The National Institute for Computational Sciences

Predicting the Future by Predicting the Past

Kraken latest weapon in high-resolution weather forecasting

by Gregory Scott Jones

Despite the comfort and confidence of the nightly weatherman, accurately predicting future forecasts is a difficult, and very important, task. Reliable weather information is critical to a number of government agencies and objectives, from the Department of Energy to the United States Air Force. When it comes to our military’s strategic, weather-dependent decisions, getting it wrong is not an option. And in order to get it right, you need computers—big ones.

To improve the accuracy of current weather models, and therefore improve forecasting, a team from the Atmospheric Technology Services Company (ATSC) in Norman, Oklahoma, is using the National Institute for Computational Sciences’ (NICS) petascale Kraken supercomputer to reforecast past weather patterns in the hopes of strengthening present models and future predictions.

Whereas traditional forecasting models rely on a deterministic approach, using a single model run that the forecaster favors, the ATSC team, led by Dr. Fanyou Kong, uses a 10-member ensemble method. By running ten different forecasts of the same model with slightly different initial conditions and physics options, the ensemble method provides weather forecasts on probabilistic framework, said Dr. Kong. Each ensemble produces ten different forecasts and the results are collectively graded using the statistical mean of the ten resulting predictions.

By retrospectively forecasting (aka reforecasting) weather conditions from 1987-2006, the team hopes to create a large statistical database that will then influence future forecasts for the entire country. By establishing statistical correlations between the reforecast dataset and the actual observed data over the 20-year reforecasting period, the team will apply various innovative algorithms to calibrate future ensemble forecasts. Such calibrations could help avoids any systematic errors that could result from the deterministic approach and drastically improve forecast skills. “We hope we can get a positive outcome,” said Kong, adding “this new (reforecast-calibration) system will yield a more valuable forecast for users.”

After the reforecast data is complete, the team will develop and test calibration algorithms, in hoping to produce forecasts with theoretically unparalleled accuracy with which government agencies such as the Air Force can make more informed strategic decisions. Because the team cannot be certain that they are using the ten best forecasts in the ensemble, the calibration methods are necessary to remedy any outliers or inaccurate projections.

Essentially, said Kong, the research represents “a combination of model and observation data combined with innovative statistical techniques.” All of these tools taken together, he said, “will lead to high-skill probabilistic forecasting at fine-scales” for a variety of phenomena such as surface wind, temperature and moisture, and accumulated precipitation.

The reforecasting for the entire 20-year period is done in increments of five days, giving the team six reforecast dates per month with which to gauge their findings with observed data. Simulating each month takes about 4.5 hours on Kraken with 840 computer cores, equivalent to 420 modern dual core desktop PCs. To finish off the entire task Kong estimates that the team needs about 1.5-2 million processor hours on Kraken, a system that Kong is highly complementary of. “My team is very happy to see the high performance that Kraken brings to our project,” said Kong, adding “We have already completed the 20-year reforecast dataset, and begin the calibration study” At this rate, said Kong, the team will be done well ahead of schedule.

But he and his colleagues aren’t waiting around. They have already begun the next step in the research: performing “sensitivity experiments” by generating small sets of reforecasts with different ensemble configuration and data availability, in order to determine the optimal ensemble-calibration system, for use in future weather forecasting, whatever they may be.

The new, more robust system can also be used in conjunction with other models, such as chemical dispersion scenarios from a possible terrorist attack, further strengthening their reliability as well. The scope and applicability of this research is vast. “Fanyou’s [Kong] work is setting a new path,” said Vicki Rose, the Managing Business Director and Owner of ATSC, a company with a history of consulting for and managing various government-owned weather stations. Rose, a Native American, won the current research contract through the Small Business Innovation Research (SBIR) contract to improve weather prediction for the Air Force and other government agencies.

According to Rose, the project’s primary objective involves improvement in weather forecasting at a local scale. The data acquired from Kong’s simulations are the main ingredient. “You need the data in order to do the calibration,” said Rose. But, she added, the new model will be a boon to all sorts of other weather-related research, research that is largely dependent on Kraken’s raw computing power.

“As a small business person I have been overwhelmed with gratitude by this level of support,” said Rose. “We couldn’t do this research without the support of NICS.”