In the past few years, countries worldwide have demonstrated a great deal of interest in intelligent transportation systems (ITSs), with some notable results. Numerous academic institutions have focused on traffic prediction in ITSs. However, because of their instability and complex geographical topology, traffic characteristics are difficult to forecast. Most existing studies rely on historical data from the previous hour to make predictions and separate both the temporal and spatial elements of traffic flow. We argue that the shortcomings of this approach are its inability to consider spatio-temporal unity and the difficulty of extracting additional hidden features for model learning. This paper proposes a deep learning-based spatio-temporal aggregation graph attention network (STAGAT) that simultaneously captures spatio-temporal information by first constructing spatio-temporal adjacency matrices cross-space-time and stacking feature matrices on the basis of daily features and contemporaneous features, followed by a graph attention network (GAT) to capture the degree of correlation between nodes. The experimental results on the publicly available datasets PeMSD4, PeMSD7, and PeMSD8 demonstrate that our STAGAT model outperforms other baseline models.