Koopman-based NMPC for Virtually Coupled Train Control System

Technical Summary: Koopman-based NMPC for Virtually Coupled Train Control System

Problem Statement
The paper addresses the control challenges inherent in virtually coupled train systems, where multiple train units operate cooperatively at short distances with consistent speeds via train-to-train (T2T) communication. While virtual coupling promises high-efficiency and high-density rail operations, it imposes stringent requirements on real-time information exchange, precise speed coordination, and strict safety assurance.

Existing control strategies face a trade-off:

• Linear MPC approaches often rely on local linearization, which may fail to capture the intrinsic nonlinearities of traction and braking dynamics, potentially compromising smooth operation and resource utilization.
• Nonlinear Model Predictive Control (NMPC) offers a unified framework for handling nonlinear dynamics and constraints but suffers from high computational costs due to nonconvex optimization, making real-time implementation difficult for safety-critical applications.
• Data-driven methods (e.g., neural networks, deep reinforcement learning) lack interpretability and struggle to enforce hard operational constraints reliably.

The core problem is to develop a control framework that simultaneously handles nonlinear train dynamics and complex safety constraints (specifically collision avoidance involving velocity-squared terms) while ensuring the computational efficiency required for real-time execution.

Methodology
The authors propose an analytical Koopman-based Nonlinear Model Predictive Control (K-NMPC) approach. The methodology proceeds as follows:

1. System Modeling: The train dynamics are modeled using Newton's laws, incorporating position, speed, control inputs (tractive/braking forces), running resistance, and additional resistance. Constraints include speed limits, control input bounds, jerk limits, and collision avoidance constraints (which contain nonlinear terms related to braking distances and reaction times).
2. Analytical Lifting: Instead of data-driven identification (e.g., EDMD), the authors employ an analytical lifting procedure. They define a set of Koopman observables that include the original states (position, speed) and the nonlinear terms appearing in the constraints and dynamics (specifically squared velocities, $v_i^2$).
   • The initial observable set is $\Phi_{in} = \{p_i, v_i, v_i^2\}$.
   • To ensure a complete basis for the dynamics, higher-order monomials ($v_i^n$ for $n \ge 3$) are theoretically included, though a reduced basis with a maximum polynomial degree $\bar{n}$ is used for practical dimensionality reduction.
3. Koopman Lifted System: The nonlinear system is transformed into an infinite-dimensional bilinear Koopman lifted system. This is approximated by a finite-dimensional Affine Parameter-Varying (APV) system.
4. Prediction and Optimization:
   • The APV model is discretized using a closed-form solution of a convolution integral.
   • Crucially, the authors freeze the affine parameters of the lifted predictor along the shifted predicted trajectory (based on the previous time step's solution).
   • This freezing transforms the online optimal control problem from a nonconvex nonlinear program into a Quadratic Program (QP).
5. Baseline Comparison: The proposed K-NMPC is benchmarked against a standard time-discrete NMPC scheme that solves the original nonlinear optimization problem directly using the CasADi framework and IPOPT solver.

Key Contributions

• Analytical Lifting for Safety Constraints: The paper demonstrates how to systematically lift nonlinear safety constraints (specifically collision avoidance involving $v^2$ terms) into the Koopman observable space analytically, avoiding the need for data-driven dictionary selection which can lead to poor interpretability or high dimensionality.
• Computational Efficiency via QP Formulation: By freezing the APV parameters along the trajectory, the method converts the inherently nonconvex NMPC problem into a convex QP. This significantly reduces the online computational burden compared to solving nonconvex NLPs.
• Unified Framework: The approach provides a unified framework that handles both nonlinear train dynamics and operational constraints without resorting to local linearization or black-box data-driven approximations.

Results
The proposed method was validated through numerical simulations using realistic operating data from the Beijing Metro Yizhuang Line (Tongji Nan to Jinhai Lu section) with a two-train formation.

• Tracking Performance: K-NMPC achieved tracking performance comparable to the baseline NMPC. Specifically, it reduced the average absolute spacing deviation by approximately 49% (from 0.237 m to 0.122 m) while maintaining nearly identical speed deviations and energy consumption indices.
• Computational Time: The K-NMPC demonstrated superior computational efficiency.
  • For prediction horizons ($N_p$) ranging from 6 to 20, the K-NMPC average computation time ranged from 5.4 ms to 17.0 ms, with a maximum below 23 ms.
  • In contrast, the baseline NMPC required 20.2 ms to 29.1 ms on average, with maximum times reaching 64–89 ms.
  • The K-NMPC reduced average computation time by 40–70% and maximum time by up to 85%.
• Constraint Satisfaction: Both controllers satisfied force, rate, and collision avoidance constraints, though K-NMPC exhibited smoother control variations after the initial transient.

Significance and Claims
The paper claims that the proposed K-NMPC approach offers a viable solution for the real-time implementation of virtually coupled train control systems. By transforming the nonlinear optimal control problem into an efficiently solvable quadratic program through analytical lifting and parameter freezing, the method reconciles the need for high-fidelity nonlinear modeling with the strict real-time requirements of safety-critical railway operations. The authors assert that the approach maintains control performance comparable to standard NMPC while offering substantial gains in computational speed, making it a strong candidate for practical deployment in high-density rail scenarios.