When the next hurricane hits the Gulf states, a new prediction model aims to anticipate the areas that will be hardest hit by power outages, enabling the accurate pre-placement of repair crews.

Gulf-state electric utilities were hit this decade by five devastating hurricanes, prompting them to search for smarter ways to shorten the resultant power outages--one of which lasted for 11 days. The best bet so far is the new outage prediction model, which combines machine learning with real-time data monitoring, in hopes of reducing the length of power outages when next a hurricane hits.

"We give forecasts about when and where a hurricane will strike, and how bad it will be," said engineering professor Seth Guikema at Johns Hopkins University in Baltimore. "We have a better way of forecasting how many power outages there will be and where they will be, so that utility companies can know how many crews to call up and where to station them."

Crafted in cooperation with Steven Quiring, a professor of geography at Texas A&M, the new outage prediction model learned from historical statistics amassed during the power outages caused by past hurricanes including Dennis (1995), Danny (1997), Georges (1998), Ivan (2004) and Katrina (2005). The hope is to lessen the length of power outages. For instance, Ivan created 13,500, Katrina more than 10,000, Dennis 4,800, Georges 1,075 and Danny 620.

The learning algorithm, called a generalized additive model, learns the patterns of past outages by dividing the service area into thousands of grid cells. Using the historical data about outages in those cells, it creates a predictive statistical model. Then when a new hurricane is headed that way, the team is able to predict the wind speed in each grid cells from the past history, allowing utility companies to pre-position repair teams in what are likely to become the worst hit areas.

"What is unique about our approach, is it is based on a lot more than just poles and wind," said Guikema. "For instance, it considers long-term average precipitation, the current soil moisture, the topology of the landscape, the amount and kind of vegetation, and other factors. For example, if you have a really wet summer and then have a hurricane, you are going to have poles that are easier to blow down and cause outages."

The generalized additive model learns not only from historical data, but during each future hurricane will continually refine its accuracy by comparing its predictions to what actually happened, then updating the model so that it makes even better predictions next time.

Just recently completed, the researchers are currently waiting until the next Gulf state hurricane to put the outage prediction model to the test.