Proof, not promises · residential edition

One smart meter. Every appliance, pinpointed.

Every number on this page is computed from 17,64,76,836 real meter readings — a full year of a real London home (the public UK-DALE dataset, Kelly & Knottenbelt, Scientific Data 2015 (CC-BY 4.0)). The house has 52 plug-level meters, but our algorithms only get to see the one main meter — the plug meters are used purely to check our answers. Nothing is mocked.

Meter readings

17,64,76,836

one every 6 seconds

Meters

1 mains + 52 ground-truth plugs

Days recorded

366

2013-03-01 → 2014-03-01

Energy metered

3,077 kWh

peak 6.14 kW

Data: a UK household (House 1: end-of-terrace Victorian home, London (built ~1905)). Costs on this page use an illustrative Indian residential tariff — adjust the slider above. The aggregate CT records apparent power (VA); plug meters record real power (W). For a home dominated by resistive loads the two track closely.

01 · The raw material

A home's life, read from one wire

One week of whole-house power. Morning kettle spikes, evening cooking peaks, the overnight flatline that never reaches zero. A factory has shifts; a home has habits — and both are legible from a single meter.

Whole house, one week starting 2013-09-23, 15-minute averages

Amber band = weekend. Each thin spike is an appliance switching on — that texture is exactly what the disaggregation below feeds on.

The home's fingerprint: typical weekday vs weekend

Median ± interquartile band over 52 weeks. The double hump — breakfast and dinner — is the universal residential signature.

366 days at a glance — every hour, colored by average kW

Mar 2013 → Feb 2014 · amber ticks = weekends · scale 01.5+ kW

Columns = days, rows = hour of day. Dinner time glows all year; the dark band is the sleeping house; gaps are recorder outages — flagged, not hidden.

data quality — Aggregate

99.55%

50,35,718 readings, minutes covered

data quality — Kitchen lights

99.55%

49,78,162 readings, minutes covered

data quality — Lighting circuit

96.74%

48,97,849 readings, minutes covered

02 · Why data quality decides everything

A utility meter vs. our meter — the same day

Utility AMI rollouts record 15/30-minute billing data. Appliance signatures live at seconds. Same day (2013-08-26), same house, two resolutions — the research maps sampling frequency → which loads are detectable (Armel et al. 2013, Energy Policy). This difference is why the disaggregation below works at all.

What a billing meter sees

Whole house, 30-min blocks

What our meter sees

Whole house, 1-min (from 6-second readings)

Zoom to one appliance: the kettle, raw 6-second data (from its check-meter)

A kettle boil is a ~2.4 kW rectangle lasting a couple of minutes. In 30-minute billing data (grey) it dissolves into a barely-visible bump — undetectable, unbillable, uncoachable.

03 · The centerpiece — disaggregation

We split the house from ONE meter. The plug meters agree.

This home has 52 plug-level check-meters. Our models never see them: they are trained on history, then given only the main meter for 91 held-out days (2013-11-29 → 2014-02-28) they have never seen. Below, every claim we make from one wire is graded against the plug-meter truth — per second, per event, per kWh. Activation rules follow the standard Neural NILM definitions (Kelly & Knottenbelt 2015).

Replay: Tuesday 31 Dec 2013 — a held-out day, from the top down

Top: the only signal our models are given. Below: for each appliance, what we said (bright) vs what its plug meter recorded (dim). Overlap = proof.

Kettle

ours

plug meter

Microwave

ours

plug meter

Toaster

ours

plug meter

Fridge

ours

plug meter

Washing machine

ours

plug meter

Dishwasher

ours

plug meter

00:0006:0012:0018:0024:00

The fridge cycles all day; the kettle fires in short bursts; the washing machine runs its programs. All read from the one grey trace at the top — the colored rows are our output, the grey rows beneath them are the check-meters we never looked at.

best appliance — washing machine

F1 0.967

precision 0.983 · recall 0.951

events pinpointed (test window)

3,259

individual switch-ons matched to plug-meter truth

energy attributed

253 kWh

truth: 248 kWh over 91 days

The honest scorecard — 91 unseen days, graded every 6 seconds

Appliance	F1	Precision	Recall	Events found	kWh ours / truth	Energy error
Kettle	0.89	0.90	0.87	390/412	26.5 / 29.9	-11.3%
Microwave	0.58	0.61	0.55	307/414	9.6 / 13	-26.3%
Toaster	0.77	0.73	0.82	192/210	19.6 / 17.7	+10.7%
Fridge	0.78	0.71	0.85	2277/2352	98.5 / 83	+18.6%
Washing machine	0.97	0.98	0.95	71/77	67.1 / 68.9	-2.5%
Dishwasher	0.56	0.50	0.64	22/27	31.3 / 35.1	-10.7%

Published UK-DALE benchmarks land in the same bands — strong on kettle, washing machine and fridge; harder on microwave and dishwasher, whose signatures overlap other loads. We publish the misses too: a model that is never wrong is a model nobody checked. Baseline for scale: guessing the appliance's average power at all times gives 26 W MAE on the kettle vs our 4.2 W.

The itemized bill, two ways: our single-meter estimate vs 52 plug meters

Ours — computed from the main meter only

11.8%

Truth — measured by the plug meters

10%

8.3%

4.2%

Kettle ₹186Microwave ₹67Toaster ₹137Fridge ₹690Washing machine ₹470Dishwasher ₹219grey = rest of house (586 kWh)

This is the product: an itemized electricity bill without installing a single plug meter. Here it is graded against a house that happens to have 52 of them.

Pinpointed events — spot checks against the plug meters (±60 s matching)

When	Appliance	We said	Plug meter says	Timing off by
Sat, Nov 30, 11:29 AM	Fridge	264 W · 22 min	182 W · 113 min	0 s
Sun, Dec 1, 03:00 PM	Kettle	2,380 W · 4 min	2,417 W · 4 min	0 s
Mon, Dec 2, 09:30 AM	Dishwasher	2,427 W · 51 min	2,780 W · 104 min	102 s
Thu, Dec 5, 07:46 AM	Toaster	1,607 W · 4 min	2,663 W · 4 min	0 s
Fri, Dec 6, 04:10 PM	Washing machine	2,139 W · 100 min	3,888 W · 101 min	0 s
Sat, Dec 7, 01:32 PM	Microwave	1,594 W · 2 min	1,565 W · 2 min	0 s
Sun, Dec 8, 10:16 PM	Washing machine	1,462 W · 85 min	2,082 W · 96 min	0 s
Wed, Dec 11, 10:42 PM	Dishwasher	2,648 W · 102 min	2,749 W · 108 min	360 s
Tue, Dec 31, 12:18 PM	Microwave	950 W · 0 min	1,632 W · 4 min	0 s
Tue, Dec 31, 01:50 PM	Washing machine	1,973 W · 96 min	3,845 W · 100 min	0 s
Thu, Jan 2, 11:13 PM	Dishwasher	2,435 W · 98 min	3,795 W · 102 min	270 s
Sun, Jan 5, 09:39 PM	Dishwasher	2,414 W · 105 min	2,377 W · 102 min	144 s

Per-appliance gradient-boosted trees (scikit-learn HistGradientBoosting): a classifier for on/off state + a regressor for watts-when-on, using only sliding-window features of the aggregate meter. Input: Channel 1 (whole-house aggregate, 6-second apparent power). Plug-level meters used only as training labels and held-out ground truth. Offline attribution: features may look up to 60 s ahead of each instant. Aggregate is apparent power (VA), plug meters real power (W) — magnitudes match closely only for resistive loads; the model learns the mapping. Train 2013-03-01 → 2013-11-29, test 2013-11-29 → 2014-02-28. Every row above is a single physical switch-on we located from the whole-house wire.

04 · Wow #1 — the 3 a.m. line

52.3% of this home's electricity is used by nobody

“Floor” = median whole-house draw, 02:00-04:00, every night of the window. While the household sleeps, the house draws 184 W, around the clock — routers, standby electronics, chargers, an always-plugged PC. Annualized that is 1,615 kWhof vampire load. This isn't a model — it's arithmetic on the meter stream.

3 a.m. floor

184 W

median 02:00–04:00

share of all energy

52.3%

floor power × every hour

vampire energy / yr

1,615 kWh

floor × 8,760 h

that is, per year

₹11,305

at ₹7/kWh (your tariff — slider above)

Caught red-handed at 3 a.m. — what the plug meters confirm is awake

Data logger pc	13 W	₹797/yr
Boiler	12 W	₹736/yr
Gige usbhub	8 W	₹491/yr
Adsl router	6 W	₹368/yr
Unmetered remainder of the floor	145 W	₹8,891/yr

Feedback-driven household savings of 4–12% are the documented norm (ACEEE meta-review of 57 studies) — and the always-on floor is where they usually hide.

05 · Wow #2 — the itemized bill

One electricity bill, split by appliance

A utility bill is one number. Meter data turns it into a league table: what the fridge costs, what the washing machine costs, per month. Here it is computed from the sub-meters; the disaggregation section above shows how close we get to this table from the main meter alone— that's the point: this report, without installing 52 plug meters.

Appliance league table (12 months)

Meter	kWh	cost @ ₹7
Fridge	343	₹2,398
Washing machine	218	₹1,526
Lighting circuit	177	₹1,238
Kitchen lights	154	₹1,079
Htpc	140	₹981
Dishwasher	138	₹965
Boiler	125	₹872
Kettle	122	₹854
Data logger pc	112	₹781
Tv	92	₹642
Other 42 sub-meters	675	₹4,724
Rest of house (unmetered: oven circuit, sockets without plug meters)	783	₹5,478

Monthly stack, top sub-metered appliances

Rest of house excluded from the stack for scale (783 kWh of 3,077 ≈ ₹5,478). Winter lighting and the electric shower circuit live in that block.

06 · Wow #3 — the appliance nobody switched off

Found in the data: the dishwasher ran through the night of 2014-01-01

We did not plant this. Scanning every night of the year per appliance, the algorithm flagged 2014-01-01: the dishwasher consumed 1.3 kWh between midnight and 6 a.m. — its typical night is 0.006kWh. New Year's Eve dishes, started at midnight. The meter noticed; nobody else did.

Dishwasher power, the day before → the day after (15-min averages)

that night

1.3 kWh

typical night: 0.006 kWh

if it became a habit

₹3,296

per year, one appliance

our alert would fire

01:30

WhatsApp/SMS: “Dishwasher still running”

Detection: each night's 00:00-06:00 kWh per plug meter vs the appliance's own trailing median (14+ nights); largest excess reported. The same baseline logic catches a fridge whose seal is failing weeks before it dies (anomaly detection on REFIT homes). On a tariff with night rates this run was cheap — on a flat Indian tariff it is not.

07 · Does the algorithm actually work?

Backtest: day-ahead home forecast, held-out weeks

Same discipline as our live German renewables forecaster: train on the past (317 days), predict a future the model has never seen (2014-01-19 → 2014-03-01), day-ahead horizon, errors reported against an honest baseline. A single home is far noisier than a factory or a feeder — kettles are not predictable — so the win is smaller, and we say so.

First held-out week: actual vs forecast (30-min resolution)

Model	MAE kW	RMSE kW	MAPE
Seasonal naive (same slot last week)	0.237	0.405	70.2%
Our model (gradient boosting)	0.148	0.27	39.5%

error cut vs naive

−37.6%

mean absolute error

80% band coverage

74%

single-home demand has fat tails; fleet-level forecasts calibrate much tighter

Features: lag48, lag96, lag336, rollMean1d, hour, dow, isWeekend, isHoliday. Baseline: seasonal naive: same 30-min slot, previous week. Where this earns money is one level up: DISCOMs and aggregators forecasting thousands of homes, where individual noise averages out and this per-home skill compounds.

08 · Don't trust us

Every claim, sourced

Some papers say usage feedback doesn't help much — they're in this table too. The honest reading: dashboards alone under-deliver; quantified actions, alerts and measured verification deliver. And unlike most NILM demos, every accuracy number on this page is graded against physical check-meters, not eyeballed.

Appliance-level feedback enables 4–12% household savings	Ehrhardt-Martinez et al. 2010, ACEEE meta-review of 57 studies ↗
Disaggregation + feedback is the value; the meter alone saves nothing	Armel et al. 2013, Energy Policy 52 ↗
Counter-evidence: pie charts alone average ~4.5% with positive bias	Kelly & Knottenbelt 2016, systematic review ↗
The activation rules our scorecard uses (thresholds, min-on/off)	Kelly & Knottenbelt 2015, Neural NILM, ACM BuildSys ↗
Sampling frequency determines which appliances are detectable	Armel et al. 2013, Energy Policy 52 ↗
Anomaly detection on real homes (REFIT) catches failing appliances	arXiv 2023, anomaly detection on REFIT ↗
The dataset behind this page	UK-DALE — Kelly & Knottenbelt 2015, Scientific Data (CC-BY 4.0) ↗