Intelligent train stopping control for railways: A deep learning approach

Abstract

Station parking accuracy is an important indicator for the automatic control of high-speed trains; however, it is subject to many influencing factors, such as the characteristics of high nonlinearity and large time delays in the train control model, time-variant humidity, and uncertain weather conditions, leading to unsatisfying performance with existing feedback control algorithms. This study first proposes an intelligent train stopping control approach based on deep learning for high-speed railways. By collecting a large amount of historical train operation data from the Beijing–Shenyang high-speed railway, three data-driven models are developed for train stopping control. The first model is based on a deep-layered feedforward neural network (DNN), which predicts the exact train stopping position with dynamic train running states (position, velocity etc.) as input. Taking advantage of the physical train control models used in practice, the DNN to a convolutional neural network (CNN) is then improved by customizing the convolutional layers of the CNN. To overcome the issues arising from the incompleteness of data samples, a few-shot convolutional neural network (FSCNN) is further developed to enhance the prediction performance of the CNN. Compared with that of the current method used in practice, the simulation experiments show that the train station parking error can be decreased by 38.6%, 42.8%, and 49.2% by our developed DNN, CNN, and FSCNN, respectively.