High-Precision Model Predictive Control for High-Speed Trains Utilizing LSTM-Based Fixed-Lag Kalman Smoother With a Multi-Point Mass Model

This paper introduces a novel high-precision model predictive control method for high-speed trains that combines a fixed-lag Kalman smoother based on a long short-term memory network and a multi-point mass model. The long short-term memory network is trained on the state data that includes non-Gaussian noise to provide the necessary future observations for the Kalman smoother. A fixed-lag Kalman smoother is then developed, which utilizes the long short-term memory network's predictions to refine state estimation, effectively suppressing nonlinearity and non-Gaussian noise. The study's integration of a multi-point mass model offers a detailed description of the position and speed of each mass point within the high-speed train system, enhancing the accuracy of the control strategy. The model predictive control, relying on these smoothed state estimates, achieves high-precision control of the high-speed train. Validation experiments on a high-speed railway line in China show that the proposed method improves control accuracy by at least 27.58% over conventional filtering-based model predictive control methods.