Ensemble pdf creation for temperature

Dependencies

Dependencies are included in requirements.txt. Install with pip install -r requirements.txt.

Additionally, one of the dependencies is not yet included in the PyPI for Python 3. Install it through:

pip install git+https://github.com/vanife/pyfscache

Running the program.

The program and ensemble can be configured by modifying config.ini.

Verification

Rank histograms
Skill scores
CRPS
QQ plot

Algorithms for single forecast hour

Assumptions

Temperatures can be modeled by Normal distributions
Ensemble members are equally likely to represent the observation a priori
The ensemble set behaves as a randomly selected sample from the expected distribution of observations.

Pseudo code

function main() {
Input:
    Forecast hours F
    ensemble set S
Output:
    PDF for each forecast hour

For each forecast hour f in F:
    For each ensemble member m in set s:
        m.variance <- getVariance()
    f.pdf <- getPdf()

return F
}

Options for variance

Get control member variance
Derive member variance from ensemble mean variance

Options for model combination

Uniform distribution as prior

Data

The following data columns

Model name + Element name
Element observation value
Element value
Station id
Forecast hour
Issue date
Valid date

Ideas

Do something with covariance between model hours

Maybe Gaussian process fit on the ensemble members.
Or fit all hours together respecting their covariance
Maybe have a look into time series modeling to see if we can use autocorrelation

Relate variance between multiple issues for the same forecast hour