2010 American Control Conference FrA12.

4
Marriott Waterfront, Baltimore, MD, USA
June 30-July 02, 2010

Predictive Control of Complex Hydronic Systems

Vikas Chandan, Sandipan Mishra and Andrew G. Alleyne

Abstract— The control of hydronic building systems is con- a simulated chilled water system, revealing that an MPC
sidered in this paper, using a simulated chilled water system scheme is advantageous in several aspects, including energy-
as a case study. In this context, model-based predictive control efficiency and reliability. However, the computational com-
strategies have been proposed and compared with traditional
feedback control schemes. The advantages and limitations plexity associated with it may render it impractical for very
associated with these methodologies has been demonstrated. large scale building systems. To overcome this limitation,
The cornerstone of this work is the development of a novel, dis- a novel, distributed scheme for such systems has been
tributed predictive scheme which provides the best compromise developed in this work, which retains the attractive features
in the multidimensional evaluation framework of ‘regulation’, of MPC with sufficient conformity and significantly low
‘optimality’, ‘reliability’ and ‘computational complexity’.
computational requirements.
I. INTRODUCTION The organization of this paper is as follows. The physical
description of the system, its modeling framework and con-
The building sector consumes around 41% of total energy trol objectives are discussed in section II. The main ideas
in the United States, and accounts for nearly one-third of behind the traditional control schemes are summarized in
total greenhouse emissions. Around 35% of this contribution section III. Section IV provides details on the development
can be attributed to the heating and cooling systems present of centralized and distributed MPC schemes for the system.
in buildings [1]. The problem of efficiency enhancement of A detailed comparitive analysis of the performance of the
building HVAC systems presents diverse opportunities from proposed controllers has been made in section V for a chosen
a research perspective in several different areas of technology case study. Lastly, important conclusions and avenues of
such as design, architecture, alternative energy and control future research have been identified in section VI.
systems. In this regard, the opportunities offered by the field
of controls engineering are particularly important because II. SYSTEM DESCRIPTION
controls has a huge ‘retrofit potential’ in the sense that it A. Physical Layout
can be successfully applied to improve both the efficiency
A centralized building HVAC system is a physical net-
and performance of older, existing HVAC systems. Strong
work of interacting thermo-hydraulic components. A build-
arguments for energy efficiency in the exisiting building
ing cooling system, in particular, consists of four different
stock have recently been made [2], therefore underlining the
fluid loops - (i) the condenser water loop, (ii) the refrigerant
usefulness of controls in achieveing such goals.
loop in the chillers, (iii) the chilled water (hydronic) loop,
Central building and district hydronic HVAC systems and, (iv) the air-flow loop. The desired objective of cooling
have become popular in recent years due to the associated is realized by energy exchange between these fluid loops,
operational and energy related benefits. As a consequence, the building zones and the ambient. This work specifically
research aimed towards lowering the energy consumption of focuses on the control of the chilled water loop, which is
such central systems through advanced control has picked a subsystem of the larger HVAC system. Fig. 1 shows the
up considerable momentum. In this regard, the general area schematic layout of a system which emulates the chilled
of optimal control is the most promising (see the survey [3] water loop architecture of a 2-story building having three
and specific examples in [4], [5]). Model Predictive Control clusters of zones in each story. This network has been used
(MPC) [6] is one such useful and practical optimal control as a test system for the tools presented in this paper.
philosophy that has gained popularity in several applications.
In this paper, we attempt to investigate the role that MPC B. Reduced order dynamical model
can play in building HVAC systems for achieving various The predictive control strategies explored in this paper
desired control objectives, the most important being the require a succinct but satisfactory mathematical model of the
need to consume less energy without compromising other underlying dynamics. A formal procedure for obtaining state-
performance related goals. We are particularly interested in space models for hydronic HVAC system was introduced
its use for the control of the hydronic loop, motivated by in [7], and has been summarized in Fig. 2. This modeling
the fact that traditional local feedback strategies are still methodology was applied on the test system of Fig. 1
widely employed for their control. A comparitive analysis for suitably chosen nominal operating conditions, and the
against traditional control strategies has been performed on ensuing reduced order model obtained is of the form (1). The
manipulated variables, u(t) are the instantaneous capacities
Alleyne ([email protected]), Chandan and Mishra are with the Depart-
ment of Mechanical Science and Engineering, University of Illinois, Urbana, of chillers 1 and 2, and the opening factors of valves 3-6.
IL 61801, USA The state variables, x(t) are the coil (structure) temperatures

978-1-4244-7425-7/10/$26.00 ©2010 AACC 5112

 CHILLER 1
the choice of a control strategy is the amount of hardware
m16 and computational resources that needs to be deployed.
m13
Junc 10 Valve 1

Z1 III. TRADITIONAL CONTROL SCHEMES

Primary Pump 1
CHILLER 2

m17 Traditional liquid-loop control schemes rely on the local

m14
Junc 11 Valve 2
Junc 1
feedback control (On-off/P/PI) of the chillers and the heat
Junc 8

Primary Pump 2
Z2 exchangers ([8], [9]). The objective of the chiller controller
m5 m1 is to regulate the outlet chilled water temperature to a
m15 Valve 3
prescribed set-point via capacity control of the chiller. The
LAHX 1
controller for the heat exchanger then seeks to achieve the
m11 m6 m3 Junc 2
Junc 9

Junc 6
required instantaneous energy demand by manipulating the
Valve 4 Junc 4

LAHX 2
Z3 chilled water flow rate through it via local valve control. In
Booster Pump 1
m7 the past, the chilled water temperature set-point in building
m12 Valve 5
m2
cooling systems was usually fixed at a suitable low value
LAHX 3
between 5 to 7 deg C based on peak load requirements [10].
m8
During off peak operation, the chilled water would be diluted
Valve 6

LAHX 4
with a bypass stream of the return warm water so as to
m9 m4 Junc 3 meet the variable energy demands. However, as is evident,
Junc 7
Valve 7 Junc 5
this strategy is highly inefficient from an energy standpoint.
Z4
LAHX 5
Booster Pump 2
Therefore, a new principle called ‘chilled water temperature
m10
reset’ (CWTR) has been advocated in recent years ([11])
Valve 8

LAHX 6
wherein the chilled water set-point is adjusted during the
course of the day based on the net energy requirements of the
Fig. 1. Schematic layout of test system building. In this work, a CWTR strategy has been assumed
for specification of the chilled water set-point schedule in
Hydronic System both the on-off and PI traditional control schemes considered.
Operating
Topology
Conditions IV. PREDICTIVE CONTROL SCHEMES
Graph Component level
Representation linearization Model Predictive Control (MPC) schemes are becoming
Connectivity
Matrices
Coefficient
Matrices
increasingly popular for a wide variety of processes, which
Concatenation
can be attributed to their ability to handle constrained mul-
Full-order state
space model tivariable problems and the fact that they are intuitively tun-
Time-scale
Analysis
able. A building HVAC system is a particularly suitable can-
didate for the application of predictive control methodologies
Reduced order
state space model because of multiple control objectives, inherent complexity
due to coupled and multivariable nature of the problem, and,
Fig. 2. Summarized modeling procedure [7] presence of physical constraints.
in the chillers and Liquid Air Heat Exchangers (LAHXs), There has been significant interest lately in using MPC
whereas, the outputs, y(t) are the instantaneous cooling for various aspects of building HVAC control. Most research
capacities of the LAHXs. The model is open loop stable. efforts have focused on optimal operating strategies in the
context of VAV systems (optimal air flow rate and air
ẋ(t) = Ax(t) + Bu(t) temperature set-points) [12], [13], thermal-storage (optimal
y(t) = Cx(t) (1) charging and discharging schemes) [14], and load-side analy-
sis (optimal zone temperature and ventilation set-points) [14],
C. Control Objectives [15]. The common underlying theme in these efforts is the
The zonal cooling loads in a building change during the supervisory control of the HVAC system or its constituent
course of the day, mainly depending on the occupancy and subsystems. Control of the chilled water flow and tempera-
ambient conditions. The primary goal associated with the ture in the hydronic loop to achieve the setpoints dictated by
control of the hydronic subsystem is that the heat exchangers supervisory controller(s) is still assumed to be conventional
must satisfactorily achieve these varying energy demands (local on-off or P/PI schemes). In this work, we extend
at all instants of time. These demands must be met in the MPC framework to this ‘inner loop’ control problem,
an energy efficient manner for lower operating costs and which pertains to control of flow rate in the hydronic loop
reduced emissions. Enhanced life-span of the HVAC system subsystem together with chiller control, in order to meet the
is also an important consideration so as to maximize the various control objectives. Two versions of model predictive
return on the initial investment. This mandates that the scheme viz. (i) centralized and, (ii) distributed are being
mechanical components must be subjected to minimal fatigue proposed in this paper, whose details have been provided
loading during their operation. Another concern governing in the rest of this section.

5113
A. Centralized MPC For details of Hk , fk , Gk and wk appearing in this
The MPC scheme presented, seeks to minimize the fol- formulation, the reader is directed to [16]. Solution to this
lowing objective functional (2) at any given time instant ’k’. optimization problem can be obtained online using standard
The notation for the variables used here is as per the standard solvers available in applications such as MATLAB which
convention in MPC literature, e.g. [6]. use the active set or barrier function methods. The optimal
control at sample instant k, u∗k is given by (6)
Ny T
N
 Nu
−1  N 
 u∗k = [vk∗ (1) vk∗ (N + 1) ...vk∗ ((Nu − 1) × N + 1)] (6)
2
Jk = αj uj (k|k + i) + β [zj (k|k + i)]
i=0 j=1 i=1 j=1 B. Distributed MPC
Ny
N  A major drawback of centralized implementation of MPC

+γ [yj (k|k + i) − yj,ref (k|k + i)]2 for large scale systems is the requirement of high computa-
i=1 j=1 tional and memory resources. This justifies the development
N
 Nu
−1  of other computationally efficient control schemes which
+ψ [uj (k|k + i) − uj (k|k + i − 1)]2 (2) still retain the attractive features of MPC (see [17] for a
i=1 j=1 review). In this context, distributed MPC is a promising
Here, the recursive relation (3) defines the signal z as the alternative, which solves large scale optimization problems
integral of the output regulation error. in a decentralized manner through the solution of smaller
optimization problems, handled by local agents [18].
zj (k|k+i+1) = zj (k|k+i)+yj (k|k+i)−yj,ref (k|k+i) (3) An analysis of the model obtained in section II.B reveals
The objective functional is a weighted sum of various sub- useful information about the nature of interaction among
objectives, with weights β, γ and ψ that can be tuned. The the the various physical components in the system. Most
first term seeks minimization of energy consumption, while importantly, it is observed that the interactions between a
the second and third terms are aimed at achieving satisfactory given sink element (LAHX) and other sink elements are
regulation of the outputs. The last term was included to arrest relatively weak when compared to its interactions with any of
abrupt changes in actuation signals. the source elements. Therefore, the coupling architecture of
The linear system model, explained in section II.B, used these systems can be described by the paradigm of a leader-
by the controller is reliable only within a suitable operating follower dynamical network such as ant-colonies and bird-
range, say, ±25% of the nominal operating conditions, flocks [19]. Here, the role of leader and followers are played
about which the linearization was performed. Therefore, by the source and sink elements respectively.
the proposed controller is designed to meet the control A LGORITHM: Communication based distributed MPC
objectives only within that operating range, which necessi-
N OTATION: The subscripts m and i are used to denote
tates the imposition of saturation constraints on the inputs.
the leader (master) agent and the ith follower (slave) agent
Furthermore, slew rate constraints are also forced on the
respectively. The subscript k in (5) denoting the current time
input channels, with the same objective of limiting abrupt
instant is dropped for convenience. For definitions of the
changes in the signals as the last term in (2). Here, the
matrices and vectors Hm,m , Hi,i , Hm,i , fm , fi , Gm , Gi ,
underlying assumption in the control philosophy is that the
wm and wi , the reader is directed to [16].
entire operating range of the system shall be decomposed
into narrow, (and preferably overlapping for smoothness of S TEPS:
switching) segments with an MPC controller designed for 1. Initialization: The optimal lifted vectors of inputs cor-
∗
each such segment. For an operating condition beyond the responding to the leader agent, vm and all the follower
∗
range covered by a particular controller, a transition to a agents, vi are initialized to feasible values lying within the
different controller can be facilitated through a supervisory prescribed constraints.
switching logic. However, such details are a subject of future 2. Leader agent optimization: Following local optimization
work. problem is solved for currently known values of vi∗ :
Using the discretized, linear plant model (4), it is easy to 
∗ T
vm = argmin [vm Hm,m vm +( vi∗T Hm,i +fm
T
)vm ]
restructure the optimization problem of minimizing Jk to a {vm |Gm vm ≤wm } i
standard quadratic programming form given by (5).
3. Follower agent optimization: For each follower, i, the
x(k + 1) = Ax(k) + Bu(k) following local optimization problems are solved (in parallel)
∗
y(k + 1) = Cx(k + 1) (4) for currently known value of vm :

vk∗ = argmin (vkT Hk vk + fkT vk ) (5) vi∗ = argmin [viT Hi,i vi + ( vm∗T T
Hm,i + fiT )vi ]
{vk |Gk vk ≤wk } {vi |Gi vi ≤wi } i

Here, vk is defined as the following lifted vector of 4. Cooperative iteration: Steps 2 and 3 are repeated in
dimension N × Nu : sequence, until convergence.
[u1 (k|k)..u1 (k|k + N − 1)..uNu (k|k)..uNu (k|k + N − 1)]T . The distributed architecture for the example system of
Figure 1 has been shown in 3. An important remark must be

5114
 TABLE I
CHILLERS 1&2
P OWER C ONSUMPTION C OEFFICIENTS FOR C ONTROL I NPUTS
MASTER
AGENT
j Explanation of uj λj
1 Valve 3 opening factor 1.69×10− 3
2 Valve 4 opening factor 2.36×10− 3
3 Valve 5 opening factor 1.07×10− 3
4 Valve 6 opening factor 2.29×10− 4
Slave
agent 1
Slave
agent 2
Slave
agent 6
5 Valve 7 opening factor 5.76×10− 4
6 Valve 8 opening factor 1.80×10− 4
LAHX 1 LAHX 2 LAHX 6 7 Chiller 1 operating capacity 1/3.0
8 Chiller 2 operating capacity 1/2.0
Communication pathway

Fig. 3. Distributed MPC architecture for example system A simulation test-bed based for a nonlinear model of the
system, which was based on the principles described in [22],
110
was used to investigate the performance of the controllers.
Zonal cooling load (% of nominal)

105

100 B. Controller parameters

95
1) Centralized MPC: The dependence of the predicted
90
10 11 12 13 14 15 16 17 18
instantaneous power consumption, P (t) on the control inputs
Hour of the day
can be approximated by a linear combination, λj uj (t) for
Fig. 4. Discretized 8-hour load profile
near-nominal range of system operation. For the test system,
made about this algorithm. The computational complexity the values of the coefficients, λj have been presented in
per iteration is expected to be significantly low due to Table I. Note that the coefficients corresponding to the
the parallel optimization of small-scale problems in step chiller capacities are the inverse of their nominal COPs.
3. However, the use of large number of iterations can The coefficients corresponding to the valve opening factors
jeopardize the computational advantage of this scheme over represent their contribution to the total pump work done, and
centralized MPC. Therefore, it is recommended to terminate can be obtained by a linear hydraulic model of the system
the algorithm after relatively few iterations, which renders about the nominal conditions [7]. The weighting coefficients,
this scheme sub-optimal in practice. αj that appear in the first term of (2) are now decided by a
simple and intuitive order of magnitude analysis. We choose
V. RESULTS AND ANALYSIS αj = λj for the chiller inputs and αj = 103 λj for the valve
A. Test Case and underlying assumptions inputs. This is equivalent to saying that the pump power
The objective of this test case is to provide a basis for consumption is penalized 103 times more than the chiller
the relative performance evaluation of the various controllers power consumption which is reasonable as the total power
proposed above. The system under consideration is the consumption by the pumps is small compared to the power
chilled water loop subsystem (Fig. 1) that was introduced consumed by the chillers.
in section II. In this test case, the cooling load profile for The other parameters relevant to the MPC formulation
each of the six heat exchangers is assumed to follow the were intuitively chosen or tuned and are as follows:
general trend schematically shown in Fig. 2.14 in [20].
The nominal operating condition for the test system is one β = 5 × 10−2
where both the chillers are in operation - one at close to full γ=1
capacity, and the other at partial capacity. In practice, this ψ = 1 × 10−2
situation corresponds to peak or near peak cooling loads
N = 15 with sampling interval, Ts = 5s
during the day which generally occurs from 10 am to 6
pm. The nominal cooling loads are therefore thought to be
Note that the total prediction horizon (same as control
indicative of the mean loads during that 8 hour period, based
horizon) is 75s which is of the order of the slowest time
on which a discretized zonal load profile (Fig. 4) has been
constant of the system, and is therefore a sufficiently long
assumed. The loads in this profile have been normalized in
prediction window. Also, saturation constraints of ±25% of
terms of the nominal load with hourly sampling.
the nominal value and slew rate constraints of ±1 units per
Both the chillers in this system incorporate twin-screw
sample interval were forced on all input channels.
compressors, for which the performance curves at part load
2) Distributed MPC: Two distinct distributed architec-
conditions are assumed to be governed by Fig. 5(b) of [21].
tures were analyzed for the test system. The two chillers,
Therefore, at the specified ambient temperature of 35 deg
taken together, represent the leader in both these variants.
C, the performance dependence of these chillers on their
However, the first architecture has two followers corre-
operating capacity can be modeled by a linear relationship
sponding to two clusters of heat exchangers: LAHXs 1-
(7), where Q̇ is the capacity (in kW) at which the chiller is
3 and LAHXs 4-6. In the second architecture, each heat
running, and a and b are constant coefficients.
exchanger is treated as an independent follower. The number
COP = a + bQ̇ (7) of iterations used, Niter was varied from 2 to 6.

5115
 C. Simulation Results

Chilled water setpoint temperature (deg C)

10

8
The above controllers - localized on-off (L-OF), decen-
7
tralized PI (D-PI), centralized MPC (C-MPC) and distributed
6
MPC schemes 1 and 2 (D-MPC1 , D-MPC2 ) with Niter = 3
5 were implemented on the simulation test bed for the test
4
10 11 12 13 14 15 16 17 18
case described in section V.A. Based on the results obtained,
Hour of the day
the various control schemes were evaluated with respect to
Fig. 5. CWTR strategy for traditional local feedback control
the control objectives outlined in section II.C. The important
43
findings from this exercise have been reported below.
1) Demand matching: Fig. 6 shows the cooling capacity
Cooling capacity achieved (kW)

42

41
achieved by LAHX 6 under the action of L-OF, D-PI, C-
MPC and D-MPC2 for the step change in the reference at
40 noon (see Figure 4). Similar observations were made for the
REFERENCE
39
L-OF five other heat exchangers in the system. The behvaior for
D-PI
C-MPC
D-MPC1 was observed to be almost identical to D-MPC2
38 D-MPC 2 and therefore is not shown. It is evident that satisfactory
steady state regulation was achieved for all these schemes
37
11.9 11.95 12 12.05
Hour of the day
12.1 12.15 with tight error bounds. The transient characteristics exhibit
Fig. 6. Demand response comparison (LAHX 6) differences, but transient behavior of the response is of little
significance for such systems.
3) Localized On-off: The allowable error window to de- 2) Valve loading: The DFT of the control signals acting
termine if the corresponding control input should be on or on valve 6, during the time window around noon have been
off was set to ± 0.1. Furthermore, heuristically decided rate plotted in Fig. 7. Once again, the bahavior for D-MPC1 was
limits were applied on the actuators. observed to be very similar to D-MPC2 and therefore is not
4) Decentralized PI: The following controller gains were shown. It is evident that the critical frequencies for L-OF
arrived at, by tuning on the simulation test-bed, till satisfac- and D-PI schemes are around 2000 and 10 times higher,
tory response characteristics in terms of stability, overshoot respectively than C-MPC and D-MPC schemes, and hence
and settling time were achieved: adverse effects on the life-cycle performance of the system
Chiller control: P-gain = 30; I-gain = 1 can be expected through traditional control strategies. The
Valve control: P-gain = 7.5 × 102 ; I-gain = 4 possible reason for this observation is that localized feedback
The CWTR strategy used is shown in Fig. 5, which was is incapable of counteracting the effects of dynamical inter-
decided based on the the load profile of Fig. 4. Saturation actions (which act as a disturbance from a local perspective)
limits of ±25% were imposed on the actuator signals due to that occur between the various components of the system.
stability and robustness considerations. This highlights the importance of dynamic models in the
design of controllers for hydronic systems over the common
1200 practice of using static models.
L-OF
1000 D-PI 3) Energy consumption: Table II provides a compari-
Lomb normalized Periodogram

C-MPC
D-MPC2
son of the average energy consumption by the chillers,
800 pumps and the overall liquid loop subsystem for the various
600 High frequency content present
schemes. It is evident that the C-MPC scheme is the most
in on-off control signal
energy efficient, resulting in 7.1% and 5.8% reduction in
400 the chiller and pump power consumption respectively over
the traditional D-PI scheme, despite the fact that a modern
200
CWTR strategy was employed for the latter.
0
0 0.02 0.04 0.06 0.08 0.1
The distributed schemes are observed to be suboptimal.
Frequency (Hz) The chillers consume almost the same energy as in C-MPC,
Fig. 7. Control signal frequency content comparion (Valve 6) but the pump energy performances are significantly different.
7 This asserts that dynamic coupling among the sink elements,
even though small, has a strong bearing on the hydraulic
Time for one run of code (sec)

6 D-MPC (N = 5)
2 iter

5 D-MPC2 (Niter = 4) (pump) energy performance. However, the observation that

4
D-MPC2 (Niter = 3) D-MPC2 is less optimal than D-MPC1 with regard to pump
D-MPC2 (Niter = 2) energy consumption, reveals that the choice of the leader-
3 C-MPC
follower architecture for distributed control is important and
2
presents an interesting research problem where tools such
1 as combinatorial and cluster analysis can be applied. An
0
100 150 200 250
analysis of energy consumption for the L-OF scheme was
N.Nu

Fig. 8. Comparitive study of computational complexity

5116
 TABLE II
scheme exhibits significant potential in achieving the most
C OMPARISON OF AVERAGE POWER CONSUMPTION (kW )
ideal conformity to the control objectives. Future work shall
D-PI C-MPC D-MPC1 D-MPC2 focus on robustness analysis, extension of the proposed ideas
Overall subsystem 82.67 76.88 76.93 77.00 to the framework of whole-building optimization and further
Chillers 78.32 72.78 72.70 72.68
Pumps 4.34 4.09 4.23 4.31 theoretical development of the decentralized MPC scheme to
exploit its fullest potential.
not performed, because of the possibility of high prediction
errors considering their large operation regime. R EFERENCES
4) Computational complexity: Due to their trivial con- [1] Energy Information Administration, Annual Energy Outlook 2009.
trol logic and decentralized architectures, the D-PI and L- [Online]. Available: http://www.eia.doe.gov/oiaf/aeo
[2] G. Holness, “Sustaining Our Future By Rebuilding Our Past: Energy
OF schemes shall have negligible real-time computational Efficiency in Existing Buildings - Our Greatest Opportunity for a
complexities when implemented on microprocessors. In the Sustainable Future,” ASHRAE J, vol. 51, no. 8, pp. 16–21, Aug 2009.
predictive case, however, the computational effort required [3] S. Wang and Z. Ma, “Supervisory and optimal control of building
HVAC systems: A review,” HVAC&R Research, vol. 14, no. 1, pp.
for the solution of the optimization problem (2) is determined 3–32, Jan 2008.
by its dimension, i.e. N × Nu . Fig. 8 shows a computational [4] W. Jian and M. Zaheeruddin, “Sub-optimal on-off switching control
complexity comparison of C-MPC and D-MPC2 with differ- strategies for chilled water cooling systems with storage,” Applied
Thermal Engineering, vol. 18, no. 6, pp. 369–386, Jun 1998.
ent values of Niter , using a desktop computer with a 2.0 GHz [5] N. Nassif and S. Moujaes, “A cost-effective operating strategy to
processor and 960 MB RAM. In this case, to demonstrate reduce energy consumption in a HVAC system,” International Journal
the effect of variation in problem size, the value of N was of Energy Research, vol. 32, no. 6, pp. 543–558, May 2008.
[6] J. Rossiter, Model-based predictive control: a practical approach.
varied with Nu fixed. It follows that for large scale HVAC CRC, 2003.
systems, C-MPC can lead to significant time delays which [7] V. Chandan, G. Zak, and A. Alleyne, “Modeling of Complex Hydronic
can be avoided by the use of a D-MPC scheme with few Systems for Energy Efficient Operation,” in Proceedings of the 2009
ASME Dynamic Systems and Control Conference.
iterations. For the simulated system, convergence of both [8] M. Zaheer-Uddin and P. Monastiriakos, “Hydronic heating systems:
the D-MPC schemes was obtained at Niter = 4. Use of transient modelling, validation and load matching control,” Interna-
barrier function methods over active set methods is expected tional Journal of Energy Research, vol. 22, no. 1, 1998.
[9] Danfoss. Regulation for Hydronic Comfort Cooling Systems. [Online].
to result in even faster convergence. Higher computational Available: http://ogrevanje.danfoss.com/PCMPDF/VB36A102.pdf
and memory requirements for the C-MPC scheme results in [10] ASHRAE Handbook, HVAC Systems and Equipment, 2004.
costlier hardware, which may offset the cost benefit obtained [11] J. Piper, Operations and maintenance manual for energy management.
ME Sharpe, 1999.
by reduced energy consumption. [12] S. Yuan and R. Perez, “Multiple-zone ventilation and temperature
control of a single-duct VAV system using model predictive strategy,”
D. Discussion Energy & Buildings, vol. 38, no. 10, pp. 1248–1261, 2006.
[13] G. Huang, S. Wang, and X. Xu, “A robust model predictive control
In the light of the above results, the following general strategy for improving the control performance of air-conditioning
conclusions can been arrived at: systems,” Energy Conversion and Management, vol. 50, no. 10, pp.
2650–2658, 2009.
1. The on-off scheme is simple and easy to implement, but [14] G. Henze, D. Kalz, C. Felsmann, and G. Knabe, “Impact of forecasting
is disadvantageous from long term reliability considerations accuracy on predictive optimal control of active and passive building
due to significant fatigue loading of the components. thermal storage inventory,” HVAC & R Research, vol. 10, no. 2, pp.
153–178, 2004.
2. The Traditional PI schemes perform better than on- [15] D. Kolokotsa, A. Pouliezos, G. Stavrakakis, and C. Lazos, “Predictive
off, in terms of mechanical reliability but can consume control techniques for energy and indoor environmental quality man-
significantly higher energy in some cases when compared agement in buildings,” Building and Environment, vol. 44, no. 9, pp.
1850–1863, 2009.
to more advanced predictive strategies. [16] V. Chandan. Modeling and Control of Hydronic Building HVAC
3. The centralized MPC scheme is the most optimal in Systems. 2010 MS Thesis, Univ. of Illinois.
terms of energy comsumption, and yields ‘smoother’ actua- [17] R. Scattolini, “Architectures for distributed and hierarchical model
predictive control–A review,” Journal of Process Control, vol. 19,
tion signals but can be computationally very demanding. no. 5, pp. 723–731, 2009.
4. Distributed MPC strategies offer a compromise in terms [18] E. Camponogara, D. Jia, B. Krogh, and S. Talukdar, “Distributed
of energy consumption, reliability and computational effort. model predictive control,” IEEE Control Systems Magazine, vol. 22,
no. 1, pp. 44–52, 2002.
However, proper choice of the distributed architecture is very [19] C. Reynolds, “Flocks, herds and schools: A distributed behavioral
important to achieve the best tradeoff. model,” in Proceedings of the 14th annual conference on Computer
graphics and interactive techniques. ACM New York, NY, USA,
VI. CONCLUSIONS AND FUTURE WORK 1987, pp. 25–34.
[20] L. Jayamaha, Energy-Efficient Building Systems. McGraw-Hill Pro-
In this work, the control of hydronic building HVAC fessional, 2006.
systems was studied in detail. Both traditional (On-off and [21] F. Yu and K. Chan, “Optimum load sharing strategy for multiple-chiller
systems serving air-conditioned buildings,” Building and environment,
PI) and advanced (MPC) control schemes were developed vol. 42, no. 4, pp. 1581–1593, 2007.
and tested by means of simulation on an example sytem. [22] T. McKinley and A. Alleyne, “2008-01-0386 Real-Time Modeling of
The novel application of distributed MPC for such systems Liquid Cooling Networks in Vehicle Thermal Management Systems,”
SAE SP, vol. 2152, p. 1, 2008.
is the main feature of this work. The results indicated that
though both traditional and centralized MPC schemes have
relative advantages and disadvantages, the distributed MPC

5117