A Hybrid LSTM-Attention Approach to Short-Term Load Forecasting in Island Microgrids

juc_3_20261011162444 20261011162444 JURNAL UJI COBA purwan78@gmail.com Karoteh Utama Publisher JURNAL UJI COBA JUC 1234-273X 09212026 1 1 A Hybrid LSTM-Attention Approach to Short-Term Load Forecasting in Island Microgrids Yohanes Manurung Politeknik Digital Borneo Aisha Rahman Universiti Teknologi Selatan Island microgrids run on diesel and solar, and poor demand forecasts waste fuel. We combine an LSTM encoder with temporal attention and weather covariates to forecast load 24 hours ahead for three microgrids. Mean absolute percentage error drops to 4.8%, against 7.9% for SARIMA and 6.1% for a plain LSTM. Simulated dispatch with the new forecasts reduces diesel consumption by 6.4% per year. 09212026 3042 https://creativecommons.org/licenses/by/4.0/ https://uji.purwantoro.my.id/article/3