Towards replacing precipitation ensemble predictions systems using machine learning

Forecasting precipitation accurately poses significant challenges due to various factors affecting its distribution and intensity, including but not limited to subgrid variability. Although higher resolution simulations are often considered to improve precipitation forecasts, it is crucial to note that simply increasing resolution may not suffice without appropriate adjustments to parameterization schemes or tuning. Traditionally, ensembles of simulations are used to generate uncertainty predictions associated with precipitation forecasts, but this approach can be computationally intensive. As an alternative, there is a growing trend towards leveraging neural networks for precipitation prediction, which offers potential computational advantages. We propose a new approach to generating ensemble weather predictions for high-resolution precipitation without requiring high-resolution training data. The method uses generative adversarial networks to learn the complex patterns of precipitation and produce diverse and realistic precipitation fields, allowing to generate realistic precipitation ensemble members using only the available control forecast. We demonstrate the feasibility of generating realistic precipitation ensemble members on unseen higher resolutions. We use evaluation metrics such as RMSE, CRPS, rank histogram and ROC curves to demonstrate that our generated ensemble is almost identical to the ECMWF IFS ensemble, on which our model was trained on. We propose a novel method for generating precipitation ensemble members from deterministic weather forecasts. Prediction is done using a generative adversarial network in an image-to-image style. The neural networks is trained on low-resolution data but can be applied on unseen higher resolution data. image