A multi-level architecture to facilitate MPC implementation in

commercial buildings: basic principles and case study

José A. Candanedo1, Vahid R. Dehkordi2, and Phylroy Lopez3
1
CanmetENERGY-Varennes, NRCan, Varennes, Québec
2
CanmetENERGY-Ottawa, NRCan, Ottawa, Ontario

Abstract
This paper presents a methodology aimed at facilitating the de-
ployment of model-based predictive control (MPC) in buildings. MPC
has shown promise as an effective way to reduce utility costs associat-
ed with peak demand, and to better manage the interaction between
“smart buildings” and the “smart grid”. However, steps are needed to
streamline the implementation of MPC in buildings and thus encour-
age its adoption in building operation. The proposed architecture in-
tends to contribute to this goal by enabling a “compartmentalized”,
distributed, hierarchical approach to building modelling and controls.
The proposed multi-level methodology allows formulating control
problems so that the planning time horizon fits the scale of the system.
A model of a commercial building, including thermal energy storage
devices at different control levels, is used to demonstrate the method-
ology. Low-order resistive-capacitance models for the thermal spaces
are obtained from a detailed model created in EnergyPlus.

1 Introduction
The application of model-based predictive control (MPC) to the operation of buildings has
received a great deal of attention in recent years (Ma et al., 2010, Nghiem and Pappas, 2011,
Siroký et al., 2011, Candanedo and Athienitis, 2011, Kim and Braun, 2012, Corbin et al.,
2013). MPC has come to be recognized as an effective technique for improving load man-
agement in high-performance buildings, and as a promising approach to the incorporation of
renewable energy sources. MPC and similar methods are expected to play a key role for the
integration of smart buildings in the smart grid.
Despite these promising prospects, the practical implementation of a formal MPC
strategy in a building –understood as the application of a model-based optimization algo-
rithm– is a rather daunting task today. Reaching the state in which online MPC strategies can
be applied in buildings in a timely and cost-effective manner requires advances in several are-
as. These areas include appropriate modelling (Prívara et al., 2012, Eisenhower et al., 2012,
Candanedo et al., 2013a); automatic formulation and solution of optimization problems
(Cigler et al., 2013); tools for obtaining weather forecast information (Candanedo et al.,
2013c); data-collection and modelling of occupancy (Oldewurtel et al., 2012, Gunay et al.,
2013), among others.
This paper presents the outline of an approach aimed at facilitating the testing of pre-
dictive control strategies and the implementation of MPC in buildings (Candanedo and
Dehkordi, 2013). This methodology is one of core components of a four-year project at our
institution1; it is based on the “dissection” of the complex structure of a commercial building
into smaller control areas arranged hierarchically (e.g., room/thermal zone, group of zones,
whole building), which may be nested into each other. With this method, relatively simple

1
Multi-level Control for Buildings (MLCB), EcoEII program.
models (RC, state-space representations, simple ANN, etc.) suffice to represent each control
region satisfactorily. The methodology, still under development, is illustrated in this paper
with a case study building with five zones with active and passive thermal energy storage.

Principles of Multi-level MPC Control

The proposed methodology, summarized in Figure 1, is based on the following principles:
• Multi-level approach. Dividing the building spaces and the HVAC system in
control regions arranged in different hierarchical levels.
• Simplified models. Using simple models at each control region, both for the
thermal response of the spaces and the mechanical equipment. Simpler models
will be preferred, provided that they are accurate enough for the task at hand.
These models will be adjusted and fine-tuned with online collected data during
normal operation.
• Local control. Setting-up local “controlling agents” at each control region. These
local “agents” will operate as decision-making entities which will use the models
corresponding to their jurisdiction, together with forecasts of input data (weather,
occupancy, etc.), for load predictions and selection of control actions.
• Communication. Enabling communication and “negotiation” protocols between
the controlling agents, so that discrepancies in operation policies, which will inev-
itably occur, are resolved.

Multi-level structure

Building level GM

Branch
Group level M-1 M-2 without
(e.g., wing of building) model

Branch
Thermal zone level M-1.1 M-1.2 without M-2.1 M-2.2
(e.g., room) model

Controlling agents (MPC controllers) Time scales

GM

48 hr

M-1 M-2

8 hr

M-1.1 M-1.2 M-2.1 M-2.2

30 min

Figure 1: Conceptual description of the proposed multi-level control approach.

2 Methodology
Case study building
This investigation uses a small commercial building model created in EnergyPlus. This build-
ing model (codenamed “Chipmunk”) was designed to test and develop the multi-level control
approach methodology. The Chipmunk is a single-story building with a rectangular plan of
800 m2 (40 m × 20 m), as shown in Figure 2.

CONTROL
BUILDING
LEVEL 1

CONTROL
Z1 Z2 Z3 Z4 Z5
LEVEL 2

(a) (b)

Figure 2. (a) EnergyPlus model of “Chipmunk” building; (b) control levels.

Other key features of the case study building are:
• Double-glazed windows, window/wall ratio ≈ 40%
• R-20 (RSI-3.53) in walls
• 0.80 ACH (infiltration + ventilation)
• Façades oriented towards the cardinal points

Proposed RC structures for thermal models

A previous preliminary paper by our team made use of sets of transfer functions (i.e., SISO
systems) to model the response of the building and its zones (Candanedo and Dehkordi,
2013), which were obtained by using the MATLAB System Identification Toolbox. Such an
approach, based on sets of transfer functions, has been used in previous studies on MPC for
radiant floor heating systems (Candanedo and Athienitis, 2011) and ice storage devices
(Candanedo et al., 2013b).
Instead of transfer functions, the present study uses 3rd order, grey-box, RC thermal
networks to model the response of the building and its zones. Low-order RC networks repre-
sent a compromise between accuracy and physical significance (Candanedo et al., 2013a).
Two slightly different RC structures are proposed for the two control levels (Figure 3).
Building model Zone model
R2,ext R3,ext

R1,ext 1 R1,2 2 R2,3 3 R1,ext 1 R1,2 2 R2,3 3

0.6 ( qIG + qh )

+ qh )

qh ,c + qIG

+ +
0.4 ( qIG

Text _ C1 C2 qSG C3 Text _ C1 C2 0.7qSG 0.3qSG

Figure 3. RC thermal networks for the building and zone levels.

 These structures are not fixed. The selection of the structure and complexity level of
the thermal network is left to the good judgment of the user, depending on the application.

From RC thermal network to state space representation

To find the values of the RC parameters in Figure 3, the first step was to develop a MATLAB
code to convert the RC network into an equivalent canonical state-space representation. This
state-space formulation was presented in a previous paper (Candanedo et al., 2013a). Three
elements are needed in order to carry out this conversion: (a) information on the RC configu-
ration; (b) information on the inputs (type of input, on which node they are acting); (c) infor-
mation on the outputs (the information we are interested about).
First, the RC circuit configuration is described by following this convention in a text
file: type of device (R or C), first node, second node and value of the element. The units for
each of the elements are given in a consistent set of units. In this case, resistance is given in
K/kW, and capacitance in kJ/K (resistances could be given in K/W and capacitances in J/K).
For example, the zone model is described as:
C 1 0 99845.92
C 2 0 5751.61
C 3 0 1000.00
R 1 4 5.00
R 1 2 50.00
R 2 3 1.72
R 3 4 8.16

In this case, “C 1 0” refers to the capacitance connected between node 1 and the
ground (namely C1). Node #4 is the one to which the outdoor temperature (Text) signal is con-
nected. The ground node or earth node (node 0) is the reference. The thermal capacitances
connected to the ground node define the states of the system.
After describing the circuit, it is important to indicate: inputs (whether they are tem-
perature or heat sources, and whether they act totally or partially on each of the nodes; and
desired outputs. For the case of the inputs of the zonal model, the text file has this format:
T,4
Q,2,0.65,3,0.35
Q,3,1
Q,3,1

In this case, the file indicates that there are four inputs: (a) a temperature source con-
nected to the 4th node; (b) a heat source of which 65% goes to node 2 and 35% goes to node 3;
(c) a heat source that is received in its entirety (100%) by node 3 and; (d) another heat source
received by node 3. These sources represent, respectively, outdoor temperature, solar gains,
internal gains and heating (the last two sources are interchangeable).
Finally, the desired outputs are defined by the following brief text file which simply
means that there is a single output: the temperature of node 3.

T,3

Formulation of optimization problem

Once the circuit structure and the position inputs and outputs are defined, the state space rep-
resentation can be readily found, and put into an optimization loop to determine the optimal
RC values according to a given objective function.
 The Euclidean norm is often used to compare vectors. The Euclidean norm (or Euclid-
ean distance) is defined as the square root of the sum of the square of the elements, i.e.:

x= x1 + x2 +  x M =
2 2 2
∑x 2
j
(1)
j

The Euclidean norm of the difference between the reference and the output of the sim-
plified model can be used as one of the criteria to select the “best match” for an RC circuit
response. In this case, the objective function is defined based on a comparison of the response
of the RC and the EnergyPlus benchmark under three situations: (i) free floating conditions
(outdoor temperature, solar gains and internal gains acting together); (ii) response to solar
gains; (iii) response to internal gains.
J =0.50 y ff,E+ − y ff,RC + 0.25 y SG,E+ − y SG,RC + 0.25 y IG,E+ − y IG,RC (2)
In this case, more weight is given to the response under free floating conditions. Fig-
ure 4 summarizes the system identification procedure for the identification of RC parameters.

Define RC structure, inputs

and outputs (circuit.txt,
inputs.txt, outputs.txt)

Propose guess values for

the RC parameters to
initialize the optimization

Convert to State Space

representation by finding
the matrices A,B,C and D.

Apply input signals and

Create new circuit with new
obtain the response
RC parameters
(MATLAB lsim function)

Propose new RC values

Evaluate the response
according to an
according to a predefined
optimization algorithm
objective function )
(e.g., fmincon function)

Satisfactory
values?
NO

YES

Final RC values

Figure 4. System identification procedure for RC parameters.

 Figure 5 shows a comparison between the EnergyPlus and the RC models under free-
floating conditions. The distribution of the error approaches a normal distribution with a mean
value of μ ≈ -0.15 °C and a standard deviation of σ ≈ 1°C.
20

EnergyPlus RC model
15

10
Temp (°C)

-5

-10

-15
0 10 20 30 40 50 60
Time (days)
25
∆T, evaluation over a year
20 µ =-0.14924
σ =1.0082
Freq (%)

15

10

0
-6 -4 -2 0 2 4 6
°C

Figure 5. Free floating response of EnergyPlus and simplified RC model, zone 1.

Additional considerations in model performance

The performance of the identified RC circuits is quite satisfactory. However, there is signifi-
cant room for improvement to reduce the discrepancies observed.
• Updating of model states. A direct comparison between EnergyPlus and the
RC circuit, while acting independently from each other, is a conservative crite-
rion to assess the performance of the RC circuit. In practice, the predictions of
the RC circuit would be compared periodically with on-site measurements. In
other words, by collecting data from the real building the new states (and out-
puts) predicted by the RC model can be corrected. This correction could be
done either by simply introducing the new values or by using a more sophisti-
cated method to weigh the confidence on the sensor measurements (e.g. a
Kalman filter). This approach of continuous state correction would effectively
result in better predictions, especially at shorter prediction horizons.
• Design of system identification experiments. While there has been interest in
using simplified controls for many years, there is a need to develop a systemat-
ic procedure for the selection of the model order, structure and value of param-
eters. In particular, apart from the criteria proposed in Equation (2), other “vir-
tual experiments” could be designed to test the validity of the model. These
 experiments could include the response at particular frequencies (response at
one-cycle per day), steady-state response (“effective R-value”), etc.
• Free-floating versus controlled-conditions response. The vast degree of
fluctuation of a free-floating system (i.e., without a control system) implies
that a model most perform very well under a wide range of conditions. In prac-
tice, the control model would operate in a much more restricted range of states.
• Control-oriented model vs. simulation model. The simplified RC circuit
proposed is intended as a control model, in the sense that it will be used for
decision-making over a limited prediction horizon (ranging from a few
minutes up to a couple of days). Unlike a building simulation model, a control
model is not intended to replicate the behaviour of a real building, but to serve
as a tool to make better informed decisions.

Development of a multi-level architecture: prototype of emulator

Once the models for the building and the zones have been identified, they can be arranged in a
multi-level architecture in a simulation environment. In this study, the models have been in a
Simulink platform, with the goal of developing an emulator for testing and developing control
strategies. The Simulink environment, an offshoot of MATLAB used for dynamic simula-
tions, contains several drag-and-drop blocks for signal processing and control decisions. Oth-
er advantages include the possibility of selecting the method for the treatment of the differen-
tial equations (Euler, Runge-Kutta, etc.), the large availability of control tools (e.g., PID con-
trols, system identification). While Simulink is a general tool that does not contain specific
models building systems, programs developed in MATLAB can easily be incorporated.
This environment is comprised of the following elements:
• Simulation models, representing the response of the real building and zone. As
a first attempt, these models are slight modifications of the previously identi-
fied RC circuits. Ideally, and planned for a later development, the simulation
model will be replaced by an EnergyPlus simulation model, or at the very least
a higher order state-space representation.
• Control agents, in charge of decision making. Each of these control agents
contains its own control model (i.e., a previously identified RC model), which
is then used to calculate loads, and then make control decisions for the alloca-
tion of resources in the energy storage elements. Figure 6 compares the two
types of models to be used in the emulator.

Simulation Model Control-oriented Model

(used as benchmark) (used by control agents)
• Intended to represent reality, • Used for decision-making and
i.e. the behaviour of an optimization
Data
actual building, as accurately • Meant to facilitate calculations
as possible • Meant to provide insight, not
• Used for “virtual super-accurate.
experiments” and validation • Model contained in the “brain”
Decisions
• Not required if real data/real of the control system
building is available

x Ax + Bu
e.g., low-order =
e.g., =
y Cx + Du
Figure 6. Types of models used in emulator.
 • Metronome or “time-keeper”. This is a critical element that sends trigger sig-
nals to each of the control agents at regular intervals. These signals are sent to
the control agents, which then proceed to execute an optimization exercise
within its own jurisdiction and for its respective control horizon time scale.
Table 1 shows typical time scales for different control levels.

Table 1. Example of time scales used.

Control Number of
Control level Control horizon Sampling time
level rank intervals
1 Building 24 h 1h 24
2 Group control horizon 6h ½h 12
3 Room ½h 5 min 6
NA Simulation time step 1/24 h (2.5 min) ---- ----

A screenshot of the prototype emulator under development is shown in Figure 7.

Figure 7. Screenshot of emulator prototype (under development).

 The prototype emulator proposed in this study makes use of a Real-Time Pacer block
(Vallabha, 2010), which allows the human operator to “slow down” the simulation to take a
closer look at the response of the controllers and the building. For example, the simulation
can be made to run so that 1 second of simulation time is equivalent to 1 day of “real time”.

Treatment of uncertainty

Several sources of uncertainty are present in the modelling of dynamic response of buildings
for predictive control applications:
• Uncertainty in the model
o Source information. construction details, thermal properties of materials, opti-
cal properties of surfaces, geometric details, furniture layout, etc.
o Simplifying modelling assumptions. Even if perfect information were available,
physical modelling –even the most detailed one– requires simplifications and
approximations (e.g., , level of spatial discretization, etc.).
• Uncertainty in the dynamic inputs
o Weather forecast. Although increasingly reliable (especially over a period of a
few days) weather forecasts are obviously not exact. However, abundant in-
formation on the degree of confidence, based on the output of several weather
forecast models, is becoming available.
o Occupancy and user loads forecast. Perhaps the most difficult input to predict
is the number of occupants and their energy consumption patterns. Neverthe-
less, in the case of control applications in existing buildings, occupancy pat-
terns can be observed (e.g., how many people are present during weekdays).
User loads (lighting or electric appliances) can be monitored if sub-metering is
in place. This information becomes more reliable as more data is gathered dur-
ing normal operation. Both occupancy and user load information can be pre-
sented in terms of probabilistic descriptions (Figure 8). Recent publications fol-
low a stochastic approach (Widén and Wäckelgard, 2010).

Figure 8. Example of occupancy modelling (sample values). A different probability dis-

tribution can be used for each hour of the day.
A key advantage of a distributed, hierarchical approach based on low-order models is
that it facilitates the incorporation of uncertainty, and crucially, how to deal with it when mak-
ing a decision concerning the operation of the building. A model with fewer parameters
makes it easier to compare scenarios (for example, by running Monte Carlo simulations).
Ongoing work
At the time of writing, a number of tasks are under development for the emulator:
• Implementation of models for active thermal energy storage devices (e.g.,
brick thermal energy storage for heating and ice-storage for cooling).
• Implementation of automatic optimization algorithms that will be triggered at
regular intervals according to signals received from the “metronome”, accord-
ing to the different prediction horizons. These optimization algorithms have al-
ready been tested “off-line” with M-files (Candanedo et al., 2013b).
• Formulation of the “negotiation” protocol between levels. Different “priori-
ties” will be considered.
• Investigation of different approaches to deal with uncertainty, due to model in-
accuracies and forecast errors; development of input signals related to occu-
pancy and occupant behaviour.
• Implementation in a larger-scale model (“Elephant” model).

3 Conclusions
This paper has presented an overview of the basic principles of a multi-level architecture for
the study of MPC and other advanced control concepts in building systems. This architecture
is currently being applied in the development of an emulator tool. Also, this architecture will
also be used in the development of “control agents” in charge of a control region.
Model-based control applications require continuous verification with data obtained
on-site. Models for the building and its equipment should be made so that they can be easily
modified when online data collection is available. Given the inherent complexity of buildings,
it is difficult to assess a priori every possible factor having an impact on the building re-
sponse. Rather than attempting to create a complex model with a “perfect” prediction, it is
preferable to rely on a model with relatively few parameters. These parameters may receive
kick-start guess values that may be easily corrected as soon as on-site data from the real build-
ing is available. Simpler models facilitate the mathematical treatment of uncertainty. As men-
tioned in the paper presented at ICEBO “rather than perfect accuracy, the goals pursued are
logical structure and coherence” (Candanedo and Dehkordi, 2013).
Although low-order RC networks have been presented, the proposed multi-level archi-
tecture does not depend on a given kind of model. First-principle models, grey-box or artifi-
cial neural networks may be applied, although grey-box models may be more amenable for
practical implementation. Using simplified models also allows shifting the attention of the
control engineer towards the often-neglected issue of the quality of the dynamic input data.
The emerging interest in the development of better models for prediction of occupancy and
occupant behaviour is encouraging.
Apart from thermal models represented with linear networks, it is also necessary to in-
corporate models for other relevant phenomena (e.g., electric energy use, lighting, etc.). The
multi-level control scheme will also be useful in this case.
This multi-level, distributed architecture also enables the study of fault scenarios. For
example, if one of the zone control agents fails, the heating/cooling power calculated by the
building model can be used to provide an estimate of the requirements of the zone.
An essential element in the concept of multi-level control is the incorporation of the
different time scales and prediction horizons for each of the control levels. This methodology
acknowledges the presence of patterns or cycles (weather, occupant behaviour) which repeat
themselves at different periodic intervals, forming a juxtaposition of different “rhythms”.
4 Acknowledgements
Funding for this work was provided by Natural Resources Canada through the Office of En-
ergy Research and Development (OERD). The authors would like to thank Jacques Martel
and Justin Tamasauskas for their valuable comments.

5 References
Candanedo, J. & Dehkordi, V. R. 2013. Multi-level approach for model-based predictive
control (MPC) in buildings: a preliminary overview. International Conference for
Enhanced Building Operation (ICEBO) 2013. Montreal, Quebec.
Candanedo, J., Dehkordi, V. R. & Lopez, P. 2013a. A control-oriented simplified modelling
strategy. 13th International Building Simulation Conference (IBPSA 2013).
Chambéry, France.
Candanedo, J., Dehkordi, V. R. & Stylianou, M. 2013b. Model-based predictive control of an
ice storage device in a building cooling system. Applied Energy, 1111032–1045.
Candanedo, J., Paradis, E. & Stylianou, M. 2013c. Building simulation weather forecast files
for predictive control strategies. SimAUD 2013. San Diego, California.
Candanedo, J. A. & Athienitis, A. K. 2011. Predictive control of radiant floor heating and
solar-source heat hump operation in a solar house. HVAC & R Research, 17(3), 235-
256.
Cigler, J., Gyalistras, D., Siroký, J., Tiet, V.-N. & Ferkl, L. 2013. Beyond theory: the
challenge of implementing model predictive control in buildings. 11th REHVA World
Congress Clima 2013. Prague, Czech Republic.
Corbin, C. D., Henze, G. P. & May-Ostendorp, P. 2013. A model predictive control
optimization environment for real-time commercial building application. Journal of
Building Performance Simulation, 6(3), 159-174.
Eisenhower, B., Gasljevic, K. & Mezic, I. 2012. Control-oriented dynamic modeling and
calibration of a campus theater using Modelica. SimBuild 2012. Madison, Wisconsin.
Gunay, H. B., O'Brien, W. & Beausoleil-Morrison, I. 2013. A critical review of observation
studies, modeling, and simulation of adaptive occupant behaviors in offices. Building
and Environment, 7031-47.
Kim, D. & Braun, J. E. 2012. Reduced-order building modeling for application to Model-
Based Predictive Control. SimBuild 2012 Madison, Wisconsin.
Ma, Y., Borrelli, F., Hencey, B., Coffey, B., Bengea, S. & Haves, P. 2010. Model predictive
control for the operation of building cooling systems. 2010 American Control
Conference. Baltimore, Maryland, USA.
Nghiem, T. X. & Pappas, G. J. 2011. Receding-horizon Supervisory Control of Green
Buildings. 2011 American Control Conference. San Francisco, California.
Oldewurtel, F., Sturzenegger, D. & Morari, M. 2012. Importance of occupancy information
for building climate control. Applied Energy, In Press.
Prívara, S., Cigler, J., Vana, Z., Oldewurtel, F., Sagerschnig, C. & Zacekova, E. 2012.
Building modeling as a crucial part for building predictive control. Energy and
Buildings, In Press.
Siroký, J., Oldewurtel, F., Cigler, J. & Prívara, S. 2011. Experimental analysis of model
predictive control for an energy efficient building heating system. Applied Energy,
883079-3087.
Vallabha, G. 2010. Real-Time Pacer for Simulink (Simulink block).
Widén, J. & Wäckelgard, E. 2010. A high-resolution stochastic model of domestic activity
patterns and electricity demand. Applied Energy, 871880-1892.