Reduction of systematic differences of LSASAF shortwave solar radiation fluxes using neural networks

To improve the process of solar energy production, we can utilize the downward

shortwave flux (DSSF) measurement, which constitutes a part of the satellite

derived total and diffuse downward surface shortwave flux (MDSSFTD) product.

MDSSFTD is issued by the Satellite Application Facility on Land Surface Analysis (LSA SAF).

However, its direct application in this area is inhibited by potential systematic

errors in the DSSF product. Therefore, this has to be addressed before the DSSF can be used downstream.

To this end, we implemented a neural network-based post-processing procedure

that uses previous temporal DSSF observations and additional predictors, such

as cloudiness and time of day, to generate a corrected DSSF value. The ground

truth for this regression task are the in-situ measurements across a variety of

locations in Slovenia. Additionally, the neural network produces DSSF estimates

in terms of quantiles, providing an uncertainty estimate of the corrected prediction itself.

We verified our new method on the aforementioned region over a period of
four years. We found that our neural network approach successfully reduces

the presence of systematic differences present in the DSSF. Additionally, the

neural network method outperforms a baseline look-up-table approach in terms
of multiple criteria, such as mean absolute error, bias, and error variability.