Solar Forecast ML (SFML) provides a fully local, privacy-focused AI solar forecasting solution for Home Assistant users. It addresses the limitations of cloud-based forecasts by building a digital twin of a user's specific solar setup, learning site-specific factors like roof geometry, shading, and microclimate to deliver highly accurate, hourly forecasts. This empowers users with precise energy production predictions directly on their own hardware, eliminating data leakage and cloud dependencies.

How It Works

SFML employs a custom-built "Hubble" AI Stack, featuring a native Transformer architecture with Attention mechanisms, designed to run efficiently on Home Assistant hardware without requiring external AI services or heavy frameworks like TensorFlow/PyTorch. It creates a digital twin by analyzing roof geometry, local shading patterns, microclimate, and inverter behavior. Version 32 introduces a "Source of Truth" architecture, where SFML manages its own transactional SQL database for critical production and energy calculations, ensuring data integrity and reliability independent of Home Assistant's recorder. This approach allows for dynamic seasonal shadow mapping and environmental awareness, adapting to local anomalies for superior forecast accuracy.

Quick Start & Requirements

Installation is recommended via HACS (Home Assistant Community Store) under custom repositories. Manual installation is also supported. Key requirements include Home Assistant 2026.3.0+, configured panel-group power sensors (W), system capacity (kWp), and detailed panel group configurations (Power/Azimuth/Tilt/PowerSensor). The integration requires approximately 50 MB disk space and 200 MB RAM during AI training. It runs on x86_64, ARM, and Raspberry Pi 4/5, though the SFML Stats companion module is limited to x86_64.

Highlighted Details

• Achieves up to 97% accuracy after a calibration period.
• Operates 100% locally, ensuring complete user privacy with no external AI or cloud dependencies.
• Incorporates dynamic seasonal shadow mapping and real-world environmental awareness, including snow, fog, pollution, and altitude effects.
• Features an optional integration with the Toorox ForeSight HA add-on, a 20.5M-parameter Transformer, for enhanced forecast blending.
• The "Source of Truth" architecture centralizes critical solar data calculations within SFML's own database.

Maintenance & Community

The project is developed by Zara-Toorox. Community support and discussions are primarily hosted on the German-speaking HA Forum "simon42 Community," with links to GitHub Issues and Discussions also available.

Licensing & Compatibility

SFML is released under a Proprietary Non-Commercial license, making it free for personal and educational use. The source code is protected with PyArmor to prevent unauthorized AI training and commercial application, while maintaining full functionality for end-users.

Limitations & Caveats

The source code is obfuscated using PyArmor, which, while having minimal runtime overhead, restricts its use to non-commercial purposes and prevents direct AI model training. The system requires a learning lifecycle of up to 30 days to reach peak accuracy, and historical Home Assistant data cannot be directly imported for training. The advanced SFML Stats companion module is exclusively available for x86_64 architectures.