Abstract
In this paper, examined is the modeling of power system dynamics through the direct approximation of power system input/output mapping by neural network (NN) instead of through the conventional method by differential equations. The NN used is of multilayer type with delayed signals, which is suitable for dealing with time series data, and it is trained by the error back propagation algorithm. Two sample systems are modeled by the NN; one is a numerical simulation model, and the other is an experimental system, both of which are a one-machine infinite-bus system. The input signal and the output signal to the NN are the reference value of the generator terminal voltage and the terminal voltage itself, respectively. The parameters in the learning algorithm are adjusted so that the training develop smoothly and converge in 30, 000 times for the numerical simulation model. Thus obtained values of parameters are used for the identification of the experimental system, and the development of training is evaluated. NNs with different input structure are trained for both sample systems, and high approximation accuracy was found achieved. The performance of the NN trained by the experimental system data is compared with that of the conventional differential equation model, and the possibility of the power system dynamics modeling by NN is shown.
