Grant-funded project to tackle uncertainty of weather forecasting

How certain can we be that a weather forecast is true? It’s a very big question for aircraft pilots and ship captains, for it involves uncertainty about temperature, precipitation, pressure and so on at a range of locations, altitudes and times in the future. Current ways of assessing and communicating these complex uncertainties are just not good enough.

Enter an interdisciplinary team of UW faculty from the Statistics, Sociology, Atmospheric Sciences and Psychology departments and the Applied Physics Laboratory (APL). They’ve just won a $5 million grant to tackle this problem. The proposal was submitted by the UIF-funded Center for Statistics and the Social Sciences (CSSS); center director Adrian Raftery is the principal investigator. The funding comes from the Multidisciplinary University Research Initiative (MURI) of the Department of Defense.

One aspect of the problem that the group will tackle is that major weather models give predictions only at a grid of widely spaced locations that are often 30 miles or more apart. Local, or mesoscale, forecasts are commonly obtained from these subjectively, using rules of thumb. “During the past decade, increased computer resources and improving models have made mesoscale numerical weather prediction possible at scales of 1 to 10 miles,” said Atmospheric Sciences Professor Clifford Mass, a co-investigator on the project. “Thus the potential for direct forecasting of regional weather now exists.”

The project team will develop statistical methods for assessing uncertainty, based on the Bayesian melding approach originally developed by Raftery, professor of statistics and sociology, for whale populations and environmental risk assessment. According to co-investigator Tilmann Gneiting, assistant professor of statistics, the group intends to develop much faster ways of implementing Bayesian melding by using ideas from spatial statistics.

Skilled forecasters combine a great deal of data, and information overload has become a big challenge for them. The team will develop statistical tools to help forecasters.

But there are human factors as well as statistical ones at work here, making weather forecasting an interesting challenge for cognitive psychology. “We will address the problem of how best to communicate uncertainty in the simplest terms to an undertrained forecaster who is already overwhelmed by a large volume of information,” said co-investigator Susan Joslyn, lecturer in Psychology.

Meanwhile, co-investigator Robert Miyamoto, principal physicist at APL, will develop visualizations of the uncertainty. “The representation should depend on the user,” explained Miyamoto. “Visualizations may use animation, multivariate displays, or interactive tools such as drill-downs.”

A problem for local weather prediction is that large-scale weather models must be used for initialization. There are several well-established large-scale models in common use, and they can yield quite different forecasts. What to do?

Rather than choose a single model, the team will address this by running an ensemble of models and combining the results. To do this, they will apply the statistical approach of Bayesian model averaging, originally developed for model uncertainty in social science.

“A spirited debate has developed over whether to continue the trend to higher resolution or to run a large number of ensemble forecasts,” said Mass. “My research group in Atmospheric Sciences has concluded that more emphasis should be given to short-term ensemble predictions.”

The project will start May 1 and continue for five years.