DYNAMIC DIMENSIONING OF BALANCING POWER WITH FLEXIBLE FEATURE SELECTION

This paper proposes a novel dynamic design for control reserve dimensioning. In contrast to the current statistical analytic design we present a data driven approach with methods of computational intelligence. The chosen k-nearest neighbor algorithm is one of the most successfully used methods in machine learning. The model is able to predict complex nonlinear behavior by assuming that similar observations have similar outcomes. A condition for the success of this method is to determine the salient features. Therefore the core of this paper is to compare different methods of feature selection for the prediction task of control reserve. Numerical experiments for the year 2012 show that a machine learning approach has specific advantages over traditional approaches.