A Scalable Machine Learning Framework for Hybrid Solar–Wind Site Suitability Screening

Accelerating the energy transition requires scalable and reproducible tools for identifying suitable locations for hybrid renewable energy deployment. This study develops and evaluates a cloud-based GIS–ML surrogate framework for hybrid solar–wind suitability screening in Nigeria. Multi-source geospatial datasets, including wind speed, solar irradiance, land use, topography, and infrastructure proximity, were harmonised in Google Earth Engine to generate a 30 m national predictor stack. Approximately 30,000 spatial samples were labelled into four suitability classes using literature-informed GIS–MCDM thresholds. These GIS–MCDM-derived labels were then used to train machine-learning models as surrogate learners of the criteria-based suitability surface. Model evaluation compared a baseline Random Forest under random sampling with spatially validated Random Forest (RF), Decision Tree (DT), and Logistic Regression (LR) models. The baseline RF achieved high apparent performance, with an accuracy of 0.97, while the north–south spatial holdout reduced RF accuracy to 0.76, indicating that random splitting overestimated predictive performance. Grouped spatial k-fold validation was further used to assess spatial generalisation. Under spatial validation, tree-based models outperformed LR, with DT and RF accuracies of 0.78 and 0.76, respectively, compared with 0.43 for LR. The predicted suitability surfaces reveal a strong north–south gradient, with the most favourable zones concentrated in northern Nigeria and along major infrastructure corridors. Overall, the spatially validated RF provided the strongest balance among the tested models in terms of predictive performance, spatial consistency and planning relevance. Because the training labels were derived from GIS–MCDM criteria, the framework is best interpreted as a reproducible and transferable surrogate suitability-screening tool for evidence-informed renewable-energy planning under data-limited conditions.