A 3D Optimal Control Problem Related To 2017 Journal of Mathematical Analys

J. Math. Anal. Appl.
446 (2017) 1571–1605
Contents lists available at ScienceDirect
Journal of Mathematical Analysis and Applications

www.elsevier.com/locate/jmaa
A 3D optimal control problem related to the urban heat islands

F.J. Fernández a , L.J. Alvarez-Vázquez b,∗ , A. Martínez b , M.E. Vázquez-Méndez c
a
C.U.D. Escuela Naval Militar, 36920 Marín, Spain
b
E.I. Telecomunicación, Universidade de Vigo, 36310 Vigo, Spain
c
E.P.S. Universidade de Santiago de Compostela, 27002 Lugo, Spain
a r t i c l e i n f o a b s t r a c t
Article history: Within the framework of numerical simulation and optimal control of partial
Received 13 April 2016 differential equations, in this work we deal with the mathematical modelling and
Available online 28 September 2016 control of the processes related to the urban heat island effect. In particular, we
Submitted by X. Zhang
are interested in finding the optimal locations of green zones inside metropolitan
Keywords: areas in order to mitigate the consequences of this harmful phenomenon. So, we
Urban heat island consider a three-dimensional climate model and formulate a constrained optimal
Green zones control problem, that is extensively analyzed in the first part of the paper. Then,
Optimal control we propose a complete numerical algorithm for its resolution, interfacing the interior
Numerical modelling point algorithm IPOPT with the FreeFem++ software package. Finally, we present
several numerical tests for a simple realistic case, where the advantages of our
approach are shown.
© 2016 Elsevier Inc. All rights reserved.
1. Introduction
The phenomenon known as urban heat island (usually denoted as UHI) is characterized by higher tem-
peratures in urban environments than in surrounding rural areas, mainly due to anthropogenic reasons.
According to observations during past century, this temperature difference, which is primarily attributable
to the urban built environment, ranges from 2 to 12 ◦ C (most investigations refer to a 5 ◦ C difference), and
may pose particular risks to urban populations. The phenomenon is more remarkable during summer than
in the other seasons of the year [25], and these temperature differences are larger at night than during the
day, especially in the case of very weak winds.
Nowadays, UHI is considered as one of the major environmental challenges in this century as an undesired
side effect of urbanization and industrialization of humankind, and many cities are adopting strategies to
mitigate UHI effect, especially metropolises with large population and intensive economic activities (in many
countries heat is the primary weather-related cause of death, and thus promotion of strategies for mitigating
* Corresponding author.
E-mail addresses: fjavier.fernandez@cud.uvigo.es (F.J. Fernández), lino@dma.uvigo.es (L.J. Alvarez-Vázquez),
aurea@dma.uvigo.es (A. Martínez), miguelernesto.vazquez@usc.es (M.E. Vázquez-Méndez).
http://dx.doi.org/10.1016/j.jmaa.2016.09.048
0022-247X/© 2016 Elsevier Inc. All rights reserved.
1572 F.J. Fernández et al. / J. Math. Anal. Appl. 446 (2017) 1571–1605
the UHI effect is a big concern for government agencies). The main reason for this night-time warming is
that the shortwave radiation is still within concrete and asphalt (absorbed during the day), unlike rural
areas, and this energy is slowly released during the night as longwave radiation, preventing a rapid cooling
process. With a decreased amount of vegetation, cities often lose the shade and cooling effect of trees,
the low albedo of their leaves, and the removal of carbon dioxide. Moreover, materials commonly used in
urban areas for pavements and roofs, such as concrete and asphalt, have significantly different thermal bulk
properties (including heat capacity and thermal conductivity) and surface radiative properties (albedo and
emissivity) than the surrounding rural areas, leading to higher temperatures.
Mitigation of the UHI effect can be accomplished through the use of green roofs or of light-coloured
surfaces in urban areas (which reflect more sunlight and absorb less heat), and also – as will be addressed in
this paper – through the increasing of vegetation cover inside cities, mainly in the form of urban forests and
parks, in order to maximize the multiple vegetation benefits in controlling the temperature rises. During
last decades a vast research effort has been addressed and a wide range of literature is available for the
subject, mainly from an engineering point of view (see, for instance, the review articles [28,4,27,20,18] and
the numerous references therein). However, as far as we know, the mathematical approach to the problem
has been much more poorly attended [17,1,21], and we can emphasize the recent work of the authors [12],
where a bidimensional control problem has been formulated to mitigate in an optimal way the UHI effect
by means of shadow trees zones.
This current study was carried out in order to introduce and develop a mathematical tool to deal with
this environmental problem in a more realistic 3D framework. So, in section 2 we formulate a well-posed
three-dimensional model for the UHI phenomenon, which is mathematically analyzed and numerically solved
in sections 3 and 4, respectively. Sections 5 to 7 are devoted, respectively, to present the formulation of the
optimal control problem, prove the existence of optimal solutions, and introduce a numerical algorithm
for its computation. Finally, last sections of the paper are devoted to show several numerical results and
conclusions for a realistic example.
2. Mathematical formulation of the problem
In this section we present the three-dimensional mathematical model that we will need to solve in order
to address the optimal control problem. Basically, we consider four coupled systems of partial differential
equations (mainly, those describing the velocity and the temperature of air, and the temperatures of soil and
buildings) posed on three different domains whose union represents the whole domain under study (these
three domains are, respectively, the air, the soil and the buildings).
2.1. The physical domain
We start this subsection introducing several notations related to the definition of the domain and its
corresponding boundaries, whose graphical representation can be seen in Fig. 1.
We define the 2D domain Ω2D = {(x, y) ∈ R2 : 0 < x < a, 0 < y < b} where we consider the
positive functions HS , HA : Ω2D −→ R+ representing, respectively, the height of the layers of soil and
air, with R+ denoting the set of nonnegative real numbers, that is, R+ = {α ∈ R : α ≥ 0}. We also
consider a subdomain Ω2DB ⊂Ω
2D
(the place over where buildings are constructed), and a positive function
HB : ΩB −→ R representing the height of these buildings (obviously, we assume that HB (x, y) < HA (x, y)
2D +
for any (x, y) ∈ Ω2D

B ). Then, we define the following 3D domains:
• The domain occupied by soil:
S = {x = (x, y, z) ∈ R : (x, y) ∈ Ω
Ω3D 3 2D
, 0 < z < HS (x, y)}
F.J. Fernández et al. / J. Math. Anal. Appl. 446 (2017) 1571–1605 1573
Fig. 1. Scheme of the material domains involved in the problem and their different boundaries.
• The domain occupied by buildings:
B = {(x, y, z) ∈ R : (x, y) ∈ ΩB , HS (x, y) < z < HS (x, y) + HB (x, y)}

Ω3D 3 2D
• The domain occupied by air:
A = {(x, y, z) ∈ R : (x, y) ∈ Ω
Ω3D , HS (x, y) < z < HS (x, y) + HA (x, y)} \ Ω3D
3 2D
B
Related to previous domains, we consider the following parts of the boundaries (as a general indication to
the different notations, we will usually include as subscript an indicator of the lower domain of the boundary,
and as superscript an indicator of the upper one). So, we define:
• Lower boundary of the soil:
ΓS0 = {(x, y, z) ∈ R3 : (x, y) ∈ Ω2D , z = 0}
• Interface boundary between air and soil (this one will be the essential boundary when considering the
optimal control problem):
S = {(x, y, z) ∈ R : (x, y) ∈ Ω
ΓA 3 2D \ Ω2D , z = H (x, y)}
B S
• Interface boundary between soil and buildings (for simplicity we assume that the buildings have no
subterranean areas):
S = {(x, y, z) ∈ R : (x, y) ∈ ΩB , z = HS (x, y)}

ΓB 3 2D
• Boundary of buildings associated with roofs (we distinguish between roofs and walls because they usually
present quite different thermal properties):
B = {(x, y, z) ∈ R : (x, y) ∈ ΩB , z = HB (x, y)}

ΓR 3 2D
• Boundary of buildings associated with walls:
B = {(x, y, z) ∈ R : (x, y) ∈ ∂ΩB , HS (x, y) ≤ z ≤ HS (x, y) + HB (x, y)}

ΓW 3 2D
• Upper boundary of the air:
A = {(x, y, z) ∈ R : (x, y) ∈ Ω
ΓH 3 2D , z = H (x, y) + H (x, y)}.
S A
• Wind inflow boundary:
A = {(x, y, z) ∈ R : (x, y) ∈ (∂Ω

ΓIN )IN , HS (x, y) ≤ z ≤ HS (x, y) + HA (x, y)}
3 2D
• Wind outflow boundary:
ΓOU
A
T
= {(x, y, z) ∈ R3 : (x, y) ∈ (∂Ω2D )OU T , HS (x, y) ≤ z ≤ HS (x, y) + HA (x, y)}
• Lateral boundaries for the air:

A = {(x, y, z) ∈ R : (x, y) ∈ ∂Ω
ΓN \ (∂Ω2D )IN ∪ (∂Ω2D )OU T ,
3 2D
HS (x, y) ≤ z ≤ HS (x, y) + HA (x, y)}
• Lateral boundaries for the soil:
S = {(x, y, z) ∈ R : (x, y) ∈ ∂Ω
ΓN , 0 ≤ z ≤ HS (x, y)}.
3 2D
Thus, taking into account above notations, we have that the full boundaries of the 3D domains are given
by:
S = Γ0 ∪ ΓS ∪ ΓS ∪ ΓS ,
∂Ω3D S A B N
A = ΓS ∪ ΓB ∪ ΓB ∪ ΓA ∪ ΓA ∪ ΓA
∂Ω3D ∪ ΓN
A R W H IN OU T
A,
B = ΓS ∪ ΓB ∪ ΓB .
∂Ω3D B R W
2.2. The state systems
Once established the three different geometric domains, we proceed to write the state systems associated
with the mathematical model used here, corresponding to a time interval I = (0, T ).
• Velocity uA (x, t) (m s−1 ) and pressure pA (x, t) (m2 s−2 ) of air: Since we are dealing with a relatively
close to the ground atmospheric layer, we can consider the incompressible Navier–Stokes equations:
⎧
⎪ ∂uA θA
⎪
⎪ + (uA · ∇)uA − ∇ · (νA ∇uA ) + ∇pA = REF g A × I,
in Ω3D
⎪
⎪ ∂t θA
⎪
⎪
⎪
⎪ ∇ · u = 0 on Ω 3D
× I,
⎪
⎨
A A
uA · nA = uINA on A × I,
ΓIN (1)
⎪
⎪
⎪
⎪ uA · n A = u A
OU T
on ΓOU T × I,
⎪
⎪ N AR
⎪
⎪ uA · nA = 0 on ΓA ∪ ΓB ∪ ΓW B ∪ ΓS × I,
A
⎪
⎪
⎩ 0
uA (0) = uA in Ω3D
A ,
REF
with νA the kinematic viscosity coefficient, θA a reference temperature, g the gravity acceleration,
nA the unit outward normal vector to the boundary ∂Ω3D IN OU T
A , and uA , uA
0
and uA given boundary and
initial conditions.
• Temperature θA (x, t) (K) of air: We neglect radiation heat transfer provided by soil, buildings and sun
to the air, only considering heat transfer occurred by convection:
⎧
⎪ ∂θA
⎪
⎪ + uA · ∇θA − ∇ · (KA ∇θA ) = FA in Ω3D
A × I,
⎪
⎪ ∂t
⎪
⎪
⎪
⎪ θA = θAIN
A × I,
on ΓIN
⎪
⎪
⎪
⎪ ∂θ
⎪
⎪ KA
A
= 0 on ΓN A ∪ ΓA
OU T
∪ ΓH
A × I,
⎪
⎪ ∂n
⎪
⎨ A
∂θA
KA = bS,A
1 (θS − θA ) on ΓAS × I, (2)
⎪
⎪ ∂n A
⎪
⎪
⎪
⎪ ∂θA
= bW,A (θB − θA ) on ΓW
⎪
⎪ KA B × I,
⎪
⎪ ∂n A
1
⎪
⎪
⎪
⎪ ∂θA
⎪
⎪ KA = bR,A (θB − θA ) on ΓRB × I,
⎪
⎪ ∂n A
1
⎩ 0
θA (0) = θA , in Ω3DA ,
with KA the diffusion coefficient, bS,A W,A

1 , b1 and bR,A
1 the convection coefficients (their complete char-
IN 0
acterization can be found in next subsection), FA a heat source term, and θA and θA given boundary
and initial conditions.
• Temperature θS (x, t) (K) of soil: In this case, since radiation effects are relevant, we consider heat
transfer due both to convection and to radiation:
⎧
⎪ ∂θS
⎪ ∂t − ∇ · (KS ∇θS ) = FS in ΩS × I,
⎪ 3D
⎪
⎪
⎪
⎪
⎪
⎪
∂θS
= bA,S A,S
1 (θA − θS ) + b2 ((Tr ) − θS4 ) on ΓA
S × I,
A,S 4
⎪
⎪ KS
⎪
⎪ ∂n S
⎪
⎨ ∂θS
KS = bB,S
1 (θB − θS ) on ΓB S × I, (3)
⎪ ∂n S
⎪
⎪
⎪
⎪
∂θS
⎪
⎪ KS = 0 on ΓN S × I,
⎪
⎪ ∂n S
⎪
⎪
⎪
⎪ θS = θSSU B on Γ0S × I,
⎪
⎩
θS (0) = θS0 in Ω3DS ,
with KS the diffusion coefficient, bA,S

1 and bB,S
1 the convection coefficients, bA,S
2 the radiation coefficient,
TrA,S the radiation temperature induced by solar radiation (this temperature depends on the type of
surface and the angle of incidence, as detailed in next subsection), FS a source term, and θSSU B and θS0
given boundary and initial conditions.
• Temperature θB (x, t) (K) of buildings: Here we also consider heat transfer by convection and radiation:
⎧
⎪
⎪
∂θB
− ∇ · (KB ∇θB ) = FB in Ω3D
⎪
⎪ B × I,
⎪
⎪ ∂t
⎪
⎪ ∂θB
⎪
⎪ = bA,R (θA − θB ) + bA,R ((TrA,R )4 − θB B × I,
4
⎪ KB ) on ΓR
⎪
⎨ ∂n B
1 2
∂θB (4)
⎪ KB = bA,W (θA − θB ) + bA,W ((TrA,W )4 − θB4
) on ΓWB × I,
⎪
⎪ ∂n 1 2
⎪
⎪
B
⎪
⎪ ∂θ
⎪ KB B = bS,B
⎪ 1 (θS − θB ) on ΓBS × I,
⎪
⎪ ∂n
⎪
⎩
B
0
θB (0) = θB in Ω3D
S ,
with KB the diffusion coefficient, bA,R

1 , bA,W
1 and bS,B
1 the convection coefficients, bA,R
2 and bA,W
2 the
A,R A,W 0
radiation coefficients, Tr and Tr the radiation temperatures, FB a source term, and θB a given
initial condition.
2.3. Parameters and data
The characteristic parameters that define the thermal behaviour of the materials involved in the problem
are the following:
• ρA , ρS and ρB (g m−3 ) are the densities of air, soil and buildings, respectively.
• cpA , cpS and cpB (W s g−1 K−1 ) are the specific heat capacities of air, soil and buildings.
• αA , αS and αB (W m−1 K−1 ) are the thermal conductivities of air, soil and buildings.
• S , W and R (dimensionless constants) are the emissivities of the surfaces corresponding to soil, walls
and roofs, respectively.
• aS , aW and aR (dimensionless constants) are the albedos of soil, walls and roofs, representing the ratio
of reflected radiation from the surface to incident radiation upon it.
−2 −1
• hA B A A
S , hS , hW and hR (W m K ) are the convective heat transfer coefficients between soil/air, soil/build-
ings, walls/air and roofs/air, respectively.
From above generic coefficients, we define the following parameters appearing in the mathematical for-
mulation (2)–(4) of the thermal subproblems:
• KA , KS and KB (m2 s−1 ) are the thermal diffusivities of air, soil and buildings, defined from above
data in the following way:
αA αS αB
KA = , KS = , KB = .
ρA cpA ρS cpS ρB cpB
• bS,A A,S S,B

1 , b1 , b1 , b1
B,S
, bW,A
1 , bA,W
1 , bR,A
1 and bA,R
1 (m s−1 ) are the coefficients related to convective heat
transfer, obtained from following relations:
– for the temperature of air:
ρA cpA bS,A
1 = hA
S, ρA cpA bW,A
1 = hA
W, ρA cpA bR,A
1 = hA
R;
– for the temperature of soil:
ρS cpS bA,S
1 = hA
S, ρS cpS bB,S
1 = hB
S;
– for the temperature of buildings:
ρB cpB bS,B
1 = hB
S, ρB cpB bA,W
1 = hA
W, ρB cpB bA,R
1 = hA
R.
• bA,S
2 , b2
A,W
and bA,R
2 (m s−1 K−3 ) are the coefficients related to radiative heat transfer for soil, walls and
roofs, respectively, obtained from following relations:
ρS cpS bA,S
2 = σB S , ρB cpB bA,W
2 = σB W , ρB cpB bA,R
2 = σB R ,
with σB (W m−2 K−4 ) the Stefan–Boltzmann constant.

• Finally, in order to compute the radiation temperatures TrA,S , TrA,W and TrA,R (K) on the different solid
boundaries (soil, walls and roofs) we use the following expressions (involving the corresponding solar
radiations, albedos and emissivities):
σB S (TrA,S )4 = (1 − aS )Rsw,net (x, t) + Rlw,dow (x, t),

σB W (TrA,W )4 = (1 − aW )Rsw,net (x, t) + Rlw,dow (x, t),

σB R (TrA,R )4 = (1 − aR )Rsw,net (x, t) + Rlw,dow (x, t),
where Rsw,net (x, t) denotes the net incident shortwave radiation on the surface, and Rlw,dow (x, t) denotes
the downwelling longwave radiation, both measured in W m−2 .
Remark 1. In Section 5 we will see that the control problem consists, in essence, in changing the thermal
behaviour of a given region of the soil, which will translate into a change in the characteristic parameters
associated with that part of the soil (density, specific heat capacity, thermal conductivity, emissivity and
albedo). Specifically, we will consider the soil paved by two different materials: asphalted zones (denoted by
a subscript C) and green areas (denoted by a subscript G), and corresponding to these materials we will
have, instead of the generic soil parameters (indicated by a subscript S), the following specific ones:
• ρC and ρG , densities of asphalt and green zones.

• cpC and cpG , specific heat capacities of asphalt and green zones.
• αC and αG , thermal conductivities of asphalt and green zones.
• C and G , emissivities of asphalt and green zones.
• aC and aG , albedos of asphalt and green zones.
• hA A
C and hG , convective heat transfer coefficients between asphalt/air and green zones/air.
As expected, these coefficients will redefine the parameters appearing in the state systems (1)–(4) in the
following way:
• KC and KG , the thermal diffusivities of asphalt and green zones, will be defined by:
αC αG
KC = , KG = .
ρC cpC ρG cpG
• bC,A
1 , bG,A
1 , bA,C
1 and bA,G
1 , the coefficients related to convective heat transfer, will be obtained from:
ρC cpC bA,C
1 = hA
C, ρG cpG bA,G
1 = hA
G, ρA cpA bC,A
1 = hA
C, ρA cpA bG,A
1 = hA
G.
• bA,C
2 and bA,G
2 , the coefficients related to radiative heat transfer, will be given by:
ρC cpC bA,C
2 = σB C , ρG cpG bA,G
2 = σB G .
• On the other hand, since the generic material (soil) will turn into two different materials (asphalt – or
another similar building material – and green zones), it is necessary to redefine the coefficients related
to radiative heat transfer between soil and buildings. In our case, we will assume that the soil on which
the buildings stand is of asphalt type, therefore, instead of the coefficient hB S , we will consider the
coefficient hB C (coefficient related to convective heat transfer between asphalt and buildings). Thus, we
will have:
ρC cpC bC,B
1 = hB
C, ρB cpB bB,C
1 = hB
C.
Obviously, this ratio will not be affected by the control, and its value will depend only on the type of
material on which the buildings lay down. For simplicity, in the theoretical development we will continue
considering a generic soil.
• Finally, for the radiation temperatures we will have:
σB C (TrA,C )4 = (1 − aC )Rsw,net (x, t) + Rlw,dow (x, t),

σB G (TrA,G )4 = (1 − aG )Rsw,net (x, t) + Rlw,dow (x, t).
3. Mathematical analysis of the state systems
This section is devoted to the mathematical study of the existence of solution for the state systems (1)–(4),
a set of four coupled, nonlinear partial differential equations (with their corresponding initial/boundary
conditions, necessary for the well-posedness of the problem). For the sake of simplicity and without loss
of generality, in our analysis we will assume the existence of only one convex building constructed over a
convex region of soil, and that all the Dirichlet boundary conditions appearing in the state systems are null.
3.1. The concept of solution of the state systems
We start this subsection defining the functional spaces used in the search of solutions for the different
state subsystems. In particular,
• for the velocity of the fluid (air) we consider the following space:

VA = u ∈ [H 1 (Ω3D
A )] : u · nA = 0 on ∂ΩA , ∇ · u = 0 in ΩA
3 3D 3D
,
and we look for solutions of the Navier–Stokes equations in the space:
W1 = W 1,2,4/3 (I; VA , VA ) ∩ L∞ (I; [L2 (Ω3D 3

A )] ), (5)
where we denote (cf. the monograph of Roubíček [24] for more details on the functional space):

∂u
W 1,p,q (I; X, X ) = u ∈ Lp (I; X) : ∈ Lq (I; X ) ;
∂t
• for the temperature of the fluid we consider the following space:

XA = θ ∈ H 1 (Ω3D IN
A ) : θ = 0 on ΓA ,
and we look for solutions of the heat equation in the space:

W2 = W 1,2,4/3 (I; XA , XA ) ∩ L∞ (I; L2 (Ω3D
A )); (6)
• for the temperature of the soil we consider the following space:

XS = θ ∈ H 1 (Ω3D
S ) : θ |Γ A ∈ L5
(ΓA
S ), θ = 0 on Γ0
S
S
endowed with the usual norm

θ
XS =
θ
H 1 (Ω3D
S )
+
θ|ΓA
L5 (ΓA
S)
, and we look for solutions of the heat
S
equation in the space:

W3 = θ ∈ W 1,2,5/4 (I; XS , XS ) : θ|ΓA ∈ L5 (I; L5 (ΓA
S ))
S
∞
∩ L (I; L 2
(Ω3D
S )); (7)
• and, finally, for the temperature of the building we consider the space:

XB = θ ∈ H 1 (Ω3D
B ) : θ|ΓW ∪ΓR ∈ L (ΓB ∪ ΓB ) ,
5 W R
B B
endowed with the norm

θ
XB =
θ
H 1 (Ω3D
B )
+
θ|ΓW ∪ΓR
L5 (ΓW R , and we look for solutions of the
B ∪ΓB )
B B
heat equation in the space:

W4 = θ ∈ W 1,2,5/4 (I; XB , XB ) : θ|ΓW ∪ΓR ∈ L5 (I; L5 (ΓW
B ∪ Γ R
B ))
B B
∞
∩ L (I; L 2
(Ω3D
B )). (8)
If we denote by Cw ([0, T ]; Y ) the space of weakly continuous mappings from the time interval [0, T ] to an
arbitrary functional space Y , then, related to above functional spaces, we know several standard regularity
properties (see, for instance, [24]), which we summarize in the following result:
Lemma 1. The following inclusions are compact:
• W1 ⊂⊂ [L10/3− (Ω3D A × I)] , W1 ⊂⊂ Cw ([0, T ]; [L (ΩA )] ).

3 2 3D 3
• W2 ⊂⊂ L10/3− (Ω3D A × I), W2 ⊂⊂ L (I; L

2 4−
(ΓAS ∪ ΓB ∪ ΓB )),
W R
W2 ⊂⊂ Cw ([0, T ]; L (ΩA )).

2 3D
• W3 ⊂⊂ L10/3− (Ω3D S × I), W3 ⊂⊂ L (I; L (Γ2 )),

5 5
W3 ⊂⊂ Cw ([0, T ]; L2 (Ω3D
S )).
• W4 ⊂⊂ L10/3− (Ω3D B × I), W4 ⊂⊂ L5 (I; L5 (ΓR
B ∪ ΓB ∪ ΓS )),
W B
W4 ⊂⊂ Cw ([0, T ]; L2 (Ω3D
S )).
Given two matrices A = (Aij ), B = (Bij ) ∈ M3×3 (R), we denote by A : B their standard Frobenius
3
inner product, that is, A : B = tr(AT B) = i,j=1 Aij Bij . Then, we can specify what we understand as a
solution of the state systems:
Definition 1. We say that (uA , θA , θS , θB ) ∈ W1 × W2 × W3 × W4 is a very weak solution of the state systems
(1)–(4) if it satisfies the following variational formulations:
• for the velocity of the air:
T
dv
− , uA dt + ∇uA uA · v dx dt
dt VA ,VA
0 I×Ω3D
A

θA
+ νA ∇uA : ∇v dx dt = REF
g · v dx dt (9)
θA
I×Ω3D
A I×Ω3D
A

+ 0
uA · v(0) dx, ∀v ∈ W 1,∞,∞ (I; VA , VA ) such that v(T ) = 0;
Ω3D
A
• for the temperature of the air:
T
dz
− , θA dt + uA · ∇θA z dx dt
dt ,X
XA A
0 I×Ω3D
A

+ KA ∇θA · ∇z dx dt + bS,A
1 θA z dγ dt + bW,A
1 θA z dγ dt
I×Ω3D
A I×ΓA
S I×ΓW
B

+ bR,A
1 θA z dγ dt = FA z dx dt (10)
I×ΓR
B I×Ω3D
A

+ bS,A
1 θS z dγ dt + bW,A
1 θB z dγ dt + bR,A
1 θB z dγ dt
I×ΓA
S I×ΓW
B I×ΓR
B

+ 0
θA z(0) dx, ∀z ∈ W 1,∞,∞ (I; XA , XA ) such that z(T ) = 0;
Ω3D
A
• for the temperature of the soil:
T
dz
− , θS dt + KS ∇θS · ∇z dx dt + bA,S
1 θS z dγ dt
dt ,X
XS S
0 I×Ω3D
S I×ΓA
S

+ bB,S
1 θS z dγ dt + bA,S 4
2 θS z dγ dt = FS z dx dt (11)
I×ΓB
S I×ΓA
S I×Ω3D
S

+ bA,S
1 θA z dγ dt + bB,S
1 θB z dγ dt + bA,S A,S 4
2 (Tr ) z dγ dt
I×ΓA
S I×ΓB
S I×ΓA
S

+ θS0 z(0) dx, ∀z ∈ W 1,∞,∞ (I; XS , XS ) such that z(T ) = 0;
Ω3D
S
• and, finally, for the temperature of the building:
T
dz
− , θB dt + KB ∇θB · ∇z dx dt + bA,R
1 θB z dγ dt
dt ,X
XB B
0 I×Ω3D
B I×ΓR
B

+ bA,W
1 θB z dγ dt + bS,B
1 θB z dγ dt + bA,R
2
4
θB z dγ dt
I×ΓW
B I×ΓB
S I×ΓR
B

+ bA,W
2
4
θB z dγ dt = FB z dx dt + bA,R
1 θA z dγ dt
I×ΓW
B I×Ω3D
B I×ΓR
B

+ bA,W
1 θA z dγ dt + bS,B
1 θS z dγ dt (12)
I×ΓW
B I×ΓB
S

+ bA,R
2 (TrA,R )4 z dγ dt + bA,W
2 (TrA,W )4 z dγ dt
I×ΓR
B I×ΓW
B

+ 0
θB z(0) dx, ∀z ∈ W 1,∞,∞ (I; XB , XB ) such that z(T ) = 0.
Ω3D
S
3.2. Mathematical analysis of the existence of solution
In order to demonstrate the existence of (not necessarily unique) solutions of the coupled system of
variational formulations (9)–(12), we need to recall the standard concept of multi-valued function M :
V ⇒ V , that is, a function M : V → 2V , where the powerset 2V denotes the set of all subsets of V . So, we
will show that the multi-valued function:
M : W1 × W2 × W3 × W4 ⇒ W1 × W2 × W3 × W4 (13)
(wA , ξA , ξS , ξB ) → (uA , θA , θS , θB )
has a fixed-point, where (uA , θA , θS , θB ) ∈ W1 × W2 × W3 × W4 are solutions of the following variational

formulations:
• for the velocity of the air (Oseen equations):
T
dv
− , uA dt + ∇uA wA · v dx dt
dt VA ,VA
0 I×Ω3D
A

ξA
+ νA ∇uA : ∇v dx dt = REF
g · v dx dt (14)
θA
I×Ω3D
A I×Ω3D
A

+ 0
uA · v(0) dx, ∀v ∈ W 1,∞,∞ (I; VA , VA ) such that v(T ) = 0;
Ω3D
A
T
dz
− , θA dt + wA · ∇θA z dx dt
dt ,X
XA A
0 I×Ω3D
A

+ KA ∇θA · ∇z dx dt + bS,A
1 θA z dγ dt + bW,A
1 θA z dγ dt
I×Ω3D
A I×ΓA
S I×ΓW
B

+ bR,A
1 θA z dγ dt = FA z dx dt (15)
I×ΓR
B I×Ω3D
A

+ bS,A
1 ξS z dγ dt + bW,A
1 ξB z dγ dt + bR,A
1 ξB z dγ dt
I×ΓA
S I×ΓW
B I×ΓR
B

+ 0
θA z(0) dx, ∀z ∈ W 1,∞,∞ (I; XA , XA ) such that z(T ) = 0;
Ω3D
A
T
dz
− , θS dt + KS ∇θS · ∇z dx dt + bA,S
1 θS z dγ dt
dt ,X
XS S
0 I×Ω3D
S I×ΓA
S

+ bB,S
1 θS z dγ dt + bA,S 4
2 θS z dγ dt = FS z dx dt (16)
I×ΓB
S I×ΓA
S I×Ω3D
S

+ bA,S
1 ξA z dγ dt + bB,S
1 ξB z dγ dt + bA,S A,S 4
2 (Tr ) z dγ dt
I×ΓA
S I×ΓB
S I×ΓA
S

+ θS0 z(0) dx, ∀z ∈ W 1,∞,∞ (I; XS , XS ) such that z(T ) = 0;
Ω3D
S
T
dz
− , θB dt + KB ∇θB · ∇z dx dt + bA,R
1 θB z dγ dt
dt ,X
XB B
0 I×Ω3D
B I×ΓR
B

+ bA,W
1 θB z dγ dt + bS,B
1 θB z dγ dt + bA,R
2
4
θB z dγ dt
I×ΓW
B I×ΓB
S I×ΓR
B

+ bA,W
2
4
θB z dγ dt = FB z dx dt + bA,R
1 ξA z dγ dt
I×ΓW
B I×Ω3D
B I×ΓR
B

+ bS,B
1 ξS z dγ dt + bA,W
1 ξA z dγ dt (17)
I×ΓB
S I×ΓW
B

+ bA,R
2 (TrA,R )4 z dγ dt + bA,W
2 (TrA,W )4 z dγ dt
I×ΓR
B I×ΓW
B

+ 0
θB z(0) dx, ∀z ∈ W 1,∞,∞ (I; XB , XB ) such that z(T ) = 0.
Ω3D
S
To prove the existence of the above mentioned fixed-point, we will use a generalization of the Kakutani
fixed-point theorem for multi-valued functions (as can be found, for instance, in the classical works of
Fan [11] or Glicks [15]) stating that an upper semicontinuous multi-valued mapping M : V ⇒ V with
nonempty closed convex values, which maps a nonempty compact convex set K of a locally convex space V
into its powerset 2K has a fixed point, i.e., there exists a point u ∈ K such that u ∈ M (u).
The existence of solutions (and their a priori estimates) for equations (14), (15), (16) and (17) can be
found in the book of Roubíček [24]. For matters related to the uniqueness of parabolic equations with low
regularity in the time derivative, the interested reader is referred to the paper of the authors [3]. The details
related to the treatment of the non-local radiation terms in the heat equations, can be found, for instance,
in several articles [30,19,26]. Then, all above results can be summarized in the following lemma:
Lemma 2. For the multi-valued function M defined in (13) we have that:
1. Given uA 0
∈ L2 (Ω3DA ), wA ∈ W1 and ξA ∈ W2 , there exists, at least, an element uA ∈ W1 solution
of (14), that satisfies the following estimate:

uA
W1 ≤ C1 1 +
wA
W1 +
ξA
W2 , (18)
where C1 is a positive constant such that C1 = C1 (Ω3D

A , T, νA , θ
REF 0
, uA , g).
0
2. Given θA ∈ L2 (Ω3D
A ), FA ∈ L (I; L (ΩA )), wA ∈ W1 , ξS ∈ W3 and ξB ∈ W4 , there exists a unique
2 2 3D
element θA ∈ W2 solution of (15), that satisfies the following estimate:

θA
W2 ≤ C2 1 +
wA
W1 +
ξS
W3 +
ξB
W4 , (19)
S,A W,A
where C2 is a positive constant such that C2 = C2 (Ω3D
A , T, KA , b1 , b1 , bR,A
1
0
, FA , θA ).
3. Given θS ∈ L (ΩS ), FS ∈ L (I; L (ΩS )), ξA ∈ W2 and ξB ∈ W4 , there exists a unique element
0 2 3D 2 2 3D
θS ∈ W3 solution of (16), that satisfies the following estimate:

θS
W3 ≤ C3 1 +
ξA
W2 +
ξB
W4 (20)
where C3 is a positive constant such that C3 = C3 (Ω3D A,S

S , T, FS , Tr , θS0 , KS , bA,S B,S
1 , b1 , bA,S
2 ).
4. Given θB ∈ L (ΩB ), FB ∈ L (I; L (ΩB )), ξA ∈ W2 and ξS ∈ W3 , there exists a unique element
0 2 3D 2 2 3D
θB ∈ W4 solution of (17), that satisfies the following estimate:

θB
W4 ≤ C4 1 +
ξA
W2 +
ξS
W3 (21)
where C4 is a positive constant such that C4 = C4 (Ω3D A,W

B , T, FB , Tr , TrA,R , θB
0
, KB , bA,R
1 , bA,W
1 ,
bS,B
1 , bA,R
2 , bA,W
2 ).
We can prove now the following technical results:
Theorem 1. The multi-valued function M , as defined in (13), is a weakly upper semicontinuous mapping.
Proof. We consider the sequence {(wA,n , ξA,n , ξS,n , ξB,n )}n∈N ⊂ W1 × W2 × W3 × W4 such that
{(wA,n , ξA,n , ξS,n , ξB,n )} (wA , ξA , ξS , ξB ) ∈ W1 × W2 × W3 × W4 . To demonstrate the desired result,
we only need to prove that, for any sequence {(uA,n , θA,n , θS,n , θB,n )}n∈N ⊂ W1 × W2 × W3 × W4 such
that (uA,n , θA,n , θS,n , θB,n ) ∈ M (wA,n , ξA,n , ξS,n , ξB,n ), ∀n ∈ N, we have that (uA,n , θA,n , θS,n , θB,n )
(uA , θA , θS , θB ), with (uA , θA , θS , θB ) satisfying that (uA , θA , θS , θB ) ∈ M (wA , ξA , ξS , ξB ).
In fact, for each n ∈ N, let (uA,n , θA,n , θS,n , θB,n ) ∈ W1 × W2 × W3 × W4 be a solution of the variational
formulations (14)–(17) associated to (wA,n , ξA,n , ξS,n , ξB,n ). Since {(wA,n , ξA,n , ξS,n , ξB,n )}n∈N is a weakly
convergent sequence, it is bounded. Thus, from the estimates obtained in Lemma 2, we obtain that the
sequence of images {(uA,n , θA,n , θS,n , θB,n )}n∈N is also bounded in the space W1 × W2 × W3 × W4 . In
particular, there exists an element (uA , θA , θS , θB ) ∈ W1 × W2 × W3 × W4 , and a subsequence (still denoted
in the same way, for the sake of simplicity) such that:
• uA,n → uA in [L10/3− (Ω3D A × I)] , for any > 0 small enough,

3
and in Cw ([0, T ]; [L (ΩA )] ),

2 3D 3
• uA,n uA in L2 (I; VA ),
• θA,n → θA in L10/3− (Ω3D A × I), in L (I; L
2 4−
S ∪ ΓB ∪ ΓB )),
(ΓA W R
and in Cw ([0, T ]; L (ΩA )),

2 3D
• θA,n θA in L2 (I; XA ),
• θS,n → θS in L10/3− (Ω3D S × I), in L (I; L (ΓS ∪ ΓS ∪ ΓS )),
5 5 A B N
and in Cw ([0, T ]; L (ΩS )),

2 3D
• θS,n θS in L2 (I; XS ),
• θB,n → θB in L10/3− (Ω3D B × I), in L (I; L (ΓB ∪ ΓB ∪ ΓS )),

5 5 R W B
and in Cw ([0, T ]; L (ΩB )),

2 3D
• θB,n θB in L2 (I; XB ).
Using above convergences, we can easily pass to the limit in the variational formulations (14)–(17)
associated to (wA,n , ξA,n , ξS,n , ξB,n ), obtaining in a straight manner that (uA , θA , θS , θB ) is a solution of
(14)–(17) associated to (wA , ξA , ξS , ξB ). 2
Theorem 2. The multi-valued function M , as defined in (13), is convex in the sense that, given an element
(wA , ξA , ξS , ξB ) ∈ W1 × W2 × W3 × W4 , the image set M (wA , ξA , ξS , ξB ) is convex.
Proof. We must demonstrate that, given a pair of elements (uA 1 1

, θA , θS1 , θB
1 2
) and (uA 2
, θA , θS2 , θB
2
) ∈
M (wA , ξA , ξS , ξB ), for each α ∈ [0, 1], the convex, linear combination α(uA , θA , θS , θB ) + (1 − α)(uA , θA
1 1 1 1 2 2
, θS2 ,
2
θB ) ∈ M (wA , ξA , ξS , ξB ).
On one side, from the uniqueness of solution of the problems (15)–(17) (given in Lemma 2), we have that
1 2
θA = θA , θS1 = θS2 , and θB 1 2
= θB . Thus, each linear combination of those elements lies trivially in the set.
1 2
On the other side, for the elements uA and uA , the linearity of the problem (14) allows us to state that
any convex, linear combination of them keeps lying in the set M (wA , ξA , ξS , ξB ). 2
Then, and related to the estimates established in Lemma 2, we can state the following straightforward
algebraic result, whose proof can be left to the reader:
Lemma 3. Let us denote
ΔC1 ,C2 ,C3 ,C4 = 1 + 3 C1 C2 C3 C4 + C1 C2 C3 + C1 C2 C4 + C1 C3 C4

− C1 − C1 C2 − C2 C3 − C2 C4 − C3 C4 − 2 C2 C3 C4 ,
k , k = 1, . . . , 4, such that:
and define the constants C
1 = C1 (1 − C3 C4 ) (C2 + 1) ,
ΔC1 ,C2 ,C3 ,C4 C
2 = C2 (1 + C3 + C4 + C3 C4 − C1 C3 − C1 C4 − 2 C1 C3 C4 ) ,
ΔC1 ,C2 ,C3 ,C4 C
3 = C3 (C4 + 1) (C2 + 1) (1 − C1 ) ,
ΔC1 ,C2 ,C3 ,C4 C
4 = C4 (C3 + 1) (C2 + 1) (1 − C1 ) .
ΔC1 ,C2 ,C3 ,C4 C
If we assume that the parameters and data of our problem are such that the constants ΔC1 ,C2 ,C3 ,C4 and
Ck , k = 1, . . . , 4, are strictly positive, then, for any element (wA , ξA , ξS , ξB ) ∈ W1 ×W2 ×W3 ×W4 satisfying
that
wA
W1 ≤ C 1 ,
ξA
W ≤ C 2 ,
ξS
W ≤ C 3 and
ξB
W ≤ C 4 , the solutions of the problem (14)–(17)
2 3 4

verify
uA
W1 ≤ C1 ,
θA
W2 ≤ C2 ,
θS
W3 ≤ C3 and
θB
W4 ≤ C 4 .
We finish this section setting the result that guarantees the existence of solution of the state system
(9)–(12), whose proof is a direct application of the generalized version of the Kakutani fixed-point theorem
to the multi-valued function M :
Theorem 3. Under the hypotheses of Lemma 3, there exists a fixed-point (uA , θA , θS , θB ) of the multi-valued
function M (as defined in (13)), which is a solution of the state systems (9)–(12), and satisfies the following
estimates:
1 (1 +
uA

uA
W1 ≤ C 0

[L2 (Ω3D 3 ),
A )]
2 (1 +
θ0
L2 (Ω3D ) +
FA
L2 (I;L2 (Ω3D )) ),

θA
W2 ≤ C (22)
A A A
3 (1 +
θ0
L2 (Ω3D ) +
FS
L2 (I;L2 (Ω3D )) ),

θS
W3 ≤ C S S S
4 (1 +
θ0
L2 (Ω3D ) +
FB
L2 (I;L2 (Ω3D )) ),

θB
W4 ≤ C B B B
where C1 , C
2 , C
3 and C
4 are positive constants (obtained from those found in Lemma 3) depending on all
the corresponding space/time domains, and on all the coefficients and data appearing in the formulation of
the problem.
Proof. The closed, bounded, convex set
K = {(vA , zA , zS , zB ) ∈ W1 × W2 × W3 × W4 :
1 ,
zA
W ≤ C

vA
W1 ≤ C 2 ,
zS
W ≤ C
3 ,
zB
W ≤ C
4 }
2 3 4
is weakly compact in W1 × W2 × W3 × W4 , and, due to Lemma 3, the multi-valued function M maps this
set K into its powerset 2K . From Theorem 2 the image sets of M are convex, and from Lemma 2 these
image sets are also nonempty.
Then, taking into account that due to Theorem 1 the function M is weakly upper semicontinuous, we can
use Kakutani fixed-point theorem to obtain the existence of a point (uA , θA , θS , θB ) ∈ M (uA , θA , θS , θB ),
that is, a solution of (9)–(12), which will satisfy in a standard way estimates (22). 2
Remark 2. Even if in the mathematical analysis for the existence of solution of the state systems we have
assumed that all the parameters are constant, our reasoning can be extended without any problem to the
general case of non-constant coefficients, as long as they remain bounded.
Although non-constant coefficients will be used in next sections for the mathematical analysis of the
control problem, we have preferred to deal with constant coefficients in this section, in order to simplify the
already tedious notations used in the estimates.
Remark 3. A further step could be to analyze the uniqueness of solution for the state systems (1)–(4).
However, with this aim in mind, we should face a extremely difficult problem: the uniqueness of solution uA
of the Navier–Stokes equations in domains contained in R3 .
The objective of this work is very far from studying uniqueness of solution of the Navier–Stokes problem
(we refer the interested reader, for instance, to the classic monograph of Temam [29]), and, fortunately, in
the setting of the optimal control problem, we will not need the velocity of the air uA to be unique.
4. Numerical resolution of the state systems
In this section we present a numerical algorithm to solve the state systems (1)–(4). With this aim in mind,
we propose a semi-implicit space/time discretization based on the method of characteristics (for the time
semi-discretization), and on the finite element method (for the space semi-discretization), in the spirit of that
presented in previous papers of the authors (see, for instance, [13] and the references therein), and that –
for the sake of completeness – will be described in the following subsections. It is worthwhile remarking here
that this time semi-discretization will solve the difficulty of the coupling between the systems, and that,
finally, nonlinearities will be worked out with a suitable fixed point technique. All these discretizations are
designed to be practically implemented in any scientific software (FreeFem++ [16], in our particular case).
4.1. General considerations on the time semi-discretization
In order to set the time semi-discretization of the state systems (1)–(4) for the interval [0, T ], we choose
a number N ∈ N and define the time step Δt = N T
. Then we consider a set of N + 1 discrete times
{tn }n=0 ⊂ [0, T ] given by tn = nΔt, for n = 0, . . . , N .
N
In view of the theoretical analysis of the state systems, we note that the existence of solution can be
only proved for a certain range of parameter values. This restriction on the parameters has also an impact
at the numerical level, that we will avoid proposing a space/time semi-implicit discretization. We will see
that the proposed discretization uncouples the resolution of the systems, which will be very advantageous
in order to solve numerically the control problem.
1
We denote α = Δt , and consider the material derivative of a generic escalar field φ:
Dφ ∂ ∂φ
(x, t) = φ(X(x, t), t) = (x, t) + uA (x, t) · ∇φ(x, t),
Dt ∂t ∂t
where X represents the characteristic line, that is,
∂X
(x, t) = uA (x, t).
∂t
Thus, we can approximate the material derivative in the following way:
Dφ
(tn+1 ) α(φn+1 − φn ◦ Xn ), (23)
Dt
where φn represents an approximation to φ(tn ), and Xn (x) = X(x, tn+1 ; tn ) (i.e., the position at time tn+1
of a particle that at time tn was located at point x), is the solution of the following trajectory equation:
⎧
⎨ dX = u (X(x, t; τ ), τ ),
A
dτ
⎩ X(x, t; t) = x.
Thus, the latter term in (23) can be approximated by:
(φn ◦ Xn )(x) φn (x − uA
n
(x)Δt).
The main idea on which our discretization is based is the following: Let us suppose that we have a system
of partial differential equations as follows:
⎧
⎪ Dφ1
⎨ = h1 (t, φ1 ) + g1 (t, φ2 ),
Dt
⎪
⎩ Dφ2 = h2 (t, φ2 ) + g2 (t, φ1 ),
Dt
where h1 , h2 , g1 and g2 are generic differential operators. We propose the following semi-implicit scheme:
Given (φ01 , φ02 ), we will compute, for each n = 0, . . . , N − 1, (φn+1
1 , φn+1
2 ) as the solution of:
αφn+1
1 = h1 (tn+1 , φn+1
1 ) + g1 (tn , φn2 ) + αφn1 ◦ Xn ,
αφn+1
2 = h2 (tn+1 , φn+1
2 ) + g2 (tn , φn1 ) + αφn2 ◦ Xn ,
that is, we will treat in an implicit way the differential operators related to the main part, and in an explicit
way the differential operators corresponding to the coupling terms.
4.2. The full space/time discretization
In this subsection we face the numerical resolution of our particular state systems (1)–(4). So, for the
several domains appearing in the formulation, we consider regular meshes of size h (no necessarily of the
same characteristic size for the different domains Ω3D A , ΩS
3D
and Ω3D
B ), from which we define the finite
element spaces corresponding the different fields involved in the mathematical formulation of our problem:
VhA for the air velocity uA , MhA for the pressure pA (we assume that VhA and MhA are consistent, in the sense
that they satisfy the inf-sup conditions for the Stokes equations), XhA for the air temperature θA , XhS for
the soil temperature θS , and XhB for the temperature of the buildings θB .
Then, if we apply the general scheme exposed in previous subsection to our particular problem, starting
from given initial data (uA 0
, p0A , θA
0
, θS0 , θB
0
), we are interested in finding, for each n = 0, . . . , N − 1, the
elements
, pA , θA , θS , θB ) ∈ VhA × MhA × XhA × XhS × XhB

n+1 n+1 n+1 n+1 n+1
(uA
such that satisfy the following variational formulations:
• for the velocity and the pressure of the air:

α n+1
uA · v dx + νA ∇uA
n+1
: ∇v dx − A ∇ · v dx
pn+1
Ω3D
A Ω3D
A Ω3D
A

n
θA
= n,REF
g · v dx + α n
uA ◦ Xn · v dx, ∀v ∈ VhA , (24)
θA
Ω3D
A Ω3D
A

∇ · uA
n+1
q dx = 0, ∀q ∈ MhA ,
Ω3D
A

α n+1
θA z dx + KA ∇θA
n+1
· ∇z dx + bS,A n+1
1 θA z dγ
Ω3D
A Ω3D
A ΓA
S

+ bW,A
1
n+1
θA z dγ + bR,A
1
n+1
θA z dγ = FAn+1 z dx (25)
ΓW
B ΓR
B Ω3D
A

n
+α θA ◦ Xn z dx + bS,A n
1 θS z dγ + bW,A
1
n
θB z dγ
Ω3D
A ΓA
S ΓW
B

+ bR,A
1
n
θB z dγ, ∀z ∈ XhA ,
ΓR
B

α θSn+1 z dx + KS ∇θSn+1 · ∇z dx + bA,S
1 θS
n+1
z dγ
Ω3D
S Ω3D
S ΓA
S

+ bB,S
1 θSn+1 z dγ + bA,S n+1 4
2 (θS ) z dγ = FSn+1 z dx (26)
ΓB
S ΓA
S Ω3D
S

+α θSn z dx + bA,S n
1 θA z dγ + bB,S
1
n
θB z dγ
Ω3D
S ΓA
S ΓB
S

+ bA,S A,S,n+1 4
2 (Tr ) z dγ, ∀z ∈ XhS ,
ΓA
S

α n+1
θB z dx + KB ∇θBn+1
· ∇z dx + bS,B n+1
1 θB z dγ
Ω3D
B Ω3D
B ΓB
S

+ bA,W
1
n+1
θB z dγ + bA,R
1
n+1
θB z dγ + bA,W
2
n+1 4
(θB ) z dγ (27)
ΓW
B ΓR
B ΓW
B

+ bA,R
2
n+1 4
(θB ) z dγ = FBn+1 z dx + α n
θB z dx
ΓR
B Ω3D
B Ω3D
B

+ bS,B n
1 θS z dγ + bA,W
1
n
θA z dγ + bA,R
1
n
θA z dγ
ΓB
S ΓW
B ΓR
B

+ bA,W
2 (TrA,W,n+1 )4 z dγ + bA,R
2 (TrA,R,n+1 )4 z dγ, ∀z ∈ XhB .
ΓW
B ΓR
B
Remark 4. We must note that, although we have decoupled the resolution of the four systems, we still have
to solve for each time step a Stokes problem and two nonlinear problems (related to the Stefan–Boltzmann
boundary conditions).
It is worthwhile recalling here that in order to solve the Stokes problem we will use standard techniques
(such as those found, for instance, in [14]), while the numerical solution of the nonlinearities will be faced
using a fixed-point method [2].
5. The optimal control problem
We are interested in finding, in an optimal way, the location of a given number M of green zones over the
soil surface Ω2D \ Ω2D
B (obviously, not including the surface occupied by buildings), in such a way that the
cooling effect produced by these green zones causes a temperature reduction inside a set of NB buildings
under study and, consequently, minimizes the economic cost due to air conditioning.
5.1. The control variables
As already commented, the main characteristic of green zones is the fact that they present a thermic
behaviour completely different from the rest of paving materials, which, from a mathematical viewpoint,
translates into the controls (or design variables) to enter the control systems through the coefficients related
to this thermic behaviour. In particular, this means that the controls will affect the coefficients KS , bS,A1 ,
bA,S
1 , bA,S
2 and T A,S
r .
So, let us consider M green areas (assumed to be rectangular, for the sake of simplicity), characterized
by their location inside the 2D domain Ω2D (the height coordinate z remains determined by the height HS
of the soil, that will not be modified anyway). Thus, we have the following control vector:
b = (b1 , b2 , . . . , bM ) ∈ R4M ,
where, for each k = 1, 2, . . . , M , bk = (p1,k 1,k 2,k 2,k

1 , p2 , p1 , p2 ) ∈ R represents the location of the k-th green
4
area:
Gk = [p1,k 1,k 2,k 2,k

1 , p2 ] × [p1 , p2 ] ⊂ Ω
2D \ Ω2D .
B
Thus, the parameters KS , bS,A

1 , b1
A,S
and bA,S
2 , affected by the controls in a straight manner, can be
∞
considered as bounded functions from Ω2D \ Ω2D
B into L (Ω
2D
\ Ω2D
B ) defined by:
M
M
KS (b) = k=1 χGk KG + 1 − k=1 χGk KC
M
M
bS,A (b) = χ G bG,A
+ 1 − χ G bC,A
1
k=1 k 1 M
k=1 k
1
(28)
M
bA,S (b) = χ G b A,G
+ 1 − χ
k=1 Gk bA,C
1
k=1 k 1 M 1
A,S M A,G
b2 (b) = k=1 χGk b2 + 1 − k=1 χGk bA,C 2 ,
∞ ∞
while the radiation temperature TrA,S can be understood as a function from Ω2D \ Ω2D
B into L (I; L (Ω
2D
\
2D
ΩB )) defined by:

M
M
TrA,S (b) = χGk TrA,G + 1− χGk TrA,C , (29)
k=1 k=1
(where we are assuming that the radiation temperatures TrA,G and TrA,C are bounded functions from
L∞ (0, T )).
Remark 5. We must note that the dependence relation control–parameter occurs via the indicator functions
of the rectangles Gk , k = 1, 2, . . . , M , associated to the corresponding green zones. Moreover, it is necessary
remarking that, given a control b, functions KS (b), bS,A A,S A,S
1 (b), b1 (b) and b2 (b) only show dependence on
the space variable, while function TrA,S (b) depends both on space and on time.
Unfortunately, the mappings given by (28)–(29) are not continuous in the functional spaces where they
are defined (they could be continuous if we would consider Lp (Ω2D \ Ω2D B ), for a suitable p < ∞, instead of
L∞ (Ω2D \ Ω2DB )). This remark means that, although we are able to show the existence of solution for the
state systems in the case of non-continuous coefficients, in next section we would find very serious difficulties
in order to prove the existence of optimal solutions for the control problem. In order to avoid this trouble
in advance, we will regularize the indicator functions χGk in such a way that we will obtain not only a
continuous regularized approximation, but even a differentiable smooth function.
Motivated by previous remark, we consider the following approximation (depending on a small parameter
> 0) of the indicator function of an interval [a, b] ⊂ R (in the technical paper [31], for instance, the
interested reader can find the whole mathematical details of this regularization):
⎧
⎪
⎪ 0 if x < a − ,
⎪
⎪
⎪
⎨ ν((x − a)/) if a − ≤ x ≤ a + ,
Ra,b, (x) = 1 if a + < x < b − , (30)
⎪
⎪
⎪
⎪ ν((b − x)/) if b − ≤ x ≤ b + ,
⎪
⎩
0 if b + < x,
where ν(ξ) = 12 + 32
1
(45ξ−50ξ 3 +21ξ 5 ). We have that this function is smooth: Ra,b, ∈ C 1 (R) and, in addition,
it is differentiable with respect to the interval (in the sense that the mapping (a, b) ∈ R2 → Ra,b, ∈ L∞ (R)
is differentiable, with δRa0 ,b0 , (a, b) ≡ ∇(a,b) Ra0 ,b0 · (a, b) ∈ L∞ (R), for all (a, b) ∈ R2 ).
Bearing in mind this regularization, we can approach the indicator function of the set Gk by χGk (x, y)
gk, (x, y), where:
gk, (x, y) = Rp1,k ,p1,k , (x) Rp2,k ,p2,k , (y),

1 2 1 2
and we can then redefine the coefficients KS , bS,A A,S

1 , b1 , b2
A,S
and TrA,S just by replacing in expressions
(28)–(29) the indicator functions χGk by the smoother functions gk, . In this way, we obtain that these new
mappings control–coefficient are continuous in L∞ (Ω2D \ Ω2DB ) (in fact, they are also differentiable).
5.2. The cost functional
With respect to the objective function, we are interested in minimizing the economic cost related to the
air conditioning in the buildings. We will suppose that the global cost of the energy consumption is given
by the following functional:
NB T

1 W,l A W,l
J(b) = λB hB
l (θ̄B , t) + λA hl (θ̄A , t) dt, (31)
NB T
l=1 0
where λB , λA ∈ [0, 1] are weight parameters such that λB + λA = 1, and where

W,l 1 W,l 1
θ̄B = θB (x, t) dγ, θ̄A = θA (x, t) dγ
μ(ΓW,l
B ) μ(ΓW,l
B )
ΓW,l
B ΓW,l
B
denote, for each building l = 1, . . . , NB , the mean temperatures of the building facade ΓW,l
B ⊂ ΓB and of
W
the air in contact with that facade, respectively. (In above expressions, μ(Γ) denotes the Euclidean measure

of a generic set Γ, i.e., μ(Γ) = Γ 1 dγ, and the fields θB and θA are the solutions of the respective state
systems (4) and (2) considering the regularized coefficients introduced in previous subsection.)
Finally, for each l = 1, . . . , NB , the functions:
l : R × [0, ∞[ −→ R
hB
W,l W,l
(θ̄B , t) −→ hB
l (θ̄B , t),
l : R × [0, ∞[ −→ R
hA
W,l W,l
(θ̄A , t) −→ hA
l (θ̄A , t),
represent the effective economic cost associated to air conditioning inside the l-th building (being the latter
related to the presence of windows in the facade and the corresponding lack of insulation). We should
note that not all buildings have the same thermal insulation, or the same energy efficiency of the heating,
ventilation and air conditioning system. Thus, the functions associated with energy cost may depend on the
particular building. The explicit time dependence of these functions is motivated by the temporal variation
experienced by the energy cost.
5.3. Mathematical formulation of the optimal control problem
As mentioned at the beginning of the section, the problem is essentially related to placing in an optimal
way the M green zones in order to minimize the energy costs associated with air conditioning of the living
areas of the buildings. So, with this general purpose in mind, we perform the following considerations before
setting the control problem:
• We suppose the existence of a number L (L ≥ M ) of possible locations for the green areas (characterizing
their situation by their relative position in the bidimensional domain Ω2D ). These possible locations
S,l ⊂ Ω
will be denoted by Ω2D 2D \ Ω2D , for l = 1, . . . , L, and are assumed to be disjoint (Ω2D ∩ Ω2D =
B S,i S,j
∅, ∀i = j) and to be larger than the minimum area allowed for the green zones. In order to simplify the
formulation of the problem we will assume that the possible locations Ω2D S,l are rectangular:
S,l = [a1 , b1 ] × [a2 , b2 ],

Ω2D l l l l
l = 1, . . . , L,
but any other convex configuration would remain valid.

• The extent of each green area must be always larger than a minimum allowed value μmin and smaller
than a maximum allowed value μmax :
μmin ≤ μ(Gk ) ≤ μmax , k = 1, . . . , M,
where, obviously, 0 < μmin < μmax . Moreover, we will impose that the sum of all the areas associated
to the green zones must be lower than a given measure:

M
μ(Gk ) ≤ μM
max ,
k=1
with μM max > M μmin . (This latter assumption is essential in order to analyze if it is a better option the
choice of only one large zone than the choice of several smaller ones.)
• Each green area must be located in only one possible region, and must be fully contained inside it, that
is, for each k ∈ {1, . . . , M } there must exist a unique l ∈ {1, . . . , L} such that Gk ⊂ Ω2D
S,l .
• We will not allow two different green areas to lie in the same possible region (even if geometrically
feasible).
Taking into account above considerations, our real-world problem can be formulated as the following
optimal control problem:
min J(b), (32)

b∈Bad
where J(b) is the cost functional defined in (31), Bad is the following subset of R4M :

Bad = b ∈ R4M : ∀k = 1, . . . , M, Gk ⊂ Ω2D
S,lk , for any lk ∈ {1, . . . , L},

M
with Ω2D
S,li ∩ Ω2D
S,lj = ∅, ∀i = j; μmin ≤ μ(Gk ) ≤ μmax ; μ(Gk ) ≤ μM
max
k=1
and (uA , pA , θA , θB , θS ) is the solution of the state systems (1)–(4).

If we introduce a new binary control variable (say, for instance, y ∈ {0, 1}M L ) representing the inclusion of
a particular green zone inside a given possible region, the optimal control problem (32) can be reformulated
as a problem of type MINLP (mixed integer nonlinear programming):
min f (b, y)
b,y
s.t. h(b, y) ≤ 0 (33)
Ay ≤ c
y ∈ {0, 1}M L ,
where y ∈ RM L , b ∈ R4M , f : R4M +M L → R, h : R4M +M L → Rm1 , A ∈ Mm2 ×(M L) , and c ∈ Rm2 . In

particular:
• The integer variable y = {ykl } ∈ {0, 1}M L indicates the location of the green areas inside the possible
regions (that is, for k = 1, . . . , M , l = 1, . . . , L, ykl = 1 means that the k-th green zone Gk lies inside
the l-th possible region Ω2D l
S,l ; otherwise, yk = 0).
• The real variable b ∈ R 4M
remains showing the relative position of the M green areas within the
possible regions in which we want to locate them.
• The new objective function is such that f (b, y) = J(b). (It is worthwhile recalling here that each
evaluation of this objective function f needs the numerical resolution of the state systems (1)–(4).)
• The linear constraints Ay ≤ c represent the fact that each green zone has to be inside only one possible
region, and that each possible region can only have one green zone. The former is written as

M
ykl ≤ 1, ∀l = 1, . . . , L, (34)
k=1
and the latter as

L
ykl = 1, ∀l = 1, . . . , M. (35)
l=1
Thus, m2 = 2(M + L) (it is necessary to consider the double of the expected dimensions, since each
equality constrain gives origin to two inequality constraints). In order to simplify the notation, we will
denote the set of feasible discrete controls by Yad = {y ∈ {0, 1}M L : Ay ≤ c}.
• The nonlinear constraints h(b, y) ≤ 0 correspond to three different types of limitations: First, the points
characterizing each green zone must be coherent, that is:
p1,k 1,k
1 − p2 ≤ 0, p2,k 2,k
1 − p2 ≤ 0, ∀k = 1, . . . , M. (36)
On the other hand, each green area needs to be completely included inside only one possible region

L
L
ykl (al1 − p1,k
1 ) ≤ 0, ykl (p1,k
2 − b1 ) ≤ 0,
l
(37)
l=1 l=1

L
L
ykl (al2 − p2,k
1 ) ≤ 0, ykl (p2,k
2 − b2 ) ≤ 0,
l
∀k = 1, . . . , M.
l=1 l=1
(We must remark here that, due to the constraints on the binary variable, only one term is non-null in
each summation.) Next, the maximal/minimal bounds related to the extent of each green zone must be
observed:
μmin − (p1,k 1,k 2,k 2,k

2 − p1 )(p2 − p1 ) ≤ 0, (38)
(p1,k 1,k 2,k 2,k
2 − p1 )(p2 − p1 ) − μmax ≤ 0, ∀k = 1, . . . , M.
Finally, the limitation on the global extent of the green areas must be also satisfied:

M
(p1,k 1,k 2,k 2,k
2 − p1 )(p2 − p1 ) − μmax ≤ 0.
M
(39)
k=1
So, from above constraints, we deduce that m1 = 2M + 4M + 2M + 1 = 8M + 1. Fixed an integer

control y∗ ∈ Yad , we denote Uad = {b ∈ R4M : h(b, y∗ ) ≤ 0}, corresponding to the set of feasible real
controls. Evidently, the set of feasible real controls varies with the chosen integer control y∗ , but always
remaining convex, closed and bounded in R4M .
We should also note here that, given a fixed element y∗ ∈ Yad , the nonlinear programming (NLP)
problem:
min f (b, y∗ )
b (40)
s.t. h(b, y∗ ) ≤ 0
can be addressed as an optimal control problem. So, denoting by b∗ any solution of (40), it is clear that,
y
if the set of feasible discrete controls Yad is finite (and small, as in our case), then any element (b, ) such
∗ ∗ ∗
) = min {f (b , y ) : y ∈ Yad } is a solution of problem (33). Thus, looking to the mathematical
that f (b, y
analysis of the control problem, we must focus our efforts on the problem (40). Obviously, from a numerical
point of view, approaching the resolution of the problem by following the same strategy (solving an optimal
control problem for all possible values of the discrete control) may become impracticable in the case of
larger sets Yad , making necessary the use of more sophisticated algorithms.
6. Mathematical analysis of the optimal control problem
As outlined in previous section, we will focus our efforts on the theoretical analysis of the NLP prob-
lem (40), corresponding to fixing the discrete control variable y in the MINLP problem (33).
Theorem 4. Under the hypotheses of Theorem 3, we assume that the set of feasible discrete controls Yad is
nonempty, and that the functions {(hA l , hl ) : l = 1, . . . , NB }, associated to the air conditioning economic
B
cost inside each building, are continuous and bounded. Then, for each discrete control y∗ ∈ Yad there exists,
at least, a real control b∗ ∈ Uad such that:
f (b∗ , y∗ ) = min f (b, y∗ ) (41)

b∈Uad
Proof. Given a discrete control y∗ ∈ Yad , let us consider a minimizing sequence {bn }n∈N such that
lim f (bn , y∗ ) = min f (b, y∗ ).

n→∞ b∈Uad
Then, since Uad is a closed and bounded set in R4M , there exists a subsequence of {bn }n∈N (still denoted
in the same way for the sake of simplicity) satisfying that limn→∞ bn = b∗ , with b∗ ∈ Uad .
Corresponding to each continuous control bn , for n ∈ N, we consider the associated states (uA,n , θA,n , θS,n ,
θB,n ) ∈ W1 × W2 × W3 × W4 , solution of the corresponding state systems (9)–(12).
Now, taking into account the existence of positive constants KS , KS , bS,A S,A A,S A,S A,S A,S
1 , b1 , b1 , b1 , b2 , b2 ,
TrA,S and TrA,S such that, for all b ∈ Uad ,
KS ≤ KS (b) ≤ KS ,
bS,A
1 ≤ bS,A S,A
1 (b) ≤ b1 , bA,S
1 ≤ bA,S A,S
1 (b) ≤ b1 , (42)
bA,S
2 ≤ bA,S A,S
2 (b) ≤ b2 , TrA,S ≤ TrA,S (b) ≤ TrA,S ,
we obtain, from estimates (22) for the solutions of the state systems, that the sequence of associated states
1 , C
{(uA,n , θA,n , θS,n , θB,n )}n∈N is bounded, in the sense that there exist positive constants C 2 , C
3 and C
4
(only depending on the coefficients of the problem, on the space/time domains, and on the bounds (42))
such that:
1 (1 +
u0
[L2 (Ω3D )]3 ),

uA,n
W1 ≤ C A A
2 (1 +
θ0
L2 (Ω3D ) +
FA
L2 (I;L2 (Ω3D )) ),

θA,n
W2 ≤ C A A A
3 (1 +
θ0
L2 (Ω3D ) +
FS
L2 (I;L2 (Ω3D )) ),

θS,n
W3 ≤ C S S S
4 (1 +
θ0
L2 (Ω3D ) +
FB
L2 (I;L2 (Ω3D )) ).

θB,n
W4 ≤ C B B B
From above estimates, and arguing in a similar way to Theorem 1, we can assure the existence
of a new subsequence of {bn }n∈N (still denoted in the same way), such that the associated states
∗ ∗
{(uA,n , θA,n , θS,n , θB,n )}n∈N converge to (uA , θA , θS∗ , θB
∗
) in the corresponding spaces (it is worthwhile re-
marking here that, since the exponents for the spaces Lp where the time derivatives lie are strictly higher
than 1, we can assure compactness inside the space of weakly continuous functions). As a direct consequence,
we also have that:
• KS (bn ) → KS (b∗ ), bS,A S,A ∗ A,S A,S ∗

1 (bn ) → b1 (b ), b1 (bn ) → b1 (b ),
and b2 (bn ) → b2 (b∗ ) in L∞ (Ω2D \ Ω2D
A,S A,S
B ),
• TrA,S (bn ) → TrA,S (b∗ ) in L∞ (I; L∞ (Ω2D \ Ω2D
B )).
Using above convergences – in a completely similar way to the proof of Theorem 1 – we can pass to the
limit in the variational formulations (9)–(12) for the associated states (uA,n , θA,n , θS,n , θB,n ), and achieve
∗ ∗
that the limit (uA , θA , θS∗ , θB
∗
) ∈ W1 × W2 × W3 × W4 is a solution of the state systems corresponding to
∗
the control b . Finally, the pass to the limit in the cost function is straightforward, thanks to the continuity
and boundedness of functions hA B
l , hl , for l = 1, . . . , NB , to the strong convergences in the spaces of weakly
continuous functions, and to the classical dominated convergence theorem.
So, we can conclude that the control b∗ ∈ Uad is a solution of the problem (41), with (uA ∗ ∗
, θA , θS∗ , θB
∗
)∈
W1 × W2 × W3 × W4 its associated state. 2
7. Numerical resolution of the optimal control problem
As we introduced in previous sections, the continuous control problem can be formulated as an optimal
control problem in which the control variable has a real component (related to the position of green areas),
and an integer component (related to the possible location of green areas within a set of given regions).
This section describes how to formulate the discretized control problem and how to address its numerical
solution. We begin by setting the discretized control problem.
7.1. The discretized control problem
We try to find a discrete formulation of the MINLP problem (33). Essentially, we only need to discretize
the cost function, since the control variables and the constraints are the same as in the continuous problem.
So, we only need to consider the following discretized cost function:
N
NB
1
f (b, y) = λB hB,n
l
W,l,n
(θ̄B ) + λA hA,n
l
W,l,n
(θ̄A ) (43)
NB N
l=1 n=1
0
where, for the given initial values (uA 0
, θA , θS0 , θB
0
) ∈ VhA × XhA × XhS × XhB , the elements (uA
n+1 n+1 n+1
, p A , θA ,
θS , θB ) ∈ Vh × Mh × Xh × Xh × Xh , for n = 0, . . . , N − 1, are the numerical solutions of the
n+1 n+1 A A A S B
full time/space discretization of the state systems (24)–(27) corresponding to the smoothed version of the
coefficients (28)–(29).
7.2. Computation of the gradient of the cost function
As we noted in previous sections, for each integer variable y∗ ∈ Y, we have to solve the NLP problem (40).
Since these nonlinear optimization problems have no state constraints, if we want to use the standard
algorithms for solving such problems, we need to calculate the gradient of the discretized cost functional
f (b, y∗ ) introduced in previous subsection, with respect to the real variable b. To compute this cost function
gradient δf (b, y∗ ), we have two options: using the equations associated to the adjoint states, or using
the linearized state systems. For the sake of simplicity, we will present here only the first option (in our
particular case, this option presents a simpler numerical implementation, since it avoids the solving of a
different linearized system for each time step).
In order to compute the gradient of the discretized cost function f , we need first to derive the adjoint
state systems. So, using standard techniques, we consider the set of elements {(wA n
, snA , rA
n
, rSn , rB
n
)}Nn=0 ⊂
Vh × Mh × Xh × Xh × Xh , take them as test functions in the linearized state systems, add to-
A A A S B
gether all these equations, and reorder in a suitable way the terms appearing in the sum. Thus, defining
(wA , sA , rA , rS , rB ) = 0, we have that, for each n = N − 1, . . . , 0, the element (wA
N N N N N n
, snA , rA
n
, rSn , rB
n
) must
be solution of the following adjoint state systems:
• for the adjoint state of the air velocity:

α n
wA · v dx + νA ∇wA
n
: ∇v dx + snA ∇ · v dx
Ω3D
A Ω3D
A Ω3D
A

=α (v ◦ Xn+1 ) · wA
n+1
dx − n+1
(∇uA ◦ Xn+1 )v · wA
n+1
dx
Ω3D
A Ω3D
A

− n+1
(∇θA ◦ Xn+1 ) · v rA
n+1
dx, ∀v ∈ VhA . (44)
Ω3D
A

− ∇ · wA
n
q dx = 0, ∀q ∈ MhA ;
Ω3D
A
• for the adjoint state of the air temperature:

α n
rA z dx + KA ∇rA
n
· ∇z dx + bS,A n
1 (b)rA z dγ + bW,A
1
n
rA z dγ
Ω3D
A Ω3D
A ΓA
S ΓW
B

+ bR,A
1
n
rA z dγ = α (z ◦ Xn+1 )rA
n+1
dx + bA,S n+1
1 (b)rS z dγ (45)
ΓR
B Ω3D
A ΓA
S

z
+ bA,W
1
n+1
rB z dγ + bA,R
1
n+1
rB z dγ + n+1,REF
g · wA
n+1
dx
θA
ΓW
B ΓR
B Ω3D
A

λA
NA
n+1 W,l,n+1 1
+ δhl (θ̄A ) zdγ, ∀z ∈ XhA ;
NB N
l=1 μ(ΓW,l
A )
ΓW,l
A
• for the adjoint state of the soil temperature:

α rSn z dx + KS (b)∇rSn · ∇z dx + bA,S n
1 (b)rS z dγ
Ω3D
S Ω3D
S ΓA
S

+ bB,S
1 rSn z dγ + 4bA,S n+1 3 n
2 (b)(θS ) rS z dγ = α rSn+1 z dx (46)
ΓB
S ΓA
S Ω3D
S

+ bS,A n+1
1 (b)rA z dγ + bS,B n+1
1 rB z dγ, ∀z ∈ XhS ;
ΓA
S ΓB
S
• for the adjoint state of the buildings temperature:

α n
rB z dx + KB ∇rB
n
· ∇z dx + bA,W
1
n
rB z dγ + bA,R
1
n
rB z dγ
Ω3D
B Ω3D
B ΓW
B ΓR
B

+ bS,B n
1 rB z dγ + 4bA,W
2
n+1 3 n
(θB ) rB z dγ + 4bA,R
2
n+1 3 n
(θB ) rB z dγ (47)
ΓB
S ΓW
B ΓR
B

=α n+1
rB z dx + bW,A
1
n+1
rA z dγ + bR,A
1
n+1
rA z dγ + bB,S
1 rSn+1 z dγ
Ω3D
B ΓW
B ΓR
B ΓB
S

λB
NB
W,l,n+1 1
+ δhn+1
l ( θ̄ B ) zdγ, ∀z ∈ XhB .
NB N
l=1 μ(ΓW,l
B )
ΓW,l
B
Finally, taking as test functions in above equations (44)–(47) the elements {(δuA
n+1
, δpn+1 n+1 n+1
A , δθA , δθS ,
n+1 N −1
δθB )}n=0 ⊂ Vh × Mh × Xh × Xh × Xh , and adding all together, we obtain the following expression for
A A A S B
the gradient of the discretized cost functional:
δf (b)(δb)

N NB
−1
1
= λB δhB,n+1
l
W,l,n+1
(θ̄B W,l,n+1
)δ θ̄B + λA δhA,n+1
l
W,l,n+1
(θ̄A W,l,n+1
)δ θ̄A
NB N n=0 l=1

N −1

= δbS,A
1 (b)(δb)(θS − θA )rA dγ
n n+1 n
n=0
ΓA
S

+ δbA,S
1 (b)(δb)(θA − θS
n
)rS dγ −
n+1 n
δKS (b)(δb)∇θSn+1 · ∇rSn dx
ΓA
S Ω3D
S

A,S,n+1
+ δbA,S
2 (b)(δb) (Tr (b))4 − (θSn+1 )4 rSn dγ
ΓA
S

+ 4bA,S A,S,n+1
2 (b)(Tr (b))3 δTrA,S,n+1 (b)(δb)rSn dγ .
ΓA
S
Fig. 2. Soil surface for the numerical tests.
The calculation of the gradient δf (b)(δb) using above adjoint states requires solving systems (44)–(47)
only once, which could be an advantage with respect to the computation of the gradient using the linearized
state systems, which should be solved N times (once for each step of time).
Remark 6. The obtaining of error estimates between the state systems and the discretized systems, and
error estimates between the optimal control problem and the discretized control problem are currently being
considered. However, due to the highly nonlinear character of the problem, they are not straightforward,
and exceed the objectives of this article. So, they will be addressed in a forthcoming paper of the authors,
in the spirit of the recent papers of Neizel and Vexler [22] or Casas and Chrysafinos [9] and the references
therein.
8. Numerical results
In this section we present some numerical results we have obtained for an academic (although realistic)
test. To do this, we consider a three-dimensional domain composed of a set of four equally spaced buildings,
and analyze the different solutions we have obtained solving the problem for various scenarios. The goal we
want to achieve in this section is to analyze if it is more convenient optimally locate only one large green
area or more than one (assuming that they occupy the same total area). Both the domain meshes as the
numerical solution of the problem have been developed with Freefem+ [16]. The resolution of the nonlinear
optimization problems has been performed employing the interior-point algorithm IPOPT [32]. For the time
semi-discretization we have employed the method of characteristics, and for the space one the finite element
method (P1b − P1 for the Navier–Stokes equations, and P1 for air, soil and buildings temperatures).
8.1. Space/time description of the computational domain
As we advanced above, we consider, over a soil domain of 15 m × 15 m, a symmetrically distributed set

of four buildings (each one of 3 m × 3 m) with the geometry shown in Fig. 2.
So, we have nine possible locations where positioning the green areas: three of 45 m2 (Ω2D S,l , l = 1, 2, 3),
and six of 9 m2 (Ω2DS,l , l = 4, . . . , 9). For the remaining elements of the geometric configuration we have
chosen the following functions: the height of the soil layer over a reference level HS (x, y) = 2 m, the height
of buildings HB (x, y) = 3 m, and the height of the air layer over the ground HA (x, y) = 6 m. For the
generation of the meshes associated to soil and buildings we consider two vertices per meter on the edges
of the domain, while in the case of the domain occupied by air, we consider only one vertex per meter (see
Fig. 3). With respect to the time discretization, we choose a time step size Δt = 900 s, and we solve the
problem for N = 96 time steps (corresponding to a time interval of T = 24 hours).
Fig. 3. Meshes for the air domain: 1744 vertices (left), for the soil: 4805 vertices (centre), and for the buildings: 1372 vertices (right).
Table 1
Characteristic parameters for the thermal behaviour of the materials involved in the
problem.
Coefficient Air Asphalt Green areas Buildings
Density ρA = 1.16 103 ρC = 2.11 106 ρG = 9.99 105 ρB = 2.3 106
Specific heat cpA = 1.007 cpC = 0.92 cpG = 4.181 cpB = 0.88
Conductivity αA = 0.0263 αC = 0.062 αG = 0.58 αB = 1.4
Emissivity C = 0.95 G = 0.98 B = 0.95
Albedo aC = 0.05 aG = 0.6 aB = 0.08
Remark 7. We must bear in mind that our main goal is related to the numerical solution of a nonlinear
optimization problem with mixed integer and real variables. The algorithms used to solve this MINLP prob-
lem require the numerical resolution of a large number of systems of coupled partial differential equations,
necessary in order to evaluate the cost and its gradient. That is why we consider for our tests a domain of
small size: in order to obtain results in reasonable computational times.
8.2. Coefficients and data for the state systems
Listed below are the values associated with the coefficients and data we used in the numerical solution
of the problem. All coefficients related to the thermal behaviour of the different materials appearing in the
numerical problem, as given in Table 1, were taken from the classical reference [5]. (Since the characterization
of the thermal behaviour of green areas depends largely on the characteristics of their particular vegetation,
for the sake of simplicity we considered the values associated with water to represent the green areas. This
should not be a problem as we would be dealing with the optimal location of ponds instead of gardens,
which it is a completely similar situation.)
The convective heat transfer coefficients for the corresponding interfaces were hA A A
C = 100 on ΓC , hG = 100
A B B A W A R
on ΓG , hC = 100 on ΓC , hW = 100 on ΓB , and hR = 1 on ΓB .
To compute the radiation temperatures appearing in the heat equations for soil and buildings, we assumed
that Rsw,net (x, t) = (RMsw,dir + RMsw,dif f ) σ(x, t), and Rlw,dow (x, t) = RMlw,dow σ(x, t), where, for our
particular problem, we considered RMsw,dir = 850 W m−2 , RMsw,dif f = 550 W m−2 , and RMlw,dow =
650 W m−2 (the accurate adjustment of these values can be performed using models such as those described
in [6,7]). The function σ(x, t) ∈ [0, 1] models the attenuation of the previous maximum values, taking into
account the movement of the sun: effect of shadows, night and day, and so on. In our simplified case, we
have assumed that σ(x, t) = max{sin(2πt/86400), 0}, ∀(x, t) ∈ (ΓA C ∪ ΓG ∪ ΓB ) × [0, T ] (that is, over soil
A R
and roofs we considered no attenuation due to shadows, and radiation only depending on time), and that
σ(x, t) = γ max{sin(2πt/86400), 0}, ∀(x, t) ∈ ΓW B × [0, T ] (walls of the buildings are partially shaded with
a ratio γ that, in this particular case, has been taken equal to zero).
Finally, we considered the initial values uA 0
= (0, 0) m s−1 , θA
0
= θS0 = θB 0
= 300 K, and the boundary
IN OU T −5 −1 IN SU B
conditions uA = uA = (0, 1.0 10 ) m s , θA = θS = 300 K.
8.3. Parameters associated to the resolution of the MINLP problem
In Section 5 we have formulated our realistic problem as a MINLP problem, where two different kinds of
control variables appear: a real variable b related to the position of green areas within a given sequence of
possible locations, modelled with an integer variable y. The resolution of such problems can be performed
using algorithms of the type Branch and Bound, Generalized Benders Decomposition or Outer Approxi-
mation (see, among others, [10,23,8]). In our case, since the size of the set of admissible controls for the
discrete variable Yad is relatively small, we will solve the problem by performing an exhaustive search:
y
f (b, ) = min {f (b∗ , y∗ ) : y∗ ∈ Yad }, where, for each y∗ ∈ Yad , the element b∗ ∈ Uad is the solution of the
NLP problem (40).
To solve above NLP problem, as we mentioned at the beginning of this section, we have used the interior-
point algorithm IPOPT [32]. To use this algorithm, it is necessary to have functions that calculate the
gradient of the cost functional (see Subsection 7.2) and the Jacobian of restrictions (trivial here for their
simplicity). The data used for solving the problem were as follows:
• Number of green areas: M = 1 and M = 2. We will solve the two cases in order to analyze what is the
best strategy: one large green area or two smaller ones.
• Minimum and maximum dimensions of the green zones: We implemented the constraint (38) introducing
four parameters μx,min = 1 m (minimum dimension axis x), μx,max = 10 m (maximum dimension on
the axis x), μy,min = 1 m (minimum dimension axis and) and μy,max = 10 m (maximum dimension on
the axis and), so that:
μx,min ≤ p1,k 1,k

2 − p1 ≤ μx,max , ∀k = 1, . . . , M,
μy,min ≤ p2,k 2,k
2 − p1 ≤ μy,max , ∀k = 1, . . . , M.
Thus, in constraint (38), μmin = μx,min μy,min = 1 m2 , and μmax = μx,max μy,max = 100 m2 .
• Bounds associated with the location of the green areas: Since we are using the approximation (30) for
the indicator functions, we considered the following modification of the constraints (37):

L
L
ykl (al1 + − p1,k
1 ) ≤ 0, ykl (p1,k
2 − (b1 − )) ≤ 0,
l
∀k = 1, . . . , M,
l=1 l=1

L
L
ykl (al2 + − p2,k
1 ) ≤ 0, ykl (p2,k
2 − (b2 − )) ≤ 0,
l
∀k = 1, . . . , M.
l=1 l=1
In our case, we considered = 0.5.

• The maximum dimension for the sum of the areas associated with the green zones has been μM max =
20 m2 . In order to reduce the search for possible green areas, we only considered those whose total area
exceeds μM 2
min = 15 m , which, from a practical viewpoint, involves performing the following modification
of the constraint (39):

M
min ≤
μM (p1,k 1,k 2,k 2,k
2 − p1 )(p2 − p1 ) ≤ μmax .
M
k=1
• The set of feasible discrete controls Yad : In the case of a single green zone, because of the restrictions
imposed on the minimum area of the green zones, the set of feasible discrete controls is reduced to
those associated with regions Ω2DS,l , with l = 1, 2, 3. Therefore, we must solve three NLP problems for
finding the optimal location. In the case of two green areas, we have a total of 21 possible combinations
corresponding to the pairs (Ω2D 2D 2D 2D

S,1 , ΩS,l ), with l = 3, . . . , 9, and (ΩS,3 , ΩS,l ), with l = 4, . . . , 9 (the other
combinations are not feasible since they do not respect the constraint associated with the minimum
area of all the green zones). Although not solved in this work, in the case of three green areas, we would
have a total of 64 possible combinations for the location of green areas. In this latter case, it begins to
seem necessary to use a more sophisticated algorithm rather than exhaustive search.
• In this paper, we assume that the functions hB A
l , hl , l = 1, . . . , NB , associated to the energy consumption
for each building do not depend on time. Dependence on the average temperature on the walls and the
W,l W,l 2
surrounding air is assumed proportional to the square of itself. That is, we consider hB l (θ̄B , t) = (θ̄B ) ,
W,l W,l
hA 2
l (θ̄A , t) = (θ̄A ) , for all l = 1, . . . , NB .
The values for the parameters λB and λA are:
– Numerical test 1: λB = 1, λA = 0 (we only take into account the temperature of the buildings).
– Numerical test 2: λB = 0, λA = 1 (we only take into account the temperature of the air surrounding
the buildings).
In addition, we will consider a new test (not included in previous theoretical development of this work)
in which the mean temperature of the air layer that is in contact with the ground will be optimized:
– Numerical test 3: We consider as cost function:

NA
N
1 1 n
J(b) = A,l
θA dγ,
NA N
l=1 n=1 μ(ΓS )
ΓA,l
S
with NA = 9, and where, for each l = 1, . . . , NA ,
ΓA,l
S = {(x, y, z) ∈ R : (x, y) ∈ ΩS,l , z = HS (x, y)}
3 2D
Remark 8. Computation of the gradient of above cost functional is carried out analogously to the other cost
functionals considered in previous sections. The motivation for considering this new functional lies in the
interest of minimizing the impact of UHI effect on pedestrians’ comfort. Instead of considering temperature
of the air layer near the ground, we could also consider the averaged temperature to a certain height as
done in [12], with the immediate modifications.
Finally, in Figs. 4–6 we show the results obtained in these above detailed three numerical tests. When
presenting the results for the first two numerical tests, we incorporate a graph of the evolution of the
averaged temperatures (of the walls in the first case, and of the air surrounding the walls in the second one)
corresponding to the building in which the thermal drop is more pronounced, giving a comparison of the
different thermal behaviours in the cases of 0, 1 and 2 optimally located green zones. Additionally, for the
step of time in which the radiation intensity is higher (corresponding to central step n = 48), we also show
the thermal behaviour of the walls and the soil in the first numerical test, and the thermal behaviour of the
air at the boundaries and the soil in the second and third numerical tests.
In all these figures we can see how the temperature of the walls, the air surrounding the buildings,
and the soil decreases as the number of green areas increases. With respect to the optimal locations of
these green zones, it is worthwhile remarking here that, due to the symmetries of our problem, in the
computational process we also achieve as optimal solutions the symmetric configurations of those shown
here (with analogous values of the cost functionals).
9. Conclusions
In this paper the authors have formulated, analyzed and solved an environmental control problem, posed
in a realistic 3D framework, related to the optimal location of green zones in metropolitan areas in order to
mitigate the UHI effect.
Fig. 4. Numerical results for test 1. Left to right and top to bottom, we show: (a) the mean temperatures on the walls of first
building for the cases of 0, 1 and 2 green areas; (b) the temperature of walls, at time step n = 48, without green areas; (c) the
temperature of walls with one optimally located green area; (d) the temperature of walls with two optimally located green areas;
(e) the optimal location for one green area; (f) the optimal location for two green areas; (g) the temperature of soil with one
optimally located green area; and (h) the temperature of soil with two optimally located green areas.
Fig. 5. Numerical results for test 2. Left to right and top to bottom, we show: (a) the mean temperatures of the air surrounding
fourth building for the cases of 0, 1 and 2 green areas; (b) the air temperature on the boundaries, at time step n = 48, without
green areas; (c) the air temperature on the boundaries with one optimally located green area; (d) the air temperature on the
boundaries with two optimally located green areas; (e) the optimal location for one green area; (f) the optimal location for two
green areas; (g) the temperature of soil with one optimally located green area; and (h) the temperature of soil with two optimally
located green areas.
Fig. 6. Numerical results for test 3. Left to right and top to bottom, we show: (a) the mean temperatures of the air near the soil for
the cases of 0, 1 and 2 green areas; (b) the air temperature on the boundaries, at time step n = 48, without green areas; (c) the
air temperature on the boundaries with one optimally located green area; (d) the air temperature on the boundaries with two
optimally located green areas; (e) the optimal location for one green area; (f) the optimal location for two green areas; (g) the
temperature of soil with one optimally located green area; and (h) the temperature of soil with two optimally located green areas.
In a first part of the work the authors set a mathematically well-posed formulation of the ecological
problem in terms of an optimal control problem of partial differential equations, taking as design variables
the location and size of the green areas. This model is completely analyzed from a mathematical viewpoint.
The second achievement of the paper concerns the design and effective implementation of a complete
numerical algorithm, intended for solving the fully discretized optimization problem, which includes the reso-
lution of the underlying climatic model, the computation of the adjoint state variables, and the optimization
process.
Finally, the efficiency of the method presented here is assessed by the numerical tests developed for an
academic – although realistic – case, where the authors interface the FreeFem++ software for the simulation
of air, soil and buildings temperatures, with the IPOPT algorithm for the resolution of the discretized
optimization problem.
As a summary, we can remark that our results show that, by means of a suitable choice of the locations
and sizes of the green areas, we can mitigate in a significant way the UHI effect inside the region under
study, in order to maximize the multiple vegetation benefits in controlling temperature rises.
Acknowledgments
This work was supported by funding from project MTM2015-65570-P of Ministerio de Economía y
Competitividad (Spain) and FEDER.
References
[1] M. Agarwal, A. Tandon, Modeling of the urban heat island in the form of mesoscale wind and of its effect on air pollution
dispersal, Appl. Math. Model. 34 (2010) 2520–2530.
[2] L.J. Alvarez-Vázquez, F.J. Fernández, A. Martínez, Optimal management of a bioreactor for eutrophicated water treat-
ment: a numerical approach, J. Sci. Comput. 43 (2010) 67–91.
[3] L.J. Alvarez-Vázquez, F.J. Fernández, R. Muñoz-Sola, Analysis of a multistate control problem related to food technology,
J. Differential Equations 245 (2008) 130–153.
[4] A.J. Arnfield, Review. Two decades of urban climate research: a review of turbulence, exchanges of energy and water, and
the urban heat island, Int. J. Climatol. 23 (2003) 1–26.
[5] T.L. Bergman, A.S. Lavine, F.P. Incropera, D.P. DeWitt, Fundamentals of Heat and Mass Transfer, John Wiley & Sons,
2006.
[6] R.E. Bird, A simple, solar spectral model for direct-normal and diffuse horizontal irradiance, Solar Energy 32 (1984)
461–471.
[7] R.E. Bird, C. Riordan, Simple solar spectral model for direct and diffuse irradiance on horizontal and tilted planes at the
earth’s surface for cloudless atmospheres, J. Climate Appl. Meteor. 25 (1986) 87–97.
[8] P. Bonami, L.T. Biegler, A.R. Conn, G. Cornuéjols, I.E. Grossmann, C.D. Laird, J. Lee, A. Lodi, F. Margot, N. Sawaya,
A. Wächter, An algorithmic framework for convex mixed integer nonlinear programs, Discrete Optim. 5 (2008) 186–204.
[9] E. Casas, K. Chrysafinos, Analysis of the velocity tracking control problem for the 3D evolutionary Navier–Stokes equations,
SIAM J. Control Optim. 54 (2016) 99–128.
[10] M.A. Duran, I.E. Grossmann, An outer-approximation algorithm for a class of mixed-integer nonlinear programs, Math.
Program. 36 (1986) 307–339.
[11] K. Fan, Fixed-point and minimax theorems in locally convex topological linear spaces, Proc. Natl. Acad. Sci. USA 38
(1952) 121–126.
[12] F.J. Fernández, L.J. Alvarez-Vázquez, N. García-Chan, A. Martínez, M.E. Vázquez-Méndez, Optimal location of green
zones in metropolitan areas to control the urban heat island, J. Comput. Appl. Math. 289 (2015) 412–425.
[13] N. García-Chan, L.J. Alvarez-Vázquez, A. Martínez, M.E. Vázquez-Méndez, On optimal location and management of a
new industrial plant: numerical simulation and control, J. Franklin Inst. 351 (2014) 1356–1371.
[14] V. Girault, P.A. Raviart, Finite Element Methods for Navier–Stokes Equations, Springer-Verlag, Berlin, 1986.
[15] I.L. Glicksberg, A further generalization of the Kakutani fixed theorem, with application to Nash equilibrium points, Proc.
Amer. Math. Soc. 3 (1952) 170–174.
[16] F. Hecht, New development in Freefem++, J. Numer. Math. 20 (2012) 251–265.
[17] K.J. Mann, Computer simulation of an urban heat island using finite elements, Math. Comput. Simulation 35 (1993)
203–209.
[18] A. Martos, R. Pacheco-Torres, J. Ordóñez, E. Jadraque-Gago, Towards successful environmental performance of sustainable
cities: Intervening sectors. A review, Renewable Sustainable Energy Reviews 57 (2016) 479–495.
[19] M. Metzger, Existence for a time-dependent heat equation with non-local radiation terms, Math. Methods Appl. Sci. 22
(1999) 1101–1119.
[20] P.A. Mirzaei, F. Haghighat, Approaches to study urban heat island – abilities and limitations, Build. Environ. 45 (2010)
2192–2201.
[21] M.N. Neema, A. Ohgai, Multi-objective location modeling of urban parks and open spaces: continuous optimization,
Comput. Environ. Urban Syst. 34 (2010) 359–376.
[22] I. Neitzel, B. Vexler, A priori error estimates for space-time finite element discretization of semilinear parabolic optimal
control problems, Numer. Math. 120 (2012) 345–386.
[23] I. Quesada, I.E. Grossmann, An LP/NLP based branched and bound algorithm for convex MINLP optimization problems,
Comput. Chem. Eng. 16 (1992) 937–947.
[24] T. Roubíček, Nonlinear Partial Differential Equations with Applications, Birkhäuser, Basel, 2013.
[25] T.S. Saitoh, T. Shimada, H. Hoshi, Modeling and simulation of the Tokyo urban heat island, Atmos. Environ. 30 (1996)
3431–3442.
[26] R.M. Saldanha da Gama, Existence, uniqueness and construction of the solution of the energy transfer problem in a
rigid and non-convex blackbody with temperature-dependant thermal conductivity, Z. Angew. Math. Phys. 66 (2015)
2921–2939.
[27] M. Santamouris, Heat island research in Europe – state of the art, Adv. Build. Energy Res. 1 (2007) 123–150.
[28] H. Taha, Urban climates and heat islands: albedo, evapotranspiration, and anthropogenic heat, Energy Build. 25 (1997)
99–103.
[29] R. Temam, Navier–Stokes Equations, North-Holland, Amsterdam, 1984.
[30] T. Tiihonen, Stefan–Boltzmann radiation on non-convex surfaces, Math. Methods Appl. Sci. 20 (1997) 47–57.
[31] A.K. Tornberg, B. Engquist, Regularization techniques for numerical approximation of PDEs with singularities, J. Sci.
Comput. 19 (2003) 527–552.
[32] A. Wächter, L.T. Biegler, On the implementation of an interior-point filter line-search algorithm for large-scale nonlinear
programming, Math. Program. 106 (2006) 25–57.

A 3D Optimal Control Problem Related To 2017 Journal of Mathematical Analys

Diunggah oleh

Informasi Dokumen

Judul Asli

Hak Cipta

Format Tersedia

Bagikan dokumen Ini

Bagikan atau Tanam Dokumen

Opsi Berbagi

Apakah menurut Anda dokumen ini bermanfaat?

Apakah konten ini tidak pantas?

Hak Cipta:

Format Tersedia

A 3D Optimal Control Problem Related To 2017 Journal of Mathematical Analys

Diunggah oleh

Hak Cipta:

Format Tersedia

J. Math. Anal. Appl.

446 (2017) 1571–1605

Contents lists available at ScienceDirect

Journal of Mathematical Analysis and Applications

A 3D optimal control problem related to the urban heat islands

2. Mathematical formulation of the problem

2.1. The physical domain

for any (x, y) ∈ Ω2D

• The domain occupied by soil:

• The domain occupied by buildings:

B = {(x, y, z) ∈ R : (x, y) ∈ ΩB , HS (x, y) < z < HS (x, y) + HB (x, y)}

• The domain occupied by air:

• Lower boundary of the soil:

ΓS0 = {(x, y, z) ∈ R3 : (x, y) ∈ Ω2D , z = 0}

S = {(x, y, z) ∈ R : (x, y) ∈ ΩB , z = HS (x, y)}

B = {(x, y, z) ∈ R : (x, y) ∈ ΩB , z = HB (x, y)}

• Boundary of buildings associated with walls:

B = {(x, y, z) ∈ R : (x, y) ∈ ∂ΩB , HS (x, y) ≤ z ≤ HS (x, y) + HB (x, y)}

• Upper boundary of the air:

• Wind inﬂow boundary:

A = {(x, y, z) ∈ R : (x, y) ∈ (∂Ω

• Wind outﬂow boundary:

• Lateral boundaries for the air:

HS (x, y) ≤ z ≤ HS (x, y) + HA (x, y)}

• Lateral boundaries for the soil:

2.2. The state systems

with KA the diﬀusion coeﬃcient, bS,A W,A

with KS the diﬀusion coeﬃcient, bA,S

with KB the diﬀusion coeﬃcient, bA,R

2.3. Parameters and data

• bS,A A,S S,B

– for the temperature of soil:

– for the temperature of buildings:

with σB (W m−2 K−4 ) the Stefan–Boltzmann constant.

σB S (TrA,S )4 = (1 − aS )Rsw,net (x, t) + Rlw,dow (x, t),

σB W (TrA,W )4 = (1 − aW )Rsw,net (x, t) + Rlw,dow (x, t),

• ρC and ρG , densities of asphalt and green zones.

• Finally, for the radiation temperatures we will have:

σB C (TrA,C )4 = (1 − aC )Rsw,net (x, t) + Rlw,dow (x, t),

3. Mathematical analysis of the state systems

3.1. The concept of solution of the state systems

and we look for solutions of the Navier–Stokes equations in the space:

W1 = W 1,2,4/3 (I; VA , VA ) ∩ L∞ (I; [L2 (Ω3D 3

• for the temperature of the ﬂuid we consider the following space:

and we look for solutions of the heat equation in the space:

• for the temperature of the soil we consider the following space:

endowed with the usual norm

endowed with the norm

Lemma 1. The following inclusions are compact:

• W1 ⊂⊂ [L10/3− (Ω3D A × I)] , W1 ⊂⊂ Cw ([0, T ]; [L (ΩA )] ).

• W2 ⊂⊂ L10/3− (Ω3D A × I), W2 ⊂⊂ L (I; L

W2 ⊂⊂ Cw ([0, T ]; L (ΩA )).

• W3 ⊂⊂ L10/3− (Ω3D S × I), W3 ⊂⊂ L (I; L (Γ2 )),

• for the velocity of the air:

• for the temperature of the air:

• for the temperature of the soil:

• and, ﬁnally, for the temperature of the building:

3.2. Mathematical analysis of the existence of solution

has a ﬁxed-point, where (uA , θA , θS , θB ) ∈ W1 × W2 × W3 × W4 are solutions of the following variational

• for the velocity of the air (Oseen equations):

• for the temperature of the air:

• for the temperature of the soil:

• and, ﬁnally, for the temperature of the building:

σB S (TrA,S )4 = (1 − aS )Rsw,net (x, t) + Rlw,dow (x, t),

σB W (TrA,W )4 = (1 − aW )Rsw,net (x, t) + Rlw,dow (x, t),

σB C (TrA,C )4 = (1 − aC )Rsw,net (x, t) + Rlw,dow (x, t),

• uA,n → uA in [L10/3− (Ω3D A × I)] , for any > 0 small enough,

In our case, we considered = 0.5.