This paper presents a hybrid framework integrating Artificial Intelligence (AI), Data Envelopment Analysis (DEA), and Explainable AI (XAI) for smart grid optimization. The central challenge addressed is the rarely solved difficulty of simultaneously combining predictive accuracy, operational efficiency assessment, and decision-making transparency within a single coherent system. The proposed solution relies on a sequential five-stage pipeline: energy prediction via Long Short-Term Memory (LSTM) networks and Gradient Boosting, node efficiency evaluation through DEA-CCR and DEA-BCC models, and algorithmic decision interpretation via SHAP and LIME. Experiments conducted on a synthetic dataset of 120 smart grid nodes, validated on the real-world PJM hourly energy consumption dataset, yield a predictive coefficient R² = 0.967 and reveal a mean efficiency improvement potential of 23% for underperforming nodes. The local accuracy property of SHAP is verified, and DEA score stability is confirmed by bootstrap analysis.
Darha et al. (Wed,) studied this question.