作者
Chaojie Zhang,Lei Zhang,Liping Zeng,Naixing Luo,Xingyu Liu,Mengmeng Jing
摘要
With the widespread deployment of photovoltaic (PV) technology in smart grids, accurate PV power forecasting has become essential for ensuring grid security and optimizing dispatch operations. However, under complex meteorological conditions and multi-source heterogeneous feature scenarios, PV power generation is jointly driven not only by historical power sequences but also by multiple weather variables such as irradiance, temperature and wind speed. Existing approaches often struggle to simultaneously characterize the dynamic dependencies among features and the multi-scale temporal properties of PV sequences, leading to redundant feature interactions and insufficient capture of critical patterns. To address these issues, this study proposes a multi-scale and multi-feature fusion model for PV power forecasting. The model employs a hierarchical multi-scale decomposition block to separate short-term fluctuations from long-term trends, and incorporates a cross-temporal–spatial channel interaction block to jointly model temporal and feature dependencies, thereby effectively integrating multi-source and multi-scale dynamic information. In addition, a unified feature selection framework is developed, which enhances the robustness of feature screening through score renormalized and weight-fusion strategies. Experimental results based on PV power data collected from a photovoltaic power plant in Guizhou Province, China, and a public benchmark dataset demonstrate that the proposed method achieves the best performance across all four evaluation metrics, including MSE, MAE, RMSE, and R 2 . On Dataset 1, the proposed model achieves an MSE of 0.0067 and an R 2 of 0.9908 for one-step forecasting. At the 12-step forecasting horizon, it reduces the MSE and MAE by 16.8% and 18.8%, respectively, compared with the best-performing competing method. On Dataset 2, the proposed method achieves an MSE of 0.0256 and an R 2 of 0.9801 for one-step forecasting. When the forecasting horizon extends to 16-step, it reduces the MSE and MAE by 11.7% and 6.9%, respectively, compared with the best baseline model. These results demonstrate the effectiveness of the proposed method and its potential to provide reliable forecasting support for the operation, control, and planning of PV power plants.