A hybrid clustering–regression approach for predicting passenger congestion in a carriage at a subway platform

This paper describes a novel integrated deep-learning framework that uses accessibility and time-varying patronage demand data to forecast the passenger congestion levels of individual carriages at a subway platform. The forecasting task involved the following challenges: (1) preprocessing spatiotemporal multivariate patronage data, (2) defining the effects of accessibility at platforms and time-series passenger demand on carriage congestion levels, and (3) designing an integrated deep-learning framework to manage heterogeneous spatiotemporal data. To address these challenges, an integrated deep-learning mechanism, namely a Conv-LSTM, was developed, which consisted of a convolutional neural network and long short-term memory (LSTM) framework to manage spatial and temporal features, respectively. Multidimensional datasets for testing and training the Conv-LSTM framework were collected from line one of the metropolitan subway systems in Busan, Korea. These datasets comprised (1) accessibility data corresponding to the entrance and exit locations at a subway platform relative to a carriage, (2) time-varying passenger demand data for a station, and (3) time-varying congestion data for a carriage. The performance of the Conv-LSTM framework was compared with those of other deep-learning approaches, namely a recurrent neural network, an LSTM, and a gated recurrent unit. The Conv-LSTM framework outperformed the other deep-learning approaches on the test dataset. This research can promote the application of deep-learning algorithms for addressing the challenges associated with handling spatiotemporal multivariate datasets and defining the relationships between congestion levels, accessibility, and passenger demand patterns for a platform in a subway station.