Experimenting Model Blend at the Finnish Meteorological Institute

Finnish Meteorological Institute (FMI) has successfully used manual editing for several years to produce operational weather forecasts. Editing has mainly been done using physical-based diagnostic algorithms developed by the duty forecasters. The continuously increasing amount of weather data has made it more difficult for forecasters to be able to interpret and effectively use all the information available. Also the temporal and spatial resolution of weather models has increased during the past years. This has led FMI to seek new methods to provide better base information for the forecasters to produce operational weather forecasts.

FMI has implemented a method to create a model blend where information from multiple models is combined to one consensus forecast. Method is the same as in NOAA’s National Blend of Models (NBM). First we bias correct the raw model data using decaying average method and then calculate weights for each model based on recent MAE verification results. All calculations are done in grids and for individual models and lead times separately. Observation analysis field produced by Local Analysis and Prediction System (LAPS) is used as the ground truth that against all the model errors are calculated. The method doesn’t need archives of past model data and is fast and reliable to run operationally.

FMI has produced operational model blend for the Scandinavian domain for 2m temperature and dew point temperature since March 2019. The blend is calculated twice a day after ECMWF model forecast data is available. The weather models included to the current FMI blend are GFS, ECMWF, Hirlam, Harmonie and MOS corrected ECMWF output. Hirlam and Harmonie are limited area models using ECMWF’s forecasts for boundary conditions.

Verification results for the FMI blend have been very promising. The model blend provides seamless and consistent forecasts and has in general lower RMSE than its components. The next step is to add more parameters to FMI blend calculations and to research bias correction methods for different parameters and possibly for different seasons.