Data-driven predictive control with reduced computational effort and improved performance using segmented trajectories

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

Abstract

Data-enabled Predictive Control (DeePC) provides a powerful model-free framework for automation by using historical input-output data to predict future system behavior. Despite its potential, standard DeePC formulations often struggle with measurement noise and the computational burden of long prediction horizons. This paper introduces a novel "segmented" trajectory approach that divides the long-term prediction horizon into smaller, interconnected segments. This structural innovation enforces causality, enhances robustness against stochastic disturbances, and shifts computational complexity from a non-linear to a linear growth pattern. Experimental results on a mechanical benchmark and a high-fidelity building energy simulation show that the segmented approach reduces tracking errors by over 30% and significantly improves real-time feasibility.

Introduction

Traditional Model Predictive Control (MPC) relies on explicit mathematical models, which can be difficult and costly to develop for complex systems. Data-driven methods like DeePC bypass this by using Willems’ fundamental lemma to project future trajectories directly from data. However, in practical applications involving noise, these methods often become "over-parameterized," leading to acausal behavior where future inputs appear to influence past outputs. Furthermore, as the prediction horizon increases—a necessity for systems with slow dynamics like buildings—the computational cost typically explodes.

Methodology: Segmented Trajectories

The core contribution of this work is the restructuring of the prediction trajectory. Instead of predicting a single long trajectory, the horizon is divided into $F$ shorter segments. The terminal state of one segment serves as the initialization for the subsequent one, creating a linked chain of predictions.

Comparison of (a) unsegmented and (b) segmented trajectory structures, illustrating how the horizon $N$ is partitioned into linked segments.

This segmentation enforces a lower-block triangular structure on the predictor matrix, physically re-enforcing the causality that noise often obscures. Additionally, the sequential linking results in a block-diagonal Hessian matrix within the optimization problem, enabling the use of efficient sparse quadratic programming solvers.

Experimental Evaluation

The researchers evaluated the segmented formulation using two distinct scenarios:

  1. Two-Mass-Spring-Damper System: A mechanical benchmark used to assess set-point tracking under 100 different noise realizations.
  2. Building Energy Management: A 102-state simulation of a six-room apartment using real-world weather data and occupancy profiles to test long-horizon (24-hour) performance.

Key Results and Insights

The segmented approach demonstrated superior performance across several metrics:

  • Robustness to Horizon Length: Unlike the standard formulation, where performance often degrades as the look-ahead window grows, the segmented version remains stable. In the two-mass system, it outperformed the standard version in 86% of long-horizon scenarios, reducing average errors by 32-36%.
  • Computational Efficiency: The transition from non-linear to linear computational scaling is a critical advantage. For a horizon of 100 samples, the segmented approach solved significantly faster, making it far more suitable for real-time industrial applications.
  • Building Performance: In the energy management case study, the segmented controller reduced occupant discomfort by 72% and energy costs by 5% compared to the unsegmented baseline.
Sum of set-point errors vs. prediction horizon, showing the stability of the segmented approach compared to the increasing error of the unsegmented method.

Conclusion

By decoupling the prediction horizon from the complexity of the underlying data relationship, the segmented trajectory approach addresses the primary bottlenecks of data-driven predictive control. Future work will focus on automated hyperparameter tuning and extending the framework to handle highly non-linear systems and persistent disturbances through offset-free control methods.

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 reduced computational effort and improved performance using segmented trajectories}},
  journal = {IEEE Trans. Control. Syst. Technol.},
  year = {2023},
  doi = {10.1109/tcst.2022.3224330},
}
Figure