Satellite-Derived Solar Irradiance Nowcasting using Spatiotemporal Deep Learning

Abstract:

In recent years, integrating renewable solar energy into modern electrical grids has led to unprecedented research into surface solar irradiance forecasting. Due to solar being a variable energy resource, it is essential to have accurate solar irradiance forecasts to ensure system operators can maintain the stability of the power grid.

The resource variability is explained by a deterministic component (Earth's movement with respect to the Sun) and a stochastic component (for example, the presence of clouds). The latter is widely regarded as the primary reason behind uncertainty in current solar irradiance forecasts, which has been identified as one of the critical challenges in widespread photovoltaic integration across the globe (Antonanzas et al., 2016). By extension, we are primarily interested in improving short-term forecasts of cloud properties.

We investigate several state-of-the-art video prediction algorithms from the deep learning literature for nowcasting satellite-derived surface solar irradiance. We include several state-of-the-art methods from the meteorology and renewable energy literature to compare our model with current methods.

We use the Surface Radiation Data Set – Heliosat (SARAH) 2-1, which is derived from the European Meteosat Second Generation geostationary satellites (Pfeifroth et al., 2019). Specifically, we use the effective cloud albedo, which directly measures clouds' impact on the solar surface irradiance by filtering away the ground albedo, thus enabling our model to focus exclusively on cloud dynamics.

In addition to satellite-derived solar irradiance measurements, we include four ground stations in Germany from the popular Baseline Surface Radiation Network (BSRN) to serve as an external validation of our results.

The experimental comparison is based on a wide range of metrics from the solar irradiance literature, including the Kolmogorov Test Integral. Initial results show the spatiotemporal deep learning models achieving higher accuracy than current state-of-the-art solar irradiance nowcasting techniques for both satellite-derived surface solar irradiance and ground-based pyranometer measurements for 2017. Examples of spatial and temporal accuracy will be presented along with an analysis of the accuracy between NWP and purely satellite-based methods