Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A state reduction approach for learning-based model predictive control for train rescheduling

Caio Fabio Oliveira da Silva, Xiaoyu Liu, Azita Dabiri, Bart De Schutter

TL;DR

The paper tackles real-time train rescheduling in large urban networks where mixed-integer optimization is computationally challenging. It introduces a multi-resolution state design that uses a reduced passenger-flow state for learning discrete decisions and a full state for MPC to compute continuous timings, combined with an ensemble of neural networks for robust inference. The approach yields dramatic reductions in optimality gaps (e.g., from ~7–8% to ~0.15–0.25%) and substantial memory savings (state dimensionality drop from 4002 to 748 with $N_s=4$, $H=9$), while maintaining fast online computation. This enables scalable, constraint-satisfying learning-based MPC for real-time train rescheduling, with future work exploring PCA and autoencoders for alternative reductions.

Abstract

This paper proposes a state reduction method for learning-based model predictive control (MPC) for train rescheduling in urban rail transit systems. The state reduction integrates into a control framework where the discrete decision variables are determined by a learning-based classifier and the continuous decision variables are computed by MPC. Herein, the state representation is designed separately for each component of the control framework. While a reduced state is employed for learning, a full state is used in MPC. Simulations on a large-scale train network highlight the effectiveness of the state reduction mechanism in improving the performance and reducing the memory usage.

A state reduction approach for learning-based model predictive control for train rescheduling

TL;DR

The paper tackles real-time train rescheduling in large urban networks where mixed-integer optimization is computationally challenging. It introduces a multi-resolution state design that uses a reduced passenger-flow state for learning discrete decisions and a full state for MPC to compute continuous timings, combined with an ensemble of neural networks for robust inference. The approach yields dramatic reductions in optimality gaps (e.g., from ~7–8% to ~0.15–0.25%) and substantial memory savings (state dimensionality drop from 4002 to 748 with , ), while maintaining fast online computation. This enables scalable, constraint-satisfying learning-based MPC for real-time train rescheduling, with future work exploring PCA and autoencoders for alternative reductions.

Abstract

This paper proposes a state reduction method for learning-based model predictive control (MPC) for train rescheduling in urban rail transit systems. The state reduction integrates into a control framework where the discrete decision variables are determined by a learning-based classifier and the continuous decision variables are computed by MPC. Herein, the state representation is designed separately for each component of the control framework. While a reduced state is employed for learning, a full state is used in MPC. Simulations on a large-scale train network highlight the effectiveness of the state reduction mechanism in improving the performance and reducing the memory usage.
Paper Structure (8 sections, 9 equations, 2 figures, 2 tables, 3 algorithms)

This paper contains 8 sections, 9 equations, 2 figures, 2 tables, 3 algorithms.

Table of Contents

  1. Introduction
  2. Problem description
  3. State Reduction Approach for Learning-based MPC
  4. Decoupling of the decision variables
  5. State reduction
  6. Training and inference
  7. Case study
  8. Conclusions

Figures (2)

  • Figure 1: A representation of the control framework with state reduction.
  • Figure 2: Three-line railway network in Beijing.