BART—Physical Damage Approach for Power Outages Forecast

Hurricanes cause significant damages to the electrical power infrastructure. A recent example is the passage of hurricane Maria in the Caribbean which reached a peak intensity of maximum sustained winds of 175 MPH (280 km/h) in September 2017 destroying most of the 2,400 miles of transmission lines. After 1-month, more than 3 million people (80% of population) were still without power. The preparedness for such events can be improved by anticipating the likely damage that extreme weather events may cause. Primarily, this will help the utility companies and emergency managers making restoration plans in advance of the extreme weather event.

The aim of this work is the development of a power outage forecast model using the experience of Hurricane Maria as a learning and training event. The approach is based on machine learning tools for power outage forecasting, specifically a Bayesian additive regression tree (BART). The BART multivariable approach improves over models that consider only one variable, mainly wind speed, with physical and mechanical information of the power distribution infrastructure. BART is a data mining, fully Bayesian probability model, with a prior and likelihood. As with any other machine learning algorithm, BART requires a large amount of data to produce accurate predictions, and uses a set of explanatory variables to explain the behavior of the response variable. As a result of catastrophic events, lack of data is a typical scenario. Our approach uses a risk assessment (referred Impact Model) of one of the most critical components of the power network, the high voltage transmission lines. The impact model estimates the risk of failure of the structure due to extreme wind conditions, using mechanical stress analysis and power tower design criteria. The impact model results, along with the following explanatory variables are used to make predictions: Wind Speed; Precipitation; Soil Moisture; Elevation, and MODIS Land Classification.

For the case study of Hurricane Maria, weather data was simulated using the Weather Research and Forecasting Model (WRF). The BART model was trained using a preliminary dataset, that was gathered for one of the 230 kV transmission lines. The island was divided into four grid cells, for the purpose of comparison between the actual failure data and those estimated by BART. Future work will include further refinement of the Impact Model using actual power outages data for multiple past storms including hurricanes Irene and Georges.