Real-Time Traffic Management of High-Speed Trains: A Novel Approach Combining Optimization with Machine Learning

The real-time traffic control of high-speed trains (HSTs) is an important task in railway networks to reduce the delay time of trains in case of unexpected disruptions. Traditional approaches are based on mixed-integer programming (MILP) formulations to be solved by optimization methods such as branch-and-bound (B&B), which however require a very long time in real-world problems. Our study introduces a novel methodology that employs machine learning (ML) techniques to accelerate the optimization of MILPs. Specifically, our approach develops a feedforward network, which can be trained off-line, to predict the optimal MILP parameter configurations for the control of HSTs. Additionally, we develop a differential evolution (DE) algorithm, which fine-tunes these parameters in response to dynamic conditions. We conduct numerical experiments based on the real-world data of Beijing-Zhangjiakou high-speed railway (BZHR). The results demonstrate that our ML-accelerated approach significantly outperforms traditional optimization methods, improving the solution performance by 21.47%.