基于灰关联-贡献度删减法的神经网络结构优化

摘要

考虑到神经网络结构对网络应用性能的影响,提出一种基于灰关联-贡献度的神经网络隐含层结构优化方法.计算隐节点输出与网络输出的灰关联度,以此为依据删除对网络输出影响较小的隐节点,直到剩余隐节点的灰关联度大于灰关联删减阈值;计算隐节点对网络输出的贡献度,删除贡献度较小的隐节点,直到剩余隐节点的贡献度大于删减阈值,得到较为精简的神经网络.将该方法用于风电场风速的神经网络预测模型中,仿真结果表明,采用该结构能够精简网络结构,与其它方法相比,该方法优化后的神经网络性能能够得到有效改善.%In view of the influence of the structure of neural network on the application performance of network, a structure optimization method based on grey correlation-contribution was proposed.The grey correlation degrees between the output of hidden nodes and the output of network were calculated, and hidden nodes which made a little influence on the output of network were deleted until the grey correlation degrees of the remaining hidden nodes were greater than grey correlation pruning threshold.The contribution degrees of hidden nodes for network output were calculated, and hidden nodes which have less contribution degrees were deleted until the contribution degrees of the remaining hidden nodes were greater than contribution pruning threshold, so a simpler neural network was built.The structure optimization method was applied to wind speed forecasting models based on neural network.Simulation results show that the grey correlation-contribution pruning method can effectively simplify structure of neural network.Compared with other optimization methods, the performance of neural network based on the proposed optimization method is improved effectively.