An interpretable approach to passenger flow prediction and irregular passenger travel patterns understanding in metro system

Metro passenger flow prediction is an essential aspect of intelligent transportation systems. However, despite the emergence of deep learning technologies and the development of numerous models in a competitive manner to enhance prediction accuracy, interpretability often remains overlooked. This paper investigates the interpretability of passenger demand and irregular passenger behaviors within the metro system, considering both macro and micro-level perspectives. The first segment of the study focuses on predicting metro station passenger flow through an integration of k-means clustering and XGBoost models. This method utilizes historical passenger flow data, Points of Interest (POIs) information, transportation facility information, and external factors like weather. Moreover, Shapley Additive (SHAP) values and Accumulated Local Effects (ALE) are employed to interpret the model, identifying key factors that significantly impact station passenger flow, which serves as the foundation for constructing the RP dataset for micro-level analysis. The second segment advances the predictive capabilities of destination choice among passengers with irregular travel patterns using an enhanced Multinomial Logit (MNL) model. By incorporating residual networks and self-attention mechanisms into the traditional MNL framework, dubbed DELogit, the model offers improved expressiveness and interpretability. Focusing on the Hangzhou Metro Line 1, this study extracts the OD distribution across 12 stations, applying DELogit to predict destination choices, and uses elasticity analysis to visualize how external factors influence individual travel decision-making. The results reveal superior predictive accuracy and interpretability, with a 11 % increase in precision over MNL, underscoring the effectiveness of these advanced methodologies in understanding metro passenger behavior. © 2024 Elsevier Ltd