ML does not detect a weekly regular high load

[ppmt]

This is an issue that I had for some time and talked about it a few month ago with @gcoan but life got in the way and I had to sidetrack that issue but now I would like to may be get to the bottom of it.

May be someone has a similar situation and can chime in :)

I have been using ML for some time (more or less from the beginning) and it worked for the most part except for one thing. It doesn't anticipate a regular high load created by my baking oven.

Here is my house situation:

• I have 2 EVs (one getting charged every other day, one about once a week at the week-end).
• I have a solar system and battery (4kW solar, 9.6Kwh battery).
• I also have 1 bread baking oven which is in use 2 nights a week every Thursday and Friday. Consumption is roughly around 14kW per night.

Not counting the oven I would say that the ML is doing a pretty good job at forecasting the need of the house but in my experience it has never been very good at anticipating the nights that the oven is in used.

As I said it is turned on every Thursdays and Fridays so I would expect that the ML would have caught on that.
I normally turn on the oven around 5pm and turn it of around 22h30.

I have read the documentation several times and believe that my config is correct, but I reached a point where I would appreciate some feedback.

Note: Rereading it today it says that I should turn on the temperature component for better accuracy. I am not sure it would help since I don't have electric heating or a heat pump (yet) but still I activated it this morning.

In my config:

• load_ml_enable and load_ml_source are both set to true.
• load_today is set correctly as well as pv_today
• I have also removed the cars so that they are not factored in for the calculation.

For the ML components the load_ml_max_days_history is set to 28 and load_ml_database_days to 90
In HA the database retention is set to 30

Is there anyone who is using ML and also has an "odd" config like mine that seems to confuse the forecasting?

I insert the ML graph which shows the last week. It shows that the forecast(+8h) got it wrong on the 21st and again is going to be wrong for the 28th.
What happens is that the forecast(+1h) eventually catches up with the real consumption.

[gcoan]

Hi @ppmt see what response to get to the Q&A, and if you don't get any good insights, then please do log it as a bug. Include the logfiles, debug file and your load energy sensor history showing the pattern that Load ML should have learnt from.

My guess is that it's putting too much weight on the daily load cycle and less on the more periodic load, but this is a guess.

There's an AI auto-triage running now which is doing a good job of looking at the code and finding obscure issues like this

[ppmt]

Ok I will do that this week-end depending on the responses...let's not forget that it could just be a me problem :)

When you say the load energy sensor history, do you mean a graph of the oven energy sensor?

[gcoan]

When you say the load energy sensor history, do you mean a graph of the oven energy sensor?

I mean the sensor predbat is using for load energy today in apps.yaml. Predbat wouldn't have any knowledge of the oven energy, should be working off total house load energy.

And I don't think temperate input will bring any more accuracy.

[ppmt]

Not adding anything but another ML graph taken while the oven has been on for about 4 hours

[gcoan]

so the good thing is that load ML is recognising that the load is increasing more than expected, and is adjusting the forecast (yellow and red lines) to account for that. And from the look of following days its not polluting future forecasts too much

the gap is that load ML is not predicting in advance (the purple +8 hour forecast) that your load will rise for your ovens

[ppmt]

Yes exactly. And because of that it is not charging my batteries to prepare for it, since according to its forecast the battery should last me long enough to wait for the next day solar power

And obviously by the time it detects it, it is too late to do anything about it!

[gcoan]

since you know when your ovens are going to be on, you can always add some manual force charges to the plan, or add a load override so your batteries are charged sufficiently

but predbat should be able to do that automatically for you

[ppmt]

@gcoan, I am about to write the bug report and was wondering which predbat.log file I should include.

There is a button to dowload the predbat.log , is that the one I should include or should it be one that covers the time before the oven were on and the forecast is wrong?

[gcoan]

The button will only download the latest log

Really you need the log before the ovens were on when load ml wasn't predicting the weekly load. Use samba share or file editor to find the right logfile in the Predbat directory. It'll be Predbat.N.log

[browellm]

My untested feeling is similar to @ppmt: that LoadML overweights daily periodicity over longer range periods such as weekly. I don't have the clear example he has to illustrate but my typical weekend usage is higher than weekday and I think LoadML leans on the previous 24 hours for its forecast and underweights weekly patterns. Will be interested to see what tuning options come out of the bug report.