Guckst Du hier. SFML arbeiten die Daten von ML graphisch auf.
ARM Home Assistant Architekturen werden aktuell nicht unterstützt - das wird aber kommen!
Der Grund ist ein rein technischer, ARM Architekturen haben eine “eigene” HA-Version die nicht in der Lage ist, alle aktuelle (originalen Home Assistant) Bibliotheken zu kompilieren.
Bitte nicht mit dem PI nativ verwechseln, PI kann es sehr wohl! - Es liegt NICHT an der Hardware sondern der Home Assistant Version! - die Frage ob HA für ARM Prozessoren vorgefertigte Versionen aktueller Laufzeitumgebungen anbieten wird, kann ich nicht beantworten. Schaut man sich aber den aktuellen Core an, ist meine persönliche Einschätzung eher “nein”.
Die Version für ARM-Prozessor Home Assistant wird (Stand jetzt) das Dashboard bekommen, aber nicht die umfangreiche Statistik Funktionen, diese benötigt spezielle Bibliotheken die leider nicht auf Home Assistant für ARM lauffähig sind.
Also bitte etwas Geduld für Raspberry, HA Green und co.. ich suche nach einer Lösung!
1 „Gefällt mir“
Heute hatte die Integration mit besseren Wetter gerechnet obwohl nur Nebel und Wolken waren.
Deshalb auch possible Obstruction?
Habe heute auch gleich mal umgestellt vom Pi4 mit SD auf Generic x86-64 weil die SD-Karte am Pi4 die Flocke machen wollte, konnte da gerade noch ein Backup machen so das ich fast kein Datenverlust (ca. 30min) hatte 
Mal schauen wie weit ich mit dem Ram komme, werde da eventuell noch aufrüsten.
1 „Gefällt mir“
Wenn du nicht Frigate und Konsorten mit 10 Kameras betreibst.
Ich habe keine ungewöhnlichen Integrationen installiert und so sieht mein Systemstand nach fast 3 Jahren aus. CPU mit 4 u. mehr Kernen ist gut für HA. Arbeitsspeicher habe ich viel zu groß gewählt. Eine SSD sollte es auch schon sein.
1 „Gefällt mir“
Da es mehrfach von den Testern den Wunsch gab " Bitte auch eine helle Version" .. gibt es diese nun auch in SFML-Stats es kann einfach per Schalter zwischen hell/ dunkel umgeschaltet werden..
1 „Gefällt mir“
Wer ist denn so ein masochist und nutzt freiwillig keinen dark Mode? 
5 „Gefällt mir“
Also das du Proxmox ausschliesst kann ich ja verstehen, aber Light Mode, da bin ich raus 
2 „Gefällt mir“
Seit der „0“ Prognose und dem Neustart hängt die Prognose „oben“ fest, und ich weiß nicht wieso. Habe nichts geändert. Hoffe mit v12 klappts dann wieder
Hier war auch besseres Wetter vorhergesagt, aber den ganzen Tag dann nur Nebel und Nieselregen. Trotzdem war die Prognose fast eine Punktlandung:
Balken - Ist / Linie - Prognose
Auch gestern fast eine Punktlandung, so kann es gerne weitergehen.
VERSION V12 incl Stats und Grid-Price Monitor is out…
Solar Forecast ML V12 “Sarpeidon” - 1st HA Full AI & ML Solar Forecast
Intelligent Solar Forecasting for Home Assistant with Physics-First ML Architecture
Imagine your smart home not just reacting, but predicting - using machine learning that adapts daily to your unique solar installation. That’s exactly what Solar Forecast ML does. With Version 12.0.0 “Sarpeidon” we introduce enhanced Multi-Weather Blending, Per-Panel-Group Learning, and Hourly Correction Factors for unprecedented accuracy.
What Makes This Integration Special?
Unlike simple weather API services, Solar Forecast ML learns from your real data and uses actual solar physics. This means:
- Day 1: Accurate forecasts using real physics (POA irradiance, temperature correction)
- Week 1: System learns your panel geometry (tilt/azimuth) automatically
- Week 2+: ML enhances physics predictions with local patterns (shadows, obstructions)
- Month 1+: 93-97% accuracy with full Physics+ML ensemble
The integration adapts to your specific roof orientation, local weather conditions, shading patterns, and seasonal characteristics - all without manual configuration.
New in Version 12.0.0 “Sarpeidon”
Multi-Weather Source Blending
- Open-Meteo + wttr.in - Combines multiple weather sources for better cloud predictions
- Adaptive Weight Learning - System learns which weather source is more accurate for your location
- Trigger-based Fetching - Only queries secondary source when cloud cover > 50%
Per-Panel-Group Learning (NEW)
- Individual Energy Sensors - Configure separate kWh sensors per panel group
- Group-specific Efficiency - Each panel group learns its own hourly efficiency factors
- Shadow Detection per Group - Identifies which panels are affected by shadows
Hourly Correction Factors (NEW)
- Hour-specific Weather Corrections - Instead of daily averages, corrections are applied per hour
- Morning/Afternoon Optimization - Addresses systematic forecast biases at different times
- 7-day Rolling Learning - Continuously improves based on recent performance
Physics-First Architecture
The core architecture - a complete redesign from traditional solar forecasting:
- PhysicsEngine - Real solar physics calculations (POA, temperature correction, cell efficiency)
- GeometryLearner - Automatically learns your panel tilt & azimuth using Levenberg-Marquardt optimization
- AI Neural Network - Pure NumPy LSTM AI that captures temporal patterns
- Residual Learning - ML learns the difference between physics and reality
- Weighted Ensemble - Confidence-based blending of physics + ML predictions
- Captain’s LOG - Monthly system health report (tilt, azimuth, seasonal tracking)
Intelligent Detection Systems
- Correlation-based Shadow Detection - Distinguishes clouds from real obstructions
- Physics-based Frost Detection - Uses Magnus formula (dew point) for accurate frost warnings
- Cloud Layer Physics - Separate transmission models for low/mid/high clouds
Unique Dashboard and statistics " SFML-STATS" BETA
- First Home Assistant interactive Dashboard* - Self hosted web-dashboard for local access
- Beautiful UX - stunning dashboard with dark and white mode
- Advanced statistic - Download stunning png charts about your solar-system
Unique Grid Price Monitoring “Grid Price Monitor” BETA
- Germany & Austria only*
- Triggers automaticaly advanced Grid Power Automations - Automatic charging of you EV, Battery when price is low
- Advanced statistic - together with SFML-STATS
Core Features
Forecasting
- 3-day yield forecasts (today, tomorrow, day after) with Physics+ML optimization
- Hourly predictions for detailed planning
- Physics-based calculations using GHI, DNI, DHI from Open-Meteo
- Automatic panel geometry learning - no manual tilt/azimuth configuration needed
- Temperature-corrected efficiency - accounts for cell temperature effects
Machine Learning supported by AI
- Adaptive Algorithm Selection:
- < 100 samples: Ridge Regression (fast, stable)
- 100+ samples: TinyLSTM Neural Network (+5-10% accuracy)
- 14 optimized features per hour (time, weather, astronomy, lag)
- 24-hour sequence learning for temporal pattern recognition
- Zero external dependencies - pure NumPy implementation
Analytics & Detection (AI)
- Shadow Detection - Correlation-based analysis separates weather from obstructions
- Frost Detection - Physics-based dew point calculation with Magnus formula
- Peak hour detection - When do you produce the most?
- Production time calculation - Optimized sunrise to sunset tracking
- Self-sufficiency calculation - How independent are you really?
Data Privacy
- 100% local processing - Everything calculated on your system
- 100% local AI and ML - No ChatGPT, Claude, Gemini, Grok needed
- No cloud services - All data stays on your disk
- No API keys required - Uses free Open-Meteo API
Companion Integrations
Solar Forecast ML works best with these companion integrations that extend its capabilities:
SFM-Stats (Statistics Dashboard)
What it does: Provides comprehensive statistics, visualizations, and analytics for your solar production data. Creates beautiful Lovelace cards showing historical performance, accuracy metrics, and trend analysis.
Features:
- Historical production charts (daily, weekly, monthly)
- Forecast vs. actual comparison graphs
- ML model performance metrics
- Panel group efficiency visualization
- Weather correlation analysis
Installation via Service:
service: solar_forecast_ml.install_sfm_stats
data: {}
After running this service:
- Restart Home Assistant
- Add the new cards to your dashboard
- SFM-Stats entities will appear under
sensor.sfm_stats_*
Platform Compatibility for SFM-Stats:
Platform Status Notes x86_64 (Intel/AMD) Fully Supported Recommended Home Assistant OS (x86) Fully Supported Native installation Docker on x86 Fully Supported Standard HA container Proxmox VE Not Compatible NumPy/SciPy compilation issues in virtualized environment. CPU instructions not properly passed through. Raspberry Pi (ARM) Not Compatible Insufficient RAM and slower NumPy operations cause timeouts. SBC / ARM Processors Not Compatible Single-board computers (Odroid, Orange Pi, etc.) lack computational power for statistics processing. 32-bit Systems Not Compatible Requires 64-bit architecture. Note: These restrictions apply only to SFM-Stats, not to Solar Forecast ML itself. The main integration runs on all platforms including Raspberry Pi and ARM devices.
Grid-Price Monitor
What it does: Integrates with dynamic electricity pricing (aWATTar, Tibber, Nordpool) to optimize your energy usage based on solar forecast and grid prices.
Features:
- Real-time electricity price tracking
- Optimal charging/discharging recommendations
- Price forecast integration with solar forecast
- Automation triggers for low-price periods
- Cost savings calculations
Installation via Service:
service: solar_forecast_ml.install_grid_price_monitor
data:
provider: "awattar" # Options: awattar, tibber, nordpool
country: "DE" # Your country code
After running this service:
- Restart Home Assistant
- Configure your electricity provider credentials (if required)
- Grid-Price entities will appear under
sensor.grid_price_*
Note: Both companion integrations require Solar Forecast ML to be installed and configured first. They share data seamlessly without additional configuration.
Sensors & Diagnostics
Core Sensors
| Sensor | Description |
|---|---|
sensor.solar_forecast_ml_today |
Today’s forecast (kWh) |
sensor.solar_forecast_ml_tomorrow |
Tomorrow’s forecast (kWh) |
sensor.solar_forecast_ml_day_after_tomorrow |
Day after tomorrow (kWh) |
sensor.solar_forecast_ml_next_hour |
Next hour prediction (kWh) |
sensor.solar_forecast_ml_production_time |
Production hours today |
ML & Analytics Sensors
| Sensor | Description |
|---|---|
sensor.solar_forecast_ml_model_state |
ML model training status |
sensor.solar_forecast_ml_model_accuracy |
Current prediction accuracy (%) |
sensor.solar_forecast_ml_training_samples |
Number of training samples |
sensor.solar_forecast_ml_shadow_current |
Current shadow detection |
sensor.solar_forecast_ml_performance_loss |
Shadow-related losses (%) |
Weather & Trend Sensors
| Sensor | Description |
|---|---|
sensor.solar_forecast_ml_cloudiness_trend_1h |
1-hour cloud trend |
sensor.solar_forecast_ml_cloudiness_trend_3h |
3-hour cloud trend |
sensor.solar_forecast_ml_weather_stability |
Weather volatility |
Services (Developer)
ML Services (Developers only or on advice)
| Service | Description |
|---|---|
force_retrain |
Retrain ML model with all available data |
reset_model |
Reset ML model to initial state |
Bootstrap Services (Developers only or on advice)
| Service | Description |
|---|---|
bootstrap_physics_from_history |
Train Physics+ML from HA history (up to 6 months) |
bootstrap_from_history |
Bootstrap pattern learning from history |
Multi-Weather Services (Developers only or on advice)
| Service | Description |
|---|---|
refresh_multi_weather |
Force refresh of all weather sources |
learn_weather_weights |
Trigger weight learning from today’s data |
Companion Integration Services
| Service | Description |
|---|---|
install_sfm_stats |
Install the SFM-Stats companion integration |
install_grid_price_monitor |
Install the Grid-Price Monitor companion integration |
Astronomy Services (Developers only or on advice)
| Service | Description |
|---|---|
build_astronomy_cache |
Build comprehensive sun position cache |
refresh_cache_today |
Refresh astronomy data for current week |
Testing & Maintenance (Developers only or on advice)
| Service | Description |
|---|---|
test_morning_routine |
Analyze 6 AM predictions (read-only) |
run_all_day_end_tasks |
Manual trigger for 23:30 workflow |
run_weather_correction |
Rebuild corrected weather forecast |
send_daily_briefing |
Send formatted solar forecast notification |
Installation
HACS (Recommended)
- Open HACS in Home Assistant
- Go to “Integrations”
- Click the three dots menu → “Custom repositories”
- Add:
https://github.com/Zara-Toorox/ha-solar-forecast-ml - Select category: “Integration”
- Install “Solar Forecast ML”
- Restart Home Assistant
- After Setup please wait 10-15 min and perform a 2nd Restart (needed to fill the caches)
Manual Installation
- Download the latest release
- Copy
custom_components/solar_forecast_mlto yourconfig/custom_components/ - Restart Home Assistant
- After Setup please wait 10-15 min and perform a 2nd Restart (needed to fill the caches)
Configuration
- Go to Settings → Devices & Services
- Click “Add Integration”
- Search for “Solar Forecast ML”
- Configure:
- Power Sensor (required): Current solar power in Watts
- Daily Yield Sensor (required): Daily yield in kWh (must reset at midnight)
- System Capacity (optional): Your system size in kWp
- Panel Groups (optional): Format
Power(Wp)/Azimuth(deg)/Tilt(deg)/[EnergySensor] - Additional Sensors (optional): Temperature, Lux, Radiation, etc.
Panel Group Configuration (NEW in V12)
For systems with multiple panel orientations, you can configure individual groups:
Format: Power(Wp)/Azimuth(deg)/Tilt(deg)/[EnergySensor]
Example:
1425/180/9/sensor.pv_south_kwh_today
870/180/47/sensor.pv_roof_kwh_today
This creates:
- Group 1: 1425Wp, facing South (180deg), 9deg tilt, with its own energy sensor
- Group 2: 870Wp, facing South (180deg), 47deg tilt, with its own energy sensor
The system will learn individual efficiency factors for each group, improving overall accuracy.
Quick Start After Installation
For immediate full accuracy, run the bootstrap service (Developers only or on advice):
service: solar_forecast_ml.bootstrap_physics_from_history
data:
days: 180 # Uses up to 6 months of history
This will:
- Fetch your production history from Home Assistant
- Get historical weather from Open-Meteo Archive
- Train the GeometryLearner (learns panel tilt/azimuth)
- Train the ResidualTrainer (ML corrections)
- Your system is fully calibrated!
How It Works
Architecture: “Physics-First, ML-Enhanced”
+-------------------------------------------------------------------+
| LAYER 1: DATA SOURCES |
| Open-Meteo + wttr.in (Blended) + Astronomy + Local Sensors |
+-------------------------------------------------------------------+
| LAYER 2: PHYSICS ENGINE |
| POA Calculation -> Temperature Correction -> Power Output |
+-------------------------------------------------------------------+
| LAYER 3: LEARNING ENGINE |
| GeometryLearner + ShadowLearner + WeatherPrecision + PanelGroups |
+-------------------------------------------------------------------+
| LAYER 4: PREDICTION ENSEMBLE |
| Physics (Rule-Based) + TinyLSTM (ML) -> Weighted Combination |
+-------------------------------------------------------------------+
| LAYER 5: ANOMALY DETECTION |
| Frost Detection + Shadow Detection + Cloud Physics |
+-------------------------------------------------------------------+
System Lifecycle
| Phase | Duration | What Happens |
|---|---|---|
| Fresh Install | Day 0 | Physics engine active with defaults, ~70% accuracy |
| Initial Learning | Day 1-7 | Geometry converges, Ridge ML available |
| Geometry Convergence | Day 7-14 | +/-3 deg tilt, +/-8 deg azimuth accuracy, ~85-90% |
| ML Activation | Day 14-30 | TinyLSTM enabled, ensemble: 70% physics + 30% ML |
| Production | Day 30+ | Full calibration, 93-97% accuracy |
With bootstrap_physics_from_history: Skip to “Production” phase immediately!
Requirements
- Home Assistant 2024.1.0 or newer
- Power sensor (Watts)
- Daily yield sensor (kWh, resets at midnight)
- ~50 MB disk space for data files
- ~100-150 MB RAM during ML training
Note: Solar Forecast ML runs on all platforms including Raspberry Pi and ARM devices. Only the optional SFM-Stats companion integration requires x86_64 architecture.
Optional but Recommended
- Lux sensor (improves shadow detection)
- Temperature sensor (improves efficiency calculation)
- Solar radiation sensor (W/m2, best for precision)
Troubleshooting
Check Logs
# Dedicated log file
/config/solar_forecast_ml/logs/solar_forecast_ml.log
Common Issues
Predictions too low?
- Check if you have used the correct parameters (kWp = the sum of all your installed panels)
- Check if
yield-sensoris set correctly (kWh of your panels! must reset at midnight) - Check if
power-sensoris set correctly (W of your panels!)
ML not training?
- Need 10+ samples for Ridge, 100+ for TinyLSTM
- Check
sensor.solar_forecast_ml_training_samples
Shadow detection wrong?
- Ensure lux sensor is configured
- System needs clear-sky days to learn patterns
SFM-Stats not working on Proxmox/ARM?
- The SFM-Stats companion integration is not compatible with Proxmox or ARM-based systems
- Solar Forecast ML itself works fine on these platforms
- For SFM-Stats, please use x86_64 hardware
Contributing
Contributions are welcome! Please read our contributing guidelines and submit pull requests to the dev branch.
License
This project is licensed under the GNU Affero General Public License v3.0 - see the LICENSE file for details.
Credits
- Developer: Zara-Toorox
- Architecture Design: Physics-First ML in Python , SFML-STATS in VUE, Grid Price Monitor in Python
- Weather Data: Open-Meteo + wttr.in (free, no API key required)
Support
- Issues: GitHub Issues
- Discussions: GitHub Discussions
Made with solar power in mind
1 „Gefällt mir“
ANLEITUNG ZU STATS
WICHTIG: Nach der Konfiguration oder Neukonfiguration bitte HA Neuladen!
Hinzufügen des Dashboard (wenn gewünscht)
-
Einstellungen
-
Neues Dashboard
3.Website
URL: http://DEINE_IP_ADRESSE:8123/api/sfml_stats/dashboard als Dahboard einbindet
- Name: SFML
5.Speichern
Bei mir wird der Punkt Panel Groups in der Konfig nicht angezeigt
Neukonfigurueren auswählen… unten direkt auf okay klicken dann kommt es.. anders lässt HA es nicht zu, da es ein 2. config_flow ist,. der lässt sich nicht in den 1. einbinden - Einschränkung von HA-
Lief es sauber durch via HACS oder hat HACS wieder was verschluckt?
1 „Gefällt mir“
Bitte bei Problemen zunächst das Handbuch lesen! Wie immer *txt durch *pdf ersetzen!
Handbuch und FAQ DE.txt (633,5 KB)
1 „Gefällt mir“
SML Update über HACS lief durch, keine Fehlermeldungen im HA log nach Neustart
Beim Anwählen Neukonfiguration kommt nach dem OK der Punkt Panel Groups funktiioniert soweit
1 „Gefällt mir“
Danke Chris! Viel Spaß beim Spielen
Guten Morgen zusammen!
Darf man fragen, wie auf einem RPI das GPM installiert wird? Wie wird der beschriebene Service gestartet? Dankeschön
Version12 ging über HACS ging es fehlerfrei zu installieren.
Perfekt!
Grüße, dieter
Guten Morgen Dieter und einen schönen 3 Advent.
Du kannst auf einem PI den Service zur Installation aufrufen. Einmal Neuladen und dann wie gewohnt die Integration hinzufügen. Nur SFML stats kannst du leider ( noch nicht) auf dem PI installieren