Explainable AI for harmonic fingerprinting and voltage sag diagnosis in decentralized power grids: trends, challenges and future directions

Abstract

The evolution of decentralized power grids has increased the complexity of power-quality monitoring, particularly harmonic fingerprinting and voltage sag diagnosis. Artificial intelligence improves disturbance detection and classification, yet black-box models limit transparency, engineering validation, and operator trust. This review synthesizes 108 selected studies on explainable artificial intelligence (XAI) for power-system diagnostics, focusing on SHapley Additive exPlanations (SHAP), local interpretable model-agnostic explanations (LIME), attention-based interpretability, visual analytics, and physics-informed learning. The review integrates harmonic fingerprinting with voltage sag diagnosis through their shared requirements for source attribution, temporal interpretation, physical consistency, and operator-oriented explanation. Four major deployment gaps are identified: data quality, computational latency, physical grounding, and trustworthiness. Future priorities include real-time embedded XAI, physics-informed neural networks, federated learning, standardized trustworthiness metrics, and adaptive model lifecycle management. The findings indicate that reliable autonomous diagnosis requires explainability to be integrated with predictive performance, electrical-system physics, computational efficiency, and field validation. This integration provides a stronger foundation for transparent, resilient, and trustworthy diagnostic systems in decentralized power grids.