Machine learning two-step algorithm for prediction of proton exchange membrane water electrolyzer cell performance under variable power inputs

The fluctuating nature of renewable energy sources, such as solar and wind, introduces dynamic input conditions that are difficult to replicate and analyze using conventional laboratory approaches. This work presents a machine learning-based approach for predicting proton exchange membrane water electrolyzer (PEMWE) performance under dynamic power inputs, aiming to reduce experimental complexity, and accelerate control system development and PEMWE deployment. A novel two-step machine learning algorithm was developed using a feedforward neural network for PEMWE current estimation and a long short-term memory architecture for hydrogen production forecasting. Experimental data, gathered from eight distinct power profiles under variable temperature regimes, was used to train and validate the models. The algorithm demonstrated strong predictive capabilities and generalization across unseen operational voltage profiles, achieving a mean absolute error of 0.0183 for current prediction and 0.1833 for hydrogen production. Additional validation on quasi-random and constant-power profiles confirmed the robustness of the proposed approach, even under noisy conditions. This study highlights the potential of machine learning to serve as a digital surrogate for complex PEMWE experiments, enabling accurate performance predictions and paving the way for advanced control strategies and real-time system optimization of green hydrogen production.