The important role of feature selection when clustering load and generation scenarios

Abstract

Power market models can generate load and generation scenarios, for a given market regulation. The generated scenarios can be interpreted as a sample of the future utilisation of the power network, and be used as a basis for a contingency and reliability analysis. However, to use all the generated scenarios as input in a contingency and reliability analysis can lead to quite extensive computational requirements. A data reduction framework, which finds groups of similar scenarios, and only uses the group characteristics as input in a contingency and reliability analysis, is presented and discussed. It is shown that the data reduction framework can reduce the computational requirements by about 90% with little loss of accuracy. However, the success of this approach is highly dependent on which features that are used to quantify similarity between scenarios, and it is shown that choosing a set of nonoptimal features leads to large errors. The feature selection is compared with the choice of clustering algorithm, and shows that the feature selection process has a much large impact on the results than the choice of clustering algorithm.