Multilayer Perceptron architecture optimization for peak load estimation

Abstract

Since the development of the Multilayer Perceptron, many types of artificial neural networks (ANNs) have emerged, each having best performances in solving particular types of problems. Current research developments focus on hybrid neural models, which combine neural and symbolic computation elements. In power engineering, ANNs are used today in a variety of applications, including optimization, approximation, forecast and classification tasks, for which an optimized ANN architecture is essential in obtaining the best results. Genetic Algorithms (GAs) can be used for identifying this architecture. While the general assumption when training a Multilayer Perceptron is that all neurons from one layer have the same activation function, this paper uses a genetic algorithm to search for the best mixed activation function configuration for the hidden layer, using as test bench a peak load estimation study.