A human-in-the-loop active learning framework for scalable wind energy potential suitability assessment

High-resolution wind energy potential (WEP) suitability assessment typically requires expert judgements on a large number of candidate sites. Such labelling is subjective, costly, time-consuming and difficult to scale, especially when siting criteria and policy constraints evolve over time. To address this challenge, this study proposes a human-in-the-loop (HITL) active learning framework that aims to (i) minimise expert annotation effort, (ii) enable scalable, high-resolution WEP mapping, and (iii) provide interpretable, expert-consistent model outputs. Within the framework, a limited set of expert-style labels is first generated for a subset of locations to train an intelligent classifier. An active learning strategy then iteratively selects the most informative locations for additional expert annotation, and the classifier is updated until the desired performance is reached. Shapley Additive exPlanations (SHAP) are integrated as an interpretability tool to quantify the contribution of each siting criterion and to explain model behaviour at both global and local scales. The proposed framework is demonstrated for the Beijing–Tianjin–Hebei region using GIS-based spatial datasets. Among a suite of candidate machine learning models, the Transformer model coupled with entropy-based active learning achieves the best trade-off between accuracy and labelling effort, requiring only about one quarter as many location labels as random sampling to reach a comparable level of WEP assessment performance. The resulting high-resolution suitability maps are cross-validated against existing wind farms and further explained using SHAP to reveal criterion-level and location-specific driving mechanisms. The case study demonstrates that the proposed HITL active learning framework can provide label-efficient, adaptive and transparent decision support for wind energy planning in complex metropolitan settings.