基于TG–LSTM神经网络的非完整时间序列预测

The prediction of incomplete time series via TG–LSTM neural network

中文关键词  数据缺失  非完整时间序列  长短期记忆神经网络  前向传播算法  学习算法

英文关键词  data missing  incomplete time series  long short-term memory neural network  forward propagation algorithm  learning algorithm

中文摘要

      针对传统模型对含数据缺失的非完整时间序列预测精度不高的问题, 利用长短期记忆(LSTM)神经网络强 大的时序建模能力, 提出一种带时间门的长短期记忆(TG–LSTM)神经网络. 首先, 提出一种能同时对输入值在线估 计和输出值实时预测的TG–LSTM单元结构; 其次, 基于TG–LSTM结构设计一种网络的前向传播算法, 实现输入填 补和输出预测同步进行; 然后, 建立TG–LSTM神经网络的学习算法来对输入填补和输出预测任务整体训练; 最后, 通过在Mackey-glass基准数据集, 月平均气温数据集和污水处理出水氨氮预测中的实验结果表明: 与传统方法相比, TG–LSTM神经网络模型能以更高精度对非完整时间序列进行填补和预测.

英文摘要

      In order to solve the problem that the poor prediction accuracy of traditional model for the time series with missing data, the strong time series modeling ability of long short-term memory (LSTM) neural network is used and the long short-term memory neural network with time gate (TG–LSTM) is proposed. Firstly, the TG–LSTM unit structure is proposed, which can be used to realize the online estimation of network inputs and real-time prediction outputs, simultaneously. Secondly, the forward propagation algorithm is designed according to the TG–LSTM structure to realize the synchronization of input filling and output prediction. Furthermore, the learning algorithm of TG–LSTM neural network is established to uniformly train input filling and output prediction tasks. Finally, the experimental results of the Mackey-glass benchmark data set, monthly average temperature data set and the ammonia nitrogen concentration prediction of waste water treatment process show that compared with the traditional methods, the incomplete time series can be filled and predicted by the TG–LSTM neural network model with higher accuracy.