End-to-end learning of representative PV capacity factors from aggregated PV feed-ins

Energy system models rely on accurate weather information to capture the spatio-temporal characteristics of renewable energy generation. Whereas energy system models are often solved with high abstraction of the actual energy system, meteorological data from reanalysis or satellites provides rich gridded information of the weather. The mapping from meteorological data to renewable energy generation usually relies on major assumptions as for solar photovoltaic energy the photovoltaic module parameters. In this study, we show that these assumptions can lead to large deviations between the reported and estimated energy, as shown for the case of photovoltaic energy in Germany. We propose a novel gradient-based end-to-end framework that can learn local representative photovoltaic capacity factors from aggregated PV feed-ins. As part of the end-to-end framework, we compare physical and neural network model formulations to obtain a functional mapping from meteorological data to photovoltaic capacity factors. We show that all the methods developed have better performance than commonly used reference methods. Both physical and neural network models have much better performance than reference models whereas operational use cases may prefer the neural network due to higher accuracy while interpretable, physical models are more suited to academic settings.