<![CDATA[Accurate rainfall prediction is crucial for mitigating hydrometeorological disasters in Indonesia, where weather patterns are highly volatile and heavily influenced by topography. This study evaluates the performance of MLP algorithm in predicting rainfall intensity using high-resolution AWS data. The evaluation compares model accuracy across three contrasting topographical characteristics: coastal area (Serang Maritime Station), open lowland area (Cengkareng Meteorological Station), and mountainous area (Citeko Meteorological Station). The MLP model utilized historical rainfall, temperature differences, and humidity differences as input variables, trained over a three-year dataset (2022-2024). The results indicate that the MLP model possesses high nowcasting capabilities with an average MAE of 0.29 mm. The coastal area yielded the best prediction performance with RMSE 1.10 mm and MAE 0.22 mm, due to stable wind circulation patterns. Conversely, the open lowland area showed the highest RMSE of 2.14 mm due to sudden extreme rainfall spikes, while the mountainous region proved the most challenging to model with MAE 0.34 mm, due to highly dynamic orographic effects. Furthermore, the study identifies a structural limitation where the MLP underestimates extreme peak amplitudes, caused by data sparsity and the conservative nature of the MSE loss function. Ultimately, while the MLP is highly reliable as an early warning system for light to moderate rainfall, it requires further architectural modifications to accurately estimate extreme heavy rainfall.]]>
