CGForecast: Radar conditioned graph neural network for ultra-short-term wind power forecasting

Deep learning methods have shown promising results in multi-station ultra-short-term wind power forecasting yet remain challenging under extreme convective weather conditions due to over-reliance on coarse-resolution numerical weather prediction (NWP). To tackle this limitation, we propose CGForecast, a multimodal dual-branch framework for probabilistic forecasting across multiple wind farms. First, to address the insufficient utilization of NWP field data, CGForecast employs a graph convolutional network (GCN) to learn time-varying inter-site relations while injecting regional NWP wind field features into node states to couple local measurements with large-scale meteorological forecasts. To improve forecast stability, a Transformer-based branch is accompanied to extract global features across stations for deterministic forecasts. Then, to enhance responsiveness to convective events, we condition the GCN features by real-time radar observations via a conditional variational autoencoder, enabling the model to adjust its internal type representation according to ongoing weather without resorting to discrete event classification. Finally, we jointly optimize two branches to gain probabilistic forecasts. Extensive experiments on a year-long dataset from 10 wind farms in the Hubei Province of 2024 indicate that CGForecast improves aggregate probabilistic accuracy relative to strong baselines in most seasons where CRPS is slightly worse than scalar-input baselines in an Autumn regime, together with competitive gains under convective and rapidly changing weather types.