Data-driven estimation of design heating loads for HVAC equipment sizing

1.1. Issues with oversized heat pumps

• Oversized heat pumps typically cost more upfront and operate less efficiently, while peak electricity demand can rise by up to 13%.The paper notes that peak demand can strain building or grid infrastructure.Source: §1.1 · p. 2

1.2. Existing methods for equipment sizing

• Manual J and similar methods can be costly, rely on difficult-to-specify inputs, and return point estimates without uncertainty; commercial tools lacked neutral third-party accuracy evaluation.Practitioners may also rely on rules of thumb or like-for-like replacement instead of formal load calculations.Source: §1.2 · pp. 2–3

1.3. Contributions and potential impact

• The new open dataset covers 74 detached single-family houses in five United States climate zones, with housing, equipment, Manual J, utility-bill, thermostat, and weather data.The authors describe it as the largest and richest open dataset of this type of which they are aware.Source: §1.3 · pp. 3–4
• The thermostat method uses existing equipment capacity and smart-thermostat time series; the bill method uses equipment efficiency and monthly utility bills.Both methods require less human-specified input data and could be fully automated.Source: §1.2 · p. 3, §1.3 · p. 4
• 2.75 and 2.96 times: existing heating equipment averaged these ratios versus thermostat- and bill-method estimates; practitioner Manual J averaged 2.26 and 2.35 times those estimates, respectively.These averages cover 74 houses.Source: §1.3 · p. 4
• The bill method may be better suited to scale because nearly every household has utility bills, while relatively few have smart thermostats that can export data.Building the 74-house dataset required reaching out to more than 5,000 people; Manual J quotes ranged from $300 to $2,500.Source: §1.3 · p. 4
• Data-driven sizing could inform standards and shift industry practice, potentially reducing oversizing, equipment costs, utility bills, emissions, and infrastructure strain.These are potential effects described by the authors, not outcomes measured in this study.Source: §1.3 · p. 4

2.1. Existing design-load estimation methods

• Design-load estimation spans floor-area rules of thumb, steady-state methods such as Manual J, and transient simulators that produce time-series loads but require extensive inputs and expertise.Rules of thumb often lead to oversizing; inaccurate inputs can reduce the accuracy of physics-based methods.Source: §2.1 · p. 5

2.2. Data-driven building thermal modeling

• Data-driven models use measured building data to capture as-built performance, while black-box models can lack interpretability and residential sensor data may be limited or behavior variable.Gray-box models combine data and physics.Source: §2.2 · p. 5
• Building on PRISM-style load lines, this paper applies Bayesian bootstrapping for confidence intervals, extends regression to thermostat time series, and targets equipment sizing.The tradition models thermal load as linear in the indoor-outdoor temperature difference.Source: §2.2 · p. 6

2.3. Residential thermostat field studies

• Prior thermostat studies often modeled building dynamics or retrofit potential; many used Donate Your Data records that lacked HVAC equipment capacities and efficiencies.Those studies commonly fit lumped resistance-capacitance parameters.Source: §2.3 · pp. 6–7
• Unlike Baasch et al.'s daily-average runtime balance-point method, this paper selects overnight data to reduce uncertainty from internal gains and improve model fits.Zeifman et al. also integrated time-series data, while this paper restricts attention to nights when solar effects and occupant behavior variability are minimal.Source: §2.3 · pp. 6–7

• The methods section covers data collection, a shared thermal circuit model, and the thermostat and utility-bill methods for estimating design thermal loads.Source: §3 · p. 7

3.1. Data collection

• 125 respondents met the initial suitability criteria, including having eligible heating equipment and a smart thermostat or willingness to install one.Participants were recruited through university email lists, social media, and online building energy forums.Source: §3.1 · p. 7
• Manual J recruitment was difficult: practitioners were hard to find and quotes reached $2,500.The highest amount paid for a calculation was $1,200.Source: §3.1 · p. 7
• 316 people responded to outreach, 103 were selected to join, and 71 completed the full process.Participants were asked to complete each task sequentially before receiving the next.Source: §3.1 · p. 7, Figure 1 · p. 8
• Data from 74 participants included bills, thermostat records, and equipment photos; 71 provided Manual J calculations, and 70 datasets passed completeness checks.Some missing or corrupted data could not be resolved.Source: §3.1 · p. 8

3.2. Thermal circuit model

• The model represents indoor air with temperature Tin and capacitance C, linked by resistance R to outdoor air at Tout.Heating power Q̇h and exogenous power Q̇e supply heat to the indoor air.Source: §3.2 · p. 8
• The model includes both equipment-supplied heat and exogenous gains such as body heat, lights, appliances, and sunlight.Source: §3.2 · p. 8
• The model’s first-order linear differential equation governs indoor air temperature dynamics.The displayed equation passage identifies the heating and exogenous-power terms.Source: §3.2 · p. 8

3.3. Thermostat method

• The method fits thermal resistance and exogenous heat from thermostat records, then evaluates the model at design indoor and outdoor temperatures.It uses indoor and outdoor temperatures and heating-equipment runtime from five-minute measurements, assuming constant equipment power while running.Source: §3.3 · pp. 8–9, Figure 2 · p. 9

3.3.1. Time-averaging to eliminate C

• Time-averaging intervals with equal initial and final indoor temperatures eliminates thermal capacitance C from the model.Source: §3.3.1 · p. 9

3.3.2. Data pre-processing

• Overnight periods reduce solar effects and occupant-behavior variability, and align with hours when design heating loads typically occur.For households using night setbacks, selected intervals run from before the setpoint reduction until indoor temperature returns to the daytime setpoint.Source: §3.3.2 · pp. 9–10, Figure 3 · p. 10

3.3.3. Model fitting

• Linear regression on nightly observations estimates 1/R and exogenous heat, with nonnegative constraints on R and exogenous heat.Source: §3.3.3 · p. 10

3.3.4. Design load estimation

• The design load estimate is the model’s predicted thermal power needed to maintain indoor temperature under design conditions.It uses the estimated inverse resistance and exogenous heat together with the design indoor-outdoor temperature difference.Source: §3.3.4 · pp. 10–11

3.3.5. Uncertainty quantification

• The method reports confidence intervals for both the fitted heating-load relationship and the estimate at the design temperature difference.The example figure distinguishes the 95% confidence band across temperature differences from the 95% interval at the design condition.Source: §3.3.5 · p. 11, Figure 4 · p. 11

3.4. Utility bill method

• The bill method converts fuel consumption to supplied heat using equipment efficiency and applies the shared 1R1C model.The described dataset uses monthly natural-gas bills; the approach could also use electricity bills for electric resistance heating or heat pumps with accurate COP curves.Source: §3.4 · pp. 11–12
• The method assumes indoor temperature equals a representative setpoint θ and that heating operates only below an outdoor-temperature threshold.Equipment efficiency is obtained from specifications and assumed time-invariant.Source: §3.4 · p. 12

3.4.1. Data pre-processing

• Each billing period supplies one observation using local hourly weather data and monthly fuel-derived heat supply.The input data typically span a year, so summer months are filtered out by the heating-month selection procedure.Source: §3.4.1 · p. 13
• The method selects a threshold τ⋆ to separate inactive months from heating months, fitting only months above that threshold.The threshold minimizes combined error for a constant non-heating-month heat-supply fit and a linear heating-month fit.Source: §3.4.1 · pp. 13–14, Figure 5 · p. 13

3.4.2. Uncertainty quantification

• For bill-method uncertainty, Bayesian bootstrap intervals address the very wide analytical intervals associated with few heating months.A one-year dataset typically has no more than seven heating months; the largest analytical interval in the dataset spanned over 23.5 kW.Source: §3.4.2 · p. 14

4.1. Survey demographics

• 43 of 74 houses (58%) were in climate zone five; most had 165–262 m² of floor area and were built between 1960 and 2005.Participating houses spanned U.S. climate zones two through six.Source: §4.1 · pp. 14–15, Figure 7 · p. 15

4.2. Thermostat method results

• Thermostat-method R² averaged 0.89 and had a median of 0.90, indicating that the first-order temperature-difference model explained about 90% of observed heat-supply variation on average.Nighttime averaging improved model fit; hourly averaging yielded mean and median R² values of 0.56 and 0.57.Source: §4.2 · p. 15

4.3. Bill method results

• Bill-method R² averaged 0.90 and had a median of 0.95, higher than the thermostat method, but its distribution had more low-end outliers.The authors attribute the outliers to fewer data points and greater variability in monthly bills across seasons and participants.Source: §4.3 · pp. 15–16, Figure 8 · p. 16
• The Bayesian bootstrap decreased 95% confidence-interval width by about 50% on average relative to the t-distribution analytical method.The authors state that narrower intervals should enable more accurate heating-equipment sizing.Source: §4.3 · p. 16

4.4. Method comparisons

• At the 95% confidence level, the methods’ central estimates were equal for 36% of houses; agreement within ±1.76 and ±3.52 kW held for 71% and 80%, respectively.The two tolerances correspond to typical residential HVAC capacity increments of 6,000 and 12,000 Btu/h.Source: §4.4 · p. 16
• Existing equipment averaged 2.75 and 2.96 times the thermostat- and bill-method estimates, while Manual J averaged 2.26 and 2.35 times those estimates, respectively.The commercial software estimates averaged 1.61 and 1.70 times the thermostat- and bill-method estimates; the rule of thumb produced the highest estimates.Source: §4.4 · pp. 16–17, Figure 9 · p. 17

5.1. Lessons learned from data collection

• Only 74 of roughly 5,000 initial contacts supplied complete data, after more than six months and over 1,000 individual emails.Sequential requests and incremental gift cards were effective; Manual J was requested last, and an online thermostat-data community brought in about 75% of participants who completed the study.Source: §5.1 · pp. 17–18

5.2. Implementation practicalities

• The thermostat method is viewed as more robust and near-measurement for about three-quarters of homes, but needs exportable smart-thermostat data and existing equipment capacity.Capacity was often obtained from nameplate photos, among the most challenging data to collect.Source: §5.2 · pp. 18–19
• The bill method is simpler to implement because it uses monthly bills and equipment efficiency rather than heating capacity.Households may estimate efficiency from equipment age and quality when nameplate efficiency is unavailable.Source: §5.2 · p. 19
• Six or seven monthly fuel-use observations can make bill estimates volatile; the Bayesian bootstrap mitigates but does not eliminate this, and design loads require extrapolation.Despite these constraints, the authors found bill-method estimates similar to thermostat-method estimates and view bills as better suited to low-cost implementation at scale.Source: §5.2 · p. 19
• Commercial software estimates average 1.61 times the thermostat-method estimates and 1.70 times the bill-method estimates, while requiring only a remote 10-to-20-minute housing survey.Source: §5.2 · p. 19

5.3. Limitations and future work

• Recruitment through an energy-enthusiast community may have selected participants with more efficiency retrofits than typical households, possibly lowering their design loads.The authors note that Manual J and the tested commercial software should in principle reflect the house as built, including retrofits.Source: §5.3 · p. 19
• The dataset contains detached U.S. houses, disproportionately in cold climate zone five, all heated by gas furnaces; attached and multifamily homes remain untested.Future work could test other housing types, climates, and equipment and compare methods with standards from other countries.Source: §5.3 · pp. 19–20
• Actual design-load ground truth was unavailable, although thermostat observations at design conditions made estimates nearly measurements for 72.9% of houses.Accurate thermal instrumentation can be costly and invasive; circuit-level metering is a proposed alternative for electric-resistance heating.Source: §5.3 · p. 20

• Both data-driven methods—one using thermostat data and one using monthly energy bills—had strong goodness-of-fit statistics and agreed reasonably well.Source: §6 · p. 20
• As implemented by practitioners, Manual J overestimated design heating loads by more than a factor of two relative to the investigated data-driven methods.Source: §6 · p. 20
• Other existing methods, including floor-area rules of thumb and like-for-like replacement, produced even higher design-load estimates than the data-driven methods.Source: §6 · p. 20