A Constraint-Tightening Approach to Nonlinear Stochastic Model Predictive Control for Systems under General Disturbances

Henning Schlüter, Frank Allgöwer

Abstract

This paper presents a nonlinear model predictive control (MPC) strategy for stochastic systems with general (state and input dependent) disturbances subject to chance constraints. Our approach uses an online computed stochastic tube to ensure stability, constraint satisfaction, and recursive feasibility in the presence of stochastic uncertainties. The shape of the tube and the constraint backoff are based on an offline computed incremental Lyapunov function (ILF).

Introduction

Model Predictive Control (MPC) is a cornerstone of modern control engineering, prized for its ability to handle system constraints directly. However, real-world applications often involve nonlinear dynamics and random disturbances. While Robust MPC accounts for worst-case scenarios, it frequently results in overly conservative performance. Stochastic MPC (SMPC) offers a more flexible alternative by employing chance constraints, allowing for a predefined probability of constraint violation.

The primary hurdle for Nonlinear SMPC is computational complexity. Existing methods often simplify noise to be purely additive or require solving high-dimensional optimization problems unsuitable for real-time applications. This work bridges that gap by providing a computationally efficient framework for nonlinear systems under general disturbances.

Methodology

The core of the proposed approach is a Constraint-Tightening mechanism that utilizes an online stochastic tube. Rather than calculating exact probability distributions, the method bounds uncertainty within a geometric tube.

Technical Components

  1. Incremental Lyapunov Function (ILF): Computed offline, the ILF defines the geometry of the stochastic tube by measuring the potential divergence between nominal and actual system trajectories.
  2. Online Stochastic Tube: The controller predicts a nominal path and constructs a tube around it that contains the true state with a specified probability. As shown in Figure 1, this tube evolves over the prediction horizon.
  3. Constraint Backoff: To guarantee safety, the controller enforces stricter limits on the nominal trajectory. This "tightening" ensures that even with stochastic deviations, the actual state remains within original safety boundaries.
  4. Deterministic Reformulation: By using the ILF to track uncertainty via a single scalar variable, the stochastic problem is transformed into a deterministic optimization. This allows the controller to maintain speeds comparable to standard, non-stochastic MPC.
Schematic of the stochastic tube propagation and constraint tightening over the prediction horizon, illustrating how the nominal path is constrained to ensure safety.

Experimental Evaluation

The approach was evaluated using a numerical simulation of a DC-DC converter regulation problem. This benchmark involves inherent instability and parameter uncertainty modeled as a truncated Gaussian distribution. The control objective was to regulate the converter while respecting a chance constraint on electric power: $P[|x_1x_2| \leq 2] \geq 0.8$.

Key Results and Discussion

The simulation results demonstrated that the proposed controller is both safe and efficient. In a 14,000-step simulation, the system satisfied the power constraint in 95% of the steps, comfortably exceeding the 80% requirement.

Phase plot of the DC-DC converter states ($x_1, x_2$) under the proposed SMPC. The black trajectories remain largely within the safety boundaries defined by the red curves.

Insights

  • General Disturbance Handling: Unlike methods limited to additive noise, this framework handles disturbances that vary with state and input, making it applicable to complex systems like power grids or robotics.
  • Computational Scalability: The method only adds a minimal number of states and inputs to the optimization problem, ensuring feasibility for high-frequency control loops.
  • Reduced Conservatism: While the 95% satisfaction rate indicates some residual conservatism, the method is significantly more efficient than traditional robust approximations or simple Lipschitz-based bounds.

Conclusion

This paper introduces a scalable Nonlinear SMPC framework that provides rigorous safety guarantees for systems with complex disturbances. By leveraging offline ILF computation and online scalar propagation, the method achieves a practical balance between safety and computational performance. Future work will focus on further reducing conservatism by refining disturbance bounds and exploring applications in highly non-convex systems.

Summary updated

Abstract

This paper presents a nonlinear model predictive control strategy for stochastic systems with general (state and input dependent) disturbances subject to chance constraints. Our approach uses an online computed stochastic tube to ensure stability, constraint satisfaction and recursive feasibility in the presence of stochastic uncertainties. The shape of the tube and the constraint backoff is based on an offline computed incremental Lyapunov function.
BibTeX
@inproceedings{schluter2020constraint,
  author = {Henning Schlüter and Frank Allgöwer},
  title = {{A Constraint-Tightening Approach to Nonlinear Stochastic Model Predictive Control for Systems under General Disturbances}},
  booktitle = {IFAC-PapersOnLine},
  year = {2020},
  doi = {10.1016/j.ifacol.2020.12.518},
}
Figure