Scalable regional joint wind power forecasting method supporting dynamic incorporation of new farms

As wind farm clusters expand, reliable forecasting for newly constructed wind farms becomes increasingly important. However, these farms typically have limited historical power data during early operation, while their incorporation alters the structure and data distribution of the existing cluster. This paper proposes a scalable joint wind power forecasting method with a teacher-student architecture for dynamic cluster expansion. A teacher model is trained on existing wind farms to learn spatial adjacency and temporal attention patterns, which are used to construct a Behavior Infrastructure Graph (BIG). When new wind farms are incorporated, the BIG is expanded based on wind power correlations and used to initialize the expanded teacher model. The teacher model is then adapted using limited data from the new farms. Finally, knowledge distillation trains a student model under the joint supervision of the expanded teacher outputs and observed power data from the new farms. This allows the student model to inherit the teacher’s forecasting capability while learning farm-specific information. Case studies involving 35 wind farms demonstrate that the proposed method effectively expands the forecasting model, improves forecasting accuracy over longer forecast horizons for newly constructed wind farms, and maintains stable forecasting performance for existing wind farms.