【摘要】 Transformer在时间序列分析中展现巨大潜力,但其注意力机制常因聚合语义不相关的查询-键对而影响预测性能,同时,时间序列中存在的周期性、突发波动等复杂模式也增大建模难度.为此,文中提出全局-局部特征融合的多变量时序预测方法(Multivariate Time Series Forecasting Method with Global-Local Feature Fusion, MTSGLFF).首先设计Top K-Transformer,根据可学习的传感器嵌入动态生成稀疏掩码,保留关键序列,进行特征聚合.在此基础上,构建双分支时间序列预测框架,包含全局分支网络和局部分支网络.全局分支网络通过跨变量注意力机制捕获全局的交互特征;局部分支网络采用多尺度架构,将时间序列分解成多粒度模式,精细化建模局部相关性.在10个基准数据集上的实验表明,MTS-GLFF在长短期预测任务中的性能均较优.更多还原

【Abstract】 The Transformer demonstrates considerable potential in time series analysis. However,the attention mechanism of the Transformer often aggregates semantically irrelevant query-key pairs,thereby resulting in the degradation of prediction performance. Moreover, complex patterns in time series, including periodicity and abrupt fluctuations, pose additional challenges for effective modeling. To address these issues, a multivariate time series forecasting method with global-local feature fusion(MTSGLFF) is proposed. First, a TopK selection operator is designed. It dynamically generates sparse masks based on learnable sensor embeddings, thereby retaining key sequences for subsequent feature aggregation. Next, a dual-branch time series forecasting framework comprising global and local branch networks is constructed. The global branch captures global interaction features through a cross-variable attention mechanism, while the local branch adopts a multi-scale architecture that decomposes time series into multi-granularity patterns for fine-grained modeling of local dependencies. Experiments on 10 benchmark datasets demonstrate that MTS-GLFF achieves competitive performance.更多还原