Artificial intelligence increasingly decides when electricity flows, how much it costs and when electric cars charge. But so far the decisions made by the algorithms have often remained opaque. Researchers at the Karlsruhe Institute of Technology have developed a method called SHAPformer that makes AI predictions in the energy system directly comprehensible for the first time. What’s behind it and why it’s also relevant for consumers.
The management of energy supply is becoming increasingly complex. Wind and solar power fluctuate depending on the weather, while electric cars, battery storage and heat pumps change consumption patterns. Network operators and energy suppliers use artificial intelligence to operate their systems efficiently and stably.
In order to precisely coordinate electricity generation and consumption, many factors must be taken into account at the same time. This includes weather forecasts, load forecasts, network and distribution capacities as well as the behavior of consumers. However, artificial intelligence must not remain a black box in these critical infrastructures. In addition, human supervision is a regulatory requirement through the European Union’s AI Act.
How AI controls our power grids
Researchers at the Karlsruhe Institute of Technology have a method called SHAPformer that is intended to make AI decisions more transparent. It is specifically designed for time series forecasts based on consecutive data such as electricity consumption or electricity prices. The method combines transformer models, known from modern language models, with methods of explainable artificial intelligence.
The approach uses concepts from game theory to make the influence of individual factors such as temperatures, holidays, wind forecasts or previous consumption data visible. When training the model, the working group specifically ignored individual pieces of information.
This makes it possible to understand the contribution of individual influencing variables to a prediction. The team trained the system with real data from the transmission system operator TransnetBW. The aim was to predict electricity consumption and prices over periods of up to a week while displaying the influencing factors.
Why the KIT approach is more efficient than previous methods
Many previous methods only offer explanations after the fact and require additional computing power to do so. The new development from Karlsruhe, on the other hand, integrates explainability directly into the training process. The accuracy of the predictions is maintained while the efficiency of the analysis increases. With this work, the scientists are creating a methodological basis in order to transfer such approaches into practice in the future.
In addition to technical precision, trustworthiness and acceptance among users play a role. This applies, for example, to intelligent systems for charging and discharging electric cars or home storage systems. Tenure-track professor Benjamin Schäfer from the Institute for Automation and Applied Computer Science at KIT explained:
Users are likely to be more accepting of an intelligent charging system if it is clear why an electric car charged later at night than usual – for example because electricity prices were particularly high at a time and costs could be saved.
Also interesting: