Research equals innovation

We research the future

Weather and climate forecasts, high-performance computing, and machine learning

As part of the EU’s MAELSTROM research and development project, 4cast has been passionately and consistently driving innovations in machine learning and high-performance computing with a focus on recording weather data. Generating live weather data from social media is one aspect of that. We are involved in three research initiatives under the MAELSTROM umbrella, thereby contributing to the energy transition and more sustainable energy generation.

The forecasting accuracy of weather models depends on the quantity and quality of the weather measurements used to initialize the weather models. With the help of social media users acting as “on-the-ground” weather stations, we hope to improve weather predictions. As part of an initial explorative study, we’re trying to forecast rain on the basis of tweets. We’ve developed a classifier built on the DeBERTa architecture to leverage the text content of tweets in order to predict whether it’s “raining” or “not raining” in the tweeter’s location at the time of their post.
We initialize our model with pretrained weightings and adjust it for our specific application. Our best model achieves an F1 score of 0.66 for the minority class (rain) with an AUC of 0.77.

Macro weather situations are large-scale weather patterns over a particular country. These can provide insights into local weather and the assumption is that the energy yields of wind and solar parks can also be deduced in this way. As part of the MAELSTROM project, we’re researching precisely that hypothesis and trying to determine whether macro weather situations can help us make our forecasts even more accurate.

High-performance computing is a great way of running highly complex calculations at record speed. However, access to supercomputers is limited, and the work steps required of machine learning scientists are extremely complicated and arduous. Furthermore, researchers in the field of weather and climate physics work with enormous volumes of data, and processing this data slows down the performance of machine learning processes.

We work on an online platform where users can develop machine learning projects and discuss them with other users. The platform provides an overview of all the steps in the workflow: data, coding, experiments, and trained models. Since the platform is connected directly to the interfaces of the supercomputers, models can be trained and used for forecasts and other applications straight from a browser. The results from training or applying the models are displayed on the platform in real time.