A cross-city knowledge integration framework for short-term metro passenger flow forecasting

Abstract

Accurate short-term forecasting of metro passenger flow is crucial for metro operation management and service planning. However, conventional prediction models are typically developed using data from a single city, which may limit their ability to capture diverse passenger flow patterns and reduce forecasting performance when historical data are insufficient. To address this issue, this study proposes METcross, a cross-city knowledge integration framework for short-term metro passenger flow forecasting. The proposed framework incorporates both static and dynamic covariates, including economic indicators, weather conditions, and passenger flow characteristics, to comprehensively represent station-level features. METcross consists of two stages. In the first stage, passenger flow patterns and feature representations are learned from a source city. In the second stage, feature embeddings from the source and target cities are integrated to enhance the representation of metro passenger flow characteristics and generate forecasting results for the target city. By incorporating knowledge from multiple cities, the framework can exploit shared mobility patterns while preserving city-specific characteristics. To evaluate the effectiveness of the proposed framework, experiments are conducted using metro passenger flow datasets from Wuxi and Chongqing. The results demonstrate that METcross consistently outperforms conventional single-city forecasting models. Compared with the No Fusion (NF) baseline, the proposed framework reduces the Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) by 22.35% and 26.18%, respectively. The findings indicate that cross-city knowledge integration can effectively improve short-term metro passenger flow forecasting accuracy and provide a practical solution for metro systems with limited historical observations.