Urban congestion is a big problem in our cities. It leads to commuter delays and economic inefficiency. More tragically, though, it leads to a million deaths annually worldwide. Research in the International Journal of Reasoning-based Intelligent Systems shows how artificial intelligence (AI) might be able to carry out real-time traffic forecasting and so provide a way for the authorities to manage our road networks better.
Road vehicles do not behave as individual entities, traffic flow is a dynamic system in which there are no truly isolated events at individual locations, but conditions that ebb and flow over time. The researchers describe this phenomenon as spatiotemporal dependency. Events at one point at a given time influence conditions elsewhere on the roads. For example, a slowdown on a motorway might trigger congestion further down the route or in areas fed by the motorway some time later.
The researchers explain that capturing these delayed and distributed effects has long proved difficult for conventional forecasting models. Existing systems rely on simplified assumptions or short-term data patterns. The new approach using a hybrid deep learning system known as STG-Former. This brings together two computational approaches: graph neural networks and transformer models. A graph neural network represents the road system as a network of connections. The model can thus learn about traffic conditions over an area. The transformer component uses an attention mechanism to identify the most relevant information at any given time. It can thus detect patterns as they change through time.
Tests with this new system on standard traffic datasets show the model is much more accurate in its predictions than even the leading rivals and works well during periods of peak congestion when those other models often fail. The improvement is significant in the context of urban congestion, where even a small improvement in predictions can help traffic management improve its operational decisions and so avoid gridlock or major stalls in the flow of traffic.
Cheng, H., Cao, Y. and Li, W. (2026) ‘Transformer-GNN hybrid architecture for optimising real-time traffic forecasting on highways’, Int. J. Reasoning-based Intelligent Systems, Vol. 18, No. 9, pp.38–50.