CrossGFA: wind power prediction with a multi-scale cross-graph network via a Frequency-Enhanced Channel attention mechanism

Haoyu Zhang, Daoli Wang, Xuchu Jiang

Abstract: Wind power generation data exhibits non-periodic and non-stationary characteristics coupled with significant noise levels, posing challenges for conventional forecasting models. Existing time series prediction techniques struggle to handle the instability, high sampling frequencies, and inherent noise present in wind power data. To address these issues, we propose a novel Multiscale Cross Interaction Graph Neural Network with a Frequency-Enhanced Channel Attention Mechanism (CrossGFA). The CrossGFA effectively captures wind power trends across multiple scales via cross-scale GNN modules while reducing noise. Simultaneously, the cross-variable GNN component leverages both homogeneity and heterogeneity among variables, enhancing the detection of potential associations between different wind power characteristics. Furthermore, the frequency-enhanced channel attention mechanism complements the GNN framework by mitigating frequency domain noise. Extensive evaluations on four real-world wind power station datasets demonstrate that CrossGFA outperforms state-of-the-art time series forecasting methods, validating its effectiveness in handling the complexities of wind power data.