Hydrogen fuel cells for green shipping: A machine learning-based power prediction framework

Hydrogen fuel cells offer a promising route toward low-emission maritime electrification, but reliable prediction of electrical power is essential for their integration into shipboard energy-management systems. This study develops a compact, explainable machine-learning framework using 119 observations from a publicly available stationary hydrogen fuel-cell installation. Ambient, fuel-cell gas, and coolant outlet temperatures; power-generation and consumption setpoints; auxiliary consumption; and operating status were used to predict power supplied to the grid. Linear regression (LR) and Extreme Gradient Boosting (XGBoost) were evaluated using a chronological 80:20 training–testing split, with parity plots, residual diagnostics, and SHapley Additive exPlanations (SHAP) supporting model assessment. LR fitted the training observations closely (R² = 0.9960) but failed to generalize to testing data (R² = -33.8364; RMSE = 65.651). In contrast, XGBoost achieved consistent training and testing performance (R² = 0.9740 and 0.9770, respectively), with a testing RMSE of 1.687 and MAE of 1.273. SHAP identified auxiliary consumption (mean absolute SHAP = 9.82) and the power-consumption setpoint (2.45) as dominant predictive features. The analysis addresses grid-delivered power rather than propulsion demand or directly measured hydrogen utilization. The findings demonstrate the value of interpretable nonlinear regression for fuel-cell power prediction and offer a methodological foundation for future maritime applications.