Hybrid Approach Combining Modified Gravity Model and Deep Learning for Short-Term Forecasting of Metro Transit Passenger Flows

Abstract

Short-term forecasting of metro transit passenger flows is of great importance to the urban subway system in the various aspects of train and crew scheduling, congestion mitigation strategies, operational decision-making, and dynamic information provision. In this paper, a hybrid short-term forecasting approach is developed by combining the modified gravity model and deep learning models (e.g., convolutional neural networks [CNN] with auto-encoder). There are three components in this hybrid forecasting approach: (a) the modified gravity model that incorporates both the geographic information surrounding metro stations and station-level inflows/outflows as regression attributes; (b) the convolutional auto-encoder that tackles the sparsity issues of origin–destination (OD) matrices of passenger flows; and (c) the fusion of physical regression results and the decoder matrix, where the backpropagation algorithm is applied to tune the optimal fusion weight parameter matrix. The combination enables the proposed approach to achieve the trade-off between model interpretability and forecasting accuracy. The proposed approach is applied to the short-term forecasting of passenger flows for the metro transit network in Beijing, China. The experimental results show that the hybrid approach is promising and outperforms the benchmark models, for example, time-series models, long short-term memory, and CNN. The application demonstrates that the proposed hybrid short-term forecasting approach is suitable in both the station-level trip generation/attraction and the inter-station OD passenger flows.