Smarter Forecasts: Fusion Forecasts

When I first started forecasting solar production for our installation the approach was simple: physics. The GTI model (Global Tilted Irradiance) takes into account the orientation of the panels, their tilt, the movement of the sun across the sky, and irradiance forecasts provided by weather services. Out of that comes a clean curve of expected daily production. It is steady and rooted in solar geometry, and for a while it gave me a reliable baseline.

But physics alone is not the full story of solar energy in the messy real world. Local haze that is never picked up by the forecast, sudden bursts of rain that dim the light for just half an hour, or even temperature effects that change panel efficiency all chip away at the neat predictions the GTI curve provides. The model can tell you how much energy you should get on a clear June day, but it has no idea how dust on the panels or a rogue cumulus cloud at 3 p.m. will change the outcome. In practice GTI forecasts often looked elegant on paper yet diverged noticeably from what the inverter reported in the evening.

The next step was to teach a machine. Instead of relying purely on geometry and physics, I built an AI model and fed it years of production data alongside detailed weather records. This gave the system not just formulas but memory. It could learn from real-world patterns, from how our panels behave under Belgian skies, and from the countless quirks that no irradiance table can ever capture.

Almost immediately the difference became clear. The AI forecast had a knack for spotting local nuances: the way morning haze tends to burn off around ten, how windy days subtly cool the panels and improve efficiency, or how certain cloud patterns in the forecast almost always mean rain in our valley. In many cases it simply outperformed the GTI model, producing forecasts that matched the inverter with uncanny accuracy.

But there was a catch. When the weather turned unusual the AI sometimes lost its footing. A sudden storm front or a week of atypical haze could send the model swinging wildly, overshooting or undershooting by large margins. In other words, the AI was brilliant at recognising familiar situations, but when pushed outside its comfort zone it could get carried away.

So I let the two talk. Instead of choosing between physics or machine learning, I combined them. Enter the Hybrid forecast, a weighted blend of both AI and GTI. The idea was simple: let the AI take the lead, but let physics have a steadying hand on the wheel. After experimenting with different weights, I found that a 70/30 ratio worked best: seventy percent from the AI, thirty percent from GTI.

The effect was immediate. The AI’s ability to learn from local quirks still shaped most of the forecast, but the GTI acted as a stabilizer whenever the machine tried to overreach. The result was a curve that kept the sharpness of AI but avoided its worst swings.

On many days the Hybrid landed noticeably closer to reality than either model on its own. When the AI overshot, GTI pulled it back. When the physics underestimated, AI nudged it up. It was like pairing intuition with discipline — the creative learner and the calm scientist working side by side. For the first time I felt like the forecast had become more than just a tool; it had turned into a dialogue between two very different ways of seeing the sun.

Now comes the latest stage in this little evolution: Fusion. Where Hybrid relied on a fixed 70/30 rule, Fusion takes the idea one step further by actually learning the best blend from history. Instead of me hard-coding the weights, Fusion looks back at how AI and GTI have performed against actual production and figures out for itself how to combine them most effectively.

It does this by training a simple regression model on past days, treating the AI forecast, the GTI forecast, and even a constant offset as inputs. The result is a set of weights that can shift over time. On some stretches, Fusion leans more on AI because it has been consistently close to reality. On others, it gives GTI a stronger hand when physics proves steadier than the machine. The offset allows Fusion to correct for systematic drifts, like when both models tend to overshoot slightly in midsummer or undershoot in winter.

And importantly, Fusion knows its limits. If the training window is too thin — say, only a handful of recent days with valid data — it does not try to be clever. In those cases it gracefully falls back to the familiar 70/30 split, ensuring that the forecast is never worse than what has already proven to work well.

In practice, Fusion feels less like a formula and more like a living balance. It adapts, it remembers, and it corrects. Each morning the number it shows is no longer just physics or just machine learning, nor even a compromise between the two. It is the best of both, fused into a single forecast that evolves with the weather and with the seasons.

Why does this matter? Because solar forecasting is a moving target. The weather is quirky (especially here in Belgium), the seasons shift, and the system itself ages. Fusion is my way of letting physics and AI play as a team, not rivals. On test days Fusion has already proven to be the most reliable column on my dashboard — less extreme, more balanced, and much closer to what the inverter actually reports in the evening.

From GTI to AI to Hybrid to Fusion, this has been an iterative journey of making the forecasts not just smarter, but also more trustworthy. And while perfection is still impossible (the clouds don’t read my code), the progress is obvious: what started as a rough guess is slowly becoming a reliable daily guide.

Visit the experimental fusion forecast here.