A Multimodal Dataset for Large Language Model Applications in the Energy Domain

Research Motivation

The energy sector is undergoing a rapid digital transformation, yet AI-driven decision-support tools are often hindered by fragmented data sources. While numerical data, geospatial maps, and policy documents exist, they are typically siloed in different formats and repositories. This paper introduces the mAIEnergy dataset to bridge this gap, providing a unified, multimodal foundation for Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) systems in the energy domain.

Approach and Methodology

The authors developed a three-stage pipeline—Identification, Retrieval, and Preparation—to aggregate data across four modalities: textual (policy, scientific, news), imagery (satellite, infrastructure, labels), numerical (market prices, weather, load), and geospatial (grid networks, power plants).

Key to the dataset's utility is its focus on harmonization and cross-modal linkage. Rather than simply co-locating raw files, the authors implemented a standardized schema with consistent attribute names and normalized units. They also created a semantic enrichment layer, embedding records for hybrid retrieval and constructing a property graph that links disparate data points (e.g., connecting a specific power plant's geospatial location to its generation time series and relevant regulatory documents).

By adhering to FAIR (Findable, Accessible, Interoperable, and Reusable) principles, mAIEnergy provides a reproducible and extensible resource for researchers. It moves beyond single-modality datasets by enabling complex reasoning tasks—such as correlating weather patterns with electricity prices or mapping infrastructure assets to policy compliance—which are essential for modern energy-system modeling and decision-making.