Yet a new study suggests that the way we talk about these districts, whether through expert-curated databases or machine-learning analysis of project communications, systematically narrows the concept to its technical core while leaving governance, finance, and social dimensions in the shadows.
Their work, published in Energy Informatics, applied both expert-guided and machine-derived topic modeling to two distinct English-language corpora: a structured database of expert-labeled PED case descriptions and an unstructured collection of project websites and open-access reports. The goal was not to crown one approach as superior but to understand where they converge, where they diverge, and what those patterns reveal about how PED knowledge is produced and communicated.
The study employed a three-way analytical design. First, a Pseudo-document-Guided LDA model was trained on structured case descriptions from the PED-DB, a curated database where entries are pre-labeled across seven thematic dimensions: Energy, Mobility, Stakeholder, Technology, Financial, Social, and Environmental. After preprocessing, 66 documents remained available for training. Second, an unsupervised Latent Dirichlet Allocation model was applied to an unstructured corpus of 35 documents retained from 44 identified PED project sources, totaling approximately 149,420 words. Third, an unsupervised control LDA was trained on the same structured corpus as PG-LDA to serve as a methodological baseline, allowing the researchers to separate the influence of modeling approach from the influence of corpus type.
The PG-LDA model, implemented using the tomotopy Python library, was guided by expert-defined seed terms anchored to the seven PED thematic categories. Thirty-five seed pseudo-documents—five per label—were appended to the training corpus to steer topic formation toward domain-specific vocabulary such as heat pumps, smart grids, and electric vehicle charging. The unsupervised LDA on the unstructured corpus, by contrast, was configured with five topics selected through coherence optimization over a range of candidate topic counts. Both models used fixed random seeds to ensure reproducibility.
To compare topics across the two corpora, the researchers faced a fundamental challenge: the expert-guided model produced bigram-based topics while the unsupervised model generated corpus-derived unigrams, resulting in near-zero direct vocabulary overlap. They turned to sentence-embedding cosine similarity using the SBERT model all-MiniLM-L6-v2, which maps topic keywords into a shared dense semantic space regardless of surface form. For each PG-LDA category, the maximum similarity across all LDA topics was selected as the best-match alignment score.
The results revealed moderate semantic alignment across most thematic dimensions, with SBERT cosine similarity values ranging from 0.348 to 0.535. Energy showed the strongest alignment at 0.535, followed closely by Environmental at 0.521 and Mobility at 0.507. Social, Financial, and Technology exhibited moderate but lower correspondence. Stakeholder was the only category exhibiting weak alignment, indicating that governance and multi-actor coordination are framed in distinctly different ways across structured and unstructured corpora. The cross-model comparison between PG-LDA and control LDA yielded a mean SBERT of 0.453, while the cross-corpus comparison between control LDA and unstructured LDA produced a mean of 0.462—broadly comparable scores suggesting that neither modeling approach nor corpus type alone dominates the observed thematic differences.
A complementary term-frequency analysis provided direct textual evidence corroborating these model-based findings. In the unstructured project corpus, energy-related terms occurred at 28.1 instances per 1,000 words, compared to 8.7 for Financial and 5.7 for Mobility dimensions. In the structured expert corpus, the disparities were even more pronounced: energy vocabulary appeared at 79.0 per 1,000 words versus 11.4 for Financial and 14.5 for Mobility. At the document level, energy-related terms appeared in over 91 percent of documents in both corpora, while Financial vocabulary was present in only 40.5 percent of structured documents and Mobility-related terms in 52.4 percent.
The researchers also consulted five domain experts from the PED-ACT project, who reviewed the PG-LDA topic labels, keywords, and interpretations across a series of workshops held during 2024 and 2025. The experts considered the structured-data-based topics interpretable and broadly relevant, particularly for Energy, Environmental, Technology, and Social dimensions. However, their feedback indicated that Financial, Mobility, and Stakeholder dimensions require more explicit contextualization in project documentation and evaluation frameworks, as these themes appeared less consistently articulated in the modeled textual corpus.
The study’s central finding is that PED discourse across both expert-curated documentation and public project communication converges systematically on techno-environmental dimensions despite the framework articulating seven thematically equal pillars. Financial feasibility, mobility integration, and stakeholder coordination show the weakest representation and lowest cross-corpus alignment scores, while Energy and Environmental categories consistently lead. The researchers suggest this pattern reflects a process of conceptual narrowing in which a multidimensional policy concept is being reduced to an energy-technical construct. If these dimensions remain systematically underreported, decision-makers and funding bodies may have less evidence to assess and invest in the social and institutional aspects of Positive Energy Districts, potentially reinforcing the very narrowing the data reveals. The authors note that standardized reporting templates are one possible response but emphasize that this recommendation is provisional and should be assessed through broader document samples and consultation with projects, municipalities, funders, and other institutions.
Subject of Research: Comparative topic modeling of Positive Energy District characterization using expert-guided and machine-derived approaches
Article Title: Thematic alignment across expert-guided and machine-derived topic models in Positive Energy District characterization
Article References: Shah, J., Han, M., & Zhang, X. (2026). Thematic alignment across expert-guided and machine-derived topic models in Positive Energy District characterization. Energy Informatics, 9(1), Article 74. https://doi.org/10.1186/s42162-026-00688-9
Image Credits: AI Generated
DOI: 10.1186/s42162-026-00688-9
Keywords: Positive Energy Districts, topic modeling, LDA, PG-LDA, semantic alignment, urban energy transition, renewable energy, governance, sustainability, text mining, SBERT, climate neutrality
News Source: Denise Maddox. (October 29, 2026). AI and experts map Positive Energy Districts, revealing shared blind spots. Scienmag.
