Real-time urban traffic flow prediction and congestion cause analysis based on multimodal open-source data and deep learning

Abstract

Accurate prediction of urban traffic flow and timely identification of congestion causes are essential for intelligent transportation systems. However, existing approaches typically rely on single-modality traffic data and fail to incorporate contextual factors such as weather conditions, special events, and road network topology. This paper proposes a multimodal deep learning framework that integrates open-source traffic speed and volume data, road network topology, weather records, and event information for real-time traffic flow prediction and congestion cause analysis. The framework employs a Graph Convolutional Network (GCN) to capture spatial dependencies among road segments, a Temporal Convolutional Network (TCN) to model temporal traffic patterns, and a cross-attention fusion module to integrate heterogeneous contextual features. A multi-task learning head jointly predicts traffic flow values, congestion levels, and congestion causes. Experiments on the PeMS-BAY and METR-LA benchmark datasets augmented with weather and event metadata demonstrate that the proposed model achieves a Mean Absolute Error (MAE) of 14.6 and Root Mean Squared Error (RMSE) of 20.8 on PeMS-BAY, outperforming state-of-the-art baselines including GCN-LSTM (MAE 17.8) and standalone TCN (MAE 19.2). The congestion cause classification achieves 87.3% accuracy across five categories.