Speaker
Description
For about two decades, operational postprocessing of NWP output for short-range and medium-range forecasts at Deutscher Wetterdienst (DWD) has been mainly based on linear/logistic regression Model Output Statistics (MOS) approaches. However, Artificial Intelligence (AI) has nowadays become a central key for many modern applications and research areas, and it also gains more and more importance in weather and climate related studies. Consequently, we want to investigate its potential also for postprocessing in an operational weather forecasting environment based on NWP output.
In this study, we start first experiments using machine learning approaches based on neural networks for postprocessing the output of the limited area short-range ensemble forecasting NWP model COSMO-DE-EPS with a resolution of 2.8km. For the beginning, we focus on postprocessing wind forecasts for a 2° x 2° (220 km x 140 km) region around the Harz mid-mountain range in Germany. For training the neural network we have hourly forecasts from a COSMO-DE-EPS archive for the years 2011-2018 with 63 surface and 7 pressure level variables (features) and local Synop station observations (labels).
First results indicate that dimensionality reduction using different regularization techniques turn out to be a challenging issue in order to avoid overforecasting for this configuration. Another challenge is the shape and tuning of the neural network, e.g. we apply various configurations for the number of hidden layers and neurons trying to improve the stability of the results. We also focus on computing individual feature importance using the neural network in order to gain better insight into the role that individual COSMO-DE-EPS variables play for the performance of the postprocessed forecast.
Our first results indicate the potential of the approach; a further step will however be to obtain a stable output allowing a direct comparison to the quality of our traditional MOS approaches.
