As a critical component of Intelligent Transportation Systems (ITS), traffic forecasting serves as a fundamental tool for urban management. Nevertheless, the inherent spatiotemporal complexity of traffic flow data continues to present significant challenges for accurate prediction. On the one hand, the extant methods not only fail to account for the multi-scale temporal dependencies within the time domain, but also struggle to fully leverage the rich information in the frequency domain. On the other hand, spatial features display complex dynamic dependencies due to variations in both location and time. To address the above issues, we propose multi-scale dynamic graph convolutional networks for spatial-temporal traffic forecasting. In the temporal dimension, we introduce a multi-scale mechanism that utilizes dilated convolutions and attention mechanisms to learn features in both the time and frequency domains. In the spatial dimension, we propose a method for building dynamic graphs. Then, we use a novel dynamic graph convolution module and a spatial attention layer to capture the dynamic spatial dependencies in the traffic network. Extensive experiments on five real-world datasets demonstrate the effectiveness of our model, which outperforms the current state-of-the-art baseline methods in terms of forecasting performance.