After 256 days of comparison, it is still too early for definitive conclusions, but clear patterns are already visible. By combining a physical GTI model with a data-driven AI approach, the Fusion forecast proves to be more stable and reliable over time than either method on its own.
Why my Fusion forecast is more reliable than AI or GTI alone
For some time now, I have been comparing three different approaches to forecasting my daily PV production: a physical model based on Global Tilted Irradiance (GTI), a purely data-driven AI approach, and a hybrid method I deliberately call Fusion. The idea behind it is simple. If both models have different strengths and weaknesses, why not combine them?
256 days is still too short a time span to draw definitive conclusions. But it is long enough for patterns to emerge, and those patterns are already quite clear.
Three models, three ways of thinking
GTI is explainable, stable, and physically sound. It describes what is theoretically possible under idealised assumptions. What it does not capture well is local reality: cloud dynamics, microclimates, partial shading, inverter clipping, or those days that simply behave “differently”.
The AI model learns exactly that local reality. It recognises patterns, typical daily profiles, and recurring deviations. In return, it is more sensitive to outliers and situations that occur rarely or not at all in the training data.
The Fusion forecast combines both. It does not try to be smarter than its components, but more robust. It dampens the extremes of the AI model and compensates for the GTI model’s blindness to local effects.
What the results show
Over the 256 analysed days, the difference shows up clearly in the numbers. The AI model was most often the closest on individual days, with 103 days, compared to 97 days for Fusion and 56 for GTI. But when looking at overall accuracy, Fusion performs best. Its average absolute error is 4.87 kWh, compared to 5.30 kWh for the AI model and 9.79 kWh for GTI. In other words, AI wins slightly more often on good days, but Fusion stays closer to reality over time.
This is where the difference becomes clear.
The Fusion forecast has the lowest average absolute error. In other words, over time it stays closer to reality than either of the individual models. While the AI model occasionally shines and occasionally misses by a wide margin, Fusion remains remarkably consistent. GTI shows the expected behaviour: valuable as a reference, but too inaccurate as a standalone forecast.
Why this matters
In practice, reliability matters more than occasional perfection. A forecast that is spot-on on some days but wildly off on others is hard to use. For planning, expectation management, or automation, the worst days matter more than the best ones.
Fusion reduces exactly that risk. It smooths extremes without losing structure. It is not a spectacular approach, but it is an honest one.
What I take away from this
These results confirm something that appears again and again in technical systems: hybrid approaches tend to win where reality is more complex than any single model can handle. Neither physics alone nor statistics alone is sufficient to reliably model a real, local energy system.
The Fusion forecast is therefore not a compromise, but a deliberate choice against ideology. It uses what can be explained and corrects it where experience knows more than theory.
Or put differently: AI can shine. GTI can explain. Fusion is what you can rely on.