應用QPESUMS高解析降雨資料改良洪水預報模式之研究

Flood Forecasting Model Using the High-resolution Rainfall Products from QPESUMS

黃鵬豪(Peng-Hao Huang)

指導教授 ： 許銘熙

國立臺灣大學/生物資源暨農學院/生物環境系統工程學研究所/碩士(2008年)

https://doi.org/10.6342/NTU.2008.03190

摘要

台灣位於亞熱帶地區，氣候深受季風、颱風及洋流之影響，其平均年降雨量可達2510毫米，大約集中於每年5~10月的梅雨及颱風季節，再加上坡陡流急的自然環境因素，使得每逢颱風所挾帶之強勁風勢與豐沛雨量，常引發下游地區嚴重的洪水災害，導致人民生命與社會經濟的損失。因此，有效利用氣象及水文資訊以提升洪水預報之精度，做為減災及應變之用即顯得相當重要。
然而降雨的預報資訊在許多洪水預報模式之研發過程中是很重要的前置工作，其預報精確度關係到整個洪水預報系統之結果。近年來，台灣隨著觀測技術的進步及分析系統的開發，由中央氣象局完成全台都卜勒雷達網之建置，並於2002年起與美國劇烈風暴研究室(National Severe Storm Laboratory；NSSL)合作開發多重觀測工具之定量降水估計與分類技術(Quantitative Precipitation Estimation and Segregation Using Multiple Sensors；QPESUMS)後，顯著提升了對於劇烈或突變之天氣系統監測與預警。故本文首先利用類神經網路建立雨量-水位模式，探討加入未來1小時、1~2小時及1~3小時雨量資訊後之水位差異；此外，由於空間中的雷達網格擁有較高解析度之特性，故本文利用徐昇多邊形法劃分每個雨量站之控制面積，嘗試地將空間中的雷達網格對應至地面的徐昇多邊形網格，間接將雷達雨量銜接至地面雨量站，進而計算控制面積內之雷達平均雨量，由於目前缺乏精準的定量降水預報(Quantitative Precipitation Forecasting；QPF) 資訊，因此本文採用定量降水估計(Quantitative Precipitation Estimation；QPE)進行位移處理，並假設為QPF，以估算雷達平均雨量隨時間之增減率，進而推測地面雨量站之未來雨量資訊，而後進一步探討推測雨量對於雨量-水位模式、及河川洪水演算模式之初始值修正及整合類神經網路模式，於預報未來1~3小時水位之精度改良成效。為驗證模式的效能，本文衡量各水文測站及氣象雷達資料的完整性下，取得近年三場颱洪事件進行模式測試。
由模擬結果顯示，雨量-水位模式加入含有未來雨量資訊為輸入時，確實有助於模式之演算，其中又以未來1小時雨量資訊最為關鍵，故得知應用QPESUMS推測未來雨量資訊之方法是可行的，而後結合河川洪水演算模式於洪峰時刻預報未來1~3小時水位，其河道水位縱剖線更能趨近於觀測值，故本文所建立的模式，無論在雨量-水位模式或河川洪水演算模式，確實可在颱洪期間提供合理及準確的河川洪水資訊。

關鍵字

QPESUMS ； 類神經網路 ； 雨量預測 ； 初始值修正 ； 整合類神經網路水位預報 ； 洪水位預報

並列摘要

Taiwan located in the subtropical area, and the climate often influenced by the monsoon, typhoon and ocean current. The average annual rainfall of 2510 mm approximately concerntrates on the plum rains and typhoon seasons during May to October. In addition, the natural environmental factor of the high terrain and short river course results in severe flood inundation at the downstream area, which causes disastrous losses of life and the economy. This study aims to develop a flood forecasting model by utilizing the meteorological and hydrological information for flood mitigations and emergency resposes.
The prediction of rainfall is quite important for development of flood forecasting model, and its accuracy concerns the consequences of river stage simulation. In recent years, with the progress of observation techniques and analysis system, Taiwan finished the island wide Doppler radar network by Central Weather Bureau since 2002, and cooperated with National Severe Storm Laboratory (NSSL) to develop the Quantitative Precipitation Estimation and Segregation Using Multiple Sensors (QPESUMS) system, which has improved the monitoring and prewarning weather system. The study built a rainfall-stage forecasting model using artificial neural networks (ANN), comparing the accuracy of stage forecasting with 1 hr / 1~2hr and 1~3hr ahead leading rainfall. Moreover, the radar grid in the space has higher resolution. For comparison, this study uses Thiessen Polygons Method to divide several control polygon areas for each surface rainfall gauge station, while attempting to link up the radar precipitation to rainfall station indirectly. Then the precipitation in the control polygon area is calculated by the average from the QPESUMS. Because the lack of precise quantitative precipitation forecasting (QPF) information at present, the lead time shifting from quantitative precipitation estimation (QPE) is taken place to calculate the increasing and decreasing rate of the average rainfall from radar observation. After rainfall forecasting estimation, the effects of stage forecasting with / without rainfall forecasting among rainfall-stage forecasting model, initial stage correction for forecasting model, and integrated flood forecasting model with ANN are investigated in this study. With the consideration of the complete data of each station and radar data, the three recent typhoon events are simulated to verify the efficiency of the forecasting model.
The results reveal that it’s certainly helpful to improve the accuracy of rainfall-stage forecasting model by inputting the future information of rainfall, especially for the one-hr ahead leading rainfall. It’s feasible to estimate the future information of rainfall by using QPESUMS. With the future rainfall, the forecasting stage profile from flood forecasting model is better agreement to the observational value. For a flash flood or storm, this present model can provide the reliable and satisfactory river-stages forecasting information.

並列關鍵字

QPESUMS ； Artificial Neural Network ； rainfall forecasting ； initial stage of forecasting ； integrated flood forecasting model with ANN ； flood forecasting

參考文獻

39. 林國峰、蔡斐毓、吳明璋，2006，類神經網路於颱風降雨與流量預報之研究，2006農業工程研討會論文集，pp.1201-1212。

41. 吳建均，2005，河川洪水位演算模式之研究，國立台灣大學生物環境系統工程學研究所碩士論文。

48. 張智昌，2006，整合氣象雷達與即時降雨資料於颱風降雨推估之研究，碩士論文，國立台灣大學地理資源環境學研究所。

54. 陳昶憲、楊朝仲、王益文，1996，類神經網路於烏溪流域洪流預報之應用，中華水土保持學報，第二十七卷，第四期，pp.267-274。

59. 陳信中，2006，蘭陽溪洪水預報模式之研究，國立台灣大學生物環境系統工程學研究所碩士論文。