(457i) Nonlinear Model Predictive Control with Performance-Index-Triggered Model Re-Identification

Chemical processes usually undergo changes during their operations. Those changes can be related to several different aspects, such as feedstock quality, incrustation, spent catalyst, and economic factors, among others. Process control approaches commonly consider that processes can operate inside a certain desired region that encapsulates the setpoints. However, changes in operating conditions may take the process outside of this region, bringing instability and performance degradation, mainly when one considers highly nonlinear processes, which are very common in industrial chemical and energy processes. Some control formulations, such as robust MPC, min-max MPC, and MPC for multiple scenarios [1], [2], [3], have been created to overcome this challenge.

In this work, an advanced control framework is proposed to minimize the impact of changes in control performance. This framework uses a nonlinear model predictive controller (NMPC) with a data-driven model that can be re-identified online by evaluating a control performance index. Specifically, a Gaussian process (GP) model is coupled to the MPC, and GP regression is employed to generate a set of hyperparameters for a given specific kernel function [4]. To activate the model re-identification, a trigger is defined, which is based on the integrated error between the overall process behavior and model prediction [2]. The system re-identification procedure generates a new set of hyperparameters given the new dataset. This dataset is defined by performing a steady-state detection analysis to select sufficient dynamic and steady-state data. After evaluating the performance of the new model, which is compared with the current and the original model, the new hyperparameters and dataset are replaced.

To demonstrate the proposed framework, a continuous stirred tank reactor (CSTR) is considered, where a first-order reaction occurs. The CSTR input, state, and output variables are stored using the Industry 4.0 infrastructure developed at WVU [5]. In this system, the main output is the concentration of A which is controlled to track a variable setpoint. Results of the case study will be shown, including the presence of a permanent disturbance to generate a prediction mismatch and trigger the re-identification. The augmentation of the Industry 4.0 infrastructure with online system identification and advanced control capabilities proposed here is expected to contribute to smart manufacturing approaches for industrial chemical processes of the future.

References

[1]. E. F. Camacho and C. Bordons, Model Predictive Control, 2nd ed. Springer, 2007, isbn: 978-1-85233-694-3.

[2] M. Maiworm, D. Limon, and R. Findeisen, “Online learning-based model predictive control with Gaussian process models and stability guarantees,” International Journal of Robust and Nonlinear Control, vol. 31, no. 18, pp. 8785–8812, Dec. 2021, doi: 10.1002/rnc.5361.

[3] D. Q. Mayne, J. B. Rawlings, C. V. Rao, and P. O. Scokaert, “Survey Paper Constrained model predictive control: Stability and optimality,” Automatica (Journal of IFAC), vol. 36, pp. 789–814, 2000.

[4] J. Kocijan, Advances in Industrial Control Modelling and Control of Dynamic Systems Using Gaussian Process Models. [Online]. Available: http://www.springer.com/series/1412

[5] D. Kestering, S. Agbleze, H. Bispo, and F. V. Lima, “Model predictive control of power plant cycling using Industry 4.0 infrastructure,” Digital Chemical Engineering, vol. 7, 2023, doi: 10.1016/j.dche.2023.100090.