Traffic Flow Prediction on the Spatial Graph Convolutional Networks and Temporal Transformer Model

Abstract

This study investigates traffic flow forecast using a hybrid deep learning architecture that combines Spatial Graph Convolutional Networks (GCN) with a Temporal Transformer. Traditional deep learning models, like LSTM and STGCN, have demonstrated promising performance in spatio-temporal forecasting. But the biggest limitation is that they face restrictions in modeling long-range temporal dependence and they catch complex spatial correlations. To address these issues, the proposed model employs a spatial GCN to extract topological and spatial relationships among traffic sensors from the road network, while a Transformer-based temporal module models sequential dependencies through self-attention, enabling dynamic weighting of historical observations across time steps. The METR-LA dataset is used to evaluate the model, which contains short-term traffic speed record from sensors in Los Angeles. Experimental results indicate that the Spatial GCN + Temporal Transformer achieves better prediction accuracy compared to conventional baselines, especially in multi-step forecasting scenarios, benefiting from its ability to capture local temporal dependencies, as well as global temporal dependencies. The findings highlight the potential of combining graph-based spatial modeling with attention-driven temporal modeling for intelligent transportation systems, and suggest promising directions for extending the approach to large-scale, real-time traffic prediction and route optimization tasks.