Optimization of deterministic and probabilistic forecasting for wind power based on ensemble learning

Abstract

High-performance wind power forecasting (WPF) is crucial for wind farm and grid management, particularly for tasks such as dispatch and storage planning. However, the inherent uncertainty in wind power poses significant challenges to reliable forecasting. While self-attention mechanisms and kernel density estimation (KDE) have been widely utilized in WPF, further improvements in model performance remains a tough task. To address these gaps, this paper proposes a novel deterministic WPF model that incorporates Lorenz noise-based data augmentation and the probsparse self-attention-based Informer. The model interpretability is further enhanced by providing insights into the contribution of input features through shapley additive explanations (SHAP). For probabilistic forecasting, the model is optimized using power scenarios and the multi-bandwidth kernel density estimation (MBKDE) method. Finally, a case study involving 5 wind farms is conducted. The results demonstrate a 13.44% improvement in deterministic forecast accuracy compared to the benchmark, with an additional 1.35% improvement following data augmentation. Interpretability analysis shows that adding 7 iterations of Lorenz noise enhances forecasting accuracy. Furthermore, the probabilistic forecasting model shows an improvement of at least 2.89% in overall performance.