On the use of neural networks and statistical tools for nonlinear modeling and on-field diagnosis of solid oxide fuel cell stacks

The paper reports on the activities performed within the European funded project GENIUS to develop black-box models for modeling and diagnosis of solid oxide fuel cell (SOFC) stacks. Two modeling techniques were investigated, i.e. Neural Networks (NNs) and Statistical Tools (STs). The deployment of NNs was twofold: Recurrent Neural Networks (RNNs) and an NN classifier were developed to simulate transient operation of SOFCs and identify some specific faults that may occur in such devices, respectively. On the other hand, STs are based on a stepwise multiple regression. Data for model development were obtained from experiments specifically designed to reach maximal information content. The final aim was to obtain highly general models of SOFC stacks' operation in both transient and steady state. All the developed black-box models exhibited high accuracy and reliability on both training and test data-sets. Moreover, the black-box models were also proven effective in performing real-time monitoring and degradation analysis for different SOFC stack technologies.