M&V tutorial · interval-meter baseline
Baseline from interval meter: how a VPP reduction is measured
Every demand response and virtual power plant (VPP) payment rests on a number nobody can meter: what the site would have used. Here is how interval-meter data turns into that number, and how far to trust it.
VPP Economics has no affiliation with, sponsorship from, or endorsement by any utility, aggregator, platform, or device maker named here.
The short answer
- A baseline is the site’s own history, projected onto event day. Recent interval-meter readings, sometimes with weather, estimate what it would have used. The reduction is the baseline minus the meter (U.S. Department of Energy).
- It is the default almost everywhere. NAESB calls it Baseline Type I, and PJM, NYISO, ISO-NE, MISO, CAISO, ERCOT and IESO all settle demand response with it (NAESB, NYISO survey).
- The measurement is only as good as the baseline. The meter is exact; the baseline is a forecast. For one home the best California baselines were off by 20–24% in a typical hour, against 1–3% for the whole program (CPUC).
- Error is measured on load but paid on the reduction. If a fleet cuts 25% of its load, a 3% error in the baseline is a 12% error in the credited kW.
From meter data to measured reduction
Ten days of meter data become a baseline
A worked example for one home and a 4–7 p.m. event, using CAISO’s standard 10-in-10 rule.
Swipe to see the full chart →
Event day: baseline minus meter
The dashed line is the adjusted baseline; the bars are what the meter recorded. Average reduction: 1.79 kW over three hours.
Swipe to see the full chart →
Three steps, every time: pick the baseline days, average them hour by hour, then nudge the result toward the event day’s own conditions. Here the morning ran 16% hotter than usual, so the baseline rose 16%. Without the adjustment, the same home would be credited about 1.41 kW instead of 1.79 kW.
Two ways to build it: averaging days or a model
Average recent similar days
“X of Y” rules: the highest 4 of 5, the middle 8 of 10, all 10 of 10. Simple to audit and to settle quickly. Weak on the hottest days, which are exactly when events happen, so most rules add a same-day adjustment.
Model load against weather and time
Fits each hour’s load to temperature, day type and season, then predicts event day. ERCOT’s default regression needs about 270 days of history. With a good fit it handles heat waves better than averaging, as long as the event day is not far outside the data (ERCOT).
A regression baseline, one event hour
Each dot is a past non-event weekday at 5–6 p.m. The fitted line predicts event-day load from the temperature.
Swipe to see the full chart →
Which is more accurate depends on the load. In a Lawrence Berkeley National Laboratory (LBNL) study of commercial buildings, a morning adjustment improved every model, and weather models were the only ones that consistently avoided bias (LBNL). A later LBNL test on 453 meters found every method ran low on peak days, by a median of 4.5% to 18.7%, and regression was no better than averaging (LBNL).
Which grid operators use it
How each grid operator builds the baseline
Each square is a recent eligible day; filled squares are the days the rule averages.
Swipe to see the full chart →
Two patterns stand out. Rules that pick the highest days (PJM, NYISO, IESO) set a higher baseline and credit more for the same behavior than a straight average. ERCOT drops the highest and lowest days. SPP is the outlier: the participant forecasts its own baseline, which SPP checks against a live meter snapshot (SPP).
Accuracy: one site or a whole fleet
Baseline error for one site versus a whole program
Typical hourly error (CVRMSE or RRMSE) as a share of baseline load, plus measured bias on hot days.
Swipe to see the full chart →
Random error shrinks as sites are added; bias does not. That is why the same baseline can be poor for paying one household and fine for crediting a fleet. California’s evaluators found that even unbiased baselines systematically over-credit some homes and under-credit others (CPUC).
One baseline per meter
Needed when each customer is paid on their own performance. Fine for steady commercial loads; noisy for homes. PJM requires each site to pass a 20% error test or use an alternative baseline (PJM).
Add the meters, then baseline
Sums every site’s interval data and builds one baseline for the total. Site-level noise averages out, which is why VPP aggregations are settled this way where rules allow (DOE).
One regression across many sites
Estimates the weather response from all sites at once. Individual fits carry more estimation error than a pooled fit, though the pooled model hides how individual sites differ (DOE).
How the VPP Program Designer uses these ranges
Interval-meter accuracy ranges in the designer
Assumed bias and accuracy, as a share of the credited reduction, for homes and for commercial and industrial (C&I) sites.
Swipe to see the full chart →
In the VPP Program Designer, Step 5 applies these ranges to the fleet’s dependable kW. A downward bias lowers the credited MW and the pay that follows. A wider range lowers how much capacity a planner can count on.
Where it goes wrong
Gaming the adjustment hours
Running equipment harder in the adjustment hours raises the baseline. That is why rules set those hours well before the event.
Heat-wave days
Event days are usually the hottest of the season. Unadjusted averages run low, under-crediting real reductions.
Back-to-back events
Event days are excluded from the baseline, so a run of events pushes the rule back to older, cooler days.
Batteries already cycling
A battery that discharges every evening already lowers the meter history, so an event shows less “reduction.” Device metering avoids this.
Rooftop solar
Net meters mix load and solar. CAISO lets participants baseline gross load and meter the generator separately (CAISO).
Variable loads
Sites whose load swings day to day fail accuracy tests and move to custom baselines or maximum base load.
All six ways performance is measured
Baseline from interval meter
You are here. Past meter data, sometimes with weather, predicts what the site would have used.
This pageFor short eventsMeter before / meter after
Usage just before an event against usage during it.
See the tutorial →For fleetsStatistical sampling
A metered sample stands in for the whole fleet.
See the tutorial →For batteriesMetering device output
Measures what the device itself delivers.
See the tutorial →For firm loadsMaximum base load
Pays for staying below a pre-agreed level.
See the tutorial →For residential programsControl group
Compares dispatched homes with homes held back.
See the tutorial →Method and caveats
A tutorial for program designers, utilities, aggregators and other readers. It is not a substitute for a tariff or settlement manual. Worked examples use illustrative numbers. Operator rules are weekday energy or reliability rules as published by each operator; IESO and MISO summaries come from 2020–2021 stakeholder materials and may have been updated. Checked October 23, 2026.
Ninety-percent confidence ranges are 1.645 times the typical error and assume errors are roughly normal. The load-to-reduction conversion divides the baseline error by the reduction as a share of load.