Offshore wind speed forecasting remains challenging due to strong variability, nonlinear dynamics, and the limited use of physical constraints. Existing models rarely incorporate key physical factors such as the pressure gradient and wind speed relationship, the effects of temperature and humidity on air density, and historical wind patterns, while underutilizing relevant past meteorological scenarios. To address these limitations, we propose PhysDiffWind, a physics-constrained retrieval-augmented diffusion framework for accurate probabilistic offshore wind forecasting. The framework integrates a multi-module cooperative architecture. Specifically, the retrieval-augmented module employs a multi-head temporal convolutional network to encode historical multivariate sequences and dynamically retrieves representative historical patterns from a database via similarity-based embedding matching. Meanwhile, the physics-and-context-aware side information module incorporates temporal features with environmental variables such as pressure gradient and air density to provide structured guidance for the diffusion process. Furthermore, the diffusion module adopts a transformer-based conditional architecture to implement a two-stage modeling approach: forward perturbation and reverse denoising, and generates physically consistent and probabilistically expressive wind speed distributions under the modulation of retrieval samples and physical side information. Extensive experiments on real-world offshore wind datasets show that PhysDiffWind reduces the mean squared error (MSE) by up to 55.1 % and the continuous ranked probability score (CRPS) by 49.3 % compared with state-of-the-art baselines. These results confirm the framework’s effectiveness in capturing nonlinear atmospheric dynamics and improving forecasting reliability for wind farm operations.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号