Statistical and machine learning methods for modeling the next-generation PEM water electrolysis systems: response surface methodology and artificial neural networks

• Green hydrogen production via proton exchange membrane water electrolysis was investigated. • Machine learning and computational modelling shows efficiency in predicting hydrogen production rate. • Response surface method models water electrolysis with R 2 = 0.9971 with small error. • Optimized artificial neural network models electrolysis with R 2 = 0.9928. • The voltage-current interaction and the power have the most significant correlations. Green hydrogen production is gaining substantial interest in today’s movement towards energy transition. Research efforts have been exploring water electrolysis technology, powered by renewable energy sources, to produce pure hydrogen. Proton exchange membrane water electrolysis (PEMWE) is considered one of the promising options. Due to the complexity of hydrogen production via PEMWE, arising from the multiple process variables and phases involved, it is challenging to model hydrogen production using classical modeling methods. Statistical and machine learning models have been used to model and predict the hydrogen flow rate. However, existing PEMWE modeling studies typically employ either classical statistical models with a small number of operating variables or ML models trained with limited physics-based feature engineering, and there is almost no systematic comparison made between interpretable statistical surrogates and machine-learning models. This work seeks to fill that gap by applying artificial neural networks (ANNs) and response surface methodology (RSM). The RSM is a technique that models the response both linearly and quadratically using the input features. The RSM model, in this study, outperformed existing models in the literature, with a high R2 value of 0.9971. The model also had a normalized mean absolute error (NMAE) of 0.0031 and a normalized root mean square error (NRMSE) value of 0.0074. The data is also trained on multiple artificial neural network (ANN) configurations that use K-fold stratified cross-validation, with the best config being (8,8). The best ANN model yielded the highest cross-validated R2 value of 0.9928, NRMSE of 0.0058, and NMAE of 0.0015, indicating it to be the most optimal given its layer size. The use of the RSM and ANN techniques provides insights into modeling the output and its respective interactions. The results demonstrate the effectiveness of computational methods in facilitating the transition to green hydrogen production. In terms of novelty, this study is the first PEMWE study to compare both RSM and physics-informed shallow ANNs on a multi-source dataset for predicting hydrogen flow rate, achieving state-of-the-art accuracy while retaining interpretable feature interactions.