GRAY-BOX VERSUS BLACK BOX

[petrotrinsmerv]

Currently we are preparing to implement a new MPC controller on one of our process plants. Generally, the approach used is where the relationships between manipulated variables (MVs, also called 'handles), controlled variables (CVs, or targets) and disturbance variables (DVs) are defined based on available (or generated data) , identified and modelled, thereby characterising the overall process or unit op as a 'Black Box'. It has been suggested that there is an advantage to a 'Gray Box' approach where the relationships between CVs may also be identified and modelled. However, I haven't heard (yet) what possible advantages may be derived from this. I believe, with my limited experience, that a well built 'Black Box' model, especially used in conjunction with good soft-sensors should be more than sufficient for any but the most difficult of processes. Does anyone have another perspective on this?

 

[knji]

Is it always best to user a priori information about a process whenever possible. In the Gray Box approach, you will be using some prior knowledge about the process to come up model that can be used for controller synthesis. This model can serve for parameter identification.

Another advantage of the GB approach is the parameters to be identified, using some optimization technique, often have a physical meaning. Also, you can use the knowledge about the process to properly design and excitation signal.

In the Black Box approach, all you have input output data and assume no previous knowlege of the process. Selecting a model structure would be a major challenge for there are too many questions to answer: Such as the model order, the number of parameters etc. Also, selecting an excitation signal could also be tricky.....

Go with Gray Box.