Machine learning-based framework for predicting green hydrogen production efficiency under variable environmental conditions
Abstract
Predicting green-hydrogen electrolyser efficiency with useful accuracy remains a practical obstacle to running these plants economically, since irradiance, temperature, humidity and water-feed quality interact in ways classical correlations capture poorly. In this simulation-based study we build a five-stage machine-learning pipeline — data collection, preprocessing, model benchmarking, an Adaptive Correction Module (ACM), and SHAP-based explanation — to estimate the specific production efficiency of a 500 kW PEM electrolyser. Four regressors (Random Forest, XGBoost, SVR and an MLP) are compared after Bayesian tuning on a season-stratified 75/25 split of a twelve-month synthetic dataset combining a physics-based electrolyser model with NASA POWER records for three Moroccan sites. XGBoost generalises best (R 2 = 0.978-0.993, RMSE = 0.006–0.010 Nm 3 /kWh, MAPE = 1.8–3.0 %), and the ACM — blending the global model with a local k-NN correction once an observation's Mahalanobis distance passes a data-driven threshold — lowers RMSE by 10-18 % during heat waves, dust events and high-humidity fog. SHAP analysis attributes 35–45 % relative importance to global horizontal irradiance, followed by stack temperature and humidity. Because the dataset is entirely synthetic, these results are best read as a feasibility demonstration rather than a field-validated tool; Section 6 sets out the real-plant validation we consider necessary before the framework could support operational decisions.