Methods and Technologies
Numerical Weather Prediction (NWP) is able to predict the general circulation pattern for about one week ahead with impressive accuracy.
Also elements which are reported from surface stations like T2m, wind, precipitation and cloud cover - so called Direct Model Output (DMO)
elements - are predicted by the numerical models by parametrization.
The quality of this parametrization by physical means is very inhomogenious in dependence on the specific station, element, season and weather
situation. Sometimes it is of very good accuracy, sometimes it has systematic errors and sometimes it is of no use at all.
Statistical Weather Forecasting provides an intelligent feed-back between numerical model forecasts and observations. It removes
systematic and random errors from the DMO. When DMO has no or very little predictive quality, statistical interpretation of
the circulation pattern - using other input variables than DMO - leeds to the best result. Screening regression is a good tool
to combine the model variables with the best predictive information into one equation. A comprehensive statistical weather
forecasting system has to be able to do both:
-
Correct DMO when it has conditional systematic errors (depending on lead time, season, weather situation and/or other
weather elements) and
-
Ignore DMO and parametrize the elements of interest using other predictors.
Basicially, there are three types of statistical weather forecasting methods:
-
PPM - Perfect Prog Method: Relations (based on analogy or regression) between analyzed fields and observed weather elements are
derived and applied to model forecasts.
-
MOS - Model Output Statistics: Relations (usually regression equations) between predicted numerical model fields and observed
weather elements are derived.
-
Adaptive Methods - e.g. Kalman Filter: Relations change every day after most recent observations are present.
Details about the advantages and disadvantages of these methods can be found on the
page publications.
An ideal statistical weather forecasting system would consist of PPM, MOS and adaptive methods. Most of the real-world
systems omit the PPM part since long archives are not available and/or it is more easy to develop just a MOS system.
Kalman Filtering of MOS forecasts may be useful if severe changes in the numerical model occurred. Kalman Filtering
without MOS leeds to less accurate forecasts since the Kalman Filter method and equations are too simple. The loss of
accuracy - measured in terms of reduction of error variance of the forecasts - of pure Kalman Filtering as compared to
pure MOS is for standard elements (T2m, wind, clouds) about 20 %.
|
