Correcting the Bias in the CFS Version 2 Precipitation Forecasts

Bias in the CFS Version 2 precipitation forecasts over the global land is examined and corrected against the CPC unified daily precipitation analysis for an 11-year period from 1999 to 2009. The bias in the CFS Version 2 precipitation forecasts presents regional variations and seasonal changes, and appears as a non-linear function of the target precipitation intensity. In addition, the bias also varies with the forecast leading time especially in the first several days of the forecasts.

A prototype technique is developed to correct the bias in the CFS Version 2 precipitation forecasts through matching the probability density function (PDF) of the forecasted daily precipitation against that of the co-located CPC unified daily gauge analysis. Cumulated PDF tables for the CFS Version 2 forecasts and the gauge analysis of daily precipitation are constructed for each T126 grid point for each calendar day and for each leading time step using matching pairs of forecast and gauge data collected over a region centering at the target grid point over sliding time windows of 31 days centering at the target Julian day over the 11-year time period. Only data pairs over grid box with one or more reporting gauge are included in the construction of the PDF tables. The spatial domain of data collection is expanded until sufficient number of data pairs are secured to ensure stable PDF tables.

The constructed PDF tables are then applied to remove the bias in the CFS Version 2 precipitation forecasts, assuming that the gauge analysis presents the truth of precipitation intensity frequency distribution. Preliminary examination demonstrated the effectiveness and feasibility of the PDF matching method to correct the bias in the CFS Version 2 precipitation forecasts. Bias-corrected CFS Version 2 precipitation forecasts presents overall magnitude and PDF of precipitation intensity in very close agreements with those of the gauge observations. Once the bias is removed, performance of the CFS Version 2 precipitation forecasts is then examined through comparison against the gauge analysis. Performance statistics (e.g. anomaly correlation, probability of detection) is examined for the CFS Version 2 forecasts of daily precipitation for different region, season, forecast leading times, and target precipitation intensity. Overall, CFSv2 exhibits some skills in predicting the mean precipitation anomaly and the occurrence of heavy rainfall in the first 15 days or so. The skills, however, diminish for forecasts with leading time of 15 days or longer. Detailed results will be reported at the AMS Conferences.