基于神经网络的大跨度屋盖非高斯风压场模拟方法

Non-Gaussian Wind Field Simulation Method for Large-span Roofs Based on Neural Networks

采用径向基函数神经网络(Radical Basis Function Neutral Networks,简称RBF神经网络)来模拟大跨度结构的非高斯风压场.根据某大跨度结构的形式特点,将结构风场看成是屋面位置和时间的函数,将风压场分解为一系列径向基函数.再利用单调非线性无记忆转换映射和RBF中获得的风场函数定义向量过程,从而将非高斯场的模拟转换为互相关高斯过程的模拟.将RBF神经网络应用于一大跨度屋盖的非高斯场模拟,得到结构上非高斯风压场的分布.结果对比表明,RBF神经网络模拟非高斯风压场具有较高的准确性.该方法可直接利用RBF神经网络的输出结果,避免推导高斯过程和非高斯过程的关系式,因此具有较高的效率.RBF神经网络模拟非高斯风压场在准确性和效率上均具有显著优势.

Radical basis function neural networks（RBF neural networks for short） are adopted to simulate numerically non-Gaussian wind field of large-span roofs.According to properties of a large-span roof,the wind field is considered as the function of position and time,decomposed into a series of radical basis functions.And monotonic nonlinear memoryless transformation mapping and wind field function obtained from RBF neural networks are combined to define a transformation vector process,with which non-Gaussian process is transformed to Gaussian process for simulation.The proposed RBF neural networks are applied to the simulation of non-Gaussian wind field of a large-span roof.And the non-Gaussian wind field distribution on the roof is obtained.Comparison of results shows that RBF neural networks are highly accurate when simulating non-Gaussian wind filed.The method can make direct use of the outputs of RBF neural networks,without deriving formula between non-Gaussian process and Gaussian process.Thus,RBF neural networks have obvious edge in simulating non-Gaussian wind field of large-span roofs,both in accuracy and efficiency.