Uncertainty-aware normal behavior modeling with spatiotemporal graph learning for health monitoring of pumped storage units

Pumped storage units (PSUs) are vital for energy balancing and frequency regulation in power systems with high renewable penetration. Ensuring their operational reliability through effective health condition monitoring is essential not only for the PSUs themselves but also for grid-wide stability. However, the degradation of PSUs is governed by the complex spatiotemporal interaction of hydraulic, mechanical, and electromagnetic factors under frequent switching of operational modes, along with inherent uncertainties. These characteristics challenge existing monitoring approaches in accurately capturing abnormal behaviors. To address this, we propose a novel normal behavior model (NBM) that combines hypervariable spatiotemporal graph (HSG) learning with relevance vector machine (RVM)-based uncertainty quantification. This model learns the expected healthy behaviors of PSUs across diverse operating conditions, leveraging HSG’s strength in unified spatiotemporal feature representation and RVM’s capability for sparse modeling and uncertainty estimation. Within the proposed framework, an entropy-regularized Wasserstein distance measures the deviation between actual vibration signals and the NBM-generated baseline, which is defined as the performance degradation index (PDI). A dynamic weighting strategy, based on signal fluctuation characteristics, integrates PDIs from multiple monitoring points to form a holistic health indicator. Subsequently, a Gaussian Mixture Model (GMM)-based method establishes adaptive thresholds for detecting abnormal degradation. The method is validated using data from a PSU in Zhejiang Province, China. Experimental results demonstrate its superior performance compared to existing approaches, and alignment with real-world maintenance records confirms its practical applicability and effectiveness.