Data-Driven Predictive Control With Improved Performance Using Segmented Trajectories

Edward O'Dwyer, Eric C. Kerrigan, Paola Falugi, Marta A. Zagorowska, Nilay Shah

TL;DR

This paper introduces a novel "segmented trajectory" approach to data-driven predictive control (DeePC). By breaking down the prediction horizon into smaller segments, the method significantly improves tracking accuracy in the presence of unmeasured disturbances and noise, especially for longer prediction horizons, while maintaining computational efficiency.

What Problem Does This Paper Address?

Data-driven predictive control (DeePC) methods offer a way to design controllers directly from data, bypassing the need for explicit system modeling. This is attractive for complex systems where modeling is difficult. However, existing DeePC methods can be sensitive to unmeasured disturbances and noise, leading to performance degradation, particularly when predicting far into the future (long prediction horizons). This sensitivity arises because the direct data-driven approach can implicitly learn acausal relationships between inputs and outputs, which are problematic in real-world systems. The paper aims to address this performance drop and sensitivity issue in DeePC, especially for longer prediction horizons.

Methods

The core contribution is the segmentation of the prediction trajectory. Instead of predicting the entire future trajectory as one long sequence, the prediction horizon (N) is divided into F segments, each of length Tini. This segmentation is inspired by Willems' fundamental lemma, which provides a theoretical basis for data-driven control.

The key idea is that each segment's prediction can be treated as an initialisation trajectory for the next segment. This structure helps to:

  • Reduce acausal dependencies: By breaking the problem into smaller, sequential segments, the method implicitly discourages the controller from relying on future inputs to explain current outputs, which is a common source of instability or poor performance in unsegmented DeePC.
  • Improve scalability: While the number of decision variables might increase with segmentation, the block-diagonal structure of the resulting optimization problem allows for efficient solving using sparse quadratic programming techniques. This leads to a more linear increase in computation time with horizon length compared to the nonlinear increase in the unsegmented approach.

The paper reformulates both the indirect (least-squares predictor) and direct (optimization-based) DeePC frameworks to incorporate this segmentation.

Datasets

The paper uses two main datasets for evaluation:

  1. Simulated Two-Mass-Spring-Damper System: This is a standard benchmark system used to illustrate the performance of the segmented vs. unsegmented DeePC under various simulated disturbance and noise scenarios. The system parameters (masses, spring constants, damping constants) are provided.
  2. Simulated Building Energy Management System: A detailed simulation of a six-room apartment was created using EnergyPlus software. This simulation incorporates realistic weather data (ambient temperature, solar irradiance) and occupancy profiles. The building model is represented as a 102-state, linear, state-space system.

Key Insights and Ideas

  • Segmentation is key for robustness: Breaking the prediction horizon into segments is crucial for improving DeePC performance in the face of unmeasured disturbances and noise. This is the primary takeaway.
  • Longer horizons become viable: The segmentation strategy makes longer prediction horizons practical and effective, whereas unsegmented DeePC typically suffers significant performance degradation.
  • Computational efficiency scales better: The segmented approach exhibits a more linear scaling of computation time with prediction horizon length, making it more suitable for applications requiring long-term predictions.
  • Implicit causality enforcement: The segmentation implicitly enforces a form of causality by structuring the problem sequentially, which is a significant advantage over unsegmented DeePC.
  • Real-world applicability demonstrated: The successful application to a building energy management problem highlights the practical benefits of the proposed method in a complex, real-world scenario.

Key Results

  • Improved Tracking Accuracy: In the two-mass system case study, the segmented formulation consistently outperformed the unsegmented version in terms of set-point tracking error, especially under disturbance and noise. For longer horizons (N=100), segmentation improved performance in approximately 85% of simulated cases with an average error reduction of over 30%.
  • Reduced Discomfort and Cost in Building Management: In the building energy management case study, for a one-day-ahead prediction horizon (N=95), the segmented approach reduced occupant discomfort by 72% and energy cost by 5% compared to the unsegmented approach.
  • Computational Time: For shorter prediction horizons, the segmented formulation had slightly higher computation time. However, for longer horizons (N ≥ 60), the segmented approach became faster, with its computation time scaling approximately linearly with horizon length, while the unsegmented approach scaled non-linearly.

Limitations and Future Work

  • Hyperparameter Tuning: The paper acknowledges that parameter selection (e.g., regularization weight $\lambda_g$, initialisation length $T_{ini}$) is not always trivial and suggests that automated tuning methods would be beneficial for real-world implementations.
  • Extensions to DeePC: The authors suggest further work on applying segmentation to other extensions of DeePC, such as formulations with robustness guarantees, time-varying parameters, and nonlinear systems.
  • Offset-Free Control: Investigating methods for offset-free control in the presence of disturbances would be a valuable addition.
  • Computational Efficiency: Further focus on computational efficiency and hyperparameter selection is needed to tailor algorithms for specific applications.

Summary updated

BibTeX
@article{odwyer2023data,
  author = {Edward O'Dwyer and Eric C. Kerrigan and Paola Falugi and Marta A. Zagorowska and Nilay Shah},
  title = {{Data-Driven Predictive Control With Improved Performance Using Segmented Trajectories}},
  journal = {IEEE Trans. Control. Syst. Technol.},
  year = {2023},
  doi = {10.1109/tcst.2022.3224330},
}
Figure