Machine learning approach for study on subway passenger flow

Abstract We investigate regional features nearby the subway station using the clustering method called the funFEM and propose a two-step procedure to predict a subway passenger transport flow by incorporating the geographical information from the cluster analysis to functional time series prediction. A massive smart card transaction dataset is used to analyze the daily number of passengers for each station in Seoul Metro. First, we cluster the stations into six categories with respect to their patterns of passenger transport. Then, we forecast the daily number of passengers with respect to each cluster. By comparing our predicted results with the actual number of passengers, we demonstrate the predicted number of passengers based on the clustering results is more accurate in contrast to the result without considering the regional properties. The result from our data-driven approach can be applied to improve the subway service plan and relieve infectious diseases as we can reduce the congestion by controlling train intervals based on the passenger flow. Furthermore, the prediction result can be utilized to plan a ‘smart city’ which seeks shorter commuting time, comfortable ridership, and environmental sustainability.