Evaluation study of an efficient output feedback nonlinear model predictive control for temperature tracking in an industrial batch reactor

Zoltán K. Nagy, B. Mahn, Rüdiger Franke, Frank AllgöwerControl 2006 · DOI

TL;DR

This paper presents an efficient nonlinear model predictive control (NMPC) strategy for temperature tracking in an industrial batch polymerization reactor. By combining a computationally efficient optimization algorithm with a parameter-adaptive Kalman filter for state and parameter estimation, the authors demonstrate robust and accurate control, even with unobservable states and model uncertainties.

What Problem Does This Paper Address?

Batch processes, common in high-value chemical manufacturing, present unique control challenges due to their transient nature and the need to precisely control product quality at the end of a batch. While nonlinear model predictive control (NMPC) is well-suited for these complex systems, its computational demands and the difficulty in obtaining accurate, observable models have limited its industrial adoption, especially compared to continuous processes. This paper addresses the gap by developing and evaluating an efficient NMPC implementation that can handle real-time constraints, model uncertainties, and unobservable states in an industrial setting.

Methods

The core of the paper lies in its efficient NMPC implementation. Key methods include:

  • Efficient Optimization Algorithm: A multiple shooting technique is employed to solve the on-line optimization problem. This method breaks down the optimization into smaller, manageable subproblems, significantly reducing computational time and enabling real-time feasibility, even with long prediction horizons.
  • Grey-Box Modeling: To overcome the issue of unobservable states in the detailed first-principles model, a "grey-box" approach is used. This involves combining a reduced set of first-principles equations with empirical nonlinear functions derived from the detailed model. This hybrid model is designed to be observable with the available measurements.
  • Parameter Adaptive Extended Kalman Filter (PAEKF): To handle model uncertainties and plant-model mismatch, a PAEKF is used for state estimation. This filter simultaneously estimates the system states and a subset of the model parameters, allowing for on-line adaptation of the model. The process noise covariance matrix is also adaptively estimated, improving robustness.
  • Real-time Formulation: The NMPC implementation accounts for computational delay by injecting the control input from the previous optimization step while the current optimization is running.

Datasets

The study uses data from a pilot-scale industrial batch polymerization reactor at BASF. This includes:

  • Experimental Data: Data from actual batch runs on the pilot plant were used for model tuning, parameter estimation, and evaluating the performance of the NMPC controller.
  • Simulation Data: A detailed first-principles model of the reactor was used for simulation studies, including generating optimal temperature profiles and testing the NMPC implementation under various scenarios.

Key Insights and Ideas

  • Real-time NMPC is Achievable for Complex Batch Processes: The paper demonstrates that with efficient algorithms like multiple shooting and careful model development, NMPC can be implemented in real-time for challenging batch reactor control.
  • Grey-Box Modeling is Crucial for Observability: When direct measurements of all states are not available, combining first-principles knowledge with empirical relations is a practical way to create observable models suitable for NMPC.
  • On-line Model Adaptation is Essential: The PAEKF's ability to adapt model parameters in real-time significantly improves robustness and control performance, especially when dealing with variations in process conditions (e.g., heat transfer coefficients changing due to viscosity variations).
  • Computational Delay Management is Key: The strategy for handling computational delay by using the previous step's control input while the current optimization runs is vital for maintaining continuous control.
  • NMPC Offers Superior Performance but at a Cost: While NMPC significantly outperforms classical cascade PI and linear MPC in terms of tracking accuracy and robustness, its higher implementation complexity and computational cost may not always be economically justifiable for processes with less stringent quality requirements or lower temperature sensitivity.

Key Results

  • Robust Setpoint Tracking: The NMPC controller achieved tight temperature tracking, with deviations within ±0.5°C in experimental runs.
  • Improved Control Performance: Compared to cascade PI and linear MPC, NMPC showed significantly lower integral square errors (ISE) and maximum errors, particularly during periods of high heat generation.
  • Efficient Computation: The multiple shooting algorithm ensured that the computation time for the NMPC optimization was consistently below the sampling time (20s), even with a large number of control intervals (400).
  • Model Adaptation Benefits: The PAEKF successfully adapted to variations in heat transfer coefficients, demonstrating the importance of on-line model tuning.
  • Reduced Batch Time: Off-line optimization studies showed that the NMPC-derived optimal temperature profiles could reduce batch time by over 5% compared to a benchmark recipe.

Limitations and Future Work

  • Economic Justification: For this specific process, the authors conclude that the economic benefits of NMPC might not outweigh its implementation costs, suggesting that cascade PI control offers acceptable performance from a practical viewpoint.
  • Model Detail for Coagulation: The paper notes that to avoid issues like coagulation, even more detailed models incorporating the effect of coagulants and potentially additional sensors would be necessary. The NMPC could be extended to control these aspects.
  • Multi-variable Control: The authors suggest that the NMPC approach would be more advantageous for multi-variable processes or those where product quality is highly sensitive to temperature variations.
  • Further Model Refinement: The study highlights the need for accurate modeling, and future work could focus on further refining the control-relevant model, potentially by incorporating more detailed physics or using advanced identification techniques.

Summary updated

BibTeX
@inproceedings{nagy2006evaluation,
  author = {Zoltán K. Nagy and B. Mahn and Rüdiger Franke and Frank Allgöwer},
  title = {{Evaluation study of an efficient output feedback nonlinear model predictive control for temperature tracking in an industrial batch reactor}},
  booktitle = {Control},
  year = {2006},
  doi = {10.1016/j.conengprac.2006.05.004},
}
Figure