Distributionally robust two-stage model predictive control: adaptive constraint tightening with stability guarantee

TL;DR

This paper develops a distributionally robust two-stage MPC framework that adaptively tightens constraints under unknown, time-varying disturbances, ensuring stability and performance guarantees with a tractable real-time solution.

Contribution

It introduces a novel two-stage distributionally robust MPC scheme using Wasserstein ambiguity sets, enabling adaptive constraint tightening and stability guarantees under unknown disturbance distributions.

Findings

The proposed method achieves adaptive constraint tightening for unknown disturbances.

It guarantees recursive feasibility and stability in closed-loop control.

Numerical results demonstrate improved robustness and performance.

Abstract

Model Predictive Control (MPC) is widely recognized for its ability to explicitly handle system constraints. In practice, system states are often affected by disturbances with unknown distributions. While robust MPC guarantees constraint satisfaction under worst-case scenarios, it tends to be overly conservative. Stochastic MPC balances conservatism and performance but relies on precise knowledge of the disturbance distribution, which is often unavailable. To address this challenge, this paper introduces Distributionally Robust Optimization (DRO) into the MPC framework and proposes a novel Two-Stage Distributionally Robust MPC (TSDR-MPC) scheme. The key innovation lies in formulating constraint violation penalties as a second-stage optimization problem, which, combined with the first-stage quadratic cost, constitutes a two-stage distributionally robust program. This structure enables adaptive constraint tightening against disturbances with unknown time-varying means and covariances. Utilizing a Wasserstein ambiguity set, we derive a tractable reformulation via strong duality and develop a cutting-plane algorithm that converges in a finite number of iterations, suitable for real-time implementation. To ensure closed-loop stability even under non-zero mean disturbances, we introduce a terminal constraint applied solely to the nominal system; this constraint is proportional to the current state and independent of distributional uncertainty, thus preserving overall feasibility. We provide rigorous theoretical guarantees, including recursive feasibility, finite-time algorithm termination, and an asymptotic performance bound on the average closed-loop cost. Numerical simulations validate the adaptability and robustness of the proposed framework under various disturbance scenarios.