以輻狀基底類神經網路模擬順風向風力頻譜之方法精進

The Refinement of Alongwind Spectrum Simulation Methods using RBFNN

張心銓(Hsin-Chuan Chang)

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(2019年)

https://doi.org/10.6846/TKU.2019.01031

摘要

近年來因應人口密度的增加，使得高樓大廈不斷興建的現況下，建築物耐風設計成為建設大樓的重要過程受到越來越多人的重視。在建築物的耐風設計上，計算設計風載重所需之風力頻譜一般要經由風洞試驗而取得，其取得過程相當的耗時且費用昂貴。
在智慧型系統發展的現況下，類神經網路模擬之風力頻譜雖然有些微誤差，但節省了大量人力以及實驗時間。起初順風向、橫風向、扭轉向的資料皆有做類神經網路預測，而之後資料不斷更新的同時，前人是以橫風向及扭轉向做研究，目前建築設計風力專家系統中僅有橫風向及扭轉向根據類神經所預測的頻譜進行風力的計算，不包含順風向，但早期的順風向類神經系統資料過於老舊，許多新的實驗資料並未納入網路中，因此本研究除了增添未納入分析的順風向資料外，並改良類神經網路的分析架構，以期達到更好的效果。
在本論文中，利用風洞實驗所求得之風力頻譜，並沿用淡江大學風工程研究中心的相關研究作為基礎，在之前的研究中，應用類神經網路預測風力頻譜之方法有不錯的結果，因此本研究參考其模式，對類神經網路進行深寬比的分組，在三、四、五個深寬比一組的網路中，五個深寬比一組為最佳方法，之後加上網路邊界深寬比當做訓練案例，並從50~500之中心點數間進行RMSE值比較，發現以200中心點數最為合適。同樣以類神經網路所得之的模擬頻譜，在與實驗數據進行誤差比對，並加上驗證回饋之方法的嘗試，以重要頻譜段之誤差值乘上權重，進而得到預期之風力頻譜，篩選過後之網路再經由測試案例的檢驗。
期望未來使用時，能應用於耐風設計之風力計算，甚至更進一步能用於類似頻譜數據的分析。

關鍵字

類神經網路 ； 輻狀基底函數 ； 風力頻譜 ； 風工程

並列摘要

Because of the increasing in population density, high-rise building constructions keep booming in recent years. Wind resistant design of buildings becomes an important issue and is valued more than before. To obtain the required wind spectra from wind tunnel experiments for calculating design wind loads, the process it is very time consuming and expensive
Under the current situation of developing smart systems, neural network simulation of wind spectra has slight errors, but it saved a lot of manpower and experimental time. At first, neural network predictions of alongwind, acrosswind and torsional spectra have all been attempted. The wind tunnel test data was continuously updating and expanding subsequently. The following studies focused on the acrosswind and torsional directions. The current building design wind load expert system only uses ANNs in the acrosswind and torsional directions for wind load calculations, not including alongwind data. Early research results for alongwind are outdated, and many new experimental data is not included. Therefore, this study not only adds alongwind data that is not included in the analysis, but also improves the neural network system to achieve better results.
The aerodynamic database and the relevant researches of the Wind Engineering Research Center at Tamkang University were used as the basis to develop this thesis. The applications of neural networks to predict wind spectrum had good results in previous studies. Therefore, this study referred to their model, ANN architecture and data grouping methods. The five-aspect-ratio method, selected from tryouts of three, four and five aspect ratios a group, was considered the best, and then the cases of the aspect ratios at the network boundary were add to form the training datasets, The RMSE value comparisons between the central points of 50~500 were conducted to found that the 200-center-point was the most suitable. In addition, the predicted spectra obtained by the neural networks were compared with the experimental data, and an attempt to add a verification feedback loop was investigated. The errors of the important spectrum segment were multiplied by a weight to select the most desired network. The network afterwards was tested by test cases.
It is expected that the trained ANNs can be applied to the calculation of design wind loads afterward, and even further the model and programs can be used for the analysis of similar spectrum data.

並列關鍵字

Neural network ； Radial basis function ； Wind spectrum ； Wind engineering

參考文獻

1. 鍾欣潔，2011，「預測高層建築之風力係數與風力頻譜的模式探討」，淡江大學土木工程學系碩士班 論文。

2. Holmes,john D.,2001,“Wind loading of structures”,SPON PRESS.Zhang,Aishe and Zhang,Ling,2004,“RBF neural networks for the prediction of building interference effects”,Computers and Structures 82,pp.2333-2339.

3. 陳信甫，2003，「紊流邊界層對高層建築設計風力的影響」，淡江大學土木工程學系碩士班論文。

4. 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。

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