You are on page 1of 8

CONSTRUCTION OF A PREDICTIVE MODEL FOR PRODUCT HEAT LOAD

DURING CHILLING USING AN EVOLUTIONARY METHOD

L. M. DAVEY AND Q. T. PHAM


School of Chemical Engineering and Industrial Chemistry,
University of New South Wales, Sydney 2052, Australia.

1. SUMMARY

Modelling the change in heat load with time is difficult for products with variable and
complex geometries and compositions, such as beef carcasses. However, accurate heat
load prediction is important for efficient design and operation of refrigeration systems. A
simple method for constructing a model to predict the time-variable product heat load
profile during chilling applications is proposed. The heat transfer parameters in the model
are found by curve-fitting using an evolutionary error minimisation method. The model
was tested against pseudo-data, generated by finite difference calculations, for simple
shapes as well as published experimental data for more complex shapes.

2. INTRODUCTION

The determination of heat load is important for optimal design and operation of food
refrigeration systems. Refrigeration plant size and capital cost both increase with
increasing heat load, making over-design an uneconomical option. The consequences of
under-design can be equally undesirable, leading to inadequate cooling of the product,
quality problems such as uncontrolled microbial growth, and increased weight loss. For
chilling and freezing processes, the product heat load is usually the largest load, and in
batch systems the heat release is a dynamic process with the "peak" heat load occurring
during the initial stages of cooling. Hence it is important to have accurate methods to
predict not only the mean product heat load, which is commonly used for refrigeration
design, but also the dynamic change in product heat load with time.

Several dynamic models1 predicting product heat load during food cooling have been
developed /1-7/. These have generally evolved from applications of eq. (1), which applies
to situations where boundary conditions between the cooling medium and the product are
of the convective kind.

d Tma
q M cp h A ( Ta Tma ) (1)
dt

A disadvantage of this approach is that eq. (1) implies that there is no internal temperature
gradient, which is only accurate for low Bi conditions, and hence gives inadequate
prediction for most food cooling situations. Predictions can be improved if an empirical
factor is included to account for the temperature gradient.

Lovatt et al.’s /6/ model provides the most accurate method for predicting product heat
load during freezing, outside finite difference (FD) or finite element (FEM) techniques.
The cooling process is divided into three stages: chilling, freezing, and subcooling. For

I.I.F. - I.I.R. - Commissions E2, E1, B1, B2 - Melbourne (Australia) - 1996 -


the chilling stage, the heat release calculation is based on the work of Pflug et al. /8/ for
predicting cooling rates for simple shapes, as extended by Cleland and Earle /9/ to account
for complex geometries, and applied to eq. (1). The model does not provide an exact
solution as only the exponential part of the first term in the Fourier series is included.
This causes under-prediction of the heat load at the start of the chilling process, when the
heat load is most important, as initially further terms in the series are significant. Also,
infinite thermal conductivity is still assumed, and the methods used for determining the
shape factor are strictly correct only for phase change processes.

The objective of the present work is to determine whether a simple but accurate ordinary
differential equation (O.D.E.) model to predict product heat load during chilling can be
developed in which the heat transfer parameters can be found using a curve-fitting method.
The model must be applicable to complex geometries, such as beef carcasses, as well as
time-variable cooling conditions which are commonly encountered in the food industry.

3. THEORY

The simplest model for predicting heat load variation with time is a lumped parameter
model, or single "CSTR" (continuously stirred tank reactor) model, in which infinite
thermal conductivity is assumed, as expressed in eq. (1). The drawbacks of this approach
have already been outlined above. To achieve more realistic predictions a model is
proposed in which a number of CSTRs or "tanks" are networked to allow heat flow
between the tanks, and between the tanks and the cooling medium. The model is
described by eq. (2) and (3).
n
d Ti
Mc , i Ka , i ( Ta Ti ) Ki , j ( Tj Ti ) (2)
dt j 1

n
q Ka , i ( Ta Ti ) (3)
i 1

If a very large number of tanks are used, with heat flow between tanks limited to adjacent
tanks only, and model parameters derived from physical theory, the method can be
equated with the FD method. To retain reasonable accuracy, while using considerably
fewer tanks than the FD method, the heat transfer parameters Mc, Ka, and K, for eq. (2)
and (3) are found empirically from experimental data using an evolutionary curve-fitting
method.

Curve-fitting involves the selection of model parameters to minimise an error function.


"Classical" optimization methods involve stepwise calculations in which the error function
moves towards its minimum value with each step, according to the local gradient.
However, because these methods follow a single search path, they commonly end up at a
local minimum instead of the global minimum. Other disadvantages include the difficulty
of the methods to handle the random errors present in measured data, and multi-variable
problems.

Pham /10/ developed a stochastic curve-fitting method to eliminate the problems outlined
above. It is based on the evolutionary principles of "reproduction", "mutation" and
"selection" of members of a randomly appointed "population" of trial points. At each
stage new trial points are created through the reproduction of two existing trial points or
by mutation. Elimination of trial points follows by selecting for the points providing the
best fit, so that the population size remains constant.

4. METHOD TESTING

The model in the form of eq. (2) and (3) was tested for simple cases using analytical
solutions and FD data. The model accurately predicted the heat load in these tests.

The model was then linked to the evolutionary curvefitting program of Pham /10/ and
tested against simple shapes for its ability to fit the FD data. A series of FD data was
generated for infinite rod shapes, which could represent a row of meat cartons stacked in a
blast freezer. Twenty seven data sets were generated varying the Biot number (Bi) in the
range 0.5 to 6.25, and the ratio of side lengths of the rods (b) in the range 1 to 10. The
objective function to be minimised by the genetic algorithm was the relative sum of the
squared error (SSE) between the curvefitted and the FD heat loads, as given by eq. (4).
The summation in eq. (4) was stopped once qFD,m /qFD,1 < 0.1, i.e. when the instantaneous
heat flow becomes less than 10% of the initial heat flow.
qFD,m/qFD,1<0.1

SSE ( q FD , m q Curvefit, m ) / q FD , m 2 (4)


m

5. RESULTS AND DISCUSSION

As the simplest possible model was required a one tank model was tried first, however
this did not provide adequate fit to the FD data. Fig. 1 shows the improvement in the fit
in moving from a one tank network to a two tank one. Quantitatively the improvement
can be assessed from the average ratio of errors, SSE2 Tank /SSE1 Tank , which was equal to
0.016, indicating an improvement in the SSE of 98.4%. As the errors were small for the
two tank case, there was no need to move to a more complex tank network.

Fig.1- Heat loads for an infinite rod (Bi= 1.56, b= 3), as curve-fitted for both a 1 and 2
tank network, and as predicted by FD calculations.
The curvefitted parameters for the two-tank network were regressed using a standard
statistical package, and the resulting equations are presented in eq. (5) to (7). The total
thermal mass Mc,Total and total heat transfer parameter Ka,Total can be calculated from the
known geometry and heat transfer coefficient. The remaining parameters needed to
describe the two-tank model, Mc,2 and Ka,2, could be calculated arithmetically from the
parameters for the first tank.

Mc , 1 Mc , Total [ 0.806 0.0306 ln b 0.00756 Bi ] (5)


2
(R 61.4%)

Ka , 1 Ka , Total [ 6.11 11.8 ( Mc , 1 / Mc , Total ) 2 0.0286 Bi 0.0578 ln b


17.0 ( Mc , 1 / Mc , Total ) 0.0375 ( 1 / Bi ) 0.00824 b ] (6)

(R 2 97.4%)

K1 , 2 Ka , Total [ 0.149 1.20 ( 1 / Bi ) 0.0013 b ] (7)


2
(R 83.5%)

Although the SSE for each of the twenty seven runs was very small, random variation in
parameters led to regression difficulties. Replicate runs showed similar SSE values for
each replicate, but large variations in the values of the curve-fitted parameters. This
reduced the precision of the regression equations, and is particularly evident in eq. (5).

Eq. (5) to (7) were used for model predictions under the same conditions as the curve-
fitting runs, to assess the error introduced from using the regression equations. The
absolute average percentage error between FD and regressed model predictions was less
than 5.5% for all conditions tested. The magnitude of the error increased with increasing
Bi, as shown in fig. 2, but was not clearly correlated to b.

Fig.2- Correlation of the average percentage error, between FD and two-tank regressed
model predictions, with Bi (correlation coefficient = -0.871).
a) Step change in Ta from 10°C to 0°C at 300
min

b) Ramp change in h as indicated

Fig.3- Heat load prediction by FD calculations and the proposed two-tank Model for an
infinite rod under time-variable conditions.

The regressed model was then used for predicting heatload under time variable cooling
medium conditions. Fig. 3 shows the results of predictions for a step change in the
cooling medium temperature and a ramp change in the heat transfer coefficient. There
appeared to be no loss in accuracy in predictions for time-variable conditions compared to
the constant condition predictions. This indicates that the model can be readily applied to
design situations with time-variable conditions even if the curvefitting data has been
collected under constant conditions. There have been no models reviewed in the literature
directly addressing the problem of time-variable conditions on heat load predictions,
although multi-stage processes are commonly used in industry for meat chilling, so this is
a useful feature of the model.
Fig.4- Heat load of cartons of tylose as measured by Lovatt et al. /7/, as curvefitted using
the tank model, and as predicted by Lovatt et al.’s /6/ O.D.E. model.

Fig.5- Heat load of a lamb carcass as measured by Lovatt et al. /7/ (Run 4), as curvefitted
using the tank model, and as predicted by Lovatt et al.’s /6/ O.D.E. model.

Although the model appeared to give good predictions for FD data generated for simple
geometries, it remained for the model to be tested against experimental data. Lovatt et al.
/7/ carried out experiments to determine the heat load profiles obtained on complex
shapes, for cartons of tylose and lamb carcasses undergoing freezing. The freezing trials
were undertaken in an experimental wind tunnel, with heat load measured using a form of
flow calorimetry. The model was modified to allow for the change in specific heat
capacity during freezing by recalculating the specific heat capacity at each temperature
change, using a differentiated form of Schwartzberg’s enthalpy model /11/. Pham et al.
/12/ published a series of data for the constants in Schwartzberg’s model for various
foodstuffs, including lamb and tylose.
Three replicate runs were carried out by Lovatt et al. for the carton tests, freezing the
tylose from an initial temperature of 8°C, with air at -21°C. Two cartons were used in
each test. The model was fitted to the experimental data using the evolutionary algorithm.
It was required to have the simplest possible model; a three tank model was found to
provide sufficient accuracy. The curve-fitting results are shown in fig. 4. It can be seen
that the curvefitting more closely follows the experimental heat load curve at the start of
the chilling than Lovatt et al.’s /6/ O.D.E. model, which was expected as only the first
term approximation to the Fourier series was used in their model, although more than one
term is initially significant. The curvefit shown in fig. 4 was carried out for Run 3 of the
experimental trials, and fits the data well. Curvefits were also carried out for Runs 1 and
2, however the total thermal mass of these runs was higher than the total thermal mass
calculated by the model for the given volume, and chosen density and enthalpy data,
hence the heat load was consistently underpredicted, although the slope of the heat load
curves were the same. The values for the tylose density and the percentage water in the
tylose were not given by Lovatt et al. /7/ and hence were assumed, which perhaps caused
the discrepancy.

Lovatt et al. /7/ carried out six lamb freezing trials with energy balances within 10% of
their expected values. After slaughter the lambs were placed in an 8°C temperature
controlled room for 6.5 hours prior to freezing. The initial temperatures of the carcasses
varied from 8 to 18°C, and the cooling medium temperature was -21°C. The curve-fitting
results for the lamb carcass data for Run 4 are shown in fig. 5. The model fits the data
well, once again providing a better fit to the heat load curve initially than Lovatt et al.’s
O.D.E. model. In fact the proposed model gives a closer fit for the whole heat load curve,
except between 1000 to 1500 min.

It may be noted that the freezing curves for the Tank model, for both the carton and
carcass runs, are wavy, while smooth curves are obtained in chilling runs. This is
probably due to the release of the latent heat of freezing from each successive tank.

6. CONCLUSION

A method has been developed for constructing empirical models to predict the change in
heat load with time during chilling. The heat transfer parameters in the model are found
by curvefitting the model to experimental data. Accurate curvefitting was possible using
the model for both simple shapes and complex geometries. The model was also successful
in predicting heat load under time-variable cooling conditions. Because the model
involves the solution of a small number (two or three) O.D.E., it is much faster than the
finite difference or finite element methods.

7. REFERENCES

1. MARSHALL S.A., JAMES R.W. Dynamic analysis of an industrial refrigeration


system to investigate capacity control. Proc. Inst. Mech. Engrs. 189 (1975) 437-
445.
2. CLELAND A.C. Simulation of industrial refrigeration plants under variable load
conditions. Int. J. Refrig. 6 (1983) 11-19.
3. WADE N.L. Estimation of the refrigeration capacity required to cool horticultural
produce. Int. J. Refrig. 7 (1984) 358-366.
4. CLELAND A.C. Experimental verification of a mathematical model for simulation
of industrial plants. Int. J. Refrig. 8 (1985) 275-282.
5. REYNOSO R.O., DE MICHELIS A. Simulation of cryogenic batch freezers. Int. J.
Refrig. 11 (1988) 6-10.
6. LOVATT S.J., PHAM Q.T., LOEFFEN M.P.F., CLELAND A.C. A new method of
predicting the time-variability of product heat load during food cooling - Part 1:
theoretical considerations. J. Food Eng. 18 (1992) 13-36.
7. LOVATT S.J., PHAM Q.T., LOEFFEN M.P.F., CLELAND A.C. A new method of
predicting the time-variability of product heat load during food cooling - Part 2:
experimental testing. J. Food Eng. 18 (1992) 37-62.
8. PFLUG I.J., BLAIDSELL J.L., KOPELMAN I.J. Developing temperature-time
curves for objects that can be approximated by a sphere, infinite plate, or finite
cylinder. ASHRAE Trans. 71 (1965) 238-248.
9. CLELAND A.C., EARLE R.L. A simple method for prediction of heating and
cooling rates of various shapes. Int. J. Refrig. 5 (1982) 98-106.
10. PHAM Q.T. (1994). Competitive evolution: a natural approach to operator
selection. In Progress in Evolutionary Computation, Lecture Notes in Artificial
Intellegence, X. YAO (ed.), Springer-Veilag, Heidelberg 956 (1994) 49-60.
11. SCHWARTZBERG H.G. Effective heat capacities for the freezing and thawing of
foods. J. Food Sci. 41 (1976) 152-156.
12. PHAM Q.T., WEE H.K., KEMP R.M., LINDSAY D.T. Determination of the
enthalpy of foods by an adiabatic calorimeter. J. Food Eng. 21 (1994) 137-156.

8. NOMENCLATURE

A product surface area (m3) n number of tanks


Bi Biot number q heat load (W)
b ratio of rod side lengths t time (s)
cp specific heat of product (J/kgK) T tank temperature (°C)
h heat transfer coefficient (W/m2K) Ta air temperature (°C)
K parameter for heat transfer between tanks (W/K) Tma product mass average
Ka parameter for heat transfer to air (W/K) temperature (°C)
Ka,Total theoretical total heat transfer to air (W/K)
m step number Subscripts
M product mass (kg) FD Calculated by finite
Mc thermal mass (J/K) differences
Mc,Total theoretical total thermal mass (J/K) i,j tank i, tank j

RESUME

CONSTRUCTION D’UN MODÈLE POUR PREDIRE LA CHALEUR DU PRODUIT


AU COURS DU REFROIDISSEMENT PAR METHODE EVOLUTIONNELLE

La modélisation de l’évolution de chaleur du produit avec le temps est difficile quand le


produit est de forme et composition compliquées, par example avec des carcasses de
boeuf. Cependant une prédiction précise est importante pour une conception et opération
efficaces du système de refrigération. On propose une méthode simple pour construire un
modèle de l’évolution de la chaleur du produit. Pour calculer les paramètres d’échange de
chaleur, on utilise la minimisation des erreurs par une méthode évolutionnelle. Le modèle
a été vérifié contre des résultats de différences finies pour des formes simples, ainsi que
des données expérimentales four des formes compliquées.

You might also like