Anomaly Detection and Repair of Drop-out Fuses in Alpine Areas Using β-TCVAE and Dynamic Bayesian Networks

Abstract

In response to the frequent break and misbreak of drop-out fuses caused by weather changes and seasonal factors in alpine areas, this article combines β-TCVAE and dynamic Bayesian network (DBN) to achieve efficient detection and repair of power load fluctuations. Firstly, β-TCVAE trains the power load data under normal operating conditions, identifies abnormal points through reconstruction errors, and locates the misbreak situations caused by fluctuations. Then, when an anomaly is identified, DBN uses time series data and models based on historical normal operating statuses to infer the normal data at that moment. By studying the temporal correlation between fuse status and power load fluctuations, DBN predicts the normal load value in the absence of anomalies. Finally, the erroneous power parameters are corrected in real-time to repair the fuse to normal. By combining two technologies and constructing a comprehensive framework, real-time monitoring and repair of power load fluctuations in alpine areas during winter peak periods, rainy season, sudden snowstorm, equipment aging, and load surge can be achieved. The results show that the method reduces the average misbreak rate of fuses in alpine areas by 8%, and improves the accuracy of abnormal data repair by an average of 27.6%. The model has stable comprehensive performance under extreme weather conditions, which is conducive to the operation and maintenance of the power system.