FMI - Weather Forecast Production and NWP Postprocessing
Jussi Ylhäisi, Leila Hieta, Mikko Rauhala
Finnish Meteorological Institute
10 May, 1:30 pm, in 2890
Finnish Meteorological Institute (FMI) employs ca. 700 people. FMI
provides weather forecasts for the public audience and professional
customers and is the sole authority in Finland providing official
weather warnings and services for other authorities. A part of the
applied research and development done in the institute is aimed at
improving the data quality of our forecasts. A brief overview of these
development projects and our production system is given in this
presentation. An emphasis is given to the various postprocessing
activities and projects that we currently develop.
The primary local area model (LAM) currently used at FMI is a 10-member
HARMONIE-AROME-based MEPS with a 2.5km horizontal resolution. The
ensemble consists a unique setup between Nordic countries, where the
members jointly share supercomputing and development resources between
each other. The LAM is run for the whole Scandinavian domain and
different ensemble members are spread over several super computers. FMI
has had an open data policy since 2016: Most of our observations and
model forecasts are available for the general audience.
In recent years FMI has put substantially more effort in postprocessing
of NWP forecasts. As a part of our operative forecasts, we incorporate
both various statistical postprocessing frameworks and diagnostic
postprocessing algorithms written by our experienced duty forecasters.
Our main tool for diagnostic postprocessing is an open-sourced
postprocessing package HIMAN, which is currently able to calculate more
than 60 diagnostic variables from the output of any NWP model. For
statistical postprocessing, we have since February 2017 operatively
produced point-based and Kriging-gridded Model Output Statistics (MOS)
forecasts over the European domain, based on ECMWF HRES model. As a
part of our long-term strategy, we are currently building a grid-based
framework for model blending, where we aim to generate a consensus
forecast over Scandinavian area through blending our MOS forecasts with
the direct model output (DMO) from other NWP models.
Despite having a heterogeneous training sample for our MOS forecasts,
the operative verification clearly shows the improvement over the ECMWF
DMO forecasts. Our preliminary results for the blended forecasts also
show an improvement over the MOS forecasts, which again includes the
potential to reduce the data sources the forecaster needs to take into
account in shift work.