The ability to predict forecast error has traditionally been measured by the strength of the correlation between ensemble forecast variance and the ensemble-mean forecast error. The usefulness of this type of deterministic forecast error prediction has been limited. The forecast error prediction problem could be considered probabilistically, such that forecast confidence is measured by the sharpness of the forecast distribution, provided this distribution is accurately depicted. In practice, current ensemble forecasts are biased, uncalibrated, and limited in size, particularly for mesoscale prediction of near-surface weather variables, making the true forecast probability distribution difficult to obtain. Until these obstacles are overcome, a more concerted effort should be placed into extracting the uncertainty information obtainable from current ensemble forecasts and used to make reliable forecast error predictions when possible.
To that end, the mesoscale forecast errors of sensible surface weather variables from two suboptimal, but highly effective short-range ensembles developed specifically for the U.S. Pacific Northwest are analyzed over two cool seasons. As part of the research, a simple stochastic model was developed to establish the practical limit of forecast error predictability from perfect ensembles of finite size. A variety of predictors and error metrics are examined to address the user-dependent nature of forecast error estimation. The potential benefit of applying a simple method of bias correction to the direct ensemble forecasts, before performing the calculations, is investigated. Additionally, the feasibility of constructing a secondary forecast error predictor from the temporal variability contained within an ensemble composed of lagged short-range forecasts with common valid times is explored. Finally, a new method of probabilistic forecast error prediction is evaluated.
Supplementary URL: