基于灰关联和灵敏度的BP网络隐含层结构优化

Hidden Layer Structure Optimization of BP Network Based on Grey Incidence Degree and Sensitivity Degree

在优化BP神经网络隐含层结构时,采用灰关联剪枝法是每次删除灰关联度小于灰关联阈值的隐节点,该方法学习时间短,但由于灰关联阈值的选取具有一定的主观性,可能会导致误删节点或不能完全删除冗余节点;而采用灵敏度剪枝法是每次只删除灵敏度最小的一个隐节点,放学习时间较长;因此,提出一种基于灰关联——灵敏度的BP神经网络隐含层结构调整算法;首先在网络前期学习过程中,采用灰关联法对隐含层节点进行"粗删",直到剩余隐节点的灰关联度都大于动态灰关联闽值,然后在网络后期学习过程中,采用灵敏度剪枝法对隐含层节点进行"细删",直到删除后的学习误差增大,则保留该节点,并结束学习;文章将结构优化后的神经网络应用于风电功率预测,仿真结果表明,该方法在满足学习误差要求的同时,不仅精简了神经网络结构,而且避免了灰关联剪枝法中灰关联阈值精确选取困难所带来的问题。

When using grey incidence method to optimize the hidden layer of BP neural network structure, this method takes a short learning time to delete the hidden redundant nodes whose grey incidence degree is less than grey incidence threshold. But the selection of grey incidence threshold has certain subjectivity, which may result in deleting useful nodes by mistake or being unable to delete redundant nodes completely. When using sensitivity pruning method, only the node with the minimum sensitivity will be deleted each time, therefore it takes a long learning time. In view of this, this paper presents a BP structure optimization method based on the grey incidence and the sensitivity degree. At the early stage of the network learning, this paper uses grey incidence method to delete the redundant nodes rapidly, until the grey incidence degrees of the remaining hidden nodes are greater than dynamic grey incidence threshold value. Then in the later learning process of the network, it uses sensitivity pruning method to delete the hidden nodes precisely, until the learning error increases after the deletion of the node. Then keep the node and stop the learning. In this paper, the neural network with the optimized structure is applied to wind power prediction. The simulation results show that this method can meet the requests for forecasting error. And it can not only simplify the structure of neural network, but also solve the problem brought by grey incidence threshold＇ s precise determination of grey incidence method.