摘要
针对回声状态网络（Echo State Networks，ESNs）输入序列延迟时间（和嵌入维数D的选择以及储备池的适应性问题，利用自相关性分析法从被预测样本序列构建ESNs网络输入，并通过移动通信话务量的预测问题，采用实验分析的方法讨论了储备池参数选择对于时间序列预测性能的影响.与采用ARMA和BP神经网络的预测方法相比，新方法在保证预测精度和效率的情况下，具有更好的泛化能力.
The choice of delay time and embedded dimension in time series modeling and prediction and the problem of reservoir adaption are challenges for Echo State Networks （ESNs）. Correlation analysis is introduced to construct the inputs vector from the time series in ESNs networks. Moreover, the effects of different parameters settings on prediction performances in reservoir is analyzed by experiments of mobile communication traffic prediction. Compared with ARMA and BP neural networks, the pro- posed method can ensure not only the accuracy and efficiency but also the good generalities.
出处
《电子学报》
EI
CAS
CSCD
北大核心
2010年第B02期148-154,共7页
Acta Electronica Sinica
基金
装备预研重点基金(No.9140A17040409HT01)
教育部高等学校博士学科点专项科研基金(No.20092302110013)
关键词
回声状态网络
自相关系数
时间序列
移动通信话务量
echo state networks
self-correlation coefficient
time series
mobile communications traffic
