Evaluation of the photovoltaic power prediction performance of a neural network based on input data

Abstract

This paper aims to show the influence of the data size and the number of meteorological data used in the prediction of the output power of a photovoltaic installation with neural networks. We trained with different input data with the 2019a MATLAB Neural Network Start (NNS) tool, three feedforward networks. To train these networks, we used the algorithm of Levenberg-Marquardt and as data meteorological data such as wind speed at 10m from the ground, air temperature at 2m from the ground, position of the sun, direct radiation on an inclined plane and diffuse radiation on an inclined plane downloaded in the PVGIS database for a period from January 1, 2005 to December 31, 2016 for Natitingou city in the Republic of Benin. The first network was trained with wind speed at 10m, air temperature at 2m and sun position as input, the second network with wind speed at 10m, air temperature at 2m, sun position and direct radiation on an inclined plane and the third network with wind speed at 10m, air temperature at 2m, sun position, direct radiation on an inclined plane and diffuse radiation on an inclined plane. For the three networks we took the best results from 10 trainings. Thus, we obtained for the three networks respectively as mean square error 6186, 191 and 0.46 and as regression values 0.95, 0.998 and 0.999 respectively. In descending order according to the data used, the best performance was obtained with: • wind speed at 10m, air temperature at 2m, position of the sun, radiation, direct on an inclined plane and diffuse radiation on an inclined plane; • wind speed at 10m, air temperature at 2m, position of the sun, and the direct radiation on an inclined plane; • wind speed at 10m, air temperature at 2m and position of the sun.