Deep Learning for Ensemble Forecasting

3. Science Challenge While both climate and weather forecast systems have continued to improve due to substantial efforts to improve computational capabilities, observations, and numerical models, the atmosphere is a chaotic system, and this puts a fundamental limit on our ability to make predictions. Forecasts made by high-resolution models initialized with only slightly different atmospheric states can quickly diverge. Quantifying uncertainty in forecasts is essential to adequately understand them and to make the best-informed policy decisions particularly when it comes to hydrology, extreme weather (including extreme precipitation events), and climate. Currently, forecast uncertainty at the time of prediction is assessed using ensemble systems. These systems initialize multiple forecast models with slightly different initial states or different representations of atmospheric physics and use the spread of the resulting predictions to estimate uncertainty. Additionally, the ensemble mean is typically more predictive than a single deterministic forecast. Ensemble forecasts come at extreme computational cost that limits their potential applications: typical ensembles involve 5-100 members that each use the computational resources of running a single forecast model. Nonetheless, quantifying forecast uncertainty is valuable enough that many of the world’s premier weather forecasting agencies (NOAA [1] and ECMWF [2] for instance) provide operational ensemble weather forecasts. Ensembles have recently been applied to climate modeling [3] where the need for assessing confidence in climate predictions is critical, and thus the need for informing ensemble details for climate is also critical. Ensembles also have potential for use in sensitivity studies as a way to isolate changes associated with the chaotic nature of the system from changes associated with a variable of interest. More efficient methods for accurately generating ensemble spread and improving the overall ensemble forecast would be a boon for weather forecasting, climate modeling, and our understanding of atmospheric dynamics and processes. Recently, deep convolutional neural networks have been demonstrated as very effective for many tasks in the field of atmospheric science including forecasting [4], satellite retrievals [5], downscaling/super resolution [6], atmospheric state classification, and many others. They have also been successfully applied as post-processing for ensemble forecasts [7,8]. These machine learning tools are particularly well suited for problems that involve very large, gridded datasets where complex non-linear relationships must be learned, and more conventional methods may not be up to the task. They could be a powerful tool for predicting forecast uncertainty or estimating likely ensemble spread using fewer ensemble members and for understanding parameters and conditions that lead to growth of uncertainty in predictions. A Deep Learning (DL) based approach to ensemble modeling and uncertainty estimation has the potential for massive computational savings for existing ensemble systems and would allow for