[2K1-GS-10] AI application: Weather / Fluids

[2K1-GS-10-02] Development of a short-term solar irradiance forecasting method using cloud dynamics and deep learning

〇Jun Sasaki Sasaki¹, Kenji Utsunomiya¹, Maki Okada¹, Koji Yamaguchi¹ (1. Japan Weather Association)

Keywords:weather forecasting, deep learning, cloud dynamics, Physics-informed neural network

In recent years, forecasting solar irradiance has become crucial for the efficient operation of expanding solar power systems. This study proposes a novel model designed to forecast solar irradiance six hours ahead. Traditionally, high-resolution numerical weather predictions (NWP) and satellite observation-based video frame prediction methods have been employed but have exhibited limitations in the accuracy of initial values and the ability to describe complex changes in clouds. To overcome these limitations, we have developed a model that combines deep learning with cloud dynamics. This approach achieves high accuracy in initial value estimation by directly using satellite observation data. Furthermore, the model describes the temporal evolution of clouds by incorporating equations of atmospheric dynamics with minimal parameters and neural network-based cloud microphysics. Compared to NWP and a video frame prediction method, our model shows superior performance in forecasting accuracy. The analysis of specific events has revealed that the model can accurately reproduce diverse changes in clouds, including small-scale cloud advection, advection due to vertical wind shear, cloud formation and dissipation, and precipitation. This significant advancement contributes to optimizing the efficiency of power grid operations.