การจำแนกหน่วยสูญเสียที่ไม่ใช่ทางเทคนิคสำหรับมิเตอร์อ่านหน่วยอัตโนมัติในระบบจำหน่ายการไฟฟ้าส่วนภูมิภาค (กฟภ.) ด้วยปัญญาประดิษฐ์

Abstract

Non-technical loss (NTL) in the distribution system of Provincial Electricity Authority (PEA) is a loss calculated from the difference between the total loss and the technical loss (TL), not considering the actual source of occurrence. Partly caused by loss not being read through meter, a defective meter, measurement error, electricity theft, etc. The Automatic Meter Reading (AMR) system has been installed for major electricity customers for monitoring, inspection, and to detect possible abnormalities. One of the limitations encountered by AMR systems is that the system can provide alarms for abnormalities but cannot classify anomaly patterns. With the advancement in technology and the ability of artificial intelligence today has been applied to solve this problem. In this research, AMR data from PEA database and the actual on-site inspection results were explored to visualize, analyze, extract, and classify data using machine learning and deep learning into three classes: normal conditions, defective meters, and energy theft patterns. The key points in this research are data extraction, which includes electrical signals and physical data. It is divided into three characteristics: (1) extracted as tabular features; (2) extracted by considering electrical signals in terms of time series as coefficients, frequency domains, and wavelet transforms; and (3) extracted by converting from time series to images. The results obtained have an accuracy of 60–70%, 70–80%, and 80–90%, respectively. Additionally, the data were balanced using an anomaly model, adaptive synthetic sampling, and image data augmentation to enhance model learning efficiency and reduce overfitting and bias. Moreover, we also provide customer categorization using K-means clustering, reducing multiple customer groups, and improving classification accuracy. Finally, the model has been exported to be tested with new data from PEA's website and case studies for real events to evaluate the prediction results from the actual on-site inspection. The results obtained are about 80–85% effective overall.