Data-driven modeling and prediction of PEM fuel cell voltage response to load transients for energy applications

Polymer electrolyte membrane fuel cells (PEMFCs) encounter significant challenges in adapting to dynamic load transients, with voltage fluctuations affecting both immediate performance and long-term durability. This study addresses the critical gap in predicting these transient voltage responses, which remains underexplored in existing literature. A data-driven model based on long short-term memory neural networks has been developed to predict the behavior of PEMFC's voltage under dynamic load conditions. The model's novelty lies in its millisecond-resolution transient prediction capability, comprehensive input parameterization (11 input variables versus 2–5 in previous studies), and iterative optimization framework balancing computational efficiency with predictive accuracy. Two distinct fuel cell systems are employed: an experimental 14.5 kW forklift system and a 130 kW automotive system with simulated datasets exceeding 5 million data points—providing unprecedented validation scope. Three data scaling techniques —Normalization, StandardScaler, and RobustScaler—were evaluated, with Normalization achieving superior performance across datasets (MSE of 0.05 and MAPE of 0.35 % for experimental data; MSE of 0.44 and MAPE of 0.15 % for high-voltage simulated data). StandardScaler offers comparable accuracy with higher computational requirements, while RobustScaler demonstrates higher error rates despite faster convergence. This model enables proactive management of voltage fluctuations under dynamic conditions, advancing PEMFC control strategies beyond existing studies that predominantly address steady-state or degradation prediction over longer timeframes.