摘要：

长期时序预测旨在辅助决策者进行长期规划与战略制定，然而现有预测模型在进行长期时序预测任务时忽视了经独立变量通道的特征提取的重要性，且未能有效融合多尺度时频信息进行深层特征提取。针对上述问题，提出一种多尺度时频融合与稀疏网络的长期时序预测模型WDNet。首先，通过重塑嵌入模块将一维时序数据分块后转换为二维时序图，并经过连续小波变换得到时频信息图，对时频信息图进行不规则稀疏卷积操作以获取融合的时频特征；然后，与时序数据的原有模式融合后，在集成了多层特定卷积的模块中，经独立变量通道分别从块内和块间提取深层特征；最后，经全局线性和局部非线性预测头合并输出最终结果。在6个真实公开数据集上的实验结果表明，WDNet模型能有效捕捉时序数据中的复杂依赖关系，并在长期时序预测任务上有较好的预测效果。

关键词: 长期时序预测, 时频融合, 稀疏网络, 双预测头策略, 深度学习

Abstract:

Long-term time series forecasting aids decision-makers in strategic planning. Existing models often neglect feature extraction from independent variable channels and fail to integrate multi-scale time-frequency information for deep feature extraction. To address this issue, this study proposes WDNet, a model that uses multi-scale time-frequency fusion and a sparse network to forecast long-term time series. WDNet reshapes one-dimensional time series into two-dimensional temporal graphs through chunking and an embedding module and uses continuous wavelet transform to obtain time-frequency information. Irregular sparse convolution is then applied to these graphs to capture the fused features. These features, combined with the original time series data, are processed through independent variable channels to extract deep features from within and between blocks. The final output is generated by merging the global linear and local nonlinear prediction heads. Experimental results on six real-world public datasets demonstrate that WDNet consistently reduces prediction errors across various models, effectively capturing complex dependencies and improving the long-term forecasting performance.

Key words: long-term time series forecasting, time-frequency fusion, sparse network, dual prediction head strategy, deep learning