Constrained Model Identification - Use Knowledge of the Process to Get Better Models from Less Data

Accurate dynamic process response models are a key element in a successful refining MPC application. For most refinery processes, these response models are derived from data obtained during a specially designed plant test where the process independent variables is excited to cause movement in the MPC application dependent variables.

Typically, the derived responses are not suitable for use in a controller without applying some post identification modifications to ensure the resulting model represents the target process well. These modifications can include: 1) Enforcing Gain and dead time constraints on specific responses. 2) Enforcing collinear gain relationships between related responses such as a product analyser and an associated pressure compensated temperature. 3) Removal of inverse process responses where these are judged to be infeasible. 4) Development of variable transformations to better match the response data.

This paper describes how the above post identification modifications can be input as constraints to model identification phase. Approaching the problem in this way shortens the model development time and results in better models from less data. A side benefit is that future revamps of the controller are easier as the control engineers expertise is retained within the model identification package.