Karlsruhe Institute of Technology: Better predict wind gusts with AI

In order to be able to better protect people and the environment, a more accurate forecast of extreme weather phenomena such as winter storms is essential. KIT scientists have now compared statistical and machine learning methods for forecasting wind gusts in order to make them more accurate and reliable. The inclusion of geographic information and other meteorological variables such as temperature leads to significant improvements in forecast quality, especially through modern AI methods based on neural networks.

Strong gusts of wind, such as gusts of wind with a speed of more than 65 kilometers per hour, can cause major damage and pose a danger to people, animals and infrastructure. In order to be able to issue effective warnings, early and reliable forecasts are crucial. “However, gusts of wind are difficult to model because they are driven by small-scale processes and occur very locally,” says Benedikt Schulz, doctoral student at the Institute for Stochastics at KIT. “Therefore, their predictability for numerical weather forecasting models used by weather services is limited and fraught with uncertainties.”

In order to be able to better assess such uncertainties in forecasts, meteorologists create ensemble forecasts. Starting from the current state of the atmosphere, they carry out several model calculations in parallel, each of which relates to slightly different framework conditions. In this way, they can record various scenarios about the future development of the weather. “Despite continuous improvements, these ensemble weather forecasts still show systematic errors, since local conditions, some of which vary over time, cannot be included in the models,” says Schulz. “With the help of artificial intelligence, we want to correct these systematic errors in order to improve forecasts and predict dangerous weather phenomena more reliably.”

Geographic information and other meteorological variables improve wind gust forecasts

Together with Dr. Sebastian Lerch, who heads the junior research group “AI Methods for Probabilistic Weather Forecasts” funded by the Vector Foundation at the KIT Institute of Economics, Schulz compared a large number of different statistical and AI methods for post-processing ensemble forecasts for wind gusts for the first time. “We considered both existing and new methods for statistical post-processing of numerical weather forecasts and carried out a systematic comparison of their forecast quality,” says Lerch.

It was shown that basically all post-processing methods generate reliable predictions for the speed of the wind gusts. “However, AI methods are clearly superior to classic statistical approaches and deliver significantly better results, since they allow new sources of information such as geographical conditions or other meteorological variables such as temperature and solar radiation to be better included,” summarizes Lerch. “The predictions of the AI ​​methods reduce the forecast errors of the weather models by around 36 percent on average,” says Schulz. Based on forecasts by the weather model of the German Weather Service (DWD) at 175 observation stations in Germany, the AI ​​methods delivered better forecasts at more than 92 percent of the stations than all reference models for statistical post-processing. A central role is played by the ability of neural networks to learn complex and non-linear relationships from the large amounts of data available in order to correct the systematic errors in the ensemble predictions. “By analyzing which of the information is particularly relevant for the methods, conclusions can also be drawn about meteorological processes,” says Schulz. so as to correct the systematic errors in the ensemble predictions. “By analyzing which of the information is particularly relevant for the methods, conclusions can also be drawn about meteorological processes,” says Schulz. so as to correct the systematic errors in the ensemble predictions. “By analyzing which of the information is particularly relevant for the methods, conclusions can also be drawn about meteorological processes,” says Schulz.

With their work, the researchers want to contribute to the development of methods for weather forecasting at the interface between statistics and AI. “The methods examined could be used, for example, by weather services to improve forecasts,” says Lerch. “We are in active exchange with the German Weather Service and other international weather services.”

National research for better weather forecasts

The research is part of the Collaborative Research Center/Transregio 165 “Waves to Weather” (W2W). Scientists from KIT, the Ludwig Maximilian University of Munich as the coordinator, and the Johannes Gutenberg University (JGU) are working together on a national and interdisciplinary basis to make weather forecasts even more accurate and reliable. With this, W2W is facing the current biggest challenge in weather forecasting: identifying the limits of predictability in different situations and creating the physically best possible forecast in each case.

Comments are closed.