Modeling the relation between the AVR setpoint and the terminal voltage of the generator using artificial neural networks

This paper deals with representing the relation between the voltage setpoint of the AVR (Automatic Voltage Regulation) system and the terminal voltage of the synchronous generator using artificial neural networks (ANNs). The training of the ANN is carried out using the Levenberg-Marquardt algorithm that minimizes the mean-square error between actual output data and estimated output. The experiments that serve for the training, as well as the validation of the ANN, are conducted on a real 120 MVA generator in the hydroelectric power plant Piva. Firstly, the ANNs with three different structures are trained and validated, and the optimal structure, i.e., the one that provides the lowest value of mean-square error, is adopted. Secondly, the adopted structure of the trained ANN is compared with other frequently used models, such as NARX (nonlinear autoregression model with exogenous input) and TF (transfer function) model. The validation data are obtained by carrying out experiments that comprise the different values of step disturbances on the AVR setpoint voltage signal. The presented results clearly show that the ANN model proposed in this paper ensures obtaining extremely accurate and precise results, that match almost perfectly with the corresponding experimental results. Comparative analysis proves that the ANN model is superior compared with the other two representative models. Finally, the test procedure required to obtain both training and validation datasets is very simple to conduct, mainly because it only includes adding step disturbance to the setpoint signal. Such a procedure does not have an impact on the normal operation mode of any component of the AVR system, nor does it require disconnection of the generator from the grid.