<p>Traffic flow forecasting is a critical task in intelligent transportation systems, with wide applications in urban traffic planning and smart transportation systems. In recent years, advancements in spatio-temporal models have significantly enhanced the ability to model complex spatio-temporal correlations. Despite the commendable performance of existing methods, they still face three major limitations: (i) periodic fluctuations in traffic flow data; (ii) inadequate learning of interaction patterns between spatial nodes and temporal dynamics; (iii) temporal heterogeneity caused by outlier interference. To address these limitations, this paper proposes a model integrating hybrid spatio-temporal features with dynamic diffusion graph convolution, termed HySTDG. The model captures dynamic interactions between spatial and temporal dimensions through a dynamic interaction graph convolutional network, incorporates periodic patterns via an enhanced temporal embedding module, and models long-term dependencies in traffic networks using a graph diffusion process. Additionally, an anomaly-aware spatio-temporal attention mechanism is designed to effectively mitigate the impact of outliers on predictions. Experiments on multiple real-world traffic flow datasets demonstrate that HySTDG achieves superior predictive accuracy and robustness, offering an efficient and adaptable solution for traffic flow forecasting.</p>