Analysis of Spatio-temporal Heterogeneity Impact of Built Environment on Rail Transit Passenger Flow

It is of great significance to study the influence of various built environment characteristics on passenger flow for urban rail transit network planning and operational passenger flow control. This paper considers the influence of four types of built environment characteristics on rail transit passenger flow, including population economic characteristics, station characteristics, external traffic characteristics and land use characteristics. A hybrid model (GTWR-RF) is proposed, which combines the geographically and temporally weighted regression (GTWR) and random forest (RF). The model is used to capture the spatio-temporal heterogeneity and nonlinearity of the effects of built environment characteristics on passenger flow. First, the statistical indicators of built environment are refined and improved by collecting multi-source data. The GTWR was used to calculate the influence coefficient of built environment on rail transit passenger flow, and to analyze the spatio-temporal heterogeneity of the influence of built environment on passenger flow. Then, the influence coefficient is input into the RF model for training, to analyze the nonlinearity of the influence of the built environment on passenger flow. Using the GTWR-RF model, the study completed the passenger flow prediction and determined the mean relative importance of the built environment characteristics on passenger flow prediction. A case study in Beijing shows that the GTWR-RF model can describe both the spatio-temporal heterogeneity and nonlinearity of the impact of built environment characteristics on passenger flow. Of all the built environment features, the number of working population has the most significant influence on the forecast of passenger flow, followed by the number of bus connections. In the morning peak passenger flow forecast, the determination coefficient of GTWR-RF model is increased by 5.7% compared to the OLS method, increased by 6.3% compared to the RF method, increased by 0.5% compared to the GBRT method, increased by 10.1% compared to the XGBoost method, increased by 7.3% compared to the GTWR method. © 2023 Science Press. All rights reserved.