A Data-Driven Algorithm for Short-Term Prediction of Over-Voltage and Under-Voltage Events in Distribution Grids

This paper proposes algorithms for short-term over- and under-voltage prediction in distribution grids. The proposed algorithms are developed using time-series of voltage and current measurements, which does not require the knowledge of distribution grid model (topology and parameters of the components). Various algorithms based on random forest classifier (RFC) and random forest regressor (RFR) methods, two prominent machine learning methods, are developed regarding different feature selection possibilities. The developed algorithms are tested and validated on two real datasets (grid measurement data from GridEye devices in two low voltage grids in Switzerland). An algorithm based on RFR method, with recent information including the measurement data of the last week at the same time of prediction, outperforms other algorithms. The proposed algorithm can predict over- and under-voltage events with 85% accuracy four hours ahead of the real time.