Abstract
We developed and evaluated an AI-based surrogate model to simulate Hydrogen (H2) production via Proton Exchange Membrane Water Electrolysis (PEMWE). A variety of Machine Learning (ML) and Deep Learning (DL) models were tested and fine-tuned using real-world PEMWE datasets from multiple sources, ensuring model robustness and accuracy. The models included ML algorithms such as k-Nearest Neighbors (k-NN), Support Vector Machine (SVM), Decision Trees (DT), Random Forest (RF), Category Boosting (CB), Light Gradient Boosting (LGB), and Gradient Boosting (GB), with DL models Long Short-Term Memory (LSTM), Multi-Layer Perceptron (MLP), and one-dimensional (1D) Convolutional Neural Networks (CNN). Among these, the 1DCNN model demonstrated superior performance, achieving an R-squared (R2) of 0.998944, a Mean Squared Error (MSE) of 488.82, a very low Normalised Root Mean Square Error (NMSE) of 0.001055, and Pearson correlation of 0.999472, having MLP also performing exceptionally well. Conversely, the LSTM model performed the poorest. Unlike many prior studies, we employed cross-validation techniques to rigorously validate model performance and the Wilcoxon Signed-Rank Test for statistical validation, establishing the robustness and reliability of 1DCNN predictions. Comparatively, the MLP model performed well but failed to pass the Wilcoxon test, indicating its predictions were not statistically different from any other models. These findings underscore the novelty of the proposed 1DCNN-based surrogate model, which stands out in its ability to accurately simulate PEMWE behaviour for H2 production. This model can simulate PEMWE processes in a fraction of the time required by traditional methods, providing valuable operational insights and advancing technologies across H2 production, energy sectors, transportation, and sustainable energy systems.