Outage Detection in Self-Healing Smart Grids Using Reinforcement Learning with Spectral Graph Neural Networks

A new machine-learning framework that uses spectral graph neural networks to manage power outages in distribution grids has been posted to the arXiv preprint repository. The authors report that their reinforcement-learning approach achieves near-optimal restoration performance in real time across multiple test networks [1][2]. The paper, submitted on 26 May 2026, proposes a method that learns a control policy for network reconfiguration and emergency load shedding during outages [1][2]. The authors argue that traditional machine-learning techniques are poorly suited for smart grids because of slow response times and high computational demands [1][2]. While earlier studies have applied reinforcement learning with conventional graph neural networks to the problem, those models operate in the spatial domain and can miss critical frequency-domain relationships that capture global structural patterns in power systems [1][2]. To address that gap, the researchers built a spectral graph reinforcement-learning framework that incorporates frequency-domain information directly into the policy-learning process [1][2]. The framework was evaluated on three modified IEEE test systems: the 13-bus, 34-bus, and 123-bus networks [1][2]. Experimental results indicate that the approach delivers near-optimal performance in real time and generalizes across a wide range of outage scenarios [1][2]. The work appears as a preprint on arXiv, an open-access repository that hosts electronic preprints and postprints across disciplines including electrical engineering and computer science [6]. As of November 2024, the repository was receiving about 24,000 new submissions per month [6]. Papers on arXiv are moderated but not peer-reviewed [6]. The preprint’s abstract page also surfaces several community-built discovery tools through the arXivLabs framework, which allows third-party collaborators to develop experimental features that sit alongside article records [5]. Those tools include the Bibliographic Explorer, which maps citation trees, and the CORE Recommender, which suggests related open-access papers from a global network of repositories [4][5]. arXivLabs partners are required to adhere to the repository’s values of openness, community, excellence, and user-data privacy, and they receive only minimal, anonymized user data necessary for their features to function [5].