摘要
针对长短期记忆网络(LSTM)算法对时间序列预测存在的不足,考虑到样本序列如果包含线性关系或含有噪音时LSTM算法预测将不准确,同时分析了变分自编码器(VAE)对异常样本修复的原理,提出了一种改进的LSTM时间序列预测算法VAE_LSTM,将VAE网络修复样本的思想加入到传统的LSTM网络,对样本序列进行修复后再输入LSTM神经网络训练,最终建立了时间序列预测模型.阐述了模型建立的方法与步骤,详细分析了模型的原理.使用长江汉口历史水文数据序列进行仿真实验,结果表明:VAE_LSTM算法预测模型在时间序列预测方面有较好表现,满足预测精度要求,比传统LSTM时间序列预测模型的预测准确性高,尤其中短期预测更为准确;对比实验同时表明此模型准确性高于ARIMA,RNN等预测模型.
In view of the shortcomings of long short term memory(LSTM)algorithm in time series prediction,if the sample sequence contains linear relationship or noise,the LSTM algorithm prediction will not be accurate,and considering that the principle of VAE to repair abnormal samples,an improved LSTM time series prediction algorithm VAE_LSTM was proposed.The idea of repairing samples was added to the traditional LSTM network.The sample sequence was input the LSTM neural network for training after repaired,and finally established the time series prediction model.The method and steps of model building were described,and the principle of model was analyzed in detail.Using the historical hydrological data series of Hankou of the Yangtze River for simulation experiments,the results show that the VAE_LSTM algorithm prediction model has better performance in time series prediction,meets the prediction accuracy requirements,and is more accurate than the traditional LSTM time series prediction model,especially in the short and medium term;the comparative experiments also show that the accuracy of this model is higher than ARIMA,RNN and other prediction models.
作者
杨英
唐平
Yang Ying;Tang Ping(School of Information,Guangdong Communication Polytechnic,Guangzhou 510080,China;School of Automation,Guangdong University of Technology,Guangzhou 510000,China)
出处
《湖南科技大学学报（自然科学版）》
CAS
北大核心
2020年第3期93-101,共9页
Journal of Hunan University of Science And Technology：Natural Science Edition
基金
广东高校省级重点平台和重大科研项目资助(2017GKTSCX0)。
作者简介
通信作者:唐平,E-mail:racheltang7@163.com。
