Electronic Theses and Dissertation

Pembangkit Listrik Tenaga Mikrohidro (PLTMH) memiliki potensi besar sebagai energi terbarukan, namun produksinya bersifat fluktuatif akibat pengaruh dinamika hidrologi dan lingkungan sehingga menyulitkan perencanaan operasional. Penelitian ini bertujuan mengembangkan model prediksi produksi energi listrik PLTMH berbasis Deep Neural Network (DNN). Model dibangun menggunakan data historis produksi energi (kWh) serta variabel cuaca berupa temperatur udara, kelembapan relatif, dan kecepatan angin. Analisis korelasi Pearson digunakan untuk mengevaluasi kontribusi variabel cuaca. Beberapa skenario diuji, meliputi variasi jendela input (12, 24, 36, dan 48 jam), penggunaan dropout 0,1, serta optimasi learning rate adaptif. Evaluasi dilakukan menggunakan RMSE, MAE, MAPE, dan R². Hasil menunjukkan bahwa konfigurasi terbaik diperoleh pada jendela input 36 jam tanpa dropout dan tanpa optimasi learning rate adaptif dengan RMSE 145,824 kWh, MAE 49,535 kWh, MAPE 4,48%, dan R² 0,975. Variabel cuaca tidak memberikan peningkatan akurasi yang signifikan. Proyeksi model menunjukkan produksi energi sebesar 31,17 GWh (2026), 35,49 GWh (2027), dan 36,23 GWh (2028). Nilai ini digunakan sebagai dasar perencanaan operasional dan estimasi alokasi biaya energi, yang diperkirakan meningkat dari Rp68,14 miliar pada 2026 menjadi Rp80,32 miliar pada 2028.

Kata kunci: Prediksi energi listrik, PLTMH, Deep Neural Network, Jendela input, Variabel cuaca.

Micro-hydropower plants (MHPs) have significant potential as renewable energy sources; however, their electricity production is highly fluctuating due to hydrological and environmental dynamics, which complicates operational planning. This study aims to develop a Deep Neural Network (DNN)-based model for predicting MHP energy production. The model is constructed using historical energy production data (kWh) and weather variables, including air temperature, relative humidity, and wind speed. Pearson correlation analysis is employed to evaluate the contribution of weather variables to the prediction target. Several experimental scenarios are tested, including variations in input window length (12, 24, 36, and 48 hours), application of dropout (0.1), and adaptive learning rate optimization. Model performance is evaluated using RMSE, MAE, MAPE, and R². The results indicate that the best performance is achieved with a 36-hour input window without dropout and without adaptive learning rate optimization, yielding RMSE of 145.824 kWh, MAE of 49.535 kWh, MAPE of 4.48%, and R² of 0.975. The inclusion of weather variables does not significantly improve prediction accuracy. The model projects annual energy production of 31.17 GWh (2026), 35.49 GWh (2027), and 36.23 GWh (2028). These projections provide a quantitative basis for operational planning and energy budget allocation, which is expected to increase from approximately IDR 68.14 billion in 2026 to IDR 80.32 billion in 2028. Keywords: Electricity production forecasting, micro-hydropower plant, deep neural network, input window, weather variables.