Gain-Scheduled Data-Enabled Predictive Control: A DeePC Approach for Nonlinear Systems

Margarita A. Guerrero, Braghadeesh Lakshminarayanan, Cristian R. RojasIEEE Control. Syst. Lett 2025 · DOI

TL;DR

This paper introduces DeePC-GS, a novel data-driven control strategy for nonlinear systems that adaptively switches between different "local" models based on a measurable scheduling variable. This approach overcomes limitations of existing data-driven methods by providing more accurate predictions and control for systems with varying operating conditions, demonstrated effectively on a ship steering benchmark.

What Problem Does This Paper Address?

Traditional model-based control relies on accurate system models, which can be difficult and costly to obtain, especially for complex nonlinear systems. Data-Enabled Predictive Control (DeePC) offers a model-free alternative by using input-output data directly to predict system behavior. However, standard DeePC is limited to linear time-invariant (LTI) systems. Nonlinear systems often exhibit behavior that changes significantly depending on their operating point (e.g., an aircraft's aerodynamics change with speed and altitude). Applying a single DeePC model to such systems leads to poor performance as the model becomes inaccurate. While some extensions exist for nonlinear systems, they often involve complex model lifting techniques. This paper addresses the challenge of applying data-driven MPC to nonlinear systems with varying operating conditions by leveraging the established concept of gain scheduling.

Methods

The core innovation is Gain-Scheduled Data-Enabled Predictive Control (DeePC-GS). Instead of using a single dataset to build one global DeePC predictor, DeePC-GS utilizes multiple local datasets, each tailored to a specific operating region of the nonlinear system. A measurable scheduling variable (like the surge speed of a ship) is used to determine which local dataset is most relevant at any given time.

Here's how it works:

  • Data Collection: Multiple input-output datasets are collected offline, each corresponding to a distinct operating region identified by the scheduling variable. These datasets are used to construct local Hankel matrices.
  • Region Selection: A selection map determines which operating region (and thus which dataset) is active based on the current value of the scheduling variable.
  • DeePC Optimization: At each control step, the DeePC optimization problem is solved using the Hankel matrices from the currently selected local dataset. This ensures the predictor is always based on data relevant to the system's current dynamics.
  • Switching Mechanisms:
    • Hard Switching: A hysteresis condition prevents rapid switching between datasets, ensuring stability. A temporal input blend smooths out control actions during a switch.
  • Soft Switching: Region-dependent controller parameters (like weighting matrices) are smoothly interpolated based on the scheduling variable's position relative to region boundaries. This further improves smoothness and robustness.

Datasets

The paper uses a nonlinear ship-steering benchmark for evaluation. This benchmark is based on the MARINER-hull cargo ship model, which exhibits strong regime-dependent dynamics influenced by surge speed. The authors collect data by simulating the ship's motion under various input conditions (rudder angle and propeller thrust) across different operating regions defined by surge speed. They also partition the data within each region into slices based on the trend of the regulated output (rising, constant, falling) to further improve conditioning and locality.

Key Insights and Ideas

  • Gain Scheduling for Data-Driven Control: The paper elegantly bridges classical gain scheduling with modern data-driven MPC. Instead of scheduling controller parameters, DeePC-GS schedules the data used to build the predictive model, which is a novel and powerful idea.
  • Locality is Key: For nonlinear systems, a single "global" model (or dataset) is often insufficient. DeePC-GS's strength lies in its ability to leverage local models that accurately capture the system's behavior within specific operating regimes.
  • Measurable Scheduling Variable is Crucial: The effectiveness of DeePC-GS hinges on having a readily measurable variable that correlates well with the system's changing dynamics. This variable acts as the "switch" to select the appropriate local data.
  • Computational Benefits of Slicing: Partitioning data within regions into slices (e.g., by output trend) significantly reduces the computational burden of the optimization problem by decreasing the number of Hankel columns needed, while maintaining locality.

Key Results

  • Superior Tracking Performance: In simulations, DeePC-GS demonstrated significantly better heading tracking accuracy compared to baselines like regularized DeePC, RoKDeePC (with various kernels), and Koopman-MPC, especially under disturbances.
  • Reduced Input Energy: DeePC-GS also achieved lower overall input energy consumption.
  • Tractable Computation: Despite using multiple datasets, DeePC-GS's computational footprint is manageable. By using active slices of data, it achieves substantial speedups (up to 275x) and memory reductions (42-75x) compared to a single global Hankel matrix approach, especially when considering the number of regions.
  • Robustness to Operating Point Changes: The system effectively handled transitions between operating regions, with only minor, transient dips in performance.

Limitations and Future Work

  • A Priori Knowledge: The current DeePC-GS formulation requires a priori knowledge of the operating regions and the scheduling variable. Future work could focus on learning these regions and the scheduling map directly from data.
  • Data Requirements: Collecting sufficient, persistently exciting data for each region can still be a challenge.
  • Switching Dynamics: While hard and soft switching mechanisms are proposed, further research into optimal switching strategies and their impact on stability and performance could be beneficial.
  • Noise and Uncertainty: While slack variables are used to handle noise, more advanced methods for robustifying DeePC-GS against significant model mismatch or unmodeled dynamics could be explored.

Summary updated

BibTeX
@article{guerrero2025gain,
  author = {Margarita A. Guerrero and Braghadeesh Lakshminarayanan and Cristian R. Rojas},
  title = {{Gain-Scheduled Data-Enabled Predictive Control: A DeePC Approach for Nonlinear Systems}},
  journal = {IEEE Control. Syst. Lett.},
  year = {2025},
  doi = {10.1109/lcsys.2025.3647981},
}
Figure