OPTIMIZATION OF ANOMALY DETECTION IN ELECTRICITY USAGE FOR PREPAID CUSTOMERS USING MACHINE LEARNING

• Main Author: Kurniawan, Rahmat
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/87709
• Institution: Institut Teknologi Bandung

Electricity theft and anomalies in electricity usage pose signijicant challenges for utility companies, leading to financial losses, power supply disruptions, and even fire risks caused by short circuits or overloads. In contrast to previous studies that focused more on postpaid customers, this study focuses on prepaid customers who have different characteristics. Postpaid customers pay bills every month, while prepaid customers purchase energy in advance that can be used for the next few months. In addition, prepaid customers are not yet equipped with a real-time electricity consumption reading system and still use one-way/ojjline kWh meters. This study develops an anomaly detection system for electricity usage in prepaid customers with a data-driven approach and supervised learning-based machine learning algorithms namely Classical Classijier (Logistic Regression, KNN, Decision Tree) and Ensemble Classijier (AdaBoost, Random Forest, XGBoost) with data imbalance handling using SMOTE oversampling or undersampling. The data capturing technique is applied in forming the training dataset by cutting off prepaid customer transaction data before inspection. This aims to describe customer behavior before the inspection is carried out and to reduce the time span of the data used. The results show that XGBoost with the SMOTE undersampling imbalance handling technique and data capturing technique provides the best performance with high F 1-Score and hit rate values compared to other models, especially when tested on new data or testing data. This research contributes to the development of machine learning-based solutions to detect anomalies in electricity usage among prepaid customers.