Data-Driven Medium-Range Weather Prediction With a Resnet Pretrained on Climate Simulations: A New Model for WeatherBench

Numerical weather prediction has traditionally been based on the models that discretize the dynamical and physical equations of the atmosphere. Recently, however, the rise of deep learning has created increased interest in purely data-driven medium-range weather forecasting with first studies exploring the feasibility of such an approach. To accelerate progress in this area, the WeatherBench benchmark challenge was defined. Here, we train a deep residual convolutional neural network (Resnet) to predict geopotential, temperature and precipitation at 5.625 degrees resolution up to 5 days ahead. To avoid overfitting and improve forecast skill, we pretrain the model using historical climate model output before fine-tuning on reanalysis data. The resulting forecasts outperform previous submissions to WeatherBench and are comparable in skill to a physical baseline at similar resolution. We also analyze how the neural network creates its predictions and find that, for the case studies analyzed, the model has learned physically reasonable correlations. Finally, we perform scaling experiments to estimate the potential skill of data-driven approaches at higher resolutions.

Automated summary

The traditional numerical weather prediction is based on discretizing the dynamical and physical equations of the atmosphere, while the rise of deep learning has brought increased interest in purely data-driven medium-range weather forecasting. Through the WeatherBench benchmark challenge, we trained a deep residual convolutional neural network to predict geopotential, temperature, and precipitation up to 5 days ahead, resulting in forecasts outperforming previous submissions and comparable to a physical baseline. Analysis shows the model has learned physically reasonable correlations, and scaling experiments have been conducted to estimate the potential skill of data-driven approaches at higher resolutions.