Abstract
The Austrian energy sector has seen growth in the share of renewable energy sources innational primary energy production. The utilization of these renewable energy sources
could potentially lead to variable load flow situations and difficult-to-forecast grid loads.
End-consumers, who are increasingly also producing electricity, must navigate a wide
array of options and regulations within the energy domain. This complexity can make it
challenging to maintain an overview and act in an energy-aware manner, especially given
the challenges in comprehending energy data and its visualizations. To bridge the gap
between energy domain data and end-consumers, this thesis investigates the potential
of using Large Language Models (LLMs) to generate personalized, energy-data-based
textual action recommendations within the context of the INNOnet research project,
which investigates load-dependent grid tariffs. It also explores how the text-generation
capabilities of publicly accessible LLMs can be leveraged for this use case in the energy
sector. The proposed proof-of-concept approach features modular process chains that
allow customization for various user needs and data types. Documented using UML activity and turtle diagrams, these chains outline the sequence of tasks and prompt-based
interactions with an LLM, which are necessary to produce the final recommendations. A
proof-of-concept prototype application was developed using the LangChain framework,
which demonstrates the process implementation along with preprocessing steps such as
preparing textual summaries of data. The prototype was subjectively assessed through
recommendations, which were generated within test scenarios for authentic personas.
Factors, such as alignment with input data, factual accuracy and reasonableness, were
considered. The evaluation indicates that the prototype can produce outputs that align
with the provided data, while also identifying areas for refinement, such as improving
data accuracy, contextual aspects and inconsistent structures. However, these outputs
are not consistent or accurate enough to reliably create reasonable energy-data-based
textual action recommendations. A future approach might involve utilizing machine
learning (ML) techniques to analyze the data and gather insights from it. Subsequently,
the text-generation capabilities of LLMs could be leveraged to translate these analytical
results into personalized and comprehensible recommendations.
| Date of Award | 2024 |
|---|---|
| Original language | English (American) |
| Supervisor | Christoph Schaffer (Supervisor) |