变窗口神经网络集成预测模型

Neural networks ensemble based on variable-window model

针对时间序列问题,提出了一个变窗口神经网络集成预测模型。利用自相关分析方法挖掘时间序列本身蕴涵的变化特性,并利用这些变化特性构造差异度较大的个体神经网络。变窗口集成预测模型在应用于时间序列预测的同时,还可以有效地对异常序列进行筛选和分离。将该模型应用于移动通信话务量的预测。实验分析表明,该预测系统具有较高的预测精度,并能有效地对异常序列进行分离。

This paper propsoed a novel model, called variable-window neural network ensembles to improve the ability of generalization. First it took use of the self-correlation analysis method to calculate the self-correlation coefficients of the time series. Then used the coefficients to construct all the individual neural networks of the ensemble. The forecasting model could also be used to detect the outliers among the time series data. Applied the model to forecast the telephone traffic. The experiments demonstrate the proposed model is accurate and effective in the forecasting of telephone traffic.