Anomaly Prediction in Passenger Flow with Knowledge Transfer Method

Predicting anomaly in travel demand is a crucial finding from smart card data analytics. The output of these predictions is a significant contribution to planning sustainable public transport system and generating possible knowledge for transportation learning models. This paper investigates the anomaly effects of the surge in bus passengers demand and compare it with an increase in taxi demand. Indeed, both short-term and long-term demands reveal different patterns of passengers in uncertain situations. In pursuit of our goal, we estimated the similarity in stations by both selected and latent features where pre-trained knowledge are combined as an ensemble with different weights. We present Surge Prediction and Knowledge Transfer (SPKT) model that uses Seq2Seq method combined with Multi-source Transfer Learning method on travel patterns extracted from smart card data to classify source stations and target station. To illustrate the demands blueprint, we considered multiple source stations as input to the predictor, to develop a mechanism that bridges the knowledge transfer learning with the targeted stations. To exemplify our method, we use a case study of an event with passenger surge. From experiments, we found that transferring knowledge can make the surge prediction better compared to only limited training data for the target stations. The results have proved the effectiveness of surge predictions and knowledge transfer for learning models.