Sub-Seasonal Prediction With A Global Deep-Learning Weather Prediction Model

Abstract:

We present an ensemble weather prediction system based on a global deep-learning weather prediction (DLWP) model. The model is trained using ECMWF ReAnalysis 5 (ERA5) data and convolutional neural networks (CNNs) on a cubed-sphere grid. Our ensemble includes perturbations to both initial-conditions and model-physics. The initial-condition perturbations are drawn from the set of 10 perturbed ERA5 members and do not produce optimum initial spread. Stochastic model-physics perturbations are produced by randomizing the training process for the CNNs. Our 320-member grand ensemble consists of each of 32 slightly different CNNs run from each of the 10 initial conditions. At lead times up to two weeks, our DLWP ensemble lags the state-of-the-art 50-member ECMWF numerical weather prediction (NWP) ensemble by 2-3 days of forecast lead time. Despite the non-optimal set of initial conditions, our ensemble is just slightly under-dispersive.

Extending forecasts to the sub-seasonal forecast range of 2-6 weeks, a particularly challenging forecasting window, our DLWP ensemble consistently outperforms persistence forecasts of 850-hPa temperature (T850) and 2-m temperature, with skill relative to climatology as measured by the ranked probability skill score (RPSS) and the continuous ranked probability score (CRPS). Evaluating performance for twice-weekly forecasts over the period 2017-18, our CRPS matches that of the ECMWF ensemble to within 95% confidence bounds for T850 at week 4 and weeks 5-6. Our ensemble system, which predicts six 2D shells of atmospheric data at approximately 1.4 x1.4-degree-spatial and 6-hour-temporal resolution has a huge computational speed advantage over traditional NWP models; a 6-week forecast with 320 ensemble members runs in about half a second on a single graphics processing unit.

There are many potential improvements to our relatively simple DLWP ensemble system that might be expected to both yield better forecasts and require more computational effort. It remains to be seen how well models from the rapidly developing field of machine-learning-based weather forecasting will ultimately compare to the much more mature products generated with conventional NWP, but these early results are encouraging.