T4 - Development of a deep learning prototype for operational probabilistic wind gust forecasting
Project leader: Dr. Sebastian Lerch
Other researchers: Benedikt Schulz (PhD student), Jieyu Chen (PhD student)
Probabilistic weather forecasts are nowadays obtained from statistically corrected (or postprocessed) output of numerical weather prediction (NWP) models. In light of the increasing importance of accurate and reliable weather forecasts for protecting human life and property and for economic applications such as renewable energy generation, statistical postprocessing has been a focus of research interest at the intersection of statistics and atmospheric sciences. Over the past decades, a large variety of methods have been developed and successfully applied in research and operations. A particular focus has been on on probabilistic forecasts that provide information about forecast uncertainty. Estimation and communication of forecast uncertainty is especially important for forecasts of hazardous weather conditions such as severe wind gusts, which occur quite often in both Germany and the Netherlands, and can lead to damage in society.
In the past years, a key focal point of research interest in statistical postprocessing has been the use of modern methods from artificial intelligence and machine learning. These techniques have the potential to overcome inherent limitations of traditional postprocessing methods and show promising results in case studies, in particular as they offer ways to include spatial, temporal and physical information into postprocessing models and to flexibly model nonlinear relations between arbitrary predictors and target variables. Deep learning methods based on neural networks that were proposed for postprocessing in research from Phase 1 of W2W (Rasp and Lerch, 2018) have received a lot of research interest and have been adopted in a variety of research studies.
Despite the promising research results, hardly any of the novel machine learning methods have been used in an operational context at meteorological services thus far, and a number of methodological and practical challenges remain for a successful transfer from research to operations. The overarching aim of this Transfer Project thus is to combine expertise and experience from KNMI and W2W to help overcome these impediments. With a focus on improving probabilistic forecasts of severe wind gusts, we will develop a deep learningbased postprocessing models tailored to the specific needs of operational weather prediction at KNMI. The postprocessing model will be able to provide probabilistic predictions of threshold exceedances at relevant warning levels and arbitrarily defined regions. In the second year, the developed model will be implemented in a prototype forecast product at KNMI, where forecasters will be trained to use it during their shifts in a test phase. The implementation of ’explainable AI’ methods will help increase trust in the forecasting system as forecasters can see the factors leading to a specific forecast, as well as shedding light on deficiencies in the raw numerical weather prediction model.
