Abstract
New technology enables power grid operators for the first time to monitor the dynamics of the power grid from hour to hour. The flexibility in operation this facilitates is crucial for an age of unpredictable power sources such
as wind and solar. However, such advanced metering infrastructure (AMI) generates enormous amounts of data, and extracting meaning from the torrent of data represents a significant challenge. In this thesis, we explore the possibility of quantifying specific components of real power consumption data. Crucially, we do this at the resolution level available to grid operators. We test four models on relevant performance measures, one using a filtering approach, and three using artificial neural networks. The testing shows that the artificial neural networks outperform the filtering method by around 50%,
and a logistic regression baseline by around 6% as measured by F1-score. By the same metric, an ensemble of the models outperforms the best model by around 3%. The models also vary significantly in their precision/recall tradeoff, as well as their ROC-curves. Further, we perform a sanity check of the models on unseen data from the Norwegian power grid, and show that the methods proposed provide a good foundation for real-world applications.
as wind and solar. However, such advanced metering infrastructure (AMI) generates enormous amounts of data, and extracting meaning from the torrent of data represents a significant challenge. In this thesis, we explore the possibility of quantifying specific components of real power consumption data. Crucially, we do this at the resolution level available to grid operators. We test four models on relevant performance measures, one using a filtering approach, and three using artificial neural networks. The testing shows that the artificial neural networks outperform the filtering method by around 50%,
and a logistic regression baseline by around 6% as measured by F1-score. By the same metric, an ensemble of the models outperforms the best model by around 3%. The models also vary significantly in their precision/recall tradeoff, as well as their ROC-curves. Further, we perform a sanity check of the models on unseen data from the Norwegian power grid, and show that the methods proposed provide a good foundation for real-world applications.
