Electronic Theses and Dissertation

Unit Layanan Pelanggan (ULP) Sabang sebagai sistem kelistrikan terisolasi menghadapi tantangan dalam memenuhi kebutuhan energi yang terus meningkat akibat perkembangan sektor pariwisata, pertumbuhan pelanggan, serta dinamika konsumsi harian yang fluktuatif. Kondisi ini diperparah oleh keterbatasan kapasitas pembangkit berbasis PLTD sehingga diperlukan prediksi beban puncak jangka panjang yang akurat untuk mencegah risiko defisit daya maupun oversupply. Permasalahan utama penelitian ini adalah rendahnya akurasi metode prediksi konvensional dalam menangkap pola temporal dan nonlinier beban listrik. Untuk mengatasi keterbatasan tersebut, penelitian ini mengusulkan penggunaan arsitektur Recurrent Neural Network (RNN), khususnya Long Short-Term Memory (LSTM) dan Gated Recurrent Unit (GRU), sebagai pendekatan prediktif yang lebih adaptif terhadap dinamika data deret waktu. Metode penelitian meliputi pra-pemrosesan data historis beban puncak harian, konfigurasi hyperparameter, pelatihan model menggunakan sequence length tujuh hari, serta evaluasi performa menggunakan metrik Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), dan Mean Absolute Percentage Error (MAPE). Hasil penelitian menunjukkan bahwa kedua model mampu mempelajari pola temporal dengan baik, namun GRU memberikan performa lebih unggul dengan nilai MAE 0.388, RMSE 0.486, dan MAPE 8.23%, dibandingkan LSTM dengan MAPE 8.62%. Model yang telah dilatih kemudian digunakan untuk memproyeksikan beban puncak bulan Januari 2030 menggunakan pendekatan autoregressive. Hasil prediksi menunjukkan bahwa GRU lebih responsif terhadap perubahan beban yang fluktuatif, sedangkan LSTM lebih stabil dalam menangkap pola musiman. Proyeksi beban puncak Januari 2030 yang dihasilkan kedua model berada pada rentang yang realistis dan konsisten dengan tren historis, sehingga layak dijadikan acuan dalam perencanaan kapasitas pembangkit dan strategi pengelolaan sistem kelistrikan di ULP Sabang.

Kata kunci: Beban Puncak, RNN, LSTM, GRU, ULP Sabang

ULP Sabang, as an isolated electricity system, faces challenges in meeting its ever-increasing energy needs due to the development of the tourism sector, customer growth, and fluctuating daily consumption dynamics. This condition is exacerbated by the limited capacity of diesel-based power plants, requiring accurate long-term peak load predictions to prevent the risk of power deficits and oversupply. The main problem of this research is the low accuracy of conventional prediction methods in capturing temporal and nonlinear patterns of electricity loads. To overcome these limitations, this study proposes the use of Recurrent Neural Network (RNN) architectures, specifically Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU), as a more adaptive predictive approach to the dynamics of time series data. The research methods include pre-processing of historical daily peak load data, hyperparameter configuration, model training using a seven-day sequence length, and performance evaluation using Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE) metrics. The results showed that both models were able to learn temporal patterns well, but GRU provided superior performance with MAE values of 0.388, RMSE of 0.486, and MAPE of 8.23%, compared to LSTM with MAPE of 8.62%. The trained models were then used to project the peak load in January 2030 using an autoregressive approach. The prediction results showed that GRU was more responsive to fluctuating load changes, while LSTM was more stable in capturing seasonal patterns. The peak load projections for January 2030 generated by both models were within a realistic range and consistent with historical trends, making them suitable as a reference in generating capacity planning and electricity system management strategies in ULP Sabang. Keywords: Peak Load, RNN, LSTM, GRU, ULP Sabang.