基于神经网络语言模型的时间序列趋势预测方法  

Trend Prediction Method of Time Series Trends Based on Neural Network Language Model

作　　者：王慧健 刘峥 李云 李涛 WANG Huijian;LIU Zheng;LI Yun;LI Tao(School of Computer Science,Nanjing University of Posts and Telecommunications,Nanjing 210046,China)

机构地区：[1]南京邮电大学计算机学院

出　　处：《计算机工程》2019年第7期13-19,25,共8页Computer Engineering

基　　金：江苏省自然科学基金面上项目(BK20171447);江苏省高校自然科学研究面上项目(17JKB520024);南京邮电大学引进人才科研启动基金(NY215045)

摘　　要：对于时序数据的预测,传统方法多数通过分析历史数据预测出后面的一个或者多个具体值,但预测的具体数值准确率较低。为此,提出一种新的时间序列短期趋势预测方法。通过对时序数据进行离散化,用字符表示各个时间段数据的范围,并利用神经网络语言模型预测得到下一个字符,即下一段数据的范围。实验结果表明,与支持向量机、循环神经网络、随机森林等算法相比,在预测结果分为5个区间的情况下,该算法平均预测准确率为 56.7 %,具有较高的可行性,且由于字符表示带有语义信息,所得预测结果可以反映数据趋势以及趋势变化程度。For the prediction of time series data,most of traditional methods predict one or more specific values by analyzing the historical data,but the specific numerical accuracy of the prediction is low.Therefore,this paper proposes a new prediction method of time series short-term trends.It discretizes time series data,uses characters to represent the range of data for each time period,and uses the Neural Network Language Model(NNLM) to predict the next character,which is the range of the next segment of data.Experimental results show that in the circumstances where the prediction result is divided into five intervals,the average prediction accuracy of the algorithm is 56.7 %,which means it has higher feasibility compared with support vector machine,cyclic neural network,random forest and other algorithms.And because the character representation has semantic information,the prediction results can reflect the trend of the data and the degree of change.

关 键 词：数据挖掘 时间序列预测 数据离散 长短时记忆 神经网络语言模型 深度学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]