Artificial neural network model of a short stack solid oxide fuel cell based on experimental data

Abstract

Solid oxide fuel cells (SOFCs) are complex systems in which electrical conduction, heat transfer, gas phase mass transport, chemical reactions and ionic conduction take place simultaneously and are tightly coupled. Mathematical models based on conservation laws have been shown to be slow and because of some parameter estimation for physical, chemical and electrochemical properties they have less accuracy. ANN models are powerful tools that bring simplicity and real-time response to SOFC modeling. Depending on the quality of the training data, ANN models can also show greater accuracy than CFD models. In this study ANN modeling of a short stack SOFC is considered. Training data are extracted and filtered from measurements on a dedicated test set-up. Given the fuel flow and composition, air flow, oven temperature and current, the model can predict the voltage and temperature profile of the cell. An optimized structure for the network is selected as: 5–11–6 for a 5 input, 6 output network with 11 hidden neurons. Prediction results of the ANN model deviate 0.2% concerning average relative error compared to the measurements.