Now, the all noise cascade levels are initialized to zero.
It is probably better to initialize the noise with the appropriate noise field as provided by the EPS term.
The noise cascade has holds 2 time levels of noise. Both time levels should be initialized with the same noise field. If not the overall power of the noise cascade level will not converge to the correct power.
Default auto-correlations [code from STEPS1] below this fraction.
▶ Default (lag2) auto-correlations use a power law for the correlation length as a function of scale + black magic
To prevent a loss of power in the precipitation cascade levels, each level should be renormalized after an AR iteration.
However if you use a global parameters for the AR iteration, the renormalization results in a constant (in time) field.
If QPE zero and NWP nonzero ➡ NWP provides the advection.
Set radar mean and variance and advection to zero
Use default power spectrum slopes (?) and the default auto-correlations set the rate of blending from zero rain to NWP forecasts through the calculations of nowcasting skill as a function of lead time and scale.
If I run an ensemble using multiple cores (so, dask will parallellize the ensemble members over the cores), it seems that the ensemble order is lost, resulting in weird transitions (due to noise from a different member that ends up in that member). See for instance (these are three 15-min instance in the forecast):
Or maybe even clearer:
Has any of you come across this and/or know how to fix it?
If I run on 1 core, this problem does not occur, so it gives the impression that this has to do with parallellizing using dask.
Add tests for the different combinations of no rain in NWP, radar, and both. @RubenImhoff could you pull the version in my hackaton branch and check if your case still works there? Thanks a lot!
I currently get errors for my tests (more specifically the decomposed precip still contains nans)