應用遺傳規劃法進行氣候預報降尺度分析於季節性河川流量預報

Abstract

近年來臺灣受到極端水文事件衝擊，造成農業、工業、民生之嚴重損失，若能及早預測氣候與水文變化，便能及早進行因應之決策。本研究使用ECHAM4.5季節性氣候預報資料進行降尺度。降尺度模式的建立乃利用線性迴歸法與遺傳規劃法，以平均絕對誤差作為選取降尺度方法的依據，經過比較兩種方法後，對於溫度的降尺度採用線性迴歸法較有效率，對於雨量的降尺度則需分月運用遺傳規劃法分別建立各月之數學函式較為理想。氣象合成模式擁有將降尺度後之月平均氣候資料合成為多組相同統計特性日氣象資料的特點。本研究採用GWLF模擬河川流量，驗證結果顯示GWLF可有效的模擬出三峽河之流量，可進一步將氣象合成模式產生的氣象資料代入GWLF對於季節性預報之月流量進行探討，本研究中選取三筆氣候預報資料經過降尺度後得到的流量可以得知，實際流量大致位於模擬預報流量之25%百分位流量與最小值間。對於極端事件的趨勢也有不錯的效果。綜合以上的介紹，本降尺度模式對於流量的預報有相當的預報性。

Extremely hydrologic events create severe loss in agriculture, industry and lives nowadays. The purpose of this thesis is to forecast the climate and prevent the damage. Based on General Circulation Model-ECHAM4.5, this thesis is using linear regression and genetic programming as the downscaling model and mean absolute error to evaluate the downscaling method. By comparing these two methods, we find linear regression is more efficient in temperature and genetic programming is a better method for precipitation. For precipitation, we make twelve different math formulas for each month. Furthermore, a weather generator can be used to produce daily data based on downscaled monthly statistics. The GWLF model is applied to simulate streamflow, and the validation result indicates it can simulate San-Shia River effectively. Daily data producing by weather generator as the input to the GWLF model to provide seasonal streamflows. In conclusion, this thesis is choosing three downscaled climate forecast data to generate streamflow from the GWLF model. The observed streamflow is within the ranges from min to 25% percentile of simulated streamflow. The result of simulating the trend for extremely events is positive. Over all, this downscaling model is effective in predicting streamflow.