An Adaptive Multi-Source Correlation Fusion Approach for Lane-Level Traffic Flow Prediction

Autonomous Vehicles (AVs) are essential to Intelligent Transportation Systems (ITS) and the future of transportation. Accurate lane-level traffic flow prediction is crucial for AVs to assess traffic conditions and make timely decisions, such as lane changes and vehicle following. However, the complexities of traffic environments and non-linear data distributions hinder the extraction of spatial and temporal features. Many studies use convolutional structures with adjacency matrices to capture spatial dependencies, but these often focus on a single traffic state, risking biased information and ignoring interconnections among multiple states. Additionally, they primarily derive spatial features from network topology, neglecting data-driven correlations. To address these issues, we propose the Adaptive Multi-Source Correlation Fusion (AMSCF) approach, which models spatial correlations to enhance lane-level traffic prediction. We extract spatial correlations from historical data and network topology using multi-source traffic data to construct a spatial graph integrated within a Graph Convolutional Network (GCN). Furthermore, we introduce a dynamically improved adjacency matrix that accounts for both physical connections and the impact of lane changes. Extensive experiments show that AMSCF outperforms state-of-the-art methods.