Machine learning-based sensitivity of solar radiation forecasts to cloud-related parameters in WRF-Solar

초록

Accurate prediction of global horizontal irradiance (GHI) is essential for reliable solar power forecasting. However, shallow convective clouds introduce substantial uncertainty due to their pronounced spatiotemporal variability, intermittently blocking and scattering incoming solar radiation, thereby causing fluctuations in surface irradiance. Since such clouds cannot be explicitly resolved in numerical weather prediction models, they are represented by parameterization schemes, in which numerous tunable parameters often become major sources of uncertainty. To quantify parameter contributions to this uncertainty, we applied Sobol' global sensitivity analysis to the Deng shallow cumulus scheme in the Weather Research and Forecasting (WRF)-Solar model, using a Gaussian process regression-based surrogate model to emulate the model responses. We identified the four most impactful parameters; two of them directly control effective cloud fraction and exerted the strongest influence. Shapley additive explanations further revealed that the most influential parameter contributed approximately twice as much as the second most influential parameter to GHI RMSE variability. Increasing these two parameters enhanced cloud cover, reduced GHI, and lowered prediction errors. The other two also affected GHI, though less strongly. These tendencies identified by the surrogate model were consistently observed in Sobol' analysis and WRF-Solar simulations, confirming the physical relevance of the resulting sensitivities. This study presents an approach for identifying influential parameters in a shallow cumulus parameterization and provides practical guidance for parameter tuning to improve GHI prediction performance under the examined conditions, supporting more reliable solar power forecasting.

키워드