Data-Driven Tracking MPC for Changing Setpoints

Julian Berberich, Johannes Köhler, Matthias A. Müller, Frank Allgöwer

TL;DR

This paper introduces a data-driven Model Predictive Control (MPC) scheme that can track changing or unreachable setpoints without ever needing an explicit mathematical model of the system. By leveraging behavioral systems theory and optimizing an "artificial equilibrium" online, the controller guarantees stability and performance using only a single previously recorded trajectory of the system.

What Problem Does This Paper Address?

Traditional Model Predictive Control (MPC) is a powerful tool for controlling complex systems while respecting constraints (like speed limits or voltage caps). However, it relies heavily on having an accurate mathematical model (e.g., differential equations). Building these models via "system identification" is often difficult, time-consuming, and prone to errors that can lead to poor control.

While "data-driven" MPC methods have emerged to bypass model-building, they often struggle with two practical issues:

  1. Changing Targets: If the desired goal (setpoint) changes suddenly, the controller might become "infeasible," meaning it can't find a valid path to the new goal.
  2. Unreachable Targets: If a user asks the system to do something physically impossible (like a car trying to go 500 mph), standard controllers might fail or behave erratically.

Methods

The authors combine two powerful concepts to solve these issues:

  • Willems’ Fundamental Lemma: This is a cornerstone of behavioral systems theory. it proves that for linear systems, a single "persistently exciting" trajectory (a record of inputs and outputs that exercises the system's full range of motion) contains all the information needed to predict any future behavior. The authors use this data directly in a Hankel matrix to predict future trajectories, replacing the traditional $Ax + Bu$ equations.
  • Artificial Equilibrium: Instead of forcing the controller to hit a fixed target, the authors introduce an "artificial setpoint" as a decision variable in the optimization problem. The controller tries to do two things simultaneously:
  • Move the system toward the artificial setpoint.
  • Move the artificial setpoint as close as possible to the actual (potentially unreachable) target.

This "buffer" ensures that the optimization problem always has a solution (recursive feasibility) and allows the system to settle at the "best possible" reachable state if the true target is out of bounds.

Datasets

The paper does not use a large-scale public dataset. Instead, it validates the approach using a Four-Tank System simulation, a classic benchmark in control engineering.

  • Data Generation: They generated a single input-output trajectory of length $N=100$ by applying random inputs to the simulated tanks.
  • Application: This short burst of data was the only information the controller had about the system's dynamics before it began tracking a target output.

Key Insights and Ideas

  • Model-Free Stability: The biggest "aha" moment is the proof that you can guarantee exponential stability for an unknown system. The authors prove that the system will reliably converge to the optimal reachable goal without ever knowing the system's internal matrices.
  • Semidefinite Stage Costs: A major technical contribution is extending the stability analysis to cases where the "cost" of being off-track doesn't penalize every single internal state. This is crucial for input-output systems where we can't see or measure every internal variable.
  • The "DNA" of a System: The paper reinforces the idea that data isn't just a collection of points; for linear systems, a sufficiently rich trajectory acts like the "DNA" of the system, allowing us to reconstruct any possible future behavior through simple linear combinations.

Key Results

  • Guaranteed Performance: The authors mathematically proved that the closed-loop system is exponentially stable. If the target is reachable, the system hits it; if not, it stays at the closest possible stable point.
  • Numerical Success: In the four-tank simulation, the data-driven MPC successfully tracked a target output ($y^T = [1, 1]^\top$) even though the system was initially unstable without the controller.
  • Superiority over Fixed-Setpoint MPC: Compared to previous data-driven methods that used fixed terminal constraints, this "tracking" version has a much larger region of attraction. This means the controller can successfully recover the system from a wider variety of starting positions and handle much shorter prediction horizons.

Limitations and Future Work

  • Linearity Assumption: The current theory is strictly for Linear Time-Invariant (LTI) systems. While many real-world systems can be approximated as linear, highly nonlinear systems (like soft robotics or complex chemical reactions) would require further extensions.
  • Measurement Noise: While the authors mention noise, the core stability proofs in this specific paper assume relatively clean data. Developing "robust" versions that maintain these guarantees under heavy sensor noise is a primary direction for future research.
  • Computational Complexity: Using Hankel matrices of raw data can be more computationally "heavy" than using a compact mathematical model, especially as the amount of available data grows. Finding ways to compress this data without losing the "DNA" of the system is an ongoing challenge.

Summary updated

Abstract

We propose a data-driven tracking model predictive control (MPC) scheme to control unknown discrete-time linear time-invariant systems. The scheme uses a purely data-driven system parametrization to predict future trajectories based on behavioral systems theory. The control objective is tracking of a given input-output setpoint. We prove that this setpoint is exponentially stable for the closed loop of the proposed MPC, if it is reachable by the system dynamics and constraints. For an unreachable setpoint, our scheme guarantees closed-loop exponential stability of the optimal reachable equilibrium. Moreover, in case the system dynamics are known, the presented results extend the existing results for model-based setpoint tracking to the case where the stage cost is only positive semidefinite in the state. The effectiveness of the proposed approach is illustrated by means of a practical example.
BibTeX
@inproceedings{berberich2020data,
  author = {Julian Berberich and Johannes Köhler and Matthias A. Müller and Frank Allgöwer},
  title = {{Data-Driven Tracking MPC for Changing Setpoints}},
  booktitle = {IFAC-PapersOnLine},
  year = {2020},
  doi = {10.1016/j.ifacol.2020.12.389},
}

Loading related content...

Figure