An energy saving is a difference between two numbers, and only one of them is measured. After a retrofit you can meter what the site uses. What it would have used without the retrofit is the adjusted baseline: the output of a model fitted to data collected before the change, fed with the conditions of the reporting period. The saving is only as good as that model, and the model is only as good as the data at the measurement boundary.
Raw change and adjusted saving
A food plant replaced its lighting and lighting controls. Its electricity use and production for the last baseline year and the first reporting year were:
| Production | Electricity | |
|---|---|---|
| Baseline year, measured | 12,000 t | 1,200,000 kWh |
| Reporting year, measured | 11,040 t | 1,020,000 kWh |
| Raw change | 180,000 kWh | |
| Adjusted baseline from the regression model | 11,040 t | 1,142,400 kWh |
| Adjusted saving | 122,400 kWh (10.7%) | |
| Adjusted baseline with proportional scaling, for comparison | 11,040 t | 1,104,000 kWh |
| Saving with proportional scaling | 84,000 kWh |
The model was fitted to 24 monthly readings before the retrofit:
E = 40,000 kWh + 60 kWh/t × P
E is the monthly electricity use and P is the monthly production in tonnes. The intercept is the fixed load: refrigeration standing losses, compressed-air leaks, lighting and ventilation that run whatever the output. Over twelve months it is 480,000 kWh. Heating degree-days were tried as a second variable and dropped. Their t-statistic was 1.1, below the 2.0 that IPMVP asks for, because the site heats with gas outside the electrical boundary.
The raw change is about 1.5 times the real saving, because production also fell. Proportional scaling (1,200,000 kWh × 0.92) assumes that the fixed load falls with output. It does not, so that method understates the saving by 38,400 kWh, or 31%. Two defensible-looking methods give answers 46% apart, and this is why the plan must name the model before anyone sees the result.
This result is avoided energy: the baseline is projected to the conditions of the reporting period. A normalised saving projects both periods to a fixed set of conditions, such as a typical weather year or a planned production volume. Use it when you compare years with each other or with a target, and state which one you report.
Model fit and savings uncertainty
Three statistics decide whether a regression model can adjust a baseline. Each variable needs a t-statistic above 2.0. The model's R² should be at least 0.75, which is the IPMVP value that the DOE guide quotes. CV(RMSE), the standard error divided by the mean, measures the scatter that remains. The plant's model has an R² of 0.80 and a CV(RMSE) of 4.5%.
ASHRAE Guideline 14 then asks for the uncertainty of the reported saving. For a monthly model with uncorrelated residuals, its Annex B gives the fractional savings uncertainty (FSU) as:
FSU ≈ 1.26 × t × CV(RMSE) × √((1 + 2/n) ÷ m) ÷ F
Here n is the number of baseline points, m is the number of reporting points, F is the saving as a fraction of the adjusted baseline and t is the Student t value for the confidence level. For the plant, n = 24, m = 12 and F = 0.107. At 90% confidence with 22 degrees of freedom, t = 1.72:
| Confidence | FSU | Saving |
|---|---|---|
| 68% (t = 1) | 16% | 122,400 ± 19,500 kWh |
| 90% (t = 1.72) | 27% | 122,400 ± 33,400 kWh |
Guideline 14 requires the uncertainty to be less than 50% of the annual reported saving at 68% confidence, so this model passes. The two rows describe the same model and the same data. A figure without its confidence level cannot be compared with anything.
The CV(RMSE) limits most often quoted from Guideline 14, 15% with an NMBE of 5% for monthly data and 30% with 10% for hourly data, are calibration criteria for a simulation (Option D). They are not the test for a regression baseline. For a regression, the test is the savings uncertainty above.
F is in the denominator, so small savings fail first. With the same model, a 5% saving has an FSU of 58% at 90% confidence. A model with a CV(RMSE) of 15% gives an FSU of 195% for that 5% saving, which is a result that cannot be told apart from zero. The DOE limit on whole-facility analysis in the next section comes from this arithmetic.
Guideline 14 also limits where a model applies. Use it only for periods where each independent variable is within 90% of the lowest and 110% of the highest value in the baseline data. Every reporting month at the plant was inside that range. A month with a two-week shutdown would not be. Report it separately and state that it is outside the range of the model.
The four IPMVP options
The International Performance Measurement and Verification Protocol (IPMVP), published by the Efficiency Valuation Organization, defines four options. The DOE's M&V guidelines use the same four:
| Option | What is measured | Suits |
|---|---|---|
| A, retrofit isolation, key parameters | One key parameter, such as the power of the new lighting circuits. The other parameters, such as operating hours, are stipulated: agreed and fixed in the plan | Measures where one parameter carries most of the uncertainty |
| B, retrofit isolation, all parameters | All the energy of the retrofitted system. Spot or short-term metering is enough when the load does not vary; otherwise, meter continuously for the whole reporting period | Measures with a variable load, such as drives, chillers and compressors |
| C, whole facility | The main meter or a submeter for the whole building, with a regression model on the independent variables | Several interacting measures with a large combined saving |
| D, calibrated simulation | An energy model of the building, calibrated to whichever period has measured data | New buildings, where only the reporting period can be measured, or measures such as glazing that cannot be metered directly |
A stipulated value moves risk. If the plan stipulates 4,000 operating hours and the site runs 3,200, the reported saving is 25% too high, and the party that accepted the stipulation carries the difference. The DOE guide warns that stipulation can shift risk to the customer. Stipulate only values you can document, and measure the parameter that drives the saving.
The DOE guide limits Option C to projects where the saving is expected to exceed about 10% to 15% of the consumption measured monthly. It asks for at least 12, and preferably 24 or more, months of baseline data and at least 9, preferably 12, months of reporting data. The plant's 10.7% is at the bottom of that range, and its low CV(RMSE) is what makes the result usable.
Options A and B draw the boundary around one system, so they miss interactive effects. New lighting reduces the internal heat gain: the heating load goes up and the cooling load goes down, and neither is on the lighting circuit. Estimate that effect separately, or use Option C, where the main meter sees it. For a single measure on a large site, meter the system itself with option A or B.
Plan before you change anything
Write the M&V plan and start the baseline metering before the retrofit. When the old equipment goes, you can no longer measure the original operation. The plan names:
- the measurement boundary: one system, one building or the whole facility;
- the baseline period, in whole years (12, 24 or 36 months) so that no season is counted twice;
- the independent variables, their units and their source, which the model uses for routine adjustments;
- the static factors, such as floor area, shift pattern and connected equipment, and how a non-routine adjustment is calculated when one changes (a new production line, a new tenant);
- every stipulated value and who carries its risk;
- the model form, the fit criteria, the confidence level for the uncertainty and who runs the calculation.
ISO 50015 sets out the same M&V principles for an organisation, and ISO 50006 covers energy baselines and energy performance indicators in an ISO 50001 system. An M&V plan written to IPMVP fits inside both.
Data to collect with the meter readings
| Record | Why and how |
|---|---|
| Energy (kWh) and demand (kW) at the boundary, at 15-minute or hourly interval | kWh gives the saving and kW gives the avoided demand charge. Interval data shows schedule changes that a monthly total hides |
| Operating schedule and shift pattern | Separates fewer operating hours from better efficiency |
| Production in the unit that drives the energy | Use tonnes, not batches: a batch count hides batch size, and the regression needs the physical driver |
| Heating and cooling degree-days, where the boundary includes heating or cooling | Use one named weather station throughout. Treat the base temperature as a model parameter: Guideline 14's change-point model fits it from the data |
| Equipment and settings, before and after | Nameplates, setpoints and BMS time schedules are the evidence for a non-routine adjustment and for later persistence checks |
| Meter changes, data gaps and estimates, with dates and reasons | Guideline 14 requires a reason for every gap, exclusion and estimate in the baseline and reporting data |
Monthly bills give 12 to 24 points. That is enough for Option C when the saving is large and the model is tight, as at the plant. Interval data matters in three places. Option B needs it. Measures that change schedules show at night and at weekends, and those hours disappear into a monthly total. A non-routine event shows in days, not at the next bill. For the regression, the DOE guide suggests aggregating interval data to daily or monthly values. Daily residuals are usually autocorrelated, so the monthly FSU formula understates their uncertainty, and Annex B of Guideline 14 gives the correction.
Keep the raw data. Record every exclusion and every estimate, with the reason.
Reporting the saving
The report states the adjusted saving in kWh and as a percentage of the adjusted baseline, with the FSU and its confidence level. For the plant: 122,400 kWh, 10.7%, ± 33,400 kWh at 90% confidence, with the dates of the baseline and reporting periods. It includes the model with its coefficients, R², t-statistics and CV(RMSE), every non-routine adjustment with its calculation, and every excluded period with its reason.
Value the saved kWh at the reporting-period tariff, split by time-of-use band. The electricity cost calculator helps with the band split. Report the avoided kW demand charge as a separate line, and keep implementation cost and non-energy benefits out of the energy figure.
When a static factor changes, apply the non-routine method from the plan and record it. Do not rebase the baseline without a record to make a target look met. Where a grant pays for the project, the verified saving can release funding: for EXEED projects, SEAI holds back the lower of €30,000 or 10% of the grant until the asset achieves EXEED certification or the energy savings are verified. See the EXEED grant guide.
Where savings claims go wrong
| Failure | How it shows | What to do |
|---|---|---|
| Abnormal operation in the baseline, such as a month-long shutdown or a failing compressor | Large residuals in those months and a high CV(RMSE) | Record the cause. Exclude the months only with a stated reason, or add a variable that explains them. Guideline 14 allows no more than 25% of the measured data to be excluded from a whole-facility model |
| A meter or its current transformers replaced during the project | A step in the residuals on the date of the change, often a ratio error | Run the old and new metering in parallel for an overlap period, and record the ratio and the date |
| Two measures claim the same kWh | The sum of the per-measure savings is larger than the saving at the main meter | Assign each kWh to one measure, or claim interacting measures together under Option C |
| Reporting conditions outside the baseline range | Production or degree-days below 90% of the baseline minimum or above 110% of the maximum | Do not apply the model to those periods. Extend the baseline or report them separately |
| Savings decay | The monthly residuals shrink and the cumulative sum (CUSUM) of residuals flattens | Check the settings that the measure depends on |
Savings from schedules and controls decay first. A BMS time schedule is put in hand for a weekend event and left there. A drive is switched to bypass at 50 Hz during a fault and never switched back. Neither change shows in a monthly bill until several months have passed. The DOE guide asks for Option A short-term measurements to be repeated at least annually. With interval data, plot the CUSUM every month. A rise in the night-time load is often the first sign, and the overnight baseload guide shows how to read it.
Measuring savings with EpiSensor
For options A and B, install ZEM electricity monitors on the circuits to be retrofitted before the work starts, so that the baseline is metered at the boundary. A ZEM is supplied with its current sensors calibrated to it, and Class 0.5S to IEC 62053-22 applies to the complete measurement. This keeps the instrument term of the uncertainty small. It reports over the Zigbee mesh, so it can go in weeks before the project without a data cable. For Option C, a ZPC pulse counter reads the pulse output of the existing main meter, where it has one. The site then has interval data at the boundary where it previously had monthly bills. A Gateway running Edge records the readings locally, which gives a second copy of the raw data that the plan says to keep.
Demand response events use a different baseline, built from the days before each event. See the demand response baseline guide and the rest of the buildings and process performance learning path.
Common questions
How do you measure energy savings?
Fit a model of energy use against its drivers, such as production and degree-days, to baseline data collected before the change. Feed the reporting-period values of those drivers into the model to get the adjusted baseline. The saving is the adjusted baseline minus the measured consumption.
What are the IPMVP options?
Option A measures the key parameter of the retrofit and stipulates the others. Option B measures all the energy of the retrofitted system. Option C measures the whole facility at the main meter with a regression model. Option D uses a calibrated simulation.
How long should an energy baseline be?
Long enough to cover the full range of operating conditions that affect use. For whole-facility analysis the U.S. Department of Energy's guide asks for at least 12, and preferably 24 or more, months of data. For one system, spot or short-term metering is enough when its load does not vary. When the load varies, meter it continuously.
How accurate does a savings model need to be?
Each variable in the model should have a t-statistic above 2, and IPMVP suggests an R² of at least 0.75. ASHRAE Guideline 14 requires the uncertainty of the reported saving to be less than 50% of the annual saving at 68% confidence. State the confidence level with every uncertainty figure.