Localized solar radiation zoning by combining spatially continuous estimates and Gaussian mixture models

Abstract

With the increasing role of solar energy in the global decarbonization, precise geographical zoning of solar radiation becomes crucial. Traditional methods of solar radiation zoning struggle to accurately distinguish subtle spatial and temporal differences in solar radiation due to both sparse ground-based observations and the requirement for a predefined zone number, which limits their applicability for the demands of distributed photovoltaic system. This study introduces a novel method for localized solar radiation zoning, integrating spatially continuous solar radiation data with a Gaussian mixture model. High-precision spatiotemporal estimates of solar radiation are achieved by employing deep learning algorithms to analyze meteorological satellite imagery and digital elevation model data. The use of an infinite Gaussian mixture model along with variational inference allows for the adaptive determination of the number of solar radiation zones. The case study in Guangxi Province shows that incorporating Digital Elevation Model data reduces the root mean square error of global solar radiation estimates from 134.06 W/m² to 87.68 W/m² and accurately reveals temporal and spatial variability in both global and diffuse solar radiation. This approach not only prevents overfitting when the predefined upper bound surpasses the actual number of zones but also facilitates the development of zoning schemes that can range from fine-grained, capturing subtle variations, to coarse-grained, focusing on overall patterns. The outcomes lay a solid foundation for localized regional assessment and efficient utilization of solar energy resources.