Anomaly Detection for Small Hydropower Based on Deep Spatio-Temporal Modeling

Anomaly detection in small hydropower plays a crucial role in small hydropower data-driven condition monitoring systems, contributing to improved equipment durability and reduced operational and maintenance expenses. Benefiting from advancements in the industrial internet, various sensors in small hydropower stations generate co-evolving time-series data with distinct characteristics at any given moment, which are recorded in the small hydropower condition monitoring system. The joint modeling of variable associations and temporal dependencies in small hydropower multivariate time-series data presents significant challenges for anomaly detection tasks. This study characterizes the complex relationships within small hydropower data as a combination of temporal dependencies and inter-feature correlations. To address these, it introduces a novel anomaly detection framework for multivariate time series, integrating adaptive graph structure learning, graph attention mechanisms, and temporal feature pyramid networks to effectively capture spatiotemporal dependencies. The proposed method aims to dynamically capture the most significant relationships, enabling timely detection of potential anomalies for more reliable small hydropower station monitoring. Experimental results on real-world sensor datasets from small hydropower stations show that the proposed method outperforms traditional approaches in anomaly detection, with superior capability in modeling the spatiotemporal dependencies inherent in multivariate time-series data.