[Bug]: SOC predictor is not accurate.

[jacekkjw]

Home Assistant version

2026.2.1

BMW CarData HA integration version

4.8.0 and main.

Installation method

HACS

BMW vehicle model & year

Countryman Cooper SE all4 2022

What happened?

SOC integration prediction isn't accurate

Too clarify: I’m on the main now. PHEV and AC only.
I made 2 tests. First one I just charge from 17% to 90%. Picture below (I changed the order of the sensors so colors changed):

• Integration prediction sensor is too optimistic and was synced 4 (even 5) times down.
• Prediction is syncing now with battery state (perfect) but still is syncing with bmw API prediction. API predictor sucks and many times syncing with battery with big delay. I don’t think so we should sync with it.

I manually measured/calculate (from graph and using external energy meter) real charging efficiency. It shows 67%. It’s weird, definitely too low. It suggests that battery size is too low. I checked it seral times. I found few values: 7.6 (most common), 8.8 and 9.0. API shows 7.3 comparing to 7.6 is probably ok. At the end is was difficult for me to find one source of true.
For comparison I checked efficiency on my second car (audi) and it shows 99,4%. Looks good but it little too high. Probably battery sizes provided by manufacturers are not so precise.
We don’t want to find the correct size of the battery or charging efficiency. We want to provide SOC level the same like at the official app but refreshing more frequently.
It can be done even is there a problem with correct battery size. If we remove limitation for efficiency charging (between 82% and 98%) lerning algorithm should find proper efficiency even battery size is wrong. Sometimes finally it will be very low sometimes above 100% but SOC level will be correct.
I made second test. I changed a bit the code: set the limits between 20% and 98% percent and default AC efficiency 67% (as I manually calculate for my car). Charging from 3% to 56%. Second graph below:

• Soc predictor is correct
• It was syncing up by API predictor and next down by real battery SOC level . Up and down is more or less the same value. That suggest API predictor is not so correct. ;)

Some suggestion:

• Integration shouldn’t have limitation for efficiency charging. Efficiency will correct problems with correct battery size (Up or down).
• There could be a mechanism of checking correctness of last charging session. But it should base on difference to current charging efficiency. It should work after some initial number (10?) of charging without any checks. It should allow learning algorithms find the initial value of efficiency factor .
• Could you guys implement some diagnostic sensor separate for AC I DC. One with current charging efficiency factor and second with efficiency for last charging session

Steps to reproduce

No response

Screenshots

No response

Debug logs

None

Additional context

No response

Checklist

• I'm using the latest version of the integration
• I've searched existing issues to avoid duplicates
• I've included debug logs and version information

[github-actions]

The debug logs section does not appear to contain valid Home Assistant debug logs.

Please enable debug logging and paste the actual log output:

logger:
  default: info
  logs:
    custom_components.bmw_cardata: debug

Then reproduce the issue and paste the logs in the Debug logs field.
Remove or redact personal data (VIN, location, tokens).

This issue will remain labeled needs-debug-logs until valid logs are provided.

[jacekkjw]

Logs attached.

home-assistant_cardata_2026-02-13T12-32-07.721Z.log

[jacekkjw]

SOC prediction filtered:

soc_log_table_res.txt

[renaudallard]

It's great I asked you to open a new issue because we now have logs and the OP is accurate ;)

[jacekkjw]

Yes. One note: the logs are from a modified version of the integration. I hardcoded the charging efficiency to 67% and it works perfect. This is not possible in the official version of the code. The learning algorithm will never reach such a low value because the limit for it is 82%. This is just proof of concept.

[renaudallard]

That's kind of you to specify this because it changes everything.

[renaudallard]

Would you be able to provide logs from main branch instead so we can have an insight on deviation?

[kvanbiesen]

Efficiency is a hard thing to calculate. Cause 3.2KW Charging on a phev is pretty effecient while on a EV, its drops easely into the 6O%, while charging 11KW on an EV, already bumps it up to the high 80s/low 90's. Cause of the used converters.

@renaudallard wouldnt dropping the constant for min effenciency to 0.50, fix it? since it can dip lower?

i'll remove the api predictor sync, i tried that before, better for learnging, but bigger spikes.

i had an idea for learning , i scribbled some but i havent really added the code to test.

[renaudallard]

Well, that was also my first thought. But that needs to be separated between BEV and PHEV or it might create unexpected issues because BEV will never deviate that much unless there is an issue. And that's why the limit was set in the first place.

[jacekkjw]

@renaudallard low efficiency could be caused by wrong battery size even on BEV.
@kvanbiesen what about debug sensor with current and last efficiency. These can help to understand that there is an issue at charging instalation.

[renaudallard]

A real efficiency of 67% means there a real issue with the converter in your car (below 85% means it's broken). Even the worst Chinese cheap charger with 60's tech won't have that bad efficiency. (OK, unless it burns).
Unless battery size is really different from what is expected.

[kvanbiesen]

@renaudallard check out my idea in the new branch, with the giant TYPO in the name but whatever, ive been dabbling with this all day, Its basicly a matrix. Where i hold most AC charging methods. It puts them in classes so it kinda holds a table with the efficiency at a certain voltage. Then it keeps track. Otherwise if you charge it at home (2.3KW) or at work 11KW, it will always ruin the predictions for next time, this way it doesnt.

i just pushed it - Mind you lint is prolly broken and havent tested it but most should be okay. I have to check it/test it but the base is there, take a look at it if you please.

@jacekkjw i do, it removed the syncing to predition. and the rest is above

[kvanbiesen]

A real efficiency of 67% means there a real issue with the converter in your car (below 85% means it's broken). Even the worst Chinese cheap charger with 60's tech won't have that bad efficiency. (OK, unless it burns). Unless battery size is really different from what is expected.

No it isnt, it super normal, the chargers/converters in have pretty low on those low volume and mind you, he still have the AuxPower not at low value, which really messes up the values. 2kw on your car BEV, isnt at much as 2kw on our slow charging cars (cause it still substracts that. That could boost his number up to the 80's ( i dont know his auxpower but if it is as mine 2.5 or higher, i bet his efficiency is pulled down by those values).(depending on how he changed it ofc