Evaluasi Performa Model LSTM, GRU, dan Transformer dengan Incremental Learning Menggunakan Data Streaming Beban Listrik Real-Time pada Sistem Kelistrikan Kalimantan Barat

Hakim, Arif Rahman (2026) Evaluasi Performa Model LSTM, GRU, dan Transformer dengan Incremental Learning Menggunakan Data Streaming Beban Listrik Real-Time pada Sistem Kelistrikan Kalimantan Barat. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 602224118-Master-Thesis.pdf - Accepted Version Restricted to Repository staff only Download (21MB) | Request a copy

Abstract

Peramalan beban listrik jangka pendek (Short-Term Load Forecasting atau STLF) merupakan komponen penting dalam mendukung operasi sistem tenaga listrik yang andal dan ekonomis. Seiring meningkatnya ketersediaan data operasional dari sistem Supervisory Control and Data Acquisition (SCADA), metode peramalan dituntut tidak hanya memiliki akurasi yang tinggi, tetapi juga mampu beradaptasi terhadap perubahan pola beban yang bersifat dinamis. Namun demikian, sebagian besar penelitian sebelumnya masih menggunakan pendekatan offline learning sehingga model tidak dapat memperbarui pengetahuannya ketika data baru terus tersedia selama operasi sistem. Penelitian ini bertujuan mengevaluasi performa model Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), dan Transformer yang dipadukan dengan pendekatan incremental learning untuk peramalan beban listrik jangka pendek menggunakan data streaming pada Sistem Kelistrikan Kalimantan Barat. Penelitian memanfaatkan data historis SCADA sebagai pelatihan awal model dan data streaming sebagai masukan untuk pembaruan model secara berkala. Evaluasi dilakukan menggunakan metrik Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), koefisien determinasi (R²), serta Diebold-Mariano Test. Hasil penelitian menunjukkan bahwa penerapan incremental learning mampu meningkatkan kemampuan adaptasi model terhadap perubahan pola beban dibandingkan pendekatan offline learning. Di antara model yang dievaluasi, GRU memberikan performa terbaik, sedangkan pembaruan model secara berkala terbukti mampu mempertahankan dan meningkatkan akurasi prediksi tanpa memerlukan pelatihan ulang menggunakan seluruh data historis. Selain itu, hasil Diebold-Mariano Test menunjukkan bahwa penerapan incremental learning memberikan peningkatan performa yang signifikan secara statistik dibandingkan pendekatan offline learning. Penelitian ini memberikan kontribusi berupa pengembangan kerangka kerja forecasting adaptif berbasis data streaming SCADA yang mampu memperbarui model secara berkala tanpa memerlukan pelatihan ulang menggunakan seluruh data historis, sehingga dapat mendukung pengambilan keputusan operasi sistem tenaga listrik secara real-time, meningkatkan kualitas perencanaan beban, dan membantu pengelolaan cadangan daya secara lebih efisien.
=======================================================================================================================================
Short-Term Load Forecasting (STLF) plays a crucial role in supporting reliable and economical power system operation. With the increasing availability of operational data from Supervisory Control and Data Acquisition (SCADA) systems, forecasting methods are required not only to achieve high prediction accuracy but also to continuously adapt to dynamic changes in load patterns. However, most previous studies have relied on offline learning approaches, preventing forecasting models from updating their knowledge as new data continuously become available during system operation. This study aims to evaluate the performance of Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Transformer models integrated with an incremental learning approach for short-term load forecasting using streaming data from the West Kalimantan Power System. Historical SCADA data were used for initial model training, while streaming data were utilized for periodic model updates. Model performance was evaluated using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), the coefficient of determination (R²), and the Diebold–Mariano Test. The results demonstrate that the implementation of incremental learning enhances the models' ability to adapt to changing load patterns compared with the conventional offline learning approach. Among the evaluated models, the GRU model achieved the best forecasting performance, while periodic model updates were shown to maintain and improve forecasting accuracy without requiring retraining on the entire historical dataset. Furthermore, the Diebold–Mariano Test confirmed that the implementation of incremental learning produced statistically significant performance improvements over the offline learning approach. This study contributes by developing an adaptive forecasting framework based on SCADA streaming data that enables periodic model updates without requiring full retraining on the entire historical dataset. The proposed framework can support real-time operational decision-making, improve load forecasting quality, and enhance reserve power management efficiency in power system operations.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Short-Term Load Forecasting, LSTM, GRU, Transformer, Incremental Learning, Streaming Data, SCADA, Forecasting Beban Listrik
Subjects:	T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK1001 Production of electric energy or power T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK1007 Electric power systems control
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Electrical Engineering > 20101-(S2) Master Thesis
Depositing User:	Arif Rahman Hakim
Date Deposited:	03 Aug 2026 03:58
Last Modified:	03 Aug 2026 03:58
URI:	http://repository.its.ac.id/id/eprint/143116

Actions (login required)

View Item