Energy Theft Detection in Smart Grids via Explainable Attention Maps

Electricity theft is a widespread issue that has a negative impact on both electricity users and utility businesses. Smart meters record energy usage at hourly or daily frequency, creating time series which can help in fraud spotting. This study utilizes a daily consumption dataset given by State Grid Corporation of China to detect electricity fraud with mediocre class imbalance. Most of the existing studies either try to maximize the models detection quality with complex neural network architectures by sacrificing interpretability or try to extract handcrafted meaningful predictors while losing accuracy. In this work the authors proposed a method capable of providing high explainability while still preserving competitive detection quality of 0.896 ROC-AUC and 0.972 MAP@100. The key contribution of this paper is reformulating the original classification problem into time series segmentation problem with the help of attention maps. New labeling is acquired by averaging decision paths from random forest models. The authors also investigate different pitfalls during validation which can lead to overly optimistic estimates of models quality.