Accurate and reliable traffic speed prediction is crucial for enhancing traffic management efficiency and alleviating traffic congestion. To improve traffic speed prediction accuracy and capture complex dynamic spatio-temporal dependencies in traffic data, this study proposes a traffic speed prediction model based on a multi-scale dynamic spatio-temporal convolutional network (MDSTCN). First, an adaptive static spatial adjacency matrix is constructed based on graph convolutional network (GCN), and an attention mechanism is introduced to capture the directional dependencies and dynamic interactions of traffic flow, thereby uncovering intricate dynamic spatial relationships. Second, a multi-scale dilated convolution structure is adopted to extract local temporal features and long-term variation trends. Finally, the model is trained and tested on a real-world dataset, and various comparative experiments are conducted. The results show that, compared with the best baseline model (Graph WaveNet) the proposed MDSTCN achieves lower E_RMS, E_MA, and E_MAP. Furthermore, the fluctuation range of prediction errors across the 15, 30 min, and 60 min horizons are the smallest, highlighting its good adaptability and stability in long-term prediction tasks.