| 摘要: | 近年來由於人口密度不斷的增加，許多地區的大樓如雨後春筍般興建，為因應地狹人稠的人口比例，建築的高度也相對提高，然而隨著樓層越高受到風的影響越是嚴重，因此建築耐風設計即成為在建構高樓時重要的課題。在結構物的耐風設計上，設計風載重所需的風力頻譜通常是需要經由風洞試驗取得，其過程相當耗時且昂貴。
而計算流體力學(CFD)的發展日漸進步，雖然有些微誤差，但相對於風洞實驗所耗費的人力與時間較低，對於未來擴增氣動力資料庫的數據將更為方便。然而若要提升CFD的準確度，網格數量勢必會增加，導致計算時間龐大，因此近期以CFD建置氣動力資料庫的計畫中，期望以最少數量之網格來達成最大模擬效益為目標進行網格繪置，因為網格數相對減少，使得模擬結果相對變差，所以本研究中將CFD數值模擬的基底彎矩頻譜與風洞實驗數值間的比值所得出的修正係數進行類神經網路模擬，期望能利用類神經網路的預測能力與CFD模擬之便利性，來大幅降低未來進行風洞實驗所耗費的時間與人力。
本研究以類神經網路為核心，建構CFD頻譜修正系統，透過資料前處理，分析、整理資料，並進一步將其分類，而類神經網路之建置於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，參考前人之模式方法，建立類神經網路之架構。在類神經網路之分類，以三個風力作用方向為網路分區依據，撰寫輻狀基底函數類神經網路(RBFNN)程式，探討訓練和驗證案例之分配方式與不同輸入項對於類神經模擬結果的影響，並進一步嘗試不同的輸入項與資料分類方法，以便調整網路架構，得到更準確的預測結果。最後再將預測之修正係數譜乘上CFD模擬之頻譜，得出修正後的風力頻譜，並與實驗頻譜進行誤差比對分析。
未來，此模式將應用於缺乏風洞實驗之案例(如削角、或特殊斷面等)，以CFD數值模擬，並經由此模式得到修正後之模擬案例，以利後續耐風設計之計算、應用，以及氣動力資料庫之擴建。
Recently, buildings in many areas have sprung up and the heights of buildings increased as well, due to the continuous growth in population density. As buildings grow taller, wind effects become more severe and wind resistant design of buildings turn into an important subject. Wind resistant building design often needs to acquire wind coefficients and spectrums from wind tunnel tests which is time consuming and expensive.
In recent years, the development of computational fluid dynamics (CFD) has gradually progressed. Although slight errors may exist, CFD consume less time and labor relative to wind tunnel tests. Therefore, CFD is a more convenient method to expand our aerodynamic database in the future. To improve the accuracy of CFD simulations, the number of grids is bound to increase, resulting in lengthy computation time. The goal is to use minimum number of grids to build the CFD simulation model to achieve maximum efficient within our computational capacity. In order to compensate the supplementary errors, the ratios of CFD and experimental base moment spectrums, which is called correction coefficients, are the targets of artificial neural networks (ANNs) simulations. In the future, it is expected to use the predictive capabilities of neural networks and the convenience of CFD simulations to significantly reduce the time and effort required for wind tunnel tests.
Using ANN as core, this research implemented a CFD spectrum correction system. In previous studies, ANNs were used to predict wind spectrums with good results. Therefore, similar radial basis function (RBF) neural network architecture was used in this thesis. Data preprocessing, analysis, normalization and grouping were performed. RBFNN program was developed to train three networks to predict alongwind, acrosswind and torsional correction coefficients. The influence of selection of training and verification cases was explored at the early stage of the research. Further attempts were made to seek different input items and data classification methods, and adjust the network parameters to get more accurate simulation results. At the end, the predicted correction coefficients are multiplied by the spectrums from CFD simulation to get the modified wind spectrums, and then compared with the spectrums from wind tunnel tests to do error analysis.
This model will be applied to cases with insufficient wind tunnel test results (e.g., recessing, corner cut, etc.). Using CFD simulation and ANN modification model to rapidly expand the aerodynamic database and support wind resistant design of buildings. |
