A systematic procedure for selecting identifiable parameter subsets for a given set of measured outputs is proposed. The aim is to only select those parameters which can be estimated uniquely from the dataset used. The proposed procedure consists of first selecting a reduced set of most sensitive parameters by sensitivity analysis and subsequently selecting identifiable parameter subsets using the Fisher information matrix.
For a particular set of outputs obtained from a typical calibration exercise at a carrousel-type nitrogen removal plant, parameter subsets ranging from two to eight parameters were selected by this procedure. The procedure proved successful as the parameter subsets thus selected could be estimated accurately from simulated data without and with noise as well as from real data.
However, the procedure is based on a property which is local in parameter space. Consequently, as an a priori assumption on the parameter vales has to be made at the start of the procedure, the selection results might be different from the results which would have been obtained by using the a posteriori parameter values. Hence, the sensitivity towards this a priori assumption was tested explicitly. For this purpose, the parameter space was sampled according to a Latin hypercube sampling scheme and the selection procedure was applied in all sampling points as if these were a priori estimates. From this extensive study it could be concluded that the results of the procedure were not too severely influenced by the a priori assumption on the parameter values. Therefore, the proposed procedure appears to be a powerful and practical tool for efficient and reliable model calibration.