Improving numerical weather forecasts with artificial neural networks
New study shows how machine learning can be used to remove systematic errors from numerical weather forecasts
All numerical weather forecasts have systematic errors. Before they can be delivered to end users, the forecasts need to be statistically calibrated. Traditionally this has been done with simple linear regression techniques. In this study a modern artificial neural network is used. Neural networks provide a flexible way to model non-linear relations between different forecast variables and other properties specific to each weather station. The neural network approach outperforms all previous state-of-the-art approaches while being more computationally affordable. This study shows the potential of neural network for the field of postprocessing in general.
Paper: Neural Networks for Postprocessing Ensemble Weather Forecasts. Rasp, S. and S. Lerch, 2018. Monthly Weather Review. https://doi.org/10.1175/MWR-D-18-0187.1