Vision Transformer and Inpainting based Approach for Short-term Forecasting of Passenger Loads on a Transit Metro Line. Focus on Explainability and Atypical Situation prediction
Résumé
This paper presents a deep learning approach for the real-time prediction of train passenger loads. We generate an image to represent the metro line data, wherein a pixel's coordinates determine a train departure, and its color denotes the train loads. A computer vision technique called inpainting is deployed to complete a real-time metro traffic image, which is equivalent to the prediction task. First, we introduce the transformation of the metro line data to an image, which takes into account all metro line constraints, e.g., irregular time sampling of trains. Second, we present the methodology to forecast the train loads based on two main deep learning approaches: U-Transformer and Channel Vision Transformer. These models offer a multi-step forecasting process over the metro line by extracting the visual features of the images. Third, we apply the proposed models to a real test case of the Paris metro line 9 to validate and benchmark our models against various classical and deep learning-based forecasting approaches. The results show that the proposed models outperform the existing models in forecasting. Fourth, we perform an In-depth analysis based on attention scores and latent spaces to interpret the performances of the proposed methods. Furthermore, we investigate the results of our models in seven atypical scenarios (e.g., strike, lockdown, and disruptions) to evaluate the robustness of the proposed approaches.
Origine | Fichiers produits par l'(les) auteur(s) |
---|