Prediction Study of Solid Oxide Fuel Cell Performance Degradation Using Data-Driven Approaches

Abstract

Performance degradation in solid oxide fuel cells (SOFCs) leads to shorter service life and unexpected downtime. To reduce economic losses and accelerate commercialization, accurately predicting the degradation is conducted in this study. First, a comprehensive analysis of performance degradation through experiments on a real SOFC system is investigated. Then, three dada-driven robust models, that is, vector autoregressive moving average (VARMA), radial basis function neural network (RBFNN), and neural basis expansion analysis for time series (N-BEATS) models are proposed to predict the SOFC's performance degradation. Herein, the top 60–90% of the experimental datasets are used for training and the bottom 40–10% for testing. After training, the prediction performance testing of these 3 models is compared with that of the bi-long short-term memory networks (bi-LSTM) and bi-gated recurrent units (bi-GRU) models. Simulation results show that both the VARMA and N-BEATS models are superior to the bi-LSTM and bi-GRU models in predicting the performance degradation of the SOFC. While the test performance of the RBFNN model is worst, especially under the top 60% training datasets condition. These indicate it is feasible to respectively establish the VARMA model and the N-BEATS model for predicting the SOFC's performance degradation.