University of Washington - Department of Statistics
The past decade has seen a culture change in the practice of numerical weather prediction. Up to the early 1990s, numerical weather forecasting was an intrinsically deterministic endeavor.
National and international weather centers used sophisticated computing resources to run carefully designed simulation code. While this is still the case today, the introduction of ensemble prediction systems has led to a radical change. Ensemble systems generate an ensemble of initial conditions and run the simulation code forward in time from each of them in turn, thereby generating a sample from the predictive distribution of future weather states.
Ensemble forecasts are typically subject to forecast bias and underdispersion, and therefore uncalibrated. To address these issues and to obtain calibrated predictive distributions, we propose the use of ensemble model output statistics (EMOS), an easy to implement statistical post-processing technique. The EMOS technique is based on multiple linear regression and yields probabilistic forecasts in the form of Gaussian predictive probability density functions. The EMOS predictive mean is an optimal, bias-corrected weighted average of the ensemble member forecasts, with coefficients that are constrained to be nonnegative and associated with the member model skill. The EMOS predictive variance is a linear function of the ensemble spread. For estimating the EMOS coefficients, we use minimum CRPS estimation, that is, we find the coefficient values that minimize the continuous ranked probability score (CRPS) for the training data. This can be understood as robust M-estimation.
We applied the EMOS method to 48-hour ahead forecasts of sea-level pressure and surface temperature over the North American Pacific Northwest in Spring 2000, using the University of Washington mesoscale ensemble, and with good results.
Joint work with Anton Westveld, Adrian E. Raftery and Tom Goldman.