Data‐Driven Medium‐Range Weather Prediction With a Resnet Pretrained on Climate Simulations: A New Model for WeatherBench
Thuerey, Nils
DOI: https://doi.org/10.1029/2020MS002405
Persistent URL: http://resolver.sub.uni-goettingen.de/purl?gldocs-11858/9521
Abstract
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° 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.
Plain Language Summary: Weather forecasts are created by running hugely complex computer simulations that encapsulate our knowledge of how the atmosphere works. This approach has served us well but is there a different way? The paradigm of machine learning proposes learning an algorithm from data rather than building it from physical principles. For several areas like computer vision and natural language processing this has worked exceedingly well, so it just makes sense to try it as well for weather forecasting. This paper presents the latest attempt at training a machine learning weather forecasting model. It is shown that the learned model produces reasonable forecasts, approximately on par with traditional models run on much lower resolution. However, there is still a large gap to current state–of–the–art high–resolution weather models that is unlikely to be closed with a purely data–driven approach because not enough training data exists.
Key Points:
A large convolutional neural network is trained for the WeatherBench challenge.
Pretraining on climate model data improves skill and prevents overfitting.
The model sets a new state‐of‐the‐art for data‐driven medium‐range forecasting.