Station-Specific Enhanced Single-Point Mass Dynamic Model for Urban Train Velocity Prediction Using LSTM Neural Networks

Accurate train velocity prediction is essential for Automatic Train Operation (ATO) systems in urban train systems, where modeling complex train dynamics under varying operation conditions and different track geometry remains challenging. This paper presented a station-specific enhanced single-point mass dynamic model based on the LSTM neural networks for directly predicting the train velocity. Data was prepared strategically by deriving the enhanced single-point formulation from the multi-point dynamics, while preserving critical track information. The 16 independent models were separately trained for a total of 16 track sections, which can enable specialized learning of unique characteristics of track geometry, addressing the reduced accuracy of the generalized models. Comprehensive evaluation demonstrated that this approach meets ATO requirements for prediction accuracy of train speed and cumulative distance tracking as assessed through regression evaluation metrics that quantify the error between the actual and predicted data.