以支持向量機結合退火演算法及局部線性內嵌法模式推估降雨事件和河川警戒水位關係–以八掌溪為例

Abstract

洪災一直是台灣一個很重要的課題，台灣島狀偏狹長型，中央高山林立使的諸多河川都因坡陡而流急，平時非雨季時河川流量鮮少，但逢多雨量時節，河川容易暴漲造成洪災，在台灣此種河川稱為荒溪型河川，為了因應河川暴漲所造成的洪災，經濟部水利署將河川水位分級定義為一二三級警戒水位，視情況而做出應變措施。 南部多荒溪型河川，其中包括八掌溪，為此研究之案例地點，在水文領域中，水文預報多為流量預報，但水位資訊實為洪災時刻之重要準則，故此篇研究欲提出水位預報模式，而本研究使用了支持向量機(support vector machine)、模擬退火演算法(simulated annealing)及局部線性內嵌法三種方法，而本篇研究欲用機器學習的方法，直接找出雨量和水位的關係，進而可以直接透過降雨量及歷史水位之料來推估未來河川水位之變化，以用更迅速的行動來應變災害措施，而在模式當中有許多地方需要訂定參數來尋求最佳解，在以往參數的訂定總是透過密碼破解中的窮舉法來搜索最佳參數，而在此皆利用退火演算法做最佳參數的訂定。而本篇研究中特別之處為提出局部線性內嵌法取代支持向量機中核轉換之功用，並有不錯的結果以肯定此方法之改變為可行之趨勢。 南部多荒溪型河川，其中包括八掌溪，在本研究進行之案例分析地區即為八掌溪流域，以小公田、中埔、南靖、大湖、岸內等五雨量站，及八掌溪義竹(後生橋)水位站，以徐昇式法得平均雨量和水位，分析其相關性並建立模型，將來欲利用雨量藉由此模型推估河川之水位，以提升防災應變措施效率。

The issue of the floods is important in Taiwan. It is because the narrow and high topography of the island make lots of rivers steep in Taiwan. The tropical depression likes typhoon always causes rivers to flood. Prediction of river flow and river stage under the depression rainfall circumstances is important for government to announce the warning of flood. Every time typhoon passed through Taiwan, there were always floods along some rivers. The warning is classified to three levels according to the warning water levels in Taiwan. Propose of this study is to predict the level of floods warning from the information of precipitation and recorded river stage. To classify the extent of floods warning by the above-mentioned information and modeling the problems, a machine learning model, Support vector machine (SVM), is used. Simulated annealing (SA) is a probabilistic heuristic algorithm to find out the optimization parameter in SVM model. Another manifold learning algorithm which called locally linear embedding (LLE) is used to replace the function of kernel transform in SVM. The result shows that LLE is work for replacing kernel transform, and it’s a new idea for combine these three methods together. This model can predict the level of flood warning by precipitation and recorded river stage, and it can make government announce the warning of flood in the better timing that can keep the danger of flood from residents along the rivers.