Rapid Algorithm for Generating and Selecting Optimal Metro Train Speed Curves Based on Alpha Zero and Expert Experience

Abstract

According to the current research status of urban rail transit’s fully automatic operation (FAO), the train driving speed curves are usually obtained through simulation and calculation. The train driving speed curves obtained by this method not only have low efficiency but also are not suitable for complex road conditions. Inspired by AlphaZero, a reinforcement learning algorithm that utilised vast amounts of artificial data to defeat AlphaGo, an AI Go program, this paper investigates and analyses methods for rapidly generating a large number of speed curves and selecting those with superior performance for train operation. Firstly, we use the powerful third-party library in Python as the basis, combined with the idea of AlphaZero, to produce artificial speed curves for metro train driving. Secondly, we set relevant parameters with reference to expert experience to quickly produce massive reasonable artificial speed curves. Thirdly, we analysed relevant indicators such as energy consumption, running time error and passenger comfort to select some speed curves with better comprehensive performance. Finally, through the many observations with different running distances and different speed limits, we found that the speed curves produced and selected by our algorithm are more productive, diverse and conducive to the research of train driving operation than the actual data from traditional manual driving and ATO (automatic train operation) system.

Keywords

• rail transit
• speed control
• data generation
• curve optimisation
• urban rail transit
• train speed