Explainability in wind farm planning: A machine learning framework for automatic site selection of wind farms

Abstract

The process of wind farm site selection involves complex considerations spanning environmental, economic, and social factors, often handled through multi-criteria decision making (MCDM) methods. However, MCDM approaches face several challenges. To address these, integrating machine learning (ML) techniques offers promise. Nevertheless, existing studies have yet to employ a standalone ML-based approach for this purpose. This study seeks to pioneer an explainable ML-based framework for automated wind farm site selection. The framework incorporates feature selection techniques, seven ML algorithms, statistical tests, and a local explainability method. The experimental area was chosen as Balıkesir province, which hosts 13% of Türkiye's installed wind energy capacity. According to the prediction performances of the ML models, Extreme Gradient Boosting-XGBoost has the highest accuracy (0.9607), followed by Light Gradient Boosting Machine-LightGBM (0.9580), Random Forest-RF (0.9518), Histogram-based Gradient Boosting-HGB (0.9387), Classification and Regression Trees-CART (0.8946), Logistic Regression-LR (0.8856), and Naive Bayes-NB (0.8456). Additionally, McNemar’s tests revealed statistically significant differences among ML models. Based on the explanations provided by SHapley Additive exPlanations, wind speed, distance to transmission lines, distance to protected zones, and elevation were the top contributing criteria. Moreover, analysis revealed regions with elevated wind speeds, higher elevations, and closer proximity to transmission lines and protected areas as favorable for wind farm installation. The study's outcomes indicate strong agreement (97%) between current wind turbine locations and those identified as suitable by ML models. Within Balıkesir province, our research has identified a significant number of highly suitable areas where wind turbines have not yet been installed. These findings underscore the critical importance of these locations for future investments in wind farm development. The proposed framework could be highly efficient for wind farm siting studies in provinces that share similar characteristics with the study area used in this study offering a robust tool for future wind farm siting studies. Finally, the study contributes to achieving UN SDG 7 by employing ML to refine wind farm site selection, thereby facilitating the transition towards sustainable and universally accessible energy.