CPTCFS: CausalPatchTST incorporated causal feature selection model for short-term wind power forecasting of newly built wind farms

Wind energy is increasingly vital globally, requiring precise output forecasting for stable, efficient power systems. However, this becomes particularly challenging for newly built wind farms that lack historical data. Statistical models are unsuitable in this context due to their reliance on historical data. While both physical models and data-driven transfer learning methods offer some solutions, they exhibit limitations when applied to newly built wind farms. Physical models require complex parameter tuning and high computational costs, and transfer learning generally necessitate a certain amount of historical data for model transfer. More critically, existing methods fall short in capturing the causal relationships between wind power and meteorological variables, impacting both the accuracy and robustness of the models in this specialized scenario characterized by distribution shifts. To address these challenges, this study introduces an integrated wind power forecasting model named CPTCFS, comprising two core components: Causal Feature Selection (CFS) and CausalPatchTST. The CFS identifies key features with direct causal relationships to wind power output through causal inference, surpassing traditional feature selection methods like PCA and correlation coefficient analysis. CausalPatchTST, integrating a sample weighting mechanism with the advanced Transformer variant model PatchTST, effectively addresses distribution shift issues caused by the lack of historical data in newly built wind farms, ensuring prediction accuracy and robustness in data-scarce environments. In 24-hour prediction tests using hourly data from two Australian wind farm clusters, the CausalPatchTST model with the sample weighting mechanism achieved a significant 13.29% reduction in Root Mean Square Error (RMSE) compared to the PatchTST model without this mechanism. Furthermore, the entire CPTCFS model outperforms existing models on other key accuracy indicators, demonstrating its broad applicability in the wind power forecasting domain and immense potential in other renewable energy prediction areas. •Pinpoints meteorological factors impacting wind power using Causal Feature Selection.•Introduces causal deep learning CausalPatchTST with HSIC test and sample weighting.•Shows robustness and accuracy with real-world data for renewable energy forecasting.