Short-Term Passenger Flow Forecasting of Beijing Subway Based on ARIMA and KNN Models

Abstract

Urban rail transit short-term passenger flow forecasting is critical for optimizing operation scheduling and improving service quality. Taking Beijing Subway as the research object, this paper compares the prediction performance of the Autoregressive Integrated Moving Average (ARIMA) model, the K-Nearest Neighbor (KNN) model, and their hybrid model. Based on 15-minute granularity passenger flow data from March to September 2024, combined with external factors such as weather and major events, the paper construct an ARIMA-KNN hybrid framework: first using ARIMA to fit the linear trend of passenger flow and obtain residual sequences, then applying KNN to correct residuals nonlinearly, and finally summing the two results to get the final prediction. Experimental results on the Guomao Station show that the hybrid model achieves the best performance, with RMSE reduced by 23.4% and 15.0% compared with ARIMA and KNN respectively, and MAPE of 12.7%. During peak hours, the hybrid model performs more prominently, with MAPE of 9.80% and 10.50% in morning and evening peaks respectively, which can better capture nonlinear fluctuations. This study provides technical support for refined operation scheduling and congestion early warning of Beijing Subway.