Accurate traffic flow prediction is crucial for ensuring the safe operation and efficient scheduling of intelligent transportation systems. Despite significant progress, current prediction models still face challenges due to the complex spatio-temporal correlations in traffic networks and the nonlinear nature of traffic data. These challenges include insufficient accuracy in spatio-temporal dynamic modeling, difficulties in jointly representing temporal dependencies and spatial heterogeneity, and the need for deeper integration of spatio-temporal features. To address these issues, this study proposes a traffic flow prediction model, termed as DSAGACN, based on cross-scale interactive attention and dynamic graph attention convolutions. In the temporal dimension, a cross-scale interactive attention mechanism is designed. Through multi-scale local context awareness and dynamic interactions, a multi-level temporal receptive field spanning “point, trend, and global” scales is developed, thus effectively capturing dynamic traffic flow variations. In the spatial dimension, a dynamic graph attention convolution module is introduced and combined with spatial positional embedding. Attention mechanisms are leveraged to adaptively capture spatial correlations and represent heterogeneity. At the feature fusion level, an efficient fusion mechanism is proposed to enhance the joint representation of spatio-temporal features while reducing computational complexity. Experimental results on two real-world traffic datasets show that DSAGACN effectively extracts and integrates spatiotemporal features from traffic data, achieving higher prediction accuracy than existing mainstream models.