Traffic data often exhibits significant dynamic spatial correlations and sudden pattern changes. Traditional models struggle to fully capture the spatio-temporal dependencies of real-world transportation networks. To address this challenge， a Bi-directional Diffusion Dynamic Graph convolutional Network with multi-source information fusion and graph-level attention for traffic flow prediction （BDDGNet） was proposed. The network integrated spatio-temporal information and multi-source dynamic features， utilizing a multi-head attention mechanism to construct dynamic graph adjacency relationships between nodes， thereby capturing spatial connectivity strength evolving over time in the traffic network. For spatial modeling， a bi-directional diffusion dynamic graph convolution module was designed to learn the propagation dynamics of node states in both directions， thereby enabling collaborative modeling of spatial dependencies in traffic flow. At the same time， a graph-level attention mechanism was introduced to extract global semantic information of key nodes， thereby enhancing the perception capability for the overall topological structure of the traffic network. Experimental results on two real-world traffic flow datasets， PEMSD4 and PEMSD8， showed that for 15-minute， 30-minute， and 60-minute prediction tasks， on the PEMSD4 dataset， BDDGNet achieved the Mean Absolute Error （MAE）， Root Mean Square Error （RMSE）， and Mean Absolute Percentage Error （MAPE） of 18.26， 30.26， and 12.16%， representing reductions of 25.2%， 19.3%， and 26.9% compared with Diffusion Convolutional Recurrent Neural Network （DCRNN） based on static graph structure， respectively； on the PEMSD8 dataset， BDDGNet achieved the MAE， RMSE， and MAPE of 14.07， 23.44， and 9.24%， representing reductions of 16.3%， 11.1%， and 15.4% compared with DCRNN， respectively. The proposed network can capture spatio-temporal dependencies in dynamic traffic environments more accurately and significantly enhance prediction accuracy.