Interval Prediction Method for Wind Power Based on Informer-KAN

When employing the Informer model for wind power prediction, inherent noise in wind power datasets and the model's final layer's reliance on simplistic linear output have been observed to compromise the model's capacity for capturing nonlinear relationships, consequently resulting in suboptimal prediction accuracy. To address these limitations, a novel Informer-KAN wind power prediction model is proposed. The fully connected layers of the original Informer model are systematically replaced with Kolmogorov-Arnold Networks (KAN) layers constructed using cubic B-spline functions, which are designed to significantly enhance the model's capability for capturing complex nonlinear interactions and subsequently improve wind power prediction precision. Moreover, kernel density estimation is applied to the Informer-KAN model's predictions to generate probabilistic forecasts, enabling the derivation of power prediction intervals at varying confidence levels.