Traffic forecasting is crucial for location-based services. Recent studies tend to utilize dynamic graph neural networks to capture spatial-temporal correlations. However, urban traffic faces spatial heterogeneity of different lane structure at intersections, leading to different traffic patterns of lanes in different directions. It also confronts temporal heterogeneity of varying traffic trends in different time periods. Unfortunately, the influence of such spatial-temporal heterogeneity on traffic evolution are not fully considered in existing methods. To this end, this paper proposes a novel dynamic-graph-based model called HA-STGN, which integrates these heterogeneous features into spatial-learning components to model traffic networks in a finer granularity. Specifically, we design a dynamic graph model, which performs on a direction-aware road network to extract the structural information of intersections. Then, a time-sensitive attention mechanism is proposed to perceive the effect of time by introducing explicit temporal features. Moreover, an adaptive fusion module is provided to balance the spatial-temporal information adaptively. Finally, extensive experiments are conducted on two real-world datasets to verify the effectiveness of our model. The results show that our proposed HA-STGN can effectively capture spatial-temporal dependencies and outperform all the baseline methods. [ABSTRACT FROM AUTHOR]