Solar Insolation Prediction through Deep Belief Network

Abstract

The continued development of computational tools offers the possibility to execute processes with the ability to carry out activities more efficiently, exactness and precision. Between these tools there is the neural architecture, Deep Belief Network (DBN), designed to collaborate in the development of prediction technics to find information that allows to study the behavior of the natural phenomena, such as the solar insolation. This paper presents the obtained results when using the DBN architecture for solar insolation prediction, simulated through the programming tool Visual Studio C#, showing the deep level that this architecture has, how it affects the number of layers and neurons per layer in the training and the results to predict the desired values in 2014, with errors close to 2% and faster to training, respect to errors obtained through conventional methods for neural training, which are about 5% and take long periods of training.

Keywords

• Accord.Net, Afforge.Net, Back propagation (BP), Contrastive Divergence (CD), Deep Belief Network (DBN), predicción de radiación solar, Restricted Boltzmann Machine (RBM), visual studio 2010 - C#.