Existing traffic flow forecasting work has two limitations: most methods in modeling the diffusion process of traffic flow do not account for information transmission delays and lack the ability to handle information from geographically distant areas with longer propagation delays. Additionally, most methods focus solely on capturing geographical spatial features, neglecting the semantic spatial features that may arise from different traffic patterns in various regions. To tackle these issues, we introduce an Enhanced Spatio-Temporal Extended Pattern Diffusion LSTM Network (SE-PDLSTM) to improve the precision of traffic flow forecasting. Building upon the traditional LSTM model, we design a diffusion convolution module to capture medium- and short-range geographical spatial dependencies in spatiotemporal data, according to the diffusion process of traffic flow. To improve the model’s capacity for processing information with longer propagation delays from geographically distant areas, we innovatively introduce a Spatial Extended Attention Module (SEAM), which improves the efficiency of capturing and integrating long-range information. Additionally, we incorporate a Traffic Pattern Extraction Module (TPEM), enabling the model to extract distinct traffic patterns exhibited by different regions. Extensive experiments on four traffic flow benchmarks confirm our model’s superiority, achieving SOTA results on all datasets with significant performance gains.