Renewable energy research: We are shaping the future
Weather and climate forecasting, high-performance computing and machine learning
In the European research and development project Maelstrom, 4cast is consistently focusing on innovations in the areas of machine learning and high-performance computing. The focus is on collecting and analyzing weather data, including the generation of live data from social media. With three pioneering projects, 4cast is making a valuable contribution to the energy transition and more sustainable energy generation. In this way, we are driving renewable energies forward through research.
Advancing renewable energy research–this is only possible if precise weather data is available. The predictive power of weather models depends largely on the quality and quantity of weather measurements for initialization. An innovative approach is the use of social media as citizen weather stations to improve weather forecasts.
In an initial exploratory study, we investigated whether rain events can be predicted from tweets. For this purpose, a classifier based on the DeBERTa architecture was developed, which analyzes the text content of tweets to determine whether it is “raining” or “not raining” at the location and time of the tweet. The model was initialized with pre-trained weights and optimized specifically for this application.
Our best model achieved an F1 score of 0.66 for the minority class (rain) and an AUC value of 0.77. Such innovative approaches show how research can advance renewable energy by integrating modern technologies and new data sources.
Large-scale weather patterns are large-scale weather structures over Germany. These can allow conclusions to be drawn about the local weather and therefore – it is assumed – also about the energy yield of wind and solar parks. As part of Maelstrom, we are researching precisely this issue and trying to find out whether large-scale weather patterns can help us to improve our forecasts even further.
High-performance computing is a great way to perform highly complex calculations in the shortest possible time so that research can advance renewable energies and innovative solutions can be developed more quickly. However, access to supercomputers is limited and the work steps for machine learning scientists are very complex and tedious. In addition, researchers in weather and climate physics work with extremely large amounts of data, the processing of which often forms the bottleneck in the performance of machine learning processes.
To address these challenges, we are developing an online platform that enables users to develop and share machine learning projects. The platform provides an overview of all steps of the workflow: from data management and code to experiments and trained models. A direct connection to the interfaces of supercomputers makes it possible to start model training and application, for example for precise predictions, directly from the browser. Results from training or applications appear on the platform in real time, which speeds up and optimizes the entire process.
This approach demonstrates how research can promote renewable energy by simplifying access to cutting-edge technology while increasing the efficiency of machine learning processes.
