46 views

Uploaded by hamid_ouali_1

save

You are on page 1of 6

**applied to multisource multizone building temperature regulation
**

Petru-Daniel Moros¸an, Romain Bourdais, Didier Dumur, Jean Buisson

Abstract—This paper presents a distributed model predic-

tive control algorithm based on Benders’ decomposition for

temperature regulation in buildings. It is well known that

the main objective of this control problem is to minimize

the heating (cooling) energy bills while maintaining a certain

indoor thermal comfort. In order to reduce the energy costs,

many buildings are equipped with several heating sources with

different dynamics, gains and energy prices, an example is the

use of a hot water based central heating and local electric

convectors as a complementary heating source. Using a linear

system model of the controlled process, the MPC minimization

problem can be solved by linear programming. The Benders’

decomposition exploits a particular structure (block-angular) of

the constraint matrix and distributes the computational demand

among local controllers.

I. INTRODUCTION

Nowadays the optimization of the energy consumption is

a world target and it is no longer feasible to design a system

without considering it. Within the services and households

sector, the major energy consumer is the heating/cooling

system. Even if the new trends are to construct buildings

to accomplish new environmental standards, the problem

remains unsolved for buildings where thermal insulation

works are difﬁcult to be implemented. A solution to reduce

the energy bills is the use of multiple heat sources. One

of the most common example is the use of a central gas

or oil heating system as the base heating source and local

electric convectors as complementary sources, in order to

take advantage of the lower price of the kWh of the fossil

fuels and in the mean time to use the faster dynamics of

the electrical heating systems to ensure good control per-

formances. To efﬁciently control these two types of heating

sources an optimal control law is required. This is the context

in which the presented work takes place.

Even if many studies were performed in order to optimize

the energy efﬁciency of heating systems, the controllers that

are mostly used today remain the on/off and PID. To ensure

a proper regulation, auto-tuning methods of PID parameters

have been proposed in [1]. Other approaches can be found

in the literature like fuzzy logic [2], neural networks [3] or

genetic algorithms [4]. To the author’s knowledge, studies

on control strategies for multiple heating sources are almost

inexistent.

Petru-Daniel Moros¸an, Romain Bourdais and Jean Buisson are

with IETR-SUPELEC, Avenue de la Boulaie - B.P. 81127, F-

35511 Cesson-S´ evign´ e Cedex, France {petru-daniel.morosan,

romain.bourdais, jean.buisson}@supelec.fr

Didier Dumur is with SUPELEC Systems Sciences (E3S), Au-

tomatic Control Department, 91190 Gif-sur-Yvette Cedex, France

didier.dumur@supelec.fr

During the last two decades a growing interest has been

granted to model predictive control (MPC) due to its ability

to handle constraints in an optimal control environment. In

MPC, the control input is calculated by solving an optimal

control problem (minimization of a cost function) over a

given horizon. Only the ﬁrst element of the open-loop

command sequence is applied to the system. At the next

instant, a new optimization is performed based on current

measurements. The predictive control has been successfully

used in many and varied applications [5], [6]. In particular,

for heating and cooling systems, different formulations of

cost functions and constraints have been analyzed in [7] to

minimize the consumption or to guarantee a desired comfort

level. A stochastic predictive control approach is proposed

in [8] handling with chance constraints.

Especially in the case of multiple heating sources, the

optimization objective should be the minimization of the

price paid for the indoor heating to favor the cheapest heating

source. The anticipative effect of the predictive control can

be efﬁciently used when the future reference proﬁle is

known. Most of the rooms/ofﬁces have an intermittent use

(composed of occupation and inoccupation periods), which

in most cases can be well known in advance. Deﬁning

the thermal comfort as optimization constraints implies that

they will have a dynamical evolution due to the fact that

the comfort should be deﬁned only during the occupation

periods. When dealing with relatively slow dynamics, as

thermal systems, the anticipative effect of the MPC can

be a main advantage over the common controllers, because

knowing the occupation periods in advance it allows to turn

on the heating system at the optimal moment (which can vary

between some minutes and several hours before the start of

the occupation period) in order to accomplish the comfort

constraints from the beginning of the occupation period.

A fully centralized MPC for large-scale systems (build-

ings with many rooms) is often undesirable and difﬁcult

to implement. The dimension of the optimization problem

grows with the number of subsystems composing the global

process and for a large number of subsystems (rooms), the

computational demand explodes. Another drawback of the

centralized strategies is their poor ﬂexibility and reliability.

Alternatively, large-scale control problems can be solved

by several controllers which take care of the local control

parameters. In order to improve their control performances,

the local controllers communicate over so-called negotiation

iterations, leading to a distributed control strategy. In this

paper we propose a distributed MPC architecture, based on

Benders’ decomposition technique, controlling a central heat-

ing system and several local heating sources in a multizone

building.

The paper is organized as follows. Section II introduces

the control model and formulates the MPC optimization

problem, including the energy prices, the occupation proﬁle

and the temperature comfort bounds. In Section III, we

present the Benders’ decomposition used to obtain a dis-

tributed predictive control strategy, synthesized in an iterative

algorithm. The efﬁciency of the method is illustrated by

simulation results is Section IV. Conclusions and future

directions are proposed in Section V.

II. PROBLEM FORMULATION

The anticipative effect of MPC consists in using a model

of the process in order to predict its behavior during a ﬁnite

horizon. A linear discrete time state space representation of

a zone model i ∈ S = {1, 2, ..., s} can be stated as:

_

_

_

xxx

i

(k + 1) = AAA

i

xxx

i

(k) +

_

BBB

i1

BBB

i2

¸

·

_

u

i

(k)

u

c

(k)

_

y

i

(k) = CCC

i

xxx

i

(k)

(1)

where the vector xxx

i

∈ R

n

i

is the local state, u

i

, u

c

, y

i

∈ R

are the local and the shared input (electrical heating power

and power input of the boiler) and the output (measured

room temperature), respectively. Using the model (1), we

can write the prediction equation for a control horizon N

u

and a prediction horizon N

2

as:

ˆ yyy

i

(k) =

_

ˆ y

i

(k + 1|k) · · · ˆ y

i

(k + N

2

|k)

¸

T

= ΨΨΨ

i

xxx

i

(k) +ΦΦΦ

i1

uuu

i

(k) +ΦΦΦ

i2

uuu

c

(k),

(2)

where the matrices ΨΨΨ

i

, ΦΦΦ

i1

and ΦΦΦ

i2

are deﬁned in Appendix.

The control objective is to minimize the energy bill due

to the indoor heating, which is usually a linear function

of the consumed energy (our control inputs). The thermal

comfort (deﬁned here by an upper w

i

(k + j) and a lower

w

i

(k+j) temperature bounds) and the physical limitations of

the process are the constraints of our optimization problem.

As mentioned before, the thermal comfort should be deﬁned

only during the occupation periods (see Fig. 1). So denote

δδδ

i

(k) =

_

δ

i

(k + 1) · · · δ

i

(k + N

2

)

¸

T

as the future oc-

cupation proﬁle over the prediction horizon for the room i.

Intuitively, each element of this vector is deﬁned as:

δ

i

(k + j) =

_

1, k + j ∈ Occupation

i

0, k + j ∈ Inoccupation

i

.

(3)

Now we are able to formulate the linear programming

problem corresponding to our control objectives as:

min

uuu

c

(k),uuu

i

(k),∀i∈S

J(k) =

i∈S

ccc

T

i

(k)uuu

i

(k) +ccc

T

c

(k)uuu

c

(k) (4)

subject to

µ

low

i

(k + j) = δ

i

(k + j)(w

i

(k + j) − ˆ y

i

(k + j|k)) ≤ 0

µ

up

i

(k + j) = δ

i

(k + j)(ˆ y

i

(k + j|k) −w

i

(k + j)) ≤ 0

∀i ∈ S, ∀j = 1..N

2

(5)

O

c

c

u

p

a

t

i

o

n

(

δ

i

)

T

e

m

p

.

[

o

C

]

Time

Time

k

N

2

15

20

22.5

1

0

w

i

w

i

ˆ y

i

y

i

Fig. 1. Occupation proﬁle illustration

0 ≤ uuu

i

≤ uuu

i

(k) ≤ uuu

i

, ∀i ∈ S (6)

0 ≤ uuu

c

≤ uuu

c

(k) ≤ uuu

c

, (7)

with

uuu

i

(k) =

_

u

i

(k) · · · u

i

(k + N

u

−1)

¸

T

, (8)

uuu

c

(k) =

_

u

c

(k) · · · u

c

(k + N

u

−1)

¸

T

, (9)

ccc

i

(k) = [ c

i

(k) ··· c

i

(k+N

u

−2)

N

2

−1

j=N

u

−1

c

i

(k+j) ]

T

, (10)

ccc

c

(k) = [ c

c

(k) ··· c

c

(k+N

u

−2)

N

2

−1

j=N

u

−1

c

c

(k+j) ]

T

. (11)

where c

i

(k) and c

c

(k) are the prices for local and shared

inputs, respectively, at time step k. The sum of the last

element of the cost sequences is due to the fact that the

control input is considered constant out of the control horizon

and equal to the value corresponding to the time step k +

N

u

−1.

According to (3), the comfort constraints (5) are activated

only during the occupation periods of the zones, allowing to

reduce the energy consumption without affecting the comfort

of the occupants. Replacing (2) in (5) and rewriting in a

compact form we obtain

∆∆∆

i

(ΦΦΦ

i1

uuu

i

(k) +ΦΦΦ

i2

uuu

c

(k)) ≥ ∆∆∆

i

(www

i

(k) −ΨΨΨ

i

xxx

i

(k)) ,

∆∆∆

i

(ΦΦΦ

i1

uuu

i

(k) +ΦΦΦ

i2

uuu

c

(k)) ≤ ∆∆∆

i

(www

i

(k) −ΨΨΨ

i

xxx

i

(k)) ,

(12)

where matrix ∆∆∆

i

is obtained by eliminating the zero lines of

the matrix ∆∆∆

′

i

= diag(δδδ

i

(k)).

A. Relaxing the comfort constraints

In the linear programming problem (4-7) the thermal

comfort is considered as hard constraints. If, at a certain time

step, one of these constraints cannot be satisﬁed the global

optimization problem becomes unfeasible. In practice, to

avoid this issue the comfort constraints are softened, adding

a penalty to the cost criterion if they are not satisﬁed. The

minimization problem is:

min

uuu

c

(k),uuu

i

(k),∀i∈S

J

f

(k) = J(k) +

i∈S

N

2

j=1

f

i

(k + j) (13)

subject to (6) and (7), where the comfort penalty function f

i

is deﬁned as:

f

i

(k + j) =

_

¸

_

¸

_

0, µ

up

i

(k + j) ≤ 0 and µ

low

i

(k + j) ≤ 0

λ

i

µ

up

i

(k + j), µ

up

i

(k + j) > 0

λ

i

µ

low

i

(k + j), µ

low

i

(k + j) > 0.

(14)

The always feasible linear programming problem (13)

subject to (6) and (7) can be written in a standard form,

dropping k for the simplicity of notations, as:

min

uuu

′

c

,uuu

′

i

,∀i∈S

ccc

′T

c

uuu

′

c

+ ccc

′T

1

uuu

′

1

· · · + ccc

′T

s

uuu

′

s

subject to DDDuuu

′

c

= ggg

EEE

1

uuu

′

c

+ FFF

1

uuu

′

1

= hhh

1

.

.

.

.

.

.

.

.

.

EEE

s

uuu

′

c

+ FFF

s

uuu

′

s

= hhh

s

uuu

′

c

, uuu

′

1

· · · , uuu

′

s

≥ 000

(15)

where the matrices and the vectors composing this formula-

tion are deﬁned in Appendix. The optimization variables uuu

′

c

and uuu

′

i

are obtained by adding the required number of slack

variables to the uuu

c

and uuu

i

, respectively. One of the most com-

mon solver for linear programming problems is the Simplex

algorithm. To distribute the computational demand among

local controllers, we will use the special block structure of

the constraint matrix. This block-angular structure can be

efﬁciently exploited by the Benders’ decomposition, which

will be detailed in the next section.

III. BENDERS’ DECOMPOSITION

Bender’s decomposition, also known as the dual of

Dantzig-Wolfe decomposition [9], uses the block-angular

structure of the constraint matrix in order to parallelize the

computation of a linear optimization problem. This decom-

position method splits a single large-scale linear program-

ming problem into several independent problems which are

coordinated by a single master problem (MP). The optimal

solution of the original large-scale problem can be shown to

be identical to the solution obtained after a ﬁnite number of

iterations, solving sequentially the MP and the subproblems

[10].

Even if any linear programming problem can be solved

applying this decomposition technique, the method is re-

commended for structured linear programs. Consider a linear

optimization problem having a block-angular constraint ma-

trix structure in a standard form (15). Optimization variable

uuu

′

c

, also called complicating variable, prevents obtaining the

optimal solution by solving each subproblem independently.

But, for a ﬁxed value of this complicating variable, we

know that solving independently the subproblems leads to the

global optimum. This is the main idea of this decomposition,

where at each iteration l the master problem optimal solution

uuu

′l

c

tends to the optimal value uuu

′∗

c

. In order to write explicitly

the master problem and the subproblems, we will ﬁrstly

rewrite the original linear programming problem (15) as:

min

uuu

′

c

ccc

′T

c

uuu

′

c

+

s

i=1

z

i

(uuu

′

c

)

subject to DDDuuu

′

c

= ggg

uuu

′

c

≥ 000,

(16)

where

z

i

(uuu

′

c

) = min

uuu

′

i

ccc

′T

i

uuu

′

i

subject to FFF

i

uuu

′

i

= hhh

i

−EEE

1

uuu

′

c

uuu

′

i

≥ 000

, i ∈ S. (17)

We call (17) the subproblem i, once the complicating variable

uuu

′

c

has been chosen. Applying the duality, z

i

(uuu

′

c

) can also

be computed through the dual of (17), deﬁned as:

z

i

(uuu

′

c

) = max

ppp

i

ppp

T

i

(hhh

i

−EEE

i

uuu

′

c

)

subject to FFF

T

i

ppp

i

≤ ccc

′

i

, i ∈ S. (18)

The reason of using the dual subproblem is that the

polyhedron D

i

=

_

ppp

i

| FFF

T

i

ppp

i

≤ ccc

′

i

_

that deﬁnes the feasible

region of (18) is independent of uuu

′

c

. D

i

can also be deﬁned

(via the theorem of convex combination) by a ﬁnite number

of extreme points ppp

1

i

, ..., ppp

I

i

i

and a ﬁnite number of extreme

rays rrr

1

i

, ..., rrr

J

i

i

. The solution of (18) is an extreme point

if the dual problem is feasible and bounded or an extreme

ray if the dual is unbounded. The dual problem (18) cannot

be unfeasible, because it should imply that the primal is

unbounded, which is false by its deﬁnition. On the other

hand, using the relaxed comfort constraints, none of the

primal subproblems can be unfeasible. Consequently, using

an algorithm to solve exactly the dual subproblems (18), it

will return one of the extreme points, ppp

i

, of the feasible

region D

i

and the subproblem objective function will be:

z

i

(uuu

′

c

) = ppp

T

i

(hhh

i

−EEE

i

uuu

′

c

) = max

k=1,...,I

i

(ppp

k

i

)

T

(hhh

i

−EEE

i

uuu

′

c

). (19)

Now we are able to write the MP at iteration l, knowing the

solutions of all dual subproblems at every previous iteration

l

p

< l:

min

uuu

′

c

,z

ccc

′T

c

uuu

′

c

+ z

s.t. DDDuuu

′

c

= ggg

uuu

′

c

≥ 000

z ≥ 0

s

i=1

(ppp

l

p

i

)

T

EEE

i

uuu

′

c

+ z ≥

s

i=1

(ppp

l

p

i

)

T

hhh

i

, ∀l

p

= 1..l −1.

(20)

The Algorithm 1 that solves the original problem (15)

using the Benders’ decomposition is iterative. Each iteration

will have two main steps, consisting of solving the master

problem and then solving (in parallel) the subproblems using

the current value of the complicating variable. We add the

new constraint to the master problem (also called Benders

cut) and a new iteration can begin. Note that over the

iterations, the dimension of subproblems is constant while

the constraint matrix of MP increases after each iteration.

The algorithm reaches the optimal solution of (15) when

z

l

=

s

i=1

z

l

i

. In practice at each iteration l an upper and a

lower bound of the optimum is computed:

V AL

l

up

= ccc

′T

c

uuu

′l

c

+

s

i=1

ccc

′T

i

uuu

′l

i

, (21)

V AL

l

low

= ccc

′T

c

uuu

′l

c

+ z

l

, (22)

and the stop condition of the iterative algorithm will be

V AL

l

up

−V AL

l

low

< ǫ.

Algorithm 1 Distributed MPC version for multi-source

heating system based on Benders’ decomposition

Require: Global problem formulation (15), ǫ, xxx

i

, ∀i ∈ S

Ensure: Optimum control input sequences uuu

′∗

c

and uuu

′∗

i

, i ∈

S

1: Initialization: l = 1

2: Solve the master problem (20), obtaining uuu

′l

c

and z

l

3: Send uuu

′l

c

to all local controllers MPC

i

4: Solve (in parallel) the subproblems (in both primal (17)

and dual (18) forms), obtaining uuu

l

i

and ppp

l

i

5: Compute the current upper and lower bounds (21) and

(22)

6: if V AL

l

up

−V AL

l

low

≤ ǫ then

7: uuu

′∗

c

= uuu

′l

c

and uuu

′∗

i

= uuu

′l

i

, ∀i ∈ S, Stop

8: else

9: Update the constraints of the master problem (by

adding the new Benders cut)

10: Update current iteration l = l + 1 and Goto step 2

11: end if

The resulting control scheme has the structure presented in

ﬁgure 2. The central controller, MPC

c

, acts like a coordinator

for the local controllers, MPC

i

. It also tests the stop condition

of the algorithm after every iteration. The local predictive

controllers solve their subproblems once they have received

from the master the current value of the complicating vari-

able. If the algorithm does not offer a solution within a

sample time, the last iteration solution can be applied to the

system, as it is always feasible.

u

c

u

2

xxx

1

MPC

c

MPC

1

MPC

2

MPC

s

u

s

u

1

xxx

2

xxx

s

Fig. 2. Distributed MPC scheme using Benders’ decomposition

IV. SIMULATION RESULTS

A. Empirical study of efﬁciency

Since the distributed algorithm was implemented on a

sequential machine, the computational time required by the

decomposition algorithm to solve a linear programming

problem is:

t

seq

=

l

l=1

_

_

t

MPC

c

(l) +

j∈S

t

MPC

j

(l)

_

_

, (23)

where l is the number of Benders iterations needed to

solve the problem, t

MPC

c

(l) is the time required by the

central MPC to solve the master problem at iteration l, while

t

MPC

j

(l) is the computational time to solve the subproblem

j. The computational time using a distributed computing

environment, ignoring the communication time required at

each iteration, can be expressed as:

t

distr

=

l

l=1

_

t

MPC

c

(l) + max

j∈S

t

MPC

j

(l)

_

. (24)

In order to study the complexity of the distributed algo-

rithm compared to the centralized solver, we will use a very

simple thermal model of a room and supposing (without loss

of generality) that all the subsystems have the same model.

The values of the model (1) matrices are:

AAA

i

= [

0.9921 0

0 0.9931

] , BBB

i

= [

0.2595 0

0 0.1376

] , CCC

T

i

= [

1

1

] .

The efﬁciency of the distributed algorithm will be mea-

sured regarding the equivalent distributed computational time

and the number of iterations l. The main parameters of the

algorithm are: the number of subproblems (subsystems) s,

the dimension of the subproblems d

i

= (N

u

+ 2N

i

o

) ×

(2N

u

+ 4N

i

o

), where N

i

o

∈ {0, 1, ..., N

2

} is the number of

occupation time steps within the prediction period (i.e. the

number of lines of ∆∆∆

i

) and the tolerance ǫ. The dimension

of the prediction horizon N

2

should be chosen sufﬁciently

large in order to offer enough time to the heating system to

increase the indoor temperature up to the desired setpoint in

the worst situation (low initial temperatures). In the results

presented below, we ﬁxed N

2

= 30. For the three scenarios

presented below we considered ﬁve different cases (in each

case we changed the initial state of the subsystems) for

each value of the variable parameter, in order to have more

consistent statistical results.

1) Scenario 1: s ∈ {2

1

, 2

2

, ..., 2

7

}, N

u

= 5, N

i

o

= 15,

ǫ = 10

−3

. Fig. 3 shows a very good scaling behavior

of the algorithm in a distributed computing environment,

regarding the number of subproblems. In the mean time,

for a small number of subsystems the centralized (Simplex)

method offers better performances. Concerning the number

of iterations, we observe a logarithmic dependence of l on s.

This fact shows a good convergence speed of the algorithm

and its slight dependence on the number of subproblems.

10

0

10

1

10

2

10

−2

10

0

10

2

10

4

s

T

i

m

e

[

s

]

t

centr

t

distr

10

0

10

1

10

2

0

5

10

15

20

s

N

u

m

b

e

r

o

f

i

t

e

r

a

t

i

o

n

s

Fig. 3. The inﬂuence of the number of subsystems, s

2) Scenario 2: s = 50, N

u

∈ {1, 5, 10, ..., 30}, N

i

o

∈

{1, 5, 10, ..., 30}, ǫ = 10

−3

. Here we study the algorithm

performances with respect to the subproblem sizes, d

i

. The

control horizon dimension, N

u

, is a tuning parameter, while

N

i

o

depends on the occupation proﬁles. As Fig. 4 shows, N

i

o

has a more important inﬂuence over the computational time

than N

u

, which is normal as N

i

o

has a greater weight on the

dimension of the subproblem.

1 5 10 15 20 25 30

0

50

100

150

N

u

(N

o

i

=15) / N

o

i

(N

u

=15)

T

i

m

e

[

s

]

t

centr

(N

o

i

=const)

t

distr

(N

o

i

=const)

t

centr

(N

u

=const)

t

distr

(N

u

=const)

1 5 10 15 20 25 30

0

50

100

150

N

u

(N

o

i

=15) / N

o

i

(N

u

=15)

N

u

m

b

e

r

o

f

i

t

e

r

a

t

i

o

n

s

N

o

i

=const

N

u

=const

Fig. 4. The inﬂuence of the control horizon, N

u

, and of the occupation

time steps, N

i

o

3) Scenario 3: s = 50, N

u

= 10, N

i

o

= 15, ǫ ∈

{10

−6

, 10

−5

, ..., 10

−1

}. Fig. 5 shows that the tolerance ǫ

inﬂuences very slightly the computational demand of the

distributed algorithm. The number of Benders iterations has

a logarithmic dependence on the tolerance which shows the

exponential convergence of the method.

B. Simulation study

This second part of the experimental results will focus

on testing the proposed algorithm on a virtual building com-

posed of three rooms, as shown in Fig. 6, in order to compare

10

−6

10

−5

10

−4

10

−3

10

−2

10

−1

0

10

20

30

40

ε

T

i

m

e

[

s

]

t

centr

t

distr

10

−6

10

−5

10

−4

10

−3

10

−2

10

−1

10

15

20

25

ε

N

u

m

b

e

r

o

f

i

t

e

r

a

t

i

o

n

s

Fig. 5. The inﬂuence of the stop condition tolerance, ǫ

the performances of the proposed control strategy with the

commonly used PI controllers. The simulation of the thermal

building behavior was made using the MATLAB toolbox

called SIMBAD. All the rooms of the simulated building

have the same surface (12m

2

) and are heated by a hot

water-based underﬂoor heating system and local independent

convectors of 1200W maximal power. Each room has a

double glazed window of 2m2 surface on the larger external

wall of each zone. The external wall sandwich consists of

1cm of gypsum, 8cm of extruded polystyrene and 20cm of

concrete. The internal wall is 7.2cm thick of gypsum board.

The simulator supposes a well mixed indoor air.

The control inputs of the system are the electrical power

of the local convectors and the water temperature increment

(the difference between the water temperature which enters

the boiler and the outlet water temperature). Here we suppose

that the boiler is controlled by a fast regulator which receives

∆T as reference. This inner control loop is considered in

steady state in our simulations. The water transport delay

is not considered, being negligible compared to the system

time step T

s

= 600s. Neglecting the heat loss through the

hydraulic system and considering the efﬁciency of the boiler

η = 0.9 we can express the central and local costs in terms

of electrical c

e

and gas/oil c

g

prices in kWh as:

c

i

(k + j) = c

e

(k + j)

T

s

3600 · 1000

,

c

c

(k + j) = c

g

(k + j)

T

s

3600 · 1000η

C

pw

ρ

w

q

w

,

(25)

where C

pw

, ρ

w

and q

w

are the speciﬁc heat, the density

and the mass ﬂow of the water through the boiler. Table

I compares the heating energy costs for a single day. The

weather conditions are those measured at Rennes on the

1

st

January 1998. The occupation proﬁle and the thermal

comfort bounds are those shown in Fig. 7. The electricity

price considered is c

e

= 0.0742 e/kWh, while the gas

price is c

e

= 0.063 e/kWh. Note that, for the three control

strategies compared, the thermal comfort is similar.

TABLE I

HEATING COST COMPARISON

Control law Cost [e]

dMPC with Benders decomposition 6.35

local PI for convectors and open loop control for ∆T

1

7.84

local MPC for convectors and ∆T = 0 8.80

u

3

u

1

u

2

Zone1 Zone2

Zone3

y

1

y

2

y

3

6m

2m

4m

∆T

Fig. 6. Conﬁguration of the virtual building

V. CONCLUSION

A distributed model predictive control strategy has been

proposed for multisource multizone building temperature

regulation. In order to reduce the heating energy bill, the

MPC criterion minimizes the heating energy cost, subjected

to the comfort constraints and physical limitations of the

system, resulting in a linear programming problem. The cen-

tralized computational demand of the optimization problem

grows exponentially with the number of subsystems (rooms).

Then, for large-scale buildings, we proposed a distributed

control strategy based on Benders’ decomposition, which

allows the decrease of the computational demand by using

a network of local controllers, coordinated by a master

controller.

Future work will focus on the adaptation of the Benders’

decomposition algorithm in order to handle coupling sub-

systems (considering the thermal coupling between adjacent

rooms in the control model). In this case the global constraint

matrix will not have anymore a block-angular structure.

VI. APPENDIX

ccc

′T

i

= [ ccc

T

i

000

1,N

u

000

1,N

2

000

1,N

2

λ

i

111

1,N

2

λ

i

111

1,N

2

] ,

1

The open loop control of ∆T have the following behavior: ∆T(k) = 2,

for k ∈ [t

o

−5h, t

o

] and ∆T(k) = 0, otherwise. t

o

is the moment

when the ﬁrst room is occupied. An anticipation of 3h is used also for

the PI reference. Note that this type of control strategy is currently used at

SUPELEC in Rennes.

3 5 8 9 10 14 17 18 19 24

16

18

20

22

Time [h]

T

e

m

p

e

r

a

t

u

r

e

[ o

C

]

y

1

y

2

y

3

3 5 8 9 10 14 17 18 19 24

0

0.5

1

Time [h]

P

o

w

e

r

[

x

1

2

0

0

W

]

u

1

u

2

u

3

3 5 8 9 10 14 17 18 19 24

0

2

4

Time [h]

∆

T

[

o

C

]

Fig. 7. A one day simulation for the three zones building using the

distributed MPC based on Benders’ decomposition (N

u

= 20, N

2

= 30,

T

s

= 600s)

ccc

′T

c

= [ ccc

T

c

000

1,N

u

] , DDD = [

III

N

u

,N

u

III

N

u

,N

u

] , ggg = [ uuu

c ] ,

Φ

ij

=

_

¸

¸

_

φ

0

ij

0 ··· 0

φ

1

ij

φ

0

ij

0 ···

.

.

. ···

.

.

.

.

.

.

φ

N

2

−1

ij

··· φ

N

2

−N

u

+1

ij

N

2

−N

u

k=0

φ

k

ij

_

¸

¸

_

,

φ

k

ij

= C

i

A

k

i

B

ij

, Ψ

i

= [ (C

i

A

1

i

)

T

··· (C

i

A

N

2

i

)

T

]

T

,

EEE

i

=

_

000

N

u

,N

u

000

N

u

,N

u

−∆∆∆

i

ΦΦΦ

i2

000

N

2

,N

u

∆∆∆

i

ΦΦΦ

i2

000

N

2

,N

u

_

, hhh

i

=

_

uuu

c

−∆∆∆

i(www

i

−ΨΨΨ

i

xxx

i)

∆∆∆

i

(www

i

−ΨΨΨ

i

xxx

i

)

_

,

FFF

i

=

_

III

N

u

,N

u

III

N

u

,N

u

000

N

u

,N

2

000

N

u

,N

2

000

N

u

,N

2

000

N

u

,N

2

−∆∆∆

i

ΦΦΦ

i1

000

N

2

,N

u

III

N

2

,N

2

000

N

2

,N

2

−III

N

2

,N

2

000

N

2

,N

2

∆∆∆

i

ΦΦΦ

i1

000

N

2

,N

u

000

N

2

,N

2

III

N

2

,N

2

000

N

2

,N

2

−III

N

2

,N

2

_

.

REFERENCES

[1] C. G. Nesler, “Adaptive control of thermal processes in buildings,”

IEEE Control Systems Magazine, pp. 9–13, 1986.

[2] M. Hamdi and G. Lachiver, “A fuzzy control system based on the

human sensation of thermal comfort,” IEEE International Conference

on Fuzzy Systems, pp. 487–492, 1998.

[3] J. Liang and R. Du, “Design of intelligent comfort control system

with human learning and minimum power control strategies,” Energy

Conversion and Management, no. 48, pp. 517–528, 2008.

[4] N. Nassif, S. Kajl, and R. Sabourin, “Optimization of HVAC control

system strategy using two-objective genetic algorithm,” HVAC&R

Research, no. 3, pp. 459–486, 2005.

[5] M. Morari and J. Lee, “Model predictive control: past, present and

future,” Computers and Chemical Engineering, vol. 23, pp. 667–682,

1999.

[6] E. F. Camacho and C. Bordons, Model Predictive Control. Springer,

2004.

[7] R. Z. Freire, G. H. Oliveira, and N. Mendes, “Predictive controllers

for thermal comfort optimization and energy savings,” Energy and

Buildings, no. 40, pp. 1353–1365, 2008.

[8] F. Oldewurtel, C. Jones, and M. Morari, “A tractable approximation

of chance constrained stochastic MPC based on afﬁne disturbance

feedback,” in Conference on Decision and Control, CDC, dec 2008.

[9] G. B. Dantzig and P. Wolfe, “Decomposition principle for linear

programs,” Operation Research, no. 8, pp. 101–111, 1960.

[10] J. F. Benders, “Partitioning procedures for solving mixed-variables

programming problems,” Numerische Mathematik, vol. 4, pp. 238–

252, 1962.

If. respectively. k + j ∈ Inoccupationi . one of these constraints cannot be satisﬁed the global optimization problem becomes unfeasible. 1. allowing to reduce the energy consumption without affecting the comfort of the occupants. The minimization problem is: N2 u u c (k). So denote T δ i (k) = δi (k + 1) · · · δi (k + N2 ) as the future occupation proﬁle over the prediction horizon for the room i. According to (3). (8) (9) (10) (11) where the vector x i ∈ Rni is the local state. . [o C] 22. (3) where ci (k) and cc (k) are the prices for local and shared inputs.∀i∈S min J(k) = i∈S T u u c T (k)u i (k) + c c (k)u c (k) (4) i subject to µlow (k + j) = δi (k + j)(wi (k + j) − yi (k + j|k)) ≤ 0 ˆ i up µi (k + j) = δi (k + j)(ˆi (k + j|k) − wi (k + j)) ≤ 0 y ∀i ∈ S.ing system and several local heating sources in a multizone building. synthesized in an iterative algorithm. The sum of the last element of the cost sequences is due to the fact that the control input is considered constant out of the control horizon and equal to the value corresponding to the time step k + Nu − 1. The paper is organized as follows. the comfort constraints (5) are activated only during the occupation periods of the zones. wi wi yi ˆ Time N2 Time Occupation proﬁle illustration 0 ≤ u i ≤ u i (k) ≤ u i . A linear discrete time state space representation of a zone model i ∈ S = {1. ui . uc . the thermal comfort should be deﬁned only during the occupation periods (see Fig. (12) w Φ ∆ i (Φ i1u i (k) + Φ i2u c (k)) ≤ ∆ i (w i (k) − Ψ ix i (k)) . 1). II. i A. at time step k.u i (k). In practice.u i (k). with u i (k) = ui (k) u c (k) = uc (k) c i (k) = [ ci (k) ··· c c (k) = [ cc (k) ··· ··· ··· ui (k + Nu − 1) uc (k + Nu − 1) N2 −1 j=Nu −1 (6) (7) T . we can write the prediction equation for a control horizon Nu and a prediction horizon N2 as: ˆ ˆ ˆ y i (k) = yi (k + 1|k) · · · yi (k + N2 |k) Ψixi (k) + Φi1ui (k) + Φi2uc (k). The thermal comfort (deﬁned here by an upper wi (k + j) and a lower wi (k+j) temperature bounds) and the physical limitations of the process are the constraints of our optimization problem. where matrix ∆ i is obtained by eliminating the zero lines of δ the matrix ∆ ′ = diag(δ i (k)). k + j ∈ Occupationi 0. The control objective is to minimize the energy bill due to the indoor heating. s} can be stated as: ui (k) x (k + 1) = A x (k) + B i1 B i2 · i i i uc (k) (1) yi (k) = C ix i (k) Occupation (δi ) Temp. ∀i ∈ S 0 ≤ u c ≤ u c (k) ≤ u c . In Section III. Φ i1 and Φ i2 are deﬁned in Appendix. each element of this vector is deﬁned as: δi (k + j) = 1.N2 (5) min Jf (k) = J(k) + i∈S j=1 fi (k + j) (13) .. ... the occupation proﬁle and the temperature comfort bounds. P ROBLEM FORMULATION The anticipative effect of MPC consists in using a model of the process in order to predict its behavior during a ﬁnite horizon.∀i∈S Now we are able to formulate the linear programming problem corresponding to our control objectives as: u u c (k). including the energy prices. adding a penalty to the cost criterion if they are not satisﬁed. at a certain time step. 2. Using the model (1). . Replacing (2) in (5) and rewriting in a compact form we obtain Φ w ∆ i (Φ i1u i (k) + Φ i2u c (k)) ≥ ∆ i (w i (k) − Ψ ix i (k)) . ∀j = 1. we present the Benders’ decomposition used to obtain a distributed predictive control strategy. Conclusions and future directions are proposed in Section V. .5 20 yi 15 1 0 k Fig. Section II introduces the control model and formulates the MPC optimization problem. Relaxing the comfort constraints In the linear programming problem (4-7) the thermal comfort is considered as hard constraints. which is usually a linear function of the consumed energy (our control inputs).. = T T ci (k+Nu −2) cc (k+Nu −2) ci (k+j) ] T N2 −1 j=Nu −1 T cc (k+j) ] (2) where the matrices Ψ i . As mentioned before. The efﬁciency of the method is illustrated by simulation results is Section IV. Intuitively. respectively. to avoid this issue the comfort constraints are softened. yi ∈ R are the local and the shared input (electrical heating power and power input of the boiler) and the output (measured room temperature).

also called complicating variable. also known as the dual of Dantzig-Wolfe decomposition [9]. using an algorithm to solve exactly the dual subproblems (18). B ENDERS ’ DECOMPOSITION Bender’s decomposition..... We call (17) the subproblem i. . The optimal solution of the original large-scale problem can be shown to be identical to the solution obtained after a ﬁnite number of iterations..l − 1. the method is recommended for structured linear programs.. III. . We add the new constraint to the master problem (also called Benders cut) and a new iteration can begin. r Ji . This is the main idea of this decomposition. Consider a linear optimization problem having a block-angular constraint matrix structure in a standard form (15).. (18) The reason of using the dual subproblem is that the polyhedron Di = p i | F T p i ≤ c ′ that deﬁnes the feasible i i region of (18) is independent of u ′ . ∀lp = 1. zi (u ′ ) can also c be computed through the dual of (17). i ∈ S. (19) Now we are able to write the MP at iteration l. Even if any linear programming problem can be solved applying this decomposition technique. . Di can also be deﬁned c (via the theorem of convex combination) by a ﬁnite number of extreme points p 1 . Applying the duality. Note that over the iterations. none of the primal subproblems can be unfeasible. knowing the solutions of all dual subproblems at every previous iteration lp < l: min c ′T u ′ + z c c ′ u c . where the comfort penalty function fi is deﬁned as: up 0. Consequently. the dimension of subproblems is constant while the constraint matrix of MP increases after each iteration. deﬁned as: uc zi (u ′ ) = max p T (h i − E iu ′ ) i h c pi = hs ′ ··· ≥ 0 . consisting of solving the master problem and then solving (in parallel) the subproblems using the current value of the complicating variable. solving sequentially the MP and the subproblems [10]. The solution of (18) is an extreme point i i if the dual problem is feasible and bounded or an extreme ray if the dual is unbounded. prevents obtaining the c optimal solution by solving each subproblem independently. .∀i∈S c ui rewrite the original linear programming problem (15) as: s min c ′T u ′ + c c ′ uc i=1 uc zi (u ′ ) (16) subject to Du ′ = g c ′ uc ≥ 0. To distribute the computational demand among local controllers. . where at each iteration l the master problem optimal solution u ′l tends to the optimal value u ′∗ . E su ′ c u′ c + c ′T u ′ 1 1 + F 1u′ 1 ··· + c ′T u ′ s s = g = h1 . The dual problem (18) cannot be unfeasible.t. where uc zi (u ′ ) = min c ′T u ′ i i ′ ui subject to F iu ′ = h i − E 1u ′ . subject to F T p i ≤ c ′ i i . i=1 l p (p ip )T E iu ′ + z ≥ c i=1 l The Algorithm 1 that solves the original problem (15) using the Benders’ decomposition is iterative. . On the other hand. The optimization variables u ′ c and u ′ are obtained by adding the required number of slack i variables to the u c and u i .subject to (6) and (7). of the feasible region Di and the subproblem objective function will be: uc E c zi (u ′ ) = p T (h i −E iu ′ ) = i h k=1. Optimization variable u ′ . i ∈ S. c i u′ ≥ 0 i (17) c ′T u ′ c c subject to Du′ c E 1u′ c . it will return one of the extreme points. Du ′ = g c u′ ≥ 0 c z≥0 s s (20) p (p ip )T h i . F su ′ s u′ s . One of the most common solver for linear programming problems is the Simplex algorithm.. which is false by its deﬁnition. p Ii and a ﬁnite number of extreme i i rays r 1 . we will use the special block structure of the constraint matrix. This decomposition method splits a single large-scale linear programming problem into several independent problems which are coordinated by a single master problem (MP). Each iteration will have two main steps. which will be detailed in the next section.. because it should imply that the primal is unbounded. we know that solving independently the subproblems leads to the global optimum. In order to write explicitly c c the master problem and the subproblems. once the complicating variable uc u ′ has been chosen. . µup (k + j) > 0 i i λi µlow (k + j). This block-angular structure can be efﬁciently exploited by the Benders’ decomposition. for a ﬁxed value of this complicating variable. p i . dropping k for the simplicity of notations.z s. µlow (k + j) > 0. we will ﬁrstly . uses the block-angular structure of the constraint matrix in order to parallelize the computation of a linear optimization problem.u ′ . using the relaxed comfort constraints. respectively. i i (14) The always feasible linear programming problem (13) subject to (6) and (7) can be written in a standard form. But. u1 (15) where the matrices and the vectors composing this formulation are deﬁned in Appendix. + . µi (k + j) ≤ 0 and µlow (k + j) ≤ 0 i fi (k + j) = λi µup (k + j). as: min u ′ ...Ii pi h E c max (p k )T (h i −E iu ′ )..

This fact shows a good convergence speed of the algorithm and its slight dependence on the number of subproblems. j∈S (24) In order to study the complexity of the distributed algorithm compared to the centralized solver. The values of the model (1) matrices are: 0 0 A i = [ 0.. we observe a logarithmic dependence of l on s. It also tests the stop condition of the algorithm after every iteration. Empirical study of efﬁciency Since the distributed algorithm was implemented on a sequential machine.9921 0. ∀i ∈ S. 1. i 0 0 1 MPC1 MPC2 MPCs uc u1 x1 u2 x2 us xs Fig. 3 shows a very good scaling behavior of the algorithm in a distributed computing environment. the computational time required by the decomposition algorithm to solve a linear programming problem is: l V ALl = c ′T u ′l + up c c i=1 c ′T u ′l . . C T = [ 1 ] . 27 }. B i = [ 0.9931 ] .2595 0. ∀i ∈ S Ensure: Optimum control input sequences u ′∗ and u ′∗ . Nu = 5. MPCc . x i . The computational time using a distributed computing environment. . N2 } is the number of occupation time steps within the prediction period (i.. we will use a very simple thermal model of a room and supposing (without loss of generality) that all the subsystems have the same model. ignoring the communication time required at each iteration. regarding the number of subproblems. as it is always feasible. The local predictive controllers solve their subproblems once they have received from the master the current value of the complicating variable. Concerning the number of iterations.. obtaining u l and p l i i 5: Compute the current upper and lower bounds (21) and (22) 6: if V ALl − V ALl up low ≤ ǫ then 7: u ′∗ = u ′l and u ′∗ = u ′l . tM P Cc (l) is the time required by the central MPC to solve the master problem at iteration l. the last iteration solution can be applied to the system. In the mean time.e. (23) where l is the number of Benders iterations needed to solve the problem. MPCc tseq = l=1 tM P Cc (l) + j∈S tM P Cj (l) . where No ∈ {0.. If the algorithm does not offer a solution within a sample time.1376 ] . The central controller. we ﬁxed N2 = 30. For the three scenarios presented below we considered ﬁve different cases (in each case we changed the initial state of the subsystems) for each value of the variable parameter. Fig. obtaining u ′l and z l c 3: Send u ′l to all local controllers MPCi c 4: Solve (in parallel) the subproblems (in both primal (17) and dual (18) forms). up low Algorithm 1 Distributed MPC version for multi-source heating system based on Benders’ decomposition Require: Global problem formulation (15). acts like a coordinator for the local controllers. the number of lines of ∆ i ) and the tolerance ǫ. i the dimension of the subproblems di = (Nu + 2No ) × i i (2Nu + 4No ). in order to have more consistent statistical results. The dimension of the prediction horizon N2 should be chosen sufﬁciently large in order to offer enough time to the heating system to increase the indoor temperature up to the desired setpoint in the worst situation (low initial temperatures). for a small number of subsystems the centralized (Simplex) method offers better performances. 22 . In practice at each iteration l an upper and a lower bound of the optimum is computed: s IV. ǫ. S IMULATION RESULTS A. i ∈ c i S 1: Initialization: l = 1 2: Solve the master problem (20).. Distributed MPC scheme using Benders’ decomposition The efﬁciency of the distributed algorithm will be measured regarding the equivalent distributed computational time and the number of iterations l. In the results presented below. MPCi . 2. i i (21) (22) V ALl = c ′T u ′l + z l . The main parameters of the algorithm are: the number of subproblems (subsystems) s. .. low c c and the stop condition of the iterative algorithm will be V ALl − V ALl < ǫ. i 1) Scenario 1: s ∈ {21 . Stop c c i i 8: else 9: Update the constraints of the master problem (by adding the new Benders cut) 10: Update current iteration l = l + 1 and Goto step 2 11: end if The resulting control scheme has the structure presented in ﬁgure 2. −3 ǫ = 10 . can be expressed as: l tdistr = l=1 tM P Cc (l) + max tM P Cj (l) .The algorithm reaches the optimal solution of (15) when s l z l = i=1 zi . No = 15. while tM P Cj (l) is the computational time to solve the subproblem j.

while i i No depends on the occupation proﬁles. being negligible compared to the system time step Ts = 600s. . is a tuning parameter. = 15. Nu . The number of Benders iterations has a logarithmic dependence on the tolerance which shows the exponential convergence of the method. Each room has a double glazed window of 2m2 surface on the larger external wall of each zone. ǫ i 2) Scenario 2: s = 50. Neglecting the heat loss through the hydraulic system and considering the efﬁciency of the boiler η = 0. 4 shows. 3600 · 1000 Ts Cpw ρw qw .9 we can express the central and local costs in terms of electrical ce and gas/oil cg prices in kWh as: Ts . while the gas price is ce = 0. for the three control strategies compared. 6. No (25) 3) Scenario 3: s = 50. No ∈ −3 {1. 10−5 . All the rooms of the simulated building have the same surface (12m2 ) and are heated by a hot water-based underﬂoor heating system and local independent convectors of 1200W maximal power. 7.0742 e/kWh. in order to compare i No where Cpw . The external wall sandwich consists of 1cm of gypsum. This inner control loop is considered in steady state in our simulations. 4.. The electricity price considered is ce = 0. Simulation study This second part of the experimental results will focus on testing the proposed algorithm on a virtual building composed of three rooms. The internal wall is 7. Nu ∈ {1. Nu . s Fig. . 3. The occupation proﬁle and the thermal comfort bounds are those shown in Fig. Note that.. Table I compares the heating energy costs for a single day. and of the occupation i time steps. The inﬂuence of the number of subsystems. . The inﬂuence of the stop condition tolerance.2cm thick of gypsum board. The control inputs of the system are the electrical power of the local convectors and the water temperature increment (the difference between the water temperature which enters the boiler and the outlet water temperature). 5.. The water transport delay is not considered. No has a more important inﬂuence over the computational time i than Nu .. The simulator supposes a well mixed indoor air. the thermal comfort is similar. 10. Here we study the algorithm performances with respect to the subproblem sizes. di . which is normal as No has a greater weight on the dimension of the subproblem. ρw and qw are the speciﬁc heat. B. 8cm of extruded polystyrene and 20cm of concrete. cc (k + j) = cg (k + j) 3600 · 1000η ci (k + j) = ce (k + j) Fig. Nu = 10. . 10. ǫ = 10 . The weather conditions are those measured at Rennes on the 1st January 1998. 30}. 5. The inﬂuence of the control horizon. 150 tcentr (No=const) i Time [s] 100 50 0 t t t distr centr distr (N =const) (N =const) u i o (N =const) u 1 5 10 i 15 i 20 25 30 Nu (No=15) / No (Nu=15) Number of iterations 150 Ni =const o u 100 50 0 N =const 1 5 10 u i o 15 i o u 20 25 30 N (N =15) / N (N =15) the performances of the proposed control strategy with the commonly used PI controllers. 30}. Fig. as shown in Fig.. 5 shows that the tolerance ǫ inﬂuences very slightly the computational demand of the distributed algorithm. The simulation of the thermal building behavior was made using the MATLAB toolbox called SIMBAD. Here we suppose that the boiler is controlled by a fast regulator which receives ∆T as reference. As Fig. The control horizon dimension. the density and the mass ﬂow of the water through the boiler. 10−1 }....063 e/kWh. 5.10 4 40 t centr tcentr tdistr Time [s] Time [s] 10 2 tdistr 30 20 10 10 0 10 −2 10 20 Number of iterations 15 10 5 0 10 1 10 2 0 −6 10 25 Number of iterations 10 −5 10 −4 s ε 10 −3 10 −2 10 −1 20 15 0 0 10 10 1 10 2 10 −6 10 10 −5 10 −4 s ε 10 −3 10 −2 10 −1 Fig.. ǫ ∈ {10−6 .

H.N2 I ∆ iΦ i1 0 N2 . ··· N −Nu +1 ··· φij2 . Sabourin.” Energy Conversion and Management. CDC.” IEEE International Conference on Fuzzy Systems. pp.N2 0 Nu .Nu ∆ −∆ iΦ i2 0 N2 .N2 I ∆ −∆ iΦ i1 0 N2 .80 22 20 18 16 y1 y 2 y3 3 5 8 9 10 Time [h] 14 17 18 19 24 Power [x1200W] 1 0. 487–492. Oliveira.N2 λi1 1. Lee. [6] E. resulting in a linear programming problem.Nu I N2 . Du. [2] M. T . VI. 2005. .Nu 0 N2 .5 0 3 4 5 8 9 10 Time [h] 14 17 18 19 u u 1 y1 u1 Zone1 Zone2 u2 3 24 ∆ T [ C] 4m y2 u2 o 2 0 3 5 8 9 10 Time [h] 14 17 18 19 24 y3 u3 ∆T 6m Zone3 2m Fig. Nesler. 23.N2 0 Nu . Springer. . ].84 8. C. 459–486. pp. [10] J. 1962. In order to reduce the heating energy bill. 667–682. ] . “Predictive controllers for thermal comfort optimization and energy savings.Nu 0 φ0 ij ··· 0 I Nu . Freire. . Ψi = [ (Ci A1 )T i ij i Ei = Fi = 0 Nu . In this case the global constraint matrix will not have anymore a block-angular structure. subjected to the comfort constraints and physical limitations of the system. and R. 1986. G. [9] G. 238– 252. . hi = uc Ψ ∆ −∆ i (w i −Ψ ix i ) w Ψ ∆ i (w i −Ψ ix i ) I Nu . 2004. Morari. Kajl. Z. F. S. coordinated by a master controller. 101–111. pp. 7. pp. Camacho and C. pp. 1353–1365. dec 2008.Nu Φij = φ0 ij φ1 ij 0 ··· φij2 N −1 .N2 R EFERENCES [1] C. Liang and R. “A tractable approximation of chance constrained stochastic MPC based on afﬁne disturbance feedback. Hamdi and G. Wolfe. Note that this type of control strategy is currently used at SUPELEC in Rennes. no. vol.N2 φk = Ci Ak Bij . pp. Oldewurtel. N2 −Nu k=0 . the MPC criterion minimizes the heating energy cost. φk ij ··· (Ci Ai N2 T ) .Nu 0 1. 3.Nu Fig.Nu 0 Nu . . pp.” Numerische Mathematik. 8. Lachiver. Jones. [4] N.” Operation Research. [8] F. 6. D = [ I Nu . C ONCLUSION A distributed model predictive control strategy has been proposed for multisource multizone building temperature regulation. Mendes. vol. “Decomposition principle for linear programs.35 7. The centralized computational demand of the optimization problem grows exponentially with the number of subsystems (rooms).” in Conference on Decision and Control. 517–528. A PPENDIX c ′T = [ c T i i 0 1. we proposed a distributed control strategy based on Benders’ decomposition. ] . N2 = 30. 48. 40. An anticipation of 3h is used also for the PI reference. A one day simulation for the three zones building using the distributed MPC based on Benders’ decomposition (Nu = 20. otherwise.N2 λi1 1. [5] M. 9–13. . to ] and ∆T (k) = 0. B. 2008. which allows the decrease of the computational demand by using a network of local controllers. “Model predictive control: past. 1999.N2 0 N2 . Dantzig and P. Morari and J. 1 The open loop control of ∆T have the following behavior: ∆T (k) = 2. for k ∈ [to − 5h. g = [ uc ] . no.Nu 0 Nu . Model Predictive Control. . “Optimization of HVAC control system strategy using two-objective genetic algorithm.N2 −I N2 . Nassif. 4. 2008. [3] J. Benders. pp. 1960. Bordons.N2 0 1. for large-scale buildings.” Energy and Buildings. G. “A fuzzy control system based on the human sensation of thermal comfort.N2 0 N2 . F. .N2 I N2 .N2 −I N2 .Temperature [oC] TABLE I H EATING COST COMPARISON Control law dMPC with Benders decomposition local PI for convectors and open loop control for ∆T 1 local MPC for convectors and ∆T = 0 Cost [e] 6. no.Nu I Nu . “Design of intelligent comfort control system with human learning and minimum power control strategies. Ts = 600s) c ′T = [ c T c c 0 1. and M. 1998. [7] R. no. to is the moment when the ﬁrst room is occupied..Nu ] .Nu ∆ iΦ i2 0 N2 .N2 0 Nu . “Partitioning procedures for solving mixed-variables programming problems. “Adaptive control of thermal processes in buildings. Then.” Computers and Chemical Engineering. present and future.” IEEE Control Systems Magazine.N2 0 N2 . Future work will focus on the adaptation of the Benders’ decomposition algorithm in order to handle coupling subsystems (considering the thermal coupling between adjacent rooms in the control model). and N.” HVAC&R Research. Conﬁguration of the virtual building V.

- em8779-e.pdfUploaded byJoe Padayatty
- data leakage detectionUploaded byAshifAli
- Mpc (Webcast0jhl)Uploaded byYang Gul Lee
- Integer LindoUploaded byXiimena Rojas
- 2629509[1]Uploaded byfuckyou321
- Transportation Problem A Special Case for Linear Programming Problems.pdfUploaded byCaz Tee-Nah
- Nonconvex Separable Programming Problem for Optimal Raw Material Mix in Flexible Polyurethane Foam ProductionUploaded byAlexander Decker
- Quiz 5 QuestionsUploaded byChun Yu Poon
- Modeling and Optimization of Integrated Production Inventory Distribution Network for a Tyre Retread IndustryUploaded byInternational Journal of Research in Engineering and Science
- EggUploaded byOm Prakash Pal
- MPC Matlab UserGuideUploaded byanncar1987
- SNP OptimizerUploaded byPankaj Shukla
- Chapter 3cUploaded byBanshi Jakhar
- Operations-Research.pdfUploaded bymhel vianney bariquit
- minimum break pointUploaded byKléberVásquez
- Profitability on Operations ResearchUploaded byCrystal Padilla
- Bertasimas Odoni-Or2011-Optimal Selection of Airport Runway ConfigurationsUploaded bydharamkar
- Appendix C1 Optimization Method for Determination of Solubility Parameters 2014 Cleaning With SolventsUploaded byMenee Love U Too
- optimizationUploaded bylynndong0214
- Optimization Mesh Morphing Optimization 8Uploaded byVikram Mangalore
- MPC of Condesate Distillation ColumnUploaded byArief Noor Rahman
- GA - DEUploaded byZellagui Energy
- 1 Ejemplo Diseño Procesos - HidroalquilacionUploaded byMilagrosKaterineMuñozSalome
- 175922 Ch2A Linear ProgramingUploaded byThenes Kumar
- optimal power flowUploaded byDummyofindiaIndia
- Lect_ModConvOpt - Best1 Book on OtimizationUploaded bydartheamon
- Theory MetodeSimplexUploaded byBisma Reza
- Applied Mathematics & Information SciencesUploaded bySpin Fotonio
- IJERTV1IS9388Uploaded bySai Siva
- SNP_Optimization___Training_slides.pdfUploaded byjagannadha.birakayala6893