Wind Speed Scenario Generation under Typhoon Conditions via Prediction-Anchored Diffusion

Wind speed sequences under typhoon conditions exhibit complex spatio-temporal non-stationary characteristics, making it challenging for existing methods to control key features. This paper proposes a method for generating conditional wind speed scenarios based on a prediction-anchored diffusion model (PADM). First, a typhoon-aware time-series graph neural network (TA-TGNN) is developed, which constructs typhoon-aware adjacency matrices through trajectory similarity, temporal decay, and intensity similarity to characterize non-stationary dependencies. Second, a dual-dimensional adaptive attention mechanism (DDAM) is introduced, combining phase detectors and physical constraints to adaptively select and fuse key information across temporal and feature dimensions, thereby enhancing phase-aware prediction. Finally, PADM leverages deterministic prediction results as anchors and integrates hierarchical conditional encoding with temporal awareness to generate physically consistent and diverse wind speed scenarios. Case studies demonstrate that the proposed method accurately reproduces the structural characteristics of typhoon wind speeds and captures variations under different conditions, while maintaining stability and consistency across various hyperparameter settings.