<p>Traffic prediction is essential for intelligent transportation systems, yet achieving accurate prediction still presents challenges. Current methods predominantly revolve around spatiotemporal graph modeling, employing predefined or dynamically learned adjacency matrices to represent spatial relations and focusing on aggregating spatial information for individual time steps. However, the cross-temporal spatial correlation present in traffic flow data should be more noticed in these studies. This study proposes a cross-temporal global spatial attention network (CTGSAN) to address cross-temporal spatial dependency challenges. In CTGSAN, our designed global spatial attention module (GSAM) effectively captures cross-temporal spatial dependencies across the entire spatiotemporal domain. The key innovation of CTGSAN lies in its ability to explicitly model the dynamic spatial correlations that evolve over time, which are often overlooked by traditional methods. Additionally, we integrate a long and short-term convolution module to capture temporal correlations at the individual node level. This dual-module design allows CTGSAN to not only capture the complex spatial dependencies but also effectively model the temporal dynamics, leading to more accurate predictions. We conduct experiments on four public traffic datasets. The results show significant advantages of our method over other state-of-the-art baselines. Specifically, CTGSAN demonstrates superior performance in long-term prediction tasks, where the ability to capture cross-temporal spatial dependencies is crucial. Moreover, ablation experiments are performed on the global spatial attention module, proving the crucial role of cross-temporal spatial correlations.</p>