The integration of renewable energy sources, particularly solar photovoltaics (PV), has increased the complexity of real‐time data management. Heterogeneous devices with varying control cycles and frequent missing data further complicate this process. Traditional rule‐based or device‐specific methods require extensive data collection and parameter tuning, making them impractical for large‐scale systems with diverse devices and time resolutions. Recent advances in deep learning enable zero‐shot forecasting, even when target devices are absent from training data. However, their generalization performance in zero‐shot cross‐frequency forecasting—predicting unseen devices and time resolutions—remains largely unexplored. To address this gap, a large‐scale multiresolution dataset is constructed with over 100 million data points, including solar irradiance and PV generation data across several time resolutions. An empirical analysis is conducted to evaluate various state‐of‐the‐art forecasting approaches, training on multiple time resolutions (e.g., 1 s, 1 min, 1 h) and testing on untrained resolutions (e.g., 5 min, 30 min). The results show that forecasting performance varies with time resolution, with different architectures excelling at different granularities. Motivated by this finding, a mixture‐of‐experts framework is proposed to dynamically combine models by time resolution, enhancing robustness and generalization. Our findings offer insights into scalable, efficient real‐time forecasting for renewable energy.
Suemitsu et al. (Sun,) studied this question.