融合分解与转置策略的多变量时间序列预测模型  

A multivariate time series prediction model integrating decomposition and invert strategies

作　　者：张金涛 程明月 刘芷町 Zhang Jintao;Cheng Mingyue;Liu Zhiding(State Key Laboratory of Cognitive Intelligence,University of Science and Technology of China,Hefei,230026,China)

机构地区：[1]认知智能全国重点实验室,中国科学技术大学,合肥230026

出　　处：《南京大学学报(自然科学版)》2025年第1期58-70,共13页Journal of Nanjing University(Natural Science)

基　　金：安徽省自然科学基金(2408085QF193);中央高校基本科研基金(WK2150110032)

摘　　要：时间序列预测是一项重要的数据分析技术,在交通、经济、气候等领域有广泛的应用,可以辅助资源合理分配、重大风险决策以及规划未来走向.近年来,随着机器学习和深度学习方法的发展,多变量时间序列预测问题受到广泛关注,然而,现有多变量时间序列预测方法无法同时捕捉复杂的时序间和变量间的依赖关系.提出一种融合转置嵌入方法与时间序列分解的时间序列预测模型DItrans.首先,针对时间序列进行趋势项、周期项和残差项分解,并在此基础上,分别执行转置嵌入,利用不同的编码器结构来学习表征.转置嵌入方法使DItrans模型可以更好地获取多变量之间的相关性,而趋势项、周期项和残差项的分解有助于捕获邻近时间点的信息.同时,DItrans模型引入一种新的编码器结构,其结构更灵活,使模型能够捕获更复杂的时间序列特征.在三个真实数据集上对提出的模型进行了性能评估.实验结果表明,DItrans模型的均方误差和平均绝对误差均取得了最佳效果,和对比算法相比,其均方误差下降了1.71%~79.28%,平均绝对误差下降了0.72%~57.52%.Time series forecasting is a crucial data analysis technique with wide⁃ranging applications in transportation,economics,climate studies,etc.It aids in the rational allocation of resources,major risk decision⁃making,and future planning.Recently,the development of machine learning and deep learning methods has brought significant attention to multivariate time series forecasting.However,existing methods often fail to simultaneously capture the complex dependencies between time points and variables.In this paper,we propose DItrans,a novel time series forecasting model that integrates a transpose embedding method with time series decomposition.Initially,we decompose the time series into trend,periodic,and residual components.Following this,we apply transpose embedding to learn the representations using different encoder structures.The transpose embedding method allows DItrans to better capture correlations between multivariate variables,while the decomposition into trend,periodic,and residual components aids in capturing information from neighboring time points.Additionally,DItrans introduces a new,more flexible encoder structure,enabling the model to capture more complex time series features.We evaluate the performance of the proposed model on three real⁃world datasets.Experimental results demonstrate that DItrans outperforms existing methods in terms of both mean square error(MSE)and mean absolute error(MAE).Specifically,DItrans achieves reductions in MSE ranging from 1.71%to 79.28%and reductions in MAE ranging from 0.72%to 57.52%compared to benchmark algorithms.

关 键 词：时间序列预测 多变量时间序列 深度学习 时间序列分解 

分 类 号：O211.61[理学—概率论与数理统计] TP18[理学—数学]