Local rebalancing
© Siemens AG
Intelligent, automated management and control mechanisms as well as reliable load and generation forecasts are essential for distribution grids in highly decentralized energy infrastructures. Siemens provides scalable systems to automate, display, and control all elements in grids that are increasingly challenged by the medium- and low- voltage infeed of power from intermittent renewable sources. Innovative solutions such as EnergyIP DEMS for the management of demand response and virtual power plants as well as trendsetting energy storage solutions help balance generation and load to maintain and enhance the reliability and availability of distribution grids.

AI Increases Accuracy of Forecasting Models

For a power grid to be stable, the amount of energy fed into it and the amount of energy drawn from it must match. Reliable forecasts are needed to plan this. In the AGENS project, we are using Artificial Intelligence (AI) to develop algorithms that optimize such forecasts.

AGENS: Neural Networks Predict Energy Consumption

BMBF Project AGENS – Analytical-Generative Networks for System Identification

Germany will probably consume around ten percent more electricity in 2030 than previously assumed – announced the Federal Ministry of Economics in July 2021. Energy forecasts provide an idea of the future development of energy supply so that better planning can be made. The forecasts are created with the help of mathematical models that depict complex relationships as realistically as possible.

In the AGENS project, which is funded by the German Federal Ministry of Education and Research (BMBF), we are developing smart forecasting models together with partners from industry and research. The basis of this: Using the latest mathematical methods based on neural networks (NN), the algorithms are able to predict overall complexity based on large amounts of data.

In the electricity market in particular, there are crucial uncertainty factors in the correct forecasting of energy demand, because it depends on the feed-in from renewable energy sources, which fluctuate greatly from season to season – wind and solar energy are dependent on the weather. Together with classic thermal power plants, they supply the electricity in Germany.

To ensure that energy demand can be met, energy companies (balancing group managers, or BKVs for short) must submit a forecast of energy demand and generation for the coming day in their area of responsibility. This is transmitted to the associated network operator with quarter-hour accuracy in order to ensure network stability. This means that the better the models and forecasts, the better the companies can plan.

Use Case: Energy Company EWE

We are developing such a model for forecasting the energy demand of individual companies on the basis of time-recurrent neural networks (RNNs). These provide multiple possibilities for modeling complex dynamic behavior patterns based on time series. The large number of energy consumers provides a huge amount of data, optimal for training Artificial Intelligence. Currently, all commercial customers with a consumption > 100 MWh per year are subjected to a recording power metering (RLM) and an individual consumption forecast is generated for each of them – also at the project partner EWE. EWE AG is a utility company in the field of electricity, natural gas, telecommunications and information technology with headquarters in Oldenburg.

The benchmark for the procedures developed in the project is a standard forecasting procedure that is already in use for day-ahead forecasts at many energy suppliers and also at EWE. In this so-called comparison day procedure, the companies evaluate the consumption on comparable days in a certain time window (usually three to four weeks) for the day to be forecast.

In order to be able to map the complexity with mathematical models, flexible data-driven solutions are necessary. At EWE, a moving average model (MA) is currently used to predict consumption values. We are putting this to the test and optimizing it.

Our Expertise: Neural Networks and Time Series in Focus

Our »Financial Mathematics« team not only brings knowledge from energy industry practice, but also on the modeling of electricity prices based on global demand models and risk management in the electricity market. In particular, our expertise in time series analysis is used in the project. The team takes a leading role in the statistical analysis of electricity demand data.

Specifically, we are developing algorithms and a reference model for energy forecasting. The approaches are based on artificial Neural Networks that emulate the human brain. They are trained with a large amount of data and thus learn to predict more reliably. In cooperation with EWE and Siemens, we are evaluating the method with new data collected by EWE in the course of the project against the defined criteria. Time series analysis methods also look at changes over time. They can detect seasonalities, for example.

Save Costs and Risk for Energy Companies

Through the synergies of research with the application companies EWE and Siemens, the project creates a transfer of novel mathematical methods to the energy industry. Mathematics plays a key role here as a cross-sectional science.

The resulting models then serve the companies directly as decision support in practice. For them, optimizing their own forecast quality means reducing the compensation amounts to be paid and, above all, reducing the associated extreme risk of Value at Risk (VaR). A realistic improvement of the existing forecast models by around 40 percent would already mean a reduction in VaR of a high seven-digit euro amount per year – for the test region considered alone.

Actual energy consumption of an EWE customer company measured in MWh (in red) compared with the two forecast models. Blue are the calculations of our reference model developed in the project and black are the results of the forecast model currently in use at EWE.
© Fraunhofer ITWM
Actual energy consumption of an EWE customer company measured in MWh (in red) compared with the two forecast models. Blue are the calculations of our reference model developed in the project and black are the results of the forecast model currently in use at EWE.

Our Partners in Research and Industry

  • Technical University Kaiserslautern, Workgroup Financial Mathematics, Department of Mathematics (Coordination Prof. Dr. Ralf Korn)
  • University of Bremen, Center for Technomathematics, Faculty of Mathematics/Computer Science
  • Aschaffenburg University of Applied Sciences, Faculty of Mathematics/Computer Science/Technomathematics
  • Siemens AG, Digital Industries, Solution Operations; Bremen
  • EWE AG, DataLab; Oldenburg

 

Project Duration

01.04.2020 until 31.03.2023