excerpt: A useful solar forecast earns trust only when it is checked against real production day after day. By combining irradiance-based modelling with learned system behaviour, Fusion Forecast turns prediction into an auditable process rather than a polished curve.
Fusion Forecast comparison page reworked
During Pentecost I finally had time to rework the Fusion Forecast comparison page:
https://f97.be/fusion_comparison.html
Fusion Forecast is my attempt to improve daily solar production forecasts by combining two different views of the same photovoltaic system. One view comes from physics: the GTI Forecast, based on the expected solar radiation on the tilted panel surface. The other view comes from learned behaviour: the AI Forecast, based on historical production and weather-related patterns from my own installation.
The Fusion Forecast page shows the operational forecast:
The comparison page is the test bench behind it. It brings together three different ways of estimating the same thing: how much electricity my photovoltaic system will produce on a given day. The page compares the GTI Forecast, the AI Forecast and the Fusion Forecast with the measured production from my Fronius inverter.
A forecast is only useful if it can be checked. It is not enough to draw a smooth line into the future. The model has to show, day by day, how far it was away from actual production, where it performed better than the individual methods, and where it still missed.
That is what the comparison page now does. It shows the forecasts over time, compares absolute errors, separates clear, mixed and overcast days, checks for systematic overestimation or underestimation, and provides a daily audit table. Fusion Forecast is not meant to be a black box. It is an inspectable blend of physical modelling and learned behaviour, tested against real production data from a real installation.
For the technical background, I have also written a short paper:
Fusion Forecasting for Residential PV: An Adaptive Regression Blend of AI and GTI with Online Calibration
https://f97.be/download/fusion.pdf
The GTI Forecast
The GTI Forecast is based on Global Tilted Irradiance. GTI describes how much solar radiation reaches the panels at their actual angle and orientation. For a PV system, this is more useful than a general weather forecast, because solar panels do not receive sunlight on a flat horizontal plane. They receive it on a tilted surface facing a specific direction.
The GTI Forecast is close to the physics of solar production. It starts from expected solar radiation and converts that into an estimated daily energy yield. This makes it a strong reference model. On clear days, when irradiance and production follow a clean physical pattern, GTI can be very reliable.
Its weakness is that it can miss local and operational effects. It may not fully capture passing clouds, small-scale weather differences, dirt on panels, inverter behaviour, temperature effects, shading or other site-specific deviations. GTI knows a lot about available sunlight, but less about how one specific installation behaves every day.
The AI Forecast
The AI Forecast is based on historical production and weather-related patterns. It uses past data to learn how the PV system usually behaves under certain conditions. Instead of calculating production only from physical irradiance values, it looks for recurring relationships in the available data.
This makes the AI Forecast useful for capturing local effects that are difficult to describe with a simple physical model. Shading, seasonal behaviour, inverter limits, panel orientation, local cloud patterns and systematic deviations can all leave traces in historical data. A learning model can pick up some of these patterns.
Its weakness is that it depends on the quality and representativeness of the training data. If the weather situation is unusual, if the data history is too short, or if the model has learned a pattern that does not apply to the next day, the forecast can be wrong. AI is good at recognizing patterns, but it does not automatically understand the physical limits of a solar installation.
The Fusion Forecast
The Fusion Forecast combines both approaches. It takes the AI Forecast and the GTI Forecast and blends them into one daily prediction. In the page shown here, the Fusion model can use a learned model with coefficients for AI and GTI, or fall back to a fixed hybrid weighting when no learned model is available. The page also applies practical limits, for example by preventing physically impossible production values.
The point is not that AI is always better, or that GTI is always better. Both models contain useful information. GTI contributes the physical solar reference. AI contributes learned behaviour from the real installation. Fusion tests whether combining both produces a forecast that is more stable than either source alone.
This matters because forecast errors are not evenly distributed. One method may perform well on clear days, another may handle mixed or cloudy conditions better. Fusion reduces dependence on a single view of the system.
What the page measures
The overview chart shows actual production, GTI prediction, AI prediction and Fusion prediction over time. This makes visible whether Fusion follows the real production curve more closely than either input model alone. The individual lines can be switched on and off, so the behaviour of each method can be inspected separately.
The error statistics compare median absolute error, average absolute error and the larger errors at the edge of the distribution. These tail values matter because a model with a good average can still be unreliable if it occasionally misses badly. For energy planning, the difficult days are often more important than the easy ones.
The page also separates the results by weather regime. Clear days, mixed days and overcast days are treated differently, because forecast models do not fail in the same way under all conditions. A GTI-based forecast may perform well when sunlight and production follow a clean physical pattern. AI may compensate for recurring local effects. Fusion tries to combine both strengths.
The signed error chart shows whether a model tends to overestimate or underestimate production. This is important for practical use. A forecast that is often too optimistic creates different problems than one that is systematically too cautious.
The calibration view compares predicted production with actual production. A well-calibrated model should stay close to the diagonal line where prediction and reality match. This helps reveal whether a method behaves proportionally across low, medium and high production days.
Finally, the daily table provides the audit trail. For every date it shows actual production, the three forecasts, their errors, the closest method and the relative error margin. The page does not ask the reader to trust the model. It exposes the model’s behaviour in detail.
Fusion Forecast is not a black box. It is an inspectable blend of physical modelling and learned behaviour. GTI provides the physical reference, AI captures patterns in the historical data, and the Fusion layer tests whether combining both produces a more robust daily forecast for a real photovoltaic installation.