Machine learning-based Direct Normal Irradiance (DNI) forecasting using satellite data for Concentrated Solar Power (CSP) plants with Thermal Energy Storage (TES)

Abstract

Efficient operation of next-generation concentrated solar power (CSP) plants with thermal energy storage (TES) requires reliable direct normal irradiance (DNI) forecasting to optimize dispatch strategies and reduce the levelized cost of energy (LCOE). While ground-based forecasting methods offer high precision, their implementation is often cost-prohibitive for many plants. In this study, we propose a cost-effective forecasting framework using Exponential Gaussian Process Regression (Exp-GPR) trained on geostationary Himawari-8 satellite data. To capture temporal dependencies, 17 meteorological and radiative variables were utilized across lead times ranging from 30 to 360 minutes. Our results demonstrate that the Exp-GPR model achieves robust accuracy, with an of 0.89 and an nRMSE of 0.09 at the 60-minute horizon. To ensure model interpretability, Shapley Additive Explanations (SHAP) were applied, identifying the Solar Zenith Angle and cloud properties as the primary predictive drivers. Cross-site evaluation across diverse South Asian regions confirms the model’s transferability, although performance varies with topographic complexity. This SHAP-augmented pipeline provides an interpretable and scalable tool for CSP operators to mitigate forecast risk and leverage the thermal inertia of TES systems effectively.