Day Ahead works but MPC fails with Cannot add/subtract NaN/inf values

[Tecnologic]

Hi All,

my day ahead optimisation is working but mpc is not. With the same input parameters. Its complaining about
Cannot add/subtract NaN/inf values but i see no point where my input values could lead to div by 0 or something like that.

can someone give me a point where to look at?

Maybe important: EPEX Spot changed the unit from Eur/MWh to Eur/kWh. After that change the MPC started complaining.

mpc.log

[davidusb-geek]

Please share your command template used to launch an MPC optimziation call

[Tecnologic]

Here we are again. It worked till 6pm yesterday but then had the same error as before:

5b918bf2_emhass_2025-01-29T07-25-39.543Z.log

Is there any other data emhass takes from HA or is it just the data i pass with the Rest command?

[davidusb-geek]

Here we are again. It worked till 6pm yesterday but then had the same error as before:

5b918bf2_emhass_2025-01-29T07-25-39.543Z.log

Is there any other data emhass takes from HA or is it just the data i pass with the Rest command?

It depends on your setup.
In your case you are using the naive method load power forecast. So it will fetch your load sensor data from HA and then build a very basic "naive" model on that data. So check the data on your sensor to see if there is any issue with that.
But according to the logged messages it seems that your problem is elsewhere. You have an issue with finding the solver.
There was a problem with this on v0.12.3, update to v0.12.4 to see if its solved for you. Otherwise leave the default values for the solver.
In my case the following settings for the solver works just fine on my amd64 based server:

  "lp_solver": "default",
  "lp_solver_path": "empty",

It can get complicated about this if your are on any arm-based arch

[Tecnologic]

Thanks for the answer, I'm on v0.12.4 and x86_64 hardware and i changed from naive to mlforecaster.

I started with fit and tune running every night on 14 days of data. And the predict every 5min. But is the predict call necessary when the mpc is set to use the ml forecaster?

And may be a side question is there a entity available that tells me exactly when emhass wants to charge the battery from grid and when its just charging from solar? I find i hard to identify the points in time when charging from grid is requested and when not just via p_batt, p_grid and p_pv forecasts.

Thanks, i appreciate Ur work.

[davidusb-geek]

I started with fit and tune running every night on 14 days of data. And the predict every 5min. But is the predict call necessary when the mpc is set to use the ml forecaster?

No, the predict is not necessary, it is called inside the MPC call.

And may be a side question is there a entity available that tells me exactly when emhass wants to charge the battery from grid and when its just charging from solar? I find i hard to identify the points in time when charging from grid is requested and when not just via p_batt, p_grid and p_pv forecasts.

No, currently there is no entity for this. We had recently a discussion on this, we will probably add a graph for this in the future. But in the meantime that could possibly be solved with some HA automations.

See this issue #422 with that exact same request and where some solutions are proposed using automations

[Tecnologic]

Ok, Thanks for clarification. Then i remove the predict calls.

[Tecnologic]

I have a new issue with the mpc please see the log below
logs-1.txt

Looks like the mpc uses mlforecaster and generates 48 lags so 24h forecast but then uses 12??? Where are the 12 lags_opt coming from?