Skillful Twelve Hour Precipitation Forecasts using Large Context Neural\n Networks

Abstract

The problem of forecasting weather has been scientifically studied for\ncenturies due to its high impact on human lives, transportation, food\nproduction and energy management, among others. Current operational forecasting\nmodels are based on physics and use supercomputers to simulate the atmosphere\nto make forecasts hours and days in advance. Better physics-based forecasts\nrequire improvements in the models themselves, which can be a substantial\nscientific challenge, as well as improvements in the underlying resolution,\nwhich can be computationally prohibitive. An emerging class of weather models\nbased on neural networks represents a paradigm shift in weather forecasting:\nthe models learn the required transformations from data instead of relying on\nhand-coded physics and are computationally efficient. For neural models,\nhowever, each additional hour of lead time poses a substantial challenge as it\nrequires capturing ever larger spatial contexts and increases the uncertainty\nof the prediction. In this work, we present a neural network that is capable of\nlarge-scale precipitation forecasting up to twelve hours ahead and, starting\nfrom the same atmospheric state, the model achieves greater skill than the\nstate-of-the-art physics-based models HRRR and HREF that currently operate in\nthe Continental United States. Interpretability analyses reinforce the\nobservation that the model learns to emulate advanced physics principles. These\nresults represent a substantial step towards establishing a new paradigm of\nefficient forecasting with neural networks.\n