AI-Supported Energy Management

One of the greatest challenges is the immense variety of data sources. Our methodological expertise begins with breaking down these information silos and transforming them into a unified, machine-readable knowledge base.

We develop pipelines that enable large language models (LLMs) not only to process heterogeneous data formats, but also to interpret them semantically:

• Time-series data: Load profiles, temperature curves, and pressure trends are interpreted not merely as numerical sequences, but as physical processes with characteristic patterns.
• Unstructured documents: Information from PDF data sheets, free-text maintenance reports, and technical standards is extracted and placed into context.
• Graphical information: We develop multimodal approaches to capture and interpret information from P&ID diagrams, tables, and technical schematics.

The result is a vectorized knowledge base that functions as a dynamic knowledge repository for the organization and provides the foundation for advanced AI applications.

Conventional chatbots are not sufficient. We develop proactive LLM-based agent systems capable of autonomously performing complex tasks. Rather than simply responding to user requests, these systems act as investigative assistants.

• Hypothesis-driven root cause analysis with RAG: When a monitoring algorithm detects a statistical anomaly (e.g. increased standby consumption), it initiates an analysis agent. Using Retrieval-Augmented Generation (RAG), the agent searches the knowledge base for potential causal relationships and generates evidence-based hypotheses (e.g. “The compressed air system shows elevated base consumption. A temporary seal replacement was documented in the maintenance log dated May 15. This may indicate a leakage.”).
• Workflow orchestration: We develop agent architectures in which multiple specialized agents interact and collaborate. An “analysis agent” can initiate a “RAG agent” for targeted information retrieval and subsequently activate a “planning agent” that generates a concrete, standards-compliant solution proposal, including an economic assessment.
• Human-in-the-loop validation: Critical decisions remain under human supervision. Before a recommendation is finalized, the system presents its hypotheses and supporting evidence to an expert for review and validation. This combines the analytical capabilities and speed of AI with the reliability of human expertise.

We use large language models (LLMs) to significantly lower the barriers to working with complex engineering tools. Instead of mastering every technical detail of a software environment, experts can interact with systems directly through natural language.

• Automated generation of energy system models: Preparing data for simulation models is typically a time-intensive task. Our LLM-based assistance systems streamline this process. Users provide heterogeneous raw data and define their objectives through dialogue. The assistant automatically cleans and structures the data, fills data gaps using plausible synthetic data, and generates executable, parameterized simulation models in established frameworks such as Modelica and oemof.
• Interactive operation of digital twins: We develop interfaces that enable LLMs to function as intuitive access layers for complex systems such as digital twins. Instead of navigating complex dashboards, maintenance personnel can formulate requests directly in natural language, for example: “Compare the efficiency profile of compressor 3 with the equivalent compressor at site B over the past week.” The LLM interprets the request, retrieves the relevant data, performs the analysis, and provides the results in a comprehensible form.

Our methodological research is translated into practical, application-driven solutions that we develop and prototype within research projects.

• AI assistant for ISO 50001 energy audits: We transform static audit reports into dynamic, continuous monitoring systems. The assistant continuously identifies inefficiencies, links them with evidence from internal documentation, and recommends prioritized, standards-compliant measures.
• The "compressed air agent":In this approach, we combine a digital twin of a compressed air system with a large language model (LLM). The system enables not only continuous monitoring, but also dialogue-based root cause analysis and predictive maintenance based on Physics-Informed Neural Networks (PINNs).
• AI assistance for energy system simulation:We develop assistance systems that significantly accelerate the workflow from heterogeneous data sources to executable energy system simulation models, thereby reducing barriers to adoption for SMEs.