This record contains data for the manuskript "Exploring decomposition of temporal patterns to facilitate learning of neural networks for ground-level daily maximum 8-hour average ozone prediction" by L. H. Leufen, F. Kleinert and M. G. Schultz.

We provide the complete experiment folders of the best trained networks (_network_daily). These contain the data used (data), the forecasts created (forecasts), the neural network used (model), graphics about the data and the evaluation (plots) as well as the exact results of the error analysis (latex_report). In addition, each experiment folder contains a start script (start_script.txt) with disabled train options, which can be used to restart the evaluation without overwriting or training the model. We have added an example file (run_Example.sh) how such a call could look like. To start a new experiment it required to update the start script accordingly. Furthermore, we provide the results of the uncertainty estimate of the MSE (analysis_data, stored in leufen-data-0.tar.gz) for the FCN experiments but also for the follow-up experiments with different NN architectures (FCN, RNN, CNN). All mentioned data are grouped and packed into several tar.gz archives (leufen-data-.tar.gz).

To have a better insight into the data and experiment, we provide a ready-to-run jupyter notebook to load and visualize our data and results. To run this notebook we rely on docker. Download the docker file (leufen-docker.tar.gz) and follow the instructions (instructions.md and instructions.pdf) to load data, models, and the notebook. Note that changes made by the user to the notebook well be removed on exit as long the option "--rm" is present. When following the instructions, it is not required to unpack the data file (leufen-data-*.tar.gz). If you encount issues with disk space limits and docker, it is always possible to use a reduced number of data files or to unpack data for a single experiment and just parse them to the docker container. When using a windows host system, some commands provided in the instructions might slightly deviate.

All experiments were carried out with the software MLAir in version 2.0.0 . A detailed description is given in https://doi.org/10.5194/gmd-14-1553-2021 and the source code can be found at https://gitlab.jsc.fz-juelich.de/esde/machine-learning/mlair .