Accurate hydrogen flow prediction in proton exchange membrane electrolyzers using feature selection and stacked ensemble learning

Green hydrogen produced using proton exchange membrane (PEM) electrolyzers plays a pivotal role in decarbonizing energy systems. Accurately predicting hydrogen flow rates is essential for optimizing electrolyzer performance and ensuring system reliability. This study presents a two-stage solution to enhance the accuracy, transparency, and reliability of hydrogen flow-rate prediction: (i) rank-aggregation-based identification of the most influential features, and (ii) development of an optimal ensemble framework integrating Bayesian optimization and stacking paradigms. A feature selection pipeline was designed using rank aggregation that combines statistical (correlation and regression weights) and model-agnostic (tree-based statistics and SHapley Additive exPlanations (SHAP)) techniques to capture both linear and nonlinear dependencies among input parameters. The optimized ensemble framework leverages Bayesian hyperparameter tuning and a stacked support vector regression (Meta-SVR) meta-learner to integrate the strengths of multiple traditional ML and deep tabular models (TabNet, NODE, DeepFM). Comprehensive experiments using PEM electrolyzer data demonstrate the empirical superiority of the proposed approach, achieving the highest R 2 of 0.9761 with relative improvements of 10.29%, 16.76%, and 18.75% over SVR, XGBoost, and RF, and 0.72%, 2.00%, and 2.86% over TabNet, NODE, and DeepFM, respectively. Residual-error distribution analysis confirms the reliability and unbiased performance of the proposed model, while computational and statistical analyses verify its efficiency and significance relative to baseline models. The Meta-SVR framework also achieved a near-instantaneous inference time of 0.07 s, highlighting its suitability for real-time hydrogen production monitoring and control. Overall, the proposed approach provides a reliable, interpretable, and computationally efficient solution for hydrogen flow-rate prediction in PEM electrolyzer systems.