結合遺傳演算法與運算樹模式改進雷達推估降雨量：以石門測站颱風事件為例

Abstract

本研究分析中央氣象局五分山都卜勒雷達資料，利用2000－2010年間前十大颱風降雨之雷達回波資料為輸入變數，藉以推估石門雨量站單點降雨量，利用遺傳演算法結合運算樹（Genetic Algorithms of Operation Tree，GAOT）模式優選最佳公式型式，並與氣象站雷達推估降雨經驗公式Z=aR^(b)比較，研究案例分別採用Z=aR^(b)、GAOT五層與六層運算樹結構等三種模式，結果顯示三種模式的推估誤差均方根（Root Mean Squared Error，RMSE）值無論在訓練階段或測試階段，均以六層GAOT（GAOT-6 layers）表現最佳，明顯優於傳統的迴歸分析之Z=aR^(b)，但比五層GAOT（GAOT-5 layers）進步量非常不顯著，故本研究建議較簡潔的GAOT-5 layers模式作為Z-R關係式。This study applied GAOT (Genetic AIgorithm and Operation Tree Model) to improve the rainfall estimation which is the most important forcing for hydrogeomorphic processes and natural hazards. The radar reflectivity from the Wufenshan Doppler radar data and the ground Shih-Men raingauge are the input variable and target, respectively. The 10 most torrential typhoon events between 2000 and 2010 are the input variables to estimate the rainfall of the Shih-Men raingauge station. Two genetic algorithm operation tree (GAOT) models, including five layers and six layers operation tree, are proposed and the estimations were compared with the empirical rainfall estimation formula (Z=aR^(b)). The results showed that the root mean squared errors (RMSEs) of the GAOT with six layers is minimum at both training and testing stages, which is better than those of the Z-R equation by traditional regressive method but similar to those of GAOT with five layers. Therefore, we suggest the simpler Z-R equation obtained by GAOT with five layers.