Linear Data-Driven Economic MPC with Generalized Terminal Constraint

Yifan Xie, Julian Berberich, Frank Allgöwer

TL;DR

This paper introduces a data-driven Economic Model Predictive Control (EMPC) scheme that optimizes the performance of unknown linear systems without requiring a mathematical model or prior knowledge of the optimal operating point. By combining the "Fundamental Lemma" for trajectory prediction with a generalized terminal constraint, the authors prove that the system can autonomously find and converge to an economically optimal steady state using only raw input-output data.

What Problem Does This Paper Address?

In traditional control, we usually want a system to track a specific setpoint (like keeping a room at exactly 72°F). Economic MPC (EMPC) is different: it aims to minimize a general cost—such as energy consumption or material waste—where the "best" operating point might not be known in advance.

Most EMPC strategies rely on two things that are often unavailable in practice:

  1. An accurate physics-based model of the system.
  2. Prior knowledge of the optimal steady state (the equilibrium that yields the lowest cost).

While "data-driven" methods have emerged to bypass the need for a model, they typically still require the user to provide the target optimal setpoint. This paper tackles the challenge of controlling a completely unknown system to achieve its best economic performance without knowing what that "best" performance looks like beforehand.

Methods

The authors build their framework on three main pillars:

  • Willems’ Fundamental Lemma: This is a cornerstone of modern data-driven control. It states that if you have a "persistently exciting" trajectory (data that explores the system's dynamics sufficiently), any future trajectory of that linear system can be represented as a linear combination of the collected data. This allows for prediction without an explicit $Ax + Bu$ state-space model.
  • Generalized Terminal Constraints: Instead of forcing the controller to end its prediction at a specific fixed point, the authors use a "generalized" constraint. This requires the system to reach some valid equilibrium (steady state) at the end of the horizon, but lets the optimization algorithm decide which one is best based on the economic cost.
  • Artificial Equilibrium: The terminal steady state is treated as a decision variable within the optimization problem. A penalty term ($\beta$) is added to the cost function to ensure the system stays close to this equilibrium, which helps guarantee stability.
  • Handling Unknown Costs: In a clever extension, the authors show that if the economic cost function is linear but its parameters are unknown, the Fundamental Lemma can be used to predict the cost values directly from data, treating the cost as if it were just another output of the system.

Datasets

The paper does not use a standard public dataset. Instead, it evaluates the approach using a numerical simulation of a linearized chemical reactor, a benchmark frequently used in the EMPC literature (originally from Faulwasser et al., 2018).

  • Data Generation: They generated an input-output trajectory of length $N=100$ by applying random inputs to the simulated reactor.
  • Setup: The system has 3 states, 1 input, and 2 outputs. The goal was to minimize a cost function related to the reactor's yield.

Key Insights and Ideas

  • Model-Free Optimality: The most significant "aha moment" is that you don't need to identify a model or solve for the optimal steady state offline. The controller performs both exploration (implicitly via data) and optimization (explicitly via the MPC cost) simultaneously.
  • The Power of Linearity: By treating the stage cost as an additional system output, the authors show that the same data-driven tools used for motion prediction can be used for "cost prediction." This effectively turns a learning problem into a geometric one.
  • Performance Guarantees: The authors provide mathematical proof that the "asymptotic average performance"—how the system performs over the long haul—can be made arbitrarily close to the theoretical optimum by simply increasing the penalty parameter $\beta$.

Key Results

  • Recursive Feasibility: The authors proved that if the optimization problem is solvable at the first time step, it will remain solvable for all future steps, ensuring the controller never "gets stuck."
  • Convergence: In the chemical reactor simulation, the data-driven controller successfully drove the system to the optimal stage cost of approximately -0.6396.
  • Robustness to Parameters: The simulation showed that the specific choice of the penalty weight ($\beta$) did not drastically change the final performance, suggesting the method is robust and easy to tune.
  • Comparison: Unlike previous data-driven MPC methods (e.g., Berberich et al., 2020), this scheme reached the optimal point without being told where that point was at the start.

Limitations and Future Work

  • Linearity Constraint: The current theory is strictly for Linear Time-Invariant (LTI) systems. Real-world chemical reactors and robots are often highly nonlinear, though they can sometimes be approximated as linear locally.
  • Noise Sensitivity: The proofs in this paper assume "clean" data. While the authors mention that their previous work handled noisy data for tracking, extending those guarantees to the economic case with generalized constraints is flagged as a future research direction.
  • Cost Function Shape: The "unknown cost" feature currently requires the cost function to be linear. Extending this to quadratic or general convex costs would make the tool much more versatile for complex engineering problems.

Summary updated

Abstract

In this paper, we propose a data-driven economic model predictive control (EMPC) scheme with generalized terminal constraint to control an unknown linear time-invariant system. Our scheme is based on the Fundamental Lemma to predict future system trajectories using a persistently exciting input-output trajectory. The control objective is to minimize an economic cost objective. By employing a generalized terminal constraint with artificial equilibrium, the scheme does not require prior knowledge of the optimal equilibrium. We prove that the asymptotic average performance of the closed-loop system can be made arbitrarily close to that of the optimal equilibrium. Moreover, we extend our results to the case of an unknown linear stage cost function, where the Fundamental Lemma is used to predict the stage cost directly. The effectiveness of the proposed scheme is shown by a numerical example.
BibTeX
@inproceedings{xie2023linear,
  author = {Yifan Xie and Julian Berberich and Frank Allgöwer},
  title = {{Linear Data-Driven Economic MPC with Generalized Terminal Constraint}},
  booktitle = {IFAC-PapersOnLine},
  year = {2023},
  doi = {10.1016/j.ifacol.2023.10.209},
}

Loading related content...

Figure