Real-Time Abnormal Alarm Method for Distribution Network Equipment Based on Gaussian Mixture Model

Aiming to address the limitations of traditional static threshold methods and existing machine learning models in distribution network equipment health assessment, this paper proposes a novel real-time abnormal alarm method based on Gaussian Mixture Model (GMM). Traditional methods struggle to distinguish between genuine abnormalities and random fluctuations, while existing machine learning models heavily depend on high-quality data. Our proposed method constructs multi-time-scale GMM models, leveraging historical operation data to dynamically characterize equipment status features. By combining marginal probability density calculation with an adaptive threshold adjustment mechanism, it enables real-time detection of abnormalities in multi-dimensional parameters such as voltage and current. For scenarios with monitoring blind spots and data missing, our model employs a dynamic update strategy to reduce dependence on data integrity. It designs hierarchical alarm triggering rules based on abnormality levels (general, severe, critical), effectively decreasing false and missed alarms. Experiments using ring-main unit partial discharge intensity data demonstrate the method's effectiveness. Even with partial data missing, it maintains high reliability, improves recognition accuracy for hidden and highly random abnormalities, shortens response time to seconds, and reduces the false alarm rate through a dynamic threshold mechanism. This study offers a solution that is both general and adaptable for distribution network equipment health management. It provides theoretical support for fault warning and operation optimization in smart grids. The proposed method's ability to handle data missing and its independence from data quality make it highly practical for real-world applications. Future work will focus on further optimizing the model and expanding its application scope within smart grid management.