Study on Nonlinear Identification SOFC Temperature Model Based on Particle Swarm Optimization–Least-Squares Support Vector Regression

Abstract

In order to facilitate valid solid oxide fuel cell (SOFC) temperature control scheme, a nonlinear identification method of SOFC temperature dynamic behaviors is proposed using an autoregressive network with exogenous inputs (NARX) model, whose nonlinear function is described by a least-squares support vector regression (LSSVR) method with radial basis kernel function (RBF). During the identifying process, a particle swarm optimization (PSO) algorithm is introduced to optimize the parameters of LSSVR. On the other hand, a mechanism model is developed to sample the training data to regress the NARX model. Investigations are conducted to analyze the effects of training data size and PSO fitness function on the accuracy of the NARX model. The results demonstrate that the NARX model with tenfold cross-validation fitness function and large size data is precise enough in predicting the SOFC temperature dynamic behaviors. The maximum errors of cathode and anode outlet temperature are 0.3081 K and 0.3293 K, respectively. Furthermore, the simulation speed of NARX model is much faster than the mechanism model because NARX model avoids the internal complex computation process. The training time of the NARX model with large size data is about 1.2 s. For a 20,000 s simulation, the predicting time of the NARX model is about 0.2 s, while the mechanism model is about 36 s. In consideration of its high computational speed and accuracy, NARX model is a powerful candidate for valid multivariable model predictive control (MPC) schemes.