PROCESS MODELING AND OPTIMIZATION USING FOCUSED ATTENTION NEURAL NETWORKS

Neural networks have been shown to be very useful for modeling and opt imization of nonlinear and even chaotic processes. However, in using s tandard neural network approaches to modeling and optimization of proc esses in the presence of unmeasured disturbances, a dilemma arises bet ween achieving the accurate predictions needed for modeling and comput ing the correct gains required for optimization. As shown in this pape r, the Focused Attention Neural Network (FANN) provides a solution to this dilemma. Unmeasured disturbances are prevalent in process industr y plants and frequently have significant effects on process outputs. I n such cases, process outputs often cannot be accurately predicted fro m the independent process input variables alone. To enhance prediction accuracy, a common neural network modeling practice is to include oth er dependent process output variables as model inputs. The inclusion o f such variables almost invariably benefits prediction accuracy, and i s benign if the model is used for prediction alone. However, the proce ss gains, necessary for optimization, sensitivity analysis and other p rocess characterizations, are almost always incorrect in such models. We describe a neural network architecture, the FANN, which obtains acc uracy in both predictions and gains in the presence of unmeasured dist urbances. The FANN architecture uses dependent process variables to pe rform feed-forward estimation of unmeasured disturbances, and uses the se estimates together with the independent variables as model inputs. Process gains are then calculated correctly as a function of the estim ated disturbances and the independent variables. Steady-state optimiza tion solutions thus include compensation for unmeasured disturbances. The effectiveness of the FANN architecture is illustrated using a mode l of a process with two unmeasured disturbances and using a model of t he chaotic Belousov-Zhabotinski chemical reaction. (C) 1998 Elsevier S cience Ltd. All rights reserved.