Use of Artificial Neural Networks for GHI Forecasting

Artificial Neural Networks are computational models inspired by the functioning of the central nervous system of animals. They are able to promote machine learning and pattern recognition with high quality. This study aimed to predict a step ahead of daily Global Horizontal Irradiation (GHI) using Artificial Neural Networks (ANRs) of two types: Nonlinear Autoregressive Network (NAR) and Nonlinear Autoregressive Network with Exogenous Inputs (NARX). The network structure was developed using the Matlab 2019b (R) NTSTOOL tool. The database used was provided by the network of the National Environmental Data Organization System (SONDA), covering the Brasilia station. Different models of neural networks were tested, contemplating varied inputs, and combinations of them. In addition, two training algorithms were tested, as well as several numbers of processors in the hidden layer, input delay and percentages for training, validation and testing of the network. The main point was to find the best ANR structure possible, which has a configuration capable of providing the lowest possible errors with good generalization. The results were compared with the naive and exponential damping models. A structure was obtained that provided 5.75% for the test MAPE error and 29,383.55 W/m(2) for the test RMSE, standing out in relation to the naive forecast (MAPE: 19.28%; RMSE: 67,920.73 W/m(2)) and that provided by the exponential damping model (MAPE: 9.36%; RMSE: 35,091.81 W/m(2)).