PREDICTIVE SIMULATION OF THE ELECTRICAL PRODUCTION OF A PHOTOVOLTAIC INSTALLATION BY ARTIFICIAL NEURAL NETWORKS

Abstract

This paper explores the possibility of predicting the production of a prosumer by artificial neural networks. We made a cross comparison of two ANNs architectures (looped and unlooped) with respect to multivariable regression in order to have an efficient and reliable tool for predicting the production of a photovoltaic installation from meteorological data (solar irradiance and ambient temperature).To accomplish these goals, we used monitoring data of an installation over a period of 72 days to realize, train and test two ANNs topologies (looped and unlooped) which are trained with the Levenberg-Marquardt algorithm. After training and testing, it first appears that the neural networks show the best performance compared to the multivariable regression method. Then, among the two architectures, the one with the lowest uncertainties (𝑀𝐴𝑃𝐸=24.489% 𝑅𝑀𝑆𝐸=436.08 and MBE=30.93) is the feed-forward architecture.In this article, we have shown the importance of managing energy supply and demand, but also the main drawbacks of the current electricity network in Senegal.In summary, we have developed a model capable for predicting the production of a prosumer based on meteorological data (sunshine and ambient temperature).