Zara-Toorox/Solar-Forecast-ML

What it solves

Solar Forecast ML addresses the limitations of generic solar forecasting tools, which typically rely on cloud services or static formulas that fail to account for site-specific factors like local shading, specific roof geometry, or inverter limitations. It provides a highly accurate, personalized forecast by learning the unique production patterns of your specific solar installation.

How it works

The project runs a complete AI stack directly within your Home Assistant instance. It uses a transformer-based model with attention mechanisms that trains on your local sensor data and weather inputs. By building a digital twin of your system, it continuously refines its predictions based on your actual historical performance, accounting for variables like seasonal shading, fog, and hardware constraints. All training and inference occur locally on your hardware, ensuring complete data privacy.

Who it’s for

This project is for Home Assistant users who have a solar PV system with DC power sensors and want to move beyond basic cloud-based estimates. It is best suited for users who value data privacy, want to optimize energy management (such as battery charging), and are comfortable configuring sensors and managing their own local infrastructure.

Highlights

• Local Transformer Model: Runs a transformer architecture locally on your home server without relying on external cloud APIs or large-scale frameworks like PyTorch.
• Digital Twin Approach: Learns site-specific nuances, including shading patterns, panel orientation, and inverter clipping, to provide a tailored 72-hour forecast.
• Privacy-First Design: No data, sensor readings, or learned model weights ever leave your local Home Assistant instance.
• Explainable AI: Provides transparency into forecast quality and includes an optional 'Hubble' energy copilot that explains energy recommendations in plain language.