Self-learning forecasting methods take changes in the energy system into account

"Our forecasting methods will do far better justice to the increasing dynamics on the generation and consumption side than the traditional methods," says Dominik Beinert, Co-Project Leader at Fraunhofer IEE. "We are thus providing grid operators, direct marketers, plant operators and other stakeholders with powerful tools that they can use to optimize flexibility in the energy system, for example."

The research project called "KonSEnz" - the acronym stands for continuously self-learning forecasting methods and services in smart energy markets and grids - will run for three years. Fraunhofer IEE is leading and coordinating the project. In addition to the University of Kassel and Wobben Research and Development, ENERCON's research company, the transmission system operators Amprion, TenneT and 50Hertz are involved as associated partners. The project is funded by the German Federal Ministry for Economic Affairs and Climate Protection.

Methods use continuous, adaptive learning processes

Forecasting models must be trained again and again in order to be able to provide meaningful predictions on a permanent basis. The methods used to date require this process to be triggered manually each time. This requires a great deal of effort. Above all, however, the forecast quality suffers: because the training is discontinuous, such static models cannot reflect the dynamics of the energy system. However, developments such as the rapid expansion of photovoltaics or the increasing number of flexible consumers must be taken into account in the forecasts without delay in order to achieve reliable results.

In the KonSEnz project, the partners are therefore developing innovative digital methods that immediately and automatically integrate changes on the generation and consumption side, including their interactions, into the forecasts. The researchers are using continuous, adaptive learning methods for this purpose. The methods ensure that the models are updated continuously so that they can react quickly to changing situations. This significantly improves the forecast quality.

The behavior of newly installed generators and consumers for which no data is yet available is simulated in an individually adapted manner using existing systems. The project partners are relying on the concept of machine learning operations (MLOps) to automatically transfer the model update to operations.

Forecasting methods as scalable microservices

The project partners are basing the development of the new methods on several use cases. For example, in the field of photovoltaics: the high number of PV systems requires efficient methods in order to meet the requirements of operational forecasting and training operations. The associated flexibility potential can only be exploited through reliable forecasts of supply-dependent generation and self-consumption. Another use case is the prediction of power flows between the high-voltage and extra-high-voltage grid, in which constant changes to the switching states in the grid must be taken into account. This enables grid operators to reliably forecast possible overloads of equipment, for example. The researchers want to check and demonstrate the performance of the solutions they have developed by carrying out field tests as part of the project.

The KonSEnz team will design the new, adaptive methods as scalable, resilient microservices that can be seamlessly embedded in various control, management and operating systems thanks to numerous interfaces. The microservice architecture to be developed combines continuous learning, transfer learning and machine learning operations. The services will be orchestrated in such a way that they can also cope with a very high number of forecast calculations and constant optimizations. The researchers will publish their results as an open access publication. Software components will be made available to the energy industry as open source.