Control-Oriented Extraction and Prediction of Key Performance Features Affecting Performance Variability of Solid Oxide Fuel Cell System

Performance prediction technology enhances the dependability and safety of solid oxide fuel cell (SOFC) systems. However, unreasonable prediction objects and insufficient prediction accuracy impede the advancement of prediction methodologies in SOFC systems. To better predict SOFC system performance changes, two algorithms are developed in this article: a mutual information (MI)-based feature selection algorithm that identifies variables sensitive to SOFC performance changes and a particle swarm algorithm optimized backpropagation (BP) neural network (NN) (PSO_BP). The PSO_BP algorithm predicts hard-to-measure variables (e.g., heat exchanger temperature and combustion chamber temperature) using easy-to-measure data. Real data tests reveal significant improvements in feature reduction, prediction accuracy, and speed. The MI algorithm reduces feature variables by 70% (from 40 to 12), aiding sensor layout optimization. The PSO_BP algorithm trains in 2 s and achieves less than 0.5% prediction error. It can also predict hard-to-measure temperature data, which can assist researchers in optimizing SOFC system performance.