Rainfall Prediction in East Java Using Attention-Based LSTM

Abstract

Rainfall is one of the most influential meteorological parameters affecting agriculture, transportation, infrastructure, public health, and disaster mitigation. Accurate rainfall prediction is highly important, especially in East Java, one of the largest rice-producing regions in Indonesia. This study aims to develop a daily rainfall prediction model using the Attention-Based Long Short-Term Memory (LSTM) method. The dataset used in this study was obtained from the Indonesian Agency for Meteorology, Climatology, and Geophysics (BMKG), covering the period from January 1, 2000, to December 31, 2023, with a total of 8,764 daily records. The variables include average temperature, relative humidity, rainfall, sunshine duration, and average wind speed. The research stages include data collection, preprocessing, missing value handling, outlier replacement using the Z-Score method, correlation-based feature selection, Min-Max normalization, lagged feature construction, data splitting, Attention-Based LSTM model design, model training, model testing, and Streamlit application integration. The best model was obtained using a batch size of 16 and 100 epochs, achieving a validation MSE of 0.007038 and a validation RMSE of 0.083893. On testing data, the model achieved an MSE of 0.008076 and an RMSE of 0.089871. These results indicate that Attention-Based LSTM can model daily rainfall patterns effectively and has potential as a data-driven rainfall prediction support system