A Multiscale Photovoltaic Power Forecasting Framework Using VMD and GRUFormer-Based Hybrid Deep Learning

Accurate photovoltaic (PV) power forecasting is crucial for grid stability and large-scale renewable integration. Traditional methods often struggle with the multiscale, non-stationary nature of PV signals. This study introduces a novel hybrid forecasting framework that integrates variational mode decomposition (VMD) with frequency-adaptive heterogeneous deep learning architectures, which combine diverse neural network structures—such as Transformer, GRUFormer, and gated recurrent units (GRU)—to leverage their complementary modeling capabilities for different frequency components. The raw PV power series is first decomposed into low-frequency (LF), mid-frequency (MF), and high-frequency (HF) components using VMD. For MF prediction, we introduce GRUFormer, a hybrid model that integrates GRU and Transformer architectures to capture both local periodic patterns and global temporal dependencies. A Transformer network is applied to model long-term trends in the LF component, while a GRU network is used to capture transient fluctuations in the HF component. The final output aggregates these predictions. Experiments on real-world datasets demonstrate the effectiveness of the proposed framework. Additional cross-seasonal and multi-weather evaluations confirm its robustness and generalization capability. Cross-dataset experiments further demonstrate the model’s strong transferability and robustness across different regions and data sources. Ablation studies highlight the significance of assigning models based on frequency characteristics in improving prediction performance. The proposed method achieves RMSE, MAE, and R²scores of 0.5110 MW, 0.2887 MW, and 0.9985, respectively, outperforming several state-of-the-art approaches. This work provides a systematic, interpretable, and extensible approach for multiscale PV power forecasting, supporting more reliable intelligent grid management.