Artificial Intelligence‐Driven Traffic Flow Forecasting and Spatial Road Operation Risk Assessment
TL;DR
The proposed Adaptive Graph WaveNet with Temporal Feature Fusion model provides an integrated framework for accurate multistep forecasting and spatially interpretable road operation assessment, and traffic management resource allocation.
Abstract
Accurate multistep traffic flow forecasting and spatial assessment of road operation conditions are crucial for proactive urban traffic management. Existing models can capture temporal variation and road network dependence, but their accuracy often decreases as the forecasting horizon extends, while numerical flow forecasts alone cannot directly reveal where operational pressure is concentrated. This study proposes an Adaptive Graph WaveNet with Temporal Feature Fusion model, termed AGWN TF, for short‐term traffic flow forecasting and road operation risk assessment. The model extends Graph WaveNet by adding timestamp‐derived temporal context to its spatiotemporal forecasting backbone, including time of day, day of week, and calendar type. AGWN TF was evaluated on a 294‐sensor subnetwork extracted from the LargeST GBA dataset, using the previous 6 h of observations to predict traffic flow over the following 60 min. The forecasts were further converted into a Road Operation Risk Index by integrating flow pressure, short‐term flow change, prediction uncertainty represented by node‐level forecasting error, and spatial influence. AGWN TF achieved an MAE of 36.06, an RMSE of 55.41, and a MAPE of 18.36%, outperforming LSTM, STGCN, and Graph WaveNet. Compared with Graph WaveNet, the three errors were reduced by 2.99%, 1.89%, and 2.50%, respectively. Its advantage became clearer at the middle and later forecasting steps, while representative‐node analysis showed good reproduction of recurring peaks and overall traffic evolution. The RORI maps at 08:00, 12:00, and 18:00 revealed distinct spatial changes in operational pressure across the day, with more concentrated elevated‐risk patterns during the morning and evening periods. These results demonstrate that AGWN TF provides an integrated framework for accurate multistep forecasting and spatially interpretable road operation assessment, and traffic management resource allocation.