Research on the Cause Diagnosis Method of Line Loss Anomaly Based on the Analysis of Electricity Fluctuation in the Substation Area

Abstract

This paper presents a three-stage line loss anomaly diagnosis method based on power consumption fluctuation analysis in transformer districts. The proposed approach integrates causal forest modeling, SHAP value attribution, and TCN time series convolution networks to identify anomalies. By applying threshold-based analysis of power supply rate changes and consumption rate variations, it differentiates between grid-side anomalies, user-side anomalies, and composite anomalies. The causal forest model quantifies individual user processing effects, while the TCN network calculates load curve similarity to precisely detect hidden anomalies (such as minor power surges accompanied by load pattern deviations) that traditional methods might miss. SHAP values are used to visualize key feature contributions, enhancing diagnostic interpretability and field verification efficiency. Using real-world data from typical transformer districts, this study demonstrates the effectiveness and practicality of the proposed method, discusses its limitations, and outlines future research directions.