Previsão de cargas elétricas utilizando redes neurais artificiais MLP

Abstract

This work compares two Artificial Neural Networks for forecasting future electrical loads from a historical series of residential consumption and a historical series of total consumption of electric energy in Brazil, with a network designed from a Toolbox available in the Software Matlab, and another implemented using the same software. Noting the importance of energy planning for energy utilities, regulatory agencies and the free energy market, seeking to keep the National Interconnected System (SIN) prepared for the growth of the energy market. Making a brief introduction to the concepts of Artificial Neuron and Artificial Neural Networks (ANNs), it also mentions a brief timeline on the evolution of ANNs technology. It then presents articles with a similar theme, showing the effectiveness of the methods used by the aforementioned authors through the compilation of data from the articles, citing the term used for the forecast horizon of each of them and showing calculated error, as well as the location referring to the data. studied by each. In the methodology, the historical series are characterized. The first historical series consists of seventeen years divided into monthly values, using one hundred and eighty months as a training base and making the forecast and comparison with a series of twenty-four future months with the network implemented by Toolbox, reaching MAPE of 3.97 % for residential consumption and MAPE of 2.9% for total consumption; for the implemented ANN, the forecast generated was forty-one months out of chronological order, obtaining MAPE of 3.19% for the residential consumption series and 2.19% for the total consumption. The second application had smaller errors than the first, proving to be more effective, possibly due to statistical approximation error, called overfitting, in which the statistical pattern model applies well to the entire model during ANN training, however generates larger errors when generating the forecast. The implemented ANN has better signal input and output control parameters, as it allows a better adjustment of network parameters, such as the number of neurons in each layer.