河川洪水系集預報模式

Abstract

台灣特殊的地理環境使得颱風豪雨等天然災害層出不窮，如何防洪治水向來是棘手的課題，因此準確地洪水預報對於決策單位即時應變至關重要。機率預報應用於河川洪水可進一步提供更多資訊，不只預測最可能發生之水位，更可預測其淹水位可能範圍，提供決策單位面對各種潛在狀況之參考。 本文以動力波模式(許銘熙等人，2000)為河川洪水預報之基礎，加入初始條件、邊界水位與曼寧糙度係數等不確定性進行系集預報，將原本之定率預報模式擴增為機率預報模式，再結合系集卡門濾波進行資料同化，以倒傳遞類神經網路模式於水位站之預報水位回饋修正，提升模式預報精度。 將模式應用於淡水河流域，經韋帕颱風(2007)與辛樂克颱風(2008)測試驗證後，結果顯示模式之回饋演算功能顯著提升定率預報精準度。機率預報提供95%信賴區間預報水位範圍，能有效預測洪水位之可能性。兩場颱風實際命中率分別為89.5%、78.8%，顯示尚有其他不確定性之因子影響預報水位範圍，此現象尤其在河系中游區域更為明顯，可見尚需解決低估河系中游區域不確定性的問題，並進一步考慮更完整之不確定性來源。

The special geographical and meteorological environment induced lots of natural disasters such as typhoon and flood in Taiwan. Emergency response and flood evacuation are the major non-structural measures for flood mitigation. Therefore, an accurate flood forecasting model is an indispensable tool for the decision of disaster management agencies. Probabilistic forecasting of flood stage can provide not only the most likely water level, but also the possible range, which offer the reference of a variety of potential situations for decision-makers. Based on one-dimensional dynamic wave theorem, an ensemble forecast technique has been developed in this study by considering uncertainties factors including initial condition, boundary condition, and Manning’s coefficient. The original of dynamic model is a deterministic model which converts to probabilistic forecasting model with the ensemble forecasting. The join data assimilation using the ensemble Kalman filter and back-propagation neural network are employed on gage stations which can offer better feedback estimate and model accuracy. The model is applied to the Tamsui River basin. Two typhoon events of Weipa(2007) and Sinlaku (2008) are used as model validation. The simulated results show that flood stage of the probabilistic forecasting is better accuracy than that of the deterministic forecasting. Based on the probability forecast of 95% confidence interval, the most of the observed level were located in the predicted range. From the comparison of the actual hit ratio of the two typhoon events, it can be found that the 89.5% and 78.8% of observed level fell at prediction range of confidence interval, which shown that forecast range is not enough and underestimate of the uncertainty. This phenomenon is obvious especially in the river midstream. It can be seen that the more factors of uncertainty is needed for further study.