融合全局和序列特征的多变量时间序列预测方法

Combining Global and Sequential Patterns for Multivariate Time Series Forecasting

时间序列在现实生活中具有广泛的用途,使用时间序列预测模型能够预估序列的未来变化趋势,为决策提供支撑.对于多变量时间序列的预测研究,已经提出了很多模型,但已有方法存在如下问题:不能同时考虑时间序列本身和协变量的信息;忽略了多变量时间序列中的全局信息;不能对预测结果进行解释.针对这些问题,本文提出了一个基于深度学习的多变量时间序列预测模型TEDGER,可以提取隐藏在单个时间序列中的序列模式和隐藏在多变量时间序列中的全局特征,并将序列模式和全局特征进行融合,通过残差预测的方式实现时间序列的预测.本文所提模型在真实的时间序列数据集上进行了实验评估.结果表明,本文提出的模型在预测准确度上超越了其他基准模型,同时模型拥有一定的可解释性.

Time series is widely used in real world with many applications nowadays.By developing time series forecasting models,we can predict how the time series evolve in the future,which provides support for decision making in different scenarios.Many models have been proposed in literature.Existing studies on multivariate time series forecasting either cannot take both times series and covariates into consideration,lack interpretability,or ignore global trends across multivariate time series.To solve these issues,we propose a new deep learning based multivariate time series forecasting model TEDGER(Tensorized recurrent encoder-decoder framework with global patterns and residual forecasting).TEDGER can capture sequential patterns hidden in individual time series and can extract global trends hidden across multivariate times series.Specifically,our proposed model follows encoder and decoder framework and we adopt Tensorized Long Short-Term Memory network as the basic processing unit to provide the model with possibility to distinguish the importance of different features.The encoder is designed to capture sequential pattern hidden in each time series,in which two different kinds of attentions are designed to weigh the importance of historical information and covariate information,offering interpretability to the model in the same time.Global change pattern hidden in multiple time series is extracted based on temporal regularized matrix factorization.We further propose two different ways to combine sequential pattern and global trend to make residual prediction for final time series forecasting,which correspond to two variants of our proposed model:TEDGER I and TEDGER II.We provide time complexity analysis of the proposed model.Meanwhile,we conduct comprehensive experiments on real-world time series datasets to evaluate the performance of the proposed model.We compare the performance of our proposed model with eight benchmark models,which belong to three different categories.The comparison results demonstrate superior performance of our proposed model over benchmark models.Ablation study is conducted to evaluate the necessity of the global pattern exaction module and results confirm its benefit.We also check the impact of different input window size and different future steps in prediction on the model’s performance.Case study shows that our proposed model can explain the forecasting results to some extent.