Passion project · proof of conceptAI agents as power brokers
Street-level energy agents

One street. One transformer.
An agent in every home.

Each home saves for itself. A station agent keeps the street’s cable in check. Tested on last winter’s real French data.

━ OFF-PEAK TIMERS
╍ AGENTS + STATION
Transformer limit · 400 kVA (assumed)
00:0006:0012:0018:00 · Fri 17 Jan 2025, a red day
The problem02 / 11
What we found

Smart homes, each acting alone, make the street worse.

1,080

kVAh over the limit with simple off-peak timers

2,764

with smart homes optimising alone: they all pick the same cheap hours

624

with our station agent: −77 % vs smart homes, −42 % vs timers

Winter 2024–25 · 222 simulated homes · 400 kVA transformer · median of 5 random streets

Why now03 / 11

Old streets, new loads.

22

red days each winter on EDF Tempo: every smart home reacts to the same signal

EVs

and heat pumps arriving on streets wired for yesterday’s demand

LLMs

small enough to run locally, good enough to turn plain words into settings

Meet Rue des Voisins04 / 11

222 homes and shops, one transformer.

The Martinselectric radiators, EV
Tank houseselectric hot-water tank
Frozen-food shopfreezers that can pre-cool
Click & collectcold storage, opening hours
The blocksflats matched to real meter totals
Stationthe referee

Real building stock from public data (BDNB); the households are synthetic.

How it works05 / 11
Every evening, before tomorrow

Homes plan. The station referees.

18:00

Brief

Each home reads only what is known tonight: prices, Tempo colour, forecast

18:05

Plan

An optimiser finds the cheapest plan that keeps every comfort rule

18:10

Check

The station adds up all plans against the transformer limit

18:15

Nudge

The biggest contributors re-plan; a change is kept only if the street gets better

18:20

Explain

Each home gets its plan in plain words

Optimiser · the maths engineLanguage model · plain words in and outCheckers · block anything a resident ruled out
Why agents, not a timer06 / 11
A resident’s note, as written

I leave for work at half past eight. I’m back and plugged in by eight in the evening. I need the battery about four fifths full.

What the agent sets
leave 08:30back 20:00battery 80 %laundry Mon Tue Wed Sat
98.5 %of settings read correctly from test notes by a small open model
94 %on 20 hand-written notes, 6 of them in French
0invalid settings reached a plan: a checker rejects them first
How we prove it07 / 11

Real data.
Last winter.

Public French data
RTE eCO2mixTempo calendarEnedis load curves BDNB buildingsOpen-Meteo forecastsENTSO-E prices
Per electric-heated homeTempo tariff, vs off-peak timer, per year
€293
Overload at the transformerstation vs smart homes alone, winter
−77 %
Comfort brokenhours below anyone’s minimum
0
Forbidden changes blockedwhen a model said yes to a resident’s “never”
17 / 17
No peekingagents only see what was known at 18:00 the day before
Whole seasonsnot hand-picked days
Tough baselinestimer · perfect-foresight optimiser
Results08 / 11

One hard week, whole seasons.

HERO WEEK
Cold wave · 13–19 Jan 2025 · red days

631 → 222 kVAh over

The station cuts the overload by two thirds; with a 5 % safety margin, to 51

Whole winter · Nov–Mar

€38 for the whole street

The cost of coordinating: 12 homes pay a little more, at most €36 over the winter

Summer

Little to shift

Savings come from winter. Moving off-peak hours to midday lowers bills but raises the summer peak

Tomorrow · half the homes with an EV

The cost concentrates

About 126 homes pay more, up to €81 each. Someone has to settle that fairly

What we have not shown09 / 11

Honest about the limits.

Simulated, not meteredSavings come from a backtest on real data, not from real bills
Synthetic householdsReal buildings and real meter totals, invented families
The transformer is a guessIts size, and whether the big shop shares it, only the network operator knows. Without the shop, today’s street stays under 400 kVA
Claude not yet testedAll agent results so far use small open models on our own GPU
Market10 / 11

Others manage a home.
We coordinate the street.

Voltalistop-down: switches heating off at peaks
Tibberone home at a time: smart EV and battery
Enogridshares solar inside communities
Usbottom-up agents that share one transformer
Network operatorsstreet-scale flexibility, like today’s local flexibility tenders
Residentsa free app: plain words, plain answers, a veto
Fairness ledgerwho gave flexibility, who gets paid back
What’s next11 / 11

Rue des Voisins first.
Then every stressed street.

Next

Ask the network operator for the real transformer and who shares it

Then

Claude agents vs open models, same tests

Looking for

One real street to pilot on

Aditya Sankaran · Télécom SudParis