Options
2024
Meeting Abstract
Title
Regional PV estimation based on a ground station network as a meteorological grid
Abstract
This study demonstrates a new regional PV estimation system in Germany using historical data from a network of around 100 meteorological stations. This is compared to the satellite or numerical weather data usually used in the same, variable-resolution physical model. The PV estimation with the ground stations performs better for Germany than satellite data, despite its far lower resolution, though the satellite data maintains the advantage for small, substation regions.
The solar prediction system (SPS) used in this study uses physical models to probabilistically simulate plants of many orientations and other properties at all points of a weather data grid and works modularly with various numerical weather predictions and satellite data. It takes a somewhat opposite approach of a typical upscaling, as instead of averaging several reference input values, it simulates nonlinear effects for all possible reference locations and aggregates only probabilistically at the end.
Here, we test something new with the SPS, treating the meteorological stations of the German Weather Service (DWD) like an irregular weather data grid that is comparably sparse, with the hypothesis that it may nevertheless have positive features due to the higher quality of its observations of global and diffuse horizontal irradiance. We evaluate the results against German PV meter data as well as via correlations to vertical load time series from TSO substations.
The results question the conception that machine learning outperform physical PV modelling, so long as the meteorological inputs are truly accurate. Closing this gap has practical consequences, as physical models otherwise hold several advantages, including far superior computational efficiency, e.g. for realizing digital twin models today, as well as for distinguishing between PV production, storage, and feed-in for grid operation, which empirical models based on reference plant feed-in cannot. In addition to improving PV estimation, the results thus motivate future research to improve irradiance modelling.
The solar prediction system (SPS) used in this study uses physical models to probabilistically simulate plants of many orientations and other properties at all points of a weather data grid and works modularly with various numerical weather predictions and satellite data. It takes a somewhat opposite approach of a typical upscaling, as instead of averaging several reference input values, it simulates nonlinear effects for all possible reference locations and aggregates only probabilistically at the end.
Here, we test something new with the SPS, treating the meteorological stations of the German Weather Service (DWD) like an irregular weather data grid that is comparably sparse, with the hypothesis that it may nevertheless have positive features due to the higher quality of its observations of global and diffuse horizontal irradiance. We evaluate the results against German PV meter data as well as via correlations to vertical load time series from TSO substations.
The results question the conception that machine learning outperform physical PV modelling, so long as the meteorological inputs are truly accurate. Closing this gap has practical consequences, as physical models otherwise hold several advantages, including far superior computational efficiency, e.g. for realizing digital twin models today, as well as for distinguishing between PV production, storage, and feed-in for grid operation, which empirical models based on reference plant feed-in cannot. In addition to improving PV estimation, the results thus motivate future research to improve irradiance modelling.
Funder
Bundesministerium für Wirtschaft und Klimaschutz -BMWK-