Effects of the structure of training regime on a machine-learning based power outage model

This research investigates the effects of different training regimes, based on different groupings of the training data, on an operational machine-learning based Outage Prediction Model (OPM) applied across a large study region in the Northeastern United States (Connecticut, Massachusetts and New Hampshire). Training data defines empirical machine-learning models, their quality, and their outputs, but not all datasets are homogeneous or comprehensive. The training data in the OPM contains parameters describing weather events, infrastructure, land cover, soils, and elevation from different sources; weather events in our database range from 2005 through 2018 across a domain that encompasses five historically separate power utilities across New England. There could easily be heterogeneous groups of data in the training data caused by factors such as changes over time, or unaccounted differences in regional climate, predominant tree species, or the maintenance practices of the utilities. If these different groupings are treated as the same and given to a machine learning model, it would harm the quality of the model by introducing noise. Our hypothesis is that refining and separating these groupings of data, and training a separate model for each group should yield better results.

This study evaluates several different training groupings for OPM training; the groupings are informed by political and operational boundaries as well as a clustering analysis of the training data performed with agglomerative hierarchical Ward algorithm. The results demonstrate an optimal way of structuring the training regime of the OPM, suggest weaknesses in our current parameterization scheme, and highlight aspects of the mechanics of power outages that are poorly described with our current dataset.