Predictive Maintenance Trafo Distribusi Menggunakan Machine Learning Untuk Meningkatkan Keakuratan Pengambilan Keputusan

Mahdi, Aulia (2025) Predictive Maintenance Trafo Distribusi Menggunakan Machine Learning Untuk Meningkatkan Keakuratan Pengambilan Keputusan. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6032222147-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only until 1 May 2027. Download (7MB) | Request a copy

Abstract

PT PLN (Persero) adalah BUMN yang bergerak di bidang ketenagalistrikan, yang bertugas selain untuk mencari keuntungan, juga memiliki tugas sebagai public service obligation dengan memberikan pelayanan yang terbaik kepada pelanggan di seluruh pelosok negeri. Dalam menjalankan proses bisnisnya di sisi distribusi, PLN melalui unit operasionalnya harus menjaga kinerja keandalan yang diukur dalam Key Performance Indicator (KPI) berupa SAIDI (System Average Interruption Duration Index), SAIFI (System Average Interruption Frequency Index), ENS (Energy Not Served), Gangguan Penyulang 20 kV, Gangguan Kubikel serta Gangguan Trafo. Dalam upaya penurunan gangguan Trafo Distribusi, PLN melaksanakan pemeliharaan prediktif berbasis kondisi atau Condition Based Maintenance (CBM). Predictive Maintenance yang dilaksanakan saat ini belum cukup efektif karena selain belum semuanya menggunakan alat ukur dan sensor yang canggih, database hasil inspeksi tidak dimanfaatkan dengan baik dan pengambilan keputusannya belum berbasis pendekatan scientific, sehingga menimbulkan potensi gangguan trafo yang tidak dapat diantisipasi. Dalam penelitian ini digunakan pendekatan Machine Learning untuk Predictive Maintenance Trafo Distribusi. Pada tahap pertama, pemodelan Unsupervised Learning metode K-Means Clustering digunakan untuk mengelompokkan aset trafo berdasarkan indikator kesehatannya. Hasil dari pelabelan cluster ini kemudian digunakan untuk melatih model Supervised Learning dengan metode Extreme Gradient Boosting (XGBoost), yang memberikan prediksi yang akurat terkait target pemeliharaan trafo. Variabel input yang digunakan dalam penelitian ini bertipe numerik yaitu usia trafo, suhu trafo, persentase pembebanan trafo dan persentase ketidakseimbangan beban trafo, yang didapatkan dari hasil inspeksi Gardu Distribusi yang terdapat pada database aplikasi Master Jardis PLN UID Banten. Hasil yang didapatkan dari pengolahan data 2709 Trafo Distribusi menggunakan Machine Learning dengan algoritma K-Means Clustering dan klasifikasi XGBoost, berupa hasil klasifikasi tingkat keandalan dan kesehatan Trafo Distribusi dalam 4 kelas, yaitu kelas Trafo Sehat sejumlah 1273 Trafo (46,99%), kelas Trafo Beban Tinggi dan Rawan Gangguan sejumlah 827 Trafo (30,53%), kelas Trafo Tua dan Rawan Gangguan sejumlah 475 Trafo (17,53%), dan kelas Trafo Losses sejumlah 134 Trafo (4,95%). Model menunjukkan performa tinggi dengan accuracy 97%, precision 92-97%, dan recall 95–98% pada semua kelas, tanpa indikasi overfitting ataupun underfitting. Dari sisi implikasi manajerial didapatkan hasil perhitungan total biaya Predictive Maintenance berbasis Machine Learning adalah sebesar Rp14.133.542.153, sedangkan total biaya pemeliharaan paduan Time-Based dan Condition-Based dengan perhitungan existing PLN adalah sebesar Rp73.920.146.452, sehingga didapatkan Cost Efficiency yang sangat baik sebesar 80,88%.
=================================================================================================================================
PT PLN (Persero) is a state-owned company engaged in the electricity sector, which is tasked not only to seek profit, but also has a duty as a public service obligation by providing the best service to customers throughout the country. In carrying out its business processes on the distribution side, PLN through its operational units must maintain reliability performance as measured by Key Performance Indicators (KPI) in the form of SAIDI (System Average Interruption Duration Index), SAIFI (System Average Interruption Frequency Index), ENS (Energy Not Served), 20 kV Repeater Disruption, Cubicle Disruption and Transformer Disruption. In an effort to reduce Distribution Transformer disturbances, PLN carries out condition-based predictive maintenance (CBM). Predictive Maintenance carried out at this time is not effective enough because in addition to not all of them using sophisticated measuring instruments and sensors, the database of inspection results is not utilised properly and decision making is not based on a scientific approach, resulting in potential transformer disturbances that cannot be anticipated. In this research, a Machine Learning approach is used for Predictive Maintenance of Distribution Transformers. In the first stage, Unsupervised Learning modelling using K-Means Clustering method is used to cluster transformer assets based on their health indicators. The results of this cluster labelling are then used to train the Supervised Learning model with the Extreme Gradient Boosting (XGBoost) method, which provides accurate predictions regarding transformer maintenance targets. The input variables used in this study are numeric types, namely transformer age, transformer temperature, transformer loading percentage and transformer load imbalance percentage, which are obtained from the results of Distribution Substation inspections contained in the PLN UID Banten Master Jardis application database. The results obtained from data processing of 2709 Distribution Transformers using Machine Learning with the K-Means Clustering algorithm and XGBoost classification, in the form of classification results of the level of reliability and health of Distribution Transformers in 4 classes, namely the Healthy Transformer class of 1273 Transformers (46.99%), the High Load and Disturbance Prone Transformer class of 827 Transformers (30.53%), the Old and Disturbance Prone Transformer class of 475 Transformers (17.53%), and the Losses Transformer class of 134 Transformers (4.95%). The model showed high performance with 97% accuracy, 92-97% precision, and 95-98% recall in all classes, with no indication of overfitting or underfitting. In terms of managerial implications, the results of the calculation of the total cost of Machine Learning-based Predictive Maintenance is Rp14,133,542,153, while the total cost of Time-Based and Condition-Based maintenance with existing PLN calculations is Rp73,920,146,452, resulting in a very good Cost Efficiency of 80.88%.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Predictive Maintenance, Trafo Distribusi, Machine Learning, K-Means Clustering, Extreme Gradient Boosting (XGBoost), Benefit Cost Analysis.
Subjects:	T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK6565.T7 Transformers
Divisions:	Interdisciplinary School of Management and Technology (SIMT) > 61101-Master of Technology Management (MMT)
Depositing User:	Aulia Mahdi
Date Deposited:	18 Feb 2025 04:13
Last Modified:	18 Feb 2025 04:13
URI:	http://repository.its.ac.id/id/eprint/118778

Actions (login required)

View Item