Electricity Theft Detection from Smart-Meter Data Using Wide and Deep CNN

Electricity theft (meter tampering, bypassing the meter) is the main non-technical loss of distribution utilities. Smart meters record the daily consumption of every customer, and theft changes that pattern - but a utility can inspect only a few meters, so the task is to rank customers for a field inspection. This project uses the REAL SGCC smart-meter data set of the State Grid Corporation of China (42,372 customers, daily kWh from 1 January 2014 to 31 October 2016, 3,615 labelled thieves = 8.53 %, 25.7 % of the readings missing; downloaded from GitHub henryRDlab/ElectricityTheftDetection and checked by SHA-1 / SHA-256 / CRC-32) and compares a classical rule-based consumption-drop detector and logistic regression with Random Forest and LightGBM on 28 statistical features and the wide (1-D) + deep (2-D weekly image) CNN of Zheng et al. (IEEE TII 2018). On 8,346 hold-out customers (720 thieves) the drop rule is no better than random (ROC-AUC 0.498), logistic regression reaches 0.773, the wide & deep CNN 0.778 (the published value is about 0.78), the Random Forest 0.819 and the deployed LightGBM 0.856: its first 100 inspections find 91 thieves (MAP@100 0.956) and it catches half of the thieves after inspecting only 9.3 % of the customers (drop rule 46.8 %, random 50 %). GNU Octave cross-checks (run), a complete MATLAB version (written, not run) and a web application with five databases are included. Everything runs on a standard laptop CPU (8 GB RAM, no GPU).

Category: Electrical and Electronics Engineering (Power Distribution, Non-Technical Losses, Electricity Theft, Smart Meters, Advanced Metering Infrastructure, Revenue Protection, Time-Series Classification, Deep Learning, Wide and Deep Convolutional Neural Network, LightGBM, Random Forest, Logistic Regression, PyTorch, GNU Octave, MATLAB, Python)

Electrical & Electronics Engineering Project EE13 from ProTechSC Academy, with graphical abstract, video demonstration and project guidance.