PREDICTION OF INFREQUENTLY MEASURABLE QUANTITIES IN POORLY MODELED PROCESSES

Reliable prediction of process variables requires a good model of the process. For processes that are poorly known, the generic modelling ca pability of neural networks offers an attractive alternative. However, for satisfactory performance, the conventional implementations of neu ral networks require large sets of off-line data and on-line measureme nt of crucial variables, such as concentrations. Meeting each of these requirements is often infeasible in chemical processes. Use of a more efficient neural-network structure as well as incorporation of the av ailable poor model allows the network to require only a feasible numbe r of off-line data and no on-line measurement of the crucial variables . The strategy is demonstrated for accurate prediction of on-line unme asured concentrations in a simulated batch reactor, for which only a p oor physical model accounting for 70% of the reactant consumption and 30% of the reaction-heat generation, and only a limited number of impe rfect off-line concentration measurements from a few 'modelling' runs, are available. The proposed combination of a feedback neural network and the available partial process model satisfactorily solves this rea listic prediction problem, for which the previously available methods prove inadequate. The extended Kalman filter yields, even for the nomi nal operation range, significantly less accurate predictions, is not r obust to changes in the operation regime, and requires much more effor t to tune.