energy · infrastructure · family: the model was never trained on this
the savings were never real
Building energy models overestimate retrofit savings, distorting investment decisions
Problem statement
Building energy models used to justify retrofit investments systematically overestimate the energy savings that retrofits actually deliver. This "performance gap" — the difference between modeled and actual post-retrofit energy use — is substantial: the Energies study reports that static models frequently overestimate energy savings, and it cites one earlier study in which a simplified steady-state approach overestimated primary energy savings by up to 64.7%. In the study's own comparison of five Slovenian case-study buildings, the static compliance calculation and a dynamic simulation of the same building gave energy consumption estimates that differed by 15.6% to 36.8% (a mean absolute percentage error of 24.06%), with the static figure the higher of the two in every case; that comparison is model against model, not model against metered use. The problem stems from static modeling approaches that rest on steady-state assumptions, standardized climate data, and fixed occupancy assumptions, when in reality building energy performance emerges from complex interactions between envelope, HVAC systems, weather, occupancy patterns, and human behavior that are highly dynamic and difficult to predict.
Why this matters
Retrofit investment decisions are made on the basis of these energy models. When models overestimate savings, they also predict unrealistically short payback periods, leading to overly optimistic investment decisions and misallocated resources. At the policy level, governments base building energy codes and renovation mandates on modeled performance targets. If those models are systematically optimistic, compliance doesn't actually deliver the expected emissions reductions, and renovation strategies built on them may fail to meet long-term climate commitments. National regulations aligned with the EU's Energy Performance of Buildings Directive (EPBD), such as Slovenia's PURES, rely primarily on static, standardized energy performance calculations, which research has shown often underestimate actual consumption.
What’s been tried and why it hasn’t worked
Static calculation methods (e.g., the ISO 13790 monthly method) are widely used for regulatory compliance because they're fast, standardized, and simple. But they use simplified thermal models, assume standardized occupancy, and cannot capture hourly fluctuations in weather and demand, thermal mass effects, or the dynamic interplay of energy supply and demand. Dynamic simulation tools like EnergyPlus and TRNSYS offer higher fidelity but require extensive input data, greater expertise, and more computation — practical barriers that, according to the Energies study, often hinder their use in smaller-scale renovations. Machine learning approaches trained on building operational data show promise, but depend heavily on the accuracy and completeness of historical data and often rely on large amounts of it; because the quality and availability of that data affect prediction accuracy, the AIMS Energy review concludes that such prediction algorithms are not fully applicable to all buildings.
What would unlock progress
Two complementary advances could close the performance gap. First, hybrid models that combine physics-based building simulation with data-driven calibration from sensors and utility data could capture building-specific dynamics without requiring the full data input of detailed simulation. Second, robust methods for incorporating occupant behavior into energy models — moving beyond idealized schedules to probabilistic occupancy and behavior models — would address one of the largest sources of prediction error. Occupant behavior has been shown to cause wide variations in actual energy consumption, yet most models assume standard schedules and fixed boundary conditions rather than treating occupants as dynamic agents.
Entry points for student teams
A student team could instrument 3–5 campus classrooms or offices with low-cost temperature, CO₂, and occupancy sensors, then compare actual energy consumption against predictions from a standard energy model (e.g., OpenStudio or DesignBuilder). The measured performance gap could be decomposed by source: envelope differences, system efficiency, occupancy variation, and behavioral factors. This would produce a quantitative case study of where prediction models break down. Skills in building science, data analysis, sensor deployment, and HVAC fundamentals would be most relevant.
Genes
Structural cousins: stuck the same way, in other fields
Sources
"Energy Efficiency and Decarbonization Strategies in Buildings: A Review of Technologies, Policies, and Future Directions," *Applied Sciences*, MDPI, 15(21):11660, 2025. (accessed 2026-02-11). Supplemented with "Bridging the Gap to Decarbonization: Evaluating Energy Renovation Performance and Compliance," *Energies*, MDPI, 18(5):1146, 2025; and "A review of the influencing factors of building energy consumption and the prediction and optimization of energy consumption," *AIMS Energy*, 2025. go to source ↗
Working notes
Notes made while writing and checking this brief, shown as written. They use working shorthand.
- The "Bridging the Gap to Decarbonization" paper (Energies, 2025) found that even buildings that had undergone retrofits frequently failed to meet current energy performance requirements, particularly for window and floor insulation.
- The performance gap is documented in the Energies study and the research it cites; that study also finds that a lack of standardization in renovation measures contributes to gaps between expected and actual energy savings.
- Cross-domain connection: this problem mirrors the ocean-dl-extreme-event-failure brief, where models trained on benign conditions fail in real-world variability — the underlying issue of unrepresentative training data applies to building energy models trained on idealized assumptions.
- Occupant behavior is described in the Applied Sciences review as a key piece of the energy efficiency puzzle; the review notes that models typically assume standard schedules for equipment and occupant densities, which may differ in reality.
- Space heating and water heating together account for almost half of the energy use in buildings worldwide (Applied Sciences review), and heating energy use is sensitive to occupant thermostat settings: the review says each degree (°C) change can alter heating energy use by roughly 5–7%, depending on climate and system.
Reconciliation 2026-09-30: Fact repair after reading the whole write-up against the three sources it names. The publisher pages at mdpi.com refused automated access, so the two MDPI papers were fetched on 2026-09-30 as the publisher's own PDFs from its file host: the Applied Sciences review (https://mdpi-res.com/d_attachment/applsci/applsci-15-11660/article_deploy/applsci-15-11660.pdf, 56 pages, full text searched and the modelling, occupant-behaviour and policy sections read) and the Energies study by Stegnar (https://mdpi-res.com/d_attachment/energies/energies-18-01146/article_deploy/energies-18-01146.pdf, 22 pages, read in full). The AIMS Energy review was fetched as the journal's PDF (https://www.aimspress.com/aimspress-data/aimse/2025/1/PDF/energy-13-01-003.pdf, 51 pages, full text searched and the prediction-methods section read). Crossref records for all three were also fetched; the Source line's titles, journals, volumes, issues and article numbers match them (the AIMS Energy paper is 13(1):35–85). (1) The text said the performance gap "routinely ranges from 20% to 60%" between predicted and measured use. No source gives that range. The Energies study compares a static compliance calculation with a dynamic simulation of five Slovenian case-study buildings and reports relative errors of 15.6% to 36.8% between the two (the static estimate the higher one each time), with a mean absolute percentage error of 24.06%; it cites one study finding that a simplified steady-state approach overestimates primary energy savings by up to 64.7%. The text now gives those figures, says "modeled and actual" rather than "predicted and measured", and states that the study's own comparison is model against model, not model against metered use. (2) "Idealized occupant behavior, simplified physics, and stable conditions" is now the Energies study's wording: steady-state assumptions, standardized climate data and fixed occupancy assumptions. (3) "Hundreds of thousands to millions of dollars per building" and the erosion of "investor and building owner confidence in energy efficiency as a viable investment class" are in no source and were removed; the sentence now gives the study's finding that static models predict unrealistically short payback periods, leading to overly optimistic investment decisions and misallocated resources. (4) "A hidden gap in national climate commitments" is now the study's statement that renovation strategies built on static simulations could fail to meet long-term climate commitments; the EPBD sentence now says that national regulations aligned with the EPBD, such as Slovenia's PURES, rely primarily on static calculations, which research has shown often underestimate actual consumption, as the study puts it. (5) "Cannot capture feedback loops between systems" is now the study's list of what static methods miss: hourly fluctuations, thermal mass effects and the dynamic interplay of supply and demand. (6) The text said dynamic tools need data "expensive to collect, expert calibration, and significant computation time", making them "impractical for large-scale retrofit planning across building portfolios". The study says they need extensive input data, greater expertise and more computation, and that these barriers often hinder their use in smaller-scale renovations; the sentence now says that. (7) The text said machine-learning models "require years of historical data" and "don't transfer well" between buildings because each is idiosyncratic. The AIMS Energy review says the data-driven method is highly dependent on the accuracy and completeness of historical data, that optimized prediction algorithms often rely on large amounts of historical data whose quality and availability affect accuracy, and that such algorithms are therefore not fully applicable to all buildings; the sentence now says that. (8) "2–3x variation in energy use between identical buildings" is in no source; the Applied Sciences review says occupant behaviour "can lead to wide variations in actual energy consumption", and the sentence now says that. (9) In the notes above, the claim that "no standardized method exists for quantifying it during the design phase" is in no source; the bullet now gives the study's finding that a lack of standardization in renovation measures contributes to the gap. (10) The IEA EBC Annex 66 bullet is in no source and was replaced with what the Applied Sciences review says about occupant behaviour and standard schedules. (11) "HVAC contributes approximately 40% of building energy consumption" is in no source; the bullet now gives the review's figures (space and water heating together almost half of building energy use worldwide; roughly 5–7% change in heating energy per degree of thermostat setting), and the bullet's claims about control strategies, maintenance status and static models, which the sources do not make in that form, were dropped. (12) The text listed "ASHRAE Standard 90.1 compliance tools" as a static calculation method and called static methods "the dominant approach for retrofit planning" because they are "inexpensive". No source calls ASHRAE 90.1 a static method (the Applied Sciences review names it only as a standard for compliance modeling), and the Energies study says static models are widely used for regulatory compliance because of their efficiency and simplicity; the sentence now says that and names only the ISO 13790 monthly method, which the study used as its static model. Confirmed in the sources and left as written: static models overestimate savings and misalign investment decisions (Energies abstract); ISO 13790 as the static method; EnergyPlus and TRNSYS as detailed tools and OpenStudio and DesignBuilder as interfaces to them; hybrid physics-plus-data models (the AIMS Energy "gray box" method, calibration to metered data in the Applied Sciences review, and the Energies study's closing call for a hybrid approach); occupants suggested as "the main cause" of the gap in some contexts; and the first note's statement that retrofitted buildings often fail current requirements, particularly window and floor insulation. Not checked, because they are proposals rather than claims about the world: the entry-point study design and the call for probabilistic occupancy models. The 2026-08-18 note above still quotes the brief's old "20–60%" figure; it is left as written. A second reader went over these repairs against the same three sources on the same day; the wording of items 1, 4, 7 and 11 was tightened and item 12 added in that pass. No title changed. Genome tags untouched.