Selecting parameters to estimate to obtain the best model predictions

A methodology is proposed to select parameters for estimation when data contain insufficient information to reliably estimate all model parameters. Parameters are ranked from most estimable to least estimable based on model structure, uncertainties in initial parameter guesses, measurement uncertainties and experimental settings. A new mean-squared error criterion is used to determine the optimal number of parameters to estimate from the ranked list so that the most reliable model predictions can be obtained. The methodology is illustrated using a dynamic chemical reactor model.