«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

27

期数:

2017年03期

页码:

35-38

栏目:

智能、算法、系统工程

出版日期:

2017-03-10

文章信息/Info

Title:

 A Prediction Model for Time Series Based on Deep Recurrent Neural Network

文章编号:

1673-629X（2017)03-0035-04

作者:

 杨祎玥[1]; 伏潜[2]; 万定生[1]

 1.河海大学 计算机与信息学院；2.江苏省交通规划设计院

Author(s):

 YANG Yi-yue[1]; FU Qian[2]; WAN Ding-sheng[1]

关键词:

 ; 小波分析; 深度循环神经网络; 时间序列; 预测

Keywords:

 ; wavelet analysis; DRNN; time series; prediction

分类号:

TP391

文献标志码:

A

摘要:

 针对水文时间序列的高度非线性和不确定性等问题,利用深度循环神经网络的时间序列预测能力,结合小波变换方法,将原始序列分解重构为多个低频和高频序列,针对各个子序列进行网络模型训练,建立一个基于小波变换的深度循环神经网络的水文时间序列预测模型(WA-DRNN).网络训练方法采用时间进化反向传播(BPTT)算法,逐步更新网络权值.实验结果表明,WA-DRNN模型较普通的DRNN模型在预测值的均方差和绝对误差上均有较好提升,并且由于该模型的多尺度特性,能够一定程度上减少模型预测引起的滞后作用.实验结果证明,WA-DRNN模型具有预测精度高、滞后误差小的优点,对深度学习算法在水文时间序列预测的应用上有一定帮助.

Abstract:

 Aimed at the problems of high-nonlinearity and nondeterminacy for hydrology time series,a prediction model for hydrology time series based on Wavelet Analysis and Deep Recurrent Neural Network ( WA-DRNN) is put forward by using the predictive capabil-ity of deep recurrent neural network,combined with the wavelet analysis for the reconstruction of the original time series and training of high and low frequency series. The network training adopts Back Propagation Through Time ( BPTT) algorithm to update the network weight. The experiment shows that the WA-RNN model is better than the normal DRNN model in the mean square error and absolute er-ror,and for the reason of multiscale the model can decrease the lag of prediction. It turns out the WA-DRNN model has advantages of higher predictive accuracy and less lag,which is helpful for application of hydrology time series prediction of deep learning algorithm.

常用功能

更新日期/Last Update: 2017-05-12