With the inclusion of renewable energies and a growing trend towards electrification, the electrical power grid is put under heightened strain. This makes the quick identification and resolution of instabilities within the grid instrumental. With the help of machine learning, a classification of instabilities is possible. This thesis focuses on the identification of voltage instabilities, i.e. rapid voltage changes and voltage dips, by developing a machine learning model. Multiple models are tested, out of which one with optimal recall and one with optimal precision are identified. In addition, influential input factors are analysed. By using real grid data from a medium voltage grid in Sweden, it is hoped that the model developed in this thesis can help grid companies identify and predict voltage instabilities within their grids. In literature, models based on decision trees are deemed the most promising. A base model is identified, and its parameters are optimised based on recall or alternatively precision, in this case a BalancedBaggingClassifier, included in python’s Scikit-learn library. Also, a rolling horizon is applied to the input data, due to the scarcity of instabilities within the data. Different input sets are tested and it is found that the inclusion of all available phases of voltage data is necessary for good performance, while it is not needed to include current or power data. At the end, the models are tested for their ability to forecast instabilities. While short-term forecasting of 1h is very successful with the precision model, the recall model has more versatility in its forecast horizons.
Sarah Meuter (Wed,) studied this question.