本研究提出一個結合主成份分析(Linear Principal Components Analysis,PCA)、非線性主成分分析(Non-Linear Principal Components Analysis,NLPCA)及輻狀基底函數類神經網路(Radial Basis Function Neural Network,RBFNN) 之非線性主成分分析統計降尺模式(Nonlinear Principal Component for Genetic Algorithm based- Radial Basis Function Neural Network, NGRB )。首先使用PCA及NLPCA分析氣象站與GCM模式資料,其次透過基因演算法(Genetic Algorithm,GA)優化NGRB模式參數,最後再以4種GCM模式的未來情境資料預測淡水、台中、花蓮及高雄測站的短期(2020/01~2039/12)、中期(2050/01~2069/12)及長期(2080/01~2099/12)之單月總降雨量。 經模擬分析顯示:資料分析結果以NLPCA較PCA能解釋原資料之特徵,並且能有效使減少NGRB模式運算時間, GA世代數為25、族群數為25時NGRB能合理地收斂;模擬4個測站短、中及長期未來單月降雨量,與歷史平均降雨量(1950/01~2000/12)相比,淡水測站短、中及長期平均降雨變化百分率依序為-28.26%、2.02%、-28.14%;台中測站短、中及長期平均降雨變化百分率依序為-36.35%、-23.44%、101.84%;花蓮測站短、中及長期平均降雨變化百分率依序為-1.06%、77.82%、-33.99%;高雄測站短、中及長期平均降雨變化百分率依序為29.24%、39.98%、25.98% 關鍵字: 非線性主成份分析、輻狀基底類神經網路、基因演算法、氣候變遷
The study combines principal component analysis (PCA), nonlinear principal component analysis (NLPCA), and genetic algorithm (GA) based radial basis function neural network (RBFNN) to develop a statistical downscaling model (NGRB). Firstly, the PCA and NLPCA are used to analyze data from meteorological stations and GCM model. GA is then employed to optimize parameters of NGRB model . Finally, four GCM models outputs from A1B scenario are applied to predict short term(2020/01 to 2039/12), medium term (2050/01 to 2069/12) and long term (2080/01 to 2099/12) monthly rainfall of Tamsui, Taichung, Kaohsiung and Hualien stations. The simulated results show that NLPCA can extract features of data better than PCA and can reduce much computation time of NGRB; GA can converge effectively with 25 generations and populations. It reveals that the average percentage of variation of monthly rainfall compared with historical average rainfall in short term, middle term, and long term of Tamsui station are -28.26% ,2.02% , 28.14%, respectively; that ofTaichung station are -36.35% , - 23.44%, 101.84%, respectively;Hualien station are -1.06%, 77.82% - 33.99 %, respectively; that of Kaohsiung stations are 29.24% ,39.98%, 25.98%, respectively. Keywords: nonlinear principal component analysis, radial basis neural networks, genetic algorithms, climate change