Tipologi Curah Hujan Tahunan Berbasis Autoencoder Dan Prediksi Hujan Lebat Harian Menggunakan 1D-CNN di Stasiun Maritim Serang

Abstract

Coastal areas require operational climatological products for infrastructure planning and weather preparedness. This study aims to build an annual rainfall pattern typology and develop a daily heavy rainfall detection signal at the Serang Maritime Meteorological Station by integrating an Autoencoder (AE) and a 1D Convolutional Neural Network (1D-CNN). Daily rainfall data from 1995 to 2024 were aggregated into 36 decadal units per year for the typology, and then derived into accumulation features for daily detection. The Autoencoder (AE) transformed the decadal vectors into two dimensions, which were then clustered using K-Means, resulting in two stable climate patterns. Typology 1 consists of 21 years of rainfall data with an average of 1,502 mm, characterizing a moderate wet monsoon pattern, while Typology 2 covers 9 years with an average of 1,927 mm, reflecting a wetter pattern. Furthermore, the 1D-CNN was trained using a 14-day sequence with heavy rainfall accumulation features (≥50 mm/day), successfully achieving an AUROC of 0.667 on the test data. The CNN's superiority surpassed the conventional threshold baseline accuracy with a CSI of 0.029 compared to 0.021, confirming the added value of temporal pattern extraction. This research produces an efficient and reproducible single-station observation-based climatological framework for local early warning services.