How long till predictions are accurate?

[SPACKlick]

First time using Home Assistant and my Predbat is set to "Control Charge" but "Read Only" so I can keep track of the forecasting before I put it in control of everything. But its under-forecasting my solar by a little bit and over-forecasting my house load by a lot.

Yesterday the prediction was 20.58 kWh solar 20.66kWh load. But actual was 24.1kWh solar 14.10kWh load. I do have a heatpump and predheat set up. I am also trying to train load_ml but that is also predicting load of around 21kWh per day. Those are comparable with previous days errors.

Any advice would be appreciated. Configuration below.

---
pred_bat:
  module: predbat
  class: PredBat

  prefix: predbat
  timezone: Europe/London
  currency_symbols:
    - '£'
    - 'p'
  threads: auto
  load_filter_threshold: 30

  fox_key: !secret fox_key
  fox_automatic: True

  # Model inverter idle consumption when the battery rate is zero (W).
  inverter_battery_rate_min:
    - 200

  battery_charge_power_curve:
    100: 0.15
    99: 0.15
    98: 0.22
    97: 0.31
    96: 0.42
    95: 0.48
    94: 0.58
    93: 0.68
    92: 0.77
    91: 0.85
    90: 0.94

  battery_discharge_power_curve:
    4: 1.0

  inverter_clock_skew_start: 0
  inverter_clock_skew_end: 0
  inverter_clock_skew_discharge_start: 0
  inverter_clock_skew_discharge_end: 0
  clock_skew: 0

  open_meteo_forecast:
    - postcode: "[censored]"
      kwp: 7.0
      declination: 30
      azimuth: -85
    - postcode: "[censored]"
      kwp: 7.0
      declination: 30
      azimuth: 95
  open_meteo_forecast_max_age: 4.0

  num_cars: 0

  rates_import_override:
    - date: '2026-10-31'
      start: '06:00:00'
      end: '07:00:00'
      rate: 0

  rates_import:
    - start: "00:00:00"
      end: "04:00:00"
      rate: 25.46
    - start: "04:00:00"
      end: "07:00:00"
      rate: 15.46
    - start: "07:00:00"
      end: "13:00:00"
      rate: 25.46
    - start: "13:00:00"
      end: "16:00:00"
      rate: 15.46
    - start: "16:00:00"
      end: "00:00:00"
      rate: 25.46

  rates_export:
    - rate: 15.0

  days_previous:
    - 7
  forecast_hours: 48

  load_ml_enable: true
  load_ml_source: false

  battery_scaling:
    - 1.0
  import_export_scaling: 1.0

  export_triggers:
    - name: 'large'
      minutes: 60
      energy: 1.0
    - name: 'small'
      minutes: 15
      energy: 0.25

  car_charging_energy:
    - sensor.wholehome_home_device_0_arotherm_plus_heat_pump_heating_electricity_today

  load_forecast:
    - predheat.heat_energy$external

  predheat:
    forecast_days: 2
    run_every: 5
    days_previous:
      - 7
    mode: pump

    external_temperature: sensor.[censored]_home_outdoor_temperature
    internal_temperature:
      - sensor.home_zone_1_circuit_0_current_temperature
    target_temperature: sensor.home_zone_1_circuit_0_desired_heating_temperature
    weather: weather.forecast_home

    hysteresis: 0.5
    hysteresis_off: 0.1
    smart_thermostat: false

    heat_loss_watts: 136
    heat_loss_degrees: 0.025
    heat_gain_static: 200

    heat_output: 15500
    heat_volume: 65

    heat_max_power: 5000
    heat_min_power: 800
    heat_cop: 4.3

    flow_temp: sensor.home_circuit_0_flow_temperature_setpoint
    flow_difference_target: 8

    weather_compensation:
      -20: 41
      -10: 37
      0: 33
      10: 28
      20: 22

[gcoan]

Yesterday the prediction was 20.58 kWh solar 20.66kWh load. But actual was 24.1kWh solar 14.10kWh load. I do have a heatpump and predheat set up. I am also trying to train load_ml but that is also predicting load of around 21kWh per day. Those are comparable with previous days errors.

You are always going to get some variability in the solar and house load, they are after all just a forecast. I would say your solar is within tolerance. You will probably get a better forecast with Solcast integration, but what you have is reasonable

Here's a graph this month of solar forecast vs actuals for me, and as you can see sometimes above, sometimes below

Your house load, is that your entire apps.yaml? There is no load_today, import_today, pv_today or export_today sensors. If correct Predbat is only receiving your predheat forecast.

You will need a load_today sensor for your non heat pump energy. Personally I don't use predheat, prefer instead to let load_ML do it all, as this has temperature variability in its model.

You don't have days_previous_auto set to false so predbat is using your days_previous window of 7 to average house load over the past 7 days. Have a look at the load sensor (assuming you have one and its not in the apps.yaml extract you gave), how consistent it is day to day?

[SPACKlick]

Ah, I thought from documentation that PredBat would be getting load_today automatically from my FoxESS integration. Do I need to define those 4 entities in the apps.yaml to point at things from my FoxESS set up?

[gcoan]

ah yes, I see you have fox_automatic set to true, then yes, it should get these from inverter automatically. You can see what sensor has been configured by looking at the apps.yaml config in the Predbat web console and then sense check it for consistency/data span.

And I see you have excluded your heat pump energy as car energy so it doesn't double count with predheat. This will be excluding your hot water heating energy though which predheat doesn't calculate.

you might want to check input_number.predbat_car_charging_threshold in case Predbat is excluding high house load thinking it is car charging - see https://springfall2008.github.io/batpred/car-charging/#filtering-car-charging-energy-from-house-load

You should be able to see in the log file exactly how much load is being read from the sensors, how much is excluded and sense check these