APPLICATION OF MACHINE LEARNING CLASSIFICATION FOR ELECTRICITY THEFT DETECTION AMONG POSTPAID RESIDENTIAL CUSTOMERS USING TRADITIONAL METERS: A CASE STUDY ON CUSTOMER SERVICE UNITS WITH THE HIGHEST INCIDENCE AT PLN

• Main Author: Pascal Taruna, Alief
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/86679
• Institution: Institut Teknologi Bandung

Electricity theft is a major challenge for PT PLN (Persero), particularly in managing 27 million postpaid customers, most of whom still use traditional meters. Detecting and addressing electricity theft has become increasingly complex, requiring more efficient approaches. Unlike smart meters, traditional meters lack communication capabilities, making electricity theft detection reliant on manual processes. This study develops a machine learning-based electricity theft detection model to optimize the Target Operation (TO) process, which identifies customers for field verification. By analyzing monthly electricity usage, particularly in the 450 VA household segment receiving government subsidies, this model aims to optimize the Target Operations (TO) formation process, which is currently conducted manually. The objective is to minimize subjective observations and ensure that subsidies are allocated accurately. Unlike other studies that commonly use open data and smart meters, this research utilizes monthly consumption data from PLN's postpaid customers, offering a novel approach to theft detection. Various classification models were tested, including Decision Tree, Random Forest, KNearest Neighbors, Logistic Regression, and Naive Bayes, with Random Forest demonstrating the best performance across various simulation scenarios. This study also introduces a sequential evaluation method that enhances accuracy through layered filtering, where detection results from the three-theft model are further filtered using the two-theft and one-theft models, resulting in a more precise TO. The combination of Random Forest and K-Nearest Neighbors achieved the best performance, with an accuracy of 0.89, precision of 0.83, recall of 0.98, F1-Score of 0.90, and an AUC of 0.89. These findings provide practical benefits for PLN by enabling a more objective and standardized TO process, reducing human error, and improving operational efficiency.