Researchers release energy records for 2,027 Sofia buildings

Researchers have assembled energy information for 2,027 buildings in Sofia, combining official Energy Performance Certificates with cadastral records and construction characteristics. Published in Scientific Data on November 3, the peer-reviewed data descriptor presents a reusable resource for modelling annual energy consumption and its geographic patterns. The contribution is a harmonised set of building records that other teams can use to develop and assess predictive methods.[1]

Energy Performance Certificates describe a building's assessed energy characteristics, but the records are rarely formatted identically across data providers or linked seamlessly to geographic footprints. Cadastral information contains the location and physical identity of buildings, yet may omit the energy-related attributes that a heating or retrofit model needs. Combining the sources requires matching real entities, not merely joining two spreadsheets on a street name. Differences in identifiers, building extensions, construction dates and how spaces are measured can lead to false matches or missing records. The Sofia team describes a multistage workflow for cleaning, validating, standardising and enriching the records before publishing a consistent dataset.[1]

What the 2,027 records actually describe

According to the data descriptor, the combined resource includes building function, construction characteristics and energy-performance indicators, all made suitable for context-based analysis of annual consumption. These variables allow a researcher to ask whether energy demand differs between categories of buildings or correlates with age, form and assessed efficiency. They also give a basis for studying which areas might benefit from further inspection. The public dataset is explicitly designed to support machine learning, simulation and urban-energy planning. That is its intended capability, not a claim that any particular forecast model has already reached a stated accuracy or that 2,027 buildings represent the complete city stock.[1]

The 2,027 records are a sample, not a demonstrated census of Sofia's building stock. How buildings entered the certificate system matters to the interpretation: assessed properties could differ from those without records. For example, uneven coverage by building age or district could make a model's apparent geographic precision reflect data availability. Analysts should establish the sample's coverage before using it to rank the whole city.[1]

An AI model needs a ground truth it can trust

For a meaningful first modelling exercise, an analyst could split the buildings into training and hold-out sets and predict an annual energy indicator from documented attributes such as use and construction form. Evaluation should report not only average error but the distribution of error across building functions, sizes and neighbourhoods. A geographic hold-out—testing on districts that contributed no training buildings—would better reveal whether the learned relationships travel than a random split among nearly identical neighbouring properties. The underlying certificates may reflect standardised assessment conditions rather than measured bills, so a prediction of certificate-based energy demand must not be presented as verified real electricity or heating consumption.[1]

Why a local dataset can be more valuable than a giant generic one

A useful comparison comes from England and Wales, where Cambridge researchers published the NEBULA dataset in June: 613,290 neighbourhood-level samples assembled from building and environmental information for areas containing roughly five to 150 households. The Sofia resource instead centres on individual identified buildings and energy certificate attributes. The datasets differ radically in unit of observation, geographic scope and input construction. One is not demonstrably better because it has more rows. Both illustrate that choosing the right level of detail matters: neighbourhood analysis can guide broad policy, whereas building-specific records may help decide where a surveyor should inspect next.[1][2]

For decisions such as retrofit funding, a useful next step would compare certificate-based predictions with measured energy where available and examine errors across building types and districts. That would test whether the harmonised records support the particular decision being proposed. The released dataset gives independent researchers a common starting point for that work.[1]

Report an error or suggest a correction