Research project FEAT
Flexible, Explainable, Accurate - Machine learning in complex systems under uncertainty
The increasing complexity of energy systems requires the development of new methods for forecasting and analysing time series. Machine learning (ML) approaches in particular have proven to be promising. However, they often lack important properties such as flexibility, explainability, and accuracy. The FEAT project aims to close these gaps by developing new ML methods that can adapt to changes, deal with outliers, and new data points while showing the influence of individual components on the results. The Federal Ministry of Education and Research is funding the FEAT project to promote the development of these innovative solutions. Such methods can help to improve the predictive accuracy of time series in complex systems, like electricity markets with a high share of renewable energy, and thus support decision-making in various areas of application.
Research project FEAT | |
|---|
Duration | August 2022 to July 2025 |
Funded by | Federal Ministry of Education and Research |
Project participants | - University of Tübingen, Cluster of Excellence – Machine Learning for Science
- Institute of Networked Energy Systems
|
As part of the FEAT project, new methods for modular neural networks are being developed. These novel network architectures are adapted to different aspects of the time series to be predicted in order to take into account the special features of the respective forecast variables. New error metrics allow the explanation of results and their uncertainties. The learning tasks are optimally distributed across different networks, thereby increasing the accuracy and flexibility of the trained networks.
In the FEAT project, the Institute of Networked Energy Systems is working on the development of new methods for predicting electricity price time series in complex energy systems. The researchers are using the open-source agent-based electricity market model AMIRIS, which was developed in-house, to apply, test, and validate the methods developed. The combination of ML methods with agent-based models helps to better understand and predict complex systems. In addition, the institute is working on integrating the new methods into existing energy systems and analysing their impact on the energy industry.
The results of the FEAT project should help to improve the accuracy of electricity price forecasts and thus support decision-making in the energy industry. All methods, models, and data from the project will be made available as open-source after the end of the project. This transparency should help to transfer the results to other areas and thus accelerate the research and teaching in this field.
