摘要： 多维时间序列（Multivariate Time Series，MTS）的各子序列包含不同时间跨度的多尺度特征，其中蕴含着发展过程、方向和趋势等信息，对预测具有不同的重要程度。然而现有的时间序列预测模型不能有效地捕获多尺度特征并评估其重要程度。针对上述问题，提出一种基于多尺度时序特征融合与双注意力机制的多维时间序列预测网络FFANet（multi-scale temporal Feature Fusion and dual Attention mechanism based time series forecasting Network），能够有效融合多尺度特征并关注其中重要部分，实现更精确的预测结果。首先，多尺度时序特征融合模块中并行的时序膨胀卷积层使模型具有多种感受域，从而提取时序数据在不同尺度上的特征，并根据重要性对其进行自适应融合。其次，双注意力模块对融合的时序特征进行重新标定，通过分配时序和通道注意力权重并加权至对应的时序特征，使FFANet聚焦对预测有重要贡献的特征。实验结果表明，本文提出的FFANet与AR、VARMLP、RNN-GRU、LSTNet-skip、TPA-LSTM、MTGNN和AttnAR时间序列预测模型相比，在Traffic、Solar Energy和Electricity数据集上的预测精度分别提升了30.78、31.76、22.81、11.25、7.51、3.80、1.88个百分点。

Abstract: Each variable of the multivariate time series contains multi-scale features of different time spans. These features contain abundant information about the development process, direction and trend, and possess different importance for forecasting task. However, existing time series forecasting models cannot effectively capture multi-scale features and comprehensively evaluate the importance of multi-scale features. To this end, a new time series forecasting model (FFANet) based on multi-scale temporal Feature Fusion and dual Attention mechanism is established in this study. This model can effectively fuse and concentrate important multi-scale features to achieve better forecasting performance. First, the parallel dilated convolutional layers in the multi-scale temporal feature fusion module makes the FFANet have multiple receptive fields. Therefore, the FFANet is capable of extracting and fusing multi-scale features adaptively. Then, the dual attention module is introduced to recalibrate the fused features. The FFANet focuses on key features by normalizing the temporal and channel weights to the corresponding fused features. Experimental results on Traffic, Solar-Energy and Electricity multivariate time series datasets show that the forecasting accuracy index of the proposed model is satisfactory. Compared with the Auto Regressive(AR), Vector Auto Regressive Multilayer Perception(VARMLP), RNN-GRU, Long- and Short-term Time-series network(LSTNet-skip), Temporal Pattern Attention(TPA-LSTM), Multivariate Time Series Forecasting Graph Neural Networks(MTGNN), and Attention-based Auto Regression(AttnAR) models, the prediction accuracy improved by 30.78, 31.76, 22.81, 11.25, 7.51, 3.80 and 1.88 percentage points, respectively.