The problem with solar forecasts
Solar energy is abundant, clean, and predictable — at least in theory. The sun rises every day, follows a well-defined path across the sky, and delivers its energy according to celestial mechanics that can be calculated with great precision. And yet anyone who has lived with a photovoltaic system knows: the actual energy you get out of your panels on any given day can be surprisingly hard to predict. Clouds, temperature, haze, seasonal shifts, and inverter limits all conspire to turn a “sunny outlook” into a disappointing harvest.
Over the last years I have worked with two very different forecasting approaches:
- GTI forecasts – physics-based models using solar geometry and irradiance projections
- AI forecasts – data-driven machine learning models trained on years of weather and production history
Each has strengths. Each has weaknesses. And that’s exactly why I built the Fusion Forecast, which blends them intelligently into a hybrid that is more reliable than either approach on its own.
GTI: Physics on your roof
The GTI (Global Tilted Irradiance) forecast is a physics-based method. It starts from astronomical facts: the position of the sun, the angle of your panels, the orientation of your roof. It then combines this with irradiance projections (how much solar radiation will actually reach the ground) from a weather service. From there, the calculation is straightforward: apply inverter and temperature corrections, adjust for shading or clipping, and you get a predicted energy yield.
Strengths of GTI
- Transparency: Every step can be explained in physical terms. If the result looks odd, you can usually pinpoint why.
- Low data needs: No training set required—just panel specs, orientation, and a weather forecast.
- Robust under normal conditions: When the sky is clear, GTI is very good at estimating production.
Weaknesses of GTI
- Weather sensitivity: Clouds are the Achilles’ heel. Irradiance forecasts are notoriously variable, and GTI cannot “learn” how local weather patterns affect your roof.
- No correction for bias: If the weather model systematically overestimates radiation in your region, GTI will inherit that error.
- Rigidity: GTI cannot adapt when conditions deviate from the standard physics assumptions.
AI: Learning from the past
The AI forecast takes a different route. Instead of applying physics, it learns patterns from historical data. I trained machine learning models (Random Forest, Gradient Boosting) on more than 1500 days of paired weather data and measured production. The model ingests features like temperature, humidity, wind, cloud cover, and solar radiation, then outputs a predicted daily yield.
Strengths of AI
- Local adaptation: Because it trains on my inverter and my weather, the AI learns site-specific quirks that physics alone can’t capture.
- Pattern recognition: AI can capture complex, non-linear interactions between variables, e.g. how certain cloud cover plus wind speed correlates with production.
- Self-improving: As more data is fed in, the model can be retrained to become more accurate.
Weaknesses of AI
- Data hunger: It needs a long, high-quality history of production data to work well.
- Opacity: Unlike GTI, it’s often hard to explain why the AI predicted what it did.
- Instability under rare conditions: AI can struggle when it encounters weather patterns not present in its training data (e.g. a once-in-a-decade heatwave).
Fusion Forecast: The best of both worlds
The Fusion Forecast is my answer to these trade-offs. Rather than choosing between physics and data, it uses both. Technically, it’s a regression blend:
[ \text{Fusion} = b + w_{AI} \cdot \text{AI} + w_{GTI} \cdot \text{GTI} ]
Where:
- (w_{AI}) and (w_{GTI}) are weights learned from recent history
- (b) is an intercept that corrects systematic bias
If there isn’t enough recent data to train these weights, Fusion falls back to a simple 70/30 hybrid (70% AI, 30% GTI). On top of that, an adaptive calibration step rescales the forecast so it aligns with the most recent two weeks of actual performance.
Why Fusion works
- Adaptive weighting: If AI is performing better lately, Fusion leans more on AI. If GTI has the edge, it shifts toward GTI.
- Bias correction: The intercept and calibration scale adjust for consistent over- or under-prediction.
- Resilience: By combining two independent signals, Fusion avoids catastrophic misses when one method fails.
A concrete example
On 27 August 2025 the numbers looked like this:
- AI forecast (calibrated): 44.6 kWh
- GTI forecast: 30.9 kWh
- Fusion forecast: 36.8 kWh
- Actual production: 35.82 kWh
- Error margin: Fusion only 2.8 % off the truth
Fusion positioned itself between the more optimistic AI and the more conservative GTI, leaning toward the most probable outcome based on its learned weights and intercept. As the day progressed and real production data became available, the advantage of this balanced approach became clear: over time, Fusion has consistently outperformed both standalone methods in terms of average error.
Conclusion
Forecasting solar energy isn’t about picking the “right” model. It’s about recognizing that every approach has blind spots.
- GTI gives us transparency and physics, but misses local weather effects.
- AI gives us adaptability and nuance, but can overfit or misfire on rare conditions.
- Fusion unites the two, balancing their strengths and cancelling their weaknesses.
That’s why my dashboards now show all three side by side—but Fusion is the one I trust most when it comes to planning tomorrow’s energy use. And there is another reason why this hybrid approach matters: thanks to climate change, the weather we experience is becoming warmer overall, but also increasingly hazy and unstable. A purely data-driven AI tends to assume that more heat means more sun, and it will take years of new training data before the models catch up with the changing atmosphere. Fusion, by contrast, can compensate right now by anchoring the AI optimism with the physics-based GTI baseline, keeping the forecast realistic even as our climate shifts.