.
Fan energy consumption is proportional to the cube of the airflow.
Hence minimizing the peak airflow would dramatically reduce fan en-
ergy consumption. We have considered a cost function for the MPC
which comprises linear combination of the total input airflow (`1 norm
of input) and the peak of airflow (`∞ norm of input). The alternative
would be to use the actual nonlinear function of fan energy consump-
tion. However, it would lead to nonlinear MPC which is much slower
than linear MPC. We use the proposed cost function to achieve better
computational properties. Also in order to guarantee feasibility (con-
straint satisfaction) at all times, we implement soft constraints. The
predictive controller solves at each time step the following optimiza-
tion problem:
MPC Formulation
min {|Ut |1 + κ|Ut |∞ + ρ(|t |1 + |t |1 )} = (4.8a)
Ut ,¯
,
N −1 N
X X
min { |ut+k|t | + κ max(|ut|t |, · · · , |ut+N −1|t |) +ρ (|εt+k|t | + |εt+k|t |)}
Ut ,ε̄,ε
k=0 k=1
(4.8b)
s.t. xt+k+1|t = Axt+k|t + But+k|t + Edt+k|t , k = 0, ..., N − 1 (4.8c)
yt+k|t = Cxt+k|t , k = 1, ..., N (4.8d)
U t+k|t ≤ ut+k|t ≤ U t+k|t , k = 0, ..., N − 1 (4.8e)
T t+k|t − εt+k|t ≤ yt+k|t ≤ T t+k|t + εt+k|t , k = 1, ..., N (4.8f)
εt+k|t , εt+k|t ≥ 0, k = 1, ..., N (4.8g)
where Ut = [ut|t , ut+1|t , · · · , ut+N −1|t ] is vector of control inputs, and

4.7. Comparisons 59
= [εt+1|t , · · · , εt+N |t ] and = [εt+1|t , · · · , εt+N |t ] are the slack vari-

ables used to utilize soft constraints on room temperature. xt+k|t is the
vector of temperatures of all the nodes in the RC network, yt+k|t is
the room temperature vector, dt+k|t is the disturbance load prediction,
and T t+k|t and T t+k|t for k = 1, · · · , N are the lower and upper limits
on the room temperature, respectively. U t+k|t and U t+k|t are the lower
and upper limits on the airflow input by the variable air volume (VAV)
damper, respectively. Note that based on the ASHRAE Standard 62.1-
Section 6.2.6.1, during unoccupied hours, ventilation systems should be
able to maintain the required non-zero ventilation rates (U t+k|t > 0) in
the breathing zone Handbook [2009]. ρ is the penalty on the comfort
constraint violations, and κ is the penalty on peak power consumption.
A larger value for κ reduces the peak power consumption, and results
into a smoother air mass flow profile, while the system may end up
using more cumulative conditioned air. On the other hand, a larger
ρ penalizes temperature violations more and leads to higher energy
consumption compared to a case with smaller ρ where more (though
small) temperature violations may occur. κ provides a mechanism for
tuning the trade-off between peak air flow and cumulative conditioned
air consumption, while ρ provides a mechanism for tuning the trade-off
between comfort and cost (energy consumption).
At each time step only the first entry of Ut is implemented on
the system. At the next time step the prediction horizon N is shifted
leading to a new optimization problem. The prediction horizon is N =
24. This process is repeated over and over until the total time span of
interest is covered.
4.7.2 RMPC Formulation
We consider additive uncertainty to the system model as previously

described in Equation (3.12). A schematic of the robust optimal control
implementation on the nonlinear building model is shown in Figure 4.2.
In RMPC algorithm, the cost function is the same as the one in the
MPC case:
RMPC Formulation
min {||Ut ||1 + κ||Ut ||∞ + ρ(||t ||1 + ||t ||1 )} (4.9)
Ut ,ε̄,ε
Subject to:
xt+k+1|t = Axt+k|t + But+k|t + E(dt+k|t + wt+k|t ) k = 0, ..., N − 1 (4.10a)

yt+k|t = Cxt+k|t k = 1, ..., N (4.10b)
T t+k|t − εt+k|t ≤ yt+k|t ≤ T t+k|t + εt+k|t k = 1, ..., N (4.10c)
U t+k|t ≤ ut+k|t ≤ U t+k|t k = 0, ..., N − 1 (4.10d)
εt+k|t , εt+k|t ≥ 0 k = 1, ..., N (4.10e)
∀ wt+k|t ∈ W k = 0, ..., N − 1 (4.10f)
The only difference with respect to the MPC algorithm is the intro-
duction of the additive uncertainty term w in the state update equation.
Using this formulation, we derive a robust counterpart of an uncertain
optimization problem in which constraints are satisfied for all possible
uncertainties, and worst-case objective is calculated.
In Maasoumy and Sangiovanni-Vincentelli [Fort Lauderdale, FL,
2012] the open-loop constrained robust optimal control (OL-CROC) is
proven to be conservative. The closed-loop constrained robust optimal
control (CL-CROC) formulation overcomes this issue but it can quickly
lead to an intractable problem Löfberg [2003]. Next, we review the
feedback prediction concept followed by our proposed formulation to
improve upon the feedback prediction scheme.
The idea in feedback prediction is to introduce new decision vari-
ables and parameterize the future control sequences using the future
disturbances and an additive independent decision variable. Define an
affine disturbance feedback as:
i−1
X
ui := mi,j wj + ni ∀i = 1, 2, ..., N − 1 (4.11)
j=0
Therefore the input vector can be written as U = Mw + n, where

4.7. Comparisons 61
M and n are given by
   
0 ··· ··· 0 n0
 ..   .. 
 m1,0 0 . 0   . 
M :=  , n :=  (4.12)
   
.. .. .. .. .. 

 . . . .



 .


mN −1,0 · · · mN −1,N −2 0 nN −1
and the vector of disturbances is given by w = [w0 w1 · · · wN −1 ]0 .

The control sequence is parameterized directly in the uncertainty.
What we have here is basically a sub-optimal version of the closed-loop
min-max solution Löfberg [2003].
The main problem with the min-max formulation based on LTS
parametrization is the excessive number of decision variables and con-
straints. The reason is the high-dimensional parametrization of matrix
M. To resolve the issue of high-dimensional parametrization of matrix
M, we propose the following new parameterizations.
By analyzing the structure of the optimal matrix M, it was ob-
served that the parametrization of the input does not need to consider
feedback of more than past two values of w at each time, hence we
propose the following disturbance feedback.
ui := mi,i−2 wi−2 + mi,i−1 wi−1 + ni

i−1
X (4.13)
= mi,j ωj + ni ∀i = 1, 2, ..., N − 1
j=i−2
and the corresponding parametrization matrix M is an N × N matrix

that has the entries on the first and second diagonal of M below its
main diagonal as decision variables and 0 elsewhere. n remains as in
Equation (4.12). With this structure we exploit the sparsity of the
feedback gain matrix to enhance the computational characteristics of
the controller.
4.7.3 Comparing MPC and RMPC

Performance Indices
To compare the overall performance of the proposed controllers we
define indices to measure the energy consumption and comfort level
provided by each controller. In addition, we define a new index to eval-
uate the overall performance of each controller considering both the
energy and comfort indices.
• The energy index Ie in (kWh) is defined:

Z 24
Ie = [Pc (t) + Ph (t) + Pf (t)] dt (4.14)
t=0
where cooling power Pc , heating power Ph and fan power Pf are

determined by:
Pc (t) = ṁc (t)cp [Tout (t) − Tc (t)] (4.15a)

Ph (t) = ṁh (t)cp [Th (t) − Tout (t)] (4.15b)
Pf (t) = αṁ3 (t) (4.15c)
cp = 1.012(kJ/kg.o C) is the specific heat capacity of air and α =

0.5(kW.s3 /kg3 ) is the fan power constant Ma et al. [pp. 1-65,
2011]. Using these constants, the fan power values, in (kW), can
be calculated.
• The discomfort index Id in degree Celsius hour (o Ch) is defined
as the sum of all the temperature violations in the course of a
day.
Z 24
Id = min T (t) − T (t) , |T (t) − T (t)| .1B(t)c (T (t)) dt (4.16)
t=0
where B(t) = [T (t), T (t)] is the comfort zone at time t and 1 is

the indicator function.
• A good control performance means not only low energy consump-

tion, but also low resulting discomfort. To assess the overall per-
formance of the controllers, we need to examine both Ie and Id
4.7. Comparisons 63
at the same time. Using the two indices defined above we define
a third index called Overall Performance Index (IOP ). The in-
tuition behind this new index is to take into account the energy
and discomfort index in one single term. IOP is defined as:
(Id∗ − Id )/||Id ||∞
IOP = (4.17)
Ie /||Ie ||∞
where Id∗ is the maximum allowed discomfort and ||.||∞ denotes

infinity norm or the maximum value of energy indices among all
three controllers. Negative value of IOP means that the discomfort
index is not within the preferred range. The lower the Id and Ie
are, the higher the IOP will be. Therefore, the higher the IOP ,
the better the overall performance. In this study, the limit on
the allowed discomfort index is heuristically chosen to be Id∗ =
0.5(◦ Ch) to ensure adequate comfort level.
Comparison Results
To illustrate the effectiveness of the controllers proposed in subsec-
tions 4.7.1 and 4.7.2, we assess their performances for different model
uncertainty values denoted by δ and defined as
λ
δ= × 100 [%] (4.18)
||d||∞
where λ is the `∞ norm bound of the uncertainty as given by Equa-
0 0 0 0
tion (3.13) and d = [d1 , d2 , ..., dN ] is the disturbance realization vector.
d0 represents transpose of vector d.
A time constant of ∆t = 1 (hr) is used for all controllers. We im-
plement the introduced model predictive controllers with a prediction
horizon of N = 24. The choice of N = 24 is to provide a good balance
between performance and computational cost for the MPC framework.
We use the following numerical values for parameters in Equa-
tions (4.8) and (4.10). U = 63 cfm (0.03 m3 /s) is the higher limit
on air mass flow, [T .|t T .|t ] = [20 22]o C during occupied hours, and
[T .|t T .|t ] = [19 23]o C is used during unoccupied hours. For the sim-
ulations we use κ = 0.75 and ρ = 50. ε and ε are the slack variables
used to avoid feasibility problem, where and are the vectors storing
slack variables.
Figure 4.4: Control input and resulting temperature profile for the existing con-
troller on the building (Measurements), RBC, MPC, and robust MPC controllers.
The additive uncertainty bound is considered δ = 60% in this case.
Optimal controller and the resulting room temperature with the

presence of a box-constrained uncertainty in four cases are depicted in
Figure 4.4. Measurements, as shown in black, shows the air mass flow
and temperature recording for the room using a simple existing control
policy of the building HVAC system. RBC represents the result of the
rule-based control. MPC refers to the performance of a model-based
control algorithm in which no knowledge of the model uncertainty is
known a-prior to the control algorithm. RMPC refers to the simulation
of the control algorithm which considers the model uncertainty bound
and utilizes the uncertainty feedback strategy of Equation (4.11) in
designing the control policy.
We consider stochastic uncertainties with different uncertainty
bounds (λ) as introduced in Equation (3.13). The MPC does not have
any a-priori information regarding the additive uncertainty, and calcu-
lates the controller solely based on the deterministic system dynamics.
4.7. Comparisons 65
However, the RMPC integrates the uncertainty bound information in

the control derivation. Controller performances are evaluated based on
indices introduced in Section 4.7.3. Problem is solved using CPLEX
12.2 Cpl [2012] on a 2.67 GHz machine with 4 GB RAM. Here are the
discussions of the results:
Table 4.1: Comparison of LTS and TLDS uncertainty feedback parameterizations

results for the case of δ = 60%.
Number of Average
Controller feedback decision simulation time Ie Id
variables for N = 24, in (s) (kW h) (o Ch)
LTS l.m.r( N (N2+1) ) 200 16467 0
TLDS 3l.m.r(N − 1) 138 16467 0
Computational Aspects: Exploiting the TLDS structure results

in the same control law that was obtained from the LTS structure.
However, matrix M of LTS has l.m.r N (N2−1) variables (quadratic in
N ) while matrix M of TLDS has l.m.r(2N − 3) variables (linear in
N ). Hence it exhibits shorter computation time. On average, the sim-
ulation time for TLDS is 30% less than the LTS structure, as shown
in Table 4.1.
Comfort: From Figure 4.4, the RMPC is the only controller that
is able to keep the temperature within the allowed comfort zone at
all times during this test simulation, meaning maintaining minimum
level of discomfort (Id ≤ Id∗ ), while RBC still does a very good job and
MPC fails to do so with Id > Id∗ for all δ ≥ 40%. Figure 4.5 depicts
how discomfort index Id varies with additive model uncertainty δ for
MPC and RMPC. Note that different data points for one δ value refer
to simulations with different random sequences. The reason for such
a wide variation of the simulation results, specially for large values
of δ, stems from the fact that depending on the value of the random
variable at any time, the resulting disturbance vector can either lead
5
RBC
4.5
MPC
4 RMPC
Trendline of MPC data
Discomfort index, Id (°Ch)
3.5 Trendline of RMPC data
2.5 Parameter Adaptive

Building (PAB) Model
2
1.5
0.5
0
0 25 50 75 100
Additive model uncertainty, δ (%)
Figure 4.5: Discomfort index Id versus additive model uncertainty (δ). We gener-
ate a uniform random sequence based on the disturbance prediction error value δ.
The generated random sequence is used in the simulations for making this graph.
Trendlines in this figure are calculated based on least square estimation.
to temperature rise or fall with respect to the nominal disturbance

value. It is shown that RMPC manages to keep the perfect comfort
level (Id = 0) for additive model uncertainty up to δ = 75%, while the
MPC maintains the perfect comfort level for uncertainty bounds up to
δ = 20%. The discomfort index for MPC goes as high as 4.61 (o Ch)
while the value for RMPC reaches 1.2 (o Ch) in the worst case in the
simulations corresponding to δ = 100%. Since RBC is not a model-
based control technique, its performance does not depend on values of
δ, hence the straight horizontal line in Figure 4.5 (Id = 0.25o Ch).
Energy Consumption: Figure 4.6 depicts the variations of energy
index Ie versus the uncertainty bound on the unmodelled dynamics. It
is clear that the energy index for RMPC increases dramatically with
δ while the energy index for MPC only changes slightly. However, this
comes with the drawback of increased discomfort index for MPC. Fig-
ure 4.6 also shows energy consumption of RBC (Ie = 1.43 × 104 kWh).
MPC for all values of δ leads to a lower level of energy consumption
4.7. Comparisons 67
4
x 10
5
RBC
4.5 MPC
RMPC
4 Trendline of MPC data
Trendline of RMPC data
Energy index, Ie (kWh)
3.5
3
Parameter Adaptive
2.5 Building (PAB) Model
1.5
0.5
0
0 25 50 75 100
Additive model uncertainty, δ (%)
Figure 4.6: Energy index Ie versus additive model uncertainty (δ). The data points
for this graph were generated using a similar technique as in Figure 4.5. Trendlines
in this figure are calculated based on least square estimation.
than RBC, but RMPC leads to more energy consumption than RBC
soon after δ = 35%.
Comfort-Energy Trade-off: An important point to notice from
Figure 4.6 is how much more energy needs to be supplied to the HVAC
system to maintain the comfort level in the presence of imperfect and
faulty, un-modelled dynamics predictions. Consider the case where δ =
75%. MPC will lead to a discomfort index of 1.7o Ch on average, while
the RMPC is able to maintain the temperature below a discomfort
index of 0.016o Ch on average. However, this level of comfort provided
by the RMPC comes at a cost of energy consumption of 3 times more
than that of the MPC case. Note that due to the trade-off between
comfort and energy consumption, the choice of which controller to use
is made by the building HVAC operator, and it depends on various
factors such as criticality of meeting the temperature constraints for
the considered thermal zone in the building, and availability and price
of energy at that time of the day/year.
As observed from Figure 4.5 and 4.6, the behavior of controllers
vary considerably as the model uncertainty increases. For instance, the

energy required to keep the same level of comfort for RMPC in the case
of δ = 75% is almost 3 times the energy required to provide the same
level of comfort when δ = 25%. Figure 4.5 and Figure 4.6 show the
importance of a good model like PAB in minimizing the energy con-
sumption of building HVAC systems for a desirable comfort level using
model-based control techniques by accurately capturing the dynamics
of the system.
Figure 4.7: Energy saving of MPC and RMPC compared to RBC as a function
of model uncertainty. The blue eclipse shows the operating area of the PAB model
which keeps the model uncertainty very small.
4.7.4 Comparing MPC and RMPC versus RBC
The rule-based controller in this study is a conventional on-off HVAC

controller. The time constant of the control action implementation is
∆t. The controller opens the dampers of conditioned air flow to the
thermal zones when heating is required and keeps it fully open for the
duration of ∆t. In the next time step the controller checks the temper-
ature again and adjusts the damper position if the room temperature
is within the comfort zone or keeps it open if the room air temperature
4.7. Comparisons 69
is still outside the comfort zone. In on-off control, the position of the
dampers can be either the min value or the max value. When system
goes to the cooling mode, supply air temperature changes accordingly.
The experimental data presented here is for the heating mode only. To
be consistent and to perform a fair comparison, we use the same time
constants ∆t for all controllers.
Figure 4.7 demonstrates savings of MPC and RMPC versus RBC.
As shown, the maximum theoretical energy saving of MPC compared
to RBC is 36%, and that of RMPC compared to RBC is 30% for the
building studied. These values decrease as model uncertainty increases.
Energy saving of MPC versus RBC stays positive even for large val-
ues of model uncertainty, while energy saving of RMPC versus RBC
is positive only for model uncertainty values up to about 34%, and
is negative for larger model uncertainties (i.e. RMPC consumes more
energy than RBC).
The results of an extensive study in Gyalistras and Gwerder [Zug,
Switzerland, 2010] show that MPC HVAC control can potentially pro-
vide 16%-41% building energy saving compared to rule-based con-
trollers, which is in agreement with our findings. The saving depends on
various factors including climate zone, insulation level, and construc-
tion type. Stochastic MPC was shown in Gyalistras and Gwerder [Zug,
Switzerland, 2010] to be superior to the rule-based control given the un-
certainties in occupancy and weather forecast. Our findings also show
that the robust MPC outperforms the rule-based control in terms of
energy consumption and user comfort. Although these two MPC tech-
niques (robust and stochastic MPC) both address model uncertainty,
they are formulated differently and hence can lead to different perfor-
mance results. A comparative analysis of these two MPC techniques
is the subject of our future work. Given the accuracy of the PAB for
removing model uncertainty, designing MPC scheme based on PAB is
a promising solution for building control problems.
4.7.5 Selecting a Controller Based on Model Uncertainty
For simulation evaluation of energy consumption and provided comfort

level, we have compared the overall performance of the three controllers
Figure 4.8: Overall performance index for RBC, MPC and RMPC as a function of
model uncertainty. The red zone demonstrates the region which MPC outperforms
RMPC and RBC as it yields a higher IOP . The green zone represents the region
that IOP of RMPC is higher than that of MPC and RBC. RBC dominates in terms
of IOP in the blue zone. In the gray zone the resulting discomfort index is not
acceptable.
using IOP . The results, as shown in Figure 4.8, suggest that for model
uncertainties less than 30% MPC is the best controller type. For model
uncertainties above 30%, RMPC and RBC are close in performance
while for δ between 30% and 67% RMPC is the best, and for model
uncertainties larger than 67%, RBC leads to better overall performance
than model-based control techniques. This information can be utilized
in choosing a controller type for building HVAC system. As described
in the paper, proper choice of building HVAC control would depend on
the accuracy of the given building model. The range of uncertainties
for a given building model can be obtained by taking the difference of
the temperature predictions from the building model and temperature
measurements from a building. The statistics of such uncertainty can
be found once such data is available. The mean and variance of the
uncertainty from the statistical analysis can be used to select the best
controller type.
4.7. Comparisons 71
4.7.6 Comparing Energy-Based MPC with XMPC
MPC Formulation The formulation the energy-based MPC in the

comparison uses the energy index, Ie , defined in Maasoumy et al.
[2014a] to minimize energy consumption in HVAC systems. Ie in (kWh)
is defined as function of all HVAC loads in a building:
24
X
Ie = [Pc (t) + Ph (t) + Pf (t)] ∆t (4.19)
t=1
where cooling power Pc , heating power Ph and fan power Pf are deter-
mined by:
Pc (t) = ṁri (t)cpavg [Tri (t) − Tc (t)] (4.20a)

Ph (t) = ṁri (t)cpavg [Th (t) − Tri (t)] (4.20b)
Pf (t) = κ(ṁri )3 (4.20c)
where Tc and Th are the supply air temperatures in the cooling mode
and the heating mode, respectively. κ [W.s3 .kg− 3] is the fan coeffi-
cient which expresses the cubic relation between power consumption
and mass flow rate (ṁri ). Since air mass flow has a constant value and
is not a control input in the state space model, fan power (Pf ) is not
considered in the energy cost function. We formulate an MPC problem
to minimize Ie . The electrical energy consumption is proportional to
the difference between supply air temperature (Tc or Th ) and the room
air temperature Tir . Hence, the controller minimizes the required en-
ergy while keeping the room temperature within the comfort zone. We
use the objective function in Equation (4.21a) to minimize energy us-
age with low computational requirements. In addition, soft constraints
denoted by in Equation (4.21), are formulated to guarantee optimal
problem feasibility at all times. The following optimization problem is
being solved at each time step t:
min {|Ie |1 + ρ(|t |1 + |t |1 )} (4.21a)

Ut ,¯
,
subject to:
xt+k+1|t = Axt+k|t + But+k|t + Edt+k|t (4.21b)
yt+k|t = Cxt+k|t (4.21c)
U t+k|t ≤ ut+k|t ≤ U (4.21d)
δU ≤ ut+k+1|t − ut+k|t ≤ δU (4.21e)
T t+k|t − εt+k|t ≤ yt+k|t ≤ T t+k|t + εt+k|t (4.21f)
εt+k|t , εt+k|t ≥ 0 (4.21g)
where Equations (4.21b) and (4.21c) is the building state equa-
tion model. Input constraints on supply air temperature are shown
in (4.21d) and (4.21e), and (4.21f) denotes the output constraint on
room air temperature and (4.21g) is the constraint on slack variables.
Constraints (4.21b) and (4.21d) must hold for all k = 0, 1, ..., N − 1
and constraints (4.21c), (4.21e), (4.21f), and (4.21g) must hold for
all k = 1, 2, ..., N . Ut = [ut|t , ut+1|t , · · · , ut+N −1|t ] denotes the vector of
control inputs, t = [εt+1|t , · · · , εt+N |t ] and t = [εt+1|t , · · · , εt+N |t ] are
the slack variables vector. yt+k|t is thermal zone temperature vector,
the output of state model. dt+k|t denotes the disturbance vector, and
T t+k|t and T t+k|t for k = 1, · · · , N are the comfort bounds (lower and
upper bounds) on the zone temperature, respectively. U t+k|t and U de-
note the lower and upper limits on the supply air temperature provided
by the heat-pump, respectively. Since physical limit on maximum sup-
ply air temperature is not time varying, time invariant constraint U is
used. δU and δU are limits on rate of change of supply air temperature
due to dynamics of heat-pump condenser. ρ is being used to consider
penalty on the comfort constraint violations.
At each time step, the first entry of input vector, Ut , computed by
the optimization problem is applied to the building model and moves
the system forward to the next time step. Then, the prediction time
horizon, N , is shifted to form a new optimization problem. The op-
timization process is repeated until the total time span of interest is
covered. In order to find the optimal future inputs, weather prediction
4.7. Comparisons 73
for the next 24 hours, schedule of the zone (temperature bounds), and
the building energy model are used. A schematic of model predictive
control implementation is depicted in Figure 4.3
XMPC Formulation Exergy destruction (Ẋdestt ) from Equation (4.7)

is selected as the objective function. In order to minimize Ẋdestt , we
propose the objective function (4.22). For this purpose, the XMPC al-
gorithm finds the optimum inputs (supply air temperature) to minimize
the irreversible entropy generation of the HVAC system and therefore
increases the system efficiency. The reason that we use supply air tem-
perature as the control input is that in the control logic of the test-bed
that we are using for this study, air mass flow is considered constant in
occupied hours, and supply air temperature is the only control input.
To perform a fair comparison, we also use the same control input and
leave the air mass flow constant.
The following equation expresses the XMPC objective function:
XMPC Formulation
min {Ẋdestt + ρ(|t |1 + |t |1 )} (4.22a)

Ut ,¯
,
subject to:
xt+k+1|t = Axt+k|t + But+k|t + Edt+k|t (4.22b)
yt+k|t = Cxt+k|t (4.22c)
U t+k|t ≤ ut+k|t ≤ U (4.22d)
δU ≤ ut+k+1|t − ut+k|t ≤ δU (4.22e)
T t+k|t − εt+k|t ≤ yt+k|t ≤ T t+k|t + εt+k|t (4.22f)
εt+k|t , εt+k|t ≥ 0 (4.22g)
The MPC problem is subjected to the same constraints on input

and states as those in Equation (4.21). In Equation (4.22), Ẋdestt is
a nonlinear function and it leads to a nonlinear model predictive con-
trol (NMPC) problem compared to linear model predictive problem
(LMPC) in Equation (4.21).
Results The results of the comparison are summarized in the follow-

ing table:
Table 4.2: Performance comparison for three HVAC controllers.
Cumulative Cumulative Reduction in Reduction in

Controller exergy energy exergy energy
type destruction consumption destruction consumption
[kW h] [kW h] [%]* [%]*
XMPC 2.7 4.2 22 36
EMPC 2.8 4.6 18 24
RBC 3.3 5.7 0 0
* x−x0
Saving percentage is calculated by x
× 100, where x0 is the energy
consumption or exergy destruction of the RBC.
In this chapter, we used the building model that was obtained in Chap-
ter 3, we presented MPC to minimize the energy consumption of the
building while satisfying thermal comfort, and constraints on other sys-
tem states and inputs. Model uncertainty is an unavoidable challenge
for modeling and model-based control of a building HVAC system.
We characterized the impact of model uncertainty on MPC con-
trollers and presented two approaches to minimize model uncertainty
for building controls. We presented an MPC framework that is ro-
bust against additive uncertainty. The new framework is a closed-loop
RMPC utilizing uncertainty knowledge to enhance the nominal MPC.
The RMPC is capable of maintaining the temperature within the com-
fort zone for model uncertainties up to 75%. We proposed a new un-
certainty feedback parametrization of the control input, TLDS, for the
closed loop RMPC which results in the same energy and discomfort
indices as the previous parametrization, LTS, with a lower number of
decision variables, linear in time horizon N, as opposed to quadratic, for
the LTS. The new TLDS parametrization results in an average simula-
tion time of 30% less than LTS. Closed-loop RMPC outperforms nom-
inal MPC considering the provided level of comfort. However, higher
comfort comes at the cost of dramatically higher energy consumption
for RMPC. For uncertainty range of 30% to 67%, RMPC leads to bet-
ter overall performance compared to MPC and RBC, although it fails
to provide a better energy-comfort trade-off if model uncertainty is less
than 30% or more than 67%. We proposed a new performance index
(IOP ) to assess buildings’ energy consumption and comfort level si-
multaneously. The IOP index is used for evaluating different building
controllers. IOP index can be used to generate a guideline for choosing
appropriate controller type for buildings. This can be helpful to the
building control community for deciding on a proper controller type
based on how accurate an available building model is for model-based
controller design. We found that the best choice for controller type
changes from MPC to RMPC, and then finally to RBC as the model
uncertainty increases. A typical RBC controller outperforms model-
based controllers (MPC and RMPC), if building model uncertainty is
above 67%.
We also derived and formulated exergy destruction as a function
of the physical parameters of the building. An optimal control prob-
lem is formulated to minimize exergy destruction rate. The beneficial
new aspects of MPC problem based on exergy is optimizing the use of
low quality energy (low exergy) for HVAC systems and hence decreas-
ing irreversible entropy generation. Our findings show that the MPC
controllers outperform the rule-based on-off controller for HVAC sys-
tems. Compared to RBC, energy-based MPC results in 18% and 24%
reduction in exergy destruction and energy consumption, respectively.
XMPC enhances the results from energy-based MPC by offering 22%
reduction in exergy destruction and 36% reduction in HVAC energy
consumption compared to RBC. Our results show the supremacy of
XMPC compared to conventional energy-based MPC since it consumes
12% less energy and saves 4% more exergy. The benefits from XMPC
stem from reduction of sources of irreversible entropy generations by
controlling the HVAC system variables that affect heat transfer, in-
ternal entropy generation rate, and exergy flows into the zones. The
implementation of the proposed XMPC is the same as conventional
energy-based MPC. The only difference is the change in the objective
function. Details for implementation of conventional MPC are exten-
sively found in references Bengea et al. [2014]. The implementation of
XMPC on a real building test-bed is an interesting avenue for future
research.
5
Test-beds and Real-scale Experiments
As the theoretical results of using novel control algorithms were deemed

to be promising, researchers started developing large-scale test-beds
to implement such new control techniques on real buildings. In this
chapter we review some of the major efforts in this domain and present
some outstanding findings of studies such as how effective new control
techniques are when implemented on real buildings at scale.
5.1 Review of Experimental Building Studies
Various large-scale experiments have been performed to study the effect

of different building control designs. Effects of new control algorithms
can be evaluated in three ways. 1) The same test-bed is operated once
with the existing control logic and once with the novel control algorithm
implemented on the system. 2) Test-bed is used to train or calibrate a
descriptive model of the system using a preferred simulation environ-
ment/tool. The simulation model is then used to simulate two scenarios,
one with a simple control strategy and one with the proposed control
algorithm. 3) Test-bed is used to train or calibrate a descriptive model
of the system using a preferred simulation environment/tool. The sim-
77
78 Test-beds and Real-scale Experiments
ulation model is then used to simulate the building performance with

the novel control algorithm while performance of the existing controller
is obtained directly from the test-bed measurements. Using any of the
three approaches above, performances of the system under two con-
trollers are then compared, controlling for factors other than “control
action”. To evaluate the performance, metrics such as cost and con-
straint satisfaction of a particular control input are used. While the
costs and comfort constraints are mostly similarly defined across the
studies, the choice of the model is arguably the most distinctive feature
in an experiment. Sturzenegger et al. [2015] reviewed studies in which
the models ranged from building simulation software models (e.g., En-
ergyPlus or ESP-r ESPr) and Artificial Neural Networks (ANNs) to
the more commonly known linear and nonlinear state-space ones.
Earlier works on experimental building control used building sim-
ulation software models. A practical proof of this concept was first
reported in Clarke et al. [2002]. The authors used an ESP-r model to
optimize the starting time of a radiator in an unoccupied test cell in a
3-hour-long experiment. Liu and Henze [2006a] extended the idea to a
more complex HVAC setup in a four-day experiment using a TRNSYS
model TRNSYS. The primary system consisted of two chillers and an
active thermal energy storage, while the secondary system comprised a
ventilation unit serving two unoccupied test rooms. Schuss et al. [2011]
reported MPC experiments conducted over two months in two occu-
pied Austrian office buildings mainly equipped with window shading
and opening devices. In both cases, HAMBase HAM [2015] and radi-
ance Rad [2015] models were used. Experimental application of MPC
using ANN model was reported in Ferreira et al. [2012]. Several exper-
iments were conducted in four occupied rooms of a university building
equipped with independent Air Handling Units (AHUs) over a total
experiment time of around three days.
To the best of our knowledge, the experimental application of an
MPC using a state-space model has first been published in Liao and
Dexter [2010]. The authors report the control of a boiler supplying two
commercial buildings equipped with radiators over a period of 40 days.
MPC was based on a low-order linear state-space model predicting a
5.2. Review of Large-scale Test-beds for Building Studies 79
building-wide average room temperature. Širokỳ et al. [2011] report

the control of the TABS of an eight-floor building block of a uni-
versity building over three months. Their setup provided the unique
opportunity to systematically compare the MPC’s performance to a
baseline controller running in an identical nearby building block. The
authors used a low-order linear state-space model. Aswani et al. re-
port an experimental proof of concept of an adaptive MPC approach
where in each step the internal model was improved based on the mea-
surements. The scheme used a scalar linear state-space model and was
applied to an air conditioning unit of a student computer room. Ma
et al. [2010], used a nonlinear model in an MPC framework controlling
the chillers supplying a large campus-wide cold water storage tank over
two five-day periods. The buildings’ actuators were not controlled; in-
stead, the total campus cooling demand was estimated as a function of
the weather forecast and then considered as a predictable disturbance.
Castilla et al. [2014] report experimental results obtained over 6 hours
in a single test room actuating the mass flow and cooling of a fan-coil
unit using a linear state-space model. Bengea et al. [2014] report MPC
experiments over three weeks, in which a centralized AHU supplying
a 650 m2 space of a research facility was controlled using a nonlin-
ear model. All the above studies report a successful operation of MPC
and efficiency improvements when compared with baseline control. The
experiment durations and the numbers of controlled zones varied sig-
nificantly across the studies. However, all focused on the control of a
single HVAC actuator. In general, all of the studies aimed at demon-
strating MPCs benefits but lacked a discussion of the development and
implementation costs.
5.2 Review of Large-scale Test-beds for Building Studies
In what follows we describe some of the notable test-beds that have

been developed by various organizations for the purpose of testing and
validating control algorithms on buildings.
5.2.1 UC Berkeley
The campus of UC Berkeley has been subject to many studies and ex-
periments where researchers such as Maasoumy et al. [2014c], Aswani
et al. either used the data from BEMS for building model training or
implemented their control algorithms on the system by directly over-
writing the control logic or set-points of the system.
LoCal
The Berkeley LoCal project has produced a network architecture for

localized electrical energy reduction, generation, and sharing by exam-
ining how pervasive information can fundamentally change the nature
of these processes Katz et al. [2011]. A key component of this is the
use of sMAP Dawson-Haggerty et al. [2010] to exchange physical data
about the systems involved. This allows producers of physical informa-
tion to directly publish their data in a format for consumption by a
diverse set of clients.
The LoCal Energy Network is a cyber overlay on the energy dis-
tribution system in its various physical manifestations, e.g., machine
rooms, buildings, neighborhoods, isolated generation islands and re-
gional grids. Pervasive information exchange will enable a more efficient
scalable energy system with improved resilience and quality of delivered
power. LoCal’s key contributions include bringing about pervasive in-
formation about energy availability and use, interactive load/supply ne-
gotiation protocols, and controllable loads and sources. Together these
yield a system for agile, distributed, and integrated management of
energy that can buffer energy on the path to reduce peak-to-average
energy consumption, moderate infrastructure provisioning, and encour-
age power-limited design and operation. The main building block is
the intelligent power switch, logically connecting sources to loads by
bundling information (bits) with energy (electrons) flows.
Figure 5.1: Berkeley Retrofitted and Inexpensive HVAC Test-bed for Energy Effi-
ciency (BRITE).
Berkeley Retrofitted and Inexpensive HVAC Test-bed for Energy

Efficiency (BRITE)
BRITE is a system for testing different control strategies on an AC

unit that cools a computer laboratory on the Berkeley campus, and
it is shown in Figure 5.1 Though it is built using commodity parts,
the computers can be replaced with microcontrollers. The strength of
this structure is that it scales to building-wide systems. Moreover, the
proposed MPC schemes Aswani et al. are computationally scalable be-
cause of their convexity. In this test-bed, the LoCal server gathers sen-
sor data and stores this in a simple measurement and actuation profile
(sMAP) database Dawson-Haggerty et al. [2010]. A control computer
accesses the Internet and LoCal server to get weather forecasts and
sensor data, and it runs a learning-based MPC scheme that computes
a control input that is sent through the LoCal server to the thermostat.
The thermostat transmits a corresponding signal to the AC.
5.2.2 SinBerBEST
Singapore-Berkeley Building Efficiency and Sustainability in the Trop-

ics (SinBerBEST) is an interdisciplinary group of researchers from UC
Berkeley, Nanyang Technological University (NTU), and the National
Figure 5.2: SinBerBEST facilities located in both Berkeley, California and in Sin-
gapore.
University of Singapore (NUS) who come together to make an impact

with broadly applicable research leading to the innovation of energy ef-
ficient and sustainable technologies for buildings located in the tropics,
as well as for economic development.
SinBerBEST researchers are located in both Berkeley, California
and in Singapore. In addition to traditional office space, the CREATE
campus in Singapore houses dry and wet lab spaces and a state-of-the-
art cyber physical test bed.
Additionally, SinBerBEST has facilities within the Electrical En-
gineering building in Cory Hall at UC Berkeley. In this space, they
utilize the office as a living test bed in which researchers determine
what types of environmental conditions are preferable through a social
game that Konstantakopoulos et al. [2015] designed for encouraging
energy efficient behavior among building occupants with the aim of
reducing overall energy consumption in the building.
Research Areas within SinBerBEST
• Thrust 1: Sensing, Data mining and Modeling Modeling

and data mining, Distributed Sensing and Cooperative Control,
Human activity sensing and inference.
• Thrust 2: Multi-Level Optimal Control Active chilled beam

systems for tropical buildings, Renewable energy cooling system
for tropical buildings, Total dynamic model for control and opti-
mization, Dynamic control and optimization, Distributed sensing
and cooperative control, Resource management.
• Thrust 3: High Confidence Building Operating System

BAC – Modeling and controller design, Complete design flow for
BAC systems, Wireless sensor network integration to BAS, Intel-
ligent power switches and their control, Hybrid DC/AC building
power distribution, Efficient information network for intelligent
buildings, Energy efficient applications for smart buildings.
• Thrust 4: Human-Building Interaction and the Envi-

ronment Air-conditioning and mechanical ventilation system
hygiene, Human factors for energy-efficient thermal comfort,
Bioaerosols in tropical buildings, Temperature and humidity ef-
fects on pollutant dynamics, Influence of ventilation indoor envi-
ronmental quality.
• Thrust 5: Materials, Designs and Lifecycle Integrated ma-

terial and structural design approach, Structural facades and life-
cycle assessment for tropical buildings Developing sustainable ul-
tra high performance concrete (S-UHPC), Nano-scale optimiza-
tion of green concrete.
• Thrust 6: Test Bed Integration Investigating existing test-

bed opportunities in Singapore and UCB, Test-bed frameworks
for cyber-physical approach and assessment.
5.2.3 Energy Resource Station (ERS)

The experiment was carried out in the Energy Resource Station, oper-
ated by the Iowa Energy Center (IEC). The ERS is a unique demon-
stration and test facility, where laboratory-testing capabilities are com-
bined with real building characteristics. The ERS is capable of simulta-
neously testing two full-scale commercial building systems side-by-side
with identical thermal loading. The ERS building, a single-story struc-
ture with a concrete slab-on-grade, has a height of 4.6 m and a total
Figure 5.3: Layout of the test facility at the Energy Resource Station (ERS),
Ankeny, IA Liu and Henze [2006b].
floor area of 855 m2 . The building is divided into a general area (of-
fice space, service rooms, media center, two classrooms, etc.), and two
sets of identical test rooms, labeled A and B, adjacent to the general
area. The eight test rooms are organized in pairs with three sets of
zones having one exterior wall (east, south, and west) and one set that
is internal. Figure 5.3 depicts a layout of the ERS including the four
sets of identical test rooms used for the experiment. The test facility
has a central heating plant, consisting of a natural gas-fired boiler, and
a cooling plant with three nominal 35 kW air-cooled chillers that op-
erate in both chilled-water and ice-making modes. The chilled-water
loop is filled with 22% propylene glycol water solution. In addition, the
building includes a 440 kWh internal melt ice-on-tube thermal energy
storage tank as well as pumps and auxiliary equipment needed to pro-
vide cooling. Hence, several modes of operation between these sources
of cooling are possible in order to supply chilled-water to the AHUs.
5.2.4 BubbleZERO
BubbleZERO, shown in Figure 5.4, is an experimental building and

laboratory, and is conceived as part of the low exergy module devel-
opment for Future Cities Laboratory (FCL). The BubbleZERO was
constructed by ETH Zurich with concrete floor and ceiling as well as
LowEx systems installed along with an experimental chiller. It is cur-
rently installed on the NUS campus in Singapore. It provides extensive
opportunities to test and evaluate the performance of sensing and con-
trol systems as well as low energy systems.
Deployed instruments with a network of wireless sensors provides
instant and interactive information on multiple dimensions of the build-
ing environment. Figure 5.5 shows the system structure. The system
is built based on the open source wireless sensor network platform
Telosb Polastre et al. [2005], and the software is developed based on
the TinyOS Levis et al. [2005] operating system. Each sensor node is
powered with an embedded CPU, equipped with a wireless commu-
nication module for intercommunication, and various interfaces and
sensors.
The system contains 4 sub-systems as shown in Figure 5.5. The
Environment sensing system senses the indoor and outdoor ambient
parameters. Temperature and humidity are monitored for both indoor
and outdoor, while CO2 concentration is collected for indoor. The en-
vironment sense system includes 8 indoor sensors, 4 CO2 concentra-
tion sensors and 4 outdoor sensors. The 8 indoor sensors (Telosb41-
Figure 5.4: BubbleZERO test-bed
Telosb48) are hanged at the ceiling panels. They measure tempera-

ture and humidity at different locations in the room. The 4 CO2 flaps
(CO2flap31-CO2flap34) are installed at the ceiling to measure the in-
door CO2 concentration and also to control the air exhaust for the
room.
5.2.5 UC Merced
The UC Merced chilled water plant includes three water-cooled chillers
and a two million gallon chilled water storage tank. The tank is charged
during the night to minimize on-peak electricity consumption and take
advantage of the lower ambient wet bulb temperature. A model predic-
tive control algorithm was developed for the UC Merced campus chilled
water plant 5.6.
The control algorithms determined the optimal chilled water plant
operation including chilled water supply temperature set-point, con-
denser water supply temperature set-point and the charging start and
stop times to minimize a cost function that includes energy consump-
Figure 5.5: Sensor deployment in the test-bed
Figure 5.6: Schematic Diagram of UC Merced Chilled Water System.
tion and peak electrical demand over a 3-day prediction horizon.

A detailed model of the chilled water plant and simplified mod-

els of the buildings served by the plant were developed by Ma et al.
[2010] using the equation-based modeling language Modelica. Steady
state models of the chillers, cooling towers and pumps were developed,
based on manufacturers’ performance data, and calibrated using mea-
sured data collected and archived by the control system. A detailed
dynamic model of the chilled water storage tank was also developed
and calibrated. Simple, semi-empirical models were developed to pre-
dict the temperature and flow rate of the chilled water returning to the
plant from the buildings. These models were then combined and simpli-
fied for use in a model predictive control algorithm that determines the
optimal chiller start and stop times and set-points for the condenser
water temperature and the chilled water supply temperature.
The experimental results show a small improvement in COP over
the baseline policy but it is difficult to draw any strong conclusions
about the energy savings potential for MPC. With this system only
four days of suitable experimental data were obtained once correct
operation of the MPC system had been achieved. These data show an
improvement in COP of 3.1% ±2.2% relative to a baseline established
immediately prior to the period when the MPC was run in its final
form. This baseline includes control policy improvements that the plant
operators learned by observing the earlier implementations of MPC,
including increasing the temperature of the water supplied to the chiller
condensers from the cooling towers.
During the course of the project, the team uncovered various prob-
lems with the chilled water system, whose effect is estimated to lead to
as much energy waste as the result of running an inefficient controller.
Although the types of problems uncovered and the level of energy sav-
ings may differ significantly from one building to another, some of the
benefits of detecting and diagnosing problems are expected from the
use of MPC for any chilled water plant.
5.2.6 OptiControl
OptiControl Opt [2015] is a project that started in 2007 between ETH

Zurich, Siemens Switzerland Ltd., Empa Dubendorf, MeteoSwiss, and
Gruner AG . OptiControl combines the newest developments from the

fields of building technologies, numerical weather forecasting and con-
trol engineering. The project develops and tests novel, predictive con-
trol approaches plus corresponding software modules to be incorporated
in commercial building automation systems. In the first phase of the
OptiControl project, which was from 2007 to 2010, specialists from the
participating institutions worked together on the development of novel,
predictive control strategies for buildings.
The project focused on the application Integrated Room Automa-
tion (IRA). IRA deals with the automated control of blinds, electric
lighting, heating, cooling, and ventilation of an individual building
zone. For the development and performance assessment of the new
control strategies a general methodology was developed. It is based on
computer modeling and simulation and can be transferred also to other
applications.
While Phase I of the OptiControl project was entirely based on ex-
tensive simulation studies, in Phase II the developed control strategies
were implemented in a representative Swiss office building to ensure
their testing and evaluation in a realistic setting. The project team in-
cluded participants from the ETH Zurich, Siemens Switzerland Ltd.,
and Gruner AG. The weather forecasts are provided by MeteoSwiss.
Phase II of the project went from 2011 to 2013.
Sturzenegger et al. [2015] report the final results of the predictive
building control project OptiControl-II that encompassed seven months
of MPC on a fully occupied Swiss office building. In the experiments,
MPC controlled the building reliably and achieved a good comfort level.
The simulations suggested a significantly improved control performance
in terms of energy and comfort compared to the previously installed
industry-standard control strategy. However, for similar buildings and
with the tools currently available, the required initial investment is
likely too high to justify the deployment in everyday building projects
on the basis of operating cost savings alone. Nevertheless, development
investments in an MPC building automation framework and a tool
for modeling building thermal dynamics together with the increasing
importance of demand response and rising energy prices may push the
technology into the net benefit range. Experiences with state-of-the-art

integrated predictive RBC developed within the OptiControl-II project
by the Siemens engineers were reported in Gwerder et al. [2013].
The building used in this experiments is located in Allschwil, close
to Basel, Switzerland. The building was constructed in 2007 and has a
total conditioned floor area of 6000 m2 consisting of five upper floors
of offices and restaurant on the ground floor. The MPC strategy was
applied to the upper five floors, while the ground floor was separately
actuated. A TABS is the main heating and cooling actuator, supplied
by a gas boiler and a cooling tower. The entire building is served by
a single Thermally Activated Building System (TABS) zone, i.e., the
circulating water’s mass flow rate and supply water temperature are
determined globally for the entire building. A central AHU supplies
the offices with fresh air. It includes a heat exchanger for return air
heat/cold recovery, a heating coil in the supply air, and an evaporative
cooler in the return air. The supply air temperature and mass flow
rate are again determined globally for the entire building. On each
floor, the air is supplied to the offices on the outer parts of the floors
and returned from the rooms in the center of the building. Natural
ventilation by manual opening of windows is possible in all office rooms.
In addition, in the corner offices, radiators are available whose supply
water temperature is controlled. The gas boiler provides all heating
energy for the TABS, the AHU heating coil, and the radiators. The
cold water for the TABS is generated by a hybrid cooling tower.
Several wireless room temperature and window contact sensors, as
well as electric load meters, TABS, and AHU heating/cooling power
meters were installed at the beginning of the project to enable the
thorough evaluation of the control experiments and the validation of
building models as well as to support the newly developed RBC and
MPC control strategies. Moreover, the blinds control was integrated
into the Building Automation System (BAS). Finally, an industry PC
was setup for running the RBC and MPC control algorithms and an
external database was established to monitor and analyze the building’s
operation.
The study performed by Sturzenegger et al. [2015] differs from the
other studies presented earlier in this chapter in several respects. First,

they considered the integrated simultaneous control of several actua-
tors (TABS, ventilation, and blinds). Their choice was guided by the
results of numerous simulation studies that had shown that the bene-
fits of MPC increase with an increasing complexity of the control task
at hand. Second, unlike a majority of the above studies (exceptions
being Liao and Dexter [2010], Širokỳ et al. [2011], Ma et al. [2010],
and Bengea et al. [2014]), they did not consider only test cells or in-
dividual rooms, but an entire, fully operational building. Third, they
did not manipulate or replace the existing control hierarchy, but rather
we introduced an additional level of supervisory control. Fourth, they
performed long-term experiments in both, the heating and the cool-
ing seasons. Finally, the goal of the project was from the beginning
that the resulting control solutions can later be easily incorporated in
commercial workflows and BASs.
5.2.7 Michigan Tech’s Lakeshore Center
Michigan Tech’s Lakeshore Center is a three-story building with area

of 5700 square meter. This building was renovated in 2009 with new
ground-source heat pumps (GSHP) installed in each heating zones. The
Building Management System (BMS) of Lakeshore employs Niagara
AX, a web-based real-time software framework, to control the climate of
Lakeshore center building. GSHPs provide exergy required for heating
and cooling in this three-story building. Each zone is equipped with a
heat-pump to maintain comfort for occupants. Figure 5.7 illustrates a
schematic of a heat-pump in a zone in the test-bed. GSHPs exchange
heat between the earth and the heating/cooling zones of the building.
GSHPs are considered as HVAC renewable energy efficient technologies
that have a high COP. The nominal COP of our test-bed heat-pump
is 3.2. GSHP units in the test-bed have a multi-speed fan that supplies
hot or cold air to the zone when heating or cooling is needed.
The existing HVAC controller is a typical on-off controller to main-
tain desired temperature for each zone. When zone temperature exceeds
the lower limit, the compressor of heat-pump is switched on until tem-
perature reaches the comfort level. The zone temperature is measured
Figure 5.7: Testbed schematic with a ground-source heat-pump. Sensor 1-a shows
the BMS temperature sensor and sensor 1-b belongs to the data-logger.
and logged by a combination of a temperature data-logger (shown in

Figure 5.7) and a built-in temperature sensor with the accuracy of
±0.2◦ C as part of the BMS.
We reviewed some of the major efforts in building large-scale experi-

mental test-beds to study the effect of novel building technology and
control design on real buildings. We covered test-beds ranging from
room-size to multiple buildings geographically located in institutions
in the US, Europe and Asia. In Table 5.1, we summarize the test-beds
introduced in this section using metrics such as location, size, build-
ing type, institutions involved in the experiment, heating and cooling
systems and actuation system.
Test-bed Location Size Building Institutions Heating Cooling Actuation system
type involved system system
UC Berkeley Berkeley, multiple university UC Berkeley gas boiler cooling supply air temperature,
CA buildings buildings towers mass flow rate and local
reheat at VAV level
SinBerBEST Berkeley, multiple university UC Berkeley, gas boiler cooling supply air temperature,
CA and buildings buildings NTU, NUS towers mass flow rate and local
Singapore reheat at VAV level
UC Merced Merced, CA multiple university UC Merced, gas boiler chillers, chilled water supply tem-
buildings buildings UC Berkeley, chilled perature, condenser wa-
LBNL, UTRC water ter supply temperature,
storage charging start and stop
tank times
OptiControl Allschwil, 6000 m2 office ETH Zurich, gas boiler cooling thermally activated build-
Switzerland rooms Siemens, tower ing system, AHU for heat-
Gruner, Empa ing and cooling, blinds
Dubendorf,
MeteoSwiss
MichiganTech Houghton, 5700 m2 office MichiganTech ground ground heat-pump compressor
MI rooms source source
heat heat
pumps pumps
ERS Ankeny, IA 855 m2 office Iowa Energy gas boiler chillers, supply air temperature,
rooms Center, Iowa ice stor- storage system charge and
State University age tank discharge rates
BubbleZERO Singapore 20 m2 un- ETH Zurich, - hydronic decentralized air supply
occupied NUS radiant units, shades
cooling
Table 5.1: Comparison of large-scale test-beds.

6
Designing Building Control Systems as
Cyber-Physical Systems
The design of HVAC systems is crucial for reducing energy consump-

tion in buildings. As complex cyber-physical systems, HVAC systems
involve three closely-related subsystems – the control algorithm, the
physical building and environment, and the embedded implementation
platform (Figure 6.1). In the traditional top-down approach, the design
of the HVAC control algorithm is done without explicit consideration of
the embedded platform. The underlying assumption is that the compu-
tation and communication capabilities of the embedded platform are
sufficiently performing for any type of control mechanism. However,
with the advent of more complex HVAC control algorithms for energy
efficiency, the use of distributed networked platforms, and the imposi-
tion of tighter requirements for user comfort, this assumption on the
embedded platform is no longer valid. Various aspects of the platform,
including sensor accuracy and availability, communication channel re-
liability, and computing power of embedded processors, may have a
significant impact on the quality and cost of a BAS. For instance, a
BAS with MPC algorithm as its control logic has different compu-
tation power requirements compared to a BAS with simple bang-bang
controller as it’s control logic. Thus, the design of the control algorithm
95
96 Designing Building Control Systems as Cyber-Physical Systems
Physical
Components & Control
Environment Algorithm
Cyber-
Physical
System
Embedded
Platform
Figure 6.1: As complex cyber-physical systems, HVAC systems involve three

closely-related subsystems - the control algorithm, the physical environment and
the embedded implementation platform.
should take into account the configuration of the implementation plat-

form and vice versa, i.e., the control algorithm and the embedded plat-
form should be co-designed, as proposed by Maasoumy et al. [2013b].
Part of the problem that this chapter addresses is how to model
uncertainty which itself can be due to randomness in the system, un-
predictability of some system components or data quality due to sensor
placement, affect the performance of a model-based controller such as
MPC. This uncertainty propagation problem has recently been studied
by several researchers. Bengea et al. [2011] present a methodology that
establishes achievable targets for parameter estimation errors based on
closed-loop performance sensitivity. The control algorithm uses per-
turbed building-model parameters. They perform simulations to esti-
mate the dependency of energy cost and constraint infringement time
97
on the magnitude of these perturbations. The simulation results are

then used to define targets for the parameter estimation errors, which in
turn are applied to specify the character of excitation and model struc-
ture used for identification. Eisenhower et al. [2012] perform sensitivity
and uncertainty analysis of about 1000 parameters of a building model.
They extend traditional sensitivity analysis in order to decompose the
pathway as uncertainty flows through the dynamics, and they identify
which internal or intermediate processes transmit the most uncertainty
to the final output. The proposed results are presented as a method that
is applicable to many different modeling tools, and its applicability is
demonstrated on an EnergyPlus model. Behl et al. [2014] present the
Model-IQ toolbox that, given a plant model and real input data, au-
tomatically evaluates the effect of the uncertainty propagation from
sensor data to model accuracy to controller performance. They then
apply the Model-IQ uncertainty analysis for model-based controllers in
buildings to demonstrate the cost-benefit of adding temporary sensors
to capture a building model. Peleš et al. [2012] present a computation-
ally efficient process for uncertainty quantification, sensitivity analysis
and automated calibration of building models. They demonstrate their
methodology using an energy simulation model of a medium sized of-
fice building. The reduced order model (or meta-model) obtained as
part of the process is reused for automated calibration of the high fi-
delity building energy model from which it was derived. The advantage
of meta-model is that it returns function evaluation in a fraction of a
second as opposed to the high fidelity model, which takes several min-
utes to simulate. Furthermore, the meta-model is obtained in a simple
analytical form. Consequently, it can be easily manipulated and used
in various optimization algorithms. This approach provides a reliable
and fully automated way to assess first order parametric sensitivities
in building models. Higher-order sensitivities are computationally more
expensive to evaluate.
Once we identify how model uncertainty affects the performance of
a model-based controller such as MPC, then we can solve the co-design
problem that we introduced earlier in this chapter. Here we review our
co-design approach presented in Maasoumy et al. [2013b] that ana-
lyzes the interaction between the control algorithm and the embedded
platform through a set of interface variables – in particular sensing
accuracy. Six control algorithms are analyzed that take into account
the sensing error, and the relation of control performance and cost is
modeled versus sensing error. The relation of embedded platform cost
versus sensing error is captured by analysis of the collected data from a
test-bed. Based on these models, the co-design of the control algorithm
and the temperature sensing subsystem of the embedded platform is
performed to optimize with respect to energy cost and monetary cost
while satisfying the constraints for user comfort level.
6.1 The Co-design Problem
Smart buildings today have sophisticated and distributed control sys-

tems as part of a BAS. The task of a BAS is to maintain building
climate within a specified range, control the lighting based on the oc-
cupancy schedule, and monitor the system performance and failures.
To accomplish these tasks, a BAS has to deal with computation and
communication non-idealities stemming from the distributed nature of
the implementation platform.
Here we analyze the performance and energy cost of six HVAC
control algorithms under different assumptions on temperature sensing
accuracy. Indeed, we observed in our experiments that the accuracy of
sensing data has significant impact on the control algorithms, to the
extent that different algorithms should be chosen depending on the dif-
ferent sensing accuracy of the implementation platform. Based on this
observation, we propose a framework to co-design the HVAC control
algorithm with the part of the embedded platform that directly affects
sensing accuracy – specifically the choices of sensor locations and the
number of sensors. We believe this approach can be extended to a gen-
eral framework for co-designing the control algorithm and the entire
embedded platform. Figure 6.2 shows the structure of the co-design
framework, with blocks in parentheses representing the focus of this
paper. The key aspects of the co-design process are:
1. Identifying the interface variables that are significant to both the

6.1. The Co-design Problem 99
Control constraints
and objectives
(energy cost, user Design space exploration
comfort)
Control algorithm design

(controller type, parameters)
..
Platform constrains Interface variables ..
and objectives (sensing accuracy) .
(monetary cost)
Pareto front of
optimal designs
Embedded platform design
(number of sensors, locations)
Platform library
(available sensors)
Figure 6.2: Co-design framework for HVAC systems
control algorithm and the embedded platform,
2. Designing the control algorithm with consideration of the inter-

face variables and modeling the relation of control metrics (e.g.
control performance) versus the interface variables,
3. Capturing the relation of platform metrics (e.g. platform mone-

tary cost) versus the interface variables, and
4. Co-exploring the design of the control algorithm and the platform

through the interface variables.
Here we focus on temperature sensing accuracy, which on one side

significantly affects the control algorithm and on the other directly
relates to the design of the sensing system. During the exploration of
the design space, for any given monetary budget, we find the optimal
choice of sensor locations and the number of sensors to maximize the
accuracy of sensing data. Sensing accuracy is then used for the selection
of the HVAC control algorithm (from the six candidate controllers) to
minimize energy cost while satisfying user comfort requirements. We set
the monetary budget to a range of different values, and find the design
that consumes the minimal energy for each given budget. The result
of this design space exploration is a Pareto front of optimal monetary
cost and energy cost.
Various models and control algorithms have been proposed in the
literature for HVAC systems Maasoumy [2011], Ma et al. [pp. 1-
65, 2011], Oldewurtel et al. [Baltimore, MD, 2010], Maasoumy and
Sangiovanni-Vincentelli [2012], Maasoumy et al. [2011]. In a recent
work, unscented Kalman filtering has been used for online estimation
of building thermal parameter estimation Radecki and Hencey [2012].
The focus of these papers is on physical modeling and control design
without taking into account the limitations of the embedded platform.
For instance, Maasoumy and Sangiovanni-Vincentelli [Fort Lauderdale,
FL, 2012] propose a robust control algorithm against weather and oc-
cupancy prediction uncertainties. However, the uncertainties from the
embedded platform measurements such as sensor errors are not ad-
dressed. Yang et al. [2010, 2012] focus on the design of the embedded
software and hardware for a given control algorithm, thus not address-
ing design space exploration for optimal HVAC system design.
6.2 Sensing and Prediction Accuracy Modeling
We use the model that was proposed in Section3.1 where a building is

modeled as an RC network.
6.2.1 Sensing Accuracy
Here we focus on temperature sensing accuracy to determine the noise

characteristics of the indoor temperature (xk in Equation (3.11)), which
drives the design of the control algorithms (see Chapter 4).
Measurement inaccuracies of individual sensors can be categorized
into the following three types: no information, completely incorrect
information, and incorrect but in-range information. The error of the
6.2. Sensing and Prediction Accuracy Modeling 101
indoor temperature estimation, denoted by rt , is affected by accuracy

of individual sensors, number and locations of sensors in a thermal
zone, and physical properties of the building. Statistics of rt may be
extracted from historical data as shown in Section 6.3.1, and we assume
it is additive to the temperature measurement.
Accordingly, temperature measurement in Equation (3.11) is up-
dated as
zk = Cxk + F rt
k (6.1)
where F is a matrix of proper dimension and zk is the temperature
reading at time k.
6.2.2 Prediction Accuracy

Disturbance prediction (dˆk in Equation (3.11)) in our model depends
on the prediction of CO2 concentration level in the room (Ψ) and on
the prediction of ambient air temperature (Tout ) as shown in Equa-
tion (3.10). We use ck to denote the error in CO2 level prediction, and
hence the predicted CO2 level is
Ψ(k) = Ψ̂(k) + ck (6.2)
and ot
k to denote the error in ambient air temperature prediction, and
hence the predicted ambient temperature is
Tout (k) = T̂out (k) + ot

k (6.3)
Based on Equation (3.10), disturbance prediction can be expressed

using CO2 and ambient temperature prediction error, as
dˆk = āT̂out (k) + b̄Ψ̂(k) + ē

= ā(Tout (k) − ot c
k ) + b̄(Ψ(k) − k ) + ē (6.4)
= dk − (āot c
k + b̄k )
where dk denotes the predicted disturbance at time step k. The above

equation suggests that the disturbance prediction error, denoted by
wk , is a linear combination of the CO2 prediction error, ck , and the
ambient air temperature prediction error, ot
k . This leads to the fol-
lowing state update and output measurement equation (modified from
Equation (3.11))
xk+1 = Axk + Buk + E(dk − wk )

zk = Cxk + F vk (6.5)
where wk = −ξ(āot c
k + b̄k ) and the constant ξ is a function of the
discretization method and discretization step of the continuous system
dynamics, and vk = rtk (i.e. the temperature measurement error). The
CO2 and ambient air temperature forecast errors are uncorrelated ran-
dom variables with variance σc and σot , respectively. Hence, variance
of wk is calculated by
σw = E[(w − ŵ)(w − ŵ)T ]

n o
ot + b̄˜
= E [−ξ(ā˜ c )][−ξ(ā˜
ot + b̄˜
c )]T (6.6)
= ξ 2 (ā2 σc + b̄2 σot )
where ˜ot = ot − ˆot and ˜c = c − ˆc . In this paper, we assume typical
values for σot , σc and other constants in simulations: σot = 2, σc = 50,
ā = 0.01, b̄ = 0.06 and ξ = 10.
If CO2 and temperature sensors are deployed to facilitate the pre-
dictions of CO2 level and ambient air temperature, the values of σot
and σc will be largely affected by the choice of CO2 and temperature
sensor type, number and locations.
6.2.3 Control Cost and Performance vs. Sensing Accuracy

We consider three controller strategies: on-off1 , MPC and RMPC, and
combine each of these three controllers with either Extended Kalman
Filter (EKF) or UKF. For each controller, we simulate the energy cost
and performance (measured by a discomfort index)2 with respect to
different levels of sensing accuracy.
Figure 6.3 and Figure 6.4 show the energy cost and discomfort in-
dex of the six controllers with either EKF or UKF under different levels
of sensing accuracy from 0% to 100%. Note that we use a normalized
1
Other controllers such as PI could also be used here.
2
For detailed discussion on energy cost and discomfort index, refer to Maasoumy
and Sangiovanni-Vincentelli [2012].
6.2. Sensing and Prediction Accuracy Modeling 103
18000
16000
14000
Energy Cost (kWh)
12000
EKF-MPC
10000 EKF-RMPC
8000 EKF-OnOff
UKF-MPC
6000
UKF-RMPC
4000
UKF-OnOff
2000
0
0 25 50 75 100
Temperature measurement Noise (%)
Figure 6.3: Energy cost vs. measurement noise
0.6
0.5
Discomfort Index (°Ch)
0.4
EKF-MPC
EKF-RMPC
0.3
EKF-OnOff
UKF-MPC
0.2
UKF-RMPC
0.1
UKF-OnOff
0
0 25 50 75 100
Temperature measurement Noise (%)
Figure 6.4: Discomfort index vs. measurement noise

uncertainty based on the maximum value of the disturbance. For in-

stance, measurement error of 50% corresponds to ||w||∞ = 0.5 ∗ ||d|| ˆ ∞.
More details can be found in Maasoumy and Sangiovanni-Vincentelli
[Fort Lauderdale, FL, 2012].
From these results, we can see that some controllers are always
superior to other controllers in terms of both energy cost and discomfort
index. For instance, the two MPC controllers are better than the two
on-off controllers. However, in other cases, particularly among MPC
and RMPC controllers, choosing which controller to use depends on
the design requirements and the measurement noise and error, which is
affected by the embedded platform. For instance, let us assume there is a
requirement that the maximum discomfort level should be under 0.3. If
we choose a number of accurate sensors and we estimate that noise will
be under 25%, then we select EKF-MPC since it consumes the least
amount of energy while satisfying the requirement. If we choose less
accurate (or fewer) sensors to reduce platform cost and we estimate
that noise will be around 75%, then we select UKF-RMPC since it
consumes the least energy among the schemes that satisfy the comfort
level requirement (the other being EKF-RMPC).
6.3 Sensing System Design and Accuracy
The relation between sensing system design and temperature sensing

accuracy, i.e. the temperature measurement error vk in Equation (6.5),
is quite important in the overall performance of the HVAC system. We
collect measurements from multiple temperature sensors in a test-bed
and analyze their statistics to estimate sensing accuracy under different
number of sensors and locations. This approach may be applied in prac-
tice if the designer has access to the target building (or buildings/test-
beds with similar characteristics) and can deploy sensors for testing,
otherwise simulation approaches such as computational fluid dynamics
(CFD) analysis may be applied.
6.3. Sensing System Design and Accuracy 105
6.3.1 Analysis of Historical Data
We collected data from test-bed for one week period (July 16-23, 2012).
Sensor readings from 8 indoor temperature sensors located in the test-
bed were collected. The 8 sensors were spread out to cover the entire
area of BubbleZERO. The spatial sensor location variation led to dif-
ferent levels of measurement accuracy from each sensor with respect to
the average temperature in the space. To obtain the relation between
sensing accuracy and choice of number and locations of sensors, we
analyzed the data collected from the test-bed. During the experiment
sensor Telosb44 only stored a few data points due to faulty behavior;
consequently, we neglected that sensor in our analysis.
We consider the average of all the remaining 7 sensors as the actual
temperature of the room. We then select a different number of sensors
and calculate the difference between the average measure of the selected
k sensors (k = 1, 2, ..., 6) and the average measure of all 7 sensors
(i.e. the actual temperature of the room). This difference provides an
estimation of the temperature sensing accuracy under certain selection
of the sensors, and can be regarded as the measurement error of this
set of sensors. The root mean square (rms) of this difference is denoted
by ∆rms .
To further study the effect of sensor locations on sensing accuracy,
for each k value we enumerate all possible sets of sensors (with differ-
ent locations) and pick the set that provides the minimal rms value of
its measurement error, which is denoted by δm b . We also calculate the
measurement error of a randomly selected set (assuming equal prob-

ability for each sensor) denoted by δm r . The results of this analysis is
shown in Figure 6.5. For instance, when k = 1, the best sensor (i.e. the
one that provides the minimal rms value with respect to the average of
all 7 sensors) is sensor T45 , which is located in the south eastern part
of the bubble and provides ∆rms = 0.29. When k = 2, the best two
sensors are T41 and T48 , which are the two sensors located in the two
opposite corners, southwest and northeast of the bubble and leads to
∆rms = 0.18. In Figure 6.5, for the sake of simplicity, we approximate
the normal distribution for the measurement error and calculate the
corresponding mean and variance. The variance will then be used in
BEST sensor: T BEST set of 2 sensors: {T , T } BEST set of 3 sensors: {T , T , T }

45 41 48 41 45 48
RANDOM sensor: T47 RANDOM set of 2 sensors: {T47,T48} RANDOM set of 3 sensors: {T41, T42, T43}
Probability density function, Φ(ε)

5 δbm (µ=0.16, σ=0.24 RMS=0.29) δbm (µ=−0.058, σ=0.17 RMS=0.18) δbm (µ=0.015, σ=0.14 RMS=0.14)
8 8
δrm (µ=1, σ=0.68 RMS=1.2) δrm (µ=0.81, σ=0.66 RMS=1) r
δm (µ=−0.64, σ=0.55 RMS=0.84)
4
b b
N ~ (µ=0.16, σ=0.24) N ~ (µ=−0.058, σ=0.17) 6 Nb ~ (µ=0.015, σ=0.14)
6
3 Nr ~ (µ=1, σ=0.68) Nr ~ (µ=0.81, σ=0.66) Nr ~ (µ=−0.64, σ=0.55)
4 4
2
1 2 2
0 0 0
−3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3 −2 −1 0 1 2
Error, ε Error, ε Error, ε
BEST set of 4 sensors: {T42, T43, T46, T47} BEST set of 5 sensors: {T42, T43, T45, T46, T47} BEST set of 6 sensors: {T41, T42, T43, T46, T47, T48}
RANDOM set of 4 sensors: { T41, T42, T43, T47} RANDOM set of 5 sensors: {T41, T42, T43, T45, T46} RANDOM set of 6 sensors: {T41, T42, T43, T45, T46, T48}
30

12 b b b
δm (µ=−0.012, σ=0.11 RMS=0.11) 20 δm (µ=0.023, σ=0.069 RMS=0.073) δm (µ=−0.027, σ=0.04 RMS=0.048)
25
10 δrm (µ=−0.64, σ=0.55 RMS=0.84) δrm (µ=−0.32, σ=0.26 RMS=0.42) δrm (µ=−0.17, σ=0.11 RMS=0.2)
Nb ~ (µ=−0.012, σ=0.11) 15 Nb ~ (µ=0.023, σ=0.069) 20 Nb ~ (µ=−0.027, σ=0.04)

8
Nr ~ (µ=−0.64, σ=0.55) Nr ~ (µ=−0.32, σ=0.26) Nr ~ (µ=−0.17, σ=0.11)
15
6 10
4 10
5
2 5
0 0 0
−2 −1 0 1 2 −2 −1 0 1 2 −2 −1 0 1 2
Error, ε Error, ε Error, ε
Figure 6.5: Average error of k sensors for the minimal error set of sensors and a
random choose of sensors. Each figure lists in its title the best and random set of k
sensors for k=1,2,...,6, and plots the pdf of its measurement error (i.e. the difference
between the average of k sensor readings with the average of all 7 sensor readings),
b r
denoted by δm and δm , respectively. The best and random set of sensors are selected
based on their resulting ∆rms . N b and N r represent the normal distribution.
the selection of control algorithms.

The reason for assuming the normal distribution of measurement
errors is that a measurement is the result of a process. In this case, the
process includes what type of sensor to choose, manufacturing of the
sensor, where we install the sensor, acquiring data from the sensors,
and so on. Each step in the process may lead to a small error with a
probability distribution. We sum the error over all the steps to obtain
the final error, and due to the central limit theorem, that leads to the
normal distribution, no matter what the error on the individual steps
is.
6.4 Design Space Exploration
Based on the results presented above, we explore the design space for
both control algorithm and sensing system to build a Pareto front of
optimal energy and monetary cost under the constraint on user comfort
level.
Specifically, we first choose the number of sensors based on a given
6.4. Design Space Exploration 107
monetary budget. We assume the best sensor locations are known to

the designers (possibly through testing different location combinations)
and decide the sensing data accuracy according to the statistics in Fig-
ure 6.5, i.e. the measurement error δm b . Then by setting the measure-
ment error vk in Equation (6.5) to σm b , we conduct simulations for
all six controllers and choose the one that minimizes the energy cost
and satisfies the constraint on user comfort level as measured by the
discomfort index. Results are shown in Figure 6.6.
If it is required that the discomfort index be less than 0.12, the
Pareto front consists of different control algorithm choices depending
on the sensor monetary budget. In our platform, each Telosb unit costs
about $70. When the budget is set to be under $140, we may choose
1 or 2 sensors, in which case the control algorithm that can satisfy
the comfort level constraint and provide the minimal energy cost is
UKF-RMPC (with energy cost around 4500). When the budget is set
be more than 350, we may choose 5 or 6 sensors, in which case the
control algorithm that has the minimal energy cost and satisfies the
comfort constraint is EKF-MPC (with energy cost under 2500). When
the budget is around 280, we may choose 4 sensors, in which case the
best control algorithm is UKF-MPC.
Intuitively, when we have a small budget, the sensing data accuracy
is lower and we need a more robust algorithm to satisfy the comfort
level constraint; hence the RMPC controller (the MPC controllers do
not satisfy the constraint in this case). When we have a large budget,
the sensing accuracy is higher and we may choose the more energy-
efficient algorithms, hence the MPC controllers. If the discomfort index
is required to be less than 0.1, we will choose the UKF-RMPC controller
under any budget, since the other controllers either do not satisfy the
comfort constraint (i.e. EKF-MPC, UKF-MPC, on-off controllers) or
cost more energy (i.e. EKF-RMPC). If the discomfort index is set to be
0.14, we choose EKF-MPC if there is budget for more than 1 sensor.
If the sensor locations are selected randomly (i.e. theoretical best
locations might not be known or accessible in practice), the energy cost
and performance of each controller under certain budget are different
from the best location case. Figure 6.7 shows the exploration results
5000
4500
discomfort bound = 0.1
Energy Cost (kWh)
4000
3500 discomfort bound = 0.12 UKF-RMPC

UKF-MPC
3000 EKF-MPC
2500

2000
70 140 210 280 350 420
Sensor Monetary Budget ($)
Figure 6.6: Pareto front under comfort constraints with best sensor locations
8000
7000
Energy Cost (kWh)
6000
5000
discomfort bound = 0.2 UKF-RMPC
EKF-MPC
4000
3000
2000
70 140 210 280 350 420
Sensor Monetary Budget ($)
Figure 6.7: Pareto front under comfort constraints with random sensor locations
with random sensor locations. We can see that similar to the best loca-
tion case, the selection of control algorithms depends on the number of
sensors. For instance, the solid line represents the case where the dis-
comfort index is required to be less than 0.2. When the budget is under
280, we may select at most 4 sensors and the best control algorithm is
UKF-RMPC (other controllers do not satisfy the comfort constraints
except for EKF-RMPC but it has higher energy cost). When the bud-
get is more than 350, the best control algorithm is EKF-MPC. If the
discomfort index is required to be less than 0.1, the best control al-
gorithm is always UKF-RMPC. Note that the energy cost under any
number of sensors is also more than the energy cost in the best location
case with the same number of sensors, which is to be expected.
In this chapter, we demonstrated why the design of HVAC systems as

an integrated cyber-physical system is crucial for reducing energy con-
sumption in buildings. In the traditional top-down approach, the con-
trol algorithm and the embedded platform are in general designed sepa-
rately leading to sub-optimal systems. We proposed a co-design frame-
work for HVAC systems to explore control algorithm design and sensing
platform selection concurrently by analyzing their inter-dependencies.
We designed six control algorithms, namely on-off, MPC and RMPC
controllers with either EKF or UKF filtering. The estimation algo-
rithms are used to effectively address sensing and prediction errors.
We also analyzed the relation between sensing accuracy and the num-
ber and locations of sensors, using the collected data from a well-
instrumented test-bed. Based on these models and analysis, we explored
the design space of both control algorithm and sensing platform and
generated Pareto fronts with optimal energy and monetary cost.
7
Dynamic Contracts for Building-Grid Interaction
The smart grid ecosystem has been going through major upgrades in
the last two decades in three verticals:
1. New hardware such as solar panels, wind generation turbines, and

plug-in electric vehicles,
2. New sensing devices like smart meters and smart thermostats,
3. New communication and computation infrastructure such as the

broadband two-way communication and cloud computing.
Consequently, these breakthroughs in hardware and software have

paved the way for new operational schemes, such as automated demand
response and ancillary services from some unprecedented sources such
as buildings. In this new paradigm, buildings are beginning to play
new roles in the operation of the smart grid as entities for trading en-
ergy in real-time. As a result, the need for re-designing the smart grid
architecture and operation is more apparent than ever.
110
7.1. A Supply-Following Scenario for Smart Buildings 111
7.1 A Supply-Following Scenario for Smart Buildings
Consumers of energy usually do not pay attention to when they use

energy. Demand for electricity tends to rise specially at times when
it seems natural to use it; so natural in fact that we all tend to use
it at the same time and in similar ways. On days when demand for
electricity is high, extra power plants are needed to meet the demand.
An environmentally-friendly alternative to such a demand-following
strategy is offered by supply-following or demand-response programs,
where utilities provide incentives to encourage consumers to reduce
their demand during peak periods and use electricity at a less con-
gested time, as proposed by Maasoumy et al. [2014c]. Another solu-
tion would be to significantly increase the penetration of Renewable
Energy Sources (RESs), thus avoiding the introduction of expensive
power plants. Several states in the United States and countries around
the world have set ambitious targets for penetration of RESs by the
next few years. California, as an example, has targeted a 33% RES
portfolio by 2020 Helman [2010]. However, a large-scale power grid
requires continuous power balance between supply and demand; the
power flows through individual transmission lines and facilities should
also be controlled by continuously adjusting generation or load. Such
an instantaneous matching becomes challenging due to the volatility,
uncertainty, and intermittency of RESs, and makes their integration
into the grid extremely difficult. The situation is even worsened by the
uncertainties and randomness in the demand due to short-term ran-
dom switching of millions of individual loads or longer-term (e.g. daily
or seasonal) fluctuations in load and weather patterns. In the end, a
mix of such strategies will probably shape the future of the smart grid
ecosystem.
Such flexible loads with thermal storage capabilities, also denoted as
TCLs, are deemed to play an important role in regulating the grid fre-
quency and, consequently, in enabling deep penetration of RESs. It has
been reported that about 20% of the total electricity consumption in
the United States is used by residential TCLs such as air conditioners,
heat pumps, water heaters, and refrigerators bui, eia. Recently, Maa-
soumy et al. [2014c,d] showed that flexible loads such as TCLs are good
112 Dynamic Contracts for Building-Grid Interaction
candidates for providing ancillary services since their aggregate flexi-

bility can be controlled very fast, and sums up to tens of Gigawatts of
power, only in the United States.
TCLs are particularly well-suited for Direct Load Control and De-
mand Response programs that require loads to both decrease and in-
crease power consumption because they are capable of storing thermal
energy, much like a battery stores chemical energy. Despite several chal-
lenges in using loads for system services, key advantages include: (i)
reduction in the overall grid emissions Strbac [2008]; (ii) instantaneous
response of loads to operator requests, versus slow response of genera-
tors to significant output changes Kirby [2003]; and (iii) less variability
associated with a very large number of small loads with respect to that
of a small number of large generators Kirby [2003]. Therefore, while an-
cillary services have been conventionally supplied by generation units,
the increasing need for more energy storage capacity for frequency reg-
ulation, as well as more agile sources of ancillary services, makes it
attractive to also use energy reserves on the demand side. It may soon
be the case that the only technical impediment to reliable utilization
of loads for system services is the development of the necessary models
and control strategies and the development of inexpensive and scalable
communication and sensing infrastructure Woo et al. [2008].
In a supply-following “smart” grid scenario, buildings can exploit
remotely controllable thermostats and “smart” meters to communicate
with energy providers, trade energy in real-time and offer frequency reg-
ulation services, by leveraging the flexibility in the energy consumption
of their HVAC systems. The realization of such a scenario is, however,
strongly dependent on our ability to radically re-think the way both
the grid and the building control algorithms are designed.
Smart buildings are essential elements of the smart grid and can
play a significant role in its operation. A building has some inherent
flexibility in the way its HVAC system consumes electricity while re-
specting the comfort of its occupants. This flexibility could be used to
reduce costs if the electricity price is time-varying, or could be traded
(i.e., sold to the utility) if a proper contractual framework exists. Typi-
cally, the electricity consumption of a commercial building is controlled
7.1. A Supply-Following Scenario for Smart Buildings 113
by a process that aims at reducing costs. Among the many possible

ways to design such a process, MPC stands out. In order to use the po-
tential of commercial buildings as providers of flexibility to the smart
grid, we need to fundamentally redesign the way a building, and in
particular its HVAC, is controlled.
Commercial buildings are perceived to be ideal sources to provide
ancillary services due to the following reasons:
1. Commercial buildings account for more than 35% of electricity

consumption in the US.
2. More than 30% of commercial buildings have adopted BEMS

technology which facilitate the communication with the grid sys-
tem operators for providing flexibility. The majority of these
buildings are also equipped with variable frequency drives, which
in coordination with BEMS, can modulate the HVAC system
power consumption frequently (in the order of seconds).
3. Compared to a typical residential building, commercial buildings

typically have larger HVAC systems and therefore consume more
electricity.
According to Modera et al. [1999] and Westphalen and Koszalinski

[2001] about 10-15% of electricity consumption in commercial buildings
is related to the fans of HVAC systems. Fans are the main drivers
to move the conditioned air from the AHU to the rooms for climate
control. For instance, the main supply fans that feed Sutardja Dai
Hall on UC Berkeley campus can spin at variable speeds, with the
corresponding power consumption which is proportional to the cube of
fan speed, with the maximum rated power of 134 KW or about 14%
of the maximum power consumed in that building. Moreover, we can
directly control their power, upwards or downwards, making it an ideal
candidate for ancillary demand-response.
Historically, very few, high-capacity reserves, such as industrial
plants Todd et al. [2008], have been used to provide ancillary ser-
vices on the demand side. However, when a “swarm” of widespread
and smaller capacity reserves are available, these service providers are
better managed by intermediate entities called aggregators. The role

of an aggregator is to provide appropriate incentives for a swarm of
buildings at the right time, bundle the resulting capacity, and sell it in
the wholesale market for frequency regulation.
7.2 Ancillary Service from Buildings
Maasoumy et al. [2013a] via extended real-scale experiments observed

that modulating the fan speed of HVAC systems for extended peri-
ods of time with the existing control algorithms can lead to discomfort
and does not allow optimizing the amount of flexibility provided by
a building. Hence, Maasoumy et al. [2014c] proposed to re-design the
control algorithm, and considered commercial buildings whose HVAC
systems are controlled by an MPC scheme running an optimal control
problem at each time step k. Typically, as discussed in previous chap-
ters, the MPC aims at minimizing the total energy cost (in dollars). In
an SF scenario, such cost must account for the reward received from
the utility because of the building flexibility in energy consumption.
Maasoumy et al. [2014c] first identify and quantify “flexibility". Then,
a contractual framework between the utility and the building opera-
tor is designed so that the building can “declare" its flexibility and be
rewarded for it. Finally, a control algorithm was proposed that allows
the operation of building under this framework and at-scale experi-
ments are carried out to demonstrate the high potential of commercial
buildings as a source of flexibility and the feasibility of the developed
algorithm.
Clearly, by obeying the utility power consumption signals, the build-
ing may consume more or be in a worse state at the end of the H c time
slots with respect to a conventional demand-following protocol. The
flexibility declared by the building operator would then be significant
only if
• it is enough to be effectively exploited for frequency regulation

services Maasoumy et al. [2014d], and
• the reward from the utility is appropriate for the building.

7.2. Ancillary Service from Buildings 115
@ Building Side
Outside weather condition

Occupancy, Internal heat Min-Max
Model Predictive Control
Building state (temp.) Algorithm
State and input constraints
Real-time
Baseline power cons.

reward rate
energy rate
Up Flexibility
@ Grid Side
Down Flexibility
Flexibility
Per-unit
Ancillary Power Flow
Down envelope ≤ S ≤ Up envelope
Utility
Company
Baseline Power Flow Prediction for
Power Flow
next Hc time
steps
Generation
Figure 7.1: Schematic of the grid architecture and contractual framework proposed
in Maasoumy et al. [2014c].
The schematic of the entire system architecture is shown in Figure 7.1.

The solid-line arrows correspond to the baseline power flow. The an-
cillary power flow is represented by a dashed-line arrow. For the archi-
tecture in Figure 7.1 and the commercial contract summarized above,
we can state our energy management integrated control problem as
follows:
Maasoumy et al. [2015] propose a dynamic contractual framework
that in real-time analyzes the requirements of the grid on one side, and
requirements of the building on the other side, and performs optimal
operation of the whole system while taking into account the state and
input constraints of both building and grid and governs this interac-
tion through a set of interface variables and constraints. They use the
concept of assume-guarantee contracts to formalize the requirements
of the grid and the building subsystem as well as their interface. At
Figure 7.2: Flexibility trading market and wholesale energy market and the inter-
action of different components of these markets.
the building level, such formalization leads to the development of an

optimal control mechanism to determine the HVAC energy flexibility
while maximizing the monetary incentive for it. At the grid level, it al-
lows formulating a model predictive control scheme to optimally control
the ancillary service power flow from buildings, while integrating con-
straints such as ramping rates of ancillary service providers, maximum
available ancillary power, and load forecast information. To simplify
the problem, Maasoumy et al. [2015] abstract into one grid agent all
the players beyond the aggregator, such as the wholesale market play-
ers and the generation units, and denote as buildings the demand-side
service providers that deal with the aggregator. They then focus on
grid and buildings as the two sides of the supply-demand spectrum,
by abstracting all the intermediate entities involved in the chain from
power generation to power consumption. More recently, Maasoumy and
Sangiovanni-Vincentelli [2015] illustrated how the two building and grid
side controllers can interact seamlessly in a single framework with mul-
tiple time-scales of different sub-systems, as shown in Figure 7.2. They
define two markets, namely the flexibility trading market (FTM), and
the wholesale electricity market (WEM). Aggregator plays the role of
an interface between these two markets.

To address the challenges originating from this distributed and hier-
archical system, Maasoumy et al. [2015] resort to a Contract-Based De-
sign (CBD) methodology. CBD has recently emerged as a compositional
paradigm for the design of complex systems, emphasizing the concept
of interface and requirement formalization to facilitate system integra-
tion and provide formal support to the whole design flow Nuzzo and
Sangiovanni-Vincentelli [2014], Sangiovanni-Vincentelli et al. [2012].
They then provide an integrated design framework for MPC synthesis,
which can combine and subsume both the approaches in Maasoumy
et al. [2014c] and Maasoumy et al. [2014d].
The advantage of our contract-based methodology with respect to
previous works is threefold:
• it enables compositional design of the building and the grid MPC

schemes, so that they can be independently implemented while
still guaranteeing that their integration is correct;
• it allows extending the approaches in Maasoumy et al. [2014c]

and Maasoumy et al. [2014d] to highly distributed architectures,
including a large number of control areas and buildings, in a
scalable way;
• it supports automatic synthesis of embedded control software di-

rectly from assume-guarantee specifications.
There are a number of related papers in the literature. Balandat

et al. investigate the contract design problem that an energy aggre-
gator who participates in the wholesale market for Ancillary Services
faces. Specifically, they consider a situation in which commercial build-
ings agree to adjust their HVAC power consumption with respect to a
nominal schedule, according to a signal sent by the aggregator.
Vrettos et al. [2014] present a novel hierarchical control scheme to
enable provision of frequency control reserves by a pool of office build-
ings managed by an aggregator. The reserves are provided by control-
ling the consumption of the HVAC systems in a robust way. The ag-
gregator determines once a day the optimal amount of reserves and its
allocation among the participants. On the building level, a robust MPC

controller optimizes the HVAC system consumption every 15 minutes,
and a proportional controller provides the reserves in real-time.
Hao et al. [2013] show how fans in AHUs of commercial buildings
alone can provide substantial regulation service, with little change in
their indoor environments. A feed-forward architecture is proposed to
control the fan power consumption to track a regulation signal.
Aksanli et al. [2014] demonstrated a framework composed of dis-
tributed smart buildings and energy generation. In that paper Aksanli
et al. [2014] demonstrate the feasibility of such real-time communica-
tion and control among several grid nodes and building nodes. They
studied the effects of smart building controllers on the grid frequency
and stability via an experimental approach, using ad-hoc controller for
the grid node, while the controllers on the building sides ranged from
simple bang-bang controllers whose only objective was to keep the tem-
perature within a comfort zone, to more sophisticated min-max MPC
controllers that at the same time keep temperatures within the comfort
zone and maximize building capacity for ancillary service when needed
most.
Managing the aggregation of a large number of heterogeneous loads
to achieve a desired response, in the case large scale demand-response
programs still remains to be a challenge, especially at a fast time scale.
Jin et al. [2015] presents a system-level modeling and control design
approach for Demand Response (DR) management in smart grids, by
following a contract-based methodology. Given a set of flexible DR
loads, capable of adapting their power consumption upon external re-
quests, Jin et al. [2015] find a strategy for an aggregator to optimally
track a DR requirement by combining the individual contributions of
the DR components. In their framework, both the design requirements
and the DR components’ interfaces are specified by assume-guarantee
contracts expressed using mixed-integer linear constraints. They use
contracts to formulate an optimal control problem, which is solved re-
peatedly over time, in a receding horizon fashion. They illustrate the
effectiveness of our methodology for a set of DR components includ-
ing fans and pumps, showing that it enables modular development of
Ancillary power flow
Supply
Commercial Buildings
Demand
Baseline power flow
EV
Residential Buildings
Generation Transmission & Distribution Consumption
Figure 7.3: As influential players in the smart grid domain, buildings have a signif-
icant role in its operation. One of the services that buildings will provide to enhance
the operation of the smart grid is ancillary services to help frequency regulation
through automated demand response events.
non-disruptive DR control schemes.

To counteract the variability of RES, Hanif et al. [2015] proposed
provision of flexibility from the demand side as a viable option. Hanif
et al. [2015] discussed how the heating, ventilation and air conditioning
(HVAC) system, mostly installed in medium to large sized office build-
ings, can provide demand side flexibility. A model predictive control
(MPC) scheme in a receding horizon environment is deployed to pro-
vide an economic operation of the building, while respecting comfort
constraints of dwellers. Furthermore, robustness is introduced in the
MPC scheme to participate in both the energy and reserve market. In
doing so, the controller is also evaluated with respect to its sensitivity
towards economical and technical constraints. The National Electricity
Market of Singapore (NEMS) is used as a case study and the most
Figure 7.4: Fan power consumption versus volume flow rate. Data is for January
through August of 2013.
important parameters governing the challenges for integrating demand

side flexibility in the grid are pointed out.
7.3 Dynamic Contracts
7.3.1 The Baseline System
We consider a commercial building that has an HVAC system controlled

by an MPC as our baseline. The MPC has the objective of minimizing
the total energy cost (in dollar value). Let the time be slotted with τ as
the length of each time slot, and let H m be the prediction horizon (in
number of time-slots) of the MPC. System dynamics is also discretized
with a sampling time of τ . Typical values for τ and H m range from
15 min to 1 hr for τ and a few hours to a few days, e.g., 3 to 72 time
slots (with the assumption of 1 hr for each time slot) for H m . The
choice of H m depends on how far in the future the estimation of the
predicted values have an acceptable accuracy, and also on how far in
the future the required information (e.g. cost of energy) is available.
We use τ = 1 hr and H m = 24 hrs. At each time t, the controller solves
the following optimal control problem to compute the optimal vector
~ut = [ut , . . . , ut+H m −1 ], where ut+k is the air mass flow to the thermal
zones of the building.
7.3. Dynamic Contracts 121
m −1
HX
e
min Chvac (ut+k , πt+k , π ne,c , π ne,h , Tt+k
out
) (7.1)
~
ut
k=0
s.t. xt+k+1 = f (xt+k , ut+k , dt+k ), k = 0, ..., H m − 1
xt+k ∈ Xt+k , k = 1, ..., H m
ut+k ∈ Ut+k , k = 0, ..., H m − 1
The inputs to the optimal control problem are the states (i.e. zone
temperatures) xt (as initial condition), the set of electric energy prices
{πte , . . . , πt+H
e
m −1 }, the non-electric and non-fan energy prices, such as
gas price for heating π ne,h , and cooling π ne,c which are considered time-
invariant, a set of constraints Xt+k on the system states of the type:
“xt+k should be in Xt+k for all times t + k where k ∈ {1, . . . , H m })”, a
set of constraints Ut+k on system inputs for all k ∈ {0, . . . , H m −1}, and
an estimate on unmodelled disturbances dt+k (e.g., outside temperature
or building occupancy Maasoumy and Sangiovanni-Vincentelli [2012])
for all k ∈ {0, . . . , H m − 1} for the next H m time slots. T out is the
outside air temperature, f captures the system dynamics, and the total
HVAC power consumption cost Chvac (t) is the summation of fan power,
cooling power and heating power, given by:
Chvac (t) = π e (t)Pf (t) + π ne,c Pc (t) + π ne,h Ph (t) (7.2)

where with the assumption of no recirculation of air, and without loss
of generality, these power consumptions for time slot [t, t + 1] (with the
assumptions of constant air mass flow and price of energy over [t, t+1])
are calculated as follows:
Pf (ut ) = c1 .u3t + c2 .u2t + c3 .ut + c4 (7.3)

Ph (ut , Ttout ) = cp .ut .(T − Ttout )/COPh
s
(7.4)
Pc (ut , Ttout ) = cp .ut .(Ttout − T s )/COPc (7.5)
where c1 = −6.06×10−13 , c2 = 6.73×10−8 , c3 = −1.2×10−3 , and c4 =

59.2 are constants of the fan as shown in Figure 7.4, cp is the specific
heat of air, and COPh , and COPc are the coefficients of performance
for the heating system and the cooling system, respectively, and the
supply air temperature T s is considered constant. To move the coolant
fluid around, heating and cooling systems use pumps, which consume
electric power. However, we assume that electric power consumption of
pumps is negligible compared to the non-electric heating and cooling
powers of these systems. (see Maasoumy [2011], Maasoumy et al. [2011]
for more details).
Hence, the MPC solution to (7.1) is an optimal air mass flow tra-
jectory (also called an air mass flow profile) ~u∗t = [u∗t , . . . , u∗t+H m −1 ].
Only the first entry of ~u∗t is implemented at time t. At the next time
step, t + 1, the horizon of MPC is receded by one step, a new MPC is
set up, and solved to obtain ~u∗t+1 = [u∗t+1 , . . . , u∗t+H m ]. Again, the first
entry is implemented on the system, and the horizon is receded, and
this process repeats until the whole time frame of interest is covered.
At each time t, the optimal air mass flow vector for the next H time
steps is given by ~u∗t = [u∗t , . . . , u∗t+H m −1 ], with corresponding fan power
flow profile of Pf (~u∗ ) = [Pf (u∗t ), . . . , Pf (u∗t+H m −1 )]. In the following,
we will omit the reference to t in the profile and define a power profile
as a vector Pf (~u∗ ) = [Pf (u∗0 ), . . . , Pf (u∗H m −1 )].
7.3.2 The Baseline Contract
The baseline contract corresponding to this baseline system is very

simple. Namely, the building can consume what it wants at all times
and it pays πte per unit of consumed electricity at time t (i.e., if the time
span is of duration ∆, then it pays t+∆ e
P
k=t π (k)Pf (k) for electric power).
Hence, the only information being exchanged between the utility and
the BM is the real-time price vector [π0e , . . . , πH
e
m −1 ] in ($/kWh) sent
by the utility to the building (π ne,c and π ne,h are considered constant
and known). Then the building solves the MPC (7.1) to compute what
its electricity consumption profile/trajectory would be in the next H m
time slots. As described, this MPC would minimize its total cost in the
next H m time slots while respecting the building and the occupants
(comfort) constraints. Let ~u∗ denote the trajectory that is the solution
to this MPC and let C(~u∗ ) be its cost.
7.4. Flexibility-aware Contractual Framework 123
7.4 Flexibility-aware Contractual Framework
7.4.1 Definition of Flexibility

We define flexibility in terms of two envelope air mass flow profiles
e~l = [eltcs , . . . , eltce ] a lower envelope, and e~u = [eutcs , . . . , eutce ] an upper
envelope. tcs and tce are contract start and end times. Based on the
requirement from the utility that flexibility of each building has to be
determined ahead of time, we pick tcs , and consequently tcs as follows:
tcs ≥ 1 (7.6)
tcs + 1 ≤ tce (7.7)
tce − tcs = H c H m (7.8)
where H c is the length of the contract. Typical values for H c are much
smaller than H m , and can take values from one slot to a few time
slots. The BEMS computes e~l = [el0 , . . . , elH m −1 ] a lower envelope and
e~u = [eu0 , . . . , euH m −1 ] an upper envelope (using the algorithm described
later in the paper), so that any air mass flow profile ~u = [u0 , . . . , uH m −1 ]
such that for all k ∈ {0, . . . , H m − 1}, elk ≤ uk ≤ euk is feasible; i.e. no
building constraints are violated at anytime. The corresponding fan
power consumption envelopes are Pf (e~l ) = [Pf (el0 ), . . . , P~f (elH m −1 )]
and Pf (e~u ) = [Pf (eu0 ), . . . , Pf (euH m −1 )]. However, the building only
declares the first H c values of the envelope as flexibility, namely,
e~l = [eltcs , . . . , eltcs +H c ] a lower envelope, and e~u = [eutcs , . . . , eutcs +H c ]
an upper envelope, at the beginning of each contract. The reason for
declaring a subset of the obtained envelopes is that due to model mis-
match, imperfect predictions of disturbance and so on, the later values
in the H m -step envelopes may not be accurate and need to be updated
in the next time step. By declaring these two envelopes, the building
has essentially declared its flexibility for the next H c time slots. Note
that there is no objective function and energy cost here, we define
flexibility with respect to feasibility criteria.
By declaring these two envelopes, the building manager is telling
the utility: “I allow you to select any power trajectory Pf (~u) =
[Pf (utcs ), . . . , Pf (utce )] such that for all k ∈ {tcs , . . . , tce }, Pf (elk ) ≤
Pf (uk ) ≤ Pf (euk )”.
To quantify the flexibility provided at each time step k, we need to

define a single metric. The metric that is natural to use is the difference
between the upper and lower power envelope that can be consumed by
the building without violating any constraints. Hence,
Flexibility(k) , Pf (euk ) − Pf (elk ) (7.9)
7.4.2 Contract Framework

In the proposed framework, the building operator can declare its flexi-
bility contract to the utility. The flexibility declared by the building op-
erator would only be significant if the reward is appropriate. It is impor-
tant to understand that by allowing the utility to select any trajectory
Pf (~u) = [Pf (utcs ), . . . , Pf (utce )] or correspondingly ~u = [utcs , . . . , utce ]
such that for all k ∈ {tcs , . . . , tce } elk ≤ uk ≤ euk , the building might
consume more or be in a worse state at the end of the H c time slots.
We define the “flexibility" contract as follows:
The utility charges the building operator for its baseline power con-
sumption Pf (~u∗ ) which is agreed upon at the time of the contract, irre-
spective of the deviations from the baseline power consumption due to
flexibility signals from the utility, and the utility rewards the building
operator for its declared flexibility.
Hence, this contract is deterministic in the sense that from the
beginning of the contract, the utility and the building operator both
know how much money each has to pay; the building pays for consum-
ing Pf (u∗t ) at rate πte , and the utility pays the building for providing
downward flexibility at rate β t and for providing upward flexibility at
rate β t .
In short, the contract implementation steps are as follows:
1. The building and the utility agree upon a contract length, H c .
2. The utility declares [π0e , . . . , πH e

m −1 ], as per unit price of electric
energy, [β 0 , . . . , β H m −1 ] in ($/kW ), as reward for down flexibility,

and [β 0 , . . . , β H m −1 ] in ($/kW ), as reward for upward flexibility.
If the utility is not willing to commit to the flexibility rates for
the time span beyond the next immediate contract period, i.e.
[β H c +1 , . . . , β H m −1 ], and [β H c +1 , . . . , β H m −1 ], the building oper-

ator can then obtain an estimate of these values from historical
data.
3. The building operator computes the baseline air mass flow u∗k and
the two envelopes elk and euk , for the time frame k = 0, 1, . . . , H m −
1, with its overall cost if it uses the flexibility contract as follows:
Building Operator Payment
z }| {
m −1
HX
Cf = Chvac (uk , πke ) (7.10)
k=0
m −1
HX m −1
HX
− β k ψ(uk , elk ) − β k ψ(uk , euk )
k=1 k=1
| {z }
Reward for Providing Flexibility
where ψ, and ψ are defined as:
ψ(uk , euk ) , Pf (euk ) − Pf (uk ) (7.11a)

ψ(uk , elk ) , Pf (uk ) − Pf (elk ) (7.11b)
4. The building operator then declares Pf (e~l ) and Pf (e~u ), the base-
line profile Pf (~u∗ ) that it will consume and the length of the
contract H c to the utility.
5. In the next H c time slots, the utility will send signals (sk )’s, such
that Pf (elk ) ≤ sk ≤ Pf (euk ) and the building operator has to obey
the signals, i.e., has to consume power in time slot k equal to sk .
Flexibility signal sk may arrive as frequently as every few seconds,
as mentioned earlier.
7.4.3 The Control Mechanism

As discussed in Section 7.4.2, if the utility provides a vector ~π e =
[π0e , . . . , πH
e
m −1 ] of per-unit of energy prices per time slot, then BEMS
P m −1
could try to minimize its total energy cost H k=0 Chvac (uk , πk , Tkout ).
In that case, BEMS uses building flexibility selfishly to minimize its
cost. Hence, BEMS does value (and does use) its flexibility. If the utility
values the flexibility that the building HVAC system can offer, it has
to provide the right incentive and the right mechanism to declare this
flexibility.
Formalizing Flexibility and Incentives in the MPC Framework
We say that the building can offer a flexibility Ψ := {ψ, ~ in fan power
~ ψ}
~ } in air mass flow which comprises
or equivalently a flexibility Φ := {~ϕ, ϕ
down flexibility, ~ϕ, and upward flexibility ϕ ~ , from the contract start
time tcs = t + 1 to the contract end time tce := t + 1 + H c for 1 ≤
H c ≤ H m time slots (starting from x0 ) if there exist two trajectories
e~l = ~u + ~ ~ , that satisfy:
ϕ and e~u = ~u + ϕ
ϕk ≤ 0, ϕk ≥ 0 ∀k ∈ {tcs , . . . , tce } (7.12)

f (xk , uk + ϕk , dk ) ∈ Xk+1 ∀k + 1 ∈ {tcs , . . . , tce } (7.13)
f (xk , uk + ϕk , dk ) ∈ Xk+1 ∀k + 1 ∈ {tcs , . . . , tce } (7.14)
uk + ϕk ∈ Uk ∀k ∈ {tcs , . . . , tce } (7.15)
uk + ϕk ∈ Uk ∀k ∈ {tcs , . . . , tce } (7.16)
What we defined here is a flexibility over multiple time slots. As-

sume that BEMS declares ~ϕ and ϕ ~ (i.e., the utility might choose any
fan power (and consequently air flow ûk ) in time step tcs ≤ k ≤ tce as
long as u∗k + ϕk ≤ ûk ≤ u∗k + ϕk ), where u∗k is the baseline air mass flow.
Hence, we “center" the flexibility around u∗ .
The flexibility-aware optimal control problem should minimize the
cost function which is composed of the cost for the baseline HVAC
power consumption, Chvac , (i.e., the cost of problem (7.1)), minus the
reward for the flexibility, R, which is computed as follows:
T T
~ = ~β .ψ(~
~ B)
R(Φ, ~ u, ϕ
~ ) + ~β .ψ(~
~ u, ~ϕ) (7.17)
where B ~ := {~β, ~β}, and ψ(.), and ψ(.) are given by (7.11), in which
e~l = ~u + ~ ~.
ϕ and e~u = ~u + ϕ
Putting It All Together
We assume at each time step t that the current state of the building is
known. Furthermore, the prediction of the outside temperature, inside
heat generation, and constraints on the system states and inputs are
known. The outputs of the algorithm is the nominal power consumption
of the building, u∗t+k for k ∈ {0, ..., H m − 1}, the flexibility that the
building can provide, Φ∗t+k+1 for k ∈ {0, ..., H m − 1}, for future time
steps, without violating constraints. Based on this set of information,
the BEMS can decide how much flexibility to offer. It may declare very
small flexibility and hence get a cost close to Chvac .
The algorithm described above can be formulated as a min-max
MPC problem. At time t we are interested in solving the following
robust optimal control problem:
Min-max Problem
m
HX −1
min max Chvac (ut+k , πt+k ) − R(Φt+k+1 , Bt+k+1 ) (7.18a)
~ ~ t+1 w
ut ,Φ ~t
k=0
subject to: xt+k+1 = f (xt+k , ut+k + wt+k , dk )
∀ k = 0, ..., H m − 1 (7.18b)
∀ wt s.t. : ϕt ≤ wt ≤ ϕt (7.18c)
∀ wt+k s.t. : ϕt+k ≤ wt+k ≤ ϕt+k
∀k = 1, ..., H m − 1 (7.18d)
m
ϕt+k ≥ 0, ∀ k = 1, ...H −1 (7.18e)
m
ϕt+k ≤ 0, ∀ k = 1, ...H −1 (7.18f)
m
xt+k ∈ Xt+k ∀ k = 1, ..., H (7.18g)
m
ut+k + wt+k ∈ Ut+k ∀ k = 0, ..., H −1 (7.18h)
Note that ϕt and ϕt are computed in the previous time step and
are constant values in this formulation, while ϕt+k and ϕt+k for k ∈
{1, . . . , H m − 1} are optimization variables and will be computed in the
current time step by solving the optimal control problem (7.18).
The inner maximization problem make the optimization problem
robust and derives the worst-case scenario cost and constraints. The
outer minimization problem solves for its arguments (~ut , Φ ~ t+1 ) while
it is guaranteed that the constraints are satisfied for all values of un-
certainty w, as long as it is within the range ϕt+k ≤ wt+k ≤ ϕt+k for
k ∈ {1, . . . , H m − 1}.
Theorem Bertsekas [1999]:

Let C be a closed convex set and let f : C → R be a convex function.
Then if f attains a maximum over C, it attains a maximum at some
extreme point of C.
According to the theorem above, the analytic worst-case solution

can be obtained. Closed convex sets can be characterized as the inter-
sections of closed half-spaces (i.e. sets of points in space that lie on and
to one side of a hyperplane). The feasible set for states and inputs, that
are the temperature of the rooms in the building and air mass flow into
the thermal zones, respectively, are defined by:
Xk := {x | T k ≤ x ≤ T k } (7.19)
Uk := {u | U k ≤ u ≤ U k } (7.20)
where T k , and T k are the upper and lower temperature limits and
U k , and U k are the upper and lower feasible air mass flow at time t.
Therefore, the feasible set of (7.18) is closed and convex. The objective
function is also convex on w, as the max is over variable w. In fact, w
does not appear in the cost function. The objective function is a linear
function on w and hence it is concave. We also consider a linearized
state update equation as follows:
xt+1 = Axt + But + Edt (7.21)
the nonlinear system dynamics has been linearized with the forward
Euler integration formula with time-step τ = 1 hr. Hence, the min-
max problem (7.18) is equivalent to:
Minimization Problem
m
HX −1
min Chvac (ut+k , πt+k ) − R(Φt+k+1 , Bt+k+1 ) (7.22a)
~ ~ t+1
ut ,Φ
k=0
s. t.: xt+k+1 = f (xt+k , ut+k + ϕt+k , dt+k ) (7.22b)
∀k = 0, ..., H m − 1
xt+k+1 = f (xt+k , ut+k + ϕt+k , dt+k ) (7.22c)
∀k = 0, ..., H m − 1
ϕt+k ≥ 0, ∀ k = 1, ...H m − 1 (7.22d)
m
ϕt+k ≤ 0, ∀ k = 1, ...H −1 (7.22e)
m
xt+k ∈ Xt+k ∀ k = 1, ..., H (7.22f)
m
xt+k ∈ Xt+k ∀ k = 1, ..., H (7.22g)
m
ut+k + ϕt+k ∈ Ut+k ∀ k = 0, ..., H −1 (7.22h)
m
ut+k + ϕt+k ∈ Ut+k ∀ k = 0, ..., H −1 (7.22i)
The argmin of the optimization problem (7.22) is the nominal power

consumption, u∗t+k , and the maximum available flexibility, Φ∗t+k , ∀ k =
0, ..., H m −1. The building declares u∗t+k , and Φ∗t+1+k for the time slots:
∀ k = 0, ..., H c −1 to the utility. After H c time slots, the BEMS collects
the updated parameters such as new measurements and disturbance
predictions, and sets up the new MPC algorithm for the time step
k = H c , H c +1, . . . , H c +H m −1, and solves the new MPC for this time
frame and uses only the first H c values of baseline power consumption
and flexibility, i.e. for k = H c , . . . , 2H c − 1, and this process repeats.
Figure. 7.1 shows the schematic of the entire system architecture.
The solid line power flow arrows correspond to the baseline system
and contract. The ancillary power flow via the flexibility contract is
shown with dashed line arrow. Real-time state of the system such as
occupancy, internal heat, outside weather condition, building temper-
ature, as well as state and input constraints are passed as input to
the algorithm. The utility also communicates information such as per-
unit energy and upward and downward flexibility prices to the BEMS
(or effectively to the algorithm). The output of the algorithm includes
baseline power consumption, downward flexibility and upward flexi-

bility values, and cost. Within the flexibility contract framework, this
information is communicated to the utility. Consequently, a flexibil-
ity signal st is sent from the utility to the building to be tracked by
the HVAC fan. Essentially, the utility has control of the building con-
sumption for the next Hc time slots. The control strategy is actuated
by sending flexibility signals (similar to frequency regulation signals).
These signals can be sent as frequently as every few seconds. In fact,
we show through experiments on a real building in Maasoumy et al.
[2014c] that buildings with HVAC systems equipped with variable fre-
quency drive (VFD) fans are capable of tracking such signals within a
few seconds.
7.5 Computational Results
The algorithm presented in Section 7.4.3 was implemented on a building

model developed and validated against historical data in Maasoumy
et al. [2011], Maasoumy and Sangiovanni-Vincentelli [2012]. The non-
linear optimization problem solver Ipopt Wächter and Biegler [2006],
was used to solve the nonlinear optimization problem. We use sampling
time of τ = 1 hr. We assess the performance of the algorithm for the
following scenarios.
Different reward rates have been considered for upward and down-
ward flexibility at each time step. In particular, downward flexibility is
rewarded more than upward flexibility (β > β) for most of the time.
Since MPC maintains the comfort level using the least possible amount
of energy, buildings that are operated under nominal MPC such as (7.1)
have no down flexibility (for extended period of time, i.e. 1 hr or more).
However, we show that via proper incentives and by solving (7.22), it
is possible to provide downward flexibility when most needed. The re-
sults of this scenario are shown in Figure 7.5 for the case when ancillary
signals are received from the utility every minute: no system constraint
(e.g. temperature being within the comfort zone) will be violated when
the fan speed enforcement is performed for arbitrary values of fan speed
as long as fan power consumption, and consequently fan speed is within
7.5. Computational Results 131
Figure 7.5: The Per-unit energy rate, and upward and downward flexibility reward
are shown in the lowest figure. The middle figure shows the resulting flexibility at
each time, and the top figure shows the resulting room temperature. Flexibility
signals are sent every minute from the utility to the building.
the safe envelope calculated by the algorithm. It is shown in Figure 7.5

that maximum flexibility (100%) is provided at times when the room
temperature is far from the boundaries of the comfort zone. The flexi-
bility decreases as the temperature of the room approaches the comfort
zone boundary, and is minimum (about 0-15% in this case) when room
temperature is fairly close to the boundaries of the comfort zone, and
the reward for flexibility is not high enough.
After reviewing various approaches proposed in the literature to address

the problem of ancillary services from buildings to the grid, we proposed
a methodology to define and quantify the flexibility of a commercial
building HVAC systems. We also proposed a contractual framework
that could be used by the building operator and the utility to declare
the flexibility on one side, and the reward structure on the other side.
We then designed a control mechanism for the building to decide how
to declare its flexibility for the next contractual period to maximize its
reward given the contractual framework. Due to the heterogeneous na-
ture of DR resources, one of the promising approaches to this problem
is the assume-guarantee contracts, expressed using Discrete Hybrid Au-
tomata (DHA) and mixed-integer linear constraints, as proposed by Jin
et al. [2015], to formalize the interface of each DR component and spec-
ify their aggregate behaviors.
Incorporation of statistical models for the aggregate behaviors of a
large number of loads, and the deployment of robust control strategies
into existing control frameworks is one of the areas for future work in
this domain. Further investigation of the scalability of the proposed
approaches in this chapter is also another area of future work.
8
Conclusion
Buildings are significant consumers of energy. Energy efficiency in

buildings is an urgent problem. Energy consumption must decrease by
50% in all retrofits and 90% in all new buildings by 2030. Contributors
to inefficiency in buildings are:
• Gaps in design processes;
• Lack of sufficient modeling and design tools;
• Gaps in operations;
• Sub-optimal system-level controls, lack of continuous diagnostics,

and lack of robustness.
Unfortunately, over the last few decades, neither has been a focus of
Research and Development. DOE has invested in incremental improve-
ments of existing tools, methods and process, but barriers in policy,
economics and behavior has slowed down the pace of innovation in this
field.
In this article we provided an in-depth overview of research activ-
ities in the field of smart buildings modeling, design automation and
133
134 Conclusion
control. In particular, we presented simulation tools based on physics-

based mathematical model of thermal behavior of buildings, along with
a novel Parameter-Adaptive Building (PAB) model framework. The
proposed PAB Model would update the model parameters as new mea-
surements arrive to reduce the model uncertainties. Using the obtained
building model, we then presented Model Predictive Control (MPC)
to minimize the energy consumption of the building while satisfying
thermal comfort, and constraints on other system states and. As build-
ing models always inherit uncertainty, we proposed a Robust Model
Predictive Control (RMPC) to obtain a control strategy that is ro-
bust against model uncertainties. We then proposed a methodology
for selecting a controller type, i.e. RMPC or MPC, versus Rule-Based
Control (RBC) as a function of model uncertainty. We also presented
Stochastic Model Predictive Control (SMPC) and Exergy-based Model
Predictive Control (XMPC) techniques.
We then addressed the “Cyber-Physical" aspect of a building HVAC
system in the design flow. We presented a co-design framework that an-
alyzes the interaction between the control algorithm and the embedded
platform through a set of interface variables, and demonstrated how the
design space is explored to optimize the energy cost and monetary cost,
while satisfying the constraints for occupant comfort level.
The last chapter of this article centered on the role of smart build-
ings in the context of the smart grid. Commercial buildings have in-
herent flexibility in how their HVAC systems consume electricity. We
presented a dynamic contractual framework that could be used by the
building operator and the utility company to declare flexibility on one
side and reward structure on the other side. We also presented a con-
trol mechanism for the building to decide its flexibility for the next
contractual period to maximize the reward. Finally, we demonstrated
that the demand-side flexibility of the HVAC system of a typical com-
mercial building can be exploited for providing frequency regulation
service to the power grid using at-scale experiments. We then showed
how this flexibility in power consumption of building HVAC system can
be leveraged for providing regulation service.
8.1. Future Work 135
8.1 Future Work
• We proposed a methodology and framework to adapt the param-

eters of a building model in an online fashion leading to the PAB
Model. As future work, the PAB model can be utilized in an
adaptive control framework to yield better control performance.
• The inter-dependencies between the HVAC control algorithm and

the embedded platform is amenable for further research. In par-
ticular, one should analyze the relation between the prediction
error and the design of the embedded platform (e.g. the choice of
the CO2 and outdoor sensors), and leverage the findings in a co-
design framework. Our consideration of the embedded platform
design can be also broadened from the sensing system to the com-
putation and communication components, such as the impact of
communication reliability on the control algorithm.
• We discussed building flexibility, and assumed that the building

operator communicates directly with the utility. However, a more
realistic scenario should be considered when an electricity bro-
ker, an aggregator, seeks rate offers from suppliers for “bundled”
groups of customers and acts on their behalf.
• We discussed ancillary power control based on the assumption

that the load in the power gird is predictable. However, it is
very difficult to accurately forecast power consumption. As future
work, a robust MPC framework could be developed to address the
uncertainties associated with imperfect predictions of load.
• In Chapter 7 we presented the case where buildings play an active

role in the operation of the smart grid. In a smaller scope, mul-
tiple buildings which share a common resource such as a chiller
plant could also benefit from the design and control strategies pre-
sented in this article. With the rise of Internet of Things (IoT)
and machine-to-machine (M2M) communication over the cloud,
it is now more apparent than ever that in the future, building-
to-building (B2B) communications would be crucial in optimally
136 Conclusion
operating clusters of commercial buildings such as on a univer-

sity campus, not in a standalone fashion, but in a networked and
connected fashion. As future work, it is important to show how
do the methods presented in this article scale to multiple “con-
nected” buildings to optimize the objective function of the whole
cluster in an orchestrated fashion, as opposed to each building op-
timizing its operation in a silo leading to a sub-optimal operation
of a group of buildings.
Acknowledgements
The authors are grateful for the many discussions and insights from
Francesco Borrelli (UCB), Duncan Callaway (UCB), Clas Jacobson
(UTC), Manfred Morari (ETH), Pierluigi Nuzzo (UCB), Kameshwar
Poolla (UCB), Catherine Rosenberg (Waterloo), Mahdi Shahbakhti
(MTU), Costas Spanos (UCB), Michael Wetter (LBNL) and Qi Zhu
(UCR). The research has been partially sponsored by the SinBerBEST
program and by TerraSwarm, one of six centers of STARnet, a Semi-
conductor Research Corporation program sponsored by MARCO and
DARPA.
137
References
Annual Energy Outlook 2015. http://www.eia.gov/totalenergy/.

Future Cities Lab. http://www.futurecities.ethz.ch/.
Buildings energy data book. http://buildingsdatabook.eren.doe.gov/default.aspx.
U.S. Energy Information Administration, annual energy review 2010.
http://www.eia.gov/totalenergy/data/annual/consumption.
U.S. Environmental Protection Agency. Buildings and their
Impact on the Environment: A Statistical Summary.
http://www.epa.gov/greenbuilding/pubs/gbstats.pdf. EPA, Washing-
ton, DC., 2009.
IBM ILOG CPLEX Optimizer, www.ibm.com. 2012.
Building Energy Data Book of DOE. http://buildingsdatabook.eren.doe.gov/
default.aspx, February 2013.
HAMBase http://archbps1.campus.tue.nl/bpswiki/ index.php/Hamlab, April
2015.
OptiControl http://www.opticontrol.ethz.ch/index.html, April 2015.
Radiance http://www.radiance-online.org/, April 2015.
National Renewable Energy Laboratory (Jan 2013). http://www.nrel.gov/.,
May 2015.
American Housing Survey for the United States. U.S. Department of Housing
and Urban Development and U.S. Department of Commerce., September
2008.
138
References 139
C. Agbi et al. “Parameter Identifiability for Multi-Zone Building Models”.

In IEEE, 51st Annual Conference on Decision and Control (CDC), pp.
6951-6956, Grand Wailea Maui, HI, 2012.
S. Ahuja, A. Surana, and E. Cliff. Reduced-order models for control of strati-
fied flows in buildings. In American Control Conference (ACC), 2011, pages
2083–2088. IEEE, 2011.
Baris Aksanli, Alper S Akyurek, Madhur Behl, Meghan Clark, Alexandre
Donzé, Prabal Dutta, Patrick Lazik, Mehdi Maasoumy, Rahul Mangharam,
Truong X Nghiem, et al. Distributed control of a swarm of buildings con-
nected to a smart grid: demo abstract. In Proceedings of the 1st ACM Con-
ference on Embedded Systems for Energy-Efficient Buildings, pages 172–
173. ACM, 2014.
B. Arguello-Serrano and M. Velez-Reyes. Nonlinear control of a heating, ven-
tilating, and air conditioning system with thermal load estimation. Control
Systems Technology, IEEE Transactions on, 7(1):56–63, 2002. ISSN 1063-
6536.
ASHRAE. “Standard 62.1-2004”. Ventilation for acceptable indoor air quality,
2004.
A. Aswani, N. Master, J. Taneja, D. Culler, and C. Tomlin. Reducing tran-
sient and steady state electricity consumption in hvac using learning-based
model-predictive control. Proceedings of the IEEE, (99):1–14.
Maximilian Balandat, Frauke Oldewurtel, Mo Chen, and Claire Tomlin. Con-
tract design for frequency regulation by aggregations of commercial build-
ings.
Madhur Behl, Truong X Nghiem, and Rahul Mangharam. Model-iq: Uncer-
tainty propagation from sensing to modeling and control in buildings. In IC-
CPS’14: ACM/IEEE 5th International Conference on Cyber-Physical Sys-
tems (with CPS Week 2014), pages 13–24. IEEE Computer Society, 2014.
S Bengea, Veronica Adetola, Keunmo Kang, Michael J Liba, Draguna Vrabie,
R Bitmead, and Satish Narayanan. Parameter estimation of a building
system model and impact of estimation error on closed-loop performance.
In Decision and Control and European Control Conference (CDC-ECC),
2011 50th IEEE Conference on, pages 5137–5143. IEEE, 2011.
Sorin C Bengea, Anthony D Kelman, Francesco Borrelli, Russell Taylor, and
Satish Narayanan. Implementation of model predictive control for an hvac
system in a mid-size commercial building. HVAC&R Research, 20(1):121–
135, 2014.
140 References
Theodore L Bergman, Frank P Incropera, and Adrienne S Lavine. Funda-

mentals of heat and mass transfer. John Wiley & Sons Incorporated, 2011.
Willy Bernal, Madhur Behl, Truong X Nghiem, and Rahul Mangharam. Mle+:
a tool for integrated design and deployment of energy efficient building
controls. In Proceedings of the Fourth ACM Workshop on Embedded Sensing
Systems for Energy-Efficiency in Buildings, pages 123–130. ACM, 2012.
Dimitri P Bertsekas. Nonlinear programming. 1999.
Y. Bi, X. Wang, Y. Liu, H. Zhang, and L. Chen. “Comprehensive exergy
analysis of a ground-source heat pump system for both building heating
and cooling modes”. Applied Energy, 86(12):2560–2565, 2009.
Vaughn Bradshaw. The building environment: Active and passive control sys-
tems. 2010.
James E Braun. Load control using building thermal mass. Journal of solar
energy engineering, 125(3):292–301, 2003.
Bruno Bueno, Leslie Norford, Grégoire Pigeon, and Rex Britter. A resistance-
capacitance network model for the analysis of the interactions between
the energy performance of buildings and the urban climate. Building and
Environment, 54:116–125, 2012.
M Castilla, JD Álvarez, M Berenguel, F Rodríguez, JL Guzmán, and M Pérez.
A comparison of thermal comfort predictive control strategies. Energy and
Buildings, 43(10):2737–2746, 2011.
M Castilla, JD Álvarez, JE Normey-Rico, and F Rodríguez. Thermal comfort
control using a non-linear mpc strategy: A real case of study in a bioclimatic
building. Journal of Process Control, 24(6):703–713, 2014.
CBECS. Commercial Buildings Energy Consumption survey (CBECS).
http://www.eia.gov/consumption/commercial/.
Y.A. Cengel, M.A. Boles, and M. Kanoğlu. “Thermodynamics: an Engineering
Approach”. 8th Edition, Chapter 8, McGraw-Hill, New York, 2015.
Spyros Chatzivasileiadis, Marco Bonvini, Javier Matanza, Rongxin Yin, Zhen-
hua Liu, Thierry Nouidui, Emre C Kara, Rajiv Parmar, David Lorenzetti,
Michael Wetter, et al. Cyber physical modeling of distributed resources for
distribution system operations. arXiv preprint arXiv:1505.00078, 2015.
R. Chengqin, L. Nianping, and T. Guangfa. “Principles of exergy analysis
in HVAC and evaluation of evaporative cooling schemes”. J. Building and
Environment, 37(11):1045–1055, 2002.
References 141
KJ Chua, SK Chou, WM Yang, and Jinyue Yan. Achieving better energy-

efficient air conditioning–a review of technologies and strategies. Applied
Energy, 104:87–104, 2013.
JA Clarke, J Cockroft, S Conner, JW Hand, NJ Kelly, R Moore, T O Brien,
and P Strachan. Simulation-assisted control in building energy management
systems. Energy and buildings, 34(9):933–940, 2002.
Drury B Crawley, Jon W Hand, Michaël Kummert, and Brent T Griffith.
Contrasting the capabilities of building energy performance simulation pro-
grams. Building and environment, 43(4):661–673, 2008.
K Dalamagkidis, D Kolokotsa, K Kalaitzakis, and GS Stavrakakis. Reinforce-
ment learning for energy conservation and comfort in buildings. Building
and environment, 42(7):2686–2698, 2007.
Stephen Dawson-Haggerty, Xiaofan Jiang, Gilman Tolle, Jorge Ortiz, and
David Culler. smap: a simple measurement and actuation profile for phys-
ical information. In Proceedings of the 8th ACM Conference on Embedded
Networked Sensor Systems, SenSys’10, New York, USA, 2010. ACM. ISBN
978-1-4503-0344-6.
Richard de Dear and Gail Schiller Brager. The adaptive model of thermal
comfort and energy conservation in the built environment. International
Journal of Biometeorology, Springer, vol. 45, no. 2, pp. 100-108, 2001.
Richard J de Dear and Gail S Brager. Thermal comfort in naturally ventilated
buildings: revisions to ASHRAE Standard 55. Energy and Buildings, vol.
34, no. 6, pp. 549 - 561, 2002.
Kun Deng, Prabir Barooah, Prashant G Mehta, and Sean P Meyn. Building
Thermal Model Reduction via Aggregation of States. In ACC, pp. 5118-
5123, Baltimore, MD, 2010.
Ibrahim Dincer and Marc A Rosen. Exergy: energy, environment and sustain-
able development. Newnes, 2012.
DOE-2. DOE-2: http://www.doe2.com/.
AI Dounis and DE Manolakis. Design of a fuzzy system for living space
thermal-comfort regulation. Applied Energy, 69(2):119–144, 2001.
Dymola. http://www.dymola.com .
Bryan Eisenhower, Zheng O’Neill, Vladimir A Fonoberov, and Igor Mezić. Un-
certainty and sensitivity decomposition of building energy models. Journal
of Building Performance Simulation, 5(3):171–184, 2012.
EnergyPlus. EnergyPlus http://apps1.eere.energy.gov/buildings/energyplus.
142 References
EnergySTAR. American Housing Survey for the United States.

http://www.energystar.gov/buildings/about-us/how-can-we-help-you/build-
energy-program/business-case.
eQUEST. The QUick Energy Simulation Tool.
http://www.doe2.com/equest/.
ESPr. ESP-r. http://www.esru.strath.ac.uk/Programs/ESP-r.htm.
Mohammad Ali Fayazbakhsh, Farshid Bagheri, and Majid Bahrami. A
resistance-capacitance (rc) model for real-time calculation of cooling load in
hvac-r systems. Journal of Thermal Science and Engineering Applications,
2015.
Clifford C Federspiel. Estimating the inputs of gas transport processes in
buildings. IEEE Transactions on Control Systems Technology, vol. 5, no.
5, pp. 480-489, 1997.
Clifford C Federspiel and Haruhiko Asada. User-adaptable comfort control
for hvac systems. Journal of dynamic systems, measurement, and control,
116(3):474–486, 1994.
PM Ferreira, AE Ruano, S Silva, and EZE Conceicao. Neural networks based
predictive control for thermal comfort and energy savings in public build-
ings. Energy and Buildings, 55:238–251, 2012.
Roberto Z Freire, Gustavo HC Oliveira, and Nathan Mendes. Predictive con-
trollers for thermal comfort optimization and energy savings. Energy and
buildings, 40(7):1353–1365, 2008.
Mark A Goforth, George W Gilchrist, and Joseph D Sirianni. Cloud Effect
on Thermal Downwelling Sky Radiance. Proc. SPIE, Society of Photo-
Optical Instrumentation Engineers, vol. 4710, no. 1, pp. 203-213, Orlando,
FL, 2002.
Markus Gwerder, Stefan Boetschi, Dimitrios Gyalistras, Carina Sagerschnig,
David Sturzenegger, Roy Smith, and Bruno Illi. Integrated predictive rule-
based control of a swiss office building. In 11th REHVA World Congress
Clima, 2013.
D. Gyalistras and M. Gwerder. Use of Weather and Occupancy Forecasts
for Optimal Building Climate Control (OptiControl): Two Years Progress
Report. In ETH Zurich, Zug, Switzerland, 2010.
ASHRAE Fundamentals Handbook. American Society of Heating, Refriger-
ating and Air-conditioning Engineers. Inc., NE Atlanta, GA, 30329, 2009.
References 143
Sarmad Hanif, Dante Fernando Recalde Melo, Mehdi Maasoumy, Tobias

Massier, Thomas Hamacher, and Thomas Reindl. Model predictive control
scheme for investigating demand side flexibility in singapore. In 50th In-
ternational Universities Power Engineering Conference, Staffordshire Uni-
versity, 2015.
He Hao, Anupama Kowli, Yashen Lin, Prabir Barooah, and Sean Meyn. An-
cillary service for the grid via control of commercial building hvac systems.
In American Control Conference (ACC), 2013.
Udi Helman. Resource and transmission planning to achieve a 33% RPS in
California–ISO modeling tools and planning framework. In FERC Technical
Conference on Planning Models and Software, 2010.
Gregor P Henze, Moncef Krarti, and Michael J Brandemuehl. Guidelines for
improved performance of ice storage systems. Energy and Buildings, 35(2):
111–127, 2003.
Gregor P Henze, Clemens Felsmann, and Gottfried Knabe. Evaluation of op-
timal control for active and passive building thermal storage. International
Journal of Thermal Sciences, 43(2):173–183, 2004. ISSN 1290-0729.
Gregor P Henze, Jens Pfafferott, Sebastian Herkel, and Clemens Felsmann.
Impact of adaptive comfort criteria and heat waves on optimal building
thermal mass control. Energy and Buildings, 39(2):221–235, 2007.
Baihong Jin, Pierluigi Nuzzo, Mehdi Maasoumy, Yuxun Zhou, and Alberto
Sangiovanni-Vincentelli. A contract-based framework for integrated de-
mand response management in smart grids. In Proceedings of the 2nd ACM
International Conference on Embedded Systems for Energy-Efficient Built
Environments, BuildSys ’15, pages 167–176, New York, NY, USA, 2015.
ACM. ISBN 978-1-4503-3981-0. . URL http://doi.acm.org/10.1145/
2821650.2821658.
S.J. Julier and J.K. Uhlmann. “New Extension of the Kalman Filter to
Nonlinear Systems". In International Society for Optics and Photonics,
AeroSense’97, pp. 182-193, Seattle, WA, 1997.
S.J. Julier et al. “A New Approach for Filtering Nonlinear Systems". In
Proceedings of ACC, vol. 3, pp. 1628-1632, 1995.
Gour Karmakar, Ashutosh Kabra, and Krithi Ramamritham. Coordinated
scheduling of thermostatically controlled real-time systems under peak
power constraint. In Real-Time and Embedded Technology and Applica-
tions Symposium (RTAS), 2013 IEEE 19th, pages 33–42. IEEE, 2013.
144 References
Randy H Katz, David E Culler, Seth Sanders, Sara Alspaugh, Yanpei Chen,
Stephen Dawson-Haggerty, Prabal Dutta, Mike He, Xiaofan Jiang, Laura
Keys, et al. An information-centric energy infrastructure: The berkeley
view. Sustainable Computing: Informatics and Systems, 1(1):7–22, 2011.
B.I. Kilkis. “From Floor Heating to Hybrid HVAC Panel: A Trail of Exergy-
Efficient Innovations”. ASHRAE transactions, pages 343–349, 2006.
B.J. Kirby. Spinning reserve from responsive loads. United States. Department
of Energy, 2003.
Ioannis Konstantakopoulos, Costas J. Spanos, and S. Shankar Sastry. So-
cial game for building energy efficiency: Utility learning, simulation, anal-
ysis and incentive design. Technical Report UCB/EECS-2015-3, EECS
Department, University of California, Berkeley, Feb 2015. URL http:
//www.eecs.berkeley.edu/Pubs/TechRpts/2015/EECS-2015-3.html.
Michaël Kummert and Philippe André. Simulation of a model-based optimal
controller for heating systems under realistic hypothesis. 2005.
S.F. Larsen, C. Filippín, and G. Lesino. Thermal behavior of building walls
in summer: Comparison of available analytical methods and experimental
results for a case study. In Building Simulation, volume 2, pages 3–18.
Springer, 2009.
P. Levis, S. Madden, J. Polastre, R. Szewczyk, E. Brewer, et al. Tinyos: An
operating system for sensor networks. Ambient intelligence, 35, 2005.
Jian Liang and Ruxu Du. Thermal comfort control based on neural network
for hvac application. In Control Applications, 2005. CCA 2005. Proceedings
of 2005 IEEE Conference on, pages 819–824. IEEE, 2005.
Zaiyi Liao and Arthur L Dexter. An inferential model-based predictive control
scheme for optimizing the operation of boilers in building space-heating
systems. Control Systems Technology, IEEE Transactions on, 18(5):1092–
1102, 2010.
Simeng Liu and Gregor P Henze. Experimental analysis of simulated rein-
forcement learning control for active and passive building thermal storage
inventory: Part 1. theoretical foundation. Energy and Buildings, 38(2):
142–147, 2006a.
Simeng Liu and Gregor P Henze. Experimental analysis of simulated rein-
forcement learning control for active and passive building thermal storage
inventory: Part 2: Results and analysis. Energy and buildings, 38(2):148–
161, 2006b.
J. Löfberg. “Minimax Approaches to Robust Model Predictive Control". Dept.
of Electrical Eng. LinkÂĺoping University, 2003.
References 145
Y. Ma, F. Borrelli, B. Hencey, B. Coffey, S. Bengea, and P. Haves. Model pre-

dictive control for the operation of building cooling systems. In American
Control Conference (ACC), 2010, pages 5106–5111. IEEE, 2010.
Yudong Ma, Anthony Kelman, Allan Daly, and Francesco Borrelli. Predictive
control for energy efficient buildings with thermal storage. IEEE Control
system magazine, 32(1):44–64, 2012.
Y. Ma et al. “Model Predictive Control of Thermal Energy Storage in Building
Cooling Systems". In Control System Magzine, pp. 1-65, 2011.
M. Maasoumy. “Modeling and Optimal Control Algorithm Design for HVAC
Systems in Energy Efficient Buildings”. In Master Thesis, University of
California, Berkeley, 2011.
M. Maasoumy and A. Sangiovanni-Vincentelli. “Total and Peak Energy Con-
sumption Minimization of Building HVAC Systems Using Model Predictive
Control”. In Design and Test of Computers, 2012.
M. Maasoumy and A. Sangiovanni-Vincentelli. “Optimal Control of HVAC
Systems in the Presence of Imperfect Predictions". In DSCC, Fort Laud-
erdale, FL, 2012.
M. Maasoumy, A. Pinto, and A. Sangiovanni-Vincentelli. “Model-Based Hier-
archical Optimal Control Design For HVAC Systems”. In ASME Dynamic
Systems and Control Conference, Arlington, VA., pages 271–278, 2011.
M. Maasoumy, M. Razmara, M. Shahbakhti, and A. Sangiovanni-Vincentelli.
“Selecting Building Predictive Control Based on Model Uncertainty”. In
American Control Conference (ACC), 2014a.
Mehdi Maasoumy and Alberto Sangiovanni-Vincentelli. Buildings to grid in-
tegration: A dynamic contract approach. In Proceedings of the IEEE/ACM
International Conference on Computer-Aided Design, pages 473–478. IEEE
Press, 2015.
Mehdi Maasoumy and Alberto Sangiovanni Vincentelli. Comparison of con-
trol strategies for energy efficient building hvac systems. In Proceedings of
the Symposium on Simulation for Architecture & Urban Design, page 11.
Society for Computer Simulation International, 2014.
Mehdi Maasoumy, Jorge Ortiz, David Culler, and Alberto Sangiovanni-
Vincentelli. Flexibility of commercial building HVAC fan as ancillary ser-
vice for smart grid. In IEEE Green Energy and Systems Conference (IGESC
2013), Long Beach, USA, November 2013a.
146 References
Mehdi Maasoumy, Qi Zhu, Cheng Li, Forrest Meggers, and Alberto

Sangiovanni-Vincentelli. Co-Design of Control Algorithm and Embedded
Platform for HVAC Systems. In The 4th ACM/IEEE International Confer-
ence on Cyber-Physical Systems (ICCPS 2013), Philadelphia, USA, April
2013b.
Mehdi Maasoumy, Meysam Razmara, Mahdi Shahbakhti, and Alberto
Sangiovanni-Vincentelli. “Handling model uncertainty in model predictive
control for energy efficient buildings”. Energy and Buildings, 77:377–392,
2014b.
Mehdi Maasoumy, Catherine Rosenberg, Alberto Sangiovanni-Vincentelli, and
Duncan Callaway. Model predictive control approach to online computation
of demand-side flexibility of commercial buildings hvac systems for supply
following. In American Control Conference (ACC), 2014c.
Mehdi Maasoumy, Borhan Sanandaji, Kameshwar Poolla, and Alberto
Sangiovanni-Vincentelli. Model predictive control of regulation services
from commercial buildings to the smart grid. In American Control Confer-
ence (ACC), 2014d.
Mehdi Maasoumy, Pierluigi Nuzzo, and Alberto Sangiovanni-Vincentelli.
Smart buildings in the smart grid: Contract-based design of an integrated
energy management system. In Cyber Physical Systems Approach to Smart
Electric Power Grid, pages 103–132. Springer, 2015.
Mehdi Maasoumy, Barzin Moridian, Meysam Razmara, mahdi Shahbakhti,
and Alberto Sangiovanni Vincentelli. “Online Simultaneous State Estima-
tion and Parameter Adaptation for Building Predictive Control". In DSCC,
Stanford, CA, May 2013.
Mehdi Maasoumy Haghighi. Controlling energy-efficient buildings in the con-
text of smart grid: A cyber physical system approach. Number UCB, EECS-
2013-244 - PhD Thesis, Dec 2013. URL http://www.eecs.berkeley.edu/
Pubs/TechRpts/2013/EECS-2013-244.html.
MATLAB. http://www.mathworks.com.
F. Meggers and H. Leibundgut. “The reference environment: utilising exergy
and anergy for buildings”. Exergy, 11(4):423–438, 2012.
F. Meggers, V. Ritter, P. Goffin, M. Baetschmann, and H. Leibundgut. “Low
exergy building systems implementation”. Energy, 41(1):48–55, 2012.
N. Mendes, G.H.C. Oliveira, and H.X. de Araújo. Building thermal perfor-
mance analysis by using matlab/simulink. In Seventh International IBPSA
Conference, Rio de Janeiro, Brazil, 2001.
References 147
Modelica. Modelica Building Library. http://simulationresearch.lbl.gov/bcvtb/

releases/latest/doc/manual/tit-DymolaCon.xhtml.
Mark Modera, Tengfang Xu, Helmut Feustel, Nance Matson, Charlie
Huizenga, Fred Bauman, Edward Arens, and Tom Borgers. Efficient ther-
mal energy distribution in commercial buildings. LBNL-41365, Lawrence
Berkeley National Laboratory, 1999.
Truong X Nghiem, Madhur Behl, Rahul Mangharam, and George J Pappas.
Green scheduling of control systems for peak demand reduction. In De-
cision and Control and European Control Conference (CDC-ECC), 2011
50th IEEE Conference on, pages 5131–5136. IEEE, 2011.
Thierry Stephane Nouidui. Validation and application of the room model of
the Modelica buildings library. In 9th International Modelica Conference
2012, Munich, Germany, 2014.
P. Nuzzo and A. Sangiovanni-Vincentelli. Let’s get physical: Computer science
meets systems. In From Programs to Systems - The Systems Perspective in
Computing Workshop, European Joint Conferences on Theory and Practice
of Software (ETAPS), April 2014.
Frauke Oldewurtel, Alessandra Parisio, Colin N Jones, Dimitrios Gyalistras,
Markus Gwerder, Vanessa Stauch, Beat Lehmann, and Manfred Morari.
Use of model predictive control and weather forecasts for energy efficient
building climate control. Energy and Buildings, 45:15–27, 2012.
Frauke Oldewurtel, Alessandra Parisio, Colin N Jones, Manfred Morari,
Dimitrios Gyalistras, Markus Gwerder, Vanessa Stauch, Beat Lehmann,
and Katharina Wirth. Energy Efficient Building Climate Control Using
Stochastic Model Predictive Control and Weather Predictions. In Ameri-
can Control Conference (ACC), pages 5100–5105, Baltimore, MD, 2010.
F. Oldewurtel et al. “A Tractable Approximation of Chance Constrained
Stochastic MPC Based on Affine Disturbance Feedback". In CDC, pp.
4731-4736, 2008.
Slaven Peleš, Sunil Ahuja, and Satish Narayanan. Uncertainty quantification
in energy efficient building performance simulations. In Proceedings of of
2nd International High Performance Buildings Conference, number 3592,
2012.
Luis Pérez-Lombard, José Ortiz, and Christine Pout. A review on build-
ings energy consumption information. Energy and buildings, 40(3):394–398,
2008.
148 References
Glenn Platt, Jiaming Li, Ronxin Li, Geoff Poulton, Geoff James, and Josh
Wall. Adaptive hvac zone modeling for sustainable buildings. Energy and
Buildings, 42(4):412–421, 2010.
J. Polastre, R. Szewczyk, and D. Culler. Telos: enabling ultra-low power
wireless research. In Information Processing in Sensor Networks, 2005.
IPSN 2005. Fourth International Symposium on, pages 364–369. IEEE,
2005.
S. Privara, J. Cigler, Z. Váňa, F. Oldewurtel, C. Sagerschnig, and E. Žáčeková.
“Building modeling as a crucial part for building predictive control”. Energy
and Buildings, 56:8–22, 2013.
P. Radecki and B. Hencey. Online building thermal parameter estimation via
unscented kalman filtering. In American Control Conference (ACC), 2012,
pages 3056–3062. IEEE, 2012.
Meysam Razmara, Mehdi Maasoumy, Mahdi Shahbakhti, and Rush D Robi-
nett III. Exergy-based model predictive control for building hvac systems.
In American Control Conference (ACC), 2015. IEEE, 2015.
R.D. Robinett III and D.G. Wilson. “Nonlinear Power Flow Control De-
sign: Utilizing Exergy, Entropy, Static and Dynamic Stability, and Lya-
punov Analysis”. Springer, 2011.
P. Sakulpipatsin, L.C. M. Itard, H.J. Van Der Kooi, E.C. Boelman, and P.G.
Luscuere. “An exergy application for analysis of buildings and HVAC sys-
tems”. Energy and Buildings, 42(1):90–99, 2010.
Alberto Sangiovanni-Vincentelli, Werner Damm, and Roberto Passerone.
Taming Dr. Frankenstein: Contract-Based Design for Cyber-Physical Sys-
tems. European Journal of Control, 18(3):217–238, June 2012.
A. Schlueter and F. Thesseling. “Building information model based en-
ergy/exergy performance assessment in early design stages”. Automation
in Construction, 18(2):153–163, 2009.
D. Schmidt. “Low exergy systems for high-performance buildings and com-
munities”. Energy and Buildings, 41(3):331–336, 2009.
Matthias Schuss, Robert Zach, Kristina Orehounig, and Ardeshir Mahdavi.
Empirical evaluation of a predective simulation-based control method. In
Proceedings of the 12th International IBPSA Conference, pages 14–16,
2011.
Jagdev Singh, Nirmal Singh, and JK Sharma. Fuzzy modeling and control of
hvac systems-a review. Journal of scientific and Industrial Research, 65(6):
470, 2006.
References 149
Jan Širokỳ, Frauke Oldewurtel, Jiří Cigler, and Samuel Prívara. Experimental
analysis of model predictive control for an energy efficient building heating
system. Applied Energy, 88(9):3079–3087, 2011.
Joana Sousa. Energy simulation software for buildings: review and compari-
son. In Proceedings of the international workshop on information technology
for energy applications (IT4ENERGY 2012), Lisbon, Portugal. Citeseer,
2012.
ASHRAE Standard. “Standard 55-2004". Thermal environmental conditions
for human occupancy, 2004.
G. Strbac. Demand side management: Benefits and challenges. Energy Policy,
36(12):4419–4426, 2008.
D. Sturzenegger, D. Gyalistras, M. Morari, and R.S. Smith. Model predictive
climate control of a swiss office building: Implementation, results, and cost-
benefit analysis. Control Systems Technology, IEEE Transactions on, PP
(99):1–1, 2015. ISSN 1063-6536. .
David Sturzenegger, Dimitrios Gyalistras, Vito Semeraro, Manfred Morari,
and Roy S Smith. Brcm matlab toolbox: Model generation for model pre-
dictive building control. In American Control Conference (ACC), 2014,
pages 1063–1069. IEEE, 2014.
D Todd, M Caufield, B Helms, M Starke, B Kirby, and J Kueck. Providing
reliability services through demand response: A preliminary evaluation of
the demand response capabilities of Alcoa Inc. ORNL/TM, 233, 2008.
TRNSYS. TRNSYS. http://www.trnsys.com/.
J. Venkatesh, B. Aksanli, J.-C. Junqua, P. Morin, and T.S. Rosing. Homesim:
Comprehensive, smart, residential electrical energy simulation and schedul-
ing. In Green Computing Conference (IGCC), 2013 International, pages
1–8, June 2013. .
Evangelos Vrettos, Frauke Oldewurtel, Fengtian Zhu, and Göran Anders-
son. Robust provision of frequency reserves by office building aggregations.
In World Congress of the International Federation of Automatic Control
(IFAC), Cape Town, South Africa, 2014.
Andreas Wächter and Lorenz T Biegler. On the implementation of an interior-
point filter line-search algorithm for large-scale nonlinear programming.
Mathematical programming, 106(1):25–57, 2006.
Detlef Westphalen and Scott Koszalinski. Energy consumption characteristics
of commercial building hvac systems volume ii: Thermal distribution, auxil-
iary equipment, and ventilation. Arthur D. Little Inc (ADLI), 20:33745–00,
2001.
150 References
Michael Wetter, Wangda Zuo, Thierry S Nouidui, and Xiufeng Pang. Modelica
buildings library. Journal of Building Performance Simulation, 7(4):253–
270, 2014.
C.K. Woo, E. Kollman, R. Orans, S. Price, and B. Horii. Now that california
has ami, what can the state do with it? Energy Policy, 36(4):1366–1374,
2008.
Y. Yang, A. Pinto, A. Sangiovanni-Vincentelli, and Q. Zhu. A design flow for
building automation and control systems. In Real-Time Systems Sympo-
sium (RTSS), 2010 IEEE 31st, pages 105–116. IEEE, 2010.
Yang Yang, Qi Zhu, Mehdi Maasoumy, and Alberto Sangiovanni-Vincentelli.
Development of building automation and control systems. Design Test of
Computers, IEEE, 29(4):45 –55, aug. 2012. ISSN 0740-7475. .
A. Yildiz and A. Güngör. “Energy and exergy analyses of space heating in
buildings”. Applied Energy, 86(10):1939–1948, 2009.
M Zaheer-Uddin and GR Zheng. Optimal control of time-scheduled heating,
ventilating and air conditioning processes in buildings. Energy Conversion
and Management, 41(1):49–60, 2000.
Xiaojing Zhang, Georg Schildbach, David Sturzenegger, and Manfred Morari.
Scenario-based mpc for energy-efficient building climate control under
weather and occupancy uncertainty. In Control Conference (ECC), 2013
European, pages 1029–1034. IEEE, 2013.

FTN EDA

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

FTN EDA

Uploaded by

Copyright:

Available Formats

Foundations and Trends

Smart Connected Buildings Design Automation:

Mehdi Maasoumy Alberto Sangiovanni-Vincentelli

4 Building Control Design 41

4.5 Stochastic Model Predictive Control (SMPC) . . . . . . . 51

5 Test-beds and Real-scale Experiments 77

6 Designing Building Control Systems as Cyber-Physical Sys-

7 Dynamic Contracts for Building-Grid Interaction 110

Buildings are the result of a complex integration of multi-physics sub-

proposed to make buildings smarter and more efficient. Further, we

M. Maasoumy and A. Sangiovanni-Vincentelli. Smart Connected Buildings Design

The term intelligent or smart building refers to the next generation of

Figure 1.1: US Primary Energy Overview (Quadrillion Btu) EIA.

1.1 Why Buildings?

But why should we care about buildings? Americans, on average, spend

18 years, and another 25% growth is projected through 2030. Residen-

1.2 Why Smart Buildings?

U.S. Energy Flow, 2014

Figure 1.2: US Energy Flow in 2014 (Quadrillion Btu) EIA.

achieve these objectives, we need to make buildings smart about the

Figure 1.3: Breakdown of energy consumption in a typical building. Over 50% of

combined with the intelligence of the Building Energy Management

1.3 Areas of research

About 50% of the energy consumed in buildings is directly related to

• Building control design;

• Test-beds and real-scale experiments;

• Buildings as cyber-physical systems;

• Smart buildings in the smart grid ecosystem.

In this article, we provide an overview of what has been achieved

The remaining chapters of this article are organized as follows:

1.4.1 Chapter 2: Simulation Tools

We start by reviewing the simulation tools that have been developed

1.4.2 Chapter 3: Building Models

We then present our work in modeling buildings. In particular, we

1.4.3 Chapter 4: Building Control Design

We first provide an overview of classical building HVAC controllers. We

1.4.4 Chapter 5: Test-beds and Real-scale Experiments

In Chapter 5 we review some of the outstanding efforts in this domain,

1.4.5 Chapter 6: Designing Building Control Systems as Cyber-

After presenting various control strategies in Chapter 4, and reviewing

1.4.6 Chapter 7: Dynamic Contracts for Building-Grid Interaction

In Chapter 7, we address the future role of smart buildings in the con-

experiments to demonstrate the feasibility of the proposed algorithm.

1.4.7 Chapter 8: Conclusion

A variety of simulation tools have been developed for building energy

2.1 Building Simulation Tools

EnergyPlus is a whole building energy simulation software supported

calculations, glare simulation and control, luminaire controls, and the

The DOE-2 software was developed by James J. Hirsch & Associates

eQUEST is a freeware, complete, fully interactive, professional DOE-

Modelica is an equation-based modeling language that is well

zone heat transfer and multi-zone airflow and contaminant transport.

TRNSYS is a TRaNsient SYstems Simulation Program with a modular

ESPr is a general-purpose simulation environment, which supports an

assessments in which no single issue is unduly prominent. By address-

Building Resistance-Capacitance Modeling (BRCM) is a Matlab Tool-

eration of (bi-)linear resistance-capacitance type models from basic

2.2 Building-to-Grid Simulation Tools

2.2.1 HomeSim / S2 Sim

ing up to 50% reduction in grid energy through a combination of dis-

Co-simulation platforms are necessary to study the interactions of com-

Each of the building simulation software tools mentioned above has

libraries, control capability, and additional capabilities and defects. Ref-