Forecasting for Short-term Passenger Flow of Subway Based on Dynamic Graph Neural Ordinary Differential Equations

With the rapid expansion of urban rail transit networks, accurate forecasting for passenger flows has become paramount for optimizing operational services. To solve the issue of the inadequate mining for the spatiotemporal characteristics in the forecasting of current subway passenger flow forecasting and to further enhance accuracy and efficiency of forecasting methods, a forecasting method for short-term subway passenger flow based on multivariate time series with dynamic graph neural ordinary differential equations (MTGODE) is proposed. The method constructs a dynamic topological graph structure by learning the dynamic correlation strength between subway stations. Continuous graph propagation is performed on the learned dynamic graph to transmit spatiotemporal information and capture the dependencies of passenger flows. Moreover, residual convolution is employed to extract periodic patterns at multiple time scales, enabling continuous representation of spatiotemporal dynamics between stations and overcoming the limitations of traditional graph convolutional network models in capturing dynamic spatial dependencies. Furthermore, to fully uncover the spatiotemporal patterns of passenger flow distribution among different stations, a multi-source fusion model for passenger flow forecasting is developed by comprehensively utilizing data from the Beijing subway's automatic fare collection system, weather data, air quality data, and surrounding land use attributes of stations. The proposed model was tested by forecasting inbound passenger flow and origin-destination flow using historical data from Beijing Station and Jishuitan Station-Dongzhimen Station. The experimental results demonstrate that the proposed model achieves superior performance compared to multiple benchmark models across three metrics: mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE). Compared to the best-performing benchmark model, the diffusion convolutional recurrent neural network (DCRNN), the proposed model reduces MAE, RMSE, and MAPE by 9.93%, 12.30%, and 9.23%, respectively. It exhibits a better fit to the spatiotemporal distribution of subway passenger flows and possesses improved prediction accuracy, stability, and fitting capability.