Welcome to F97.BE!

Forecasting solar energy is harder than it looks. GTI forecasts, grounded in physics and solar geometry, are transparent and reliable under clear skies, but they struggle with clouds and local effects. AI forecasts learn from years of production data and adapt to site-specific conditions, yet they can misfire when patterns shift or fall outside the training range. Climate change adds another layer of uncertainty. More variability, haze events and changing atmospheric conditions can reduce output in ways that are not directly captured by temperature alone. AI models often interpret higher temperatures as higher production, even when irradiance is reduced. It takes time, and new data, before models adjust.

This is why I developed the Fusion Forecast. It combines AI with GTI and balances both dynamically. Instead of relying on a single model, it corrects systematic bias and weights the more reliable signal for the given situation. The goal is not perfect prediction, but stable behaviour under changing conditions. The model has recently been upgraded, and the Fusion layer is currently collecting new data. Results should improve over the coming weeks as the system adapts, with clearer gains expected by the end of April.

Alongside Fusion, you can explore my 24-hour energy production forecast for the House Roof System, generated with a custom AI model trained on more than 1700 days of historical weather and production data. The dashboard visualizes real-time AI predictions, GTI-based forecasts, and actual inverter output in a single chart.

Each day, two models - Random Forest and Gradient Boosting - are evaluated, with the system automatically selecting the best performer for the AI forecast. In parallel, I also calculate a GTI forecast, based purely on solar geometry and irradiance predictions. These two streams are then fused into a third column: the Fusion forecast, which learns an optimal weighting, applies an offset if needed, and falls back to a fixed blend when data is limited. This way the forecast stays stable while adapting to changing weather and seasonal patterns.

Download the paper: Fusion Forecasting for Residential PV: An Adaptive Regression Blend of AI and GTI with Online Calibration.

The AI model was last updated on 26.03.2026, the GTI model on 07.09.2025. The real-world dataset currently contains more than 1700 entries.

More in depth information:

• NEW: Historical Forecast Accuracy over Fusion, AI, and GTI. Updated version!

Model Updates