<p>Accurate traffic forecasting is pivotal in the domain of intelligent transportation systems and urban planning. Although deep learning-based methods for traffic forecasting have shown significant advancements, it remains challenging to simultaneously capture both temporal and spatial dependencies in traffic flow problems. Moreover, existing research often overlooks the complex spatio-temporal characteristics of intersections within discrete vehicle trajectories in urban traffic scenarios. To address these challenges, this paper proposes a novel spatio-temporal graph transformer (ST-Graphormer) for end-to-end traffic forecasting, which effectively extracts complex spatio-temporal features from urban traffic flow data. First, to tackle the complexities of urban traffic scenarios, we designed a bidirectional long short-term memory (BiLSTM)-based temporal trajectory encoder and a graph attention network (GAT)-based spatial topology encoder. Specifically, the temporal trajectory encoder is utilized to extract and fuse long-term and short-term temporal dependencies across various intersections from continuous vehicle trajectory information. The spatial topology encoder captures the spatial dependencies between intersections based on the intersection network topology graph. Subsequently, temporal trajectory encoding and spatial topology encoding are obtained, allowing for the effective modeling of spatio-temporal representations within complex traffic intersection networks. Additionally, a multi-head attention fusion mechanism is developed to integrate prior knowledge derived from temporal trajectory and spatial topology encodings, thereby enhancing traffic forecasting accuracy. Extensive experiments on real-world traffic datasets demonstrate that the proposed method outperforms existing state-of-the-art baseline methods.</p>