Selection of learning algorithms to improve energy prediction in a photovoltaic system

Abstract:
The integration of photovoltaic systems in the network requires that, in addition to issues related to the security and stability of the network, the generation of electricity is known with greater precision by the electrical system operators. This integration can be successful if the variation in production can be predicted with great accuracy. In this work, we compare different learning algorithms of a Multilayer Perceptron artificial neural network to predict the generation of a photovoltaic installation connected to the network. The learning algorithms used in this work are Levenberg-Marquardt, Conjugate Scaling Gradient and Bayesian Regularization. The implementation of the ANNs has been carried out using Matlab software (fitting), using recorded historical data of real power generated and other meteorological variables in a prediction horizon of 1 hour. Levenberg-Marquardt gives the best results.