Skip to content

Latest commit

 

History

History
69 lines (47 loc) · 5.84 KB

README.md

File metadata and controls

69 lines (47 loc) · 5.84 KB
Raw

ESSD-DRN

This repository provides Python code for the implementation of the distributional regression network (DRN) post-processing method, accompanying the ESSD benchmark. Provide the DRN output file (see the method's details below).

This code is provided as supplementary material with:

  • Demaeyer, J., Bhend, J., Lerch, S., Primo, C., Van Schaeybroeck, B., Atencia, A., Ben Bouallègue, Z., Chen, J., Dabernig, M., Evans, G., Faganeli Pucer, J., Hooper, B., Horat, N., Jobst, D., Merše, J., Mlakar, P., Möller, A., Mestre, O., Taillardat, M., and Vannitsem, S.: The EUPPBench postprocessing benchmark dataset v1.0, Earth Syst. Sci. Data, 15, 2635–2653, https://doi.org/10.5194/essd-15-2635-2023, 2023.

Please cite this article if you use (a part of) this code for a publication.

Details on the DRN setup

The DRN approach proposed by Rasp and Lerch (2018)1 is a neural network (NN) based method where the distribution parameters of the post-processed forecast distribution are obtained as the output of a NN. Our implementation for EUPPBench closely follows Rasp and Lerch (2018)1 and uses a NN with one hidden layer of 512 nodes. The input predictors of our NN model are listed in the following table. All predictors except for the date information and the embedding are normalized to the range [0,1] using a min-max scaler before training. We assume a Gaussian distribution for the post-processed temperature forecasts, and the NN model returns mean and standard deviation of the distribution as outputs. For each lead time, a separate NN model is estimated jointly for all stations, using the CRPS as a custom loss function. The model predictions are made locally adaptive by using embeddings of both the station identifiers and the summarized land usage information. We repeat the model estimation 10 times and take the average over the outputs of all runs as the final parameters of the post-processed distributional forecasts.

Predictor Description
t2m_mean Mean of raw 2-m temperature ensemble forecasts
t2m_std Standard deviation of raw 2-m temperature ensemble forecasts
model_altitude Altitude of the model grid point given by a high-resolution (100m) Digital Elevation Model
model_latitude Latitude of the model grid point given by a high-resolution (100m) Digital Elevation Model
model_longitude Longitude of the model grid point given by a high-resolution (100m) Digital Elevation Model
station_altitude Altitude of the weather station
station_latitude Latitude of the weather station
station_longitude Longitude of the weather station
model_orog model orography, i.e. the average grid box altitude in the model
doy Sine-transformed value of the day of the year
Predictor for embedding Description
station_id Identification number of weather station in the dataset
model_land_usage Land usage of the model grid point
station_land_usage Land usage of the weather station
land_usage Summarized land usage information combining both model_land_usage and station_land_usage, using a mapping from the 2D vector (model_land_usage, station_land_usage) to the 1D scalar land_usage

Our NN model is built using the Keras2 framework in Python. The specific hyper-parameter choices for our model are shown in the following table:

Hyper-parameter of the NN model Value
Embedding size of station_id 2
Embedding size of land_usage 4
Optimizer Adam
Learning rate 0.005
Batch size 4096
Early stopping With patience of 2 epochs and minimum delta of 0.005 on monitoring the training loss
Maximum number of training epochs 50

Data and implementation details

The scripts are built for the ESSD benchmark dataset, which can be downloaded following the instructions on the GitHub repository.

Data pre-processing

In order to run our DRN scripts on the ESSD benchmark dataset we had to adapt the data format. First, we combined reforecast and observation data and saved it to separate feather files for training and testing. The respective code is in train_netCDF_to_feather.ipynb and test_netCDF_to_feather.ipynb in the /data-preprocess/ folder. The transformed .feather files include forecast information for all 21 lead times. We further split them regarding different lead time using the data_preprocess.py script from the /data-preprocess/ folder.

Training and producing predictions of the DRN model

The DRN model can be run by executing the DRN_pp.py script in the home directory. Note that file paths have to be changed accordingly.

The output of the NN models are the parameters of a Gaussian distribution. The scripts in the /output-process/ folder are used to obtain the final predictions in the required format. output_process.py adapts the format of the predictions to prepare for the following steps: First, generate_fcst.R generates samples from the distribution as the post-processed ensemble forecasts for each lead time in the form of 51 equi-distant quantiles. Then, combine_fcst.R combines the post-processed forecasts for all lead times.

The predictions are transformed to the required netCDF format and relevant meta-data is added in feather_to_netCDF.ipynb.

Computational performance indication

The computation was done with 45 CPUs (Intel(R) Xeon(R) CPU E5-2680 v3 @ 2.50GHz) with 236 Gb of RAM. Training the 21 DRN models in the setup described above for all lead times takes approximately 2 hours in total.

Footnotes

  1. Rasp, S., and Lerch, S. (2018). Neural Networks for Postprocessing Ensemble Weather Forecasts. Monthly Weather Review 146, 11, 3885-3900, available from: https://doi.org/10.1175/MWR-D-18-0187.1 [Accessed 19 December 2022] ↩ ↩2

  2. https://keras.io/ ↩