A Robust Spatio-Temporal Graph Attention Network for Traffic Flow Forecasting Under Sudden Disturbances

Abstract

The prediction of traffic flow has evolved from an empirical extrapolation method to a spatiotemporal correlation-based method due to the development of artificial intelligence and intelligent algorithms. In order to remedy the problem that the disturbance in the future causes the prediction error to increase sharply, we put forward the robust spatiotemporal graph attention network. Based on graph attention and gated temporal modeling, we provide a framework which brings together a road network topological encoding, a disturbance factor mapping, a temporal dependency extraction and a strong loss constraint mechanism. Experimental results show that under normal conditions, the model achieves an MAE of 13.42%, an RMSE of 20.74%, and a MAPE of 8.31%, outperforming STGCN’s 14.19%, 21.86%, and 8.64%, respectively; under disturbance conditions, the MAE is 17.96%, with a robustness index of 0.874, demonstrating good prediction stability.