Time–Frequency Hybrid Model Based on Moran’s Graph Network and the SVP-RF Residual Framework for Day-Ahead Solar Irradiance Prediction

Accurate prediction of solar irradiance plays a vital role in photovoltaic (PV) power generation and grid stability. Given that solar irradiance is significantly affected by spatial factors such as temperature and cloud cover, which greatly impact prediction accuracy, and that traditional models often fail to produce precise power forecasts, this paper proposes a time–frequency hybrid model for day-ahead multisite solar irradiance forecasting. The proposed model is built on the architecture of Moran graph networks, seasonal trend decomposition using loess–variational mode decomposition–permutation entropy (SVP) decomposition-aggregation, and a random forest (RF) residual network framework. Firstly, based on the spatiotemporal characteristics of PV power station clusters, the Moran index is used to extract significant features from the spatiotemporal graph network, thereby achieving higher predictive gains. Secondly, due to the strong nonlinearity and nonstationarity of global horizontal irradiance (GHI), whose time series are influenced by weather changes, cloud cover, diurnal cycles, and seasonal factors, a combined modal decomposition method integrating seasonal trend decomposition using loess (STL) and variational mode decomposition (VMD) is adopted to better capture its underlying features. Permutation entropy (PE) is then applied to calculate the complexity of each subsequence for reconstruction. At the same time, residuals between the entropy-reconstructed sequences and the original remainders are computed and predicted using the Transformer algorithm to reduce prediction errors. Finally, considering the differences in frequency and fluctuation among subsequences, a contrastive learning strategy and a model fusion strategy are designed within the proposed ensemble forecasting framework. To train the hybrid model, a solar irradiance dataset from the Xinjiang region in China was used as a case study. Four different time-step prediction tasks (3-, 6-, 9-, and 12-h ahead) were designed. Compared with existing models, the proposed hybrid model demonstrated superior performance in both accuracy and efficiency. Specifically, the root-mean square error (RMSE) and the coefficient of determination ( R 2 ) at the 12-h horizon reached 30.05 and 0.966, respectively, representing improvements of 9.16% and 0.026 over a stacking ensemble model without residual consideration. Moreover, across all four time-step prediction tasks, the proposed model consistently outperformed other models in terms of average RMSE and R 2 scores, further demonstrating its overall forecasting superiority. These findings indicate that the proposed model offers a promising solution to the challenging task of PV power prediction.