AROME solar forecast for Home Assistant

Version française

A free photovoltaic production forecast for Home Assistant, built on Météo-France's AROME model (2.5 km grid) through the Open-Meteo API. No account, no API key, no daily call quota, an update every 15 minutes. Two ways to use it:

• the integration, installed with HACS and set up in the user interface, which also feeds the Energy dashboard forecast;
• a plain YAML package, for those who prefer no custom component.

On a real installation in the south of France, AROME's forecast issued the day before missed the daily production by 17 % on average over 17 autumn days, against 21 % for the Solcast morning forecast and 20–31 % for other weather models (benchmark). The repository includes the tool to run the same test on your own data.

Features

• Quarter-hour forecast for today and tomorrow, in kW, from irradiance on the plane of your panels and air temperature.
• Physical PV model: cell temperature, temperature coefficient, system and inverter efficiency, inverter clipping, several roof planes (east/west...), optional hourly shading factors.
• Ready-to-use sensors: energy today, tomorrow, remaining today, current and next hour; power now, in 30 minutes and in one hour; peak power and peak time today and tomorrow.
• Energy dashboard forecast (integration): the expected production is drawn next to the actual one.
• Solcast-compatible attribute: detailedForecast with period_start and pv_estimate (kW, half hours) on the today and tomorrow sensors, so templates written for Solcast work by changing the entity id. The arome_solar_forecast.get_forecast action returns the same list.
• Robust: the integration keeps the last forecast for up to 6 hours if Open-Meteo is unreachable, and keeps the large forecast attributes out of the history database.
• Built-in forecast check (package): records the previous-day and morning forecasts, the actual production and the daily error, for AROME and for any other forecast you want to compare (Solcast, Forecast.Solar...).
• Benchmark tool: ranks 8+ weather models against your own production history, using archived forecasts.
• English and French user interface and entity names.
• Tested: the integration's model and every package template are checked against an independent implementation of the formula, in CI; hassfest and HACS validation on every commit.

Requirements

• A location covered by AROME: France and neighbouring areas, roughly 37.5–55.4° N and 12° W–16° E. Elsewhere, choose another weather model (Open-Meteo best match, ICON-D2, ICON-EU, Harmonie, ECMWF); check with the benchmark which one fits your site.
• Integration: Home Assistant 2025.2 or newer and HACS.
• Package: Home Assistant 2024.10 or newer, and Python 3.10+ on any computer to run the generator (optional: the package can also be edited by hand).

Installation with HACS (recommended)

In one click, from a browser where your Home Assistant is open:

Then, after downloading and restarting:

Or by hand:

1. In HACS, open the menu (⋮) → Custom repositories, add https://github.com/MariusTechLab/arome-solar-forecast with the type Integration.
2. Search for AROME Solar Forecast, download it and restart Home Assistant.
3. Settings → Devices & services → Add integration → AROME Solar Forecast. Enter the peak power, tilt and azimuth of your panels (compass: 180 = south) and the inverter limit. The location defaults to your home.
4. Optional, in the integration's Configure menu: second and third roof planes, efficiencies, temperature coefficient, hourly shading factors, weather model, update interval.
5. Energy dashboard: Settings → Dashboards → Energy → Solar panels → your production sensor → Solar production forecast → AROME Solar Forecast.

Several installations (houses, roofs) can be added as separate entries.

Installation as a YAML package

1. Generate the package

python tools/generate_package.py \
    --latitude 45.76 --longitude 4.84 \
    --array kwp=3.0,tilt=30,azimuth=180 \
    --inverter-kw 2.5 \
    --output arome_solar_forecast.yaml

• azimuth is the compass direction the panels face: 180 = south, 90 = east, 270 = west. See the conventions table: Open-Meteo and Solcast use other conventions.
• Repeat --array for each roof plane, for example --array kwp=2,tilt=20,azimuth=90 --array kwp=2,tilt=20,azimuth=270.
• --lang fr for French entity names. All options: python tools/generate_package.py --help and the configuration reference.

Two decimals of latitude and longitude (about 1 km) are enough: the AROME grid is 0.025° (about 2.5 km).

2. Add it to Home Assistant

1. Copy the file into your configuration folder, for example config/packages/arome_solar_forecast.yaml.

2. Enable packages in configuration.yaml if not done yet:

   homeassistant:
     packages: !include_dir_named packages

3. Check the configuration (Developer tools → YAML → Check configuration) and restart Home Assistant. A restart is required the first time, because it loads the rest integration.

4. Optional: exclude the two bulky technical sensors from the history database (examples/recorder.yaml).

The sensors are available within a minute of the restart.

Without Python

Copy packages/arome_solar_forecast.yaml and edit:

• in the resource: URL: latitude, longitude, tilt and azimuth (Open-Meteo convention: 0 = south, −90 = east, 90 = west);
• at the top of the detailedForecast15 template: kwp in arrays, ac_limit (kW, or none), the efficiencies and the 24 shading factors (index = local hour).

Sensors

Integration

Entity ids follow the entry name and the language of Home Assistant; with the default name and English: sensor.arome_solar_forecast_energy_today...

Sensor	Unit	Description
Energy today / tomorrow	kWh	attributes detailedForecast (30 min), detailedForecast15 (15 min), complete
Energy remaining today	kWh	from now to midnight
Energy current hour / next hour	kWh
Power now / in 30 minutes / in 1 hour	W	long-term statistics
Peak power today / tomorrow	W
Peak time today / tomorrow	timestamp
Last update	timestamp	diagnostic; attribute stale when the last call failed

Action arome_solar_forecast.get_forecast (days, resolution: 30min or 15min) returns {model, updated, forecast: [{period_start, pv_estimate}]}.

Automatic correction

In Configure, tick Automatic correction and pick your PV production sensor (energy in kWh, total or daily, or power in W or kW, with long-term statistics). Every night at 03:17, and at startup, the integration compares the archived AROME forecasts of the last 14 days with your actual production, hour by hour, and learns a correction factor for each time slot of the day: 30 minutes by default, or 15 minutes or 1 hour (option Correction resolution; below one hour it uses the 5-minute statistics, kept 10 days by default by the recorder). It captures shading (trees, buildings), soiling or a capacity lower than declared, as Solcast's automated dampening does, without using any API quota.

It is careful while data is scarce:

• the starting point is your manual shading factors (1 by default);
• no change for an hour until 3 days of data exist for it;
• then a gradual move towards the measured value: weight = days / (days + 7), about half way after 7 days and two thirds after 14;
• the median of the daily ratios is used, so an inverter outage or one unexpectedly sunny day does not pull the factor;
• days producing almost nothing while sun was forecast are treated as outages and skipped; hours forecast below 5 % of the peak power are ignored;
• factors stay between 0.3 and 1.2;
• hours are counted in solar time (UTC), so the factors stay right across daylight saving time changes; the sensor shows them with local hours.

The diagnostic sensor Correction factor shows the mean factor; its attributes give the factor, number of days, median ratio and weight for each hour. The factors are saved and survive a restart.

Best time to run an appliance

Action arome_solar_forecast.find_best_window finds the run of consecutive quarter hours with the most usable solar energy, today (from now on) or tomorrow:

action: arome_solar_forecast.find_best_window
data:
  duration: 75          # minutes, rounded up to a quarter hour
  day: tomorrow         # today or tomorrow
  power_kw: 2.0         # optional: solar above the appliance power is not counted
  base_load_kw: 0.35    # optional: house consumption subtracted first
  after: "10:00:00"     # optional bounds
  before: "17:00:00"
response_variable: window

It returns found, start, end, duration_minutes, usable_energy_kwh and solar_energy_kwh. The earliest window wins a tie. Example automation: plan the dishwasher in the evening for tomorrow's best window, then start it with a time trigger on window.start.

YAML package

Entity	Unit	Description
sensor.pv_forecast_arome_today	kWh	forecast energy today; attribute detailedForecast
sensor.pv_forecast_arome_tomorrow	kWh	forecast energy tomorrow; attribute detailedForecast
sensor.pv_forecast_arome_remaining_today	kWh	from now to midnight
sensor.pv_forecast_arome_power_now	W	forecast power for the current quarter hour
sensor.pv_forecast_arome_power_in_1_hour	W	forecast power one hour from now
sensor.pv_forecast_arome_peak_today	W	highest quarter hour; attribute peak_time
sensor.pv_forecast_arome_curve	-	attribute detailedForecast15: 192 quarter hours
sensor.pv_forecast_arome_raw_data	-	Open-Meteo response (technical)

Details: how it works.

Using the forecast

Energy expected for the rest of the day:

condition: numeric_state
entity_id: sensor.pv_forecast_arome_remaining_today
above: 3

Forecast power of each half hour, like Solcast:

{% for x in state_attr('sensor.pv_forecast_arome_today', 'detailedForecast') %}
  {{ x.period_start }}: {{ x.pv_estimate }} kW
{% endfor %}

Ready-made examples:

• examples/best_window.yaml: best start time today for an appliance of a given power and duration;
• examples/dashboard.yaml: entities card, forecast vs actual graph, daily errors, and an ApexCharts curve;
• examples/east_west_fr_with_check.yaml: a generated package with two roof planes, French names, shading factors and a daily comparison with Solcast.

A forecast is never perfect. In automations, keep a fallback for when the sensors are unavailable, and prefer real-time measurements (grid export) for short-term decisions.

Comparing forecasts on your installation

Live, from now on

Add the forecast check when generating the package:

python tools/generate_package.py ... \
    --actual-entity sensor.solar_energy_today \
    --compare Solcast=sensor.solcast_pv_forecast_forecast_today

--actual-entity is a sensor of the day's actual PV energy (kWh, reset at midnight; a utility_meter with a daily cycle on your production counter works). Every day the package records the forecasts at 00:05 and 06:30, then the actual production and the error of each source at 23:55.

On past data

python tools/benchmark.py --csv history.csv --entity sensor.pv_power \
    --latitude 45.76 --longitude 4.84 --array kwp=3,tilt=30,azimuth=180

The CSV is the Download data export of the Home Assistant History page for your PV power sensor. See the benchmark guide.

Configuration reference

Option	Default	Description
--latitude, --longitude	required	location
--array kwp=,tilt=,azimuth=	kwp=3,tilt=30,azimuth=180	one per roof plane; compass azimuth
--inverter-kw	none	inverter AC limit in kW (clipping)
--system-efficiency	0.98	soiling, wiring, mismatch
--inverter-efficiency	0.96	DC to AC
--temp-coefficient	−0.004	power change per °C above 25 °C
--cell-heating	0.03	°C of cell heating per W/m²
--shading	none	factors per local hour, e.g. 8=0.7,16=0.5,17=0.4
--model	meteofrance_arome_france	any Open-Meteo model
--forecast-days	2	2 for AROME; up to 7 with meteofrance_seamless
--scan-interval	900	seconds between API calls (minimum 600)
--lang	en	en or fr entity names
--title	per language	name prefix of the entities
--actual-entity	none	enables the daily forecast check
--compare LABEL=entity	none	other forecasts to check (repeatable)
--morning-time	06:30:00	time of the morning snapshot
--output	stdout	file to write

FAQ

The sensors stay unavailable. Check that Home Assistant was restarted (not only reloaded) after adding the first rest sensor, and look for rest errors in the log. Open the resource URL in a browser: it must return JSON with a minutely_15 block.

The forecast is constantly too high. Compare kWp and inverter limit with your installation, then run the benchmark: the fitted scale shows the systematic gap. A gap concentrated in the morning or late afternoon is shading: use --shading.

The curve is shifted in time. Most often the azimuth sign: see the conventions table.

Can it feed the Energy dashboard forecast? Yes, with the integration (see the installation steps). The YAML package cannot: that line only accepts integrations.

Integration or package? The integration for a setup in the user interface, the Energy dashboard and resilience to API outages. The package for an installation without custom component, or to adapt the templates.

Can I use it outside France? Yes, with another model (--model best_match, icon_d2, knmi_harmonie_arome_europe...). Run the benchmark to choose.

Is there a P10/P90? No: AROME is a deterministic model.

Data sources and terms

• Weather data by Open-Meteo.com, licensed under CC BY 4.0. The free API is for non-commercial use, up to 10,000 calls a day; this package makes 96 per roof plane.
• AROME and ARPEGE models by Météo-France, open data under the Etalab Open Licence 2.0.

This project is not affiliated with Open-Meteo, Météo-France or Solcast.

Related project

Solar surplus climate control: a blueprint that runs a reversible air conditioner on PV surplus, with optional pre-cooling based on a remaining-production forecast such as sensor.pv_forecast_arome_remaining_today.

Development

python -m pip install -r requirements-dev.txt
python -m unittest discover -s tests -v
python tools/validate_repo.py

After changing the generator, regenerate the reference package with the command in CHANGELOG.md; a test checks that it is up to date.

License

MIT