Identification of Wiener models for dynamic and steady‐state performance with application to solid oxide fuel cell

Abstract

AbstractThis work is concerned with identification of Wiener models (a linear dynamic part connected in series with a nonlinear dynamic one). A neural network with one hidden layer is used as the nonlinear block of the model, two network configurations are considered. For model identification three algorithms are described. In the first case model accuracy only in transient conditions is considered, only the dynamic data is used for model training. In the next two algorithms model accuracy in both transient and steady‐state conditions is considered, dynamic and steady‐state data sets are used. The steady‐state model errors are taken into account by an additional term in the minimized cost‐function or by additional inequality constraints. For comparison of discussed algorithms and model structures, identification of a Wiener model of a solid oxide fuel cell (SOFC) process is considered. It is shown that the best results are obtained by the algorithm 2 which minimizes at the same time both dynamic and steady‐state model errors, additional constraints used in the algorithm 3 are computationally quite demanding.