摘要
针对多尺度时间序列各尺度发展趋势及整体预测问题,建立小波分解回声状态网络预测模型(wavelet decomposi-tion and echo state networks,WDESN),根据各尺度的不同性质选取与之相匹配的回声状态网络模型(echo state networks,ESN),同时,通过在各尺度条件下引入权值系数实现预测分量最优整合,提高整体预测精度。预测带噪多尺度正弦序列实验表明:WDESN模型与ESN、支持向量机及BP神经网络模型相比预测精度较高。目前,该模型已成功用于移动通信话务量的预测,并满足了现实系统的精度要求。
Based on the prediction of multi-scale time series on each scale and overall trend,a new WDESN(wavelet decomposition and echo state networks) model was proposed in this paper.Firstly,according to different property of each sale,ESN(echo state networks) were properly selected to match series of each sale.Meanwhile,though introducing the weight coefficient of respective sale,the optimal combination of the prediction components was realized.Finally,the whole forecasting precision was improved.Experiment results on simulation sequence show that,the prediction performance of proposed method was superior to that of standard ESN,SVM(support vector machine) and BP neural networks.Currently,the proposed method has been successfully applied in the prediction of real mobile traffic,meeting the precision requirements of real system.
出处
《电子测量与仪器学报》
CSCD
2010年第10期947-952,共6页
Journal of Electronic Measurement and Instrumentation
基金
哈尔滨工程大学信通学院自由探索计划(编号:GK2080260103)资助项目
关键词
多尺度时间序列
小波分解
回声状态网络
multi-scale time series
wavelet decomposition
echo state networks