一种改进的神经网络集成法预测PMV指标

为解决大样本的PMV指标预测问题,采用基于模糊聚类的神经网络集成方法,将原始样本集模糊划分为多个相交子集,通过这些模糊子集训练神经网络得到预测个体,再对个体输出加权结合获得预测结果.在进行神经网络集成过程中,采用微粒群算法有效克服了聚类和常规神经网络训练容易陷入局部最优的缺点,总结出一种更加有效的神经网络集成方法.实验结果表明:基于微粒群的神经网络集成算法有较好的全局优化性能,其集成的神经网络系统能更准确地预测PMV指标.

An Improved Neural Network Ensemble for the Prediction of PMV Index

In order to predict the PMV index more effectively on a large sample set,a neural network ensemble method based on fuzzy c-means clustering(FCM) is presented. By using the FCM algorithm,the original sample set is divided into some intersectant subsets,and with these subsets the corresponding individual neural networks can be trained as parts of an integrated system.When putting this system into use,the prediction result can be obtained by summing the products of the individual networks' outputs and weights.However,the method consists mainly of two local optimum algorithms,of which one is the hill climbing method in fuzzy c-means clustering,the other the steepest descent method in training BP neural network.As an effictive improved scheme,the particle swarm optimization is introduced to cover the shortage of local optimum.In this way,the performance of the predictive system is greatly promoted.Experiments indicate that the method has a good performance of global optimum in neural network ensemble,and the compositive network system can predict the PMV index more accurately.