Next Hour Frequency of Services Prediction for Rail Transit Network

Abstract

Travel time reliability is one of the important acceptance factors for passengers and transit agencies in the urban rail transit system. During incidents, rail network operators need to understand current and expected drops in the frequency of services (FoS) to arrange control plans to move the maximum number of travelers. In this paper, we have investigated three deep neural network (NN) architectures to predict the next hour FoS of train stations: (i) long short term memory (LSTM), (ii) convolutional NNs (CNNs), and (iii) convolutional LSTM neural networks (CNN-LSTM). The outcomes assist the train management team in evaluating impacts from minor and major incidents and optimizing their response plans against service disruptions. The algorithms have been implemented on spatially ordered input data sets to enforce graphical structures for the CNN and CNN-LSTM architectures that are Spatio-temporal models. Case studies on Sydney northern metropolitan train line network indicate that the CNN-LSTM architecture outperforms the other deep learning (DL) algorithms, and improved the performance of a typical multi-layer perceptron by up to 81%. Sydney Trains have deployed an online dashboard that uses the AI engine developed in this paper.