Title:METcross: A framework for short-term forecasting of cross-city metro passenger flow

Abstract:Metro operation management relies on accurate predictions of passenger flow in the future. This study begins by integrating cross-city (including source and target city) knowledge and developing a short-term passenger flow prediction framework (METcross) for the metro. Firstly, we propose a basic framework for modeling cross-city metro passenger flow prediction from the perspectives of data fusion and transfer learning. Secondly, METcross framework is designed to use both static and dynamic covariates as inputs, including economy and weather, that help characterize station passenger flow features. This framework consists of two steps: pre-training on the source city and fine-tuning on the target city. During pre-training, data from the source city trains the feature extraction and passenger flow prediction models. Fine-tuning on the target city involves using the source city's trained model as the initial parameter and fusing the feature embeddings of both cities to obtain the passenger flow prediction results. Finally, we tested the basic prediction framework and METcross framework on the metro networks of Wuxi and Chongqing to experimentally analyze their efficacy. Results indicate that the METcross framework performs better than the basic framework and can reduce the Mean Absolute Error and Root Mean Squared Error by 22.35% and 26.18%, respectively, compared to single-city prediction models.