Fast ground irradiance computations for agrivoltaics via physics-informed deep learning models

Developing photovoltaic tracker algorithms for bifacial solar modules in agrivoltaic systems requires computationally intensive raytracing simulations to accurately quantify irradiation. Sunlight distribution on ground and module levels is essential for optimizing the setup and operation of tiltable PV systems, maximizing crop and electrical yield under various weather conditions and tilt configurations. We introduce a deep learning-based surrogate model that computes ground-level irradiation in a complex agrivoltaic scene with PV tracking. The surrogate model is physics-informed since the training data includes raytracing outputs based on real weather data. It computes the ground irradiance map based on direct normal irradiance, diffuse horizontal irradiance, solar position, and system geometry in just 3ms, four orders of magnitude faster than standard raytracing. The presented encoding of the 3D scene allows the calculation of ground irradiance using generative regression models. Our surrogate model allows on-the-fly raytracing calculations for edge computing-based PV tracker applications, where computational efforts must be minimized to enable efficient management and optimization of PV systems.