Abstract
European and global power grids are moving towards a Smart Grid architecture. Supporting this, advanced measurement equipment such as PQAs and PMUs are being deployed. These generate vast amounts of data upon which machine learning models capable of forecasting incipient faults can be built. We use live measurements from nine PQA nodes in the Norwegian grid to predict incipient interruptions, voltage dips, and earth faults. After training ensembles of gradient boosted decision trees on spectral decompositions of cycle-by-cycle voltage measurements, we evaluate their predictive performance. We find that interruptions are easiest to predict (95 % true positive, 20 % false positives). Earth faults and voltage dips are more challenging. Our models outperform naïve classifiers. We have explored forecast horizons of up to 40 seconds, but we have indications that forecast horizons of at least a few minutes are feasible.