Study On Non-Gaussian Characteristic Analysis Methods Of Wind Pressures On Building Surface

The non-Gaussian characteristics of wind loads should be fully considered in the wind-resistant design of building structures,and cannot be ignored.Several hot issues in the field of non-Gaussian wind load are reviewed,their research statuses are summarized,and some unsolved problems are pointed out.On this basis,four problems about non-Gaussian wind pressures,i.e.,probability distribution fitting,extreme value estimation,realization simulation,and statistics interpolation,are studied in depth.Some main contents and conclusions are as follows:(1)The existing crossing rate theory-based extrema estimation algorithms for non-Gaussian wind pressure have limited application scopes,and their estimation accuracy for short-tail extrema is significantly lower than that for long-tail side.For solving these problems,a separate description(SD)algorithm is developed in this paper.In SD algorithm,Johnson transformation(JT)is used as the bridge to realize the bidirectional transform between underlying Gaussian process and target non-Gaussian process.JT is suitable for non-Gaussian distribution with arbitrary combinations of skewness and kurtosis.its applicable range covers the whole Pearson plane.Due to the data on short-tail side occupy a higher weight in the calculation of maxilmum likelihood function,SD algorithm proposes to fit the parent wind pressure distribution twice,the short-tail extrema are estimated by the fitting result of maximum likelihood method,and the long-tail extrema are estimated by the fitting result of moment method,so as to improve the estimation accuracy of the short-tail extrema.The comparison shows that the overall estimation error of SD algorithm is less than 4%,and its accuracy is higher than those of traditional algorithms.The advantages of SD algorithm are especially obvious on the short-tail side and in dealing with strongly non-Gaussian soften processes.(2)The existing extreme value theory-based extrema estimation algorithms for non-Gaussian wind pressure need a long-duration wind tunnel test to measure enough pressure records,which consumes a lot of resources.To solve this problem,a hybrid measurement and simulation-based(HMSB)algorithm is proposed in this paper.The first fourth-order moments and power spectral density(PSD)are set as targets to simulate a large number of wind pressure sequences,and then to obtain enough extreme value samples for fitting the Gumbel distribution.Considering that the marginal moments have a greater influence on the accuracy of extrema estimation than PSD.a simplified simulation method is proposed to maximize the simulation efficiency in context that the accuracy of marginal moments is guaranteed.The accurate simulation of moments and the approximate simulation of PSD can be realized.Computational example proves that the estimation accuracy of HMSB method is higher than those of traditional methods,and the overall estimation error is about 4%.A lar ge number of time histories are simulated in a short duration by HMSB algorithm,by which the test resources are effectively saved.(3)Two aspects of targets,i.e.,frequency and probability,are included in the simulation of stationary non-Gaussian process.Most of the traditional simulation methods are based on the idea that exerting frequency interference firstly and probability interference secondly.In this paper,a new simulation method is proposed based on another idea that exerting probability interference firstly a nd frequency interference secondly.In development of the new method,the transformation relationship between input and output low-order moments for linear filter system is derived,by which the problem of moment distortion is solved.The causes of two kinds of incompatibilities which are respective for the new idea and for the traditional idea are discussed.It is proved that the two kinds of incompatible ranges do not overlap completely.The new method can solve some incompatible situations of traditional methods,its validity and accuracy are demonstrated by several examples.In addition,the proposed method has the potential to simulate high-order correlated non-Gaussian processes.(4)Only a few methods can simulate nonstationary non-Gaussian processes,and they need iterative calculation which results in a low simulation efficiency.Based on the linear filtering technique,a new s-imulation method is proposed in this paper.The time-varying auto-regressive(TVAR)model is used to filter underlying non-stationary non-Gaussian white noise into target non-stationary non-Gaussian process.The new method needs no iteration and is easy to operate.During its development,the transformation relationship between the time-varying low-order moments of the input and of the output for TVAR model is derived,so that the time-varying low-order moments of underlying nonstationary non-Gaussian white noise can be calculated according to the simulation targets in advance.The traditional JT is upgraded to a time-varying version for generating nonstationary white noise inputs.A simple approach is proposed to determine the order of TVAR model.The feasibility and accuracy of the new method are proved by two examples.(5)The Gaussian process regression(GPR)technique is introduced into the interpolation of wind pressures.Compared with the artificial neural network technique,which are used in the existing methods,GPR is advantageous in self-adaptively selecting hyper-parameters and offering outputs with clear probabilistic significance.For the estimation of low-order statistics,the proposed GPR-based approach is more accurate than the traditional approach,and can supply the wind pressure information at where no measuring taps are arranged or under wind directions that are not measured.For the estimation of wind pressure time histories,the powerful ability of GPR in dealing with small sample problems is fully used in the proposed time-varying GPR-based approach.The new approach has high accuracy,and can deal with the nonstationary situation that wind field is changing with time.The higher-order statistic estimations are attempted for the first time.The third-order statistic can be estimated accurately,while the results for the fourth-order statistic are unsatisfactory,due to its stronger randomness and higher fluctuation amplitude.The approach needs to be further improved for the fourth-order statistic interpolation.The interpolation of cumulative density function is also attempted.The accuracies of several examples are on a high level and can meet the requirements of application.