| Abstract [eng] |
In this study, results concerning statistical verification of DEMETER (Development of a European Multimodel Ensemble system for seasonal to inTERannual prediction) seasonal forecasts over Lithuania is presented. The main task is assessing multimodel prediction system performance. The system comprises the global coupled ocean-atmosphere models of several institutions. In order to assess seasonal dependence on forecast skills, the DEMETER hindcasts have been started from the 1st February, 1st May, 1st August and 1st November, the analyzed period is 1990-2001. The atmospheric and land-surface initial conditions are taken from the ECMWF Re-Analysis (ERA-40) dataset. The ocean initial conditions are obtained from ocean-only runs forced by ERA-40 fluxes, except in the case of MPI that used a coupled initialization method. Ech hindcast has been integrated for 6 months and comprises an ensemble of 9 members. Multimodel forecasts were verified against observations of 17 meteorological stations obtained from the Lithuanian Hydrometerological Service. In section 1 and 2 of this paper a general definition of seasonal forecast are introduced and some characteristics of statistical forecast verification are discussed as well as the short description of DEMETER multimodel members is given. Section 3 explores some basic and the most common in use technics of continuous forecasts verification. Based on statistical verificatiom methods, described in previous section, an analysis of each seasonal forecasts follows in Section 4. The summary offers main findings of this study. The multimodel system tends to overforecast temperatures in warm season and to underforecast in cold season by increasing of false alarms. In the western and coastal region models have higher reliability scores than temperature forecasts made for other parts of Lithuania. The amount of cases, when relationship between forecasted and observed values is less than ±2 degrees Celsius, is the largest of hindcast started from 1st August integrated for 6 months (that is cold season). The detailed analysis of verification results, obtained from 1st February hindcast, shows, that Met Office model performed generally more positive and/or negative biases of forecasts during 1990-2001 than any other model. Unbiased forecasts were presented by Météo-France model; it means slight difference between observed and forecasted values. The highest estimated Percent correct of analysed forecasts for Lithuania is 68%. Results show, that most reliable forecasts from each hindcast have been performed by ECMWF model. |
