使用貝氏統計即時後處理颱風降雨量預測

Abstract

颱風是台灣常見的天然災害，其挾帶極端降雨，常常對台灣造成嚴重的損害。如果沒有適當的預防措施，突如其來的豪大雨很容易導致作物損失、洪水、交通中斷，甚至引發山崩，威脅到人們的生命安全。因此，在氣象學和政策決策中，颱風的定量降雨預報至關重要。 目前的降雨預報大多是基於系集預報系統（ensemble prediction system, EPS）輸出的預報成員去進行預測，然而這種預測系統存在著系統性偏差。在以往的研究中，常會使用系集模式輸出統計（ensemble model output statistics, EMOS）與貝氏模型平均（Bayesian model averaging, BMA）進行統計後處理，這兩種方法都能同時構建校正模型並提供機率預報。但基於颱風資料的特殊性，其提供的訓練期間較短並且含有極端的雨量觀測值，兩種方法都不適合直接應用於颱風期間的降雨預測。 本研究採用中央氣象局以WRF區域模式為基礎的系集預報資料，並針對颱風發生的事件，建構混合的預測模型，結合了系集預報成員與歷史颱風記錄的資訊，透過貝氏統計進行後處理。其中，基於系集預報成員的預測分佈，可以在系集預報產生的同時建構，可解決短期數據的問題，而加入歷史颱風記錄，則能額外提供地區相關的訊息，協助預測強降雨的發生。 本研究在CRPS和可信度圖等氣象評估指標中，皆可看出良好的校正成效。此外，校正模型的計算量不致太過繁重，展現了該模型在未來的颱風事件中能被應用的潛力。

Typhoon is a natural disaster that often brings extreme rainfall and causes severe damage to Taiwan. Without appropriate prevention measures, sudden rainfall can easily cause crop losses, flooding, transportation disruption, and even lead to landslides endangering people's safety. Therefore, a real-time and reliable probability prediction of typhoon precipitation has been essential in meteorology and policy decisions. Current precipitation forecasts are mainly based on the Ensemble Prediction Systems Outputs; however, this prediction system has a systematic bias. In previous studies, two types of post-processing methods, the ensemble model output statistics (EMOS) and Bayesian model averaging (BMA), have been commonly used to construct models and provide probabilistic forecasts at the same time. However, due to the short-termed training period and extreme-value observations, both methods are unsuitable to be directly applied to precipitation prediction during the attack of typhoons. In this research, we construct a mixture model using a fully Bayesian approach that combines ensemble member forecasts with important information from historical records of typhoons. Each member-based component distribution can be trained with the real-time member forecast; it is designed to address the issue of limited short-term data, while the historical data provides additional information for handling heavy rain events. Our model has demonstrated better predictive capabilities across various meteorological evaluation metrics, including Continuous ranked probability score (CRPS) and reliability diagram. Furthermore, the computational load is not heavy, suggesting promising application of this model to future events.