Automasi Evaluasi Kinerja Dan Optimasi Jadwal Pembersihan Untuk Sistem Fotovoltaik Terdistribusi Menggunakan Algoritma Machine Learning

Darmayasa, Made (2026) Automasi Evaluasi Kinerja Dan Optimasi Jadwal Pembersihan Untuk Sistem Fotovoltaik Terdistribusi Menggunakan Algoritma Machine Learning. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6022241105-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (17MB) | Request a copy

Abstract

Penentuan prioritas pembersihan pada armada sistem fotovoltaik (PV) atap terdistribusi lazimnya bertumpu pada hari sejak pembersihan terakhir (days since cleaning, DSC) atau deviasi performance ratio (PR). Kedua metode mengurutkan situs tanpa menaksir berapa energi yang sebenarnya akan pulih bila tiap situs dibersihkan. Penelitian ini menyajikan kerangka yang melatih stacking ensemble per klaster, gabungan LSTM, XGBoost, dan Random Forest, untuk memprediksi PR harian, lalu menerapkan inferensi counterfactual guna menaksir recoverable energy gain (ringkasnya gain) tiap situs. Kerangka dikembangkan dan divalidasi pada 124 situs PV atap di lima kawasan industri di Jawa Barat, Indonesia, dengan 15 bulan rekaman SCADA harian. Ensemble mencapai MAPE 2,6 sampai 3,5 persen pada data holdout. Prediksi counterfactual mentah meremehkan gain pembersihan dengan bias −0,016 sampai −0,043, sehingga diperlukan kalibrasi terhadap 191 event pembersihan historis. Selang kepercayaan bootstrap mengonfirmasi pemulihan PR yang positif di kelima klaster. Laju soiling bervariasi 3,0 kali antarklaster, yang mendukung pendekatan pemodelan per klaster. Pada anggaran pembersihan 20 situs, peringkat terkalibrasi memulihkan energi 87 persen lebih banyak daripada peringkat DSC dan 62 persen lebih banyak daripada peringkat deviasi PR. Keunggulan marginal atas peringkat DSC pada anggaran tersebut bernilai sekitar IDR 342 juta per tahun, diperoleh tanpa menambah kru maupun frekuensi pembersihan. Kerangka bekerja hanya dengan data SCADA dan log pemeliharaan rutin tanpa sensor tambahan.
=================================================================================================================================
Cleaning prioritization for distributed rooftop photovoltaic fleets is commonly based on days since last cleaning or performance ratio deviation. Both methods rank sites without estimating how much energy each site would actually recover if cleaned. This paper presents a framework that trains per-cluster stacking ensembles combining LSTM, XGBoost, and Random Forest to predict daily performance ratio, then applies counterfactual inference to estimate site-level recoverable energy gain. The framework was developed and validated on 124 rooftop PV sites across five industrial estates in West Java, Indonesia, using 15 months of daily SCADA records. The ensemble achieves 2.6 to 3.5% MAPE on holdout data. Raw counterfactual predictions underpredict cleaning gains with bias ranging from-0.016 to-0.043, so calibration against 191 historical cleaning events is needed. Bootstrap confidence intervals confirm positive PR recovery in all five clusters. Soiling rates vary 3.0-fold across clusters, which supports the per-cluster modeling approach. For a 20-site cleaning budget, the calibrated ranking recovers 87% more energy than days-since-cleaning ranking and 62% more than PR-deviation ranking. Against the days-since-cleaning baseline, this marginal advantage is worth about IDR 342 million per year, obtained without adding crew or cleaning frequency. The framework uses only routine SCADA and maintenance log data without additional sensors.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	sistem fotovoltaik, soiling, prioritisasi pembersihan, inferensi counterfactual, stacking ensemble, kalibrasi, SCADA, photovoltaic systems, soiling, cleaning prioritization, counter factual inference, stacking ensemble, calibration, SCADA
Subjects:	T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK1087 Photovoltaic power generation
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Electrical Engineering > 20101-(S2) Master Thesis
Depositing User:	Made Darmayasa
Date Deposited:	30 Jul 2026 03:01
Last Modified:	30 Jul 2026 03:01
URI:	http://repository.its.ac.id/id/eprint/140187

Actions (login required)

View Item