Mean shift densification of scarce data sets in short-term electric power load forecasting for special days

Short-term load forecasting plays an important role to the operation of electric systems, as a key parameter for planning maintenances and to support the decision making process on the purchase and sale of electric power. A particular case in this respect is the consumption forecasting on special days, which can be a complex task as it presents unusual load behavior, when compared to regular working days. Moreover, its reduced number of samples makes it hard to properly train and validate more complex and nonlinear prediction algorithms. This paper tackles this problem by proposing a new approach to improve the accuracy of the predictions amidst existing special days, employing an Information Theoretic Learning Mean Shift algorithm for pattern discovery, classifying and densifying the available scarce consumption data. The paper describes how this methodology was applied to an electrical load forecasting problem in the northern region of Brazil, improving the previously obtained accuracy held by the power company.