Global Solar Radiation Prediction in Colombia Using a Backpropagation Neural Network Architecture

Abstract

The main source of renewable energy available in nature is solar radiation, which is the most promising resource to replace non-renewable energy sources and reduce gas emissions into the atmosphere since it allows various forms of capture and transformation through photovoltaic and photothermal systems. For an optimum use of solar energy, it is necessary to characterize and know the solar radiation at the level of the earth's surface, but this varies with time instantaneously, hourly, daily, and during seasons, with the latitude and with the local microclimates of the site. Therefore, a backpropagation artificial neural network (ANN) has been used to develop a mathematical model to predict solar radiation and the polycrystalline temperature, as a function of the ambient in the Colombian territory, specifically in the Atlantic coast. The network has been trained with 300 of the 381 data that constituted the matrix to obtain the RMSE that has been 0.164, with a network architecture composed of 10 layers and 5 neurons per layer. In addition, it has been used as a learning constant of 0.5 for each interconnection of the ANN. The increase in the number of hidden layers and the number of neurons increases the network performance, improving the prediction of the objective variable around 13% when using an architecture with five neurons per layer (NL), and 15 numbers of layers (L). In general, the results obtained have shown an acceptable performance of the artificial neural network in the estimation of solar radiation, but with certain possibilities of being improved.