Model Deteksi Pencurian Tenaga Listrik pada Gardu di PLN UID Jawa Barat Menggunakan XGBoost dengan Mempertimbangkan Aspek Spasial

Isnaini, Fauziyah Nurul (2026) Model Deteksi Pencurian Tenaga Listrik pada Gardu di PLN UID Jawa Barat Menggunakan XGBoost dengan Mempertimbangkan Aspek Spasial. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5003221170-Undergraduate_Thesis.pdf Restricted to Repository staff only Download (4MB) | Request a copy

Abstract

Pencurian listrik merupakan penyebab utama susut non-teknis yang menimbulkan kerugian finansial signifikan bagi sektor kelistrikan di Indonesia, termasuk PT PLN Unit Induk Distribusi (UID) Jawa Barat selaku unit distribusi terbesar di Indonesia. Berbagai upaya penanganan telah dilakukan, tapi belum sepenuhnya optimal karena keterbatasan infrastruktur dan sumber daya inspeksi, sehingga diperlukan mekanisme penentuan prioritas lokasi inspeksi yang berbasis data. Penelitian terdahulu mengenai pencurian listrik umumnya banyak berfokus pada pendekatan berbasis pelanggan yang menghadapi permasalahan berupa kebutuhan data dan komputasi yang besar, sementara pendekatan berbasis gardu hanya terbatas pada estimasi susut non-teknis secara agregat. Penelitian ini hadir dengan tujuan mengembangkan model hit rate sebagai indikator keberhasilan inspeksi pada level gardu menggunakan Extreme Gradient Boosting (XGBoost), dengan mempertimbangkan karakteristik operasional, ekonomi, dan spasial gardu. Hasil analisis menunjukkan bahwa distribusi hit rate bersifat zero-inflated, dengan 30,7% gardu bernilai nol, sementara gardu dengan temuan pencurian tersebar pada rentang (0, 1] dan cenderung mengelompok pada koordinat wilayah tertentu di Jawa Barat. Pemodelan XGBoost dengan pendekatan regresi menunjukkan keterbatasan dalam menjelaskan variasi hit rate akibat dominasi nilai nol pada data target, dengan RMSE sebesar 0,2370 dan R² 29,46% pada skema terbaiknya. Sebagai penanganan, variabel hit rate ditransformasi menjadi klasifikasi biner dan diperoleh skema terbaik pada pembagian data 80:20 dengan Stratified 5-Fold CV pada model XGBoost yang menghasilkan akurasi 0,7196, recall 0,8621, precision 0,7635, F1-score 0,8098, dan ROC-AUC 0,7235. Ini menunjukkan bahwa pendekatan klasifikasi biner menghasilkan performa evaluasi yang lebih baik dibandingkan regresi dalam memodelkan data hit rate yang zero-inflated, meskipun kemampuan model dalam mengenali kelas 0 secara tepat masih perlu ditingkatkan. Interpretasi model menggunakan SHAP menunjukkan bahwa cakupan gardu, kapasitas gardu, rasio piutang, longitude, dan latitude adalah variabel paling berpengaruh, yang menegaskan bahwa faktor operasional, ekonomi, dan spasial secara simultan berperan dalam menentukan hasil prediksi.
===================================================================================================================================
Electricity theft is a major cause of non-technical losses that generates significant financial harm to the electricity sector in Indonesia, including PT PLN Unit Induk Distribusi (UID) West Java as the largest distribution unit in Indonesia. Various mitigation efforts have been undertaken, but they remain suboptimal due to limitations in infrastructure and inspection resources, necessitating a data-driven mechanism for prioritizing inspection locations. Prior research on electricity theft has predominantly focused on customer-based approaches, which face challenges related to extensive data and computational requirements, while substation-based approaches remain limited to aggregate estimation of non-technical losses. This study aims to develop a hit rate model as an indicator of inspection effectiveness at the substation level using Extreme Gradient Boosting (XGBoost), taking into account the operational, economic, and spatial characteristics of substations. The analysis results show that the hit rate distribution is zero-inflated, with 30.7% of substations having a value of zero, while substations with confirmed theft findings are distributed within the range of (0, 1] and tend to cluster in specific geographic coordinates within West Java. XGBoost regression modeling shows limitations in explaining the variation in hit rate due to the dominance of zero values in the target variable, yielding an RMSE of 0.2370 and an R² of 29.46% in its best scheme. As a remedial measure, the hit rate variable was transformed into a binary classification, and the best scheme was obtained using an 80:20 data split with Stratified 5-Fold CV on the XGBoost model, achieving an accuracy of 0.7196, recall of 0.8621, precision of 0.7635, F1-score of 0.8098, and ROC-AUC of 0.7235. This indicates that the binary classification approach yields better evaluation performance than regression in modeling zero-inflated hit rate data, although the model's ability to accurately identify class 0 still needs improvement. Model interpretation using SHAP shows that substation coverage, substation capacity, receivables ratio, longitude, and latitude are the most influential variables, confirming that operational, economic, and spatial factors simultaneously play a role in determining the model's predictions.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Gardu, Hit rate, Pencurian listrik, SHAP, XGBoost, Electricity theft, Hit rate, SHAP, Transformer, XGBoost
Subjects:	Q Science > Q Science (General) > Q325.5 Machine learning. Support vector machines.
Divisions:	Faculty of Mathematics, Computation, and Data Science > Statistics > 49201-(S1) Undergraduate Thesis
Depositing User:	Fauziyah Nurul Isnaini
Date Deposited:	01 Aug 2026 07:50
Last Modified:	01 Aug 2026 07:50
URI:	http://repository.its.ac.id/id/eprint/141907

Actions (login required)

View Item