LoadML Solar Prediction differs from PV7 Solar Forecast

[metabubble]

Describe the bug
When I look at the graphs in PV7 and LoadML I see different predictions for the next day. It might be an issue with 15 minute price slots and open_meteo_forecast

Expected behaviour
LoadML and PV7 Forecast showing the same values.

Predbat version

v8.46.6

Environment details

• Inverter and battery setup
• Standard HAOS installer

Screenshots

Log file
predbat.log

predbat_plan-3.html
apps.yaml.txt

Predbat debug yaml file
predbat_debug.yaml.txt

[chalfontchubby]

Hi - to check my understanding - you expect the loadml curve to track the solar forecast? Isn't one the load, the other the solar?
Your predicted load does seem to be a shifted copy of your solar prediction - but that would imply that you perhaps have a significant load associated with aircon which would correlate with the solar input after a delay.

Either that - or I'm missing something?

[metabubble]

I probably didn't explain it well enough because I assumed certain things. Look at the second screenshots. All forecasts (red lines) peak at 8.5 kW, the ones in the past and the ones in the future

If you then look at the first screenshot, the yellow line in the past matches the peaks from the second screenshot, but the prediction in the future is off by half.

[chalfontchubby]

Ah I see it now - sorry, I'd missed your point entirely. You're not comparing load to solar, you're comparing the same forecast against itself and the future half doesn't match the past half. That's a much more specific and useful bug report than I gave it credit for.

I went and checked Predbat's own forecast pipeline for a resolution bug (you're on open_meteo_forecast, which returns hourly data, plus plan_interval_minutes: 15 - a very reasonable place for an hourly-vs-30/15-min conversion bug to hide). Ran a controlled test through the real calibration code at both 15 and 30-minute intervals and got identical totals either way, and separately checked your own debug.yaml - the internal calibrated forecast (pv_forecast_minute) shows consistent ~8-8.8kW peaks across today/tomorrow/day-after, no halving. So that particular pipeline looks clean, at least in that snapshot.

Which means the discrepancy is most likely specific to whatever's actually feeding "PV Power (Predicted)" in your custom chart. Could you share the card YAML (or just tell us which entity/attribute that trace is bound to)? That'll tell us whether it's reading Predbat's calibrated forecast, the raw Open-Meteo output, or something else entirely.

[metabubble]

These are not custom charts, this is taken from the built-in /charts?chart=PV7 and /charts?chart=LoadMLPower

[chalfontchubby]

Sorry - I'm playing catchup ion that area of the code. Checked on my setup and the graphs don't show that mismatch - digging in to the issue. Apologies again for being slow on the uptake.

[metabubble]

No need to apologize.

While you’re at it, there are two areas of concern. The chart is one, but I also wonder if the data fed to the ML predictor is also off or actually correct. LoadML uses predicted generation as input for the forecast if I am not mistaken.

[chalfontchubby]

Found it, and thanks for pointing at the standard LoadMLPower chart - that made this traceable. Fix is up: #4374.

Root cause: predict_pv_power (feeding the "PV Power (Predicted)" trace) converts simulated PV energy into an instantaneous kW reading using a scaling constant that should depend only on the simulation's fixed 5-minute step - not on plan_interval_minutes, the user-configurable plan slot width. A PR back in November accidentally tied the two together. Anyone on the 30-minute default (like Rik, which is why his own chart looked fine) has always gotten the right number by coincidence; anyone on a shorter interval gets it scaled down by plan_interval_minutes/30 - exactly half at 15 minutes, which matches your report precisely.

Nothing to do with Open-Meteo or Solcast - purely about plan_interval_minutes.

[chalfontchubby]

Claude to the rescue - I believe it's isolated to the plot and not the calculations - but "we" are checking.