一种基于自演化预训练的多变量时间序列预测算法  

A Multivariate Time Series Forecasting Algorithm Based on Self-Evolution and Pre-training

作　　者：万晨 李文中[1,2] 丁望祥 张治杰 叶保留[1,2] 陆桑璐[1,2] WAN Chen;LI Wen-Zhong;DING Wang-Xiang;ZHANG Zhi-Jie;YE Bao-Liu;LU Sang-Lu(State Key Laboratory for Novel Software Technology,Nanjing University,Nanjing,210023;Department of Computer Science,Nanjing University,210023)

机构地区：[1]南京大学计算机软件新技术国家重点实验室,南京210023 [2]南京大学计算机科学与技术系,南京210023

出　　处：《计算机学报》2022年第3期513-525,共13页Chinese Journal of Computers

基　　金：国家重点研发计划(2018YFB1004704);国家自然科学基金(61972196,61672278,61832008,61832005);江苏省重点研发计划项目(BE2018116,BE2017152);软件新技术与产业化协同创新中心;中德社会计算研究所资助

摘　　要：时间序列预测是典型的时间序列分析任务,对于辅助决策、资源配置、提前采取止损措施等方面有重要意义,在包括电力、气象、交通、商业等领域有广泛应用.近年来,时间序列预测算法一直是机器学习的热门研究领域,其中多变量时间序列预测是一个具有挑战性的任务.本文研究多变量时间序列预测的局部变量预测精度问题,即多变量预测需要在提升整体预测性能的同时保证局部单变量的预测精度.针对现有多变量时间序列预测算法不能保障局部变量预测精度的局限性,我们设计并实现了一种基于自演化预训练的多变量时间序列预测算法SEPNets.基于预训练的思想,SEPNets首先构建和训练单变量时间序列模型作为后续建模的基准.然后,通过拓展时序卷积网络和长短记忆(LSTM)单元来建模变量间复杂的时序依赖关系.通过将预训练模型和拓展模型进行融合再训练,SEPNets可以保障多变量时间序列预测的局部变量预测精度,并提升总体的预测性能.我们在5个真实数据集上对所提模型进行性能评估.实验结果表明,本文提出的SEPNets算法比现有算法获得相对最高的预测精度,同时在保障局部变量预测精度上具有更好的性能.Time series forecasting is a typical time series analysis task,which is of great significance for assisting decision-making,resource allocation,and taking stop loss measures in advance.It is widely used in fields including power,weather,transportation,and business.In recent years,time series forecasting algorithms based on machine learning have been a hot research field,among which multivariate time series forecasting is a challenging task.This paper focuses on the local variable forecasting accuracy of multivariate time series forecasting,that is,multivariate forecasting needs to improve the overall forecasting performance while ensuring the forecasting accuracy of local univariate forecastings.In view of the above problems,this paper analyzes the limitations of the multivariate time series forecasting algorithm.The multivariate time series forecasting algorithm SEPNets based on self-evolution pretraining is designed and implemented.Inspired by the pre-trained model,SEPNets first constructs and trains a univariate time series model as a benchmark for subsequent modeling;then expands the complex temporal dependence between the time-series convolutional network(1-dimensional convolution)and the long and short memory(LSTM)unit modeling variables Relationship;using the former as the latter pre-training model for fusion and retraining,SEPNets solves the problem of local variable forecasting accuracy for multivariate time series forecasting.The experimental results show that the SEPNets algorithm proposed in this paper ensures the forecasting accuracy of local variables to a certain extent while obtaining the relatively highest forecasting accuracy.

关 键 词：时间序列预测 多变量时间序列 机器学习 神经网络 预训练 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]