結構物的耐風設計通常需要經由風洞實驗，取得各種風力係數、風力頻譜的實驗數據，其過程相當耗時且費用昂貴。近年，國外風力規範逐步朝資料庫輔助(Database-Assisted)的設計模式發展，使用回歸公式來整理分析實驗數據，常無法得到準確的風力係數，因此，如何更有效的利用風洞實驗氣動力資料庫是一個重要的課題。 在風力係數部分，先前在淡江大學風工程研究中心的相關類神經網路研究中，以倒傳遞(BP)、幅狀基底(RBF)和廣義回歸(GRNN) 類神經網路來進行風力係數的模擬，發現使用「幅狀基底函數類神經網路(RBFNN)」得到最佳之結果。前述之研究範圍為三種地況(A、B、C)下之矩形斷面建物(深寬比0.2~5，高寬比3~7)，風攻角為0度時順風向、橫風向、扭轉向的各種基底風力係數，為進一步驗證預測模式之可靠度，近來重新進行風洞實驗，增加了許多模型(高寬比1~2.5)的量測數據來作為本論文的研究範圍；在風力係數的預測結果發現，順風向迎風面風力係數(Cdw)之預測優於背風面風力係數(Cdl)，在橫風向方面，深寬比大於1時類神經網路之預測可以得到較好預測結果，在扭轉向方面，與橫風向預測結果相同，同樣在B、C地況的預測是較優於A地況。 在風力頻譜部份，先前在風工程研究中心相關研究中，曾應用類神經網路來預測風力頻譜有相當的成效，因此在這次研究上，套用前人之模式方法，在類神經網路之分類，以深寬比兩個為一組，利用類神經網路 (張斐章、張麗秋)一書之隨機選取法，重新撰寫RBFNN類神經網路程式，且運用與前人不同的低頻區內插補點方式，再以新風洞試驗數據訓練新的類神經網路，探討訓練和驗證案例之分配方式，並進一步嘗試不同的資料分類方法，然後微調網路架構，得到更準確的預測結果；橫風向與扭轉向的特定頻率最大誤差比較而言，B、C地況比A地況的結果好，個別RMSE值與頻譜圖，在三種地況下，其高寬比於5.5及6.5時預測結果較差；橫風向最大誤差多半產生於低頻率區間(無因次頻率為0.01)，扭轉向最大誤差多半產生於尖峰值附近(無因次頻率為0.1)。 最後再將預測結果之類神經網路架構，進行風載重之案例分析，並探討其分析結果。 Wind resistant design of buildings often needs to acquire wind coefficents and spectra from wind tunnel tests. Recently, the development of building design wind load standards of other countries has gradually progressed toward database-assisted design methods. Using regression formulas to process and analyze experimental data of wind coefficients usually are not very accurate. Therefore, one of the most important issue is how to use experimental wind load aerodynamic database more effectively. For wind coefficients, Back Propagation (BP), Radial Basis Function (RBF) and Generalized Regression (GR) neural networks have been used previously at the Wind Engineering Research Center of Tamkang University (WERC-TKU) for wind coefficient simulations, and the previous finding was that the RBFNNs yielded the best results. In order to further validate the reliability of the prediction models, measurement data of wind tunnel experiments of models with aspect ratio from 1 to 2.5 were added to the scope of this research. The alongwind results showed that the predictions of windward drag coefficients (Cdw) were superior to the leeward drag coefficients (Cdl). For the acrosswind fluctuating lift coefficients, the predictions were better for buildings with side ratio greater than 1. For the torsional fluctuating moment coefficients, similar to the acrosswind results, the predictions of terrain B and terrain C were better than terrain A. For wind spectra, not only previous WERC-TKU methods were followed but also several improvements were employed. Data was grouped into two adjacent side ratios for classification and random center selection was used for the RBFNN as before. On the other hand, the RBFNN program was recoded and different interpolation schema for adding data points to the low-frequency region was used. Using the new wind tunnel test data to train neural networks, adjusting the training and validation data sets and fine tuning the final model were conducted to produce improved and accurate results. For the acrosswind and torsional wind force spectra, the maximum errors of terrain B and terrain C were smaller than terrain A within the specific frequency. The worse individual RMSEs of the spectra appeared at aspect ratio 5.5 and 6.5 for all three terrains. The maximum errors of acrosswind spectra usually happened in the low frequency range (dimensionless frequency of 0.01) and the maximum errors of torsional spectra mostly happened in the vicinity of the peak value (dimensionless frequency of 0.1). Finally, the neural network architectures and trained networks were applied to perform wind load analysis of several cases, and the case study results were discussed in the thesis.