Fast Offset-Free Nonlinear Model Predictive Control Based on Moving Horizon Estimation

Rui Huang, Lorenz T. Biegler, Sachin C. PatwardhanIndustrial 2010 · DOI

TL;DR

This paper introduces a novel framework for robust control of complex industrial processes by combining Nonlinear Model Predictive Control (NMPC) with Moving Horizon Estimation (MHE). The key innovation is an "offset-free" approach that guarantees desired setpoint tracking even with plant-model mismatches, achieved by either estimating disturbances or directly estimating unknown parameters. Furthermore, by integrating "advanced step" techniques, the computational burden is significantly reduced, making real-time implementation feasible for large-scale systems.

What Problem Does This Paper Address?

Industrial processes, especially those in chemical plants, are often nonlinear and subject to uncertainties. These uncertainties, stemming from factors like sensor inaccuracies, unmodeled dynamics, or changing operating conditions, lead to "plant-model mismatches." A significant consequence of these mismatches is the inability of controllers to maintain the process outputs precisely at their desired setpoints, resulting in a steady-state "offset." Traditional NMPC, while powerful for handling nonlinearities, can struggle with these mismatches and also face computational challenges due to the complexity of solving optimization problems in real-time. This paper aims to address both issues: achieving robust, offset-free control in the presence of plant-model mismatches and reducing the computational time required for NMPC implementation.

Methods

The core of the proposed solution is an integrated framework of NMPC and MHE. MHE acts as a state estimator, reconstructing the system's internal states from available measurements. The paper presents two main variations for achieving offset-free control:

  1. State and Output Disturbances: In this approach, the MHE and NMPC models are based on the nominal plant dynamics. To compensate for unknown plant-model mismatches, the MHE estimates unmeasured state and output disturbances. These estimated disturbances are then incorporated into the NMPC's predictive model, allowing it to account for the mismatch and achieve offset-free behavior.
  2. State and Parameter Estimation: When the structure of the plant uncertainty is known (e.g., a specific parameter drifts), MHE is used to simultaneously estimate both the system states and the unknown uncertainty parameters. By adaptively updating these parameters online, the plant-model mismatch is directly removed, leading to offset-free control.

To tackle the computational burden, the authors incorporate "advanced step" (as-) techniques for both MHE and NMPC. These methods leverage the sensitivity of the optimization problem's solution with respect to its parameters. Instead of solving the full optimization problem at each time step, they use pre-computed information (like the KKT matrix) to rapidly approximate the solution. This significantly reduces the online computational delay, making the controller suitable for fast-acting systems.

Datasets

Not discussed in the paper. The paper uses simulated models of a Continuous Stirred Tank Reactor (CSTR) and a large-scale Air Separation Unit (ASU) for evaluation.

Key Insights and Ideas

  • Unified Framework for Robustness: Combining MHE with NMPC provides a powerful way to handle both state estimation and control, especially in the face of uncertainties.
  • Two Paths to Offset-Free Control: The paper elegantly demonstrates that offset-free behavior can be achieved either by explicitly estimating and compensating for disturbances or by directly identifying and correcting for known uncertainty parameters. This offers flexibility depending on the prior knowledge of the process.
  • The Power of "Advanced Step" Techniques: The integration of as-NMPC and as-MHE is crucial. It highlights that significant computational gains can be achieved by cleverly reusing information from previous optimization steps, rather than re-solving complex problems from scratch. This is a key enabler for real-time implementation of advanced control strategies.
  • Parameter Estimation for Faster Rejection: The simulation results suggest that when uncertainty structures are known, estimating parameters directly leads to faster rejection of plant-model mismatches compared to estimating disturbances.

Key Results

  • Guaranteed Offset-Free Behavior: The proposed framework is theoretically proven to achieve zero steady-state offset under certain conditions (reachable setpoint, stabilizing controller, etc.).
  • Significant Computational Speedup: The advanced step techniques reduce the online computational delay by approximately two orders of magnitude (e.g., from minutes to seconds for the ASU example), making real-time implementation feasible.
  • Effective Disturbance Rejection: Simulations on a CSTR and an ASU demonstrated that the proposed methods effectively reject plant-model mismatches and regulate outputs to their setpoints.
  • Comparable Performance to Ideal NMPC: The advanced step implementations achieve performance very close to hypothetical "ideal" NMPC (which assumes instantaneous computation), but with drastically reduced computational load.
  • Improved Robustness with Parameter Estimation: The parameter estimation approach showed better performance and faster rejection of mismatches in simulations.

Limitations and Future Work

  • Assumptions for Theoretical Guarantees: The theoretical proof of offset-free behavior relies on assumptions such as the reachability of the setpoint and the asymptotic stability of the controller. If these conditions are not met, the controller will minimize the steady-state output difference.
  • Tuning Parameter Sensitivity: While the authors mention that the ASU simulation was not overly sensitive to tuning parameters, tuning MHE and NMPC controllers can still be a complex task in practice.
  • Known Uncertainty Structure for Parameter Estimation: The parameter estimation approach requires prior knowledge of the uncertainty structure, which may not always be available.
  • Future Work: The paper implicitly suggests that further exploration of the trade-offs between the disturbance estimation and parameter estimation approaches, as well as more extensive real-world industrial validation, would be valuable. The authors also note that the framework can be extended to handle state and output noises, which is a natural next step.

Summary updated

Abstract

To deal with plant−model mismatches in control practice, this paper proposes two variations of an offset-free framework which integrates nonlinear model predictive control (NMPC) and moving horizon estimation (MHE). We prove that the proposed method achieves offset-free regulatory behavior, even in the presence of plant−model mismatches. If the plant uncertainty structure is known, the MHE can be tuned to estimate uncertainty parameters, to remove the plant−model mismatch online. In addition, we incorporate the advanced step NMPC (as-NMPC) and the advanced step MHE (as-MHE) strategies into the proposed method to reduce online computational delay. Finally, the proposed method is applied on a large scale air separation unit, and the steady state offset-free behavior is observed.
BibTeX
@inproceedings{huang2010fast,
  author = {Rui Huang and Lorenz T. Biegler and Sachin C. Patwardhan},
  title = {{Fast Offset-Free Nonlinear Model Predictive Control Based on Moving Horizon Estimation}},
  booktitle = {Industrial},
  year = {2010},
  doi = {10.1021/ie901945y},
}
Figure