SHORT-TERM TRAFFIC PREDICTION BASED ON A BI-GRU-ATTEN- ARIMA RESIDUAL FUSION MODEL

Intelligent Transportation Systems (ITS) mitigate traffic congestion through real-time planning and management, where short-term traffic forecasting is crucial. Because traffic time-series data are highly nonlinear, intricate, and history-dependent, we blend spatiotemporally correlated road-traffic datasets with exogenous factors such as holidays and major events. On this basis, we propose a residual-fusion model, Bi-GRU-Atten-ARIMA, which couples the nonlinear feature-learning capacity of a bidirectional gated recurrent unit (Bi-GRU) with an attention mechanism for adaptive feature weighting, while also exploiting the linear autocorrelation strengths of an ARIMA model. By jointly capturing nonlinear and linear patterns, the model significantly enhances forecasting accuracy. Two empirical studies on major Hong Kong region roadways—covering one-month and fifty-day datasets, respectively—validate its effectiveness, showing that the Bi-GRU-Atten-ARIMA residual-fusion model outperforms competing approaches in short-term urban-traffic prediction. Leveraging these precise forecasts, we further implement a congestion-warning module that quickly flags anomalous conditions within the traffic system.