P4.3
New Bayesian probabilistic forecasts of river discharge: Linking weather forecasts to intraseasonal prediction
carlos Hoyos, Georgia Institute of Technology, Atlanta, GA; and P. J. Webster
The optimal time scale for a forecast of river discharge is on the time scale of 20-30 days. This provides sufficient lead-time to make adjustment to strategies in both agricultural practices and water resource management. Such forecasts are especially vital for developing agricultural economies. A special problem exists in Bangladesh that is without any upstream data from India in either the Ganges or Brahmaputra catchment basins. To this end, the Climate Forecast Applications in Bangladesh project provides 1-10, 20-30 and 1-6 month discharge forecasts on the basis that the Ganges and Brahmaputra are ungauged river basins. To accomplish this goal, a mix of models, external data and statistics are employed. The purpose of this talk is to describe the intermediate (20-30 day) forecasting scheme and some new innovations that have increased accuracy and also utility through producing probabilistic forecasts. The intermediate scheme utilized by CFAB is based on the banded wavelet method of Webster and Hoyos (2004). The predictand and the predictors are separated into bands determined by spectral peaks that exist in the time series of the predictand. The predictors are chosen by their physical relevance to intraseasonal variability and are chosen following long-term and rigorous composite analysis. Regression analysis of each band does not allow errors in the high-frequency data “pollute” the slow manifold that determines the intraseasonal variability. Regional rainfall and river discharge forecasts show correlations at 20 days of >0.8. It has been realized that not all information has been included in the empirical scheme. For example, forecasts of the predictors out to 10 days are made at operational centers. We use the 51 daily realizations of the predictors using 1-10 day forecasts from ECMWF to adjust the Bayesian priors of the empirical model. First, the forecasts are rendered probabilistic which is much more useful for decision-making. Second, accuracies of the forecasts are improved. We present examples of forecasts for the summer of 2006 including probabilities of exceedance of flood level. Finally, we discuss the general utility of the new Bayesian scheme.
Poster Session 4, Weather to Climate Scale Flood Forecasting Posters
Wednesday, 17 January 2007, 2:30 PM-4:00 PM, Exhibit Hall C
Previous paper Next paper