A deep learning approach for multi-attribute data: A study of train delay prediction in railway systems

Abstract

Dynamical systems that contain moving objects generate multi-attribute data, including static, time-series, and spatiotemporal formats. The diversity of the data formats creates challenges for the accurate modeling of these systems, for example, the state/location/trajectory prediction of moving objects. We developed a deep learning (DL) approach that combines 3-dimensional convolutional neural networks (3D CNN), long short-term memory (LSTM) recurrent neural network, and fully-connected neural network (FCNN) architectures to address this problem. The proposed model, named CLF-Net, uses individual factors with different attributes as input to achieve better predictions. The spatiotemporal features are fed into the 3D CNN, the time-series variables are fed into the LSTM, and the non-time-series factors are fed into the FCNN, respectively. A case study of train delay prediction for four railway lines with different operational features shows that the CLF-Net outperforms conventional machine learning models and the state-of-the-art DL models with regard to the performance metrics of the root mean squared error and mean absolute error.