D-Bi-LSTM: A double Bi-LSTM-based method for metro ridership prediction

Metro ridership forecasting is instrumental in the operation, scheduling, and safety of public transportation. As an effective method of capturing sequential relationships, Long Short-Term Memory Network (LSTM) has been broadly adopted for traffic forecasting. However, in metro networks, due to the highly regular nature of metro ridership, the spatial features aggregated from other stations still contain sequential information, which is ignored in existing studies. To this end, this paper proposes a novel model which aims to employ Double Bidirectional LSTM (D-Bi-LSTM) to capture the sequential correlations of both temporal and spatial features of metro ridership. Firstly, based on the regularity of metro ridership, different temporal patterns are integrated as the input features. Secondly, the similarity graph and the adjacency graph are utilized to capture the spatial features between metro stations. Finally, the sequential correlations of both temporal and spatial features of metro ridership are captured by employing Bi-LSTM in spatial and temporal directions, respectively. Experimental results conducted on two large metro datasets demonstrate that the proposed model has superiority compared to four widely recognized models.