Operational SYM-H forecasting with confidence intervals using Deep Neural Networks

Authors

Armando Collado-Villaverde, Pablo Muñoz and Consuelo Cid

This repository contains the predictions made by the model in the Operational SYM-H forecasting with confidence intervals using Deep Neural Networks paper on the test and test key storms and using the model in https://doi.org/10.1029/2023SW003485 as comparison.

File Descriptions

• generate_graphs_and_metrics.ipynb: Jupyter notebook for reading the predictions, calculating the metrics and generating the graphs.
• metrics.py: functions for calculating the metrics and generating the graphs.
• storm_dates.py: dates for all the subsets.
• data/: Folder containing the predictions.
• figs/: Folder containing generated figures.
• figs_SI/: Folder containing the figures for the supplementary information.

Contact

For any questions or issues, please open an issue on the GitHub repository or contact the author.

DNN hyperparams

Input timesteps: 48 of 5 minutes averages. Output: SYM-H forecast at the next 1 or 2 hours and the 5% and 95% Quantile forecasts. Same hyperparams are used for the 1 and 2 hours models.

• Top convolutional layers:
  • kernel size: 7
  • filters: 128
  • activation function: linear, swish, swish
• Multi-head attention layers:
  • num heads: 4
  • key and value dim: 128
  • dropout: 0.2
  • bias: True
• LSTM layers:
  • units: 128
• Top output dense layers:
  • units: 512
  • activation: swish
• Bottom output dense layers:
  • units: 1
  • activation: linear

Forecasted SYM-H ouput layer trained with MSE and the quantile layers with the Quantile loss in the quantile_loss.py file setting q at 0.05 and 0.95.

Trained with the AdaBelief optimizer (https://www.tensorflow.org/addons/api_docs/python/tfa/optimizers/AdaBelief) on TensorFlow 2.14

---