Regional Fault Location of Distribution Network Based on Distributed Observation and Fusion of Multi-Source Evidence

This paper proposes a multi-source evidence generation strategy (MEGS) that utilises distributed measurements to train a multi-classification support vector machine (SVM) for each observer. An observer employs time-frequency analysis to transform local current signals into feature samples, which serve as inputs to the SVM. The output of the SVM is then subjected to grey relational analysis and a voting mechanism to determine the probability of observers in identifying faults within the section. Due to the inherent uncertainty and variability of faults, the direct application of Dempster-Shafer theory (D-S theory) may result in diagnostic inaccuracies. To address this issue, we introduce an evidence fusion approach based on propositional consistency and evidence consistency (PCEC). Simulation results demonstrate that PCEC significantly enhances diagnostic accuracy beyond that achieved by individual classifiers, with an accuracy of 99.41% under ideal conditions. Factors such as load variations, sampling errors, or single observer errors may affect the quality of the evidence. However, the PCEC is effective in improving diagnostic accuracy. Further ablation studies and comparative analyses with other fusion methods validate the proposed modifications to the D-S theory as both reasonable and superior in terms of accuracy.