Hosting capacity of low-voltage grids for distributed generation: Classification by means of machine learning techniques

Abstract

A high amount of installed distributed generators (DG) in low-voltage grids, e.g. photovoltaic generators (PV), may cause serious problems due to overloading of electrical equipment and violation of voltage limits. The assessment of low-voltage grids regarding their hosting capacity for the installation of DG is a difficult task, because grid structures may be diverse and complex. In this article, we classify grids by means of machine learning techniques, in particular support vector machines (SVM). SVM learn to assess grids by means of sample data, that is, grids represented by characteristic features that were assessed by human domain experts (i.e., distribution system operators (DSO) staff). We show that this approach can significantly better reflect domain expert assessments compared to a technique we proposed earlier which is based on a stochastic load flow simulation procedure and a subsequent parametric stochastic model estimation. One key result of this article is that SVM with grid based features significantly outperform SVM using features from load flow simulations regarding the classification accuracy if both are trained with data that were assessed (labeled) by DSO staff. Experiments are based on data for 300 real rural and suburban low-voltage grids.