A GeoAI-Based Machine Learning Framework for High-Resolution Solar Energy Potential Mapping

Abstract

This paper proposes a GeoAI enabled machine learning framework to create a high resolution map for solar energy potential estimation over the state of Maharashtra, India. Using a dataset of 2.74 million spatiotemporal observations from NASA POWER (2019-2021) that spans 2,500 grids, we explicitly incorporate the spatial coordinate into our model development. Using 12 engineered features that include geographical zoning and elevation zones along with spatial cross validation and SHAP for explainability, we compare different models and demonstrate that Random Forest Regressor achieves high accuracy (R2 = 0.9965, MAE = 0.0341 kWh/m2/day) on spatially held-out test data, with substantial improvement from linear regression (R2 = 0.5462). However, validation against independent measurements is required for assessing real world applicability. Our spatial autocorrelation test using Moran’s I statistic reveals significant clustering effect (I = 0.2695), while SHAP value shows that temperature, humidity, and coordinates are the main contributors to solar energy production. Based on our map, we show that Marathwada (drought affected area consisting of Solapur and Latur) is a high potential solar zone (≈ 5.28 kWh/m2/day).