The Competence Center for Cognitive Energy Systems (K-ES), founded in Kassel in 2020, is investigating how artificial intelligence (AI) can be applied in the energy sector. For example, it is demonstrating how intelligent agents will be able to independently control plants and grids in many areas in the future. This possibility is particularly interesting when dealing with weather-dependent energy generation. In the future, forecasting models based on artificial intelligence will be able to operate much more accurately than models using current methods.

Prof. Dr. Kurt Rohrig, Head of the Fraunhofer Institute for Energy Economics and Energy System Technology IEE, emphasizes the need for intelligent digitalization covering the entire system: “A decision-making process in a decentralized system will be significantly more complex than the ones we know at the moment, as the number of input variables is far greater. Only by using artificial intelligence will we be able to operate different, complex systems together, such as electricity and heat supply systems as well as mobility systems via automated decisions. This requires robust processes running in real-time.”

Agents learn how to trade electricity and use generation plants

AI offers practical applications in a whole variety of areas in the energy sector. In the Deep Energy Trade project, a demonstrator shows how intelligent automated electricity trading can work. An agent independently learns how to identify trading strategies and trigger purchases or sales. Deep Reinforcement Learning (DRL) reduces costs and simplifies the trading process. It also facilitates access for market participants who only want to trade small volumes of energy on the market.

AI can also be used to optimize the deployment of power plants. Deployment can be intelligently and automatically planned in the short term in the energy management of renewable plant portfolios, such as solar panel installations and wind turbines in combination with storage technologies, and of flexible consumers, such as electrolyzers or electric vehicles. The Cognition²H₂Force project is investigating the application of DRL for this purpose. To this end, power-to-gas plants for producing and storing hydrogen are being combined with wind turbines and storage systems. The aim is to develop a digital twin of such a portfolio in which self-learning optimization algorithms can be trained, and will then be ready to plan deployment autonomously.

The ARCANA project deals with fault management. Until now, wind turbine operators have been using automated predictive maintenance. The research is now going one step further and focusing on error diagnostics using DRL. As a result, the AI system not only acts as an alarm but also specifies the causes that led to an outage and how to remedy them.

AI makes irradiation forecasts more accurate

The NeuRaSat project improves the accuracy of irradiation forecasts using satellite data. Accurately forecasting solar irradiation and cloud behavior is essential to defining expected power generation, energy trading and the safe operation of power grids. This process focuses in particular on cloud position, movement and density.

The aim of the Temporal Fusion Transformers (TFT) project is to demonstrate wind forecasting. The large volume of data can be processed in a model that can calculate a probabilistic power forecast for any given wind farm. It does so by pooling spatio-temporal dependencies that are available for different locations and times. Wind speed data from one measurement location is incorporated into forecasts for other locations.

Grid operators need reliable forecasting for operations just as much as power generators. Probabilistic grid condition forecasting considers different weather models and assesses their probability using a neural network. This also includes improbable but highly relevant events. AI can also be used to make more precise estimates of network topology. The Vertical Load Forecasting project applies Deep Learning using a method based on a Long-Short Term Memory (LSTM) ANN.

Grid operation as a simulation training task

The AI OPF project investigates whether DRL is suitable for eliminating network congestion and for flexibility calls, taking forecasting uncertainties into account. To do so, a grid simulator is used to train the system. The application is aimed at new regulations for the lower voltage range. Distribution system operators can use it to support their system control in line with the Redispatch 2.0 specifications.

In the CTRL (Cognitive Train/Test System for Reinforcement Learning using Labs) project, an agent for controlling a local grid transformer is used to demonstrate how voltage stability can be automated. This involves agents gradually preparing for their areas of application, starting with simulation pretraining, followed by realistic hardware-in-the-loop methods, and ending with field tests.

New methods for detecting microgrids with AI support are being developed in the AIsland project. These methods are sensitized in a particular way so as to detect unwanted microgrids in grid structures with a high converter penetration.

It is already conceivable that the entire power grid will one day operate automatically. As part of the international competition “L2RPN Challenge” (“Learn to Run a Power Network”), launched by the French transmission system operator RTE, a self-learning agent demonstrated that fluctuating inputs and loads can be used to bypass maintenance work and malicious attacks. The agent developed as part of the KESL2RPN project took fifth place out of 50 international participants.

AI can also drive the development of new devices. Specifically, in the InvEx project, K-ES scientists are investigating whether an expert system can be used to shorten the development time for converters. A database for converter parts and components, topologies and control algorithms operates at the core of the AI application.

A cognitive system can independently determine its own state and that of its assets on the basis of the information available and learn to achieve predetermined goals independently through the ability to adapt. Cognitive energy systems represent a key technology in the energy transition. Applications in the electricity sector can be found in the field of grid management and the management of generation and consumption.