Professional Documents
Culture Documents
net/publication/330703544
CITATIONS READS
0 486
3 authors:
Ole Sigmund
Technical University of Denmark
339 PUBLICATIONS 21,830 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Christian Lundgaard on 29 January 2019.
1
2
3
4
5
6 A density-based topology optimization methodology for
7
8
thermal energy storage systems
9
10 Christian Lundgaarda · Kurt Engelbrechtb · Ole Sigmundc
11
12
13
14
15
16
17
Received: date / Accepted: date
18
19
20
Abstract As many renewable energy resources are ment of renewable energy technologies have intensified
21
prone to an intermittent production of energy and the considerably within the last decades.
22
23 electric energy demand varies on daily and seasonal The harvesting of wind and solar energy are exposed
24 time-scales, it is critical to develop technologies which to a large amount of scientific and commercial atten-
25 can reduce the residual between the production and tion, as these technologies are considered some of the
26 the consumption of electric energy. By storing and re- most promising entrants in the green energy changeover
27 leasing thermal energy and converting energy between (Lund, 2007, Wang, 2010). As the harvesting of wind and
28 thermal and electric phases, Thermal Energy Storage solar energy relies on the fluctuating and unpredictable
29 (TES) systems can be used to reduce this residual. nature of weather systems, a focal challenge of these
30 In this paper, we present a design methodology which technologies is to minimize the mismatch between the
31
can be used to improve the performance of TES systems supply and the demand of electric power, the so-called
32
33 by distributing two materials with di↵erent thermal residual load (Laughlin, 2017).
34 characteristics in a two dimensional design space. The TES systems can be used to store a surplus of energy
35 design methodology is developed with basis in density- generated from intermittent energy production technolo-
36 based topology optimization and a transient potential gies such as wind turbines or solar panels (Agyenim et al.,
37 flow model coupled with thermal heat transfer. By solv- 2010, Kearney et al., 2003). By storing and releasing
38 ing a sequence of design problems, important model and thermal energy and converting energy between thermal
39 optimization parameters are identified and the perfor- and electric phases, Thermal Energy Storage (TES) sys-
40 mance of TES systems is increased with 46% compared tems serve as energy batteries and can therefore be used
41 to benchmark designs. to reduce the residual between the consumption and
42
production of electric power.
43
44 Compared to energy storage systems such as chemi-
45 1 Introduction cal batteries, TES systems have several advantages for
46 which reason they have been seen in several applica-
47 tions such as cooling and refrigeration (Chen et al.,
48 2009), solar power plants (Gil et al., 2010), solar cook-
Motivated by the concerns about anthropogenic cli-
49 ing (Muthusivagami et al., 2010) and seasonal energy
50 mate changes, health impacts of particle pollution and
storage (Novo et al., 2010). For a detailed introduction
51 diminishing fossil fuel reserves, the surge for the deploy-
to TES systems see the work of McTigue (2016) and
52 a,c
Department of Mechanical Engineering the references therein.
53
54 Techinical University of Denmark, Nils Koppels Allé 414, DK- A packed-bed TES systems can with reference to
2800 Kgs. Lyngby, Denmark Figs. 1a and 1b be decomposed into four major parts:
55 E-mail: chrlund@mek.dtu.dka , E-mail: sigmund@mek.dtu.dkc
56 b
Department of Energy Conversion and Storage
(1) porous thermal storage mixtures composed of e.g.
57 Techinical University of Denmark, Frederiksborgvej 399, DK- gravel/rocks with relatively high permeability and sand
58 4000 Roskilde, Denmark with relatively low permeability. These material phases
59 E-mail: kuen@dtu.dk are in the figure colored with and and com-
60
61
62
63
64
65
2 Christian Lundgaard
1 ary conditions, the architecture of the charging and separately identified heat transfer types: conductive heat
2
discharging phases and the choice of objective function, transfer and convective heat transfer.
3
4 we are convinced that the “optimal” topology of the The density or compressibility of the fluid are as-
5 storage matrix is complex. Due to the complexity of the sumed constant or negligible, respectively. As the length-
6 design problem, we believe that a systematic material scales of the storage mixtures are small and the transfer
7 distribution approach is best suited for determining the of thermal energy between the heat transfer fluid and
8 topology of the sand and gravel/rocks in the storage the storage mixtures is e↵ective, it is reasonable to as-
9 matrix. With point of departure in this motivation and sume that the local temperature of a small region of the
10 hypothesis, we have derived, implemented and presented re-generator is constant (Laughlin, 2017). As there are
11 a density-based topology optimization approach for TES no temperature di↵erence between the storage matrix
12 systems in the present study. and the heat transfer fluid for a very small and local
13
Topology optimization originates from solid mechan- region of the re-generator, the temperature field can
14
15 ics (Bendsøe and Sigmund, 2003, Bendsøe and Kikuchi, be adequately modeled with a single set of tempera-
16 1988) but has been applied in many di↵erent types of ture equations, the so-called volume averaging technique
17 multi-physical problems in resent decades such as natu- (Whitaker, 1986). By further limiting the physical mod-
18 ral convection (Alexandersen et al., 2014), reacting flows eling to linear/constant material parameters, the conti-
19 (Okkels and Bruus, 2007), thermoelectric energy conver- nuity of thermal energy and momentum in the fluid are
20 sion (Lundgaard and Sigmund, 2019), fluid-structure- in an arbitrary domain, ⌦, given by:
21 interaction (Yoon, 2010) and many more. We suggest
22 ⇢r · u = F in ⌦ (1a)
the interested reader to consult the work of Deaton
23 @T
and Grandhi (2014) who provide a comprehensive lit- ⇢cp + ⇢cp u · rT = kr2 T + P in ⌦ (1b)
24 @t
25 erature study for the multidisciplinary scientific work ⇥ ⇤
26 in topology optimization. Transient topology optimiza- where ⇢ the mass density kg/m3 , r is the partial
27 tion problems have been studied for flows (Deng et al., derivative with respect to spatial directions x and y
28 2011, Nørgaard et al., 2016), crass-worthiness (Pedersen, [1/m], u is the velocity vector [m/s], F is the volumetric
29 2004), wave propagation problems (Dahl et al., 2008), body force [N], cp specific heat capacity [J/kg · K], t is
30 thermally actuated compliant mechanisms (Li et al., the time [s], T is the temperature [K], k is the thermal
31 2004) and many more. The framework presented in the conductivity [W/K · m] and P is the internal heat gen-
32 present paper is transient and multi-physical and is eration [W/m3 ]. The boundary conditions of Eq. (1) are
33 therefore related to the works listed above. given by:
34
35 A density-based topology optimization methodology
Fixed pressure P = c1 (2a)
36 for steady-state potential flows coupled with thermal
transport was presented in the work of Zhao et al. (2018). Fixed temperature T = c2 (2b)
37
38 The design methodology presented in the present pa- Stress free surface n·P=0 (2c)
39 per is related to this work, but is conceptually very Thermal insulation n·Q=0 (2d)
40 di↵erent with respect to design problems and modeling
41 approaches. Where Zhao and coworkers aimed at de- where n is a vector normal to the boundary at which is
42 veloping a computationally efficient alternative to solve boundary condition is imposed, c1 and c2 are numbers
43 the full turbulent thermally coupled Navier-Stokes equa- larger or equal 0.
44 In Eq. (1), the fluid pressure and fluid velocity are
tions, the methodology presented here is concerned with
45 related via the following equation:
transient design problems in porous media and e↵ective
46
47 thermal energy storage systems. u= rp (3)
48 µ
⇥ ⇤
49 where is the permeability m2 , µ the dynamic viscos-
50 2 Physical model ity [Pa · s] and p the pressure [Pa].
51 The heat flux, Q, is simply the sum of the convective
52
With reference to Fig. 1a the material phases in the stor- heat transfer and the conductive heat transfer:
53
54 age matrix are porous. As the inertia forces of fluid flows
Q = ⇢cp u · rT kr2 T (4)
55 in porous medias are negligible, a potential flow model
56 is used to model the fluid flow. Potential flow models As a considerable amount of abbreviations and vari-
57 are equivalent to Darcy’s law for flow in porous medias ables are used throughout the paper, we have for read-
58 (Whitaker, 1986). The potential flow model is coupled ability purposes listed the most used ones in Tab. 1.
59 to a heat transport model via the contribution from two
60
61
62
63
64
65
4 Christian Lundgaard
1 Table 1: List of important variables used throughout introducing the design field dependent interpolation
2 the paper. functions (Cook et al., 2007).
3
4 The most general form of the finite element residual
Variable Description equation is given by:
5
E
6 Eastern boundary
W
Western boundary R =M(x, S(x, t))Ṡ(x, t)
7
8 # + K (x, S(x, t)) S(x, t) F (S(x, t)) = 0 (5)
T⇤ Temperature boundary condition
9 #
P⇤ Pressure boundary condition
10 where R is the residual vector, M is the mass matrix, x
#=E Eastern boundary, E is the design variable field, t is the time, S is the state
11 #=W Western boundary, W
12 ⇤ = ch charging phase: t 2 [T0 ; TS ] field vector , Ṡ is the time derivative of the state variable
13 ⇤ = dch discharging phase: t 2]TS ; TE ] vector, K is the system matrix, and F is the load vector.
14 The state field vector and the time-derivative of the
T0 Start of charging phase [s]
15 TS End of charging phase [s] state field vector are given by:
16 TS Start of discharging phase [s] 8 @P 9
17 TE End of discharging phase [s] ⇢ >
< >
=
18 P @t
P The pressure field [Pa] S= Ṡ = (6)
19 T The temperature field [K] T >
: @T >;
20 S S = {P, T}T @t
21 t Time [s]
22 Ṡ Ṡ = @S/@t
where P is the pressure and T is the temperature. The
23 u The velocity field [m/s] discretized system of equations in Eq. (5) can be written
24 Qx , Qy The thermal heat flux field as:
25 in x and y, respectively ⇢ ⇢
x The design field [-] RP P 0 0 Ṗ
26 =
RT T 0 MT T Ṫ
27 fQ Heat flux objective function
⇢
28 ⌦ Design and computational domain KP P 0 P
29 Lx Length of ⌦ in x 0 KT T + C T T + L T T T
30 Ly Length of ⌦ in y ⇢
31 FP P
N Index set for finite elements in ⌦ =0 (7)
32 FT T
M Index set for time steps in t 2]TS ; TE ]
33 O Index set for x-directional heat fluxes on E
where the element system matrices are given by:
34
35 XZ
36 MT T (x) = ⇢(x)cp (x)NT N dV (8a)
⌦e
37 3 Finite element formulation i2N
XZ (x) T
38 KP P (x) = B B dV (8b)
39 The topology optimization approach takes basis in the ⌦e µ
i2N
40 idea of spatially distributing two di↵erent material phases XZ
41 (Material A or Material B) in a two dimensional design KT T (x) = k(x)BT B dV (8c)
42 space, ⌦, in order to optimize for a specified perfor- i2N ⌦e
43
mance measure. To cast the equations in Eqs. (1a)-(1b) XZ
44 CT T (x, P ) = ⇢(x)cp (x)NT uT B dV (8d)
45 on a form which is suitable for density-based topol- i2N ⌦e
54 P
55 The discretized finite element equations can now i2N denotes a standard finite element assembly proce-
56 be obtained by multiplying the strong forms of the dure for element i over the index set of elements to be
57 equations in Eq. (1a)-(1b) with suitable test functions; assembled, N; N is the matrix of element shape functions
58 integrating over the domain; performing integration and B is the derivative of N namely, B = rx N, where
59 by parts of higher dimensions on relevant terms; and rx denotes the di↵erentiation operation with respect
60
61
62
63
64
65
A density-based topology optimization methodology for thermal energy storage systems 5
42 and Petersson, 2003): where ... denotes the remaining terms in Eq. (15).
43 The governing partial di↵erential equations in Eqs.
44 (⇤B ⇤A )(1 x)(1 + #)
⇤(x) = ⇤A + (12) (1) are time dependent for which reason a time stepping
45 1 x+# scheme is required to solve the transient response of
46
47 ⇤ and # can take the entries ⇤ 2 {⇢, cp , , k} and # 2 the heat transfer problem. An implicit second order
48 p⇢ , pcp , p , pk and denote the material parameters for backward di↵erentiation formula scheme is used:
49 Material A and B in Fig. 1 and the interpolation function
50 (M + tKn ) Sn+1 = tF + MSn (17)
penalization parameters, respectively. The symbols A
51 and B refer to the material parameters for Material A where n [-] denotes the time step and t [s] denotes the
52
and B, respectively. Intermediate design variables in time step length.
53
54 Eq. (12) may not be physically meaningful, however As the pressure is dependent on the temperature
55 if design variables take values in the extremes of the but the pressure is independent of the temperature, the
56 set, i.e. x 2 {0, 1}, the design variables are physical governing equations in Eqs. (1) are weakly coupled. The
57 meaningful. unified architecture of the system of equations in Eqs.
58 The derivatives of the residual equation in Eq. (5) (8) is computationally inefficient compared to the corre-
59 with respect to state variables are required and given sponding separated architecture where the temperature
60
61
62
63
64
65
6 Christian Lundgaard
1 and pressure equations are solved separately. However, where ⇤T denotes the matrix or vector transpose op-
2
the computationally efficient form can easily be obtained eration. The derivative of the objective function with
3
4 by rewriting Eq. (8). We have decided to present the respect to the design field is given by:
5 design methodology in the unified architecture, as this Z TE
df @G @G @S
6 form straight forwardly supports extensions to more = + +
7 advanced physical modeling such as temperature depen- dx TS @x @S @x
!#
8 dent material parameters and compressible fluid. T @ R @ R @ S @ R @ Ṡ
9 + + + dt (21)
@x @S @x @ Ṡ @x
10
11 @⇤
4 Topology optimization where d⇤
d⇤ denotes the total derivative and @⇤ denotes
12
the partial derivative.
13
The optimization problem solved in the present study By rearranging, performing integration by parts on
14
15 can in general form be written as: relevant terms and collecting all terms with @S/@x, Eq.
16 (21) can be rewritten to:
17 min f Z TE
x
18 df @G @R
˜, S) = 0
s.t. R(x̄ (18) = + T dt+
19 dx TS @x @x
20 0 x 1 8x 2 ⌦D Z TE
@G @R ˙ T @ R @ S dt
21 + T
where f is the objective function, R is the residual TS @S @S @ Ṡ @x
22
˜ is the filtered and projected design variables
vector, x̄ T
23 @R @S
24 and x is the design variables. + T (22)
@ Ṡ @x 0
25
26 After discretization of the time integrals, Eq. (22) can
27 4.1 Adjoint sensitivity analysis be written as a time-dependent adjoint problem. The La-
28 grangian multipliers, (t) and ˙ (t), are required in each
29 The gradients of the objective function, f , with respect time-step to ensure that the terms containing @S/@x
30 to the design variable field, x, are required in order in Eq. 22 are zero. This results in an adjoint problem
31 given by:
to solve the optimization problem in Eq. (18). With L
32
33 being the general Lagrangian functional, the sensitivities,
@G @R ˙ T @R
34 dL/ dx, are computed by the discrete adjoint approach = T (23)
@S @S @ Ṡ
35 (see Bendsøe and Sigmund (2003), Dahl et al. (2008)
36 and the references therein). Assuming that and ˙ in Eq. (23) are known, the
37 The adjoint sensitivity analysis is carried out in a sensitivities in Eq. (22) reduce to:
38 di↵erentiate-then-discretize approach though this ap-
39 df X @G @R
proach may provide inconsistant sensitivities. As the = ( )T dt (24)
40 design problems have many degrees of freedom, we argue dx @x @x
j2M
41 that the error is neglectiable as discussed in the work
42 where j is the j’th time-step in the index set containing
of Jensen et al. (2014).
43
44 The objective function, f , in Eq. (18) is given by: Ptime integral limits, M. Or
the time-steps enclosed by the
explained in an other way: j2M is simply the discrete
45 Z TE
RT
46 way of writing the time-integral, TSE dt.
f= G(x, S(x, t), Ṡ(x, t)) dt (19) The sums with the index set in Eq. (24) can be
47 TS
48 exchanged by the vector LM T such that the equation
49 where G is a di↵erential function and in this work called can be written in what we call the implementation form:
50 the inner objective function. The limits of the integral,
51 df @G @R
TS and TE , specify the beginning and the end of the time = LM T ( )T (25)
52 interval at which the objective function is optimized (the dx @x @x
53 discharging phase). The objective function in Eq. (19) is
54 LM T is a vector with length N consisting of zeros except
augmented with the product of the Lagrange multipliers, for the entries of the time steps enclosed by t 2]TS ; TE ]
55
(t), and the residual, R: which have the value 1/M . The variables, M and N , are
56
57 Z TE
the number of time steps enclosed by t 2]TS ; TE ] and
⇥ ⇤
58 f= G+ T
(R) dt (20) t 2 [0, TE ], respectively. As functions cast in the imple-
59 TS mentation form simplify the implementation complexity
60
61
62
63
64
65
A density-based topology optimization methodology for thermal energy storage systems 7
1 in a scripting languages such as Matlab og Python, we LO T is a vector consisting of zeroes except for the po-
2
prefer to provide implementation details in this form. sitions k 2 O which have the value 1/ly,k , where ly,k is
3
4 Aside from the variations of the inner objective func- the height of the k’th element.
5 tion, we still require the derivative of the residual vector We need the gradients of the objective function with
6 with respect to the design variables, @R/@x, to know respect to the design field, @G/@x to compute the sen-
7 all terms in Eq. (25). This term is given by: sitivities in Eq. (25). These are given by:
8 8 @M 9
9 > PP @G @Q
< · Ṗ>
= = LTO (31)
10 @R @x @x @x
=
11 @x >
: @MT T · Ṫ> ; The inner objective function, G, with respect to the
12 @x state variables, S, is given by:
13 8 @KP P 9
>
< ·P >
= 8 @Q 9
14 @x >LT >
15 +
>
(26) @ G < O @P =
: @KT T · T + @CT T · T + @LT T · T> ; = (32)
16 @x @x @x @S >
:LT @Q > ;
17 O
@T
18 The derivative of the residual vector with respect to
19 the design and state field in Eqs. (26) is achieved by The derivative of the heat flux with respect to the
20 assembling each term separately, e.g.: design and state field in Eqs. (31) and (32) are simply
21 X @me obtained with a finite element assembly:
22 @MP P
· Ṗ = PP
· ṗe , ... (27) @ Q X @ qe @ Q X @ qe
23 @x @xe = = (33)
i2N
24 @x @x @S @s
i2N i2N
25 where ... denotes the remaining terms in Eq. (26).
26 The sensitivities in Eq. (24) are computed with the where qe is the element heat flux, x is the element design
27 time stepping schemes in Eq. (17). The state fields variable and s denotes the element state fields.
28 and the adjoint fields are computed in opposite di- To solve the design problem in Eq. (18) for the inner
29 rections in time. As the state fields are computed for objective function in Eq. (28), the terms in Eqs. (23),
30 j = {1, 2, ...N 1, N }, the adjoint fields are computed (25), (29), (30) and (31) are required.
31 Please notice that the terms presented in this section
for j = {N, N 1, ...2, 1}. To compute the state fields
32
and the adjoint fields are sometimes called the forward are dependent on the specific design problem. To derive
33
34 problem and the backward problem, respectively. the sensitivities for a design problem with a di↵erent
35 objective functions requires actions in Eqs. (28), (30),
36 4.1.1 Heat flux (31) and (32).
37
38 The objective of the design problem studied in this paper
39 is to maximize the x-directional heat flux on the eastern 4.2 Filters and Projection Strategy
40 boundary, E , during the discharge phase dch , see Figs.
41 The physical design variables, ⇢˜¯i , are used in the finite
1b and 1c.
42 element analysis and are obtained by the projection
The inner objective function, G = G(x, P, T ), for
43 (Wang et al., 2011)
44 the heat flux objective function is given by:
45 Z tanh( ⌘) + tanh( (˜
⇢i ⌘))
G= Qx dS (28) ⇢¯˜i = (34)
46 E
tanh( ⌘) + tanh( (1 ⌘))
47
48 where x refer to the x-directional heat flux. Eq. (28) can where ⌘ is the projection filter threshold. The filtered
49 be rewritten to: design variables ⇢˜i are obtained from the mathematical
50 X design variables, ⇢i , by the filter operation:
51 G= Q (29) P
52 j2N w(xj )vj ⇢j
k2O ⇢˜i = P i (35)
53 j2Ni w(xj )vj
54 where k is the x-directional heat flux for the k’th element
55 on E in the index set O and Q is the heat flux field. where vj is the area of the j’th element, Ni is the index
56 The objective function in Eq. (29) can be written in set of the design variables which are within the radius R
57 implementation form as: of design variable i, w(x) is the filter weighting function
58 and xi and xj are the spatial location of elements i and
59 G = LO T Q (30) j. The filter weighting function is given by:
60
61
62
63
64
65
8 Christian Lundgaard
1 ⇢ N ✓ ◆
R |x| 8|x| R ^ x 2 ⌦D X 2cp ⇢T E L2 tkn2 ⇡ 2 kn2 ⇡ 2 t
2 w(xj ) = (36) T (x, t) = 1+ +e L2
0 otherwise kn3 ⇡ 3 L2
3 n=1
4 ⇣ n⇡ x ⌘i
5 where R is the filter radius, |x| = xi xj . sin
L
6 Finally, gradients with respect to design variables, (38)
7 ⇢i , require a transformation of the sensitivities by the
8 chain rule: The exact solution to Eq. (38) is achieved for N ! 1.
9 By comparing the analytic solution evaluated for
10 @L X @f @ ⇢¯ ˜j @ ⇢˜j N = 1000 with the finite element model in Fig. 2, we
11 = ¯ (37) point out that the modeling approaches almost provide
@⇢i @ ⇢
˜ j ˜j @⇢i
@ ⇢
12 j2⌦ D
perfectly equivalent result. With reference to the excel-
13 lent fit between the modeling approaches, we confidently
14 conclude that the convective heat transfer problem is ad-
15 5 Implementation equately resolved with the finite element model proposed
16
in Sec. 3.
17 To validate the finite element model and the sensitivities
18 are critical parts of developing a topology optimization
19
methodology. In this section we have validated the finite
20 1.2
21 element model in Sec. 5.1, validated the sensitivities in
22 Sec. 5.2, listed important implementation details in Sec. 1
Numerical solutions
23 5.3 and commented on the reproducibility of the design Analytic solution
0.8
24 methodology in Sec. 10.
25 0.6
T [K]
26 t=1.00 s
0.4
27 t=0.1 s
5.1 Validation of the finite element model
28 0.2
29 t=0.01 s t=0.5 s
30 The finite element model laid out in Sec. 3 is validated 0
31 by separately considering the convective and conductive
-0.2
32 heat transfer in two di↵erent and analytically solvable 0 0.5 1 1.5 2
33 problems. x [m]
34 By extending the inlet and outlet over the entire side
35 lengths of the storage matrix in Fig. 1a, the y-directional Fig. 2: A comparison between the analytical and numer-
36 gradients of the pressure and temperature field can be ical models for conductive heat transfer. The tempera-
37 ture, T , is plotted as function of the spatial coordinate
neglected and the corresponding one-dimensional version
38 x for di↵erent times t. Due to the excellent fits between
of the design problem is achieved.
39 the models, we conclude that the heat transfer due
40 The validation study for the conductive and convec-
tive heat transfer takes basis in this problem and the to conduction is adequately resolved in the numerical
41
42 study is carried out for the following academic model model.
43 parameters: ⇢ = 2, cp = 0.5, k = 1 (conduction), k = 0
44 (convection) and L = 2.
45
46 5.1.2 Convection
47 5.1.1 Conduction
48 The time-dependent convective heat transfer response
49 The time-dependent conductive heat transfer response in Eq. (1b) can be solved analytically by assuming that
50 in Eq. (1b) can be solved analytically by using the the transition in temperature between the hot heat
51 separation of variables approach for boundary value transfer fluid and the initial temperature of the storage
52 matrix remains perfectly abrupt (McTigue, 2016). This
problems of partial di↵erential equations (Strauss, 2007).
53 transition is from now on denoted the transition zone.
54 Assuming that the boundary and initial conditions
of the conductive heat transfer validation study are By setting the thermal conductivity to zero, it is
55
T (0, t) = T E = 1, T (L, t) = 0 and T (x, 0) = 0 , the possible to determine the position of the thermal front,
56
57 temperature, T = T (x, t), as function of the spatial s, as function of time, t, by:
58 coordinate, x, and the time, t, is given by the following
s(t) = pt (39)
59 expression: µ
60
61
62
63
64
65
A density-based topology optimization methodology for thermal energy storage systems 9
T [K]
7 coordinate, x, for di↵erent times, t. This analytic rela- t=0.01 s Numerical solutions
8 tionship has been compared with the numerical model 0.4
Analytic solution
9 in Fig. 3. With reference to the figure, it is seen that
0.2
10 the numerical model provides solutions with consider- t=0.1 s
11 ably larger extensions of the transition zones than the 0
12 analytic model. This issue is caused by a compromise
13 -0.2
between three conflicting modeling technicalities. 0 0.5 1 1.5 2
14
15 The stabilization term in Eq. (8e) is basically an x [m]
16 artificial conduction-contribution, which is added to the
thermal problem in Eq. (1b) in order to damp non- Fig. 3: A comparison between the analytic and numerical
17
18 physical oscillations in the temperature field for convec- models for convective heat transfer in porous media. The
19 tion dominated problems. The non-physical temperature figure shows the temperature, T , plotted as function of
20 oscillations are successfully removed with this approach the spatial coordinate, x, for di↵erent times, t. Due to
21 but on behalf of an increased extent of the transition the acceptable fits between the modeling approaches, we
22 zone. conclude that the convective heat transfer is adequately
23 Implicit time-stepping schemes are generally prone resolved by the numerical model.
24
to numerical damping (Cook et al., 2007). Numerical
25
26 damping in convective heat transport problems is ex- framework. A validation study of the sensitivities is per-
27 pressed as an artificial conduction-contribution and an formed by comparing the analytic adjoint sensitivities
28 increased extent of the transition zone. Explicit schemes with a finite di↵erence approximation to the sensitivities.
29 are generally better suited for capturing large gradients The error in element i between the analytic, dL/ dxi ,
30 in the state fields, such as a narrow transition zone, how- and approximate sensitivities, dLapp / dxi , is given by:
31 ever, these schemes su↵er from non-physical oscillations
32 in the temperature field. To avoid these oscillations, dL/ dxi dLapp / dxi
33 ✏i = (40)
considerably more stabilization is required which result dL/ dxi
34 in a transition zone somewhat equivalent to the implicit
35 where ||⇤|| denotes the absolute value of ⇤. The approx-
time stepping schemes.
36 imation to the sensitivities for element i is computed
37 Implicit time stepping schemes are unconditionally
with the following finite di↵erence scheme:
38 stable and computationally cheep memory-wise for large
39 time steps, however large time steps reduce non-physical dLapp,i f (xi ) f (xi + x)
= (41)
40 oscillations in the temperature field but on behalf of an dxi || x||
41 increased amount of artificial conduction.
where x denotes a small perturbation in the design
42 The numerical examples in this paper are based on
43 field. To achieve the approximate sensitivity field it is
a comprehensive study and parameter calibration which
44 required to evaluate Eq. (41) for all elements which is
take the above considerations into account and solves
45 computational expensive compared to computing the
the state problems with an acceptable accuracy.
46 analytic adjoint sensitivities in Eq. (24).
The finite element model has also been successfully
47 The error between the approximate and analytic
48 verified for convective and conductive heat transfer with
sensitivities for a sequence of di↵erent perturbations
49 the commercial finite element software, COMSOL. With
in the design field, x = 10 1 , 10 2 , ..., 10 7 , 10 8
50 reference to the validation studies in Sec. 5.1.1 and
have been plotted in Fig. 4. The validation study is
51 5.1.2 and the successful COMSOL validation study, we
conducted on all elements (for a coarsely discretized
52 conclude that the finite element model is reliable.
problem), a design field with randomly chosen design
53
54 variables and the design problem sketched in Fig. 1. As
55 the validation study is conducted for all elements, each
56 5.2 Validation of sensitivities curve on the figure represents the convergence of error
57 for one element. To validate all elements is important
58 To validate the sensitivities is a critical part of the de- because an error in the sensitivity analysis may not
59 velopment of a topology optimization methodology and necessarily be expressed in all elements.
60
61
62
63
64
65
10 Christian Lundgaard
1 As we observe excellent convergence of the errors, Table 2: A list of model parameters used to solve the
2 design problem.
we confidently conclude that the sensitivity analysis is
3
4 correctly derived and implemented.
Material parameters
5
(rocks/gravel)
6
7 Heat capacity cp = 2.267 [J/kg·K]
8 Conductivity ks = 0.117 [W/K·m]
Viscosity µs = 6.10 · 10 4 [Pa·s]
9 Density ⇢s = 846.72 [kg/m3 ]
10 Permeability s = 10 7 [m2 ]
11
12 Material parameters
13 (sand)
14 Heat capacity cp = 900 [J/kg·K]
15 Conductivity kw = 0.1 [W/K·m]
16 Viscosity µw = 6.10 · 10 4 [Pa·s]
17 Density ⇢w = 2500 [kg/m3 ]
Permeability w = 10 11 [m2 ]
18
19 Dimensions
20 of design problem
21 Fig. 4: The error, ✏, between the analytic and approxi- Length of storage matrix, x Lx = 2 [m]
22 mated sensitivities plotted as function of the magnitude Height of storage matrix, y Ly = 1 [m]
23 Number of elements, x 200 [-]
of the perturbation in the design field, x. The analytic
24 Number of elements, y 100 [-]
25 and approximated sensitivities are computed with the Number of time steps 144 [-]
26 adjoint sensitivity analysis in Sec. 4 and a backward Time step length t = 25 [s]
27 finite di↵erence approximation in Eq. (41), respectively.
Parameters for the
28 Despite the time-dependence, weakly coupled physics charging phase
29 and a randomly fluctuating design field, the errors have E
Temperature Tch = 400 [K]
30 a linear convergence rate (in a logarithmic scale) and W
Pressure Pch = {3500,
31 a small magnitude for all elements which indicate that 6000, 9000} [Pa]
32 the sensitivity analysis is correctly derived and imple- W
Pch =0 [Pa]
33 mented. Start of charge phase T0 = 0 [s]
34 End of charge phase TS = 1800 [s]
35
Parameters for the
36
discharging phase
37
W
38 Temperature Tdch = 300 [K]
5.3 Design problem details Pressure W
Pdch = {3500,
39
6000, 9000} [Pa]
40 The material parameters of the sand and rocks/gravel E
Pdch =0 [Pa]
41 Start of discharge phase TS = 1800 [s]
in Fig. 1 and Tab. 2 are the temperature independent
42 End of discharge phase TE = 3600 [s]
43 version of the temperature dependent material param-
44 eters numerically and experimentally validated in the
45 work of Bruch et al. (2014). The material parameters are
46 suited for the design problem solved in the present pa- µ and ⇢. Please notice that the color map used to present
47 per for the following reasons: (1) the assumption about design solutions is chosen such that blue corresponds to
48 incompressibility in Eq. (1a) is adequate, as the thermal gravel/rocks and yellow corresponds to sand.
49 heat transfer fluid is in the liquid phase (oil) where the
50 bulk modulus is large. (2) As the di↵erence between the The Heaviside projection parameter, , in Eq. (34)
51 extreme limits of the temperature is approximately 100 is updated every 100th design iteration after the scheme:
52 = {4, 8, 16, 32, 64, 128}. The design process is stopped
K (Bruch et al., 2014), the importance of taking tem-
53 when = 128 and the design process is converged,
54 perature dependent material parameters into account
may be less critical. Furthermore, we argue that the i.e. when the maximum di↵erence between the design
55 variables in iteration k and k 1 is less than 0.1%.
56 assumption about temperature independent material
57 parameters is acceptable, as the di↵erence between the The interpolation function parameters in Eq. (12)
58 extreme limits of the material parameters in the relevant are pk = 1, p = 10 4 , pcp = 1 and p⇢ = 1. These param-
59 temperature range are: 14%, 4%, 55% and 9% for cp , k, eters have large influence on the well-posedness of the
60
61
62
63
64
65
A density-based topology optimization methodology for thermal energy storage systems 11
1
2
3
4
5
6
7
8
(a) P = 3500 (b) P = 6000 (c) P = 9000
9
10 Fig. 5: Design solutions solved for di↵erent magnitudes the pressure driving the flow during the charging and
11
discharging phases, P . The design solutions depend on P which indicate that this is an important model
12
13 parameter for the material layout of sand and gravel/rocks material phases in the storage matrix of TES systems.
14
15
phase and by conduction in the sand phase. As the ratio is smooth and stable despite the time dependent and
16
17 between the convective and conductive heat transfer weakly coupled physics. The discontinuities in the con-
18 is large, the surface area of the transition between the vergence plot are caused by the updates of the threshold
19 material phases is large. This feature is used to e↵ectively projection parameter, , in Eq. (34).
20 absorb and extract the thermal energy in the sand phases
21 during the re-generation cycle.
22 As the heat flow paths in the charging and discharg-
23 ing phases are di↵erent, it is necessary to model the
24 full cycle to fully capture the features of the underlying
25
physics.
26
27
28 To further validate the design solutions, we have
29 compared the objective functions for design solution in
30 Fig. 5c and a benchmark design consisting exclusively
31 of the gravel/rocks material phase. As the design con-
32 sisting exclusively of gravel/rocks perform better than
33 the design consisting exclusively of sand for the model
34
parameters in this study, we have decided to use this
35
36 design as the benchmark.
37 The temperature, T , and the x-directional heat flux
38 Qx as function of the time, t, have been plotted in Fig.
39 8 for the benchmark design and the design solution in
40 Fig. 5c.
41 By computing the objective function for the bench-
42 mark and the design solution in Fig. 5c, i.e. integrating
43
under the curves in Fig. 8 for time interval t 2]TS ; TE ],
44
45 the performance of the topology optimized design solu-
46 tion is ⇡ 46% higher than the benchmark design. The
47 better performance of the topology optimized design
48 solutions are conditioned by the more complex flow path
49 and the thermal properties of the sand phase which can
50 store more energy per unit volume compared to the
51 gravel/rocks material phase.
52
53
54 To demonstrate the behavior of the design algorithm,
55 we have in Fig. 9 plotted the normalized heat flux ob-
56 jective, fQ , as function of design iteration number, k
57 for the design solution in Fig. 5c. A sequence of snap-
58 shots of the design evolution has been plotted in Fig. 10.
59 We notice that the convergence of the design problem
60
61
62
63
64
65
A density-based topology optimization methodology for thermal energy storage systems 13
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30 Fig. 6: The temperature state fields for the design solution in Fig. 5c plotted for di↵erent time steps during the
31 charging phase.
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61 Fig. 7: The temperature state fields for the design solution in Fig. 5c plotted for di↵erent time steps during the
62 discharging phase.
63
64
65
14 Christian Lundgaard
|f|[-]
-2.2
7
T [K]
8 -2.4
40
9 Design 5c: solid lines
10 Benchmark: dashed lines
-2.6
20
11
12 -2.8
0 0 200 400 600
13
0 1000 2000 3000 k [-]
14
15 t [s]
Fig. 9: The normalized objective, f , plotted as function
16 (a) Temperature, T of the design iteration number, k, for the design solution
17 Charing phase, ch Discharing phase, dch
10 8 x 1 in Fig. 5c. The discontinuities in the convergence plot
18
19 Design 5c: solid lines are caused by the updates of the threshold projection
20 Benchmark: dashed lines parameter, , in Eq. (34).
21 0.5
22
Q x [W/m2 ]
23
0
24
25
26
-0.5
27
28
29 -1
30 0 1000 2000 3000
31 t [s]
32
(b) x-directional heat flux, Qx
33
34
Fig. 8: The temperature, T , and the x-directional heat
35
36 flux, Qx , as function of time, t, evaluated at E and W
37 for benchmark design solutions and the design solution
38 in Fig. 5c. With respect to the objective function in Eq.
39 (11), the performance of the topology optimized design
40 solution is ⇡ 46% higher than the benchmark design.
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
A density-based topology optimization methodology for thermal energy storage systems 15
1
2
3
4
5
6
7
8 (a) k = 1
9
10
11
12
13
14
15
16 (b) k = 50
17
18
19
20
21
22
23
24 (c) k = 100
25
26
27
28
29
30
31
32 (d) k = 200
33
34
35
36
37
38
39
40 (e) k = 300
41
42
43
44
45
46
47
48 (f) k = 450
49
50
51
52
53
54
55
56 (g) k = 600
57
58 Fig. 10: Snapshots of the design evolution which result
59 in the design solution in Fig. 5c.
60
61
62
63
64
65
16 Christian Lundgaard
1 3. The design solutions obtained by the topology op- Martin Philip Bendsøe and Noboru Kikuchi. Generat-
2
timization methodology outperform the benchmark ing optimal topologies in structural design using a
3
4 design solutions by 46%. We therefore confidently homogenization method. Computer methods in ap-
5 conclude that the methodology is excellently suited plied mechanics and engineering, 71(2):197–224, 1988.
6 for optimizing the storage matrix of thermal energy Thomas Borrvall and Joakim Petersson. Topology opti-
7 storage systems. mization of fluids in Stokes flow. International Journal
8 4. The study reveals new insight in physical and topo- for Numerical Methods in Fluids, 41(1):77–107, 2003.
9 logical e↵ects of thermal energy storage systems, A Bruch, JF Fourmigué, and R Couturier. Experimental
10 as it is shown that the design solutions depend on and numerical investigation of a pilot-scale thermal
11 the boundary conditions, the material parameters oil packed bed thermal storage system for csp power
12 and the objective functions. To obtain high perform- plant. Solar Energy, 105:116–125, 2014.
13
ing thermal energy storage systems, it is therefore Haisheng Chen, Thang Ngoc Cong, Wei Yang, Chunqing
14
15 critical to take these parameters into consideration. Tan, Yongliang Li, and Yulong Ding. Progress in
16 electrical energy storage system: A critical review.
The methodology can easily be extended to incor-
17 Progress in natural science, 19(3):291–312, 2009.
porate temperature dependent materials (phase change
18 Robert D. Cook, David S. Malkus, Michael E. Plesha,
materials), multiple materials phases, other objective
19 and Robert J. Witt. Concepts and applications of fi-
20 functions, boundary conditions and density variations
nite element analysis. John Wiley & Sons, 4th edition,
21 (natural convection). These studies may be conducted in
2007.
22 the future, but for now this paper serves as a demonstra-
Jonas Dahl, Jakob S. Jensen, and Ole Sigmund. Topol-
23 tion that topology optimization is a suitable methodol-
ogy optimization for transient wave propagation prob-
24 ogy for optimizing thermal energy storage systems.
lems in one dimension : DDesign of filters and pulse
25
26 modulators. Structural and Multidisciplinary Opti-
27 9 Acknowledgements mization, 36(6):585–595, nov 2008. ISSN 1615147X.
28 doi: 10.1007/s00158-007-0192-5.
29 The authors acknowledge the financial support received Joshua D Deaton and Ramana V Grandhi. A survey of
30 from the TopTen project sponsored by the Danish Coun- structural and multidisciplinary continuum topology
31 cil for Independent Research (DFF-4005-00320). optimization: post 2000. Structural and Multidisci-
32 plinary Optimization, 49(1):1–38, 2014.
33 Yongbo Deng, Zhenyu Liu, Ping Zhang, Yongshun Liu,
34 10 Replication of results and Yihui Wu. Topology optimization of unsteady
35
incompressible Navier-Stokes flows. Journal of Com-
36 As comprehensive implementation details are provided,
37 putational Physics, 230(17):6688–6708, 2011.
we are confident than the design methodology is re- Antoni Gil, Marc Medrano, Ingrid Martorell, Ana
38
39 producible, for which reason we have decided not the Lázaro, Pablo Dolado, Belén Zalba, and Luisa F
40 publish the code. If the information provided in the Cabeza. State of the art on high temperature thermal
41 paper isn’t sufficient, scientist or interested parties are energy storage for power generation. part 1concepts,
42 welcome to contact the authors for further explanations. materials and modellization. Renewable and Sustain-
43 able Energy Reviews, 14(1):31–55, 2010.
44 Jakob S Jensen, Praveen B Nakshatrala, and Daniel A
45 References
Tortorelli. On the consistency of adjoint sensitivity
46
analysis for structural optimization of linear dynamic
47 Francis Agyenim, Neil Hewitt, Philip Eames, and
48 problems. Structural and Multidisciplinary Optimiza-
Mervyn Smyth. A review of materials, heat transfer
49 tion, 49(5):831–837, 2014.
and phase change problem formulation for latent heat
50 David Kearney, Ulf Herrmann, Paul Nava, B Kelly, Rod
thermal energy storage systems (lhtess). Renewable
51 Mahoney, J Pacheco, Robert Cable, N Potrovitza,
and sustainable energy reviews, 14(2):615–628, 2010.
52 Daniel Blake, and Henry Price. Assessment of a
Joe Alexandersen, Niels Aage, Casper Schousboe An-
53 molten salt heat transfer fluid in a parabolic trough
54 dreasen, and Ole Sigmund. Topology optimisation for
solar field. Journal of solar energy engineering, 125
55 natural convection problems. Int. J. Numer. Meth.
(2):170–176, 2003.
56 Fluids, pages 699–721, 2014. doi: 10.1002/fld.
Robert B Laughlin. Pumped thermal grid storage with
57 Martin Bendsøe and Ole Sigmund. Topology Optimiza-
heat exchange. Journal of Renewable and Sustainable
58 tion - Theory, Methods and Applications. Springer,
59 Energy, 9(4):044103, 2017.
2003.
60
61
62
63
64
65
18 Christian Lundgaard
1 Ying Li, Kazuhiro Saitou, and Noboru Kikuchi. Topol- Walter A Strauss. Partial di↵erential equations: An
2
ogy optimization of thermally actuated compliant introduction. Wiley, 2007.
3
4 mechanisms considering time-transient e↵ect. Finite Krister Svanberg. The method of moving asymptotesa
5 elements in analysis and design, 40(11):1317–1331, new method for structural optimization. International
6 2004. journal for numerical methods in engineering, 24(2):
7 Henrik Lund. Renewable energy strategies for sustain- 359–373, 1987.
8 able development. Energy, 32(6):912–919, 2007. Fengwen Wang, Boyan Stefanov Lazarov, and Ole Sig-
9 Chrisitan Lundgaard, Joe Alexandersen, Mingdong mund. On projection methods, convergence and ro-
10 Zhou, Casper Schoesboe, and Ole Sigmund. Topology bust formulations in topology optimization. Structural
11 optimization for fluid-structure-interaction problems. and Multidisciplinary Optimization, 43(6):767–784,
12 Structural and Multidisciplinary Optimization, 2018a. dec 2011.
13
doi: 0.1007/s00158-018-1940-4. Zhifeng Wang. Prospectives for china’s solar thermal
14
15 Christian Lundgaard and Ole Sigmund. A density-based power technology development. Energy, 35(11):4417–
16 topology optimization methodology for thermoelectric 4420, 2010.
17 energy conversion problems. Structural and Multidis- Stephen Whitaker. Flow in porous media i: A theoretical
18 ciplinary Optimization, 57(4):1427–1442, 2018. derivation of darcy’s law. Transport in porous media,
19 Christian Lundgaard and Ole Sigmund. Design of seg- 1(1):3–25, 1986.
20 mented o↵-diagonal thermoelectric generators using Gil Ho Yoon. Topology optimization for stationary fluid–
21 topology optimization. Applied Energy, 236:950–960, structure interaction problems using a new monolithic
22 2019. formulation. International journal for numerical meth-
23 Christian Lundgaard, Joe Alexandersen, Mingdong ods in engineering, 82(5):591–616, 2010.
24
Zhou, Casper Schousboe Andreasen, and Ole Sigmund. Xi Zhao, Mingdong Zhou, Ole Sigmund, and Casper Sc-
25
26 Revisiting density-based topology optimization for housboe Andreasen. A poor mans approach to topol-
27 fluid-structure-interaction problems. Structural and ogy optimization of cooling channels based on a darcy
28 Multidisciplinary Optimization, pages 1–27, 2018b. flow model. International Journal of Heat and Mass
29 Joshua McTigue. Analysis and optimisation of thermal Transfer, 116:1108–1123, 2018.
30 energy storage. PhD thesis, Department of Engineer-
31 ing, University of Cambridge, 2016.
32 RM Muthusivagami, R Velraj, and R Sethumadhavan.
33 Solar cookers with and without thermal storagea re-
34 view. Renewable and Sustainable Energy Reviews, 14
35
(2):691–701, 2010.
36
37 Sebastian Nørgaard, Ole Sigmund, and Boyan Lazarov.
38 Topology optimization of unsteady flow problems us-
39 ing the lattice Boltzmann method. Journal of Com-
40 putational Physics, 307:291–307, 2016.
41 Amaya V Novo, Joseba R Bayon, Daniel Castro-Fresno,
42 and Jorge Rodriguez-Hernandez. Review of seasonal
43 heat storage in large basins: Water tanks and gravel–
44 water pits. Applied Energy, 87(2):390–397, 2010.
45 Fridolin Okkels and Henrik Bruus. Scaling behavior of
46
optimally structured catalytic microfluidic reactors.
47
48 Phys. Rev. E, 75:016301, 2007.
49 Claus BW Pedersen. Crashworthiness design of transient
50 frame structures using topology optimization. Com-
51 puter methods in applied mechanics and engineering,
52 193(6-8):653–678, 2004.
53 Alberto Pizzolato, Ashesh Sharma, Kurt Maute, Adri-
54 ano Sciacovelli, and Vittorio Verda. Design of e↵ective
55 fins for fast pcm melting and solidification in shell-
56 and-tube latent heat thermal energy storage through
57
topology optimization. Applied Energy, 208:210–227,
58
59 2017.
60
61
62
63
64
65
View publication stats