<![CDATA[Hourly rainfall in Banten Province changes rapidly and varies across locations; therefore, its prediction requires models capable of learning both local patterns and temporal dependencies. This study evaluates one-dimensional Convolutional Neural Network (1D-CNN), Long Short-Term Memory (LSTM), and hybrid 1D-CNN-LSTM models for one-hour-ahead rainfall prediction using GSMaP Now data. The dataset covers 1 January-31 December 2024, 155 subdistricts in eight regencies/cities, and 1,361,520 records. Preprocessing included correction of the -99 code, time-based interpolation, Min-Max normalization fitted on the training data, construction of 6, 12, and 24 hour input windows, and chronological data splitting. The experiment produced 72 training runs and used MAE, MSE, and RMSE for evaluation. Based on the configuration with the lowest testing RMSE in each region, 1D-CNN achieved average MAE, MSE, and RMSE values of 0.39 mm/hour, 1.50 (mm/hour)2, and 1.21 mm/hour. The hybrid model achieved 0.38 mm/hour, 1.43 (mm/hour)2, and 1.19 mm/hour, whereas LSTM obtained the lowest values of 0.35 mm/hour, 1.32 (mm/hour)2, and 1.14 mm/hour. LSTM was the best model in all regions, although the optimal input window varied between 6 and 24 hours.]]>
Copyrights © 2026
