Architectural Frameworks for extracting semantic and spatial references from unstructured text using AI: The case of Energy markets

Abstract

This thesis proposes an architectural framework for extracting semantic and spatial references from unstructured text using Artificial Intelligence, with application to the energy market domain. The European energy sector is rapidly digitalizing, producing vast volumes of open and semi-open data such as market indicators (ENTSO-E), transmission system measurements (ADMIE), geospatial datasets for infrastructures, and meteorological fields. Yet, structured data alone are rarely sufficient to explain market anomalies—sudden price spikes or shifts in production mix—whose causes are often described only in unstructured textual sources like news articles and policy announcements. The proposed system integrates structured time-series with AI-driven knowledge extraction from textual data to form a unified, explainable analytical framework. Its architecture consists of: (i) spatiotemporal harmonization and canonicalization of heterogeneous datasets; (ii) ontology-aligned extraction of events, entities, and locations with semantic, spatial, and temporal grounding; and (iii) storage and reasoning within a Neo4j knowledge graph. The GraphRAG methodology (Graph-based Retrieval-Augmented Generation) enables provenance-rich, explainable narratives that link anomalies in energy time-series to their plausible drivers through Cypher-first retrieval strategies. By bridging statistical anomaly detection with semantic understanding, the thesis contributes to the development of a Semantic Spatial Data Infrastructure (SSDI) for the energy sector. This framework enhances transparency, explainability, and decision support in energy data analytics, offering a foundation for future integration with predictive and causal models.