Improving Medium-range Probabilistic Quantitative Precipitation Forecast for Heavy-to-extreme Events through the Conditional Bias-penalized Regression
Abstract
NWS Meteorological Ensemble Forecast Processor (MEFP) employs a statistical mechanism named Mixed-type meta-Gaussian distribution (MMGD) to create postprocessed probabilistic quantitative precipitation forecasts (PQPF) from single-valued quantitative precipitation forecast. A common feature of postprocessed PQPFs for heavy precipitation events is the underprediction of the probability of their occurrence. To remedy the above situation, we develop a new scheme for estimating the parameter of MMGD, which is herein referred to as Conditional bias-penalized regression (CBPR). CBPR, as its name implies, employs a regression technique that minimizes a hybrid cost function that is a weighted average of the error variance and Type-II conditional bias. Through the weight on the latter, CBPR to an extent mitigates the tendency of linear regression to underestimate the magnitude of extreme rainfall amounts; it moreover introduces additional flexibility in representing heteroscedasticity. CBPR performance was comparatively evaluated with a research version of MMGD in postprocessing of 4-day accumulated precipitation forecasts from Global Ensemble Forecast System (GEFS) over 18 basins in the service area of California-Nevada River Forecast Center (CNRFC) over the lead times of 1 to 4 days. The results indicate that CBPR consistently outperforms the baseline MMGD as measured by common verification measures, and the performance gain is particularly evident in forecasting heavy-to-extreme precipitation events.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2019
- Bibcode:
- 2019AGUFM.H33P2245G
- Keywords:
-
- 1816 Estimation and forecasting;
- HYDROLOGY;
- 1833 Hydroclimatology;
- HYDROLOGY;
- 6309 Decision making under uncertainty;
- POLICY SCIENCES & PUBLIC ISSUES;
- 6344 System operation and management;
- POLICY SCIENCES & PUBLIC ISSUES