Abnormal Passenger Flow Forecasting for Rail Transit Based on Multi-Source Big Data Fusion

Abstract

To address the problem of forecasting abnormal passenger flow at urban rail transit stations caused by large-scale events, holidays, and weather changes, this paper proposes a multi-source big data abnormal spatio-temporal prediction framework, MBDASPF. Based on historical entry and exit passenger flow, the model incorporates external disturbance factors such as day of week, holidays, weather, official event website information, and Weibo super topic data, and adopts a “branch-transform-fuse” strategy to extract the historical passenger flow correlation and the abnormal disturbance correlation, respectively. The historical branch employs GRU to capture recent proximity, daily periodicity, and weekly tendency, while the disturbance branch employs LSTM to model the influence of external events. Finally, through two-stage fusion and an attention-based output layer, multi-step forecasting for the next 10, 20, and 30 min are realized. Experiments on data from the Olympic Sports Center Station of Nanjing Rail Transit Line 10 show that MBDASPF outperforms ARIMA, SVR, BPNN, and GRU overall in terms of MAE, RMSE, and MAPE, and can more effectively characterize abnormal passenger flow fluctuations during large-scale events.