A 2026 paper on a small home energy monitor shows what happens when it guesses which of four appliances is running from one main meter. Overall precision falls to 77% [1]. The fridge scores 0.84, but the dish washer drops to 0.10, so a label in your app is right about one time in ten [1]. That is where the guess fails hardest.

The big picture:

A home energy monitor holds two tools in one box. A physical clamp around a wire measures the current flowing through it. Everything else you see in the app is just math making a guess.

The software watches the total power at your main breaker. Then it tries to figure out which device caused each spike. Experts call this splitting process load disaggregation, but an owner just sees a pie chart.

That splitting job is hard. In a 2026 study, a small machine-learning model hit 90% precision and 96% recall with just two appliances [1]. Adding two more loads dropped it to 77% precision and 80% recall, costing thirteen points of precision [1].

Precision means the share of the system’s labels that are right. Recall means the share of real power events it actually finds. Both numbers come from the researchers, not from a brand’s sales pitch.

The average hides the worst misses. The four-load table scores the fridge at 0.84 precision and 0.94 recall, and the electric furnace at 0.50 and 0.15 [1]. The dish washer lands at 0.10 and 0.08, meaning nine of every ten labels are wrong [1].

The paper explains why the fridge and the microwave separate so well. A fridge makes low-power shifts over long times, and a microwave makes high-power spikes over short ones [1]. The authors print the four-load table without explaining the dish washer’s failure.

The team tested a public set of six homes, using five to train the model and holding the sixth out [1]. The whole model fit into 17 kilobytes and answered in 0.43 milliseconds on a tiny computer chip [1]. That is the exact class of software running the breakdown on a cheap home device.

A second team measured this naming problem and found a similar pattern. Their pattern-matching neural network guessed which of four appliances was running with 88.9% accuracy and an 88% F1 score, a metric balancing precision and recall [2]. Their second network design reached 80.67% [2].

Accuracy inside one home ranged from 86.58% for the fridge to 94.6% for the dish washer. But the authors note this only works “for carefully chosen values of thresholds” [2]. That is their strict warning, not mine.

An open reproduction of this model family lands between a 0.5671 F1 score for the microwave and 0.8478 for the washing machine [3]. Different test data reveals the same wide spread in accuracy.

I keep coming back to what this means for a real house. Space heating and air conditioning made up 52% of annual home energy use in 2020 [4]. The loads an owner most wants to track are the two biggest power draws.

These are also two of the hardest signals to pull from a single main wire. A physical clamp on the air handler circuit settles the debate. A software guess does not.

By the numbers

  • 77% — Four-appliance precision: Pulling four appliances from one meter dropped the 2026 model from 90% precision on two loads to 77% on four [1].
  • 0.10 — Dish washer precision: Inside that same four-load model, the dish washer scored 0.10 precision and 0.08 recall, while the fridge scored 0.84 [1].
  • 88.9% and 80.67% — Four-class naming: A neural network reached 88.9% accuracy naming which of four appliances was running; a second model reached 80.67% [2].
  • 52% — Heating and cooling share: Space heating and air conditioning took up 52% of annual home energy use in 2020, representing the main loads a monitor is bought to watch [4].

What I’d watch:

Basic math settles this argument, not brand names. A clamp measures one physical wire and reports exactly what it saw. A software line reports what the math decided, even though the app makes both numbers look identical.

  • The clamp count: A monitor with a bank of branch clamps physically covers the circuits an owner acts on. Those include the air handler, the water heater, and the dryer. Each line is a hard measurement rather than a guess.
  • The inferred lines: A whole-home monitor without branch clamps still gives an honest total power reading. The individual breakdown above that total is where the 77% guess lives. It is a ranking aid, not a true reading.
  • The threshold caveat: Several published accuracy scores improve a lot with hand-tuned detection limits [2]. A limit perfectly tuned on one house tells you nothing about another house.
  • The separable shapes: The research keeps finding that big, sharp, high-power steps separate well [1]. Slow, low-power cycling does not separate well. That is the physical reason finding a fridge is easy and a furnace is hard.

What I expect next is basic honesty. The precision and recall figures in this piece come from research datasets. A product page that published its own accuracy on held-out homes would be the first one I have seen.

The catch

A clamp is perfectly honest about the wire, but not always about the device. Two loads sitting on one circuit become one combined number again. That is what a shared bedroom circuit or a kitchen counter run gives you.

The furnace line in the 2026 table scored 0.50 precision and 0.15 recall [1]. A furnace is exactly the kind of load that shares a circuit in many houses.

The benchmark test is also small. Six homes in one dataset is enough to show the general direction of the error [1]. It is not enough data to predict how the math will work in your house.

An owner who buys a monitor strictly for the appliance breakdown is paying for a weak feature. The total main usage and the hard-clamped circuit lines are the only parts that hold up.

I would still install one. The counter-argument is not that the software guess is useless. It is that the guess and the hard measurement arrive in the exact same app.

They carry the same visual confidence. But only one of them can be checked against a physical nameplate.

At a glance

  • The Big Shift: A panel monitor’s appliance breakdown relies on a math guess, not a true meter reading. In a 2026 peer-reviewed study, separating four loads dropped precision to 77%, and the dish washer landed at 0.10.
  • Why It Matters: Heating and cooling account for 52% of home energy use, and those are the exact loads an owner most wants to track. Making decisions based on a guessed number means acting on weak data.
  • What I’d Watch:
  • Clamp coverage: The total number of branch clamps decides how much of your app is actually measured.
  • Held-out accuracy: Whether a vendor publishes true precision scores on homes their model did not train on.
  • Threshold tuning: Accuracy claims that depend on hand-tuned limits do not travel reliably between different houses.
  • The Catch: A clamp only measures a circuit, and a shared circuit still leaves you with a single combined number. The published tests only cover a handful of homes, not yours.

Related reading

Sources

[1] Wu, Tan, Duan, Yakovlev, Rahman and Shafik, “A Real-Time Tsetlin Machine-based Non-intrusive Load Monitoring System on MCUs”, arXiv 2608.18780v2, accepted at the International Symposium on the Tsetlin Machine 2026 — https://arxiv.org/html/2608.18780v2 [2] Shahab et al., “Energy Disaggregation & Appliance Identification in a Smart Home: Transfer Learning enables Edge Computing”, arXiv 2301.03018v2 — https://arxiv.org/html/2301.03018v2 [3] St. Angel, “Energy Management Using Real-Time Non-Intrusive Load Monitoring” (goruck/nilm), replication results on the REFIT household dataset — https://github.com/goruck/nilm [4] US Energy Information Administration, “Use of energy explained: Energy use in homes” — https://www.eia.gov/energyexplained/use-of-energy/homes.php