Neuro-Evolution of Frequency Controllers in the Power-Electricity System Island

Abstract

We present a methodology for designing frequency controllers in a power-electricity isolated island that is based on artificial neural networks (ANN). A less commonly employed machine learning approach based on neuro-evolution (NE) is proposed for learning the ANN, demonstrating its efficiency in dynamic system control. The power-electricity island consists of four water turbogenerators, two steam-gas cycle units, one nuclear unit, a distribution network, and consumers. The proposed architecture of neural controllers and the NE-based learning method, using a genetic algorithm, is described. The results are compared with commonly used PI controllers using different criteria functions. The comparison shows that the proposed approach achieves significantly better control performance compared to the conventional approach using PI controllers.