measuring heat loss

in a lived-in house

Problem statement

The single number that says how well a house's fabric keeps heat in — the heat transfer coefficient (HTC, in W/K) — can only be measured reliably by a co-heating test: the house is emptied, heated to a constant temperature for two to three weeks in winter, and the input power is regressed against the indoor–outdoor temperature difference. That is impossible for the hundreds of millions of occupied homes that retrofit programmes need to assess before and after work. The IEA EBC Annex 71 project (2016–2021, nine countries) set out to extract HTC instead from data occupied homes already produce — smart-meter gas and electricity, a few room temperature sensors, local weather — and found that the methods work in principle but not yet as a quality-assurance tool: static and dynamic statistical estimates on five monitored case-study houses "deviated often significantly (20% or more) from the target value," and in a blind test on five inhabited UK dwellings different participants' estimates "for some of the buildings were in close agreement with the target values (co-heating test results), while for other buildings deviations up to almost 50% were found," with "a significant underestimation of the HTC" across all five houses.

Why this matters

Retrofit finance, energy-performance certificates, heat-pump sizing, and pay-for-performance schemes all depend on knowing a building's actual fabric performance rather than the value a designer or assessor assumed — and the two routinely diverge (the "performance gap"). Without an in-use measurement, a retrofit programme cannot verify that installed insulation delivered the heat-loss reduction it was paid for, cannot rank a housing stock by real need, and cannot detect the leaky outliers that most deserve attention. Annex 71's own conclusion is that the methods "show promise" but that "a further in-depth analysis on more case studies is advisable to turn the methods into reliable tools to be used in actual performance assessment," with "specific attention" needed on "uncertainty and repeatability before moving to large scale applications." The UK's SMETER programme, which tested eight smart-meter-based technologies against co-heating baselines in 30 occupied houses, found the concept "effective" but the field still lacks an accepted accuracy standard for the occupied case.

What’s been tried and why it hasn’t worked

Annex 71 followed IEA EBC Annex 58, which had established that HTC can be identified from dynamic tests in unoccupied buildings; the step to occupied buildings is where the methods lose robustness. The Subtask 3 team derived the full heat-balance equation and tested every simplification against five well-characterised houses: static approaches (averaging, single and multiple linear regression) and dynamic approaches (ARX and state-space models). Both families gave internally consistent results but often missed the co-heating reference by 20% or more, and the analysis showed that "assumptions on almost all parameters (measurement time and period, internal heat gains, temperature averaging,…) significantly impact the outcome" — for example, using one room's temperature instead of a multi-sensor average changed the computed HTC by up to 15%, and splitting metered gas between space heating and domestic hot water, or estimating occupant and appliance gains, were among the most influential guesses. In the blind test on five occupied UK homes, static methods gave consistent answers across participants but dynamic methods — which give the analyst freedom over data frequency and model selection — diverged between users. The binding constraint is therefore not computing power or algorithms but disentangling three confounded heat sources (fabric, systems, occupants) from a handful of low-resolution signals, in a building whose occupants open windows, run showers, and switch heating on and off unpredictably.

What would unlock progress

Two things would move the field from "promising" to "usable": a validated, standardised uncertainty budget for in-use HTC (so a result comes with a defensible confidence interval rather than a point estimate that may be 50% off), and cheap ways to observe the confounders directly — a domestic-hot-water disaggregation signal, an occupancy or window-state proxy, a solar-aperture estimate — so that they stop being free parameters. Adjacent fields have solved structurally similar problems: non-intrusive load monitoring disaggregates appliance signatures from a single electricity meter, and system-identification practice in process control routinely delivers parameter estimates with credible intervals from noisy operational data. Annex 71's own framing points at the cheapest data source: the building's "on-board monitoring systems" — the controls and meters of its own heating services — rather than dedicated test instrumentation.

Entry points for student teams

A team could take one instrumented occupied house (or a public dataset such as the UK SMETER release cited in the Annex 71 report) and build a Monte Carlo uncertainty analysis over the assumption set Annex 71 identified — measurement window, gain estimates, temperature averaging, DHW split — to produce HTC estimates with intervals and to rank which assumption most inflates the error. A sensing-focused team could prototype a low-cost add-on (a single hot-water pipe temperature clip, a window-contact sensor, or a boiler-bus reader) and measure how much each shrinks the estimate's uncertainty. A data-science team could test whether a fixed, non-discretionary dynamic-model recipe removes the between-analyst spread seen in the blind test. Relevant skills: building physics, statistical system identification, time-series analysis, low-cost sensing.