Multi‐Target Stacked LSTM Framework for Dynamic Setpoint Optimization and Hydrogen Consumption Reduction in PEM Fuel Cells

Proton exchange membrane fuel cells (PEMFCs) are considered one of the most promising clean energy technologies; however, improving operational efficiency while minimizing hydrogen consumption remains a critical challenge for their large‐scale deployment. This study proposes an artificial intelligence‐driven predictive control framework based on a multi‐target stacked long short‐term memory (LSTM) network for simultaneous prediction of stack voltage and hydrogen flow rate, followed by dynamic operating setpoint optimization. The model was trained in 26 930 min of experimental PEMFC operating data, enriched through temporal feature engineering, thereby increasing the input dimensionality from 4 to 16 features. The proposed architecture achieved coefficients of determination ( R 2 ) of 0.7931 for stack voltage and 0.7135 for hydrogen flow rate, with low prediction errors (MAPE of 0.68% and 1.25%, respectively). A two‐dimensional optimization strategy identified the recommended operating condition at 71°C and 320 kPa, achieving an 87.0% voltage constraint satisfaction rate while reducing hydrogen consumption by 0.25% under a universal setpoint (0.41% under per‐sample optimization). One‐dimensional optimization yielded statistically significant hydrogen savings of 0.063 LPM (0.41%), verified using Wilcoxon signed‐rank testing and bootstrap confidence intervals (CIs). Monte Carlo dropout analysis produced mean predictive uncertainties below 0.40%, while the average inference latency of 41.4 ms confirmed real‐time applicability. Furthermore, the optimized operating strategy can reduce annual carbon emissions by up to 0.90 tCO 2 per year for a 50‐unit fleet (0.018 tCO 2 per unit) for grey hydrogen systems, demonstrating the proposed framework's potential for intelligent, low‐carbon PEMFC operation.