Features Combination Selection for Metro Ridership Prediction

Accurate prediction of metro ridership is crucial for the operation, scheduling, and safety management of metro systems. While numerous deep learning models have been proposed to enhance forecasting accuracy, the critical role of input features has often been overlooked. As a fundamental component of any model, input features determine which historical information is utilized for learning. Without selecting key features, deep learning networks struggle to fully leverage their learning capabilities and achieve optimal forecasting outcomes. To investigate the impact of historical features on metro ridership prediction, this study constructs five distinct approaches for integrating historical information and develops a base model to capture the spatial-temporal correlations of passenger flow. Through experimental analysis using various types of metro data, the effects of different features on forecasting results are systematically evaluated. The findings reveal that input features significantly influence prediction performance, providing valuable insights for addressing metro ridership forecasting challenges.