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.
Introduction
Traditional Model Predictive Control (MPC) has long been a standard for controlling complex systems while adhering to strict operational constraints. However, its efficacy is fundamentally tied to the accuracy of the underlying mathematical model, typically derived through labor-intensive system identification. In many real-world applications, building these models is not only time-consuming but also prone to errors that can destabilize the controller.
While data-driven MPC methods have emerged to bypass the modeling phase, they often struggle with two critical practical challenges: handling sudden changes in setpoints and managing unreachable targets. This paper introduces a novel data-driven tracking MPC scheme that addresses these issues by leveraging behavioral systems theory and the concept of "artificial equilibria."
Methodology
Willems’ Fundamental Lemma
The core of our approach lies in behavioral systems theory, specifically Willems’ Fundamental Lemma. This lemma proves that for linear time-invariant (LTI) systems, a single "persistently exciting" trajectory—a record of inputs and outputs that sufficiently exercises the system's dynamics—contains all the necessary information to predict any future behavior. By organizing this data into a Hankel matrix, we can predict future trajectories directly from raw data, effectively replacing traditional state-space equations ($Ax + Bu$).
Artificial Equilibrium and Tracking
To ensure the controller remains robust even when targets change or are physically impossible to reach, we introduce an artificial setpoint as a decision variable within the optimization problem. Instead of forcing the system to hit a fixed, potentially infeasible target, the controller optimizes two objectives simultaneously:
- Driving the system toward the artificial setpoint.
- Moving the artificial setpoint as close as possible to the actual target.
This "buffer" mechanism ensures recursive feasibility, meaning the optimization problem always has a valid solution, and allows the system to settle at the optimal reachable equilibrium if the true target is out of bounds.
Experimental Validation
We validated the proposed scheme using a simulation of a Four-Tank System, a classic benchmark in control engineering. The controller was initialized using a single input-output trajectory of length $N=100$, generated by applying random inputs. This short burst of data was the only information the controller possessed regarding the system's dynamics.
As shown in the figure above, the data-driven MPC successfully tracks the target output ($y^T = 1$) even when starting from an initial state far from the setpoint. The open-loop predictions (dashed lines) illustrate how the controller plans its path toward the artificial equilibrium at each iteration, eventually converging the closed-loop output to the desired target.
Key Insights and Results
- Guaranteed Exponential Stability: We provide mathematical proof that the closed-loop system is exponentially stable. The system reliably converges to the optimal reachable goal without requiring knowledge of internal system matrices.
- Semidefinite Stage Costs: A significant technical contribution is the extension of stability analysis to cases where the stage cost is only positive semidefinite. This is vital for input-output systems where internal states may not be fully observable.
- Expanded Region of Attraction: Compared to previous data-driven methods using fixed terminal constraints, our tracking version demonstrates a significantly larger region of attraction, allowing for recovery from a wider range of initial conditions with shorter prediction horizons.
Conclusion and Future Work
This work demonstrates that data is more than just a collection of points; for linear systems, it acts as the "DNA" of the system, enabling high-performance control without explicit modeling. While current results are restricted to LTI systems, future research will focus on extending these guarantees to nonlinear systems and improving robustness against significant measurement noise. Additionally, addressing the computational complexity of large Hankel matrices remains a priority for scaling this approach to higher-dimensional systems.
Summary updated