A Deep Learning Model Combining STL, GCN, and GRU for Urban Mass Transit Station Short-Term Passenger Flow Prediction

The passenger flow at urban rail transit (URT) stations is susceptible to random events, posing significant challenges to accurate prediction. This paper proposes a deep learning model based on Seasonal and Trend decomposition using Loess (STL), Graph Convolutional Neural Network (GCN), and Gated Recurrent Unit (GRU) to mitigate the influence of sample noise on the prediction task. By applying STL decomposition to attenuate sample noise, the proposed model leverages GCN and GRU to uncover deep spatiotemporal patterns within the samples and accurately forecast the future passenger flow at URT stations during peak periods. The model is validated using a real data set from the Hangzhou subway network and compared against non-deep learning models, GRU models, and GCN-GRU models. The results demonstrate that the STL-GCN-GRU model achieves the best prediction accuracy among baseline models in URT passenger flow forecasting, with MAE, MAPE, and R2 of 13.25%, 20.94%, and 93.22%, respectively.