Week 1–4 Tropical Cyclone Forecasting in the Western North Pacific: Verification of Super Typhoon Ragasa (2025) and Application of a Deep Learning-based Probabilistic Index

Bridging the gap between medium-range weather forecasting and seasonal outlooks, the Central Weather Administration (CWA) has implemented a multi-model ensemble framework to enhance tropical cyclone (TC) monitoring on sub-seasonal timescales (weeks 1–4). This operational platform synthesizes objective TC detection from leading global systems, including the 46-day ECMWF ensemble, the 32-day NCEP ensemble, and the CWA’s Global Ensemble Prediction System (GEPS), etc. We also developed a region-specific Probabilistic Formation Index, which serves as an operational Forecast Confidence Level (FCL) for the TC Threat Potential Forecast product.

The FCL is developed by using a deep learning architecture utilizing a Long Short-Term Memory (LSTM) model. The model is specifically designed to extract signals from key sub-seasonal drivers, such as the Western North Pacific Monsoon Index (WNPMI), sea surface temperature (SST), and intraseasonal oscillations including the Madden-Julian Oscillation (MJO) and Boreal Summer Intraseasonal Oscillation (BSISO). A specialized loss function was implemented during the training phase to address the inherent data imbalance of TC formation events. 

Systematic evaluations across the 1–4 week horizon demonstrate substantial forecast skill, particularly within the first two weeks. Notably, the correlation between dynamical model performance and the AI-derived FCL reveals the latter's efficacy as a proxy for forecast reliability in real-time operations. The practical value of this integrated approach is exemplified by the successful subseasonal prediction of Super Typhoon Ragasa (2025). This case study highlights the system's ability to provide early TC formation signals and reliable track outlooks, offering critical leadtime for disaster risk reduction. Complementing these efforts, probabilistic TC rainfall outlook products specifically designed for S2S timescales have been developed to provide valuable reference for water resources management and disaster mitigation. More details will be presented at the meeting.