Lightweight anomaly detection for power grid microservices based on temporal-aware fusion and graph neural networks

Abstract

Microservice architecture is widely adopted in the digital transformation of smart grids, yet its distributed nature poses dual challenges for anomaly detection: the strong temporal periodicity of power grid data and the obscurity of fault propagation paths. This paper proposes a lightweight anomaly detection method integrating temporal-aware multimodal fusion and Graph Neural Network (GNN). Its core innovations are as follows: A lightweight temporal attention fusion strategy adaptive to power grid load periodicity, which enhances fusion accuracy without complex pre-training; A simplified GNN-based root cause localization method leveraging a compact power grid domain knowledge graph. Experiments on real power grid data demonstrate that the proposed method achieves an F1-score of 92.3% and a localization latency of 2.2 seconds, outperforming traditional methods by 2.5 - 3.0 percentage points in performance. Our research provides an efficient solution for the operation and maintenance of smart grid microservices.