The design of the UMOS system addresses the problem of maintaining a MOS forecast system in an environment of frequently changing models. Following a model change, data from the old and new models are weighted according to a formula that assigns high weights to new model data, and gradually phases out the old model data as the sample size from the new model approaches 300 cases or so. The expectation is that the blending of the model data would ensure a stable and steady transition of the equation coefficients from old to new model. However, the mixing of predictor data from models with different bias characteristics might contribute significant unexplainable variance to the development datasets. The system uses forward stepwise screening regression for temperature, wind and POP; a discriminant analysis module is under development for categorical predictands such as cloud amount, ceiling, visibility and precipitation type.
Tests were done to evaluate the design of the UMOS system and to compare its performance to the existing perfect prog forecast system during a model transition period following a major model change in September 1998. Results indicate that the coefficients changed remarkably little during the transition and that the response to the new model data usually occurred with the addition of a relatively small sample from the new model. The goodness of fit of the regression equations also did not decrease during the model blending period, except for wind speed, where a slight deterioration was noticed, probably due to the change in model resolution. Comparisons with the perfect prog forecasts revealed that the UMOS forecasts were superior throughout the model transition period, but that the improvements were small after the first day of the forecast period.
Selected test results will be shown to illustrate the characteristics of the new system.
Supplementary URL: