Managing our energy supply is becoming increasingly complex. Wind and solar power fluctuate depending on the weather, while at the same time electric vehicles, battery storage, and heat pumps are changing consumption patterns. “Grid operators and energy suppliers are increasingly using artificial intelligence to operate their systems efficiently and stably,” says tenure-track professor Benjamin Schäfer from the Institute for Automation and Applied Computer Science (IAI) at KIT.

To align power generation and consumption as precisely as possible, numerous factors must be taken into account simultaneously—such as weather forecasts, load forecasts, grid and distribution capacities, and consumer behavior. “AI helps with this, but it must not remain a black box. People must be able to understand how predictions and decisions are made. Especially in the sensitive energy sector, where errors can have serious consequences, transparency and human oversight are crucial—and, with the European Union’s AI Act, they are also a regulatory requirement,” says Schäfer, who heads the Helmholtz Young Investigator Group DRACOS (short for: Data-Driven Analysis of Complex Systems) at KIT and received the 2026 Ecology Prize from the Viktor & Sigrid Dulger Foundation of the Heidelberg Academy of Sciences for his work.

“SHAPformer” combines several methods

In their latest study, funded by the Helmholtz Association through Helmholtz AI, Schäfer’s research group presents the new method “SHAPformer.” The scientists developed it for time-series forecasting—that is, for predictions based on sequentially collected data, such as electricity consumption or electricity prices. The goal is to make AI-based predictions as precise and transparent as possible.

The method combines Transformer models—known from modern language models—with techniques from explainable artificial intelligence (“Explainable AI”). The name “SHAPformer” refers to the combination of Transformer models with SHAP methods. These are based on concepts from game theory and reveal the influence that individual factors have on a prediction. These include, for example, temperatures, holidays, wind forecasts, or historical consumption data.

The influence of individual factors becomes visible

“When training our model, we deliberately masked specific pieces of information,” explains Matthias Hertel, a research associate at KIT’s IAI and the study’s first author. “This allowed us to understand the influence that certain inputs have on the model’s predictions.”

The team trained the model using, among other things, real-world data from the transmission system operator TransnetBW. The goal was to predict electricity consumption and prices over periods of up to one week—while simultaneously revealing which factors influence the forecasts. This allows for the analysis of how individual factors contribute to a prediction.

Explainability Integrated Directly into Training

Many existing methods only provide explanations after the fact and require significant additional computing power to do so. “A distinctive feature of our approach is that we integrate explainability directly into the training process,” says Hertel. This preserves the accuracy of the predictions while increasing the efficiency of the analysis.

“With our work, we’re laying the methodological groundwork for translating such approaches into practice in the future,” says Schäfer. Not only do technical precision and reliability play a role here, but also user acceptance. Schäfer cites as examples intelligent systems for charging and discharging electric cars or home energy storage units that automatically respond to electricity prices. “Users are likely to be more receptive to a smart charging system if it is clearly understandable why an electric car charged later than usual at night—for example, because electricity prices were particularly high during that time, allowing for cost savings.”

Video