Multi-Sensor Partial Discharge Prediction in Switchgear Cabinets Using Deformable Attention Transformer

Partial discharge (PD) prediction is essential for ensuring the reliability and safety of switchgear cabinets, yet it remains challenging due to the complex, multi-sensor data required to accurately capture PD events. A primary difficulty lies in dynamically extracting relevant temporal features from sensors with varied sampling frequencies. To address this, we propose a Deformable Attention Transformer (DAT) that adapts attention across temporal and feature dimensions, effectively capturing local patterns across diverse inputs. Additionally, we incorporate Gaussian Mixture Models (GMM) to introduce data-driven priors, aligning with the probabilistic distribution of switchgear states and enhancing the model's ability to distinguish true PD events from normal fluctuations. Our proposed model, particularly the TCA-Transformer variant, demonstrates a 3% improvement in accuracy and a 5% increase in AUC over standard Transformer models, along with a reduction in time prediction error from 7 seconds to 2 seconds, underscoring its suitability for real-time predictive maintenance. These advancements affirm the theoretical and practical significance of our approach, establishing a robust foundation for more reliable and proactive PD detection in power system applications.