Studi Komparatif Model Extreme Gradient Boosting dan Long Short Term Memory untuk Prediksi Jangka Pendek Daya Keluaran PLTS dengan Pendekatan Multi-site
Iftor Ilyas Rizky, Prof. Nazrul Effendy, S.T., MT.,Ph.D. ; Dr.Eng. Dwi Joko Suroso, S.T., M.Eng.
2026 | Tesis | MAGISTER TEKNIK FISIKA
Pemanfaatan
energi terbarukan dapat menjadi alternatif solusi untuk permasalahan emisi dan
keberlanjutan dalam pemenuhan kebutuhan listrik. Namun, daya keluaran Pembangkit
Listrik Tenaga Surya (PLTS) sebagai salah satu energi terbarukan yang masif
dikembangkan bersifat intermiten karena bergantung pada paparan iradiasi
matahari yang diterima. Variabilitas iradiasi menyebabkan fluktuasi daya keluaran
sehingga diperlukan prediksi daya PLTS yang akurat. Penelitian ini bertujuan
mengevaluasi performa beberapa model prediksi daya keluaran PLTS serta
membandingkan pengaruh pendekatan single-site dan multi-site pada
model XGBoost dan LSTM terhadap akurasi prediksi di Sistem Lombok. Model
prediksi yang dibangun dalam penelitian adalah model konvensional yaitu model Persistence,
Regresi Linear, Auto Regresi dan model berbasis machine learning (ML)
yaitu Extreme Gradient Boosting (XGBoost) dan Long Short Term Memory (LSTM). Semua
model diuji pada horizon prediksi jangka pendek yaitu 30, 60 dan 120 menit ke depan.
Penelitian
ini menggunakan data operasional dan meteorologi PLTS Pringgabaya, Selong, dan Sengkol yang
bersumber dari data Supervisory Control and Data Acquisition (SCADA)
sistem Lombok yang beresolusi 10 menit dan sepanjang 1 tahun. Model XGBoost dan
LSTM dioptimasi menggunakan Bayesian Optimization untuk memperoleh
hyperparameter terbaik. Evaluasi performa menggunakan Mean Absolute Error
(MAE) dan R² di mana menunjukkan bahwa pendekatan multi-site secara umum
meningkatkan akurasi prediksi, terutama pada horizon yang lebih panjang. XGBoost
multi-site memberikan performa terbaik pada prediksi 30 menit ke depan
(MAE 0,99 MW; R² 0,87), sedangkan LSTM multi-site menunjukkan performa
terbaik pada prediksi 60 menit ke depan (MAE 1,2 MW; R² 0,82) dan 120 menit ke
depan (MAE 1,3 MW; R² 0,80). Untuk menginterpretasikan model, analisis SHapley
Additive exPlanations (SHAP) menunjukkan bahwa informasi antar site berkontribusi
terhadap peningkatan akurasi model karena membantu model dalam menangkap
dinamika spasial-temporal. Berdasarkan hasil penelitian, dapat disimpulkan
bahwa pendekatan multi-site efektif meningkatkan akurasi prediksi daya
PLTS pada Sistem Lombok, dengan XGBoost memberikan performa terbaik pada
horizon 30 menit dan LSTM pada horizon 60 serta 120 menit.
Renewable energy utilization has become
a promising solution to address greenhouse gas emissions and ensure the
sustainability of electricity supply. However, the power output of Photovoltaic
(PV) power plants, as one of the most widely deployed renewable energy sources,
is inherently intermittent due to its dependence on incoming solar irradiance.
Variations in irradiance lead to fluctuations in power output, making accurate
PV power forecasting essential. This study aims to evaluate the performance of
several PV power forecasting models and compare the effects of single-site
and multi-site approaches on the prediction accuracy of Extreme Gradient
Boosting (XGBoost) and Long Short-Term Memory (LSTM) models in the Lombok power
system. The forecasting models developed include conventional models, namely
Persistence, Linear Regression, and Autoregressive models, as well as machine
learning (ML)-based models, namely Extreme Gradient Boosting (XGBoost) and Long
Short-Term Memory (LSTM). All models were evaluated for short-term forecasting
horizons of 30, 60, and 120 minutes ahead.
This study employed one year of operational and meteorological data collected at 10-minute intervals from the Pringgabaya, Selong, and Sengkol PV power plants through the Supervisory Control and Data Acquisition (SCADA) system of the Lombok power system. The XGBoost and LSTM models were optimized using Bayesian Optimization to obtain the optimal hyperparameters. Model performance was evaluated using Mean Absolute Error (MAE) and the coefficient of determination (R²), demonstrating that the multi-site approach generally improved forecasting accuracy, particularly for longer forecasting horizons. The multi-site XGBoost model achieved the best performance for the 30-minute-ahead forecast (MAE = 0.99 MW; R² = 0.87), while the multi-site LSTM model outperformed other models for the 60-minute-ahead (MAE = 1.2 MW; R² = 0.82) and 120-minute-ahead forecasts (MAE = 1.3 MW; R² = 0.80). To improve model interpretability, SHapley Additive exPlanations (SHAP) analysis revealed that information from neighboring sites contributed to higher prediction accuracy by enabling the models to better capture spatiotemporal dynamics. The results demonstrate that the multi-site approach effectively improves PV power forecasting accuracy in the Lombok power system, with XGBoost achieving the best performance for the 30-minute-ahead horizon and LSTM for the 60- and 120-minute-ahead horizons.
Kata Kunci : Prediksi daya PLTS, XGBoost, LSTM, Pendekatan multi-site, Bayesian Optimization