FAV-ASTCL: Forecasting Aware Versatile Adaptive Spatio-Temporal Context Learning Framework for Real-Time Traffic Prediction

The application of urban traffic prediction plays a vital part in the implementation of sustainable ITS systems and smart city infrastructures. However, as urban traffic datasets continue to grow in speed, the existing Spatio-Temporal Graph Neural Networks (ST-GNNs) struggle to deal with the non-stationary “volatility” of real-world traffic flow. This paper proposes FAV-ASTCL: Forecasting Aware Versatile Adaptive Spatio-Temporal Context Learning. This network model has been designed to improve the resilience to variance in predicting future traffic flow in highly-volatile environments. To overcome the problem associated with “Static Topology Trap,” the FAV-ASTCL model utilizes Learnable Context Selector that adaptively updates the urban road network via dot product similarity search and alpha blending. Moreover, to control the effect of unpredictable external factors on forecasting accuracy, Exogenous Gating Mechanism has been added to filter out the weather and calendar telemetry information.We validated FAV-ASTCL on the benchmark METR-LA freeway dataset and transitioned into a real-world deployment in Hyderabad, India, across Five key traffic hubs. Our results demonstrate a transformative reduction in Mean Absolute Error (MAE) from 9.034 to 3.633, achieving an accuracy of 91.78%. We provide evidence that explicitly modeling volatility as a learnable context, rather than a noise factor, allows ST-GNNs to survive extreme event horizons—such as the sudden monsoon shower peaks common in Hyderabad—where baseline models suffer from catastrophic error propagation.