Professional Documents
Culture Documents
1.2 Main Contributions et al. (2013)); hence, we do not take the heat transfer
between evaporator-air and air-food into consideration.
The main contributions of this paper are as follows. Furthermore, we disregard the dynamics of the expansion
• Modelling and investigation of presented refrigeration valve, and model the evaporator assuming that cooling
system with heat recovery (Chapter II). capacity can be controlled as in Shafiei et al. (2013). We
• Because of the particular configuration of the system assume that the refrigerant mass flow ṁr can be treated
a discontinuity in the input condenser temperature as a control variable. Furthermore, we assume a super-
occurs. Smooth model of temperature both for two- heat temperature of Tsh = 10 and no sub cooling. The
phase and saturated gas regions is developped (Chap- indoor temperature Tind that takes part in the descrip-
ter III). tion of the cooling demand from the refrigeration system
• It has been shown that refrigeration system with UAmt (Tind − Tmt ) is assumed to be constant and treated
heat recovery and advanced controller, that enables as a known disturbance. Moreover, it is assumed that there
high pressure control has a negative operational cost. is no pressure loss in the pipes between the compressor and
Comparison has been made of the system behaviour over the three way valve, Phi = Pc . A vapour compression
under two objective function. First denoted as L1 cycle (VCC) system with CO2 (R744) as a refrigerant,
and named as Minimum Consumption describes the which works in the sub critical mode is examined. We
overall operational cost of electrical devices. Second constrain the system and do not allow it to work above the
objective function denoted as L2 and named as Pro- critical point. In particular, for CO2 , the critical temper-
sumption describes the operational cost together with ature Tcrit = 31.95 ◦C and pressure Pcrit = 73.8 Bar. The
the revenue in connection with reclaimed and sold gas cooler electrical fan, the water pump that produces the
heat. mass flow ṁc in the cooling loop and the compressor are
• Designing a direct collocation non-linear model pre- all assumed to be of variable speed types. Furthermore,
dictive controller (NMPC) for refrigeration system constant pressure across the water pump is assumed.
with heat recovery (Chapter IV and V).
2.2 One evaporator, one gas cooler refrigeration system
1.3 Paper Organisation
Four main processes can be distinguished that describe the
This paper is organised in the following manner. In operation of a typical refrigeration unit. The given process
Section II a non standard configuration of the refriger- numbers directly corresponds to Fig. 1 and 2.
ation system with heat waste reclaim is presented and
differences compared to a typical refrigeration system are Evaporation 4-1 The refrigerant that enters the evapo-
explained. First principal models of the system are shown. rator, evaporates absorbing heat from the surroundings.
Section III takes up the problem of the discontinuous Assuming that we use a medium temperature display case
dynamics of the gas cooler input temperature and presents its air temperature Tmt dynamics can be described as in
a way of overcoming such issues. The direct collocation (2). The control over Tmt can be facilitated by deliver-
method that has been used is described in Section IV . ing the specific amount of the cooling capacity Q̇e (3)
The construction of the objective function, taken opti- which would counteract the heat load Q̇load (4). Knowing
mization procedure and some of the results are given that the refrigerated goods need to be kept in between
in Section V . Acknowledgements and conclusions can be 1 ≤ Tmt ≤ 5 [◦ C], one has flexibility of changing the
found in Section VI and Section VII , respectively. delivered Q̇e . The Tmt bounds indirectly induce upper and
2. REFRIGERATION SYSTEMS DESCRIPTION lower limits on Q̇e and ṁr . Note, that the display case
can be viewed as a thermal storage with the capacity
described by the mass Mmt and the specific heat cpmt
Biegel and Stoustrup (2012), Biegel et al. (2014), Pedersen
et al. (2014), Rahnama et al. (2013), Shafiei et al. (2013),
Vinther et al. (2013).
cpmt Mmt Ṫmt = Q̇load + Q̇e (2)
Q̇e = ṁr (hei (Pe ) − heo (Pc )) (3)
Q̇load = U Aevap (Tamb − Tmt ) (4)
where, ṁr is the refrigerant mass flow, heo and hei are the
evaporator outlet and inlet enthalpies, respectively. Tamb
is the ambient air temperature.
Compression 1-2 Saturated gas that enters the com-
Fig. 1. Schematic diagram of refrigeration system with one pressor is pressurized. The work done by the compressor
loop heat waste recovery and thermal storage (WT) Ẇcomp is given by (5). Assuming constant ∆Tsh , Ẇcomp
value depends on the evaporation and compressor output
2.1 Assumptions pressures Pe and Pc as well as the mass flow ṁr .
1
When modelling the evaporator, we assume that the tem- Ẇcomp = ṁr (hgci (Pc ) − heo (Pe )) (5)
perature of the refrigerated goods is the same as the tem- 1 − fq
perature in the display case (Larsen (2005) and Pedersen where, hgci is the enthalpy at the inlet to the gas cooler.
11750
Proceedings of the 20th IFAC World Congress
Toulouse, France, July 9-14, 2017 Tomasz Minko et al. / IFAC PapersOnLine 50-1 (2017) 11257–11262 11259
Condensation 2-3 High-pressure refrigerant flows through both of them. The ODE’s that describe the dynamics of
the gas cooler where it is being cooled down. The amount the gas cooler output enthalpy hgco (14) and air tempera-
of energy that needs to be rejected Q̇c is dictated by the ture Tgca (15) are as follows.
following first law of thermodynamics. Mgc ḣgco = ṁr (hgci − hgco ) − U Acond (Tgci − Tgca ) (14)
Q̇c = Q̇e + Ẇcomp − fq Ẇcomp (6) U Acond (Tgca − Tgci )
Mgca Ṫgca = ṁgca (Tamb − Tgca ) + ,
where, Q̇e is the cooling capacity of the evaporator, Ẇcomp cpair
is the work done by the compressor and fq is the heat loss (15)
coefficient which describes the heat fraction of the com- where ṁ
gca is a control variable and Tgci is an estimated
pressor work that is being transmitted to the surroundings temperature at the input to the gas cooler. A detailed
Sonntag, Richard Edwin and Borgnakke, Claus and Van description of the modelling approach taken for the T is
gci
Wylen, Gordon John and Van Wyk (1998) presented in chapter III.
Expansion 3-4 The refrigerant is de-pressurized through
the expansion valve. So called isenthalpic expansion dic- 2.4 Thermal storage
tates that enthalpies at points 3 and 4 are equal.
When the refrigerant flows through the HEX the heat
2.3 Refrigeration system with one loop heat waste recovery waste recovery can be obtained. Water that is pumped
through the secondary side of the HEX cools down the
The system shown in Fig. 1 contains an additional mean of refrigerant and transports the reclaimed heat to the water
cooling the refrigerant, than would standard refrigeration tank (W T ). The dynamics of the water storage tempera-
system described in 2.2 posses. An additional water cooled ture is described by the following equation:
heat exchanger (HEX), three way valve for splitting Mwt cpw Ṫwt = U Ac (Tco − Twt ) − ṁd cpw (Twt − Tind ) .
the refrigerant and an additional expansion valve have (16)
been used. This results in the following modifications of
Condensation 2-3 . The water tank temperature consist of a term that de-
scribes the heat transfer from the heat recovery water
Water Cooling 2-6 Refrigerant which has been redi- loop U Ac (Tco − Twt ) and a term ṁd cpw (Twt − Tind ), that
rected by the three way valve enters heat exchanger (HEX) describes amount of heat discharged from the water tank.
and is being cooled down. The HEX can be divided into
primary side - hot side describing the dynamics of the P [Bar]
6 5
refrigerant leaving HEX on the refrigeration system side. P'hi
hho h'hi
The ordinary differential equation that describes the dy-
namics of the enthalpy and the output of the HEX hot
side hho is as follows. hgco 3 7 hgci 2
Phi
Mho ḣho = d ṁr (hhi (Pc ) − hho ) − UAh (Thi (Pc ) − Twall ) , hhi
(7)
where, Thi and hhi are the hot side refrigerant temperature
and enthalpy, rspectively, at the output of the compressor.
Both Thi and hhi are found using regression models of
CO2 refrigerant properties. Furthermore d describes the
opening degree of the three way valve, 0 ≤ d ≤ 1. hei heo 1 Wcomp Wcomp
4
Secondary side - cold side of HEX is described by the water Pe
temperature that enters the cold side of the HEX Tci (9) h [J/kg]
and the water temperature at the outlet of HEX Tci (8).
cpw Mco Ṫco = cpw ṁc (Tci − Tco ) + UAc (Twall − Tco ) (8) Fig. 2. Vapour Compression Rankine Cycle for refrigera-
tion system with R744 (CO2 ) refrigerant with heat
cpw Mci Ṫci = cpw ṁc (Tci − Tco ) − UAc (Tco − Twt ) (9) waste recovery (d = 1 case).
cpw Mwall Ṫwall = UAh (Thi − Twall ) − UAc (Twall − Tco )
(10)
3. SUBCRITICAL CO2 PROPERTIES MODELLING
Furthermore the HEX wall temperature Twall is described
in (10) and the heat transfer coefficients U Ac , U Ah take In this paper the physical model of the systems is utilized
the following forms in combination with a model predictive method as a
U Ac = U Acnom ṁc /ṁcnom (11) means of providing supervisory control. Introduction of
U Ah = U Ahnom ṁh /ṁhnom . (12) the additional water cooled heat exchanger requires that
Three way joining and expansion valves 6-7. Because an accurate description of the systems behaviour at point
of the mass flow split at point 2, see Fig. 1 and 2, it is 2, 5, 6 and 7 is provided, see Fig. 2. The focus has been
necessary to introduce (13) that describes the enthalpy at put on creating a model of the temperature just after
the gas cooler input hgci , just after the mass flow joining. the isenthalpic expansion, after the mass flow joining Tgci ,
at point 7. Tgci exhibits significantly different dynamics
hgci = d ṁr hho + (1 − d) ṁr hhi . (13) depending on the operation point of the system. In the
Gas Cooler 7-3 The refrigerant can be cooled by either mixed liquid and gas region (L+G) Tgci is constant and
gas cooler (GC) or HEX, or the cooling can be shared by can be approximated as the linear function of enthalpy
11751
Proceedings of the 20th IFAC World Congress
11260
Toulouse, France, July 9-14, 2017 Tomasz Minko et al. / IFAC PapersOnLine 50-1 (2017) 11257–11262
TgciLIN = p1 h + p2 . In the saturated gas region the non- TNL have been acquired. The TNL regression models for
linear dynamics of Tgci has been successfully approximated different values of pressure Phi have been obtained. In
using the second order polynomial TgciNL = p3 h2 +p4 h+p5 . order to acquire a more general surface model for the
TgciLIN and TgciNL have one common point that is described saturated gas region we approximate p1 , p2 and p3 as a
by the saturation curve and this point is met only if function of pressure Phi . Furthermore, the enthalpy hsat at
Pgci = Phi (when point 2 = 5). Whenever Phi > Pgci , which the discontinuity occurs is modelled. The resulting
TgciNL will enter the mixed liquid and gas region at a models are put together in order to create a smooth non-
higher temperature than the temperature for TgciLIN . In linear model (17) that describes the temperature as a
such cases a discontinuity in temperature occurs. function of enthalpy in both regions of interest.
In order to illustrate it TgciLIN for Pgci = 50[Bar], together
with TgciNL for pressures in between 50 and 70 [Bar] as 4. DIRECT OPTIMAL CONTROL METHODS:
well as the saturation curve have been plotted in Fig. 3. DIRECT COLLOCATION
On the left from the saturation curve (black), the linear
dynamics of Tgci for mixed liquid and gas region (red) In this paper we solve an optimal control problem (OCP)
have been plotted. For Pgci = 50 TgciLIN = 14.28[◦ C] of the form (1a)-(1d) using the direct collocation method.
and is constant throughout the whole mixed liquid and Throughout the simulations, zero order hold discretization
gas region. On the right side of the saturation curve of control variables is assumed.
Tgci enters the saturated gas region where it behaves in
non-linear fashion. Assuming Phi = 70, the temperature On each discrete time interval [τi , τi+1 ] the collocation
entering the mixed liquid and gas region is 28.7[◦ C]. The method Quirynen and Diehl (2015) approximates the state
higher pressure difference the larger discontinuity jump in trajectories x(τ ) with Lagrange polynomials Pi,k (τ ) of
temperature can be expected. The saturation curve for order K. In order to approximate the dynamics ẋ =
pressures in between 50 and 70 [Bar] is approximated as F (x, u) (1b) additional time grid is introduced, such that
a second order polynomial hsat = r1 Phi 2
+ r2 Phi + r3 . {τi,0 , ..., τi,K } ∈ [τi , τi+1 ].
K
x(θi , τ ) = θi,k · Pi,k (τ ) (19)
90
k=0
P = 50
where, x(θi , τi,j ) = θi,j , j = {1, ..., K} and P is a poly-
80
P = 55 nomial basis. Note that θi,k is treated as additional opti-
70 P = 60 mization variable and can be adjusted. Furthermore the
P = 65
Temperature [ ° C]
0.25
0.4
xi
Denoting TLIN as a linear function that describe the 0.2
0.1
temperature in the mixed gas region, TNL as the function θi,0 θi,1
0.1
that describe temperature in the saturated region, one can
construct the following continuous function that describe 0.05 0
8 10 12 14 16 18 20 22
the smooth transition in between TLIN and TNL . Time [s]
11752
Proceedings of the 20th IFAC World Congress
Toulouse, France, July 9-14, 2017 Tomasz Minko et al. / IFAC PapersOnLine 50-1 (2017) 11257–11262 11261
on Fig. 4 will move towards the black line in order to 5.1 Objective Function
provide best fit. The choice of number of points defines
how well the state trajectory can be approximated. It Two objective functions L1 and L2 were taken under
is necessary to enforce the continuity on the shooting consideration.
intervals. Furthermore in order to achieve a proper fit we L1 = Φ1 + Φ3 (26)
need to impose the integration constraints (20) and (21), L2 = Φ1 − Φ2 + Φ3 (27)
which forces the derivative of the polynomial to be equal
to the value of the approximated state trajectory at the Objective function L1 is named as the Minimum Con-
equivalent points. sumption and it describes the overall operational cost
of electrical devices. The L2 is named as the Prosump-
In this paper collocation based integrator together with tion and describes the revenue in connection with re-
multiple shooting is used. Such form is often called as claimed and sold heat. The price information of heat-
direct collocation in which collocation integrator is put ing p2 is constant throughout the simulations. In Den-
together with its constraints to the NLP solver. mark, the heating prices vary depending on the region.
Its value can oscillate in between 42 and 161 [eur/M W h]
http://energitilsynet.dk (March 2015). For the purpose of
5. SIMULATIONS simulations, value of the heating price has been set to
p2 = 50;
All simulations are done using MATLAB Matlab (2015)
and the symbolic framework for algorithmic differentiation 5.2 Results
and numeric optimization CasADi Andersson (2013). The
interior point method solver Ipopt Wachter (2006) is The proposed objective functions L1 and L2 were com-
used, with the linear solver M A57. Direct collocation pared in the following manner.
method is used with three Legendre collocation points. All • A. The total power consumption described by term
simulations have been carried out with a total simulation Φ1 in (23) is compared, see Fig. 5.
time of 24 [h], a sampling time of ts = 5 [min] and a • B. The total operational cost for L1 and L2 is
prediction horizon of tp = 60 [min]. The objective function compared, assuming that in both cases reclaimed heat
is constructed such that it consists of three terms has been sold, see Fig. 6.
Φ = Φ 1 − Φ2 + Φ3 (22) The proposed construction of the L2 objective function
where, minimizes the terms Φ1 , Φ3 and maximises term Φ2 . Since
p2 has larger value than the maximum value of p1 , the
tp term Φ2 dominates the objective function. In other words,
Φ1 = (p1 (t)(Ẇcomp (t) + Ẇgcf (t) + Ẇcp (t) + Ẇdp (t))dt L2 aims at maximizing the amount of reclaimed and sold
t0 heat waste. As a result the compressor output pressure
(23) and refrigerant mass flow are kept at their maxima, and
the display case temperature at its minimum. As for L1 , it
describes the operational cost of the electrical devices, exploits the electricity price information and find optimal
tp system trajectories according to electricity price forecasts.
On average the total power consumption when using L1 is
Φ2 = p2 Q̇sold (t) dt (24)
lower by 40%. On the other hand, when assuming that the
t0 heat is sold when using both L1 and L2 objective functions,
describes the cost of the sold thermal energy, and it turns out that the system has negative operational cost.
In the case when L1 is used daily revenue is about 7.9 [eur],
tp for L2 it is 11.9 [eur].
Φ3 = (R1 (X/Xmax )2 + R2 (U/Umax )2 )dt, (25)
t0
6. CONCLUSIONS
describes the penalties for our states, controls and restric- When examining a refrigeration systems with just air cool-
tions regarding the rate of power change in our system. ing and without the heat recovery, usually the compressor
In the expressions above, Ẇcomp = β ṁr (hhi − heo ) output pressure Phi is not treated as a control variable, but
instead, it is only adjusted according to the changes in the
describes the compressor power consumption. Ẇgcf =
ambient temperature. When heat reclaim is introduced,
(κ1 ṁ2gca ) stands for the gas cooler air fan power con- the control over Phi brings up additional flexibility to the
sumption, Ẇcp = κ2 ṁc and Ẇdp = κ3 ṁd for the power system and increases the refrigeration system’s potential
consumption of two water pumps used in the heat recovery for production of heat Qc . The mechanism of heat re-
loop. p1 and p2 are the electricity and district heating covery has been explained in the paper. Moreover, the
prices. Furthermore, R1 and R2 are diagonal weight ma- continuous temperature model for two-phase and satu-
trices, X and U are matrices with elements of state vector rated gas regions has been derived. Furthermore direct
x and control vector u on the diagonal, respectively, and collocation method has been presented and used. Higher
Xmax and Umax are diagonal matrices that contain the revenue might be achieved for larger systems, with more
maximum values of states and control signals, respectively. medium display cases and low temperature freezers. Fur-
The non-convex term Ẇcomp has been replaced by a slack ther increase might be achieved by operating in the su-
variable. percritical region. More results, all constants, parameters
11753
Proceedings of the 20th IFAC World Congress
11262
Toulouse, France, July 9-14, 2017 Tomasz Minko et al. / IFAC PapersOnLine 50-1 (2017) 11257–11262
-0.15
linear predictive control of a vapour compression cycle.
-0.2 International Journal of Refrigeration, 29(5), 761–772.
doi:10.1016/j.ijrefrig.2005.12.005.
-0.25
Matlab (2015). version R2015a. The MathWorks Inc.
Pedersen, R., Schwensen, J., Biegel, B., Stoustrup, J., and
-0.3
Green, T. (2014). Aggregation and control of supermar-
L1: Cost Minimization ket refrigeration systems in a smart grid, volume 19.
L2: Prosumption
-0.35
IFAC. doi:10.3182/20140824-6-ZA-1003.00268.
0 5 10 15 20 25 Pedersen, R., Schwensen, J., Sivabalan, S., Corazzol, C.,
Time [hours] Shafiei, S.E., Vinther, K., and Stoustrup, J. (2013). Di-
rect Control Implementation of a Refrigeration System
Fig. 6. Total operational cost comparison for L1 and L2 in Smart Grid.
objective functions. Quirynen, R. and Diehl, M. (2015). Lifted implicit integra-
values together with simulation files can be found at: http : tors for direct optimal control. Conference on Decision
//kom.aau.dk/project/edge/repository/IF AC 2017/. and Control (CDC), (Cdc), 3212–3217.
Rahnama, S., Stoustrup, J., and Rasmussen, H. (2013).
7. ACKNOWLEDGMENTS Integration of heterogeneous industrial consumers to
provide regulating power to the smart grid. In Proceed-
This work was supported by the Danish Council for ings of the IEEE Conference on Decision and Control,
Strategic Research (contract no. 11-116843) within the 6268–6273. doi:10.1109/CDC.2013.6760880.
’Programme Sustainable Energy and Environment’, under Shafiei, S.E., Stoustrup, J., and Rasmussen, H. (2013). A
the “EDGE” (Efficient Distribution of Green Energy) supervisory control approach in economic MPC design
research project. for refrigeration systems. 1565–1570.
Sonntag, Richard Edwin and Borgnakke, Claus and Van
REFERENCES Wylen, Gordon John and Van Wyk, S. (1998). Funda-
mentals of thermodynamics. 6.
Andersson, J. (2013). A General Purpose Software Frame- Vinther, K., Chandan, V., and Alleyne, A.G. (2013).
work for Dynamic Optimization. Ph.D. thesis, Arenberg Learning / Repetitive Control for Building Systems with
Doctoral School, KU Leuven. Nearly Periodic Disturbances. (ii).
Biegel, B. and Stoustrup, J. (2012). Model predictive con- Wachter, A.L.T.B. (2006). On the Implementation of a
trol for power flows in networks with limited capacity. Primal-Dual Interior Point Filter Line Search Algorithm
American Control . . . , 2959–2964. for Large-Scale Nonlinear Programming. Mathematical
Biegel, B., Westenholz, M., Hansen, L.H., Stoustrup, J., Programming, 106(1), 25—-57.
Andersen, P., and Harbo, S. (2014). Integration of flex-
11754