Artificial intelligence models for predicting the performance of proton exchange membrane water electrolyzers under steady and dynamic power

Abstract Proton exchange membrane water electrolyzers (PEMWEs) are one of the leading technologies to produce hydrogen from renewable power sources. Predicting the effect of degradation in PEMWEs is crucial for determining economic feasibility, maintenance schedules, and optimal control. However, existing semi-empirical and data-driven models for degradation are typically rudimentary or system-specific. Artificial Intelligence (AI) offers a method to improve and generalize models by implicitly learning the system behavior through analysis of large experimental datasets. This paper evaluated the ability of a range of AI techniques to model PEMWE performance degradation under steady-state and dynamic operation. Data from 39 distinct experiments were collated into a training dataset with 19 input features consisting of (i) construction parameters, (ii) static operating parameters, (iii) dynamic operating parameters, and (iv) the recording time. The target for modeling was cell voltage, selected as a systemic proxy for degradation. Over 6900 data points were collected with a maximum experimental duration of 5600 H. The duration studied, the number of input features, and the variety of PEMWE constructions considered make this the most comprehensive AI model on the subject, to date. The model offered a goodness of fit or coefficient of determination ( R 2 ) value of 0.9991 for the testing data and can predict the performance of untrained electrolyzers and their operation with similar accuracy.