Development and implementation of neural network observers to estimate thestate vector of a synchronous generator from on-line operating data

This paper presents a novel technique for developing and implementing artif icial neural network (ANN) observers for estimating un-measurable rotor bod y currents of a synchronous generator from time-domain on-line disturbance data. Data for training the observers are generated through off-line simula tions of a 7.5 kVA machine model whose parameters are varied in accordance with previously determined on-line parameter estimates of the generator und er consideration. Studies show that observer robustness towards noise can b e improved by enhancing the size of the observer input vector. In order to increase observer robustness towards variations in the field-resistance, si mulated variations representative of changes in field-resistance were intro duced in the training sets. After training, the observers are tested with e xperimentally obtained on-line measurements to provide estimates of un-meas urable rotor body currents. The estimated rotor body currents are then used along with experimental measurements to estimate synchronous generator par ameters.