The use of short-term prediction strategies is an important tool for planning and operation of electrical systems, playing a crucial part in aiding the decision support process for buying and selling of electricity in the future market. For the energy market, in particular, an important component to take into account for consumption forecasting are the special days (holidays or atypical days, for example). Given its unusual behavior, the estimation of such events can be a complex task, when compared to the forecasting of ordinary days. In addition, as they are often found with only a small number of samples, it is difficult to adequately train and validate prediction algorithms. To tackle these problems, this work presents a model for short-term load forecasting using the Information Theoretic Learning Mean-Shift model to clustering and densify the sample size of special days's events on a time series, there on followed by the prediction using statistical and/or machine learning algorithms; in this work represented by artificial neural network algorithms and multiple Linear regression. The model was applied in a load forecasting problem for the electric utility in the northern region of Brazil, providing an improvement in the accuracy of results.
|
