Big-Data Approach Leads to More Accurate Hurricane Forecasting

Ask anyone who’s gotten stuck in the rain on a day forecast to be sunny, and they’ll tell you: weather prediction is an imperfect science. The stakes are raised, of course, when forecasting extreme weather events like hurricanes. Current statistical models used to predict the tropical storms are accurate only 65 percent of the time, resulting both in unexpected damages and money wasted in unnecessary preparations.

Now researchers from Northwestern University and partners have developed a new method based on big-data analytics for forecasting seasonal hurricane activity that is up to 16 percent more accurate than previous techniques. The new model uses millions of historical data points to predict the path and severity of the storms.

“Our model provides higher accuracy in hurricane prediction further in advance,” said Alok Choudhary, John G. Searle Professor of Electrical Engineering and Computer Science in the McCormick School of Engineering and professor of marketing and technology industry management at the Kellogg School of Management. “In general, knowing where a hurricane will hit and how severe it will be with higher accuracy gives policymakers more time to make decisions and reduce costs, leads to increased preparedness, and prevents economic losses from planning for weather events that don’t happen.”

A paper describing the findings, “Discovery of Extreme Events-Related Communities in Contrasting Groups of Physical System Networks,” was published Sept. 4 in the journal Data Mining and Knowledge Discovery.

As opposed to current “physics-“ or “simulation-based” modeling, the team’s data-driven “network motif-based model” relies upon masses of data mined from 60 years of historical weather records. This allows the researchers to fully evaluate a spectrum of variables (such as temperature, pressure, and cloud cover) in order to identify the combinations of factors that are most predictive of hurricane activity. The researchers found their new model to be 8 to 16 percent more accurate than current methods.

The network motif-based model may also be applied to other severe weather events, such as typhoons and droughts, as well as disease outbreaks and other extreme events.

Other authors of the paper include lead authors Zhengzhang Chen and William Hendrix, both postdoctoral researchers at Northwestern; Nagiza Samatova, an associate professor of computer science at North Carolina State; Fredrick Semazzi, a professor of marine, earth, and atmospheric science at North Carolina State; Isaac Tetteh, a lecturer at Kwame Nkrumah University of Science and Technology in Ghana; and Hang Guan, a student at Zhejiang University.

The research was supported by grants from the National Science Foundation and the Department of Energy.