Background

Austria has set ambitious climate targets. Achieving them requires a transition from fossil fuels to sustainable energy sources such as hydropower, wind, and photovoltaics. To make this transition successful energy communities are to be set up, enabling households and businesses to jointly produce and consume renewable energy.

Another critical challenge is natural disasters, which are expected to become more frequent due to climate change. First responders and authorities currently rely on real-time data, but existing guidelines and emergency systems are inconvenient, hindering fast decision-making.

In both cases, AI-driven solutions—including natural language-based data retrieval and predictive models—could significantly improve the situation. AI can make energy communities more accessible, provide early warnings when danger is imminent, and support informed decision-making. As a result, people can better prepare for emergencies, make more effective choices, and potentially reduce damage.

Project Content

Part of the plans to reach climate neutrality in Austria includes revving up hydropower production from 42 TWh to 47 TWh by 2030. However, hydropower generation—particularly in small plants, such as those along the Traisen and Kamp rivers in Lower Austria—is directly dependent on water availability and flow volumes. What adds to this difficulty is that energy and water management produces complex, high-dimensional, and dynamic datasets, making it difficult to extract actionable insights. For this context-aware information, such as flooding situation reports or energy consumption trends are required. Current tools often fail to provide intuitive, natural language-driven access to such information, limiting their usability for practical decision-making. MANTRA offers a solution by developing a multi-agent system powered by large language models (LLMs). With such a system, natural language interaction with time-series energy and water data becomes feasible. Users can query data, receive actionable insights, generate automated reports, and support data-driven decision-making—all through standard natural language interaction.

Aims

We intend to develop a natural language (LLM)-powered multi-agent system that aids in time-series data analysis. It allows users to intuitively extract insights on trends, anomalies, and forecasts through natural language interaction. Moreover, we explore how multi-agent architectures can enhance time-series processing. Specifically, we distribute tasks such as data retrieval, pattern recognition, forecasting, and natural language-based interpretation across specialized agents. The focus of the project is on energy and water management, but the methodologies developed can also be applied to data from other domains, such as finance, climate modeling, and industrial monitoring.

Results

• Automated flooding situation reports:The Hydrographic Institute of the Provincial Government of Lower Austria currently compiles textual flooding situation reports manually. Automating this process using natural language generation (NLG) could significantly improve efficiency.
• Voice-based water level inquiry system:The Provincial Government of Lower Austria operates a telephone-based water level inquiry service that currently relies on a touch-tone menu system. Replacing it with a voice-based natural language interface would make the service more intuitive and user-friendly.
• Energy community support:Many members of renewable energy communities are not digital natives and therefore strongly benefit from systems that allow interaction through natural language.
• Data access for energy managers:Renewable energy community managers, as well as energy and climate coordinators, can gain insights from data more easily without requiring expert-level knowledge in data processing.
• Energy forecasting for smart consumption:A user-friendly forecasting system enables a wider range of users to predict energy production, supporting intelligent energy use, such as optimized scheduling of electric vehicle charging.