A Hybrid STGCN-Transformer-Bootstrap Method for Short-Term Expressway Traffic Volume Interval Forecasting

Short-term traffic volume forecasting is a critical task supporting active traffic management strategies such as dynamic speed limits and ramp metering. To address the limitations of existing deep-learning models in modeling directional spatial propagation among cross-sections, suppressing noise during atypical operating periods, and quantifying point-prediction uncertainty, this study proposes an interval prediction method based on the STGCN-Transformer-Bootstrap model. A dual-stream parallel STGCN architecture is designed to decouple multi-horizon spatio-temporal features, and a gated cross-attention mechanism is introduced in the Transformer decoder to adaptively suppress the interference of irrelevant historical features. The Bootstrap method with 5 000 resampling times and the percentile method are then integrated to construct confidence intervals for the 5, 15 and 30 min prediction horizons. The method is validated on ETC gantry flow data from 26 cross-sections of the G15 Shenhai Expressway Quanzhou-Xiamen segment from June 1 to 15, 2022. At two typical cross-sections, compared with ARIMA and SVM, the RMSE, MAE, and MAPE of the proposed model decrease by 49.05%, 50.57% and 51.11% on average, with reductions ranging from 26.88% to 69.05%; compared with DCRNN, STGCN-LSTM, and GCN-Transformer, the three metrics decrease by 19.77%, 21.13% and 24.70% on average, with reductions ranging from 6.45% to 54.37%. Under the 95% confidence level, the percentile method achieves the best performance across the PICP, MPIW, and CWC metrics. Taking the 5 min prediction horizon as an example, MPIW is 59.24 pcu 5 min during stable free-flow periods, rises to 59.43 pcu/5 min during traffic phase transitions, and reaches 59.55 pcu/5 min during stable congestion, demonstrating adaptive interval widths that effectively characterize the uncertainty range under severe traffic flow fluctuations on expressways.