GAD Editorial · Published · · English study · Chinese summary · Study dated 2026-10-09
A building's operating calendar is part of its architecture in use. A library may extend evening access to one floor; a school hall may serve the neighborhood at weekends. Those changes alter the relationship between space, servicing and time. An energy forecast becomes useful to an architectural team when its evaluation can distinguish familiar operating patterns from the changed conditions that a project actually introduces.
The Building Data Genome Project 2 paper, by Clayton Miller and colleagues, appeared in Scientific Data in 2020. It describes hourly records for 2016 and 2017 from 3,053 meters across 1,636 non-residential buildings at 19 sites. The collection includes building metadata and weather information, with a subset used for the ASHRAE Great Energy Predictor III competition. BDG2 is a research dataset for evaluating methods, rather than a forecasting product or a current view of those buildings.
The official repository separates raw and cleaned meter data and links the collection's documentation. The paper's data description explains that 'raw' files have already undergone processing and that timestamps use local time. Further cleaning removes specified zero-reading sequences. A reviewer should therefore identify the exact data version and transformation history before treating a neat graph as an untouched instrument record.
Consider a proposed evaluation for a community learning building with a separately operated hall. Start with a meter diagram that shows what each reading includes. Does the hall share ventilation plant with the classrooms? Does a whole-building electrical meter include a neighboring use? The intended design question could be whether extended public opening creates a persistent evening load. An hourly total without a clear measurement boundary cannot resolve that question, however accurately a model predicts it.
Set up two different tests. For a known building, train only on earlier periods and assess later periods that were unavailable during training. For a method intended to transfer across a portfolio, hold out complete buildings or sites. Report the two results separately. Randomly scattering hours from the same operating year across training and test sets would answer a narrower question than either a future forecast or an unfamiliar-building deployment.
Describe the calendar alongside the errors. Review weekday opening, weekends, closures and periods where schedules change. Compare with a simple historical baseline, such as a corresponding previous-week pattern, under the same information limits. The preferred method should improve the decisions the team needs to make, not merely an aggregate score dominated by routine hours. Retain a separate account of performance during the opening extension or other relevant transition.
Treat missing readings and zero readings as questions about the evidence. A shutdown, a communications failure and a genuinely inactive service can lead to superficially similar charts. Preserve the original record, the reason for each exclusion and the version used for training. The specific review procedure proposed here has not been run by GAD, and it does not establish the condition of any building in the dataset.
Finally, keep prediction separate from explanation. An unexpected evening demand may justify investigation of zoning, controls or the brief, but a forecast residual does not identify which one caused it. Bring the finding back to the plan, the service boundaries and the actual operating agreement. BDG2 provides a substantial historical testbed; architectural value emerges when the evaluation is organized around a building's changing use rather than an abstract contest between models.