Wind Power Forecasting Under Extreme Weather: a Novel Few-Shot Learning Architecture

Neural networks (NNs) based wind power forecasting (WPF) under extreme weather conditions faces challenges, including limited sample sizes, domain shift problem between conventional and extreme weather, and difficulty in generalizing across diverse extreme weather conditions. To this end, a novel few-shot learning architecture is proposed for accurate WPF under extreme weather conditions. Firstly, a novel transfer learning strategy containing cross-task meta-training is presented, which reduces sample size requirements of traditional NNs and enhances generalization across shifted domains under diverse extreme weather conditions. Secondly, to address the trade-off between underfitting in shallow NNs and overfitting in deep NNs when incorporating embedded meta-training in transfer learning, a lightweight parameter layer is included in the NN structure. The layer reduces the number of parameters that need to be trained, thereby facilitating the effective utilization of deep NNs. Additionally, the layer helps mitigate the domain shift problem by aligning the data distribution. Finally, a cross-domain risk minimization loss function is developed to enhance the robustness and generalization of the model, adopting second-order gradients to ensure consistent forecasting performance across diverse extreme weather conditions. Numerical results using realistic wind farm data show the effectiveness of the proposed method (nRMSE: 6.26%-15.37%, reduced by 2.05%-43.55%; nMAE: 7.62%-18.54%, reduced by 0.85%-43.39% compared to benchmark models) and promise economic dispatch (38.69% cost reduction with equal 1.98% nRMSE improvement) due to accurately forecasting power trends and mitigating extreme errors under extreme weather conditions.