Triggered System Re-Identification Using Gaussian Process Modeling for Model Predictive Control Applications

Controlling chemical and energy processes is crucial for safety, product quality, improved performance, and higher profit. However, operational issues can lead to control problems, especially when processes undergo changes during their operation. Control methods work best when processes operate close to their designed operating conditions, but instability or other issues may occur when the process is far from such conditions. To overcome these issues, the use of online model reidentification for advanced control purposes is proposed in this article. This involves reassessing the predictive model of an advanced controller, namely model predictive control (MPC), when reidentification conditions are met based on activating a defined trigger that considers the mismatch between model predictions and process values. To reidentify the process model, in this case based on a Gaussian process (GP) model, the size of the dynamic data set and its associated sampling time need to be established, in addition to the regressors of the model. A preassessment of the GP model is used to select the appropriate data set size and time step to satisfy dynamic transitions. The selection of the regressors, past inputs, and outputs required for a GP-based nonlinear autoregressive exogenous inputs (NARX) model is also performed. Additionally, steady-state detection (SSD) analysis for data classification and data set selection is carried out to ensure a minimum amount of dynamic data over steady-state data as part of the identification data set. To ensure improved performance, the new model is verified with a regressed data set and validated with new data before being used to replace the model within the MPC. A continuous stirred tank reactor (CSTR) example is considered to demonstrate the proposed framework. The CSTR input, state, and output variables are stored using a simulated Industry 4.0 infrastructure for communication, control, and model reidentification. In this reactor system, the main output is the concentration of component A, which is controlled to track a variable set point. The results of the case study show that a model update for a significant process change was required and successful to correct the model predictions and effectively control the process.