Machine Learning to Improve Numerical Weather Forecasting

Abstract

This paper presents a brief overview of trends in numerical weather prediction, difficulties, and the nature of their occurrence, the existing and promising ways to overcome them. The neural network architecture is proposed as a promising approach to increase the accuracy of the 2m temperature forecast given by the COSMO regional model. This architecture allows predicting errors of the atmospheric model forecasts with their further corrections. Experiments are conducted with different histories of regional model errors. The number of epochs after which network overfitting happens is determined. It is shown that the proposed architecture makes it possible to achieve an improvement of a 2m temperature forecast in approximately 50% of cases.