Adaptive Passenger Flow Prediction in Metro Systems: An Incremental Learning Framework for Evolving Patterns

Intelligent transportation management in metro systems requires accurate and real-time predictions of passenger flow demand. However, current research has mainly focused on static metro systems, overlooking the system’s long-term continuous expansion and evolution. This paper introduces the Incremental Multi Graph Seq2Seq Network (IMGSN), a novel framework tailored for metro passenger flow prediction that leverages incremental learning to address the challenges posed by pattern evolution. Unlike conventional models that rely on static data sets, IMGSN is designed to adaptively integrate new sequences while preserving crucial historical knowledge. To represent the metro systems, we employ a multigraph architecture enhancing the extraction of dynamic spatiotemporal dependencies. Our approach includes a pattern evolution detection mechanism to identify and train on key nodes within the system, thus minimizing the need for retraining the entire model. Additionally, to mitigate catastrophic forgetting, we use memory-aware synapses (MAS) to evaluate the importance of model parameters, applying regularization to safeguard essential previous parameters during updates. We validated our model using a comprehensive data sets from the Nanjing Metro in China. Our experimental results show that IMGSN outperforms nonincremental learning models by an average of 45.69% at the 30- and 60-min time steps and surpasses the latest baselines by an average of 29.39%. These results demonstrate the necessity of designing deep learning frameworks with incremental learning capabilities for metro systems. The efficiency and accuracy of IMGSN make it well-suited to meet the task requirements in this context.