Forecasting Dynamic Public Transport Origin-Destination Matrices with Long-Short Term Memory Recurrent Neural Networks
Fiche du document
1 novembre 2016
- handle: 10670/1.zc133p
- hal: hal-01357641
Mots-clés
FORECASTING MOBILITY DEMAND MODELE PREDICTIF TRANSPORT EN COMMUN MATRICE ORIGINE DESTINATION ORIGINE DESTINATION ZONE URBAINE SMART CARDSujets proches
Prediction Forecasts FuturologyCiter ce document
Florian Toque et al., « Forecasting Dynamic Public Transport Origin-Destination Matrices with Long-Short Term Memory Recurrent Neural Networks », HAL-SHS : sociologie, ID : 10670/1.zc133p
Métriques
Partage / Export
Résumé
A considerable number of studies have been undertaken on using smart card data to analyse urban mobility. Most of these studies aim to identify recurrent passenger habits, reveal mobility patterns, reconstruct and predict passenger flows, etc. Forecasting mobility demand is a central problem for public transport authorities and operators alike. It is the first step to efficient allocation and optimisation of available resources. This paper explores an innovative approach to forecasting dynamic Origin-Destination (OD) matrices in a subway network using long Short-term Memory (LSTM) recurrent neural networks. A comparison with traditional approaches, such as calendar methodology or Vector Autoregression is conducted on a real smart card dataset issued from the public transport network of Rennes Metropole, France. The obtained results show that reliable short-term prediction (over a 15 minutes time horizon) of OD pairs can be achieved with the proposed approach. We also experiment with the effect of taking into account additional data about OD matrices of nearby transport systems (buses in this case) on the prediction accuracy.