蒸發量在水文循環中扮演著重要的角色,為集水區經營和水資源規劃中相當重要之參數。近年來陸續有學者應用類神經網路推估蒸發量,但皆僅針對單一測站的蒸發量進行推估,若採用建立完成的模式應用於其他測站以推估蒸發量,則會因不同測站的氣候特性,使誤差相對提高。本研究選定台灣地區16個氣象觀測站,蒐集2002到2007年的氣象資料,首先透過K-Means、Fuzzy C-Means及SOM等分群聚類的方法將16個測站依其特性分成四類,分析成果顯示SOM為最適用之模式,其分類結果大致呈現北區、中區、高山區以及南區,其後藉由The Gamma Test搜尋各區之主要影響蒸發量的氣象因子,以做為模式的輸入變數,最後利用自組特徵映射網路結合線性迴歸以建構區域性之蒸發量推估模式,而推估之成果更近一步與Modified Penman以及Penman-Monteith兩傳統經驗式進行比較。結果顯示, SOMN推估之成果優於傳統經驗式,整體而言,16個測站若各自分別架構日蒸發量推估模式,雖可較能獲得較佳的結果,但過程煩瑣複雜,而本研究所建構四個區域性之蒸發量推估模式,不僅大幅降低模式架構之個數量,其分類結果顯示16測站大致呈現出依地域性聚類之分佈;此外,研究成果亦驗證以SOMN架構之區域性蒸發量推估模式確實為一精確且有效之方法。
Evaporation is an important component for watershed management and water resources development and therefore plays a key role in hydrological cycle. In recent years, applications of artificial neural networks on the estimation of evaporation have been proposed. However, previous works merely focused on estimating the evaporation at a specific site. The accuracy may decrease if the constructed model was applied to other sites due to the difference in hydro-geo-meteorology conditions. In this study, daily data are collected from 2002 to 2007 at sixteen meteorological gauges. First of all, these gauges are classified into four clusters according to their similarities by using K-Means, Fuzzy C-Means and SOM. The results indicate that the SOM is more suitable for classification as compared with other methods, and it clustered results cshow a distribution of north region, middle region, mountain region, and south region. Second, the Gamma test is used for finding the meteorological factors that may dominate the evaporation in each cluster. Finally, the selected meteorological factors are separately taken as the inputs of four self-organizing map networks (SOMNs) and the model performance are further compared with those of Modified Penman and Penman-Monteith. The results show that the SOMNs outperform two empirical formulas. Generally speaking, it is time-consuming to build a specific evaporation estimating model for each site in a region even though better performance may be obtained; whereas the four regional SOMN models constructed in this study not only provide a meaningful distribution of each cluster but effectively decrease the number of models. Furthermore, results obtained from this study strongly demonstrate that the regional SOM is an accurate and efficient method for evaporation estimation.