Ten Minutes Solar Irradiation Forecasting on Inclined Plane using Evolutionary Product Unit Neural Networks

Abstract

This work applies evolutionary product unit neural networks (EPUNNs) to estimate global inclined irradiation at real time and predict it 10 minutes in advance. Both tasks are accomplished simultaneously, by using one single model with two outputs. One advantage of our approach is that the predictions of inclined irradiation are obtained without the need of a series of historical data. In this way, the model only considers one measured input variable, which is the horizontal global irradiation at the previous instant. Besides, the evolutionary algorithm used to optimize the network allows us to obtain the best adapted topology of the model with respect to the number of hidden neurons and synaptic connections. Very promising results are obtained, where the inclined irradiation Iβ(t) is estimated with an accuracy of 5.10% of nRMSE, while it is predicted 10 minutes in advance with an accuracy of 16.97%.