As you know from previous sections, reliable weather data is crucial to short- and long-term prediction of energy production. We should distinguish between two types of weather data: historical (from real-world measurements) and forecasted.
Just like in our future solar energy model, the weather forecast is also modelled. When creating models you should always focus on the domain of the problem and decide whether the model should be trained using historical weather forecasts or rather using past real measurements.
For example, if you have to predict the purchase of tickets to an amusement park, you better take into account the weather forecast, because consumers often consider it when they are planning such activities. The advantage of this approach is also that (assuming the stability of the weather forecast error itself) the weather forecast error will have an easier predictable impact on the final predictions.
On the other hand, if you are to predict the results of physical quantity measurements, it will be better to take into account real measurements of past predictors. Then the resultant forecast error will have a more difficult to analyse random distribution (resulting from interferences between the resultant model and weather forecast model). But, at the same time, the model will be resistant to changes in weather modelling.
In the solution described in this article, I used historical data for the weather, because available sunlight is variable depending on real physical conditions. Unfortunately, there are no weather stations near the location of the solar panels. But simulation data could be a good compromise.