Modeling next-generation PV devices, particularly two-terminal perovskite tandem configuration, presents a critical challenge: strong spectral selectivity. The performance of these devices is strongly influenced by spectral variability, which affects the current matching between subcells and, consequently, the overall energy yield. Accurate tandem modeling requires the separate prediction of direct and diffuse components of time-resolved spectral irradiance, which must be fed into optical and electrical simulation tools to determine the limiting photocurrents in each subcell [1]. Furthermore, knowing the subcell currents enables the identification and mitigation of spectral mismatch, which can drive the perovskite subcell into persistent reverse bias conditions, compromising the long-term reliability of tandem devices [2]. Addressing spectral sensitivity on an intra-hourly basis, in addition to improving modeling fidelity due to the data granularity, also enhances energy forecasting in dynamic electricity pricing markets—contexts in which emerging technologies like perovskite tandems are expected to operate. While predicting solar spectra under real sky conditions is inherently complex, most existing approaches rely on clear-sky models. Physical models offer high fidelity but are computationally intensive and impractical for real-time or large-scale applications. In contrast, AI modeling offer computational speed and efficiency, meaning they can deliver accurate spectral predictions with significantly reduced processing time and resource consumption. We present an AI-driven model based on Extra Trees Regression to generate accurate spectral distribution data under clear and cloudy sky conditions. Trained on over two million terrestrial spectra measured continuously over six years in Golden, Colorado, our model captures a high diversity of real-world spectral distributions. It predicts the direct and diffuse spectral distributions using only four readily available meteorological inputs: Air Mass, Clearness Index, Diffuse Fraction, and Precipitable Water. The model achieves high predictive accuracy, with Normalized Root Mean Square Error (NRMSE) mostly below 1% per wavelength and Spectral Angle Mapper under 5%, demonstrating strong spectral similarity and robustness across varying sky conditions. Additionally, our model also exhibits geographical robustness, successfully predicting the spectral distribution in Lima, Peru, a low-latitude site with extreme precipitable water levels compared to the training site.