A solar forecast that learns from your panels.
A Home Assistant integration that predicts your PV production, corrects itself against what your installation really produces, and publishes the result where Home Assistant already knows how to show it.
Measured 2 055 W, predicted 2 062 W. Seven watts apart, on a real installation.
Learn more Data contract Changelog
Every version before this one could take an entire hour of long-term statistics down with it, and not only its own: when the recorder's hourly compile hit the collision this integration caused, it rolled the whole hour back, so every integration on that machine silently lost that hour of history. It was not rare. On one instance, two hours of a single day kept 14 series out of 494.
The cause was here since the first version and it is fixed at the root rather than worked around, so nothing you can configure avoids it on an older version. Your archived history is moved to the new series on the first start, hour by hour, and the old one is deleted only once every hour of it has been read back at its new home. The changelog says exactly what it does.
"Is it accurate?" deserves an answer with numbers behind it, so there is an open measurement running on real installations, published as it goes: helios-ha.org/benchmark.
Taking part is one checkbox, in Settings > Devices and services > Helios Forecast > Configure > Join the community benchmark, and it needs 2026.9.6 or newer. There is no key to ask for and no account to create. What travels is the panel geometry, the predicted curve, the measured production and the reliability index, once an hour: no entity names, no consumption, no other sensor, and coordinates rounded to about a kilometre. Clear the box and it stops.
The integration itself never needs any of this. It has no key, sends nothing of its own accord and computes everything on your machine, whether you take part or not.
Helios Forecast is in the HACS default store.
- In HACS, search for Helios Forecast and install it.
- Restart Home Assistant.
- Go to Settings > Devices and services > Add integration, and search for Helios Forecast.
- Fill in your first panel line: orientation, peak power, and your PV production sensor.
The production sensor must be a cumulative energy sensor in kWh, your inverter's or meter's running total, not an instantaneous power sensor in W. Only an energy sensor carries the long-term statistics the learning reads.
About the "custom integration" warning in your log
Home Assistant logs We found a custom integration helios_forecast which has not been tested by Home Assistant at startup. That notice appears for every
integration installed outside of core, which means every HACS integration. It is
informational and does not indicate a problem.
A native provider for the Energy dashboard. The prediction appears in Home Assistant's own Energy view, right next to your real production. No card required, no extra setup.
Real entities, built for the recorder. Power now, next hour, per-day energy and peaks over a seven day horizon, energy left today, and a forecast reliability score. All usable straight away in automations, history graphs and template sensors.
A sub-hourly detail series the Helios card reads for full-fidelity visuals, so the card does not have to compute a forecast of its own.
Every entity is listed on the website.
Helios Forecast computes the whole production curve, not just the current value: one point every 15 minutes, from midnight today through the seven-day horizon, each with the predicted watts and its P10/P90 confidence band. An Energy Management System can read it to shift battery charging into the expected peak, work out how much sun is still to come, or find the latest safe charging start.
As a service (recommended). helios_forecast.get_forecast returns the curve
as response data, on demand:
# How much PV energy is still forecast for the rest of today
script:
sun_left_today:
sequence:
- action: helios_forecast.get_forecast
response_variable: helios
- variables:
remaining_kwh: >
{{ (helios.forecast
| selectattr('datetime', 'match', now().strftime('%Y-%m-%d'))
| selectattr('datetime', 'ge', now().isoformat())
| map(attribute='watts') | map('float') | sum
* 0.25 / 1000) | round(1) }}
- action: persistent_notification.create
data:
title: Solar forecast
message: "About {{ remaining_kwh }} kWh of sun still to come today."The response is a list of buckets (p10 / p90 are null until the forecast has
learned enough of your history to publish a band):
forecast:
- datetime: "2026-09-12T14:00:00+02:00"
watts: 3120.4
p10: 2610.0
p90: 3450.8
- datetime: "2026-09-12T14:15:00+02:00"
watts: 3038.1
p10: 2550.2
p90: 3380.5
# ... one point every 15 minutesWith more than one installation, pass config_entry_id (from Settings >
Devices and services) to choose which one to read.
As an attribute. The same curve rides on the power_now sensor as its
forecast attribute, handy in a template sensor:
{{ state_attr('sensor.helios_forecast_power_now', 'forecast') }}If you have a battery, Helios Forecast can project its state of charge over the next 48 hours. It runs the production forecast against your home's own consumption, which it derives from your Home Assistant Energy dashboard (no extra sensor to wire), and integrates the battery's charge from your current SoC.
Turn it on in the integration settings by filling in your battery capacity and
your state-of-charge sensor; the reserve, efficiency and charge/discharge
limits are optional. A Predicted battery state of charge sensor then appears,
carrying the full curve as its forecast attribute (plus the projected low and
high over those 48 hours, when each is reached, the forecast reliability so you
know how far to trust it, and consumption_coverage, how much of the learn window
each Energy dashboard source had data for, since a source with gaps is the usual
reason a projection runs optimistic). Four companion entities, disabled by default, carry the
same low and high with their times (battery_min_soc, battery_min_soc_time,
battery_max_soc, battery_max_soc_time) for a tile or an automation.
It predicts, it never commands. Sending the charge order stays with your own automation, which knows your inverter; Helios just tells it what's coming:
# Top the battery up from the grid tonight only if it's predicted to run low
- action: helios_forecast.get_battery_soc_forecast
response_variable: soc
- variables:
lowest: "{{ soc.forecast | map(attribute='soc') | map('float') | min }}"
- condition: template
value_template: "{{ lowest < 20 }}"
- action: switch.turn_on
target:
entity_id: switch.force_battery_chargeThe response:
forecast:
- datetime: "2026-09-12T14:00:00+02:00"
soc: 78.4
- datetime: "2026-09-12T14:15:00+02:00"
soc: 80.1
# ... one point every 15 minutes, 48 hours aheadOne honest note: home consumption is a learned average, so the projection is a good steer, not a guarantee: read it alongside the reliability figure, and it gets better as it sees more of your history.
It starts from physics: Open-Meteo horizontal irradiance transposed onto each panel orientation, the direct and diffuse split, snow cover, combined with your installation geometry and a cell-temperature derate so hot days are not over-predicted.
Then it stops trusting the physics alone. It learns a correction from your home's own recorded production, matching past hours on cloud cover, sun geometry and outdoor temperature, and reading each of them as the ratio between what your installation made and what the physics said it would, so that today's geometry stays the physics' business and the history only says how your site departs from it. Over time the prediction absorbs what no generic model can know about your site: shading, soiling, an orientation that is a few degrees off.
What it does not learn is your hardware's limits. An hour where the inverter was held back, by a full battery, a zero-export rule or a grid limit, says nothing about the sky, so it is left out of the learning (out of the sky correction when it falls short of the model, out of the analog library altogether) and the limits stay where the forecast already applies them: forward, at forecast time. A full battery is recognised on its own from the state of charge sensor and the entry-level inverter cap you configured; for what the integration cannot see (zero export, a grid limit), point the optional curtailment signal at a binary sensor, input boolean or switch that is on while the inverter is being held back.
On a DC-coupled hybrid, learn from a production reading taken before the battery tap (inverter DC power plus battery DC power, integrated to kWh), not from the AC meter: the AC side never sees the energy that went straight into the battery, and learns nightly discharge as production.
It also publishes a reliability score, which reflects how much history backs the learning, how close yesterday's forecast came to what the meter then measured, and how predictable today's sky is. It is a floor rather than a grade: a signal it cannot measure counts for nothing rather than being shared out among the others, so the number never rises because something went missing. A forecast that tells you when to doubt it is worth more than one that does not.
Everything runs on your server. No account, no API key, no browser-side maths.
A forecast on a wrong configuration is wrong with a straight face, so every field is checked at startup and after every refresh: a peak power typed in watts, an inverter limit in watts, coordinates far from the home, a production sensor that is a power sensor or that counts energy at night, a battery whose charge power is ten times its capacity, a consumption source that reports intermittently. Each problem becomes a repair in Home Assistant (Settings, then Repairs, and the integration's own page), naming the value at fault and what to do about it, and it clears itself the moment the configuration is corrected. The integration's page also offers a diagnostics download for issue reports.
A panel line is a group of co-oriented panels. There are two ways to describe your installation, and the right one depends on your sensors.
One entry per line, recommended when each line has its own production sensor. Each entry gets its own device, entities and detail series, so every line can be wired to its own card. Add the integration once per line.
Several lines in one entry, for two strings on one inverter that reports a single combined value. Add the first line, then tick add another line for each extra orientation. The lines share one production sensor and one inverter limit, and the forecast sums them by their kWp share. Use Configure to edit the shared settings or the lines later. Each line can also override the shared inverter cap and GPS coordinates, for a micro-inverter string that saturates on its own or a line mounted somewhere other than the entry's home location.
Helios turns your Energy dashboard into a live 2.5D scene: the sun crossing its arc over your home, your production, grid and battery flowing in real time. It draws this integration's prediction as the dashed curve your measured production tracks against, so you see the forecast and the reality on the same picture. From card 2026.9.4 it also reads each panel line's orientation and position from this integration to mark your arrays in the scene.
Neither needs the other. Together they are the whole story.
Early days, and moving fast. The data contract the card is built against is documented and frozen, so a change on either side cannot silently break the other.
| CONTRACT.md | The data contract between the integration and the card |
| CHANGELOG.md | What changed, release by release |
| helios-ha.org | The full entity list and how it all fits together |
Requires Home Assistant 2025.11.0 or later: the archive writes its own long-term statistics series, whose metadata carries a unit class the recorder only stores from that release.
Found a bug, or is a forecast off? Open an issue. Feedback is very welcome, and it is what shapes the roadmap.
Helios Forecast is built and maintained by one person, in the open, and given away. A star costs nothing and helps other people find it. A coffee keeps the next cycle going.
Helios Forecast, a self-learning solar forecast integration for Home Assistant. Copyright (C) 2026 Jérôme CREMOUX (ReikanYsora).
Licensed under the GNU General Public License v3.0, see LICENSE.