Predicting Passenger Train Delays using Neural Network

One of the major issues in railways is passenger train delays and this issue in railway has many reasons, therefore predicting passenger train delay is a very difficult task. The aim of this paper is to present an artificial neural network based model with high accuracy to predict the delays of passenger trains in the Islamic Republic of Iran Railway. In the proposed model three different methods to define inputs, including normalized real number, binary coding and binary set encoding inputs have been used. To find an appropriate structure for proposed neural network model, three different strategies, called quick, dynamic and multiple are investigated. In this research, the registered data of passenger train delays in Iranian railway within the period (2004-2009) have been used. To eliminate any inconsistent and noisy data which always company with real world data set, a comprehensive preprocessing on this data set was done. To get more knowledge about data, some graphs such as seasonal average of delays, monthly average of delays, and total delays since 1383 to 1387 per year were sketched. To prevent models from over fitting with training data specifications, according to cross validation, the existing passenger train delays data set were divided into three subsets called training set, validation set and testing set, respectively. To evaluate the proposed model, the result of three different data input methods and three different structures were compared to each other and also to some common prediction methods such as decision tree and multinomial logistic regression. In comparison, different neural networks, training time, accuracy of neural network on testing data set and network size were considered and to compare neural networks with other well-known prediction methods, training time and accuracy of neural network on testing data set were considered and compared. To do a fair comparison among all models, a time-accuracy graph was sketched. The results revealed the higher accuracy of the proposed model.