Enhancing Spatio-Temporal Traffic Prediction through Hybrid Deep Learning Architectures and Attention Mechanisms

Abstract

Precise traffic forecasting is an important element of Intelligent Transportation Systems (ITS), facilitating proactive traffic control, congestion relief, and enhanced mobility. The conventional statistical and machine learning approaches tend to struggle in capturing the intricate spatio-temporal interdependencies in traffic stream, restricting their utility in dynamic and congested environments. Recent developments in deep learning—e.g., Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Graph Neural Networks (GNNs)—are proving useful, but a single architecture is unable to capture the multi-faceted nature of traffic data. This paper presents a new hybrid deep learning model combining CNNs for spatial feature extraction, RNNs for temporal sequence modeling, and GNNs for modeling road network topologies, supported by attention mechanisms to weight dynamic relevant features. The model is tested on real-world highway traffic flow data and compared with ARIMA, LSTM, and Spatio-Temporal Graph Convolutional Networks (STGCN). Results show that the solution presented in this work attains better performance on various metrics (MAE, MSE, RMSE, MAPE) compared to baseline models and shows resilience under different traffic conditions. The results confirm the potency of hybrid models and attention mechanisms in improving traffic prediction and paving the way toward more accurate and efficient ITS technologies.