(574n) Development of Nonlinear Predictive Model-Based Feedforward Control Framework from Closed-Loop Freely-Existing Real Data

The control of chemical processes in industry is a very important aspect of everyday operations. The ability to maintain control has an impact on process safety, product quality and plant profitability. In recent years, many different advanced control techniques have been developed including the model-based control schemes, such as Smith predictors, feedforward controllers and model predictive control schemes. In order to use any of these model-based control schemes, several obstacles must be overcome. Many chemical and biological processes exhibit nonlinear behavior, but model-based control schemes have often used linear models, which can be sufficient if the process is operated over a small range of inputs. Real processes also often exhibit complicated dynamic responses to changes in the process inputs, including nonlinear dynamics. In the past decade, advances in computational capabilities have allowed for nonlinear models to be used for the process response prediction. Some of the types of models that have been proposed include Radial Basis Functions (RBFs), Artificial Neural Networks (ANNs), ARMAX and Block-Oriented Models (BOMs).

The performance of any model-based controller is highly dependent on the model that is used to predict process behavior. The procedures for developing the model can be time-consuming and costly, requiring the process to be perturbed in order to determine cause-and-effect behavior between the process inputs and outputs. It is desirable to be able to identify the process model without causing significant upsets to everyday operations. Ideally, historical data from the plant database could be used to develop the models to be used for prediction of process output response to changes in the inputs. The advantages of using this historical data are numerous. The data is readily available, is collected frequently, covers the “typical” operating space of the process, and does not require specific perturbations of the process inputs. However, several problems can be encountered if plant historical data is used. The process inputs are likely to be highly correlated, and the range of the inputs may not be very broad. For purely empirical models such as neural networks, this can be a significant shortcoming because the model cannot be used outside the input space that was used in the model identification procedure. The ability of the model to accurately predict behavior deteriorates if extrapolation occurs.

The purpose of this work is to demonstrate a method of developing a nonlinear process model under highly correlated inputs that can be used for a predictive model-based feedforward controller that will compensate for multiple input disturbances simultaneously. The model can be developed using historical plant data collected under closed-loop conditions and still effectively determine cause-effect behavior between the inputs and output of the process. In this work, we have present in detail a methodology for developing a Wiener block-oriented model from real (plant) data from a column that accurately predicts process response behavior to multiple input disturbances that are occurring simultaneously. The model was implemented into a predictive model-based feedforward/feedback control scheme and demonstrated marked improvements over traditional feedback control on a real distillation column.

This ability to develop the model with plant historical data under closed-loop conditions represents a significant advantage over traditional model-building techniques, which require specific perturbations of the process that can affect plant operations. This work can be extended to other types of chemical and biological process systems for further investigation. Specifically, the work done by Rollins et al. to predict glucose response in type 2 diabetics will be extended to close the loop on glucose concentration.