A Novel Long Short Term Memory Network Based Train Dynamic Identification for Virtual Coupling

In recent years, virtual coupling has been proposed as a novel railway transportation method to achieve zero capacity waste with full reliability, flexibility and accessibility. To maintain platoons stable in virtual coupling, the speed of the preceding train must be predicted precisely by combing planned control outputs and train dynamic models. Unfortunately, hindered by complicated wheel-rail interaction mechanisms, accurate train dynamic is difficult to be modeled with conventional methods. This paper proposes a long short term memory network (LSTM) based train dynamic identification method. By considering non-ideal identification situations of value errors and lack of train operation data, two specific LSTM models are designed. The models are tested using the real train operation data extracted from the on-board equipment of Chengdu Metro Line NO. 8, China. Five evaluation indicators are selected to measure the accuracy of the models. The experimental results of different non-ideal conditions show improvements of the root mean square errors of train speed predictions at least 27% when using the proposed LSTM-based identification method.