利用系集卡門濾波器建立具資料同化功能之河川洪水預報模式

Abstract

臺灣地處西太平洋亞熱帶地區，氣候型態深受季風、颱風之影響， 平均年降雨量約為2500毫米，約世界平均值的2.6倍。由於臺灣山區地形之坡度十分陡峻，因此有豐沛的降雨量就會引發下游地區之洪水災害，造成極大的損失。尤其首善之都臺北位於淡水河流域，為臺灣經濟之重鎮。本研究主要在建立一個淡水河流域洪水位預報模式，可迅速準確的提供預報洪水位，以減少颱風洪水所帶來的災害與損失。 本研究以動力波預報初始值修正模式為基礎，並整合類神經網路(Artificial Neural Network, ANN)，並且統計出過去幾場颱風事件對於類神經網路預報未來1至3小時水位之誤差，以卡門濾波修正類神經網路的預報值，而後以系集卡門濾波(Ensemble Kalman Filter, EnKF)作資料同化，結合觀測值更新河川的水位，以更準確的預報值作為模式預報的目標值，建立一套淡水河流域之河川洪水預報模式，已做為發布洪水警報、淹水疏散及防救災應變措施之參考。 以卡門濾波之方法更新類神經網路預報值，再以系集卡門濾波整合變量流模式做為洪水預報。模擬結果顯示，由於卡門濾波法增加了誤差統計之特性，因此可提升洪水預報的準確性，並且有效降低誤差隨著預報時間擴散的程度。故本研究的成果確實可在颱風期間提供更為合理及準確的河川洪水資訊。

Taiwan is located on subtropical area of the west Pacific Ocean. The weather patterns have affected by the monsoon and typhoons. The averaged annual rainfall is about 2500 millimeter which is about 2.6 times of the averaged precipitation over the world. The terrain’s steep slope of mountainous areas and heavy rainfall usually cause flooding disaster to make enormous losses in downstream plain where high-density population located. Taipei city where is situated at the Tanshui river basin is the largest city in Taiwan. A flood forecasting model for Tanshui river has been developed in this study to offer a precise flood stage forecast in advance for flood-damaged mitigation. The flood forecasting model integrated the dynamic routing methods with initial value correction and the artificial neural network(ANN) techniques. The statistical quantities are obtained by the ANN results of predicted water stages with 1-3 hours lead time for several typhoons in the past. The Kalman filter is employed to correct ANN prediction values. The stages predicted with 1-3 hours lead time by Kalman filter are taken as the target values applying in flood forecasting model. Then the ensemble Kalman filter (EnKF) river flood forecasting model is developed to provide accurate and detailed flood information for the Tanshui basin at typhoon period. The flood forecasts can be used for flood alert, evacuation and emergency response. The study uses the Kalman filter to correct ANN prediction value and the ensemble Kalman filter for data assimilation. The simulated results show that the present model can be effectively to improve the accuracy of flood forecasting and reduce the error propagation with the forecasting lead time. The study can provide more accurate and reasonable flood stages during typhoons period.