Modeling and Advanced Control

of HVAC Systems
Topic: HVAC Modeling & Control

Truong Nghiem
ESE, University of Pennsylvania
nghiem@seas.upenn.edu

January 26, 2011

Outline

I Part I: Modeling of HVAC Systems

I Part II: Advanced Control of HVAC Systems

T. Nghiem HVAC Modeling & Control 2

 Part I

Modeling of HVAC Systems

Fundamentals
Zone Model

T. Nghiem HVAC Modeling & Control 3

HVAC Modeling: Overview

supply air heat gain

set-point reheat
Thermostat VAV Zone
damper

zone temperature
Sensor

I Mathematical model of the plant (Zone block).

I HVAC system: exact models are complex (nonlinear, PDE,
stochastic, etc.).
I Focus: simplified (linearized) first-principles models derived from
heat transfer and thermodynamics theories.
I Other types of models: regression models, neural networks, look-up
tables, etc.

T. Nghiem HVAC Modeling & Control 4

HVAC Modeling: Fundamental Equation

First Law of Thermodynamics

(Conservation of Energy)

Heat balance equation: H W = E

Heat H Energy input to the system.
Work W Energy extracted from the system.
Internal heat E Energy stored in the system
(can only measure/calculate its change).

Heat input Zone Heat extracted

(Supply air, radiation, (Conduction, in-

internal heat gain, etc.) (Zone air) filtration, etc.)

T. Nghiem HVAC Modeling & Control 5

Heat Transfer: Concepts

I Heat Q: energy transferred across system boundary by temperature

difference (J).
I Heat flow (rate) Q: heat transfer rate (W).
I Heat flux: heat flow rate through a surface. Heat flux density is
heat flux per unit area (W/m2 ).
I Heat capacity C : heat needed to raise temperature of a body mass
by 1 C (J/K). Also called thermal mass, thermal capacitance.
I Specific heat (capacity) Cp : heat needed to raise temperature of
1 kg of material by 1 C (J/kg K); C = mCp = VCp .
I Energy change by temperature change E = VCp T .
I Mass flow rate m (kg/s) and volume flow rate V (m2 /s);
m = V .

T. Nghiem HVAC Modeling & Control 6

Heat Transfer: Mass Transfer

Heating

Supply air temperature Ts , return air temperature Tr < Ts , volume flow

rate V . Heat transfer to the zone is:

Q = H = V Cp (Ts Tr ) ( W)

Cooling

Similarly, with Ts < Tr , heat extracted from the zone is:

Q = W = V Cp (Tr Ts ) ( W)

T. Nghiem HVAC Modeling & Control 7

Heat Transfer: Conduction

Conduction is the process of heat transfer through a substance such as a

wall, from higher to lower temperature.

Fouriers equation (3-dimensional PDE with time):

 2 
dT T 2T 2T
Cp =k + +
dt x 2 y 2 z 2

where k: thermal conductivity ( W/mK).

Simplified equation (timeless, one-dimensional):

T Th Tl
Q = kA = kA
x l
where A: cross-sectional area ( m2 ), Th : high temperature, Tl : low
temperature, l: thickness/length of material.

T. Nghiem HVAC Modeling & Control 8

Heat Transfer: Conduction

1
Define Rth = kA (thermal resistance) then

QRth = Th Tl

Equivalent to an electric circuit: T = potential, T = voltage, Q =

current, Rth = resistance.

Th Q Rth Tl

T. Nghiem HVAC Modeling & Control 9

Heat Transfer: Convection

Convection is the heat transfer between a surface and fluid/gas by the

movement of the fluid/gas.

I Natural convection: heat transfer from a radiator to room air.

I Forced convection: from a heat exchanger to fluid being pumped
through.

Newtons law of cooling: Q = hAT

where h: heat transfer coefficient ( W/ m2 K2 ); A: surface area ( m2 ),
T : temperature difference between surface and fluid.
1
Define Rcv = hA and write QRcv = T .

Q Rcv
Tsurf Tfluid

T. Nghiem HVAC Modeling & Control 10

Heat Transfer: Radiation

Radiation is the heat transfer through space by electromagnetic waves.

Example: radiation between a radiator and a wall that faces it.

Fourth-order equation given by the Stefan-Boltzman law (cf. heat

transfer textbooks).

Approximate linearized equation:

Q = hr A(T1 T2 )

where : emissivity of the surface (0.9 for most building materials); hr :

radiation heat transfer coefficient ( W/ m2 K2 ).
1
Define Rr = hr A and write QRr = T .

T1 Q Rr T2

T. Nghiem HVAC Modeling & Control 11

Heat Transfer: Solar Radiation

Solar radiation is the radiation heat transfer by sun light.

I Direct radiation to the walls, furnitures, etc. in the room.

I Then heat transfer from walls, furnitures, etc. to room air.
I No direct heat transfer to room air but indirectly through walls,
furnitures, etc. large time lag.

T. Nghiem HVAC Modeling & Control 12

 Part I

Modeling of HVAC Systems

Fundamentals
Zone Model

T. Nghiem HVAC Modeling & Control 13

Zone Temperature Model
T33 T29 T1
Window 1

Rwindow,1 T21 T

Door 1

Door 2
T13

Window 2
T17
Room 1 Room 2
T6
T9

Door 3 T19

Door 4
T15

Window 4
T23 T

Room 3 Room 4
T35 T31 T3

Outside Source: [Deng et al., 2010]

(a)
I Ignore latent load (humidity),
Fig. 1. (a)only
Layoutsensible
of a 4-roomload (temperature).
building and (b) the RC-network representation o
I No infiltration.
I Simplified model
on the with
thermal simplified heat transfer equations.
models. Remark 1 (Conservati
The outline of this paper is as follows. In Section II, matrix are all zeros, its
I HVAC system of model
the thermal VAV type.
is formally defined and its Markov chain its non-diagonal entrie
$
representation is presented. In Section III, the KL divergence V (t) = iN Ci Ti (t)
rate and the model reduction method from [9] is briefly thermal model (1) at tim
reviewed. In Section IV, the methodology is applied to
dV
reduce the thermal models. In Section V, two examples are (t) =
T. Nghiem HVAC Modeling & Control dt 14
Body Mass Temperature

At what point is room air temperature (or wall surface temperature)

measured?

I Exact temperature distribution in a body mass is complex (PDE).

I Simplification: mean temperature of all points.
I How to measure mean temperature? Sensor placement.
I Mean temperature 6= temperature that occupants feel.

T. Nghiem HVAC Modeling & Control 15

RC Network of a Wall

Model heat transfer processes using resistance-capacitance equivalent

models (RC network).

Zone wall surface model

Rr

T Ts T R Ts
reduces to

Rcv

where T : zone air temperature, Ts : wall surface temperature, Rr :

radiation resistance, Rcv : convection resistance, and R1 = R1r + R1cv .

T. Nghiem HVAC Modeling & Control 16

RC Network of a Wall

Model heat transfer processes using resistance-capacitance equivalent

models (RC network).

T1 R1 Ts1 Rc1 Tw 1 Rc2 Tw 2 Twn Rc(n+1) Ts2 R2 T2

C1 Cw 1 Cw 2 Cwn C2

Zone 1 Wall Zone 2

More accurate model of conduction with large n.

Usually use n = 2 or simplify to a single thermal resistance between Ts1
and Ts2 (n = 0).

T. Nghiem HVAC Modeling & Control 17

 RC Network for Four Rooms

RC network of conduction, convection and radiation between rooms and

outside air.
T33 T29 T1 T2 T30 T34

T5 T7

Rwindow,1 T21 T25 T26 T22 Rwindow,2

T13 T14
Window 2

T17 T18
T37
T6 T8
T37
T9 T12

T19 T20

T15 T16
Rwindow,4
Window 4

T23 T27 T28 T24

T10 T11

T35 T31 T3 T4 T32 T36

T37

Source: [Deng et al., 2010]

(b)
m building and (b) the RC-network representation of the same building.
T. Nghiem HVAC Modeling & Control 18
Heating/Cooling and Other Gains

Add heating/cooling, internal gain and solar radiation to the network.

Qsa1
T1 R1 Ts1 Rc1 Tw 1 Rc2 Tw 2 Rc3 Ts2 = Toa

Qi1
Outside
C1 Qr 1 Cw 1 Cw 2 Air

Zone Wall

Qsa1 : HVAC heat flow; Qi1 : internal heat gain; Qr 1 : radiation heat gain

Qsa1 = Vsa1 Cp (Tsa T1 ) C1 dT

dt = Qsa + Qi
1 1
R1 (T1 Ts1 )
Qi1 , Qr 1 : disturbance/prediction 0 = Qr + R11 (T1 Ts1 ) R1c1 (Ts1 Tw 1 )

T. Nghiem HVAC Modeling & Control 19

State-space Thermal Model of Zone

Define variables:

I State variables x: all temperature variables.

I Disturbance variables w : internal gain, solar radiation, outside air
temperature, etc.
I Input variables u: defined by application.
I Supply air flow rate u1 = Vsa1 : Qsa1 = u1 Cp (Tsa x1 )
I Blind control u1b [0, 1]: Qr 1 = u1b w1
I Output variables y : e.g., y are all zone air temperatures.
I Parameters: capacitances and resistances.

T. Nghiem HVAC Modeling & Control 20

State-space Thermal Model of Zone

Gather all RC networks and all differential/algebraic equations:

d
x(t) = Ax(t) + Bu(t) + Kw (t) + (Lx x(t) + Lw w (t)) u(t)
dt

Discretize the state-space model:

 
x(k + 1) = Ax(k) + Bu(k) + K w (k) + Lx x(k) + Lw w (k) u(k)
y (k) = Cx(k)

Linearize the model at some operating point:

x(k + 1) = Ax(k) + Bu(k) + K w (k)

y (k) = Cx(k)

Model reduction techniques to reduce the dimension of the model.

T. Nghiem HVAC Modeling & Control 21
 Part II

Advanced Control of HVAC Systems

Overview
Introduction to Model Predictive Control
Model Predictive Control of HVAC Systems

T. Nghiem HVAC Modeling & Control 22

Advanced Control of HVAC Systems

In this lecture, advanced control = optimal supervisory control of HVAC

system to minimize some objective function (e.g., energy consumption,
energy cost).

General optimization problem:

minimize J = f (x0...N , u0...N , w0...N )

u0...N

subject to xk+1 = g (xk , uk , wk )

constraints on xk , uk
wk disturbance model

T. Nghiem HVAC Modeling & Control 23

Ingredients of Optimal Control

System model g

Mathematical model of the HVAC system (Part 1).

Disturbance model of wk

I Constrained in a bounded set wk Wk .

I Stochastic model, e.g., Toa N (Toa , 2 ) where Toa : predicted
outside air temperature (weather forecast).

T. Nghiem HVAC Modeling & Control 24

Ingredients of Optimal Control

Constraints

I Safety and mechanical constraints: uk Uk .

I Air quality: Vsa Vsa,min .
I Thermal comfort:
I Predicted Mean Vote (PMV) index: predicts mean of thermal
comfort responses by occupants, on the scale: +3 (hot), +2 (warm),
+1 (slightly warm), 0 (neutral), 1 (slightly cool), 2 (cool), 3
(cold). PMV should be close to 0.
I Predicted Percentage Dissatisfied (PPD) index: predicted percentage
of dissatisfied people. PMV and PPD has a nonlinear relation (in
perfect condition PPD(PMV = 0) = 5%).
I PMV/PPD can be calculated as nonlinear functions of temperature,
humidity, pressure, air velocity, etc. (cf. ASHRAE manuals).
I Constraint on PMV/PPD gives (nonlinear) constraint on xk .
I Simplified as xk Xk (convex).

T. Nghiem HVAC Modeling & Control 25

Ingredients of Optimal Control

Objective function f

I Minimize energy consumption: often a linear function of control

variables u.
I Minimize peak demand: a minimax optimization problem where

f (x0...N , u0...N , w0...N ) = max c T uk

kP

I Minimize energy cost: weighted sum of energy consumption and

peak demand.
I Maximize thermal comfort by minimizing PMV squared.

T. Nghiem HVAC Modeling & Control 26

HVAC Optimal Supervisory Control Approaches

Two approaches to optimal supervisory control of HVAC systems:

1. Optimal controller controls and coordinates all field devices (valves,

dampers, etc.); conventional local control loops are replaced.
I Is a complete re-implementation of the control system.
I Requires high computational power because of the time scale of field
devices.
I Good optimization but costly implementation.
2. Optimal controller sets set-points and modes of local control loops.
I Adds optimization software to supervisory control layer; keep local
control loops.
I Requires less computational power because of slower time scale.
I Less expensive implementation, (slightly worse) optimization.

T. Nghiem HVAC Modeling & Control 27

Difficulties of HVAC Optimal Supervisory Control

Theoretical difficulties:

I Large-scale system: hundreds of variables times dozens time steps.

I Stochastic nature of the system due to weather, occupancy, etc.
I Long optimization horizon (e.g., billing period).
I Non-linearity of system linearization.
I Complex objective function, e.g., energy cost.

T. Nghiem HVAC Modeling & Control 28

 Part II

Advanced Control of HVAC Systems

Overview
Introduction to Model Predictive Control
Model Predictive Control of HVAC Systems

T. Nghiem HVAC Modeling & Control 29

Introduction to Model Predictive Control

Source:

Model Predictive Control of Hybrid Systems

Presentation by Prof. Alberto Bemporad

T. Nghiem HVAC Modeling & Control 30

 Part II

Advanced Control of HVAC Systems

Overview
Introduction to Model Predictive Control
Model Predictive Control of HVAC Systems

T. Nghiem HVAC Modeling & Control 31

MPC of HVAC Systems: Formulation

minimize J = f (xt , . . . , xt+T , ut , . . . , ut+T 1 )

ut ,...,ut+T 1

subject to xk+1 = g (xk , uk , wk ), k = t, . . . , t + T 1

xk Xk , uk Uk , k = t, . . . , t + T
wk disturbance model

where

I Horizon T  N.
I Objective function f ():
X
c T uk (energy consumption)
k=t,...,t+T 1
 
max Dt , max c T uk (peak demand)
kP{t,...,t+T 1}

with Dt : peak demand from time 0 to t.

T. Nghiem HVAC Modeling & Control 32
MPC of HVAC Systems: Formulation

minimize J = f (xt , . . . , xt+T , ut , . . . , ut+T 1 )

ut ,...,ut+T 1

subject to xk+1 = g (xk , uk , wk ), k = t, . . . , t + T 1

xk Xk , uk Uk , k = t, . . . , t + T
wk disturbance model

where

I Disturbance wk :
I Bounded constraint wk Wk robust MPC.
I Probabilistic model wk Pk stochastic MPC.
I Optional demand-limiting constraint: for every k P

c T uk D ?

with D ? : maximum demand.

T. Nghiem HVAC Modeling & Control 33

MPC of HVAC Systems: Optimization

Solving the MPC optimization problem:

I If it is convex, use standard algorithms and tools.

I If it is non-convex, use approximation technique (Lagrangian dual
problem).

Some practical considerations:

I Large-scale system distributed optimization.

I For better performance:
I Initialize with previous solution.
I Use heuristics to guess initial solution.
I Further constrain variables using rules, e.g., during peak hours, VAV
box should close more.

T. Nghiem HVAC Modeling & Control 34

B isthatlarge.
SMPC This investigation
has smaller is violations than
and less frequent
RBC. Furthermore,
large-scale the diurnal
factorial temperature variations
simulation are
B. Practical Potential Analysis
MPC
ge ofbehavior
of
much smaller HVAC
with Systems:
SMPC, which
casesforrepresenting
Examples
is a much more favourable
different
the room comfort. With real weather predictions, it can happen that
t weather conditions as described traints are being Fig. 7. violated. Therefore,
Tradeoff for SMPC and CE betweencontroller perfor
energy use and violation and
ails see [7], [8]. comparison with PB and RBC.
is assessed in terms of both energy usage and con
this investigation we compare violation.
RBC and MPC strategies only and which low/high-cost energy sources are needed to keep
the room temperature in the required comfort levels. SMPC
m the theoretical potential study. was shown to outperform both rule-based control (RBC) as
found in [12]. well as a predictive non-stochastic controller (CE). Further
benefits or SMPC are easy tunability with a single tuning
Time step [h] parameter describing the level of constraint violation as wel
Fig. 5. Room temperature profile of RBC for one year.
as comparatively small diurnal temperature variations.
nt there were on total 1228 cases VII. ACKNOWLEDGMENTS
ere done for the HVAC system, Swisselectric Research, CCEM-CH and Siemens Building
requirements, and the weather Technologies are gratefully acknowledged for their financia
support of the OptiControl project.level
Typical violation
riants are listed here:
R EFERENCES
dered are five building system [1] A. Ben-Tal and A. Goryashko and E. Guslitzer and A. Nemirovski
A). Adjustable robust solutions of uncertain linear programs, in Mathe
matical Programming, vol. 99(2), 2004, pp. 351-376.
ary in building standard (Passive
Time step [h] Fig.
[2] P. J.4. Comparison
Goulart of and
and E. C. Kerrigan SMPC and RBC.Optimization
J. M. Maciejowski,
over state feedback policies for robust control with constraints
Fig. 6. Room temperature profile of SMPC for one year.
construction type (light/heavy), Automatica, vol. 42(4), 2006, pp. 523-533.
[3] R. E. Griffith and R. A. Steward, A nonlinear programming technique
(low/high),
Usage of CE internal gains
for the six level
selected Figure
cases generally yield 4 depicts the result
Source:
for the optimization of theprocessing
[Oldewurtel
of continuous et comparison
al., 2010] of oS
systems, Journal
much more violations than allowed in the standards (results Management Science, vol. 7, 1961, pp. 379-392.
ances; low/high; also associated
not shown). One can however tune CE by assuming a
and RBC for the
[4] M. selected
Gwerder, set of
J. Toedtli, Predictive experiments:
control for integrated roomMP auto
mation, CLIMA 2005, Lausanne, 2005.
de orientation
tighter comfort(N bandorforSthe
forcontroller,
normalwhich always clearly
results in less [5] less energy
M. Gwerder, J. Toedtli,use than Ruled-based
D. Gyalistras, RBC and instrategies
control four
in [6].
WT.for corner
violations offices).
and more casesHVAC
energy use. Thus, for different smaller
comfort
[6]amounts
D. Gyalistras, of violations.
M. Gwerder (Eds.), Use This indicates
of weather and occupancyth
band widths one gets a tradeoff curve between energy use and
Nghiem Modeling & Control 35
References

Deng, Kun, Barooah, Prabir, Mehta, Prashant G., & Meyn, Sean P. 2010.
Building thermal model reduction via aggregation of states.
Pages 51185123 of: Proceedings of the 2010 American Control Conference, ACC 2010.
Oldewurtel, Frauke, Parisio, Alessandra, Jones, Colin N., Morari, Manfred, Gyalistras,
Dimitrios, Gwerder, Markus, Stauch, Vanessa, Lehmann, Beat, & Wirth, Katharina. 2010
(Jun.).
Energy efficient building climate control using Stochastic Model Predictive Control and
weather predictions.
Pages 5100 5105 of: American Control Conference (ACC) 2010.

T. Nghiem HVAC Modeling & Control 36

Thank You!

Q&A