ConvMOS: Climate Model Output Statistics with Deep Learning

This repository contains code for the paper "ConvMOS: Climate Model Output Statistics with Deep Learning".

Required data

• REMO data
  • This data is available here (~ 5.6 GB).
  • It is already prepared for use with ConvMOS.
• GTOPO 30
  • GTOPO is provided by USGS.
  • Make sure you remap it to the same grid as the REMO data with cdo for example.
• E-OBS
  • Available from ECAD. We used V19.0e for our paper.
  • Make sure you remap it to the same grid as the REMO data with cdo for example.

Correct paths to the data have to be set in the config.ini.

Data preparation (if you have raw REMO data)

This code is designed to work with data from REMO RCM. To this end, we have REMO variables as netcdf-files at the EUR-44 grid and GER-11 grid. We base this on data with timesteps of 6-hours, but we want daily data. These files have to be preprocessed for them to be used here. This is achieved with prepare_netcdf_files.sh.

First, I made sure that the data is available in the following structure with .nc files in the leave nodes:

EUR-44 or GER-11
    {var1withdifferentpressures}
        {pressure1}
        {pressure2}
    {var2withoutdifferentpressures}
    {var3withdifferentpressures}
        {pressure1}
        {pressure2}

Sometimes, there are accidental a.nc files in these folders which have to be removed.

Then, I ran the script for each var-folder like, for example, so ./prepare_netcdf_files.sh EUR-44/var1withdifferentpressures -d -f EXCEPT for variable 260 (precipitation), which I preprocess with ./prepare_netcdf_files.sh EUR-44/260 -d -f -s, so that it calculates the daily sum and not the mean of the 6-hour parts.

While it most probably does nothing bad to leave it like that, I also commented out the remapbic cdo call and changed the output of the daily aggregation step to ${filename}_remap.nc for GER-11, as the original data provided is already in GER-11 grid, so no remapping is necessary. Note that I left it in once and compared the resulting files with ncdiff and the differences were all 0, so I'm pretty sure that this is not necessary, but I did it anyways.

I also used this script on the E-OBS target data. Make sure the data is on the same grid and that the variables are renamed, so that xarray can read the files.

Running the models

• Baselines:
  • Rename config_baseline.ini to config.ini
  • Adjust model to the baseline model you want to run (see baseline.py for available model names)
  • Run python3 baseline.py
• ConvMOS:
  • Adjust config.ini to your liking
  • Run python3 run.py

Running the models in Kubernetes

We use the following models:

• Lin (Local Linear Regression):
  • Adjust config_template_remo_prec_baseline.ini so that aux_variables is empty (if necessary)
  • Run start_k8s_job.sh my_lin_run -m linear
• NL PCR (Non-local Principal Component Regression):
  • Adjust config_template_remo_prec_baseline.ini so that aux_variables contains all variables (if necessary)
  • Run start_k8s_job.sh my_nlpcr_run -m linear-nonlocal
• NL RF (Non-local Random Forest):
  • Adjust config_template_remo_prec_baseline.ini so that aux_variables contains all variables (if necessary)
  • Run start_k8s_job.sh my_nlrf_run -m rf-nonlocal
• ConvMOS:
  • Adjust config_template_remo_prec.ini so that aux_variables contains all variables (if necessary)
  • Run start_k8s_job.sh my_convmos_run -m convmos -a gggl -s
    • gggl specifies the module composition