Method and apparatus for providing multivariable nonlinear control

A method for training an apparatus for controlling a process, the process having process inputs, process outputs comprising at least one controlled variable, and wherein the process is responsive to a manipulated variable used for changing the process as a function of the controlled variable, the method comprising the steps of:

• a) generating a model of the process to calculate predicted process outputs over a future time horizon, wherein the model is constructed using historical process data;
• b) applying the model to compute a change in the predicted process outputs that occurs as a result of a change made to the historical process data;
• c) determining a prediction time where an optimum change in the predicted process outputs occurs, the prediction time;
• being an optimum prediction time;
• d) computing a predicted process output advanced by the optimum prediction time;
• and
• e) training the apparatus with the predicted process output as advanced by the optimum prediction time and the historical process data to generate a corresponding variable for use in controlling the process.

A method for generating a nonlinear function generator for controlling a process having process inputs, process outputs comprising at least one controlled variable, and wherein the process is responsive to at least one manipulated variable for changing the process as a function of the controlled variable, the method comprising the steps of:

• a) collecting a plurality of historical values of manipulated variables and historical values of the process outputs generated from operations of the process and collected over a time span;
• b) pairing the manipulated variables with the process outputs based on the historical values in accord with an optimization criteria;
• c) retrieving training data generated from operations of the process and collected over the time span;
• d) calculating a plurality of future time steps representing a future time horizon;
• e) collecting a set of controlled variables as advanced by the future time horizon;
• f) generating a prediction model using training input data, the paired manipulated variables and the process outputs based on the historical values, the plurality of future time steps, and the controlled variables as advanced by the future time horizon to compute predicted controlled variables as advanced by the future time horizon;
• g) applying the prediction model to determine changes in the predicted controlled variables that occur as a result of changes made to the historical values of the manipulated variables, said application producing change variables;
• h) selecting from the change variables, a subset representing the greatest changes in value;
• i) storing the future time steps of the change variables in the subset produced in said selection step, the future time steps denoted as the selection times;
• j) determining a greatest change from the range of changes defined by the subset of change variables produced from said selection step;
• k) extrapolating the future times at which the greatest change occurs, the future times being the optimum prediction times;
• l) computing predicted controlled variables as advanced by the optimum prediction times;
• and
• m) training the nonlinear function generator with the predicted controlled variables as advanced by the optimum prediction times and the historical process data to compute corresponding variables for use in controlling the process.