Multi-Source Local Weather Intelligence for Home Assistant — 100% Local, 100% Private
Your location. Your data. Your AI. Weather Fusion AI combines multiple independent weather sources into an intelligent ensemble and continuously learns from your own sensors — delivering 7-day hourly forecasts tailored to your microclimate. No cloud dependencies, no subscriptions, no data leakage. Just pure local intelligence powered by the proven AI architecture from Solar Forecast ML.
I've been working on this project with a lot of passion and dedication in my free time since autumn 2025. If you find Weather Fusion AI useful and appreciate the effort that goes into developing it, please consider making a small donation. Your support helps me continue to provide updates, build new features, and improve the project. Finding motivation for developing besides a full time job is not always easy. So if you would like to sponsor my coffee while coding, I would be very happy ☕
While others display generic regional forecasts, Weather Fusion AI treats each weather source as an independent expert and blends them with cloud-type-specific learned weights. It extends the proven Hybrid AI architecture from Solar Forecast ML — the pioneer of local AI in Home Assistant — to general weather forecasting for the first time.
Your weather station no longer just shows current values — it delivers precise, location-specific forecasts that adapt to your microclimate: urban heat islands, valley cold pockets, slope effects, coastal influences. Everything runs locally on your hardware with no data leaving your network.
Most weather integrations simply display a single provider's forecast. They don't know about your local conditions or why their temperature is always 2°C off at your location. Weather Fusion AI is the evolution:
| Feature | Standard Weather Integrations | Weather Fusion AI |
|---|---|---|
| Sources | Single provider | 4 independent experts blended |
| Privacy | Data often sent to cloud | 100% Local & Private |
| Learning | No learning | Learns from YOUR sensors |
| Microclimate | Ignores local conditions | Adapts to your specific location |
| Cloud Types | One-size-fits-all | Cloud-type-specific expert weighting |
| Corrections | None | 5-layer AI correction pipeline |
The heart of Weather Fusion AI is a multi-expert blending system that treats each weather source as a specialized forecaster. Instead of trusting a single source, the system learns which expert performs best under which conditions — and blends them accordingly.
Four independent weather experts feed into an intelligent blender that produces a robust fused forecast. This raw forecast is then refined through the AI correction pipeline using your local sensor data.
| Expert | Source | Strength |
|---|---|---|
| GFS (NOAA) | Open-Meteo API | Global coverage, 7-day horizon, detailed cloud layers |
| ICON (DWD) | Bright Sky API | High resolution for Europe, excellent for low clouds |
| ECMWF | wttr.in API | Strong global model, good precipitation accuracy |
| Pirate Weather | Pirate Weather API | NOAA GFS/HRRR, good for North America |
Cloud-Type-Specific Weighting — The system classifies 6 cloud types (Clear, Cirrus, Fair, Mixed, Stratus, Overcast) and maintains separate expert weights for each. Your DWD source might be best for stratus clouds while Open-Meteo wins for clear skies.
5-Layer Correction Pipeline — After blending, forecasts pass through five correction stages: hourly bias, cloud-type corrections, combined hour+cloud adjustments, seasonal patterns, and detected pattern corrections.
Pattern Detection — The AI identifies 11 recurring patterns at your location: morning/afternoon bias, evening transitions, cloud buildup, temperature lag, precipitation false alarms, microclimate warm/cold zones, wind channeling, and more.
The AI stack extends the proven architecture from Solar Forecast ML with weather-specific modules:
| Component | Purpose | What It Does |
|---|---|---|
| Precision Tracker | Deviation Analysis | Tracks forecast deviations across 5 dimensions: hourly, cloud type, season, measurement, and source. |
| Weather Learner | Weight Optimization | Optimizes expert source weighting using EMA (Exponential Moving Average) based on real observations. |
| Pattern Detector | Pattern Recognition | Identifies 11 recurring weather patterns specific to your location and microclimate. |
| Forecast Corrector | Correction Pipeline | Applies a 5-stage correction pipeline combining all learned insights into refined predictions. |
| Actual Tracker | Sensor Recording | Records hourly observations from your local sensors for comparison with forecasts. |
| Dimension | Buckets | Purpose |
|---|---|---|
| Hourly | 24 | Time-of-day specific corrections |
| Cloud Type | 6 | Clear, Cirrus, Fair, Mixed, Stratus, Overcast |
| Season | 4 | Spring, Summer, Autumn, Winter |
| Measurement | 9 | Temperature, Humidity, Pressure, Wind, etc. |
| Source | variable | Source-specific accuracy tracking |
- 7-day hourly forecasts with intelligent multi-source blending.
- Daily forecast aggregation from hourly data.
- Cloud-type-specific expert weighting for optimal accuracy.
- Weather code mapping to Home Assistant conditions (day/night aware).
- Continuous learning from local sensor observations.
- EMA-based expert weight optimization per cloud type.
- 11-pattern detection engine for location-specific corrections.
- 5-layer correction pipeline: hourly, cloud type, combined, seasonal, pattern.
- Blends 4 independent sources (Open-Meteo, Bright Sky, Pirate Weather, wttr.in).
- 6-class cloud type classification from cloud layer analysis.
- Automatic failure cooldown preventing rapid retries on API errors.
- Configurable update intervals (10–120 min).
- Morning / Afternoon Bias
- Evening Transition / Night Stability
- Cloud Buildup Patterns
- Temperature Lag
- Precipitation False Alarms / Misses
- Microclimate Warm / Cold
- Wind Channeling
Phase 1 — Day 1: Intelligent blending active immediately. All 4 experts contribute with default weights — solid baseline from the very first hour.
Phase 2 — Week 1: Early learning begins. Basic local patterns recognized, initial weight adjustments per cloud type.
Phase 3 — Week 2+: Pattern detection active. Morning fog, afternoon thermals, microclimate effects identified and corrected.
Phase 4 — Month 1+: Full calibration. High accuracy with location-specific corrections across all dimensions.
| Phase | Timeline | What Happens |
|---|---|---|
| Fresh Install | Day 1 | Intelligent blending active, solid baseline |
| Early Learning | Week 1 | Basic local patterns, initial weight learning |
| Pattern Detection | Week 2 | Microclimate effects, recurring patterns |
| Full Calibration | Month 1+ | High accuracy, location-specific corrections |
| Entity | Description |
|---|---|
weather.weather_fusion_ai_[name] |
Main weather entity with fused hourly & daily forecast |
| Sensor | Description |
|---|---|
sensor.wf_[name]_status |
System status (Active / Learning / Error) |
sensor.wf_[name]_forecast_accuracy |
Current forecast accuracy (%) |
sensor.wf_[name]_learning_progress |
Learning phase (Day X/3 or Active) |
sensor.wf_[name]_active_sources |
Number of active weather sources |
sensor.wf_[name]_last_learning |
Timestamp of last learning update |
sensor.wf_[name]_cloud_type |
Current cloud classification |
- HACS > Integrations > Custom repositories
- Add
https://github.com/Zara-Toorox/weather-fusion-ai(Integration category) - Install Weather Fusion AI
- Restart Home Assistant
- Download the latest release.
- Copy
custom_components/weather_fusion_aito yourconfig/custom_components/. - Restart Home Assistant.
Add via Settings > Devices & Services. Key inputs:
Step 1 — Location:
| Field | Description |
|---|---|
| Name | Display name for the weather entity |
| Latitude | Your location latitude |
| Longitude | Your location longitude |
| Update Interval | How often to fetch weather (10–120 min, default: 30) |
Step 2 — Local Sensors (Optional but Recommended):
| Sensor | Purpose |
|---|---|
| Temperature | Outdoor temperature sensor |
| Humidity | Outdoor humidity sensor |
| Pressure | Barometric pressure sensor |
| Wind Speed | Wind speed sensor |
| Rain | Rain sensor (binary or mm) |
The more sensors you connect, the better the AI can learn your local conditions.
Step 3 — API Keys (Optional):
If one of the raw data sources requires an optional key (e.g., Pirate Weather for higher query rates).
- Home Assistant 2026.3.0+
- ~20 MB disk space
- Runs on x86_64, ARM, Raspberry Pi 4/5
- Optional but recommended: outdoor temperature, humidity, pressure sensors
- No forecast displayed? Restart Home Assistant after installation. Check that at least one weather source is reachable.
- Learning stalled? Ensure local sensors are connected and reporting data. The AI needs hourly observations to learn.
- Source offline? The system automatically retries after a 5-minute cooldown. Check your network and API key configuration.
- Logs: Check Home Assistant logs for
weather_fusion_aientries.
Weather Fusion AI was designed from day one with one non-negotiable principle: your data never leaves your home.
This isn't a marketing claim. It's an architectural fact:
No Large Language Models involved — There is no connection to ChatGPT, Claude, Gemini, Grok, or any other AI service. Every calculation, every prediction, every learning step happens entirely within your own Home Assistant instance. The "AI" in Weather Fusion AI is your AI — running on your hardware, trained on your data.
No telemetry, no analytics, no tracking — The integration contains no usage tracking, no error reporting endpoints, no analytics libraries, and no background callbacks of any kind.
No data shared with me or anyone else — Your sensor readings, your location, your learned model weights — none of it is ever transmitted anywhere. Not to me as the developer, not to third parties.
Free weather APIs only — The integration fetches raw weather forecasts from public APIs (Open-Meteo, Bright Sky, wttr.in). These requests contain only coordinates — no personal data, no identifiers, no usage metadata.
Fully offline-capable — Once installed, Weather Fusion AI operates entirely within your local network. No internet connection is required for the AI to learn and correct forecasts (only for fetching fresh weather data).
In short: What happens in your Home Assistant, stays in your Home Assistant.
Some files in this integration are obfuscated (encrypted) with an official PyArmor version.
Why is the code protected?
- Protection against AI Training — I want to prevent my source code from being used to train AI models like ChatGPT, Claude, Gemini, or other Large Language Models (LLMs) without permission.
- Intellectual Property Protection — The algorithms for weather blending, AI-learning, and pattern detection were developed with considerable effort and represent my intellectual property.
- Open Source with Limits — This integration is free for personal use, but the source code is proprietary and subject to a Non-Commercial License.
- Unfortunately necessary — Since code has been copied without my consent, incorporated into commercial applications, and attempts have been made to read and modify it using AI in the past, I unfortunately feel compelled to protect the source code.
- Transparency — If you have a legitimate interest, I'm happy to provide information about the code or disclose it. Just contact me via GitHub Issues or Discussions.
The obfuscation has no impact on functionality. The integration works identically to the non-obfuscated version. Runtime overhead is minimal.
Weather Fusion AI — Copyright (C) 2025 Zara-Toorox · Protected with PyArmor 9.2.3
Proprietary Non-Commercial — free for personal and educational use. See LICENSE.
Developer: Zara-Toorox
Thanks to Simon42 and the users & contributors of the German-speaking HA Forum "simon42" for their testing, feedback, and discussion.
Support-Forum: simon42 Community | Issues | Discussions
Developed with 🌦️, late-night passion, and a stiff glass of Grog during Germany's wintertime.