Add ML energy consumption predictor (experimental)

[pookey]

Summary

Adds an optional ml_prediction consumption strategy using XGBoost to predict 24h energy consumption at 15-minute resolution. Depends on #33 (configurable consumption strategies).

What's included

• Complete ML module (ml/) with training, prediction, feature engineering, and standalone CLI
• ML Report dashboard page showing model metrics, feature importance, and forecast comparison chart
• Daily model retrain scheduler (23:00)
• /api/ml-report endpoint for the dashboard
• Docker base image switch from Alpine 3.19 to Debian Bookworm (required for xgboost/scikit-learn wheels)

Honest assessment

This is experimental. In my testing so far, the ML predictions are roughly on par with or slightly worse than the simpler influxdb_7d_avg strategy from #33. The model also:

• Adds significant startup time (retrain on boot ~3-5s, weather forecast fetch ~1s)
• Pulls in heavy dependencies: xgboost (~150MB), scikit-learn (~30MB), astral
• Requires switching from Alpine to Debian base image
• Needs weather entity and InfluxDB configured

That said, it's a fun playground for experimenting with consumption prediction, and the ML Report page gives good visibility into how well (or poorly) the model performs compared to baselines.

I'd completely understand if this isn't something that belongs in core. The influxdb_7d_avg strategy in #33 is more practical for production use.

Config example

home:
  consumption_strategy: "ml_prediction"
ml:
  location:
    latitude: 59.3293
    longitude: 18.0686
    timezone: "Europe/Stockholm"
  weather_entity: "weather.forecast_home"
  feature_sensors:
    outdoor_temperature: "sensor.outdoor_temperature"
  training:
    days_of_history: 30

Test plan

• Verify ML model trains on boot when ml config is present
• Verify ML Report page shows metrics, feature importance, and chart
• Test ml_prediction strategy produces 96-period forecast
• Verify fallback to fixed when ml config section is missing
• Test daily retrain scheduler fires at 23:00
• Verify system starts correctly when ml section is absent (no ML features loaded)

[johanzander]

Been thinking about this. I've been thinking about adding some ML for energy consumption prediction but never prioritized it because the 48h average sensor works good enough for me. I also have a heat pump and EV charger that makes prediction on a more granular level difficult I believe. But nevertheless would be great to get this working and for me to learn about ML.
That said it probably makes more sense to have it outside the BESS manager. Main reason is that I have another project (https://github.com/johanzander/Theria) that would benefit from this as well. They are both HA AddOns, and I have a HA template that these are built upon: https://github.com/johanzander/ha-add-on-template. Mostly useful if one wants a python backend and node frontend, so maybe not that relevant for the energy consumption predictor.

But what do you think, would you be open to create a separate HA AddOn for this? then that could serve as another configurable consumption strategy option...

[pookey]

. I've been thinking about adding some ML for energy consumption prediction but never prioritized it because the 48h average sensor works good enough for me.

So far, it's been trash and after 8 days of training data, the 'weekly average' is working better for me (the other PR - that's probably good for merge BTW).

I think the ML stuff currently falls into the 'cool experiment, but probably not useful' category!

I hadn't seen Theria) - and I was considering doing something really similar myself! Maybe I need to hop over onto that project too and have a play! Does it support a single temperature zone?

[johanzander]

The Theria project is in a really early experiemental stage and I ran into real physical constraint wrt. my heating system which made progressing any optimization kind of useless. And there is no or little intelligence yet, it only gathers / visualizes temperature for a set of configurable zones. so im not sure you will find anything useful there, but Im sure with Claude Code you could build something useful on top of it...

[johanzander]

the other PR - that's probably good for merge BTW - here you mean #33?