Abstract: A generalized least square method with special whitening filter GLS （SF） is provided in this paper. The parameters of its whitening filter are a part of model parameters, and in consequence the computational method is simplified. The GLS（SF） is superior to the traditional GLS in estimation accuracy because it gets rid of the error due to incorrect choice of the order of the whitening filter. A lot of computational results show that the convergency of GLS（SF） is also better thean that of GLS, therefore GLS （SF） is a GLS method but with simpler algorithm, more accuracy and better convergency. It is proved that the square sum of errors between model output and observed output data is minimum for the system in which the input noise is small and the output measurement noise appximates to white noise. The mathematical model obtained by this method is better than that of the traditional GLS. An example of building up dynamic mathematical model in dynamic calibration is given. The numerical results show that this method is feasible and the modeling effect is quite well.