Improved Rain Predictions Could Improve Flood Forecasting

"The problem in Pennsylvania is flooding," says Paul G. Knight, state climatologist at Penn State. "Pennsylvania is one of the more flood prone states in the nation. It is in the top quarter." Pennsylvania is laced with rivers, creeks, streams and mountains, making it vulnerable to flooding. The State is also flood prone because flooding is not a phenomenon of only one season.

"In the spring and fall we get slow floods as the ground saturates and streams fill; in the winter, rain on frozen ground, along with snow and ice melt often causes flooding and in the summer, thunder storms and tropical storms also cause flooding," says Knight.

These varied weather conditions contribute to the difficulties in forecasting heavy rains. The computer models that help forecast weather employ convective parameterization schemes to account for atmospheric conditions that are not completely rendered in the models. In essence, these schemes are fudge factors that help the models approximate rainfall and improve forecasts.

"Thunder storms, for example, are rather small and fall between the grid points in our models," says Knight. "Convective parameterization schemes try to account for such events that are smaller than the models can see."

Reducing the size of the grid squares is not yet an option because the increase in data points dramatically increases the amount of computer time and power necessary to run the models. Working with Michael S. Evans, National Weather Service meteorologist, Knight initially looked at case studies to produce a baseline of data on heavy rainfall events in the area covered by the Middle Atlantic River Forecast Center of the National Weather Service, State College, Pa. His current project expands and updates the baseline data on climate to include more data. Problems arise because past data are not in the same time frame or even the same small river basins as recent data. The baseline data helps in reviewing how different schemes to estimate the effect of thunderstorms work in predicting flooding. The researchers looked at several recent floods to see how different schemes predicted rainfall.

For example, the August storm that broke the 1999 mid-Atlantic region drought created twin flash floods in the Bradford and Mifflinburg areas. The storm caused more than $25 million in damage in the Bradford area alone.

"These twin floods posed a serious challenge to forecasters since both operational and experimental models predicted a wide range of scenarios," says Knight.

The researchers tested three different model configurations on the storm to see which produced the best fit for the actual event. Each model kept the same grid spacing but used different thunderstorm rainfall schemes. Of the three models tested, only one came close to predicting the rainfall that actually fell during the storm.

"These results do not mean that the scheme that produced the best results is the best scheme to use all the time," says Knight. "Depending on the season and the atmospheric conditions, different schemes are the best for different sets of conditions."

The researchers are trying to determine which schemes should be employed in which situations, but that will require more case studies and model testing. Ultimately, Knight and Evans would like to be able to predict storms with flood potential a day or two in advance.

The research was funded by a Cooperative Operational Meteorological Education and Training grant administered by the University Corporation for Atmospheric Research. COMET grants are intended to foster rapid technology transfer of research in forecasting to the forecasting community. The grants are given for cooperative work between universities and the National Weather Service and are regional in focus.