Time Series based Ensemble Model Output Statistics for Temperature Forecasts Postprocessing

This repository provides supplementary material for the following paper:

Jobst, D., Möller, A., and Groß, J. 2024. Time Series based Ensemble Model Output Statistics for Temperature Forecasts Postprocessing. (preprint version available at https://doi.org/10.48550/arXiv.2402.00555)

Data

The data needed for reproducing the results is publicly available:

Jobst, David, Möller, Annette, & Groß, Jürgen. (2023). Data set for the ensemble postprocessing of 2m surface temperature forecasts in Germany for five different lead times (0.1.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.8193645

For the data license see here.

ECMWF forecasts

• Source: ECMWF (European Centre for Medium-Range Weather Forecasts)
• Gridded forecasts: 50-member ensemble forecasts
• Time range: 2015-01-02 to 2020-12-31
• Forecast leadtimes: 24, 48, 72, 96, 120 hours
• Forecast initialization time: 12 UTC
• Area: Germany
• Resolution: 0.25 degrees
• Meteorological variable: 2m surface temperature (t2m)

DWD observations

• Source: DWD Climate Data Center (German Weather Service)
• Observation data: Hourly observations of the target variable (2m surface temperature)
• Number of stations: 462
• ECMWF forecasts: Bilinearly interpolated to the SYNOP stations and reduced to its mean (t2m_mean) and standard deviation (t2m_sd)
• Metadata

Variable	Description
obs	Observation of 2m surface temperature
lt	Lead time
id	Station ID
name	Station name
lon	Longitude of station
lat	Latitude of station
elev	Elevation of station
date	Date
doy	Day of the year

Ensemble postprocessing

All models except of the EMOS and autoregressive adjusted EMOS (AR-EMOS) are estimated based on the static training data 2015-2019. For the EMOS and AR-EMOS model estimation a day-by-day sliding training window is applied which uses training data of 2019 and 2020. Finally, all models are evaluated in the whole year 2020.

R-packages and R-Scripts for the ensemble postprocessing models

• EMOS.R: Local EMOS with rolling training period.
• ensAR: Local autoregressive adjusted EMOS (AR-EMOS) with rolling training period.
• tsEMOS:
  • Local smooth EMOS (SEMOS).
  • Local deseasonalized autoregressive smooth EMOS (DAR-SEMOS).
  • Local multiplicative deseasonalized autoregressive smooth EMOS with generalized autoregressive conditional heteroscedasticity (DAR-GARCH-SEMOS ($\cdot$)).
  • Local additive deseasonalized autoregressive smooth EMOS with generalized autoregressive conditional heteroscedasticity (DAR-GARCH-SEMOS (+)).
  • Local standardized autoregressive smooth EMOS (SAR-SEMOS).

Additional R-packages

• imputeTS: For the missing value imputation.
• eppverification: For the verification of the ensemble postprocessing models.