275 — Predicting Electricity Outages Caused by Convective Storms

Tervo, Karjalainen, & Jung (1805.07897)

Read on 22 May 2018
#storm  #lightning  #tornado  #weather  #prediction  #SMOTE  #machine-learning  #neural-network  #dbscan  #meteorology  #GIS  #power-grid 

There are a ton of reasons why predicting storms is an interesting and important issue. But one of the difficulties in predicting the occurrence of strange or extreme weather patterns — aside from the fact that weather is basically a giant random number machine and the methods that we use to measure weather occurrences are extremely sparse and noisy — is that the actual occurrence of these weather systems is very rare.

Training statistical models to fit these sorts of rare-occurrence scenarios poses a major challenge: A model that predicts that extreme weather never occurs at all are usually right.

In order to convert GIS-based weather data into predictions of weather-related damage to a geographically distributed resource — namely, the power grid — the authors first classify the weather events into a four-category scale based upon damage, and then use the SMOTE technique — where minority events are over-sampled in synthetic manifestations of the dataset — to train the classifier only on relevant weather event patterns (surrounded by zeroed out null data).