A frequency-domain maximum likelihood estimation of synchronous machine high-order models using SSFR test data

The authors propose a numerical scheme for processing noisy signals originating from standstill frequency response (SSFR) tests on synchronous machines. Instead of using a univariate nonlinear least-squares procedure to fit only the weighted sum of magnitude responses, they minimize a multivariate prediction error criterion based on the determinant of the residuals covariance matrix. The algorithm pertains to a large class of prediction error methods and results in a multiresponse nonlinear regression procedure related to the maximum likelihood viewpoint when the residuals distribution is Gaussian. To demonstrate the efficiency of the proposed scheme, the implementation was tested using noisy simulated data, based on the Rockport model 3.3. It is shown, using actual data from the Nanticoke turbogenerator, dating back to the EPRI-project RP-9997-2 (1980), that the frequency-domain maximum likelihood approach can be effective for direct estimation of generalized circuits with up to five equivalent windings per axis, providing satisfactory predictions of both magnitude and phase as far as the 16th harmonic