/ResNet_Weather

This repository contains code to train a direct residual convolutional neural network on weather data.

Primary LanguageJupyter Notebook

ResNet_Weather

This repository contains the code for the following paper:

Mariana C. A. Clare, Omar Jamil, Cyril Morcrette, Using deep learning to produce a computationally efficient probabilistic weather forecast through the prediction of probability density functions, Quarterly Journal of the Royal Meteorological Society.

This code trains a direct convolutional ResNet on continuous weather data to create a probabilistic weather forecast. It does so by transforming the continuous data to categorical through binning so that a SoftMax layer can be used to predict a probability density function for each output rather than a single value. The variables predicted are the geopotential at the 500hPa level and the temperature at the 850hPa level.

The repository also contains code to train a stacked neural network to combine outputs from different neural networks.

**Acknowledgements

This code has used the DataGenerator and Padding (for the CNN) from the start notebooks in this repository. The dataset used is the WeatherBench dataset [1] which is hosted here and more information about how to download the data, as well as the dataset itself, can be found at this repository. In the folder Data Exploration, this dataset has been visualised and explored.

[1] Stephan Rasp, Peter D. Dueben, Sebastian Scher, Jonathan A. Weyn, Soukayna Mouatadid, and Nils Thuerey, 2020. WeatherBench: A benchmark dataset for data-driven weather forecasting. arXiv: https://arxiv.org/abs/2002.00469