FreGAT: An interpretable frequency-domain graph attention transformer for short-term photovoltaic forecasting

Short-term photovoltaic (PV) power forecasting is vital for grid stability but hindered by solar volatility and complex time-varying meteorological correlations. Although Transformer-based models excel in time series analysis, they face two key limitations in PV domains: (1) spectral bias prioritizing low-frequency trends over high-frequency fluctuations from rapid weather changes; and (2) unstable, uninterpretable modeling of dynamic multivariate dependencies, causing high-variance predictions. To address these, we propose the Frequency-Domain Graph Attention Transformer (FreGAT), integrating spectral calibration and interpretable dynamic graph modeling. Specifically, a hierarchical frequency disentanglement mechanism captures multi-scale temporal dynamics to retain high-frequency volatility. Meanwhile, a structured spatial unmixing module, parameterized by a low-rank learnable basis, regularizes dynamic graph generation, constraining it to a physically meaningful subspace for enhanced stability and interpretability. Experiments on two real-world PV datasets show FreGAT outperforms all compared baselines. It achieves up to 81.1% lower mean squared error than weaker Transformer baselines and maintains consistent modest advantages over the strongest competitors. Visual analysis confirms learned spatial modes and frequency responses align with meteorological drivers and physical interactions, validating model interpretability.