Regional wind power forecasting guided by physics-based wind field modeling and image fusion

• By leveraging physical diffusion equations and neural network correction, the proposed method enables spatial completion and dynamic reconstruction of the regional wind field from limited spatial observations. • By constructing a regional wind field defined by the geographic boundaries of all sites, wind speed can be represented continuously at any location within the region. • By integrating physical wind field diffusion with mask-CNN correction, the wind speed time series at each site is converted into a sequence of regional wind field image features, which improves the interpretability and physical consistency of the forecasting model. Accurate forecasting of aggregated wind power output highly relies on numerical weather prediction (NWP) data. However, due to the resolution limitations, there is a geographic scale mismatch between NWP data and wind power data, which poses significant challenges to accurate wind farm modeling. To address this challenge, this work proposes a novel modelling framework. The wind evolution is simulated using physical diffusion equations derived from fluid dynamics, and the resulting wind fields are corrected using a mask-CNN model to obtain features that integrate spatial information from the entire region. These features are then encoded as a sequence of images and fed into a sequence-to-sequence (Seq2Seq) prediction model, allowing for the fusion of spatial and temporal features. For 15-min forecasting scenario, the RMSE improvement for the proposed framework exceeded 10 % on two real-world wind farm datasets. In addition, incorporating the physical diffusion equation into the modeling process reveals the mutual influence of wind speeds within the regional wind field, improving the interpretability of predictions.