You are on page 1of 8

Journal of Applied Science and Engineering, Vol. 20, No. 1, pp. 55-62 (2017) DOI: 10.6180/jase.2017.20.1.

07

Formulation of Estimation Models for Wind Force


Coefficients of Rectangular Shaped Buildings
Jenmu Wang* and Chii-Ming Cheng
Department of Civil Engineering, Tamkang University,
Tamsui, Taiwan 251, R.O.C.

Abstract
In wind-resistant design of structures, the calculation of wind coefficients is usually based on
data from wind tunnel tests. The process is very time-consuming and expensive. In order to formulate
a model to estimate wind force coefficients of rectangular buildings, various methods including
regression analysis and artificial neural networks (ANNs) were investigated. This paper focuses on
the presentation of the various approaches with emphasis on the detailed result comparisons and
discussions of models developed for alongwind, acrosswind and tortional wind coefficient
predictions.

Key Words: Wind Force Coefficients, Regression, Artificial Neural Networks, Aerodynamic
Database

1. Introduction (Cd) was used as the target of wind coefficient predic-


tion. According to our findings, RBFNN is the most ef-
Artificial neural network (ANN) is an approach to fective method among the five candidates, namely poly-
simulate or predict the results of complex domain by nomial regression, nonlinear regression, BPNN (back
using similar (but highly simplified) models of the bio- propagation neural network), RBFNN (radial basis func-
logical structures found in human brain. Training ANNs tion neural network) and GRNN (general regression
with existing cases with reasonable answers can deduct neural network). The final goal is to trained RBF neural
multi-variable nonlinear models to predict the results of networks to predict alongwind, acrosswind and torsional
similar problems. ANNs are good candidates to be in- wind coefficients all together. Consequently, the trained
corporated in our aero-data based wind resistant design networks can be embedded in a building design wind load
approach [1,2] to perform data mining and knowledge expert system as presented in [2,3] to estimate wind force
discovery from our aerodynamic database. coefficients and calculate wind loads. This paper focuses
It is noted that the purpose of the work reported here on the presentation of the various approaches with em-
is not to replace wind tunnel experiments. The objective phasis on the detailed result comparisons and discussions
is to utilize wind tunnel experiment aerodynamic data- of the final formulation developed for alongwind, acros-
bases to develop an expert system for generating design swind and tortional wind coefficient predictions.
wind loads of tall buildings. It can be used at the pre-
liminary design stage to get reasonably accurate design 2. Scope and Data Processing
wind loads without performing costly and time-consum-
ing wind tunnel tests. The design wind load aerodynamic database de-
At first, alongwind mean coefficient of base shear veloped at the Wind Engineering Research Center of
Tamkang University in Taiwan (WERC-TKU) [4,5] con-
*Corresponding author. E-mail: wang@mail.tku.edu.tw sists of four parts: (i) isolated tall buildings, (ii) interfer-
56 Jenmu Wang and Chii-Ming Cheng

ence effects of two adjacent tall buildings, (iii) other and aspect ratio (H/B) as inputs. The neural networks
buildings and structures and (iv) CFD based. The data used include BP (back propagation), RBF (radial basis
used for this research is from the aerodynamic database function) and GR (general regression) neural networks.
for isolated tall buildings. According to the results of the investigation presented
The wind tunnel measurements and their analysis of in this paper, RBF neural network is the most effective
various generic building shapes have been performed mean to predict wind coefficients. The final formulation
at Wind Engineering Research Center of Tamkang Uni- trained three RBF neural networks to estimate along-
versity (WERC-TKU) to construct an aerodynamic da- wind, acrosswind and tortional wind coefficients respec-
tabase [5,6]. Total of 150-plus building shapes were tively.
studied. The wind force coefficients and reduced force
spectra in the alongwind, acrosswind and torsional di- 3. Initial Investigation
rections of earlier models were measured using HFFB,
whereas later were measured with multi-channel elec- At the preliminary stage of this research, alongwind
tronic pressure scanning system. mean coefficient of base shear Cd was used as an indi-
Sharing the same goal of similar researches that pre- cator for selection of the final estimation method. Two
dict wind coefficients and spectra for buildings such as regression methods and three neural network methods
[7-9] etc., this research selected pressure measurements were used for the forecast of Cd. In order to yield better
of several models in the WERC aerodynamic database results, data grouping strategies, as described in Table 3,
to investigate the prediction of wind force coefficients. were studied as well.
A total of 135 wind tunnel experiment data sets as de-
scribed in Table 1 were used. The coefficients investi- 3.1 Regression Analysis
gated and their abbreviations are listed in Table 2. Their The fitting results, root mean square errors (RMSE)
definitions and setting in wind load calculations can be and the maximum errors of the applications in Table 3
found in [10]. are summarized in Tables 4 to 6. All the analyses were
To formulate a model to estimate wind force coeffi- performed using MATLABs build-in regression analysis
cients of rectangular buildings, two regression analysis
methods, namely polynomial regression and nonlinear Table 2. Nomenclature of wind force coefficients
regression, were used to compare the results at first. In Wind coefficient description Abbreviation
addition, ANNs were used as well to train, simulate and Alongwind mean coefficient of base shear Cd
forecast wind coefficients using terrain, side ratio (D/B) Alongwind RMS coefficient of base shear Cdd
Acrosswind RMS coefficient of base shear Cld
Table 1. Wind tunnel test data selected Alongwind mean coefficient of base Cdm
Model cross-section Square and rectangular moment
Alongwind RMS coefficient of base Cdmd
Terrain exposure A, B, C (a = 0.32, 0.25, 0.15) moment
Side ratio (D/B) 0.2, 0.25, 0.33, 0.5, 1.0, 2.0, Acrosswind RMS coefficient of base Clmd
3.0, 4.0, 5.0 moment
Aspect ratio (H/ BD ) 3, 4, 5, 6, 7 RMS coefficient of base torsion Ctd

Table 3. Data grouping methods for Cd


Name Data grouping Application
Aspect ratio series 3 terrains and 9 side ratios to form 27 sets Polynomial regression
Side ratio series 3 terrains and 5 aspect ratios to form 15 sets Polynomial regression
Terrain series 3 terrains to form 3 sets Nonlinear regression & neural networks
No grouping All data in a set Nonlinear regression & neural networks
Formulation of Estimation Models for Wind Force Coefficients of Rectangular Shaped Buildings 57

functions. Note that the aspect ratio, symboled as H/B, 15 equations with different p1 and p2 parameters were
in the equations is actually H / BD . fitted. The maximum errors exceed 100% in some cases,
For the side ratio series (the 2nd equation in Table 4), which are not usable. The 1st equation in Table 4, which
Table 4. Polynomial regression (needs 27 equations for the H/B series and 15 equations for the D/B series)
Polynomial regression H/B series RMSE Max. error (%)
Cd = p1 (H/B) + p2 0.0209 3.79
Terrain Side ratio p1 p2 Terrain Side ratio p1 p2
0.20 -0.0334 2.5133 1.00 0.0419 0.9875
0.25 -0.0475 2.7995 2.00 0.0173 0.6457
0.33 -0.0215 2.1985 B 3.00 0.0029 0.5221
0.50 -0.0118 1.9867 4.00 0.0045 0.4195
A 1.00 -0.0071 0.9775 5.00 0.0003 0.3916
2.00 -0.0151 0.7719 0.20 0.0176 2.6197
3.00 -0.0099 0.5389 0.25 0.0597 1.9185
4.00 -0.0061 0.4857 0.33 0.0458 1.8350
5.00 -0.0095 0.4170 0.50 0.0380 1.5504
0.20 -0.0216 2.4320 C 1.00 0.0755 0.7937
0.25 -0.0234 2.1890 2.00 0.0424 0.4910
B
0.33 -0.0058 2.2590 3.00 0.0191 0.4236
0.50 -0.0186 1.7635 4.00 0.0109 0.3695
5.00 -0.0003- 0.4180
Polynomial regression D/B series RMSE Max. error (%)
Cd = p1 (D/B) + p2 0.3899 118.7
Terrain Aspect ratio p1 p2 Terrain Aspect ratio p1 p2
3 -0.4401 2.1504 6 -0.4467 2.1939
B
4 -0.4448 2.1586 7 -0.4462 2.1892
A 5 -0.4336 2.1123 3 -0.3987 1.9610
6 -0.4290 2.0824 4 -0.4218 2.0669
7 -0.4267 2.0678 C 5 -0.4330 2.1370
3 -0.4252 2.0875 6 -0.4384 2.1639
B 4 -0.4424 2.1622 7 -0.4377 2.1697
5 -0.4424 2.1807

Table 5. Nonlinear regression (terrain series, one equation for each terrain)
Max error
Terrain Nonlinear regression RMSE
(%)
A Cd = 0.9469 (D/B)-0.9368 - 0.0309 (D/B)-2.541 (H/B)0.0799 + 0.1879 0.091 12.92
B Cd = -62.6962 (D/B)0.0195 - 1.6082 (H/B)-1.4031 + (D/B)0.5744 (H/B)-0.1149 + 63.204 0.094 12.46
C Cd = -0.0527 (D/B)-2.1419 + (D/B)-0.8133 (H/B)0.0508 + 0.1238 0.131 16.54

Table 6. Nonlinear regression (no grouping, one equation for different terrains, aspect and side ratios)
Nonlinear regression RMSE Max. error (%)
Cd = -15.9538 (a) + 57.3295 (D/B)
-0.0339
- 22.7639 (H/B)
-0.0163
+ 0.17 (a)
-0.0186 -0.0834
0.098 19.65
Cd (D/B) - 0.0039 (D/B) (H/B) - 0.9652 (a) (H/B)0.5736 - 16.8512
58 Jenmu Wang and Chii-Ming Cheng

is the aspect ratio series, is very accurate. However, it Table 7. Errors of neural network estimations (terrain
needs 27 equations with different p1 and p2 to forecast series, one network for each terrain)
Cd with different aspect ratios, and the series of equa- Back propagation neural networks (BPNN)
tions cannot be used for cases that are not covered by the Neuron center (NC) = 2
original 9 side ratios. Max. error (%)
The three nonlinear regression equations in Table 5 Terrain RMSE
Training Validation
use aspect and side ratio as independent variables to
A 0.096 14.535 13.348
estimate Cd, which solve the problem of polynomial re-
B 0.063 12.382 12.660
gression. As shown in Table 6, adding terrain exposure
C 0.097 11.519 10.836
(a) to the independent variables further reduces the equ-
Radial basis function neural networks (RBFNN)
ation to one and retains about the same level of preci-
GOAL = 0.1; SPREAD = 1.105~1.2
sion. The RMSE is about the same and the maximum er-
ror, which occurred at terrain A this time, rises some- Max. error (%)
Terrain RMSE
what. Training Validation
A 0.038 6.120 5.771
3.2 ANN Wind Coefficient Predictions B 0.042 4.878 4.727
Three different neural network architectures were C 0.058 6.484 5.476
used, namely BP (back propagation), RBF (radial basis General regression neural networks (GRNN)
function) and GR (general regression) to forecast Cd. SPREAD = 0.1
All the training and testing were performed using Max. error (%)
MATLABs neural network toolbox, and the results are Terrain RMSE
Training Validation
shown in Tables 7 and 8. The inputs of the networks in
A 0.089 8.760 9.625
Table 7 are aspect and side ratio and an additional input,
B 0.053 5.556 6.121
terrain, is added for the networks in Table 8. All the ANN C 0.053 6.445 12.399
output is Cd.
Cases of aspect ratio 3, 5 and 7 were used as training
Table 8. Errors of neural network estimations (no
cases and cases of aspect ratio 4 and 6 were used to vali-
grouping, one network for all terrain)
date the trained ANNs. That is, a total of 81 wind tunnel
experimental cases were used for training, which is 3/5 Back propagation neural networks (BPNN)
Neuron center (NC) = 3
of the total 135 cases collected.
Various parameters were adjusted when training the Max. error (%)
RMSE
networks in order to find the best results. Different num- Training Validation
bers of neuron centers (NC) for BPNN, mean squared 0.0974 14.0648 12.2302
error goal (GOAL) for RBFNN, spread of radial basis
Radial basis function neural networks (RBFNN)
functions (SPREAD) for both RBFNN and GRNN were GOAL = 0.2; SPREAD = 1.2
investigated. For the BPNNs, the number of center points
Max. error (%)
was increased gradually. The best results usually oc- RMSE
curred when the neuron center was set to two or three. Training Validation
When the center points were over five the validation er- 0.0464 6.9680 8.3138
rors tended to increase. For the RBFNNs, the GOAL was General regression neural networks (GRNN)
set first, followed by adjusting the SPREAD to achieve SPREAD = 0.1
better accuracy. It sometimes was necessary to change Max. error (%)
the GOAL to reach a better estimation. The final settings RMSE
Training Validation
are listed in Tables 7 and 8.
0.1009 12.5415 12.7313
The errors of the three different ANNs are all ac-
Formulation of Estimation Models for Wind Force Coefficients of Rectangular Shaped Buildings 59

ceptable as indicated in Tables 7 and 8. RBFNN has the the RMS coefficient, Cdd and Cdmd, a large improve-
best results in both tables. Referring to Tables 7 and 5, ment for Cdd. However, the results of the validation
the results of BPNN are comparable to nonlinear regres- cases are extremely poor, especially for Cdmd.
sion for the terrain series. However, RBFNN and GRNN After extensive experiments of different combina-
are clearly better. tions, finally, we trained one RBF neural network to pre-
Comparing the results in Tables 6 with 8, it is con- dict all four alongwind coefficients. The root mean square
clusive that the ANN models have clear advantage over errors and the absolute maximum errors of the RBFNN
nonlinear regression when only one equation or one net- are summarized in Table 12. This reduced the number of
work is used. Especially, the RBFNN model in Table 8 network to one, which is good for practical application.
works extremely well. It is better than the other two Usually, the extension of the prediction scope tends to
ANN methods in Table 8. Also, the errors of the one-net- drop accuracy, but comparing Table 9 with 12, the
work RBFNN are just slightly higher compared with the RMSEs for Cd and Cdm are not that different. On the
RBFNN three-network terrain series. other hand, the maximum errors for Cdd and Cdmd im-
prove dramatically. The final result also proved that the
4. Predictions for Multiple Alongwind original assumption, which is that the variation trends of
Coefficients all 4 alongwind coefficients are similar therefore RBFNN
can be applied similarly, is sound.
Based on the investigation of prediction of Cd in the Analyzing the RBFNN simulation results, the fol-
previous sections, RBFNN was selected as the simula- lowing remarks can be made: (1) The RMSEs and maxi-
tion model for all four alongwind force coefficients in mum errors decrease when using one network to predict
Table 2. The assumption here is that all the alongwind four alongwind coefficients compared with using four
coefficients exhibit similar changing trend with respect
to side ratio, aspect ratio and terrain. Using four RBFNNs Table 10. Errors of single RBFNN for alongwind mean
coefficients
for the four coefficients, the training and validation re-
sults are summarized in Table 9. The mean coefficients Wind Max. error (%)
GOAL SPREAD RMSE
Cd and Cdm had the best results, which the errors are coefficient Training Validation
all below 9%. However, the maximum error of Cdd is Cd 5.043 7.467
0.1 1.12 0.0218
over 55% for both training and validation cases. Instead Cdm 4.941 4.351
of using four single-output RBFNNs, neural network
architectures with multiple outputs were studied. Table 11. Errors of single RBFNN for alongwind RMS
At first, the 4 wind coefficients were divided into 2 coefficients
categories, mean and RMS. The inputs are still terrain, Wind Max. error (%)
aspect ratio and side ratio. The results are listed in Tables GOAL SPREAD RMSE
coef Training Validation
10 and 11. For the mean coefficient Cd and Cdm, the
Cdd 9.563 31.83
errors are even less than the dedicated networks in Table 0.001 1 0.0067
Cdmd 9.550 275.6
9. The errors of the training cases are all below 10% for

Table 9. Alongwind coefficient error summary using 4 Table 12. Errors of RBFNN for alongwind coefficients
single-output RBFNNs using a single network
Wind Max. error (%) Wind GOAL/ Max. error (%)
GOAL SPREAD RMSE RMSE
coefficient Training Validation coefficient SPREAD Training Validation
Cd 0.200 1.20 0.046 6.968 8.314 Cd 6.583 5.670
Cdd 0.100 8.84 0.031 59.76 55.65 Cdd 7.908 8.525
0.1/1.136 0.0174
Cdm 0.100 1.13 0.031 7.350 6.726 Cdm 6.329 4.907
Cdmd 0.001 1.14 0.002 7.415 10.00 Cdmd 7.360 8.178
60 Jenmu Wang and Chii-Ming Cheng

single purpose RBFNNs. (2) Among the alongwind co- Table 15. Errors of RBFNN prediction for Cd (Terrain A)
efficients, Cd and Cdm can be better estimated than Cdd Side Aspect
and Cdmd. (3) The worst case scenario happens at dif- Experiment Prediction Error (%)
ratio ratio
ferent terrain for different wind force coefficients. (4) 3 2.4146 2.4154 0.034
The worse errors tend to happen with side ratio 0.2, 0.25 4 2.4055 2.3367 2.860
and 0.333 building shapes. 0.2 5 2.3054 2.2970 0.364
6 2.3094 2.3065 0.123
7 2.2955 2.3626 2.923
5. The Final Formulation and Verification
3 2.6531 2.6514 0.066
Based on the investigation of prediction of Cd and 4 4 2.6253 2.6070 0.699
0.25 5 2.5488 2.5679 0.749
alongwind coefficients in the previous section, RBFNN
6 2.5088 2.4106 3.912
was selected again because of its previous successes as 7 2.4739 2.3111 6.583
the simulation model for the acrosswind and torsional
3 2.1326 2.1343 0.078
wind force coefficients in Table 2. 4 2.0919 2.0972 0.257
Using 3 individual neural networks for the 3 across- 0.333 5 2.1167 2.1014 0.722
wind and torsional wind force coefficients, the maxi- 6 2.0842 2.1103 1.251
mum errors of Cld and Clmd were over 15% for either 7 2.0288 2.1567 6.308
training or validation cases. Therefore, the two acrosswind 3 1.8980 1.8974 0.033
coefficients were grouped together using one RBFNN. 4 2.0050 1.9129 4.594
This improved accuracy, which may be caused by the in- 0.5 5 1.9524 1.9564 0.202
crease of training cases. 6 1.8814 1.8946 0.703
7 1.9006 1.8660 1.820
To summarize the results, our final formulation
trained three independent RBF neural networks to predict 3 0.9929 0.9926 0.029
4 1.0221 0.9836 3.764
alongwind, acrosswind and torsional wind coefficients
1 5 1.0125 1.0120 0.051
respectively. The root mean square errors and the abso- 6 0.9974 1.0133 1.592
lute maximum errors of the three RBFNNs are summa- 7 1.0408 1.0438 0.294
rized in Tables 12, 13 and 14.
3 0.7327 0.7323 0.050
For alongwind coefficients, the predictions of the 4 0.6971 0.7022 0.744
mean coefficients, Cd and Cdm, are better than the RMS 2 5 0.6975 0.6974 0.010
coefficients, Cdd and Cdmd. On the whole, the estima- 6 0.6972 0.6575 5.670
tions at terrain B and C are better than terrain A. Using 7 0.6570 0.6548 0.347
Cd and Cdm (as shown in Tables 15 and 16) as examples, 3 0.5064 0.5061 0.055
the average prediction errors are around 1.5%. The error 4 0.5041 0.4835 4.086
3 5 0.4912 0.4901 0.234
Table 13. Errors of RBFNN for acrosswind coefficients 6 0.4730 0.4680 1.052
7 0.4725 0.4792 1.432
Wind GOAL/ Max. error (%)
RMSE 3 0.4670 0.4666 0.089
coefficient SPREAD Training Validation
4 0.4633 0.4443 4.116
Cld 13.870 18.706 4 5 0.4513 0.4511 0.041
0.001/1.14 0.0022
Clmd 14.803 19.091 6 0.4507 0.4293 4.755
7 0.4425 0.4406 0.432
Table 14. Errors of RBFNN for torsional coefficient 3 0.3907 0.3903 0.097
Wind GOAL/ Max. error (%) 4 0.3727 0.3700 0.718
RMSE 5 5 0.3765 0.3766 0.034
coefficient SPREAD Training Validation 6 0.3558 0.3451 3.007
Ctd 0.001/1.125 0.0035 15.787 24.868 7 0.3515 0.3471 1.259
Formulation of Estimation Models for Wind Force Coefficients of Rectangular Shaped Buildings 61

Table 16. Errors of RBFNN prediction for Cdm (Terrain A) distribution is similar for Cd and Cdm. The worst errors
Side Aspect all occur at cases with aspect ratio 7, and side ratio 0.25
Experiment Prediction Error (%) and 0.333.
ratio ratio
3 1.2917 1.2923 0.047 For acrosswind coefficients, terrain B has the worst
4 1.2822 1.2645 1.377 results with average error of 3.73%, and the overall max-
0.2 5 1.2594 1.2554 0.320 imal prediction error is over 13%. However, the error is
6 1.2363 1.2546 1.482 below 4% when side ratio is greater than 1, which holds
7 1.2399 1.2749 2.822
for all three terrains. The worst scenario for Cld, side
3 1.4254 1.4247 0.049 ratio 0.25 and terrain B, is presented in Table 17.
4 1.4068 1.4136 0.480 For torsional wind coefficients, the maximal error is
0.25 5 1.3932 1.4034 0.734
over 24% impacted by the small values of the coeffi-
6 1.3493 1.3156 2.496
7 1.3397 1.2549 6.329 cients, and terrain A is better than terrain B and C. In ge-
neral, the errors are small when the side ratio is greater
3 1.1542 1.1554 0.101
4 1.1263 1.1446 1.630 than 2. Table 18 shows the prediction results of Ctd for
0.333 5 1.1610 1.1532 0.673 side ratio 0.25 at Terrain B.
6 1.1270 1.1523 2.248
7 1.1007 1.1690 6.203 6. Conclusions
3 1.0344 1.0342 0.015
4 1.0878 1.0501 3.469 Several methods, including polynomial regression,
0.5 5 1.0761 1.0787 0.242 nonlinear regression and ANN, have been carefully
6 1.0266 1.0418 1.474
studied for the prediction of wind force coefficients.
7 1.0397 1.0200 1.899
Our investigation showed that using RBFNN yielded
3 0.5511 0.5513 0.032 the best results in terms of accuracy and usefulness. Based
4 0.5698 0.5518 3.158
on this finding, the same RBFNN architecture was applied
1 5 0.5720 0.5719 0.021
6 0.5668 0.5780 1.971 for the estimation of all the seven wind coefficients un-
7 0.5940 0.5967 0.446 der study.
3 0.4191 0.4190 0.023
Further study demonstrated that instead of using
4 0.3998 0.4059 1.529
Table 17. Worst case scenario of Cld RBFNN predictions
2 5 0.4049 0.4050 0.037
6 0.4055 0.3856 4.907 Cld for terrain B and side ratio 0.25
7 0.3857 0.3840 0.449 Aspect ratio Experiment Prediction Error (%)
3 0.2920 0.2920 0.015 3 0.0957 0.1090 13.870
4 0.2927 0.2815 3.827 4 0.0813 0.0965 18.706
3 5 0.2861 0.2855 0.213 5 0.0681 0.0754 10.729
6 0.2777 0.2763 0.523 6 0.0579 0.0650 12.269
7 0.2784 0.2829 1.599 7 0.0490 0.0546 11.428
3 0.2727 0.2726 0.048
4 0.2721 0.2625 3.527 Table 18. Ctd RBFNN predictions
4 5 0.2665 0.2666 0.012
6 0.2677 0.2566 4.153 Ctd for terrain B and side ratio 0.25
7 0.2636 0.2623 0.468 Aspect ratio Experiment Prediction Error (%)
3 0.2281 0.2280 0.046 3 0.0912 0.1048 14.884
4 0.2198 0.2189 0.369 4 0.0725 0.0877 20.965
5 5 0.2228 0.2230 0.078 5 0.0594 0.0670 12.785
6 0.2128 0.2079 2.293 6 0.0494 0.0545 10.374
7 0.2112 0.2088 1.118 7 0.0408 0.0467 14.341
62 Jenmu Wang and Chii-Ming Cheng

seven RBFNNs for the seven wind coefficients, train- pp. 329-339 (2005).
ing three RBF neural networks, one for the alongwind [4] Cheng, C. M., Lin, Y. Y., Wang, J., Wu, J. C. and
coefficients (Cd, Cdd, Cdm and Cdmd), another for the Chang, C. H., The Aerodynamic Database for the In-
acrosswind coefficients (Cld and Clmd) and the other for terference Effects of Adjacent Tall Buildings, Con-
the tortional coefficient Ctd is adequate. ference Preprints, 12th International Conference on
Analyzing the RBFNN simulation results, the fol- Wind Engineering, Cairns, Australia, Jul. 1-6, Vol. 1,
lowing remarks can be made: (1) Alongwind prediction pp. 359-366 (2007).
is better than acrosswind, and acrosswind is better than [5] Cheng, C. M., Lin, Y. Y., Wang, J., Wu, J. C. and
torsional. (2) Among the alongwind coefficients, Cd and Chang, C. H., The Aerodynamic Database for the
Cdm can be better estimated than Cdd and Cdmd. (3) Interference Effects of Adjacent Tall Buildings, Con-
Acrosswind predictions are better when the side ratios ference Preprints, 12th International Conference on
are greater than 1. (4) Torsional predictions are better Wind Engineering, July 1-6, Cairns, Australia, Vol. 1,
when the side ratios are greater than 2. pp. 359-366 (2007).
The aforementioned formulation and ANN program [6] Cheng, C. M., Wang, J. and Chang, C. H., e-wind: an
have been used in later developed programs for wind Integrated Engineering Solution Package for Wind
load calculations with decent results [10]. Sensitive Buildings and Structures, Journal of Wind
& Engineering, Vol. 5, No. 2, pp. 50-59 (2008).
Acknowledgements [7] Bitsuamlak, G. T. and Godbole, P. N., Application of
Cascade-correlation Learning Network Fordetermina-
Financial supports for this research from MOST of tion of Wind Pressure Distribution in Buildings, Wind
ROC and Wind Engineering Research Center at Tam- Engineering into the 21st Century, Balkema, Rotter-
kang University Taiwan are gratefully acknowledged. dam (1999).
[8] Chen, Y., Kopp, G. A. and Surry, D., Prediction of
References Pressure Coefficients on Roofs of Low Buildings
Using Artificial Neural Networks, Journal of Wind
[1] Wang, J. and Cheng, C. M., The Application of Arti- Engineering and Industrial Aerodynamics, Vol. 91,
ficial Neural Networks to Predict Wind Spectra for pp. 423-441 (2003). doi: 10.1016/S0167-6105(02)00
Rectangular Cross-Section Buildings, Proceedings 381-1
of Fifth International Symposium on Computational [9] Krian, J., Gaparac, G., Kozmar, H., Antoni, O. and
Wind Engineering (CWE2010), Chapel Hill, North Ca- Grisogono, B., Designing Laboratory Wind Simula-
rolina, U.S.A, May 23-27 (2010). doi: 10.5359/jawe. tions Using Artificialneural Networks, Theoretical
35.347 and Applied Climatology, Vol. 120, No. 3, pp. 723-
[2] Wang, J. and Cheng, C. M., The Role of Artificial 736 (2015). doi: 10.1007/s00704-014-1201-4
Neural Networks in a Building Design Wind Load [10] Wang, J. and Cheng, C. M., Aero-Data Based Wind
Expert System Based on Aerodynamic Databases, Resistant Design of Rectangular Shaped Tall Build-
ICWE 13, Jul. 10-15, Amsterdam, Netherlands, Paper ings, International Conference on Innovations in Civil
#191 (2011). and Structural Engineering (ICICSE15), June 3-4,
[3] Wang, J. and Cheng, C. M., Web-Enabled Design Istanbul, Turkey (2015).
Wind Load Expert System for Tall Buildings, Pro-
ceedings of the 6th Asia-Pacific Conference on Wind Manuscript Received: Apr. 20, 2016
Engineering (APCWE VI), Seoul, Korea, Sep. 12~14, Accepted: Oct. 14, 2016

You might also like