Robust speed prediction of high-speed trains based on improved echo state networks

The accurate and robust speed prediction of high-speed trains (HSTs) is a challenging task in the automatic train operation (ATO) because HSTs operate in an open-air situation with much noise and many uncertainties. This paper contributes to the development of robust speed prediction methods for the ATO of HSTs based on improved echo state networks (ESNs). Firstly, an adaptive temporal scale selection approach is introduced to improve the accuracy and efficiency of ESN modeling, due to the importance of proper temporal scale selection in relation to the sound prediction performance of ESNs. Also, a random weight scaling mechanism is employed to enhance the feasibility and robustness of the proposed method, as the learning ability of ESNs lies in the constrained random connection weights. Furthermore, several conditions for the stability of the closed-loop control system are given. Our experiment results demonstrate that the proposed method successfully achieves sound performance in terms of speed prediction accuracy, efficiency and robustness.