Proof, not promises

What a smart meter actually sees inside a factory

Every number and chart on this page is computed from 1,62,59,021 real meter readings — 3 months of a real power-electronics factory in Karlsruhe, Germany (the public HIPE dataset, Bischof et al., ACM e-Energy 2018). Nothing is mocked. This is what we would compute from your plant four weeks after installing meters.

Meter readings

1,62,59,021

one every 5 seconds

Meters

plant incomer + 10 machines

Days recorded

2017-10-01 → 2017-12-31

Energy metered

30,842 kWh

peak 68.15 kW

01 · The raw material

This is what recording usage looks like

One week of whole-plant power. Without a single site visit you can already read the factory's life: shifts start, machines cycle, nights idle, the weekend flatlines — but never to zero. That flat weekend floor is the first money insight, and it comes free with the data.

Whole plant, one week (Mon 16 – Sun 22 Oct 2017), 15-minute averages

Amber band = weekend. Notice the plant never drops below ~5 kW — that floor runs 24/7.

The plant's fingerprint: typical weekday vs weekend

Median ± interquartile band over 13 weeks. The gap between the green and amber lines is production; the amber line itself is what you pay for nothing.

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

Oct → Dec 2017 · amber ticks = weekends · scale 040+ kW

Columns = days, rows = hour of day. You can see shifts, weekends, holidays (Oct 3, Oct 31, Nov 1) and the Christmas shutdown — all from one meter.

data quality — Chip press

99.99%

14,50,066 readings, minutes covered

data quality — Chip saw

99.98%

14,50,066 readings, minutes covered

data quality — High-temp oven

99.98%

14,50,028 readings, minutes covered

02 · Why data quality decides everything

A utility meter vs. our meter — the same Monday

The government AMI rollout gives 15/30-minute billing data. Useful — but appliance and machine signatures live at seconds. Same day (Mon 4 Dec 2017), same plant, two resolutions. This difference is the product: research maps sampling frequency → which loads are detectable (Armel et al. 2013, Energy Policy).

What a billing meter sees

Whole plant, 30-min blocks

What our meter sees

Whole plant, 1-min (from 5-second readings)

Zoom to one machine: the chip press, 10:00–12:00, raw 5-second data

Every spike is a press cycle. Count cycles → energy per cycle → cost per unit produced. A 30-min meter turns this entire story into two flat blocks (grey line).

03 · Wow #1 — the 3 a.m. line

A third of the bill is burned while the plant is closed

“Closed” = nights 22:00-06:00, weekends, public holidays. This isn't a model — it's arithmetic on the meter stream. The plant never drops below 4.26 kW; the 3 a.m. median is 6.32 kW. None of the 10 metered production machines cause it — meaning it hides in ventilation, compressed air, standby and IT. Finding which is exactly what sub-metering does.

3 a.m. floor

6.32 kW

median 02:00–04:00

closed-hours share

33.8%

of all energy

closed-hours energy / yr

40,870 kWh

annualized from 92 days

that is, per year

₹3.3 lakh

at ₹8/kWh (your tariff — slider above)

Monthly energy: plant open vs plant closed

Even cutting the closed-hours load in half is worth ₹1.6 lakh/year — before touching production. Feedback-driven savings of 4–12% are the documented norm (ACEEE meta-review of 57 studies).

04 · Wow #2 — the itemized bill

One electricity bill, split by machine

A DISCOM bill is one number. Meter data turns it into a league table: which machine costs what, per month. And an honest surprise — the 10 production machines are only 4% of the bill. The other 96% (HVAC, compressed air, lighting, IT) is invisible without measurement — and that is where the idle waste from Wow #1 lives.

Machine league table (3 months)

Meter	kWh	cost @ ₹8
Rest of plant (unmetered: HVAC, compressed air, lighting, IT)	29,573	₹2,36,585
Vacuum pump 1	271	₹2,171
High-temp oven	249	₹1,988
Chip press	246	₹1,970
Soldering oven	187	₹1,495
PCB washing machine	82	₹654
Vacuum pump 2	66	₹526
Screen printer	64	₹508
Chip saw	51	₹411
Pick & place unit	43	₹343
Vacuum oven	11	₹86

Monthly stack, metered machines

Rest-of-plant excluded from the stack for scale (29,573 kWh ≈ ₹2.4 lakh). Disaggregating that block — first by cheap sub-meters, then by algorithms — is the roadmap (NILM works on industrial data: Kalinke et al. 2021).

05 · Wow #3 — the penalty meter

Two silent line items: power factor & contract demand

Indian industrial tariffs bill more than kWh: DISCOMs surcharge poor power factor (typically below 0.90) and penalize exceeding contracted kVA — or you quietly overpay for headroom you never use. Both are invisible until the bill arrives. From the meter stream we see them live.

Weekly average power factor — hugging the penalty line

December averaged 0.895— a penalty month on most Indian state tariffs. The fix (capacitor bank) is typically a sub-1-year payback; we'd size it from this exact data. Instantaneous PF is below 0.9 for 55% of all readings.

Monthly peak demand vs contract

contract60 kVA

Never exceeded — but the highest peak was 54 kVA: 6 kVA of paid headroom never used. Right-sizing the contract is a one-chart decision. Load factor here is 0.24 — the plant pays peak-shaped charges for an average-shaped load.

Load duration curve — how few hours drive the peak

Demand above 30 kW happens only a few percent of the time — classic peak-shaving / load-shifting territory under Indian time-of-day tariffs.

06 · Wow #4 — the machine nobody switched off

Found in the data: the chip press ran all weekend, 67 kWh for nothing

We did not plant this. Scanning nightly consumption per machine, the algorithm flagged the weekend of 2–3 December 2017: the chip press idled at ~1.4 kW around the clock — on a weekend when its normal consumption is 0 kWh. Someone forgot to switch it off on Friday. The meter noticed; nobody else did.

Chip press power, Friday 1 Dec → Monday 4 Dec (15-min averages)

wasted that weekend

67 kWh

≈ ₹536 at ₹8/kWh

if it happens monthly

₹6,432

per year, one machine

our alert would fire

Sat 06:00

WhatsApp/SMS: “Chip press still running, plant closed”

Detection = per-machine baseline of closed-hours consumption, flag when a night deviates. The same method catches degrading equipment (a compressor drawing 18% extra flags weeks before it fails) — demonstrated in the literature on real homes and industrial data (anomaly detection on REFIT, HIPE use-case taxonomy).

07 · Wow #5 — one action, quantified

The report ends with an instruction, not a dashboard

The counter-evidence in the research is clear: pretty pie charts alone don't save money (Kelly & Knottenbelt 2016, systematic review). Specific, quantified, verified actions do. Every monthly report leads with exactly one.

Action of the month · example from this data

Put the closed-hours base load on a schedule

6.32 kW runs 24/7; none of it is production. Walk the plant at 22:30 with our live per-circuit view, find what stays on (ventilation, compressors, chargers, IT), put the non-essential half on timers/contactors.

target

3.2 kW

half the night floor

saving / year

20,435 kWh

measured against baseline, not promised

worth / year

₹1.6 lakh

at ₹8/kWh

payback

~4 months

on ~₹60,000 of timers & controls

Assumptions shown, verifiable, and then measured: we baseline before, measure after (industry-standard M&V). Systematic energy management programs sustain ~4% savings year over year (US DOE 50001 Ready) and ~10% cumulative within two years (Natural Resources Canada) — this card is that practice, automated.

08 · The feature library

Everything we compute from one data stream

Each card is a feature computed (or computable) from the same 5-second stream you just saw — ordered from day-1 arithmetic to models. The five “wow” results above are these features, rendered well.

Base / standby load

sustained minimum (3 a.m. floor)

→ the single most reliable wow

Load duration curve

sorted demand vs % of time

→ peak charges in one chart

Peak demand + timestamp

max 15/30-min kVA

→ contract right-sizing

Power factor / kVArh

P vs Q per interval

→ penalty avoidance

Load factor

average ÷ peak

→ classic efficiency KPI

Shift / weekend profiles

median day-shape per category

→ consumption outside shifts = waste

ToD cost mapping

kWh × tariff slab

→ same usage, cheaper bill

Baseline models

expected vs actual per day

→ anomaly alerts + measured savings

Event detection

machine start/stop signatures

→ cycles, counts, maintenance flags

Disaggregation (NILM)

split one meter into machines

→ itemized bill without submeters

Day-ahead forecast

gradient boosting on lags + calendar

→ planning, solar pairing, ToD shifting

Supply quality log

sags, swells, outages

→ uniquely valuable on Indian grids

09 · Does the algorithm actually work?

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

Same discipline as our live German renewables forecaster: train on the past (57 days), predict a future the model has never seen (2017-12-04 → 2017-12-24), day-ahead horizon, and report errors against an honest baseline — not against zero.

First held-out week: actual vs forecast (15-min resolution)

Model	MAE kW	RMSE kW	MAPE
Seasonal naive (same slot last week)	2.69	4.3	19.6%
Our model (gradient boosting)	2.19	3.28	16.3%

error cut vs naive

−18.6%

mean absolute error

80% band coverage

59.9%

honest miss: test weeks include the pre-Christmas slowdown the model had never seen

Features: lag96, lag192, lag672, rollMean1d, hour, dow, isWeekend, isHoliday. Baseline: seasonal naive: same 15-min slot, previous week. The under-covered interval is exactly why pilots matter — one plant's December taught the model something no public dataset could. With a year of a plant's own data, holiday regimes become features, not surprises.

10 · 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. That is precisely what this product is designed around.

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 ↗
Industrial energy management sustains ~4%/yr savings	US DOE Better Plants / 50001 Ready ↗
~10% cumulative industrial savings within 2 years of ISO 50001	Natural Resources Canada ↗
NILM (single-meter disaggregation) works on industrial loads	Kalinke et al. 2021, ACM e-Energy (on this very dataset) ↗
The dataset behind this page	HIPE — Bischof et al. 2018, ACM e-Energy · Zenodo ↗