Battery health

The integration learns your EV battery’s usable capacity from the same recorded charging sessions used for charging history — no REST quota, all derived from the stream.

How it’s learned#

A wide-range charge (say 20 → 80 %) that adds a known number of kWh implies the capacity of the whole pack: if 30 % of the battery took X kWh, the pack is roughly X / 0.30. Averaging across many such charges cancels the noise from temperature, charging speed and measurement error.

The Battery Health sensor#

Surfaces as one Battery Health sensor per EV:

• State — usable capacity in kWh.
• Attributes — vs-new percentage, sample count, and a capacity-vs-mileage trend.

What counts as a sample#

Not every charge can say something about capacity. A session qualifies only if all of these hold:

• It spans at least 25 % of SoC. BMW streams SoC as a whole percent, so a narrow charge divides a small energy figure by a small, coarsely-rounded SoC delta. Measured on a real car against BMW’s own maxEnergy of 78 kWh: a 21 % charge implied 77 kWh (0.8 % out), a 12 % charge 82 kWh (5 %), and a 9 % charge 123 kWh (58 %). The error is modest above ~20 % and explodes below ~15 %.
• It carries battery-side energy. Sessions imported from BMW’s charging history hold only a grid figure, which includes charging losses and would overstate the pack. They are excluded however wide their SoC span — so the first charges after an install, which are usually imported, don’t count.
• It caught the whole charge. If the car was already charging when the integration noticed — a restart mid-charge, or a status transition the stream never sent — both the energy and the SoC span start late by different amounts, and dividing one by the other is meaningless. Such a session is flagged late_start and skipped. Its energy and cost still count everywhere else, as a floor.
• It ran end to end without an interruption. A charge that Home Assistant restarted in the middle of keeps its full SoC span but has a hole in its energy, filled from an estimate that already assumes a pack size — so measuring capacity with it would measure the assumption. Such a session is flagged interrupted and skipped too.

Learning mode#

It reads Learning (n/10) until it has enough good samples and the estimate agrees with BMW’s own capacity figure. A suspicious number is withheld rather than shown — a health figure that jumped around would be worse than an honest “not sure yet.”

To reach a confident number faster, do a few wide-range charges (a large SoC swing in one session) rather than many small top-ups.

If it sits at Learning (0/10), your charging pattern is probably the reason: topping up little and often from a half-full pack never produces a qualifying charge. Running the battery lower and charging it back in one go — even occasionally — is what moves the counter. Meanwhile BMW’s own State of health (SOCE) sensor gives you a figure directly, if your car streams it. Note also that the vs-new percentage needs BMW’s batterySizeMax, which some cars report as 0 (invalid); where that happens the percentage stays blank however many samples accumulate.

Viewing it#

Use the view: health card: a gauge (percentage of the as-new pack) with the capacity-vs-mileage trend below.