Hybrid Graph Convolutional Networks and Long Short-Term Memory Model based on Heterogeneous Data Sources for Traffic Speed Prediction

In the modern era, traffic speed prediction has become essential for intelligent transportation systems by utilizing diverse data sources such as global positioning system, weather, and mapping services. However, a major challenge persists integrating heterogeneous big data sources while maintaining spatial and temporal consistency, which often obstructs the prediction accuracy and system scalability. Hence, a hybrid deep learning method based on Graph Convolutional Networks and Long Short-Term Memory (GCN–LSTM) is proposed in this study. At first, traffic data is gathered from Floating Car Data (FCD), Google Maps ETA, OpenStreetMap (OSM), live weather and calendar APIs. The necessary preprocessing operations such as map matching, temporal normalization, and correction of outliers are performed to standardize the data. Pearson Correlation Coefficient (PCC) is employed to analyze linear dependency and Mutual Information (MI) to pick up nonlinear dependency with the target variable future speed. Then, GCN is employed to capture spatial dependencies across road networks, while LSTM captures the temporal traffic flow trends. The proposed GCN–LSTM model achieved improved performance with reduced Root-Mean-Square-Error (RMSE) (4.5) compared to existing LSTM-Gated Recurrent Unit (GRU) models, which making it more effective for real-time urban traffic prediction.