A CNN-LSTM Model for Short-Term Passenger Flow Forecast Considering the Built Environment in Urban Rail Transit Stations

The rapid expansion of urban rail transit necessitates accurate short-term passenger flow forecasts (STPFF) to optimize operation plans and enhance service quality. However, the existing STPFF methods do not fully consider the effect of built environment on passenger flow. In this regard, a convolutional long short-term memory neural network (CNN-LSTM) model incorporating built environment indicators has been proposed for accurately short-term passenger flow predictions. First, a system of built environment indicators (including 11 indicators), anchored in the 5Ds framework, is introduced to depict the characteristics of the built environments surrounding rail transit stations. Then, the random forest model (RF) is utilized to measure and rank the indicator importance. Finally, using historical passenger flow and key built environment indicators as input variables, a CNN-LSTM model for short-term passenger flow forecast is built. Taking Beijing city, China as an example for empirical research, the results show that CNN-LSTM model considering built environment can improve the accuracy of STPFF. Utilizing the top four key built environment indicators (the ratio of commercial land area, density of point of interest (POI) categories, and bus station density) as input variables can effectively reduce model computational complexity while concurrently enhancing predictive accuracy. The highest forecasting accuracy of the model is achieved at a time granularity of 5 min. This study can effectively support the operation and management of urban rail transit.