STLAformer: Segment-based Temporal Lag Attention effective for time series forecasting

In recent years, Transformer architectures have attracted considerable attention in multivariate time series forecasting due to their powerful sequence modeling capabilities. However, existing methods primarily focus on temporal dependencies while often overlooking complex lagged relationships among variables, where the current state of a variable is mainly influenced by the historical state of other variables at certain previous time steps, indicating strong lagged correlations among them. To address this limitation, we propose STLAformer, a novel model based on a Segment-based Temporal Lag Attention (STLA) mechanism. By partitioning multivariate time series into local segments and introducing a delay-aware attention module, STLAformer effectively captures cross-variable lagged dependencies. Additionally, we design a Temporal Preservation Segmentation Embedding (TPSE) module to independently encode each segment, preserving local dynamic features, and incorporate learnable temporal encoding to enhance the model's representation of temporal structures. Experimental results demonstrate that our approach achieves significant improvements in modeling capability and forecasting accuracy over existing Transformer variants on multiple real-world multivariate time series forecasting tasks, especially exhibiting stronger generalization ability when handling complex lag dependencies.