Smart Meter Fault Classification Method with Overlapping Samples Weight Optimization and Dual-Model Decision Under Imbalanced Datasets

Accurate classification of smart meter faults helps reduce maintenance costs and improve grid stability and safety. Existing methods in the field of imbalanced classification often fail to fully utilize the differences between overlapping and safe zones, thus impacting classification performance. The paper proposes a dual-model decision classification method, consisting of a multi-label confidence comparisons model that focuses on overlapping regions and a GAN-based data balancing model. (OMLCC-GAN). It designs a weight calculation strategy to describe the degree to which samples approach the data overlapping area. Based on constructing sample-neighbors pairs, it adjusts the sampling frequency of neighboring samples to deeply explore adequate sample neighborhood relationship information in the overlapping area for inferring the labels of test samples. Simultaneously, after balancing the dataset using a generative adversarial network, a neural network classifier is employed to predict sample categories, focusing on sample features and typical samples in the safe zone. Finally, the predicted sample labels from the two methods are integrated proportionally, significantly enhancing model robustness. Experimental results on 11 public datasets and the smart meter fault dataset show that the proposed method outperforms typical methods in terms of F-measure and G-mean metrics.