結合歸屬函數之類神經網路颱風波浪推算模式

Abstract

本研究利用類神經網路結合歸屬函數來推算颱風波浪，由交通部運輸研究所港灣技術研究中心提供2000年至2009年之安平港實測波高資料，配合日本國土交通省氣象廳(JMA)的RSMC-Tokyo Center提供之颱風資料來進行類神經模式的建立。比較實測波高評估本模式之適用性。 本模式採用颱風與目標點距離(D)、目標點的方位角(θ1)、颱風侵襲角(θ3)、目標點風速(V)及目標點風向(Vdeg)作為類神經網路之輸入參數，並將θ1、θ3及Vdeg三個角度參數透過高斯歸屬函數轉換，將角度資訊轉為影響度資訊，建立修正類神經颱風波浪推算模式。經由高斯歸屬函數修正後驗證組的模式推算結果約可提升約7%之準確度，若再加入查核組後模式驗證組亦可提升約9%之準確度。在驗證組中選取傳統半經驗波浪推算模式同樣有推算的5場颱風來比較，証實本模式之推算結果較經驗公式推算準確。整體而言，本模式較往昔使用的半經驗模式來的優異，對於颱風波浪推算有一定的準確度，未來可應用於港口上的波高預測，提供海上作業船隻或近岸娛樂活動的民眾參考應用。

The thesis is to develop an Neural Network (ANN) model imposed by membership functions to estimate the typhoon waves. Wave data observed by the Harbor and Marine Technology Center during 2000 to 2009 at Anping harbor and typhoon data collected by JMA RSMC-Tokyo Center were used to train the ANN model. The validity of the proposed ANN model is verified by measured wave heights in the test stage. Five parameters including the distance from typhoon center to the interesting point (D), the azimuth between typhoon center and the interesting point (θ1), position angle in the typhoon (θ3), the wind velocity of the interesting point (V) and its responding wind direction (Vdeg), were selected in the input layer of ANN. Low correlation coefficients between some input parameters and wave heights indicating insignificant weighting to the model doesn’t illustrate basic physical interpretation. Gauss membership functions are used in the paper to transform three angle parameters, that are θ1, θ3 and Vdeg, to remedy the disadvantages of original parameters. The corrected ANN model promotes the capacity of estimating wave heights in the test stage by 7% than the original model. An extra procedure of validation is set in the training stage can increase the model performance by 9% accuracy than the original model. The proposed ANN wave model was examined to have higher accuracy on calculating typhoon waves than traditional empirical formula. Due to good estimation on typhoon waves by the proposed ANN model, the proposed method can be applied to other positions for establishing ANN forecasting wave models to provide wave information for navigation and marine activities.