Illuminating India’s Diverse Renewable Energy Transition: Evidence from Explainable Artificial Intelligence

Abstract

India's renewable energy transition is a subnational phenomenon characterised by profound heterogeneity across 30 states. While national aggregate capacity figures have attracted considerable scholarly attention, the structural determinants of state-level adoption divergence remain analytically underexplored. This study develops an integrated analytical framework combining K-means trajectory clustering, ensemble machine learning forecasting, and SHAP-based Explainable AI to examine state-level renewable energy adoption in India from 2014 to 2024. Using a 330-observation panel constructed from open government secondary data (MNRE, CEA, RBI, NITI Aayog), four models are trained and evaluated using year-based panel cross-validation. Random Forest achieves the highest predictive accuracy (R² = 0.9943, RMSE = 0.1694). SHAP decomposition of the Random Forest model identifies prior renewable energy capacity (mean |SHAP| = 1.882) as the dominant adoption determinant, confirming strong path dependency consistent with Diffusion of Innovations theory. Real state GDP (0.047) and governance quality proxied by SDG composite scores (0.008) represent the economic and institutional channels aligned with the Technology-Organisation-Environment (TOE) framework. K-means clustering (silhouette = 0.4746) partitions states into four trajectory groups: Innovators, Early Adopters, Early Majority, and Laggards with SHAP profiles revealing distinct predictor hierarchies across groups. Results are robust across exclusion of COVID-affected years, an alternate dependent variable specification, and multicollinearity diagnostics. The study provides a replicable open-data ML pipeline for subnational energy planning and offers cluster-specific policy prescriptions for India's 2030 renewable capacity targets.