The model quality for a model predictive control (MPC) is critical for the control loop performance. Thus, assessing the effect of model–plant mismatch (MPM) is fundamental for performance assessment and monitoring the MPC. This paper proposes a method for evaluating model quality based on the investigation of closed-loop data and the nominal output sensitivity function, which facilitates the assessment procedure for the actual closed-loop performances. The effectiveness of the proposed method is illustrated by a multivariable case study, considering linear and nonlinear plants.

Systems with strong interactions among the variables are frequent in the chemical industry, and the use of model predictive control (MPC) is a standard tool in these scenarios. However, model assessment in this case is more complex when compared with fairly coupled systems, since the interactions make the system more sensitive to the model uncertainties. It means that, if the coupling is high, a small modeling error in a single variable could be spread to the entire system. As a result, all the controlled variables (CVs) of the MPC will have their performance deteriorated and the root of the model problem will not be evident. This paper presents a method of model assessment for highly coupled systems. This is an extension of the method proposed by Botelho et al. for model-plant mismatch evaluation in MPC applications, based on the use of the diagonal elements of the output sensitivity matrix. One of its advantages is that the method does not require previous knowledge regarding the systems coupling level. The effectiveness of the proposed method is illustrated by two case studies: a high-purity distillation column and the Shell heavy oil fractionator.

The longevity of each MPC application is strongly related to its performance maintenance. This work provides an overview of the methodologies available to fulfill this task, including a discussion about some special requirements of performance assessment methodologies for typical industrial MPC applications. The available methodologies were compared using these requirements. The best approaches were selected and compared to a new method proposed by the authors. These techniques have been applied in two case studies: the Shell benchmark process and the quadruple‐tank process. The results show that the control policy (setpoint, soft constraints, targets) followed in the MPC application should be the determining factor in selecting the methodology for performance assessment.

Poor model quality in model predictive controller (MPC) is often an important source of performance degradation. A key issue in MPC model assessment is to identify whether the bad performance comes from model–plant mismatches (MPM) or unmeasured disturbances (UD). This paper proposes a method for distinguishing between such degradation sources, where the main idea is to compare the statistical distribution of the estimated nominal outputs with the actual modeling error. The proposed approach relies on the assessment of three case studies: a simple SISO Linear MPC and two multivariable cases, where the linear controller is subject to a linear and nonlinear plant, respectively. Results show that the proposed method provides a good indicator of the model degradation source, even when both effects are present but one of them is dominant.

No abstract