A stacked LSTM model for day-ahead solar irradiance forecasting under tropical seasons in Java-Bali

Abstract

Accurate short-term solar irradiance forecasting is essential for the efficient management and planning of power generation, especially for solar energy, which holds a major role in the Indonesian Government’s energy transition policy. A novel day-ahead solar irradiance forecasting is proposed using a stacked long short-term memory (LSTM) model to support the energy planning in the Java-Bali grid. The proposed model utilizes the first historical solar irradiance data of Java-Bali obtained from direct measurement to forecast the next day’s hourly irradiance. The results are compared with the methods of autoregressive integrated moving average (ARIMA) and recurrent neural network (RNN). This study revealed that the proposed model outperforms ARIMA and RNN, and regarded as a highly accurate forecast since root mean square error (RMSE), mean absolute percentage error (MAPE), and R2 are 25.56 W/m2, 7.27%, and 0.99, respectively. The stacked LSTM produces better forecasting in the dry season than in the wet season. The MAPE indicates that the LSTM's lowest accuracy for the dry season was 13.99%, which is categorized as a good forecast. The LSTM’s highest MAPE in the rainy season is 34.04%, which is categorized as a reasonable forecast. The proposed model shows its superiority and capability as a promising approach for short-term solar irradiance forecasting in Java-Bali.