A multi-stage graph-conditioned probabilistic forecaster with dual-stream temporal encoding for wind speed forecasting

Accurate wind speed forecasting is a critical foundation for wind power dispatch and the grid integration of renewable energy. The non-stationary fluctuations in wind speed time series, the spatially dependent variations among multiple wind turbines, and forecasting uncertainties continue to pose significant challenges for wind farm-level forecasting. This paper proposes an artificial intelligence framework for multi-turbine wind speed forecasting, termed the Multi-Stage Graph-Conditioned Probabilistic Forecaster (MGPF).The model fuses prior graphs and data-driven graphs at the node-pair level, enabling differentiated spatial connection representations for different wind turbine pairs. Simultaneously, it sets up two candidate temporal encoding paths, a Generative Pre-trained Transformer 2 (GPT-2) and a lightweight Patch, and determines the appropriate configuration based on comparative results across different datasets. Spatio-temporal cross-attention further establishes interactions between spatial and temporal features, while the probabilistic prediction head constructs prediction intervals by estimating the mean and variance of future wind speeds. Experimental results on two wind farm datasets with markedly different scales show that MGPF reduces the root mean square error (RMSE) by 35.56% and 3.92%, respectively, compared to the corresponding optimal baselines, experiments at the Kelmarsh external wind farm further validate the model's applicability under different wind farm data conditions. Probabilistic forecast results show that MGPF outperforms the probabilistic baseline overall in terms of probability distribution fitting and forecast interval quality, encoder and pre-training control experiments also validate the effectiveness of the time-encoding design.