Application of WRF-Based Single-Point Data Artificial Intelligence Post-Processing Correction Method in Practical Short-Term Wind Speed and Power Forecasting

With the rapid development of wind energy, the imperative for precise wind power predictions has intensified, with the crux lying in forecasting wind speeds. The accurate short-term (1 to 3 days) forecast of wind speeds at the hub height in boundary layer poses a significant scientific challenge. Generating such forecasts for wind farms 1 to 3 days in lead time necessitates reliance on global weather forecast products and the WRF model. In pursuit of heightened accuracy, artificial intelligence (AI) algorithms are employed to refine WRF-predicted wind speeds based on observational data.

This study draws upon observational data from five operational wind farms over three years, employing diverse deep time-series models, to examine the effectiveness and limitations of these models in post-processing corrections for WRF-predicted wind speeds. Based on our examination, we conclude that: 1) Transformer-based models have significant untapped potential, with the Pyraformer model emerging as a well-suited temporal model for post-processing corrections in wind speed and power predictions. 2) Traditional full-attention mechanisms are less effective, highlighting the importance of sparse attention as a vital approach for capturing temporal correlations in such problems. 3) The optimal model demonstrates a reduction of approximately 20% in RMSE for single-point post-processing corrections. In addition, wind speed prediction accuracy reaches around 86%, and power prediction accuracy is approximately 82%. 4) AI-based post-processing corrections may encounter challenges, including the underestimation for high-value and difficulties in reproducing forecasts below the average value.