Comprehensive Evaluation of Optimization Algorithms and Performance Criteria for ANN-Based PEMFC Voltage Prediction

Abstract

Proton exchange membrane fuel cells (PEMFCs) are considered to be a promising solution for clean energy conversion in hydrogen electric vehicles. Accurate voltage prediction is crucial for designing efficient energy management and control strategies. While deep neural networks have shown good potential in modeling PEMFCs, the role of optimization algorithms and training performance criteria in achieving accurate voltage predictions remains unclear. This research aims to carry out a comprehensive comparative study using three popular optimization algorithms and different performance criteria including prediction accuracy, convergence speed, and training stability. A real experimental dataset for a Nexa PEMFC system has been used to train and evaluate different models of artificial neural networks (ANNs) to find out which optimization algorithm and performance criteria are best for efficient modeling of PEMFCs under varying operating conditions. The results of this study are analyzed through a comparative evaluation of different metaheuristic optimization algorithms applied within a unified ANN training framework for PEMFC voltage prediction. Particle swarm optimization (PSO) provides the highest voltage prediction accuracy and robust convergence behavior, whereas Grey Wolf Optimization (GWO) achieves the fastest convergence with reduced computational time.