Abstract: As a key component of intelligent transportation systems, traffic flow prediction plays an important role in urban traffic management, transportation planning and alleviating traffic congestion. However, due to the inherent highly nonlinear characteristics and complex spatial-temporal dependencies of traffic flow data, existing methods struggle to achieve high-precision predictions. To make more accurate traffic flow prediction, a Multi-Scale Spatial-Temporal Fusion Attention Network (MSSTFAN) is proposed to solve the issue of traffic flow prediction. In the temporal dimension, a multi-scale temporal feature extraction module is designed. It decomposes multi-scale features through Fourier transform and models long-term trends and short-term fluctuations in traffic flow based on attention mechanisms in low-frequency and high-frequency channels. In the spatial dimension, complicated spatial dependencies are fully explored through a multi-level spatial feature extraction module. A geographic spatial feature extraction module is employed to capture stable and time-varying spatial relationships within the traffic road network. Additionally, a semantic spatial feature extraction module is utilized to model similar traffic flow states among different nodes. The experimental results demonstrate that the proposed model outperforms the baselines on all three real-world traffic datasets, with the MAE metric decreasing by 2.24%, 2.67% and 4.09% respectively, compared with the suboptimal results.