Parameter ranking by orthogonalization—Applied to nonlinear mechanistic models

The paper addresses methods for parameter sensitivity analysis in a large, nonlinear, mechanistic model which is to be run in an on-line estimation scheme. The parameter sensitivity has been obtained by numeric approximation. The paper proposes and applies successive orthogonalization of the sensitivity derivative for parameter ranking. The method is easy to implement and the results are easily interpreted. Orthogonalization of the sensitivity matrix gives a triangular form of the squared sensitivity. The paper shows how the triangular form of the sensitivity derivative gives a particularly easy form of the variance contribution of individual parameters, provided the model error can be assumed Gaussian. This information has been used to decide how many parameters from the ranked set are to be selected for on-line estimation.