Pengembangan Model Prakiraan Curah Hujan per Jam Menggunakan Metode CLGAN

Abstract

Penelitian ini bertujuan mengembangkan model prakiraan curah hujan per jam di Jawa Barat yang memiliki riwayat kejadian cuaca ekstrem tertinggi di Indonesia. Model dikembangkan dengan memanfaatkan metode deep learning, yaitu Convolutional Long Short-Term Memory Generative Adversarial Network (CLGAN) berdasarkan data curah hujan dan unsur cuaca permukaan lainnya dari luaran ERA5-Land yang berhubungan secara fisis. Secara keseluruhan, hasil model ini berada pada level moderat dengan evaluasi korelasi berkisar 0,41-0,43 dan RMSE berkisar 1,46-1,67 mm/jam, serta skor di atas 90% pada evaluasi spasial. Meskipun demikian, keandalan model ini masih terbatas pada prakiraan kejadian hujan ringan dengan skor ROC AUC berkisar 87-94%, POD di atas 83%, dan CSI di atas 67%, sedangkan keandalan model pada prakiraan hujan sedang dan lebat masih perlu ditingkatkan dengan skor ROC AUC berkisar 50-66%, POD maksimal sebesar 37%, serta CSI maksimal sebesar 20%. Apabila ditinjau secara spasial dan menyeluruh terhadap data ERA5-Land maupun observasi, evaluasi keandalan model pada setiap tingkatan intensitas curah hujan bernilai lebih baik pada area tengah Jawa Barat.

This study aims to develop a nowcast model for hourly precipitation on West Java which has a history as the province in Indonesia with the highest frequency of extreme weather events. The model is constructed using advanced deep learning techniques, specifically a Convolutional Long Short-Term Memory Generative Adversarial Network (CLGAN), based on data of ERA5-Land precipitation and other surface variables that are physically related to precipitation. Overall, the model demonstrates moderate performance, with evaluation metrics showing a correlation between 0,41 to 0,43 and a root mean square error (RMSE) ranging approximately from 1.46 to 1.67 mm per hour. Moreover, spatial evaluations exceed 90% in terms of scoring. However, the model's reliability remains limited to nowcast slight precipitation, as indicated by ROC AUC scores ranging from 87 to 94%, probability of detection (POD) exceeding 83%, and critical success index (CSI) exceeding 67%. In contrast, the reliability of the model for nowcasting moderate and heavy precipitation still needs to be improved, with ROC AUC scores ranging from 50 to 66%, maximum POD over 37%, and maximum CSI over 20%. Reliability evaluations for different precipitation intensity levels exhibit better values in the central area of West Java when comprehensively reviewed and analyzed spatially against both ERA5-Land and observation data.