Numerical weather prediction models are the backbone of modern weather forecasting. They discretise
and approximate the continuous multi-scale atmosphere into computable chunks. Thus, small-scale and
complex processes must be parametrised rather than explicitly calculated. This introduces parameters
estimated by empirical methods best fit the observed nature. However, the changes to the parameters
are changing the properties of the model itself. This work quantifies the impact parameter optimisation
has on ensemble forecasts.
OpenEPS allows running automated ensemble forecasts in a scientific setting. Here, it uses the OpenIFS
model at T255L91 resolution with a 20 min timestep to create 10-day forecasts, which are initialised
every week in the period from 1.12.2016 to 30.11.2017. Four different experiments are devised to study
the impact on the forecast. The experiments only differ in the parameter values supplied to OpenIFS, all
other boundary conditions are held constant.
The parameters for the experiments are obtained using the EPPES optimisation tool with different goals.
The first experiment minimises the cost function by supplying knowledge regarding the ensemble initial
perturbation. The second experiment takes a set of parameters with a worse cost function value.
Experiments three and four replicate experiments one and two with the difference that the ensemble
initial perturbations are not provided to EPPES.
The quality of an ensemble forecast is quantified with a series of metrics. Root mean squared error,
spread, and continuous ranked probability score are used with ERA5 reanalysis data as the reference,
while the filter likelihood score is providing a direct comparison with observations. The results are
summarised in comprehensive scorecards.
This work shows that optimising parameters decreases the root mean square error and continuous
ranked probability score of the ensemble forecast. However, if the initial perturbations are included in the
optimisation the spread of the ensemble is strongly limited. It also could be shown that this effect is
reversed if the parameters are tuned with a worse cost function. Nonetheless, when excluding the initial
perturbations from the optimisation process, then a better model can be achieved without sacrificing the
ensemble spread.
