Deep Learning Based Train Delay Prediction And Its Digital Twin Prototype System Design

The High-speed railway,as an important component of the national economy,has gradually developed into the main artery of public transportation due to advantages of high-speed,safety,comfort and convenience.China high-speed railway has entered the era of networked operation.The increasing complexity of the high-speed railway network and traffic density pose heavy challenges to the efficient and stable operation of trains.When emergencies such as extreme weather or equipment failure occur,it may lead to train emergency stop and large-scale trains’ delay,affecting the normal operation of the entire network and reduce passenger satisfaction,and even cause major economic losses and negative social impacts.It is of great theoretical and practical significance to quickly and accurately predict the temporal and spatial scope of delays to improve the ability of high-speed rail to deal with emergencies under network conditions.However,the high-speed railway is an operational system with a high safety level,and it is impossible to carry out delayed deduction-related experiments.The digital twin technology can realize the state consistency between the digital space and the physical space,observe the state of the physical system in real-time,and perform deductions in the digital space.It can effectively improve the real-time and efficiency of delay prediction.This paper proposes a train delay prediction method based on deep learning and builds a train delay prediction prototype system based on digital twin technology to improve the prediction performance.The main contents are as follows:Firstly,in view of accurately predicting the train delay time under emergencies,a train delay prediction method based on deep learning is proposed.Considering the diversity of delay causes,the heterogeneity of operation information and the uneven distribution of delay samples,the K-means algorithm is used to classify the delay information and the Gaussian noise based data augement method is designed to enhance the data amount,and a feature importance evaluation method based on XGBoost is constructed to learn and extract data features.Furthermore,a multi-modal neural network model based on Transformer and FCN is proposed to predict the train delay.By comparing with the baseline model,the prediction accuracy of the proposed algorithm is improved by more than 10%,which verifies the performance of the algorithm.Secondly,aiming at the problem of complicated coupling mechanism of the rail transit network and the difficulty in efficient delay prediction in a network scale,a synchronous prediction method for networked delay based on graph neural network is proposed.Considering delay propagation,node heterogeneity of the network and time delay of asynchronous prediction,a train-track event network model with train arrival and departure events as nodes and running relationships as edges is constructed.On this basis,a spatio-temporal graph neural network is proposed to mine the spatial relationships of nodes in the high-speed railway network,and then combine with the time series model to extract the spatio-temporal characteristics of the rail network,so as to achieve the goal of synchronous prediction of railway network delays in typical scenarios.Through simulation verification combined with actual data,the prediction accuracy is increased by 9%.Finally,a prototype system for train delay prediction based on the digital twin is designed.The design goals and requirements of the system are analyzed,and key parameters such as line information,train operation schedules and operating environment in the physical system are extracted to build a digital twin based on the B/S architecture and geographic information technology.Then a multi-layer architecture of the train delay prediction system is proposed from a digital-twin system perspective,and the coupling mechanism among different layers is analyzed.Web development,Django,database,and other technologies are further used to build a system prototype,and two functional modules of operation monitoring and delay prediction are built.The function of the system and the predicted results are tested and analyzed based on actual operation data.This thesis has 45 figures,12 tables and 102 references.