Previsão de geração de energia fotovoltaica utilizando método de aprendizado de máquina

Abstract

Due to the rise of photovoltaic generation together with the growth of network connections, different methods and algorithms, which are able to predict this generation in many time horizons, have been used by researchers. The main algorithm adopts the machine learning using artificial neural networks. This method is used because of its capability of working with non-linear elements. This work has the objective of developing a photovoltaic power generation forecasting in Garibaldi, RS, with a time horizon of six months, creating an artificial neural network. The input data of this network is composed of climatic historical data sets in addition to a set of historical solar photovoltaic generation data, which will be used as targets. The input variable classification is performed through stepwise regression, which lists the highest relation between input variables and the response. The modeling is accomplished by the numerical software Matlab. The software develops the algorithm and ranks the data. To train the artificial neural network the supervised learning method of error backpropagation was used. This method applies the training function bayesian regularization, aiming to reduce the forecast error. The artificial neural network obtained a good performance, reaching a 12.97% MAPE. The comparisons of neural network performance for different time horizons are also presented in this context. The results achieved are in agreement with those found in the bibliography (sic).