Solar radiation prediction using multivariate signal decomposition and physics-informed time-frequency feature extraction.

Accurate prediction of solar radiation plays a crucial role in optimizing the solar energy system design and enhancing the efficiency of photovoltaic power grid integration. However, due to the complex dynamic characteristics of solar radiation data, the realization of such a tough task faces a formidable challenge. To this end, this study presents a solar radiation prediction method which mainly consists of the high-quality data preprocessing technique and the frequency-domain physics-informed convolutional network (FD-PICN). This method can handle the complex characteristics embedded in the solar radiation data from the spatial-temporal-frequency domain. Specifically, multivariate fast iterative filtering is first employed to synchronously decompose the multi-station solar radiation data into a series of time-frequency consistent subseries where the spatiotemporal and time-frequency correlations among multiple stations are considered. Then, FD-PICN is designed to capture the evolution pattern of solar radiation by integrating cross-attention-assisted time-frequency feature extraction and two physical coherence functions (i.e., frequency-domain coherence function and phase lock value) for high-performance prediction. Finally, numerical examples grounded in the measured data from multiple stations are utilized to validate the capability of this method. Experimental analyses demonstrate that this method outperforms other compared methods across various predictive scenarios.