Addressing insufficient fault warning accuracy in wind farm operations, this study proposes a regional monitoring framework integrating multi-source information with generative deep learning. We fuse turbine supervisory control and data acquisition (SCADA) data and high-resolution meteorological grids to construct spatio-temporally aligned inputs. A novel hybrid Variational Auto-Encoders-Generative Adversarial Network (VAE-GAN) architecture augments fault samples and imputes missing values, while spatio-temporal graph convolution fuses heterogeneous features. Regional fault probabilities are subsequently predicted through a Transformer model. Validated on real-world wind farm (Engie, France) and meteorological datasets (National Renewable Energy Laboratory, United States (NREL, U.S.)), the framework achieves 92.3% warning accuracy with 6.5-hour average lead time—significantly outperforming conventional single-source models. This approach provides robust technical support for intelligent wind power operation and maintenance.