Large-scale road network traffic state analysis faces challenges like network complexity, road coupling, and state variability. Advanced algorithms such as deep learning and reinforcement learning have shown promise. However, relying solely on neural networks often lacks interpretability. Although many existing studies focus on the spatiotemporal correlation, the abnormal state fluctuations are hardly overcome. This paper presents a novel information aggregation method, considering both spatial and temporal dimensions, inspired by the reverse K-nearest neighbor algorithm. It adaptively determines spatial relationships and temporal correlations to enhance practical applications. Using California’s PeMS data, the proposed method’s effectiveness has been validated. It has been demonstrated that spatiotemporal information aggregation can play a pivotal role in traffic predicting performance with the transformer-based method. A comprehensive congestion analysis of the California highway network can obtain the spatiotemporal distribution of congestion, the frequency of congestion for roads, and the identification of congestion regions.