Leakage‐Aware and Spatially Validated Dynamic Ensemble Learning for Explainable Renewable Hydrogen Feasibility Assessment

ABSTRACT

The assessment of renewable hydrogen feasibility requires predictive models that are not only accurate but also resistant to target leakage and capable of generalizing across geographically unseen regions. This study evaluates a publicly available global renewable hydrogen dataset containing 2,535 geographically distributed observations and develops a dynamic ensemble hybrid regression (DEHR) framework for renewable hydrogen feasibility assessment. Particular attention is given to the distinction between exogenous pre‐development predictors and downstream variables that may encode information closely related to the target. Accordingly, the revised experimental protocol introduces a leakage‐free primary analysis based only on variables that can reasonably be available before project development, while models containing derived power and hydrogen‐production variables are retained only as secondary diagnostic experiments. All preprocessing, hyperparameter optimization, expert weighting, out‐of‐fold prediction generation, and meta‐learning are performed exclusively within the training data. In addition to random hold‐out testing, repeated cross‐validation and geographically structured validation are used to quantify robustness under distribution shift. The DEHR framework combines heterogeneous base regressors using out‐of‐fold performance‐based weighting and a regularized ElasticNet meta‐learner. Ablation experiments separately evaluate the contributions of expert weighting and meta‐learning. Model interpretation is performed using SHAP, LIME, partial dependence, and individual conditional expectation analyses, with explicit consideration of feature collinearity and explanation stability. The results demonstrate that near‐perfect performance obtained with the extended predictor set is substantially influenced by downstream derived variables and should therefore not be interpreted as evidence of unrestricted real‐world generalization. In contrast, the leakage‐free and spatially structured experiments provide a more conservative and practically relevant estimate of predictive performance. These findings emphasize that methodological separation, spatial validation, and leakage‐aware interpretation are essential when machine‐learning models are used for renewable hydrogen feasibility assessment.