<![CDATA[Ensuring a reliable electricity supply is crucial for Indonesia's development. This study applies deep learning to forecast diesel backup power output. One challenge in such predictions is balancing the input sequence length and the number of features to avoid overly long input sequences, which may degrade model performance. To address this, we utilized an autoencoder to compress the input sequence, improving prediction accuracy. Additionally, given the time-consuming nature of hyper-parameter optimization in deep learning, we employed Bayesian optimization to streamline the process and achieve optimal hyper-parameter settings.The study compares a General Regression Neural Network (GRNN) optimized by FOA with Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) models optimized by Gaussian Process (GP). Results show that LSTM and GRU with encoded inputs outperform their non-encoded counterparts. The GRU, combined with an autoencoder and Bayesian-optimized hyper-parameters, achieves the lowest prediction error, demonstrating superior forecasting capability.The dataset, obtained from evaluated feeders in Kapuas District, Central Kalimantan, covers hourly power generation and distribution from October 2017 to September 2018. Data was split into 11 months for training and 1 month for testing, with the training set further divided into 70% training and 30% validation. The best performing model achieved RMSE and MAE values of 27.5824 and 14.9804, respectively. Future research may explore further optimization, feature selection techniques, and extended dataset variations.]]>
Copyrights © 2025
