Article 03: Eighteen Per Cent: What Building Characteristics Can and Cannot Explain About Office Energy Use

A series mining the PhD thesis "London and UK Office Buildings: Investigating Energy Use and Landlord-Tenant Influences" (Azhari, 2025).

Key takeaway. A linear regression across the Greater London office stock explains just 18 per cent of the variation in electricity EUI and 4 per cent of gas EUI from building characteristics alone. The other 82 per cent is where management, leases and operations live, and that is where most of the controllable savings sit.

Chart

Four candidate models, electricity EUI prediction

The most flexible model (Random Forest) and the most interpretable (Linear Regression) tie at the same ceiling. R-squared peaks at 0.18 and RMSE bottoms out at 0.28.

The headline that should change the conversation

Take a dataset of 6,038 offices in Greater London with metered electricity and gas data for 2017. Add 33 explanatory variables drawn from the 3DStock model: size, height, age, attached status, borough, primary heating fuel, EPC band, air-conditioning presence, tenancy and more. Run a linear regression against electricity energy use intensity (EUI). The model explains 18 per cent of the variation between buildings. Run the same model against gas EUI: 4 per cent.

Read those numbers again. Even with one of the richest stock-level datasets ever built for UK offices, the physical and administrative attributes of a building collectively account for less than a fifth of the variation in how much electricity it actually uses, and almost none of the variation in how much gas it uses.

The conclusion is not that the model is bad. The conclusion is that the question the model is asking (how does this building physical and administrative profile predict its energy use?) is the wrong question to ask if your goal is to understand or reduce energy use.

The model bake-off

Before drawing strong conclusions from a single model, the thesis tested four. Each was set up using R and the Tidymodels framework, with an 80/20 train-test split and twenty iterations of V-fold cross-validation for tuning.

A simple linear regression. A LASSO-penalised regression to select among the 33 variables. Multivariate Adaptive Regression Splines (MARS), which can capture non-linearities and interactions between variables. Random Forest, an ensemble of decision trees that is generally the most flexible of the four on tabular data.

The results land within a narrow band. For electricity EUI, Random Forest achieves an RMSE of 0.28 and an R-squared of 0.18. Linear regression achieves the same: RMSE 0.28, R-squared 0.18. MARS comes in at RMSE 0.29 and R-squared 0.16. LASSO at RMSE 0.31 and R-squared 0.12. For gas EUI, all four models cluster around R-squared 0.03 to 0.04.

When the most flexible model and the most interpretable model finish in a statistical dead heat, the choice between them comes down to interpretability rather than performance. The thesis chose linear regression for the reported analysis, on the principle that a model whose coefficients can be read and challenged is more useful for policy and capex decisions than a black box whose predictions cannot be interrogated. Article 13 walks through that decision in more technical detail. For the purposes of this piece, the choice is cosmetic. Whichever model you pick, you are stuck at the same ceiling.

Where the missing 82 per cent goes

If the static and administrative attributes of a building explain only 18 per cent of its electricity EUI, what explains the rest? The thesis groups the unexplained variation into three categories.

The first is dynamic factors. Things that change over time: occupant density, hours of use, plant set-points, heating and cooling regimes, the seasonal mix of electric versus gas demand. The 3DStock model does not see any of these. They live in BMS logs, EMS dashboards and half-hourly metered data that is not part of the rating list or the EPC.

The second is random factors: an unplanned chiller outage, a one-off computer-room expansion, a refurbishment year that artificially depresses or inflates one year of meter readings, a long cold winter or a hot summer. Some of these are noise. Some are signal that requires more than one year of data to disentangle.

The third, and most consequential, is management and organisational factors: who pays the bill, who maintains the plant, whether the lease incentivises efficiency, whether the landlord has the ability to invest, whether the BMS is actually being read. The interview chapter of the thesis is built on the hypothesis that these factors do a lot of the heavy lifting that fixed building attributes cannot. Reported portfolio energy reductions of 38 per cent, 42 per cent and 73 per cent over a decade in major UK landlords (anonymised, see Article 8) are consistent with that hypothesis. One landlord interviewee captured the point succinctly:

"Tenant energy use is roughly half of our total energy usage, so we have to work with them to optimise systems and meet our targets."

If half of total energy use is tenant-controlled, no model built only on landlord-visible attributes can predict it.

A modelling response: the 4DStock idea

Janda and colleagues (2022) have proposed extending the 3DStock model to a fourth dimension: the organisational layer. The intuition is straightforward. The 3D version captures geometry, geography and physical attributes. A 4D version would add the network of ownership, tenancy and management relationships that determine how a building is actually run.

In practical terms, this might mean attaching to each Self-Contained Unit (SCU) the identity of its owner, the type of lease in place, the number of tenants, the presence and granularity of sub-metering, the maturity of the building management system, the presence or absence of green clauses in the lease and the energy management policy of the operating organisation. None of this data is in the rating list. Much of it is in the heads of facility managers, energy consultants and asset teams. Capturing it at stock level is a research and data-engineering problem that has yet to be solved at scale.

The 4DStock idea is not the only candidate response. Operational rating schemes such as NABERS (covered in Article 9) take a different approach. Instead of trying to model what drives energy use, they measure it directly and let the market work out the rest. The two approaches are complementary rather than competing. A 4DStock model would help researchers and policymakers explain energy variation. NABERS helps owners and tenants reveal it.

Where the real lever sits

For practitioners reading the 18 per cent number, the inference should not be discouraging. It should be focusing. If only 18 per cent of variation comes from things that cannot easily change after a building is built (size, height, age, fabric), then the 82 per cent that can change (operations, occupancy patterns, plant tuning, lease design, sub-metering, stakeholder alignment) is where the budget should sit.

That is exactly what the qualitative evidence in the thesis confirms. The major UK landlords interviewed reported the biggest decadal energy cuts not from fabric upgrades but from optimising the use of existing kit, replacing equipment at end of life with more efficient versions, putting in sub-metering and BMS analytics, and engaging tenants. Articles 6, 7 and 8 take each of these moves in turn.

For policymakers, the inference is sharper still. If the 33 variables visible in 3DStock cannot predict operational energy use, then policy instruments tied to those variables alone (most notably EPCs, covered in Article 2) cannot drive operational savings on their own. They need to be paired with operational rating, with sub-metering requirements, with lease standards and with enforcement architecture that addresses the organisational layer.

What the data does not see

Several caveats apply to the 18 per cent figure. All four candidate models are bounded by the 33 explanatory variables available. They cannot see operating hours, plant efficiency, internal loads, occupant density, fabric U-values or BMS set-points; with richer features the ceiling would be higher. The energy data is a 2017 single-year cross-section, not a panel, so the models cannot separate idiosyncratic noise from structural variation. Linear regression assumes a linear relationship after a log10 transform of the skewed numeric variables; non-linear dynamics within categories will be absorbed by the residual. Finally, generalisation to non-London offices is bounded by the same data limits, not just by sampling. Article 11 tests how far the Greater London findings extrapolate to the rest of England and Wales.

The next article steps out of the model and back into the building. If only a handful of physical characteristics actually move the EUI dial, which ones are they, and by how much? The answer is more nuanced than the regression ceiling makes it sound.

Limitations

All four candidate models are limited by the 33 explanatory variables available. They cannot see operating hours, plant efficiency, internal loads, occupant density, fabric U-values or BMS set-points. Energy data is a 2017 single-year cross-section, not a panel, so the models cannot separate idiosyncratic noise from structural variation. Linear regression assumes a linear relationship after the log10 transform. Non-linear dynamics within categories will be absorbed by the residual. The R-squared of 0.18 is the best of the four models tested, not necessarily a true ceiling. A richer feature set could push it higher. Generalisation to non-London offices is bounded by the same data limits, not just by sampling.

References

• Azhari, R. (2025) London and UK Office Buildings Investigating energy use and landlord/tenant influences. Doctoral thesis (Ph.D), UCL (University College London). URL: https://discovery.ucl.ac.uk/id/eprint/10204821/

• Steadman, P., Evans, S., Liddiard, R., Godoy-Shimizu, D., Ruyssevelt, P., and Humphrey, D. (2020). Building stock energy modelling in the UK: the 3DStock method and the London Building Stock Model. Buildings and Cities. Available at: https://doi.org/10.5334/bc.52

• Janda, K. B., Killip, G., and colleagues (2022). Extending stock models with an organisational dimension (4DStock).

• Kuhn, M., and Silge, J. (2022). Tidy Modeling with R. O Reilly. Available at: https://www.tmwr.org/

• Friedman, J. H. (1991). Multivariate Adaptive Regression Splines. Annals of Statistics, 19(1), 1-67. Available at: https://doi.org/10.1214/aos/1176347963

• Breiman, L. (2001). Random Forests. Machine Learning, 45(1), 5-32. Available at: https://doi.org/10.1023/A:1010933404324

• Bordass, B., Cohen, R., and Field, J. (2004). Energy performance of non-domestic buildings: closing the credibility gap.

Read next

• Article 13: Why I Used Linear Regression Over Random Forest on 6,000 Buildings (the technical companion to this piece).

• Article 6: The Split-Incentive Problem (where the unexplained 82 per cent starts to be accounted for).

About this series

This article is part of a fifteen-piece series adapting the 2025 PhD thesis "London and UK Office Buildings: Investigating Energy Use and Landlord-Tenant Influences" (Azhari, 2025) for a mixed academic and industry readership. The empirical findings draw on the 3DStock model of 6,038 office Self-Contained Units in Greater London with metered energy data for 2017, supplied by BEIS under a data-sharing agreement, alongside the Better Buildings Partnership Real Estate Environmental Benchmark. The qualitative findings draw on semi-structured interviews with seven major UK property organisations, conducted during the 2021 lockdown. Interviewees and their organisations are anonymised by role and organisation type. Please cite the original thesis for academic use.

Author. Rayan Azhari completed his PhD at the UCL Bartlett School of Environment, Energy and Resources in 2025, supervised by Paul Ruyssevelt and Kathryn Janda. The research was supported by the EPSRC Centre for Doctoral Training in Energy Demand (LoLo) and UK Research and Innovation through the Centre for Research into Energy Demand Solutions.

Other articles in the series. Article 1 The 30/85/89 Problem; Article 2 Why EPCs Do Not Tell You How Much Energy a Building Uses; Article 3 Eighteen Per Cent; Article 4 Mapping the Stock; Article 5 Height, Age and the Fuel Question; Article 6 The Split-Incentive Problem; Article 7 Green Leases and Service Charges; Article 8 From 38 to 73 Per Cent Energy Savings; Article 9 NABERS for Britain; Article 10 Time to Retire ECG-19; Article 11 Can London Speak for England and Wales; Article 12 The Hybrid-Work Footprint; Article 13 Why I Used Linear Regression Over Random Forest; Article 14 Vertical Postcodes; Article 15 What Is a Building?