過去房屋的設計皆是以地震力為主，但隨著現代都市的高樓林立，高層建築已經漸漸取代了低矮建築，然而在亞熱帶島嶼的台灣，冬天有東北季風，夏天又常有颱風侵襲，由於高層建築相對於低矮建築來說勁度較低，且當建築物愈接近地面因地表粗糙度的關係風力較小，因此愈高的建築物所受到的風力會愈大，所以在設計時就必須考慮到風力對高層建築物的影響。而在風力對高層建築的影響中，建築物相互間的交互作用是一個重要的問題，影響高層建築物間干擾效應的參數很多，如上游流場、建築物間的距離及主建物及干擾建物的尺寸等。 本文主要是利用類神經網路對目前已有的干擾效應資料庫來進行干擾係數的預測，最後再利用所預測的干擾係數來計算出設計風載重。研究使用的類神經網路為輻狀基底函數類神經網路，其主要架構分為輸入層、一層隱藏層及輸出層，在使用遞增式的方法選取出中心點後透過隱藏層中的輻狀基底函數來尋求出網路模式。 由本文的結果得知，在訓練資料的預測方面絕大多數的誤差都低於3%，僅有少數幾點偏高一些，但仍然都在5%以下，在預測資料上，A(α=0.32)、B(α=0.25)、C(α=0.15)三個地況中，除了C 地況在橫風向動態背景方面平均誤差較高外，其餘在靜態及動態背景部份平均誤差皆低於10%，動態共振部份A、B 地況介於10~20%之間，但在最複雜的C 地況則大多介於20~35%之間，因此在共振部份的預測仍需加強。 In large city, tall buildings are usually built in a cluster within the crowded commercial zone, and the interference effects from the adjacent buildings cannot be ignored. Due to the complexity of the interference phenomenon, the building interference phenomenon is too complicated a problem for traditional engineering approaches. A more sensible way to deal with this problem would be using good quality aerodynamic database with IT techniques such as neural network. The aim of this thesis is to construct an aerodynamic database on building interference that can be used together with the wind code or/and single building’s aerodynamic database for wind resistant design of tall buildings. In this study the Artificial Neural Network technique was applied on an existing limited aerodynamic database to predict tall buildings’ interference factor (IF), and use the result to calculate design wind load. The ANN model is radial basis function (RBF) neural network, and the framework includes input layer, a single hidden layer and output layer. By centers and radial basis function in the hidden layer to create a neural network. The results indicate that, in the training phase, the error almost can smaller than 3%, in the predicting phase, except acrosswind background part in terrain C ,the static part and background part in all terrain are less than 10%. The resonant part in terrain A&B are within 10% to 20%. However, in terrain C the error of resonant part can be as large as 20% to 35%. In other words the resonant part need to be further improved.
