You are on page 1of 6

Proceedings of the 6th International Conference on Process Systems Engineering (PSE ASIA)

25 - 27 June 2013, Kuala Lumpur.

Machine Learning Based Modeling for Solid Oxide


Fuel Cells Power Performance Prediction
M. N. Fuad,a M. A. Hussain,b
a

Chemical Engineering Department, Faculty of Engineering, UCSI University, 56000


Cheras, Kuala Lumpur
Department of Chemical Engineering, Faculty of Engineering, University of Malaya,
50603, Kuala Lumpur

Abstract
This study applies four different types of machine learning methods to model the power
performance behaviour of a tubular solid oxide fuel cells (SOFC) under different
operating conditions. The corresponding machine learning methods are: artificial neural
network (ANN), fuzzy inference system (FIS), support vector machine (SVM), and
genetic programming (GP). By using four types of inputs of the SOFC operation: i.e.
load current, fuel utilization, inlet air temperature, and air molar flow rate, the task of
the corresponding machine learning methods is to predict the stack voltage and outlet
temperature values of the corresponding SOFC operation. 1000 input-output data
pairings that were generated from the simulations of a physical tubular SOFC model
under various operating conditions were used to train the corresponding machine
learning models. It was found out from this study that ANN method has slightly better
performance in modelling the power performance behaviour of the corresponding SOFC
system under various operating conditions.
Keywords: Solid oxide fuel cells; Machine learning methods; Power performance
prediction

1. Introduction
Solid oxide fuel cells (SOFC) are expected to play a significant role in helping to meet
the ever-increasing demands for cleaner supply energy in the near future (Stambouli,
2011). Already, SOFCs have been proposed as a potential power source for distributed
and stationary power plants and also mobile applications. The advantages of SOFCs are
their high efficiency, modularity, low noise and low environmental pollution. However,
certain challenges, including the optimum operation of the SOFC stacks, need to be
resolved before the technology can be adopted for the real-world applications.
Moreover, this issue will also take into considerations the unique nature of the SOFC
operating phenomena. Clearly, the requirement to satisfy all these challenges will
require the development of an effective control strategy that is specifically tailored to
the SOFC operation. Model-based control strategies that rely on the availability of good
modeling descriptions of the SOFC phenomena are expected to play a crucial role in
this regard.
Machine learning has shown its great utility in modeling complex phenomena in
chemical processes. This utility has brought forward its potential for applications in
advanced process control strategies such as real-time optimization and model-based

20

M. N. Fuad et al.

predictive controls. Although first-principle based modeling is very useful for design
purpose and as an aid for the understanding, black-boxed modeling that employs
machine learning principles such as neural network and support vector machine is very
useful for real-time applications that demand faster and robust computations. Moreover,
the developments of the corresponding machine learning models are less demanding as
long as sufficient collections of input-output data samples are available for training
purpose. Once trained, the corresponding models can be used effectively and efficiently
to achieve various objectives such as operation point optimization or model-based
controls. Therefore, driven by these motivations, this paper seeks to study the
application of several machine learning methods to model the power performance
behavior of a solid oxide fuel cells operation. Specifically, four types of machine
learning methods i.e. artificial neural network, fuzzy inference system, support vector
machines and genetic programming are applied in this study in order to observe their
performances in modeling the operating behavior of the corresponding SOFC system.

2. Brief Descriptions of the Corresponding Machine Learning Methods


2.1. Artificial Neural Networks
Artificial neural networks (ANN) are a computational tool modeled on the
interconnection of the neurons in the nervous systems of the human brain and that of
other animals (Bishop, 1995). The structure of a feed-forward multilayer ANN is
displayed in Figure 1. In this structure, the information is passed from the input layer to
the hidden layer via various network connections and finally to the output layer. The
training phase of ANN consists of submitting samples of input-output data (called the
training data) to the network and adjusting the connection weights until the measure of
difference between the target data and ANN output is minimized. Past study has proven
that the standard multilayer feed-forward network with a single hidden layer can be
used to approximate continuous function of arbitrary complexity (i.e. universal
approximation theorem).

Figure 1. Feed-forward multilayer neural network


2.2. Fuzzy Inference Systems
Fuzzy inference systems (FIS) are simply the applications of fuzzy logic and fuzzy set
theory for data classification, decision analysis, and pattern recognition (Takagi &
Sugeno, 1985). The general architecture of FIS consists of three parts (see Figure 2).
The first part i.e. the fuzzifier, convert the crisp input to linguistic variables by using the

Machine Learning Based Modeling for Solid Oxide Fuel Cells Power Performance
Prediction

21

membership functions stored in the fuzzy knowledge base. In the next part i.e. inference
engine, a collections of IF-THEN type fuzzy rules will convert the fuzzy input to the
fuzzy output. Finally in the defuzzifier part, the fuzzy output of the inference engine will
be converted to crisp output by using the membership function analogous to the ones
employed by the fuzzifier. Currently, two types of fuzzy inference systems are widely
employed, i.e. Mamdani and Takagi-Sugeno. Moreover, the Takagi-Sugeno fuzzy
inference system can be trained by an adaptive technique in which the parameters of
the membership functions are optimized with respect to the given samples of inputoutput data.

Figure 2. Components of a fuzzy system


2.3. Support Vector Machines
Support vector machines (SVM) are among kernel-based techniques that are very
popular for data classification and regression (Ivanciuc, 2007). More formally, a support
vector machines constructs a hyperplane or set of hyperplanes in a high- or infinitedimensional space from several points in training examples (called support vectors)
which can be used for classification or regression. Formerly developed as a linear data
classifier, the extension to nonlinear classification was made possible by using kernel
trick that maps input space into a higher dimensional feature space. The training phase
of SVM amounts to solving an optimization problem that seeks to find the largest
margin hyperplane that represents the best separation of data into its proper categories.
The extension of SVM to nonlinear regression was made possible by using an insensitive loss function. Generally, the goal of SVM regression (SVMR) is to identify a
function f(x) that for all training patterns x has a maximum deviation from the target
values y and has a maximum margin (Ivanciuc, 2007).
2.4. Genetic Programming
Genetic programming (GP) is an evolutionary algorithm-based methodology inspired by
biological evolution to find computer programs that can better perform a user-defined
task (Koza, 1992). It is a specialization of genetic algorithms where each individual is a
computer program (see Figure 3) that will be evolved according to evolutionary
principles that seeks the fittest individuals among the population of the candidate
solutions. The fittest individuals represent computer programs that can perform the userdefined task optimally. The principle of GP uses various analogs of the naturally
occurring evolutionary operations, including crossover (sexual recombination),
mutation, gene duplication, and gene deletion. Among the numerous applications of GP,
it has been used successfully for symbolic regressions. Specifically, in symbolic
regressions, the task of GP is to find both structure and parameters of a nonlinear model
that minimizes the error criterion between predictions and observed data.

22

M. N. Fuad et al.

Figure 3. A computer program (e.g. mathematical function) represented as a tree


structure in genetic programming

3. Modeling SOFC Power Performance via Machine Learning Methods


The corresponding machine learning methods (i.e. ANN, FIS, GP and SVM) are used in
this study to model the power performance behavior of a tubular SOFC under various
operating conditions. In this study, each of the corresponding machine learning models
will receive four inputs of the SOFC operation i.e. load current (20-158 A), fuel
utilization (50-90%), inlet air temperature (973-1173 K), and air molar flow rate (0.010.02 mol/s). In turn, the models will predict the stack voltage and outlet temperature
values from the corresponding inputs. The database of the SOFC operation is generated
from the simulations of a physical model of a tubular SOFC for 1000 input dataset.
These input-output data pairings are further partitioned into 800 dataset for
training/validation and 200 dataset for testing. Furthermore, the dataset for the
training/validation are corrupted by 5% measurement errors while the testing dataset
are left intact. All the input-output dataset in this study are normalized to [-1, 1] in order
to facilitate the training phase. Finally, the training parameters for each of the machine
learning models are tabulated in Table 1.

4. Results and Discussions


Figure 4 shows the parity plots that were generated from the testing data and the
corresponding machine learning model predictions at the conclusion of the trainings.
Although ANN method yield the slightly highest prediction accuracy with correlation
coefficient R = 0.99922, the other machine learning methods also exhibit satisfactory
performance in modeling the operating behavior of the corresponding SOFC system. It
should be noted that the corresponding parity plots were generated from the uncorrupted
testing data that was not used in the training phase of the corresponding machine
learning models. As such, the corresponding testing data (i.e. unseen data) can provide
information to evaluate the performance of the corresponding machine learning methods
in learning the behavior of the training/validation dataset that was corrupted by
measurement errors. Moreover, the prediction capabilities of the trained machine
learning models were also investigated further by reproducing the power performance
curves of the corresponding SOFC system at various fuel utilization and inlet air
temperature conditions as shown in Figure 5. As can be seen from the corresponding
figures, all the machine learning methods can reproduce the power performance curve
including the location of the maximum power point satisfactorily.

Machine Learning Based Modeling for Solid Oxide Fuel Cells Power Performance
Prediction

23

Table 1. Machine learning methods and their associated training parameters


Machine learning
Training parameters
Artificial neural network (ANN)
Multilayer feed-forward neural network with one hidden
layer and 10 neurons in the hidden layer
Training algorithm: Backpropagation algorithm with early
stopping (as implemented by neural network toolbox in
Matlab software)
Fuzzy inference system (FIS)

The training is based on the adaptive network based fuzzy


inference system (ANFIS) as implemented by fuzzy logic
toolbox in Matlab software
2 membership functions of generalized bell type are
implemented for the input later and linear type
membership function is implemented for the output layer
Training algorithm: hybrid training method (i.e.
backpropagation algorithm with least-squares estimation
method)

Support vector machine (SVM)

SVM type: -support vector regression (Chang & Lin,


2002) as implemented by LIBSVM package
Kernel type: radial basis function

Genetic programming (GP)

Multi-gene symbolic regression (Searson, Leahy, & Willis,


2010) as implemented by GPTIPS package
Maximum number of genes: 5
Maximum tree depth: 6
Population size: 100
Number of generations: 100

SVM, R = 0.99841
1

0.5

0.5
Predicted outputs

Predicted outputs

ANN, R = 0.99922
1

-0.5

-1
-1

-0.5

-0.8

-0.6

-0.4

-0.2
0
0.2
Actual outputs

0.4

0.6

0.8

-1
-1

-0.8

-0.6

-0.4

0.5

0.5

-0.5

-1
-1

0.4

0.6

0.8

0.4

0.6

0.8

GP, R = 0.99850

Predicted outputs

Predicted outputs

ANFIS, R = 0.99825

-0.2
0
0.2
Actual outputs

-0.5

-0.8

-0.6

-0.4

-0.2
0
0.2
Actual outputs

0.4

0.6

0.8

-1
-1

-0.8

-0.6

-0.4

Figure 4. Parity plots for the testing data predictions

-0.2
0
0.2
Actual outputs

24

M. N. Fuad et al.

SVM
7000

6000

6000

5000

5000
Power (W)

Power (W)

ANN
7000

4000
Actual (FU = 0.75, Tin = 1073 K)
Predicted (FU = 0.75, Tin = 1073 K)
Actual (FU = 0.85, Tin = 1173 K)
Predicted (FU = 0.85, Tin = 1173 K)

3000
2000
1000
20

40

60

80
100
Current (A)

120

140

4000
Actual (FU = 0.75, Tin = 1073 K)
Predicted (FU = 0.75, Tin = 1073 K)
Actual (FU = 0.85, Tin = 1173 K)
Predicted (FU = 0.85, Tin = 1173 K)

3000
2000
1000
20

160

40

60

7000

6000

6000

5000

5000

4000
Actual (FU = 0.75, Tin = 1073 K)
Predicted (FU = 0.75, Tin = 1073 K)
Actual (FU = 0.85, Tin = 1173 K)
Predicted (FU = 0.85, Tin = 1173 K)

3000
2000
1000
20

40

60

80
100
Current (A)

120

140

160

GP

7000

Power (W)

Power (W)

ANFIS

80
100
Current (A)

120

140

4000
Actual (FU = 0.75, Tin = 1073 K)
Predicted (FU = 0.75, Tin = 1073 K)
Actual (FU = 0.85, Tin = 1173 K)
Predicted (FU = 0.85, Tin = 1173 K)

3000
2000

160

1000
20

40

60

80
100
Current (A)

120

140

160

Figure 5. Comparisons of actual and predicted power performance curves (air molar flowrate =
0.012 mol/s)

5. Conclusions
In this study, we have compared different types of machine learning methods to model
the power performance behavior of a tubular SOFC operation. Among the different
types of machine learning methods that were covered in this study, it was found out that
ANN method has slighlty better performance in predicting the power performance
behavior of the corresponding SOFC system under various operating conditions. The
result from this study opens the possibility for applying the corresponding machine
learning method for SOFC performance maps constructions and operation point
optimization.

References
Bishop, C. M. (1995). Neural Networks for Pattern Recognition. Oxford: Oxford University
Press.
Chang, C. C., & Lin, C. J. (2002). Training nu-support vector regression: theory and algorithms.
Neural Computation, 14(8), 1959-1977.
Ivanciuc, O. (2007). Applications of support vector machines in chemistry. In K. B. Lipkowitz &
T. R. Cundari (Eds.), Reviews in Computational Chemistry (Vol. 23, pp. 291-400). Weinheim:
Wiley-VCH.
Koza, J. (1992). Genetic Programming: On the Programming of Computers by Means of Natural
Selection. Cambridge, MA: MIT Press.
Searson, D. P., Leahy, D. E., & Willis, M. J. (2010). GPTIPS: An open source genetic
programming toolbox for multigene symbolic regression. Paper presented at the International
MultiConference of Engineers and Computer Scientists Hong Kong.
Stambouli, A. B. (2011). Fuel cells: The expectations for an environmental-friendly and
sustainable source of energy. Renewable and Sustainable Energy Reviews, 15, 4507-4520.
Takagi, T., & Sugeno, M. (1985). Fuzzy identification of systems and its applications to modeling
and control. IEEE Transactions on Systems, Man and Cybernetics, 15, 116-132.

You might also like