Distributed photovoltaic (PV) systems are characterized by wide and dispersed deployment. However, the spatial distribution, combined with significant power fluctuations caused by meteorological disturbances, poses substantial challenges for accurate power prediction. To address these challenges, a grid-based KACNN-GATransformer distributed PV power prediction framework is proposed. First, a unified geographic grid is constructed based on numerical weather prediction (NWP) grids, enabling grid-based mapping and construction of PV panel positions. Second, the KACNN model is adopted, which utilizes its learnable spline convolution kernels to adaptively extract spatial features of key meteorological factors such as local irradiance and cloud cover, effectively suppressing noise interference. Finally, the GATransformer model is designed, which incorporates a cross-attention mechanism and a dynamic masking strategy into its encoder-decoder architecture. This allows for dynamic coupling of the long-term temporal dependencies of historical power with the driving effect of NWP data, achieving deep fusion of cross-modal spatiotemporal features. Experimental results on datasets containing various typical weather scenarios demonstrate that the proposed model exhibits good robustness in complex meteorological scenarios such as cloudy and rainy weather, with prediction errors significantly lower than those of traditional prediction methods. This study provides effective technical support for reliable dispatching decisions in high-penetration PV grid integration.