An Improved Framework for Origin-Destination Prediction in Urban Rail Transit Based on Stack LSTM

Generally, the accuracy of passenger flow prediction in urban rail transit (URT) network is low due to the notable differences among different Origin-Destination (O-D) pairs. To improve the prediction performance of O-D flows in URT network, an improved framework which combines k-means clustering and stack long short-time memory (LSTM) is proposed (CS-LSTM). Firstly, the Silhouette Coefficient is calculated based on four indices, which is applied to determine 4 categories of O-D pairs. Secondly, in order to reduce the negative influence of random factors, the input data structure is adjusted by setting dynamic minimum thresholds of passenger flow. Thirdly, the parallel stack LSTM network is trained by utilizing the advantage of LSTM in predicting variable-length sequence data, and the two output sequences are re-constructed as ultimate outcome. Finally, a case study is conducted by using the data from Chengdu URT, where the proposed method is verified for different categories of O-D pairs, and the prediction error are compared between the proposed method and other 7 state-of-the-art methods. The result indicates that the proposed method has better prediction performance, especially in capturing the trends of passenger flow. The authors also compare the prediction performance under different time granularities with 15min, 30min, 60min interval. The results show that when the time granularity increases, the prediction error of O-D pairs with noticeable morning and evening peaks decrease significantly, and the prediction accuracy could be further improved through the fusion of different time granular data.


  • English

Media Info

  • Media Type: Web
  • Features: Figures; References; Tables;
  • Pagination: 20p

Subject/Index Terms

Filing Info

  • Accession Number: 01764052
  • Record Type: Publication
  • Report/Paper Numbers: TRBAM-21-00450
  • Files: TRIS, TRB, ATRI
  • Created Date: Dec 23 2020 11:18AM