基于回归神经网络的大跨度结构风压场预测

Prediction of wind pressure field for large-span roofs based on recurrent neural networks

为提高计算效率并降低存储耗时,提出一种局部回归神经网络方法用来预测大跨度结构风压场。将原有整体优化问题分解为神经元层的子问题进行处理,基于在线学习方法中的递归预报误差法对局部回归神经网络进行训练。首先给出了整体递归预报误差算法,对所有可调权进行同步处理,然后将整体优化问题分解为子问题推导出局部回归神经网络,在更新递归过程中使用二阶信息。将该方法应用于大跨度屋盖的风压预测中,并将计算结果与传统神经网络计算结果进行了比较。结果表明,本文方法的计算误差小,收敛速度快,达到了提高计算效率和降低存储耗时的目的。提出的局部回归神经网络方法为大跨度结构风压场的预测提供了准确高效的方法。

To improve computing efficiency and reduce storage time,a local recurrent neural network is proposed to predict the wind pressure field of large-span roofs. The original global optimization problem is divided into a series of sub-problems in neuron layer,and the local recurrent neural networks are trained on the basis of the recursive prediction error among on-line training methods. Firstly,the global recursive prediction error algorithm is presented with all tunable weights disposed simultaneously. Then the global optimization problem is divided into sub-problems to derive local recurrent neural networks and the second order information is employed during the updating recurrent process. The proposed method is applied to wind pressure prediction of large span roofs. Results from the current method are compared with those from traditional neural networks. The results show that the present method exhibits small error and rapid convergence,achieving the goal of improving computation efficiency and reducing storage time. The method is an effective one for predicting wind pressure field of large-span roofs.