隨著科技的進步，人們可以使用電腦來處理解決許多工程上的問題，達到提高效率、降低成本…等目的。然而還是有許多領域的問題是需要進行實驗才可以得到近似解。風工程就是屬於後者。在風工程領域之中，最為精確的設計風載重還是需要經由風洞試驗中所取得的實驗數據加以計算，然而風洞試驗雖是較為精確，但其過程相當耗時，且費用昂貴。因此如果工程師在初步設計時可以從以往相似的案例中推估出目標建物所需的各項資訊，將能替工程師省下許多的時間。 本研究使用類神經網路模擬風力頻譜，以提供作為初步設計時計算設計風載重之用。使用者可以透過相似建物來預測目標建物的相關資訊，進而完成初步設計。本論文中應用MATLAB撰寫程式，建構輻狀基底函數類神經網路，將A、B、C三種地況下順風向、橫風向及扭轉向矩形斷面模型的深寬比、高寬比、頻率及頻率所對應的頻譜值，納入類神經網路進行訓練、驗證，以建構輻狀基底函數類神經網路(RBFNN)。 在三種地況中，訓練部分特定頻率(頻率0.15至0.4)誤差均在±3.96%以下，動態反應誤差均在±11.62%以下，驗證部份特定頻率誤差均在±4.07%以下，動態反應誤差均在±13.60%以下。以RBFNN使用風洞試驗之氣動力資料庫進行學習並建構網路，在預測各地況各方向之矩形斷面建築物風力頻譜均有不錯的效果。 With the progress of science and technology, people can use computers to deal with and solve a lot of engineering problems for the purposes of raising efficiency, lowering costs, etc. Still there are a lot of problems in many fields that require to carry out experiments and to use approximate solutions. Wind engineering belongs to the latter. Wind tunnel tests usually provide the most reliable design wind loads in current wind engineering practice. Nevertheless, wind tunnel tests are expensive and time consuming. It is very desirable to save resources and to obtain relatively accurate design wind loads at preliminary design stage. It can save a lot of time for engineers. The reported research employed artificial neural networks to predict wind spectra and calculate wind loadings for rectangular cross-section buildings. Similar buildings can be used to anticipate the relevant information of the target building, and proceed to complete the preliminary design. In this thesis, MATLAB was used to implement the Radial Basis Function Neural Networks (RBFNN) constructed. The training and validation of the RBFNN cover alongwind, acrosswind and torsional wind force spectra in all three exposure conditions. For all three kinds of exposures, the spectrum errors under the most frequently used frequencies (0.15 ~ 0.4) are less than 3.96% for the training set and less than 4.07% for the validation set. As for the dynamic responses, the errors are less than 11.62% for the training set and less than 13.60% for the validation set. These results show that using aerodynamic database constructed from wind-tunnel test data to train RBFNN is a prospective approach to predict wind spectra.
