Let Every Device In: Why Inclusive VPPs Are More Accurate

Virtual power plants get more accurate as they get bigger, which makes the instinct to perfect every device before admitting it exactly backwards. ev.energy CEO Nick Woolley argues that VPPs should let every device in, and that lab testing handles the one error scale cannot.

By 2030, the United States will have 34 million electric vehicles on its roads. Today, fewer than 4% of US EVs are enrolled in an active managed charging program. That gap is the largest untapped power plant in the country, and much of it is held back by one belief: that a home charging device must be accurate on its own before it can be trusted as a grid asset.

The instinct is understandable. Grid operators need measurements they can rely on, and a noisy device looks like a risk. But perfecting every device before admitting it to a virtual power plant (VPP), a portfolio of distributed devices dispatched together as one resource, has the statistics backwards. It holds back growth, strands capacity the grid needs, and raises the cost ratepayers carry. VPPs should let every device in. That makes them more accurate, bigger, and better for customers.

The device was never the asset. The VPP portfolio is what matters.

EV flexibility does not enter power markets one charger at a time. A grid operator never dispatches your charger. It dispatches a fleet of thousands. So the question that matters is how accurate the fleet is. Uncertainty in a VPP comes in three kinds, and the fix for one does nothing for the others.

Random error is scatter around the truth. One device reads slightly high, its neighbor slightly low, each for its own uncorrelated reason. Pool them, and the errors cancel. Aggregate random error shrinks with the square root of the number of devices. Ten devices are roughly three times more accurate than one. Ten thousand are a hundred times more accurate.

Systematic error is a shared, directional bias, such as a whole class of devices reading high for the same reason. Pooling cannot cancel it, so it has to be handled on its own terms, as covered below.

Availability is not a metering error at all. A charger can be measured perfectly and still deliver nothing, because the car is not plugged in. Plug-in rates follow strong patterns and can move for many drivers at once. A larger, more varied fleet smooths them, though not by the clean square-root law, which is why availability is forecast rather than assumed.

The random half has been measured in a live trial. In ev.energy’s Operational Metering Trial with National Grid ESO, Great Britain’s grid operator, a pool of just 10 typical devices metering at ±2% met the operator’s ±1% accuracy requirement, and beyond 100 devices, aggregate error fell below ±0.5%. When the operator moved its accuracy rules from the device to the aggregate, five times as many EV assets became eligible, with no change to the hardware. That is the logic FERC Order 2222 asks US markets to adopt: treat the aggregation as the asset, because it is.

A VPP program that is more inclusive is more accurate.

Take a utility territory with 10,000 EVs and two ways to build a program. Program A follows a common design: it admits vehicles from a handful of approved brands through telematics and turns everyone else away or pushes them onto a single approved charger. Xcel Energy’s Optimize Your Charge program, run with WeaveGrid, works this way, admitting only Lexus, Tesla and Toyota through telematics. Its own evaluation filed with the Colorado regulator estimates that this design reaches 48% of EVs on the road [1]. Program B is built on open connectivity across vehicle and charger integrations. In New York, that approach reaches about 96% of drivers, and 100% in parts of the Northeast where ev.energy integrates directly with metering [1]. On the same territory, that is about 9,600 EVs.

Which program is more accurate per megawatt delivered? Program B. Not despite its breadth, because of it.

Assume both programs enroll the same everyday hardware. Aggregate random error falls with the square root of the number of devices, so doubling the fleet from 4,800 to 9,600 improves accuracy by about 1.4 times. Program B’s aggregate error comes out roughly 30% lower than Program A’s, from identical hardware, for no reason other than that there is more of it [2]. These figures are modeled and illustrative, chosen to show the mechanism. They are not measured program results.

The broader, more inclusive program is the higher-quality product, because it is broader. More coverage means more devices, and more devices mean a more accurate VPP. A vendor that artificially constrains access hands over a smaller fleet, fewer megawatts, and a noisier aggregate. Reach is accuracy.

Excluding devices destroys value three times over.

Programs often respond to noisy device data by tightening eligibility. The impulse is reasonable. The effect is costly.

First, it worsens the customer experience. Drivers who bought a perfectly good EV or charger are turned away from a program they want to join, for reasons they cannot see and did not cause. Second, it strands capacity, because every excluded device is a megawatt the grid needed and did not get. Third, it makes the program less accurate. In our modeling, a program of 1,000 EVs achieves roughly ±0.8% fleet accuracy. Exclude the 300 devices judged “worst,” and the remaining 700 deliver ±1.0% [2].

Ratepayers pay for all three. Fixed costs spread across fewer devices, less capacity reaches the grid as a VPP, and bills rise for everyone.

Aggregation won’t cancel systematic error, so test for it and correct it.

Aggregation has a limit. If a whole class of devices reads 3% high, every one of them is wrong in the same direction, and pooling does not fix it. The bias survives at any scale. It has to be measured and corrected, and that is a solved engineering problem, cheap relative to the value at stake.

At our hardware lab at the Brooklyn Navy Yard in New York, we test for it directly. Representative devices from each class in the portfolio go on the bench against settlement-grade reference equipment, and the offset is characterized across converter accuracy, charging speed, metering behavior, telematics reporting, voltages, and real operating conditions. You cannot correct a bias you have not measured, and you cannot measure it from software alone. The result is a known correction factor per device class, applied across the portfolio.

The study repeats every year because the fleet does not stand still. Hardware generations turn over, firmware updates change metering behavior, and grid conditions shift. An annual re-test keeps the view of systematic error current, so it can be corrected before it compounds [2]. Engineering handles this everywhere else. A circuit breaker is not metered every time it trips. Its design is type-tested once, certified and audited so the population still matches its certificate.

Scale solves random error. Lab testing and annual re-tests solve systematic error. A well-run VPP does both and never confuses the two jobs.

Three steps deliver best practice in VPP accuracy.

1. Let every device in. Each added enrollment improves accuracy and adds grid value at near-zero incremental cost.

2. Design for volume and diversity, not device fidelity. A large, mixed fleet of everyday hardware, properly corrected, beats a small, curated fleet per megawatt and per dollar.

3. Test annually, correct systematically. Run yearly lab studies on representative devices from every class, and apply the measured corrections portfolio-wide.

The devices are already in the driveways. The math is already on their side. The bias is measurable, correctable and cheap to manage. Every device added makes the VPP more accurate, and the grid can bank on it.

Precision is not something you buy at the device. It is something you build at the portfolio.

How many of the EVs in your territory is your program turning away? Book a demo

Sources

1. The 48% eligibility figure and program design (Lexus, Tesla and Toyota via WeaveGrid telematics, ChargePoint Home Flex as the sole eligible charger) are from Xcel Energy’s Managed Charging Evaluation: Technical Evaluation, Hearing Exhibit 104, Attachment 10, Colorado PUC Proceeding No. 26A (Optimize Your Charge program). ev.energy coverage in New York (about 96%, up to 100% where integrated with metering) is ev.energy internal analysis.

2. Illustrative figures applying the Central Limit Theorem method set out in Appendix B of Metering Matters: Unlocking Electric Vehicle Charging Flexibility Through the Asset Meter (ev.energy, National Grid ESO, Flexitricity, 2024), where aggregate random error is proportional to 1/√N. The fleet-accuracy and exclusion numbers are illustrative, chosen to show the mechanism rather than to report measured program results.