Abstract: Time series mining plays a pivotal role in domains such as renewable energy, meteorology, and finance, with growing interest in the analysis of multivariate multi-step time series. Existing deep neural network-based approaches for multivariate multi-step time series forecasting often suffer from complex model architectures and large-scale parameterization. These characteristics lead to substantial computational demands and high training costs. Moreover, most current prediction models focus predominantly on the time domain, processing either channel-independent or channel-mixer information, which limits their ability to simultaneously capture both correlated and independent channel features. This restriction can lead to reduced prediction accuracy, particularly when training data is scarce. To overcome these limitations, we propose a lightweight dual-channel time-frequency cross-attention network for multivariate multi-step time series forecasting. The network extracts both independent and mixed channel representations in the frequency domain and integrates them with the original time-domain signals via an attention-based fusion mechanism. This design enables the model to jointly leverage time-domain and frequency-domain information, thereby capturing global spatiotemporal dependencies more comprehensively. We evaluate the proposed method against eight state-of-the-art time series forecasting models on eight publicly available datasets. Experimental results show that, for example, on the representative ECL dataset, our model achieves improvements over Autoformer (NeurIPS 2022) of 17.55%, 12.87%, and 14.72% in MSE, MAE, and SMAPE, respectively. Furthermore, compared with Crossformer (ICLR 2023), our approach reduces number of parameters by 30.82%, and achieves a 66.07% reduction in training time relative to Pyraformer (ICLR 2021). These results demonstrate that the proposed network is an effective and efficient solution for multivariate multi-step time series forecasting.

摘要： 时间序列挖掘在可再生能源、气象和金融等领域中的重要性日益凸显，其中针对多元多步时间序列的分析尤其受到业界关注。目前基于深度神经网络的多元多步时间序列预测模型，其复杂的模型结构和庞大的参数体量，通常需要大量的计算资源来支撑时间序列预测任务的完成。此外，现有预测模型过分关注时域，仅能处理通道独立或通道混合的信息，限制了同时提取相关通道信息和独立通道信息的能力，导致预测精度下降，尤其在训练数据有限的情况下。为此，一种基于双通道时频域交叉的注意力网络被设计用以处理多元多步时间序列的预测问题，该网络在频域中对通道独立和通道混合两个通道的信息进行提取后，采用注意力机制将双通道的频域信息与时域原始信息进行融合，使得模型可以有效结合时域与频域的信息，进而更全面地捕捉到数据的全局时空关系。本文在8个公开的时间序列数据集上与8个知名的高性能时间序列预测算法进行对比，实验结果表明，以代表性数据集ECL为例，本文提出的算法在MSE、MAE、SMAPE指标上较之2022年NeurIPS上发表的Autoformer算法分别提升了17.55%、12.87%、14.72%；同时，新网络的模型参数量较之2023年ICLR上发表的Crossformer降低了30.82%，训练时间较之2021年ICLR上发表的Pyraformer降低了66.07%，结果证实本文设计的双通道时频域交叉注意力网络是一种轻量且高效的处理多元多步时间序列预测问题的新工具。