RXed AI News

The power peak watcher: the 15 minutes that set your electricity bill

One quarter-hour a month prices a line you never read - this warns you while you can still stop it

Part of your electricity bill is not priced on how much you used. It is priced on your single busiest quarter-hour. In Flanders that costs roughly €53 per kW per year; in the US a demand charge of $8–15 per kW per month is normal and trade sources put it at 30–70% of a commercial bill. Nobody watches it, because by the time the bill arrives the quarter-hour is four weeks old.

The problem

You switch on the oven, the compressor and the air conditioning inside the same fifteen minutes. Nothing breaks, nobody notices, and that moment sets a charge you keep paying. It is the same logic as sizing a boiler for the coldest hour of the year and not for the average winter day: you pay for the capacity you demanded, not for the average you ran at. Ground that back to a real bill and the numbers are these.

Flanders (and SME connections, not just households). The grid operator measures every 15 minutes. Your highest quarter-hour in a month is your month peak, and what you are billed on is the average of the last twelve month peaks, with a floor of 2,5 kW. The Flemish regulator puts the 2026 average at about €53,39 per kW per year excluding VAT, roughly €4,45 per kW per month. Cut your average month peak by 5 kW and that is about €265 a year off the bill, before VAT, with no change in what you actually consume.

United States. Commercial tariffs bill a demand charge on the highest 15-minute (sometimes 30-minute) draw in the billing period, commonly $8–15 per kW and higher in some territories. Trade and utility sources put demand charges at 30–70% of a commercial electricity bill. Many tariffs also carry a ratchet clause that holds 80–90% of your recorded peak for the next 6–12 months, so one bad quarter-hour is charged for most of a year. Check your own tariff sheet for the ratchet before you assume it does not apply.

Elsewhere in Belgium, the lever is different. Wallonia moved its network tariffs to peak and off-peak hours in January 2026 (peak 07:00–11:00 and 17:00–22:00, every day including the weekend) rather than charging peak power. Brussels has a limited capacity component. Same instinct, different lever: there you move load in time rather than flatten it. Read your own bill first — that line item is the whole business case.

The architecture

One box, one reader, one push channel. The whole engine is two helper sensors and one automation — there is no model in the loop and there is nothing to train. If you already run the cold-chain alarm from this library, this is the same box doing a second job.

The one piece of arithmetic that makes it work. If you have drawn 1,0 kWh in the first five minutes of a quarter, your average for that quarter is already at least 1,0 / 0,25 = 4,0 kW. You cannot get below it by switching everything off, because the fifteen minutes are averaged, not sampled. That committed figure only ever rises inside a quarter, which is why it is the right thing to alarm on: at minute three it is still a warning, at minute twelve it is a receipt.

Tool choices — and why

Home Assistant (Apache-2.0, 87,6k stars, active) is the engine, and it is here for one specific reason: its built-in utility_meter helper accepts cycle: quarter-hourly, which is exactly the grid's billing window, plus a cron form if your operator uses something else. That is the primitive this whole blueprint stands on, and it is a checkbox in the UI rather than code you have to write. Everything else — the committed-kW sensor, the month-peak sensor, the alert — is a template and an automation on top of it.

ESPHome (dual-licensed: GPLv3 for the build tooling, MIT for the code that runs on the chip — free to use in your business either way; read the LICENSE file if you plan to redistribute a product) is the firmware for the reader. Two stock components cover almost every meter: dsmr for the P1 port found on Dutch, Belgian and Luxembourg digital meters, and pulse_counter for any meter with a blinking LED, where each pulse is a fixed fraction of a kWh. Pulse counting is the honest universal answer — it needs no port, no utility permission, and it measures energy per interval, which is precisely the quantity you are billed on.

home-assistant-glow (MIT, 1,3k stars, active) is the pre-baked pulse-LED build: an ESP32 and a photodiode taped over the meter's LED, roughly €15–25 in parts. It exists so you do not have to write the ESPHome config yourself. For a P1 port, a ready-made reader is about €30–60; check current prices, they move.

ntfy (Apache-2.0, 30,5k stars, active) delivers the push. Self-host it or use the free public server; a peak warning is not confidential, but self-hosting costs nothing extra on a box you already run. Email is the wrong channel here — you have about ten minutes of usable warning, and nobody reads email in a kitchen.

Cost. Software €0, all four. Hardware one-off: €15–60 for the reader, €120–200 for a small always-on box if you do not already have one, or a Raspberry Pi. If your meter has no P1 port and no visible pulse LED — common on US commercial services and on CT-metered supplies — you need a DIN-rail meter with clamps (about €60–120 in parts) fitted by an electrician, which is real money and real scheduling; get a quote before you commit. Running cost is the electricity the box itself burns: roughly 10 W, which is about €2 a month, and €0 extra if the box is already there for something else.

What we did not pick. Energy-management SaaS with a dashboard and a monthly fee does the same arithmetic and holds your interval data. Grafana and InfluxDB would give prettier history, and they are the wrong first move: Home Assistant already keeps long-term statistics, and a second database is a second thing to back up before you have proven the peak is even shiftable.

Setup outline

1. Spend nothing first. Find the demand or capacity line on your last bill and write down the rate per kW. Then download your quarter-hour history from your grid operator's portal — in Flanders that is Mijn Fluvius, free, quarter-hour detail, exportable; in the US, ask for interval or Green Button data. Sort by the highest quarters and look at the timestamps. What you will see when it works: your worst quarters cluster at two or three recognisable moments of the day.
2. Decide whether anything is shiftable before buying hardware. If your peak is one machine that cannot wait ten minutes, stop here — you now know a fact, and this blueprint will not save you money.
3. Install Home Assistant on a small always-on box (official Docker image or the prebuilt OS image) at, or reachable from, the meter cupboard.
4. Fit the reader: P1 cable into the port, or the glow sensor over the pulse LED. In Flanders the P1 port is yours and you switch it on yourself via Fluvius. What you will see when it works: a live power reading in Home Assistant that changes when you switch on a kettle.
5. Create the quarter-hour meter: a utility_meter helper on your energy sensor with cycle: quarter-hourly. Add a template sensor committed_kw = that helper's kWh divided by 0,25. Add a second one for the month peak.
6. Set the target. Take your last twelve month peaks, pick a number a little under the lowest of them, and set the alert there. A target nobody can ever hit is the same as no target.
7. One automation: when committed_kw rises above the target, push a message that names the likely culprits — "already at 22 kW this quarter, the oven and the compressor are both running".
8. Add a no-data watchdog: if the meter reading has not updated in 30 minutes, alert. Then run it for one full month and compare your logged month peak against the operator's own figure before you trust the number.

Pitfalls — the real ones

Alarming on live power is decoration. A 30 kW spike that lasts twenty seconds barely moves a fifteen-minute average; a steady 18 kW for the whole quarter is what bills you. Alarm on committed kW, computed from energy so far, or you will get alerts that do not matter and miss the ones that do.
Your quarter has to be the grid's quarter. The windows are aligned to the clock, so if the box's time drifts you are measuring the wrong fifteen minutes and every number you produce is subtly wrong while looking perfectly reasonable. Run NTP, and reconcile one month against the operator's portal before acting on anything.
Belgian and Luxembourg P1 streams can be encrypted. You need the decryption key, and a meter swap gets you a new one. The symptom is not an error, it is silence.
The billed peak is an average, so the payoff is slow. In Flanders it is the mean of twelve month peaks, so a bad January is still in the bill in November. Many US tariffs do the same job with a ratchet at 80–90% for 6–12 months. Do not expect next month's bill to prove the project.
Do not let this switch anything off. This blueprint alarms; it does not control. Automatic load shedding on an oven, a compressor or a cold room needs an electrician, proper interlocks and someone who knows what a mid-cycle stop does to the product and to the machine. That is a different project with a different risk profile.
Alerts nobody acts on train people to ignore alerts. One target, one phone, one person, during opening hours. If the alert fires more than a couple of times a week, the target is wrong, not the staff.

Verified repos

Home Assistant Core — Apache-2.0, 87,6k stars, active (Aug 2026)
ESPHome — dual GPLv3 / MIT, 11,5k stars, pushed 07/08/2026
home-assistant-glow — MIT, 1,3k stars, pushed 03/08/2026
ntfy — Apache-2.0, 30,5k stars, active (2026)