Extended Abstract (88K)3.6
Updateable model output statistics: an effective tool for evaluation of NWP forecasts
Marcel Vallée, MSC, Dorval, QC, Canada; and L. J. Wilson
The Canadian updateable model output statistics (UMOS) system became fully-operational in the spring of 2000. The essence of UMOS is that most of the preparation for the statistical processing of the model output is carried out in real time, allowing frequent and rapid redevelopment of the equations. . The system also includes a weighting scheme to blend data from newer and older model versions in the equation development. From the point of view of model validation and verification, the rapid response capability of UMOS assists in two ways. First, by examining the frequency of selection statistics of the various model predictors, and the overall goodness of fit of the equations on the dependent sample, one can get an idea of the quality of new model predictors compared to those from previous versions. Secondly, by assessing the performance of the equations on independent data, one can determine the impact of model changes relatively soon after a model change. MOS-related model validation allows one to examine the model performance over series of forecasts valid at points, which complements model validation using case studies.
The Canadian UMOS system currently predicts three hour spot 2m temperature, three hour 10m wind speed and direction, six and twelve hour probability of precipitation greater or equal than .2mm, and cloud amount. The last of these is predicted categorically using multiple discriminant analysis, while the others use multiple linear regression.
The most recent significant NWP model change in Canada occurred last September 2001. For the first time ever a full set of MOS equations was running operationally at the very first day of the NWP implementation. The earliest verification statistics from the UMOS forecasts were available only three weeks after the model change, and gave an early indication of the impact of this change. The presentation will include sample results from this most recent model implementation, with a focus on their use to diagnose the impacts of the model change.
Session 3, Statistical Evaluation I
Monday, 12 August 2002, 4:30 PM-5:45 PM
Previous paper