Combining Endpoint Detection and One-Dimensional CNN-Based Classifier for Non-Technical Loss Screening in Smart Grids

Automated meter reading systems (AMRS) are employed to measure and record electricity consumption data in smart grids, facilitating the construction of various consumption models for residential, commercial, and industrial electricity users. These standard electricity-consumption models (SECM), along with their associated consumption patterns, can be further used for applications in load forecasting, technical loss (TL) analysis, and non-technical loss (NTL) detection. Therefore, this study applies the SECM to NTL detection by proposed novel screening model that combines the endpoint detection (EPD) and short-time Fourier transform (STFT) with a one-dimensional (1D) convolutional neural network (1D CNN) based classifier. The EPD method is employed to preliminarily detect discrepancies between metering data and historical records, focusing on the time-domain variations in electricity consumption. Subsequently, the STFT is applied to analyze the frequency contains in the drastically changing time-domain data and then generates the visualization color feature patterns. With theses feature patterns, the 1D-CNN based classifier is used to identify the data into normal (Nor), suspected incidents (SI), fraud incidents (FI), and fault or power outage (OUT) events. In experimental tests, this study simulates the electricity consumption data involving Nor and NTL events by using SECM for training and validating the “STFT +1D-CNN-based classifier” model. Through ten-fold cross-validation, the feasibility of the proposed method is evaluated, and the experimental results demonstrate the promising performances, achieving average Recall (%) of 99.12%, average Precision (%) of 87.50%, and average Accuracy (%) of 90.48%, and average F1 Score of 0.9333 for NTL (SI and FI) events detection, and also show average Recall (%) of 99.35%, average Precision (%) of 86.36%, and average Accuracy (%) of 88.89%, and average F1 Score of 0.9268 for NTL and OUT events detection, respectively. These findings confirm the effectiveness and robustness of the proposed classifier model.