Improving ANN Training with Approximation Techniques for ROCOF Trajectory Estimation

This paper proposes approximation techniques to improve the training process of artificial neural networks (ANN). The ANN model is trained to estimate the trajectory of the rate-of-change-of-frequency (ROCOF) in a multi-machine power system model using the swing equation. The approximation techniques deal with the following approaches. First of all, a simple equivalent model of a multi-machine power system is constructed for the ROCOF analysis. On top of that, the training data of the ROCOF trajectory is approximated to be a quadratic function of time. This quadratic function can be used until the ROCOF trajectory approaches the first zero point. These methods allow a more effective ANN training when the numbers of neurons and training epoch are constrained.