Bayesian–Causal Reinforcement Learning for adaptive and interpretable solar energy policy design

Résumé fourni par la source

Accelerating solar energy adoption is vital for meeting climate targets, yet conventional policies often overlook regional heterogeneity, causal relationships, and shifting socio political contexts. We propose a Bayesian Causal Reinforcement Learning (BCRL) framework that combines Bayesian Graph Neural Networks (BGNN), causal inference, and multi agent reinforcement learning to design adaptive, region specific solar policies across U.S. states. States are first clustered via BGNN using socio economic, political, and environmental features. Within each cluster, policy agents learn optimal interventions under a partially observable Markov decision process, guided by estimated causal effects of subsidies, awareness campaigns, and workforce training. SHAP based feature attribution and counterfactual analysis ensure transparency. Across all 50 U.S. states, BCRL outperforms XGBoost, standard RL, and non causal baselines, achieving a 15.3% improvement in adoption rate prediction (AUC), a 12.7% gain in job creation forecasting, and a 9.8% increase in GDP alignment. Causal aware policies further yield more robust, equitable, and long term benefits. By integrating technical rigor with policy relevance, BCRL provides an interpretable, scalable tool for accelerating solar adoption in diverse socio economic environments.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.