<![CDATA[Accurate forecasting of rainfall intensity is critical for hydrometeorological disaster mitigation in tropical regions like Indonesia. While high-resolution AWS data provides an opportunity to improve forecasting over conventional manual gauges, processing this volatile time-series data requires advanced computational models. This study comparatively evaluates predictive performance of static feed-forward MLP and sequential memory GRU deep learning architecture. Utilizing a three-year dataset (2022–2024) from three stations representing coastal, lowland, and mountainous topographies, the study reconstructed minute-aggregated AWS data using a sliding window algorithm. This successfully validated the digital sensors against manual Hellmann-type rain gauges, achieving a strong correlation (R > 0.80) for hourly accumulations. Both deep learning models were then trained using historical rainfall, temperature differences, and humidity differences. The empirical results demonstrate that the GRU model quantitatively outperforms the MLP, achieving a higher average classification accuracy of 96.49% (compared to 95.49%) and a lower RMSE of 1.51 mm (compared to 1.59 mm). The GRU’s gating mechanism proved significantly more robust in handling sharp data fluctuations across diverse terrains. However, the analysis revealed a shared structural limitation: both architectures severely underestimated extreme peak rainfall amplitudes. This anomaly stems from the inherent sparsity of extreme weather data and the mathematical conservatism induced by MSE loss function. Ultimately, while the GRU is highly recommended as a reliable frontline trigger for early warning systems, estimating absolute extreme rainfall magnitudes necessitates future exploration of non-standard loss functions and spatial data integration.]]>
Copyrights © 2026
