Feed Forward Curve Fitting Using a Process Dataset

1. Curve Fitting to a Dataset:

As a complement to my last whitepaper, Modeling And Scaling Feedforwards, I have
written this paper to discuss one of the approaches toward creating feedforward
models.

Many times it is not possible or practical to develop a mathematical relationship
between the individual contributors to a process and the resulting value (controlled
variable or intermediate variable) of interest. Using first principles predictive approaches
may not be practical or available.

Another way to determine contributors to the feed forwards or final process value is to
perform data-mining from the process itself. The following process exemplifies an
approach to creating polynomial or function generator parameters to create an
intermediate demand to anticipate changes in the process.

1.1. Collection of Data:

Select and collect all the data that may have an influence on the modeled variable,
even those process items that do not appear to have an obvious relation. The data
collection must also include the final controlled variable. When the data is obtained
under stable process conditions, the measured process disturbances are valuable to
understand and predict their impact on the final results. Modeled data is used as a
feed forward reflecting a static intermediate process variable that can be input as the
setpoint for the inner loop controller.

Data points should be averaged over a longer time frame rather than an
instantaneous value. Depending on the speed of the process, this time frame may
be a 5 minute average, 10 minute average, 30 minute average or longer.
For high speed processes or cycling processes it may be necessary to use the
instantaneous value. In that case, many cycles of that activity must be used when
generating the database.

1.2. Processing of Collected Data:

Collect these averages over as long a time as possible, reflecting various process
operating conditions. Place these data into a spreadsheet with the rows representing
synchronized times of the data records and the columns representing the averaged
data for each parameter. The following example shows an approach used to create
the functions needed to establish feedforwards or intermediate value setpoints.

Click here to download the complete document. You do not need to sign in to download this .pdf document.

Please contact Patrick Mahoney at patrick.mahoney@exceleng.net with questions.

 

Comments are closed.