預測高層建築之風力係數與風力頻譜的模式探討

The Study of Wind Coefficient and Spectrum Prediction for High-Rise Buildings

鍾欣潔(Hsin-Chieh Chung)

指導教授 ： 王人牧

淡江大學/工學院/土木工程學系碩士班/碩士(2010年)

https://doi.org/10.6846/TKU.2010.00492

摘要

在結構物的耐風設計上，計算風載重所需的風力係數與風力頻譜通常是依據風工程規範或風洞試驗而取得，其過程相當耗時且費用昂貴。因此，工程上經常使用迴歸公式來整理分析實驗數據，利用這些公式可以得到其他無進行實驗的數據結果。
風力係數使用MATLAB內建的三種回歸方法，分別為回歸分析、多項式回歸和非線性回歸來進行回歸方法的結果比較。另以MATLAB內建之類神經網路來進行預測，將地況、深寬比和高寬比作為輸入項，納入網路中進行訓練和驗證，建立能預測風力係數的類神經網路。使用的類神經網路分別為倒傳遞、輻狀基底和廣義回歸。
對原先的風力頻譜之幅狀基底類神經網路進行架構的探討，在資料的前處理或網路建立的內部架構(如：中心點選取法和函數選擇)進行比較與分析，以期達到「減少預測所需的網路數目」和「增加準確度」兩大目標。
使用幅狀基底函數類神經網路預測風力係數為最佳方法，整理應用得到三個網路來預測風力係數。順風向風力係數類神經網路以Cd與Cdm的預測結果比Cdd與Cdmd好，且B、C地況的預測優於A地況；橫風向風力係數的預測最大誤差百分比超過10%，但當深寬比大於1時，則誤差在4%以下；扭轉向風力係數受數值小影響而誤差放大，最大誤差百分比超過24%，以A地況好過B、C地況。
風力頻譜部分將地況加入輸入項為最為符合目標的方法，預測所需的網路數由72個降低為24個。將網路預測範圍擴大會出現驗證值誤差較大，尤其以高寬比6最為明顯。整體而言，以扭轉向的預測比順風向與橫風向好，C地況比A、B地況好。
前述的研究結果建立案例式專家系統，在初步設計時可供使用者利用網路連線到系統，從以往相似的案例中查詢相關資訊並推估出目標建物的風力係數和頻譜，以減少計算設計風載重所需的時間，進而完成初步設計。

關鍵字

類神經網路 ； 輻狀基底函數 ； 風工程 ； 風力係數 ； 風力頻譜

並列摘要

In wind-resistant design of structures, the calculation of the required wind loads, coefficients and spectrums are usually based on wind codes and standards or wind tunnel tests. The process is very time-consuming and expensive. Therefore, regression analysis is often used to study experimental data in practice to produce regression formulas. These formulas can then be used to forecast results without performing experiments.
Three MATLAB build-in regression methods, namely regression analysis, polynomial regression and linear regression, were used to examine wind coefficients and to compare the results of the different methods. Also, MATLAB’s neural network functions were used as well to train, simulate and forecast wind coefficients using terrain, side ratio (D/B) and aspect ratio (H/B) as inputs The neural networks used includes BPNN(Back Propagation Neural Network), RBFNN (Radial Basis Function Neural Network) and GRNN(General Regression Neural Networks).
To extend the previous research that uses RBFNNs for wind spectrum simulation, the preprocessing of data and internal structure of the networks were compared and analyzed again in this research. The goals were to reduce the number of the networks and to increase the accuracy of predictions.
According to the results presented in this thesis, RBFNN is the best way to predict wind coefficients. The final application used three networks to predict wind coefficients. For alongwind coefficients, the predictions of Cd and Cdm are better than Cdd and Cdmd, and terrain B and C are better than terrain A. For acrosswind coefficients, the maximal prediction error is more than 10%, but the error is below 4% when aspect ratio is greater than 1. For torsional wind coefficients, the maximal error is over 24% impacted by the small values of the coefficients, and terrain A is better than terrain B and C.
The suggestion of the thesis is to add terrain condition to the inputs of the RBFNN, which reduces the total number of networks needed to forecast wind spectrums from 72 to 24. This only slightly increases the error of some validation cases. The most obvious changes occurred to H/B=6 cases.
A case-based expert system operating on the Internet was built using the above findings to estimate target buildings’ wind coefficients and spectrums for preliminary designs of building wind loads.

並列關鍵字

ANN ； RBFNN ； Wind Coefficient ； Wind Spectrum

參考文獻

14.陳姵君，2007，「利用類神經網路預測建築物在干擾效應下之設計風載重」，淡江大學土木工程學系碩士班論文。

19.許祥榕，2005，「建築物設計風力專家系統之架構探討與發展」，淡江大學土木工程學系碩士班論文。

20.陳冠廷，2008，「類神經網路於高層建築設計風載重案例式專家系統之應用」，淡江大學土木工程學系碩士班論文。

1.Holmes, John D., 2001, “Wind Loading of Structures”, SPON PRESS.

2.Zhou, Y., T. Kijewski and A. Kareem, 2003, “Aerodynamic load on tall buildings: an interactive database”, Journal of Structural Engineering, ASCE, Vol. 120, pp. 394-404.