Research on data augmentation and synthetic sample quantity uncertainty in few-shot wind power prediction based on the adaptive CRITIC-HLICRVFL method

Data augmentation is crucial for few-shot wind power prediction, but an improper sample size may lead to overfitting or underfitting. Therefore, this paper proposes an adaptive data augmentation method to dynamically determine the optimal number of synthesized samples. This study introduces a hidden layer incremental cascaded random vector function link (HLICRVFL) model, which dynamically generates hidden nodes to adapt to new training samples and determines the optimal synthesis quantity based on validation performance. It also connects historical and newly added nodes to mitigate forgetting of historical information. To ensure synthesized samples cover a broader sample space, the study employs the criteria importance through intercriteria correlation (CRITIC) method and integrates three data augmentation techniques: Gaussian noise, uniform noise, and improved exponential noise. Specifically, based on the characteristics of wind power generation data, the traditional exponential probability density function is extended to the negative half-axis, proposing an improved exponential noise. Experimental results show that compared to the original random vector function link (RVFL) model, this method improves RMSE by over 23.88 % and R value by over 5.08 % on four datasets, demonstrating its accuracy in few-shot wind power prediction and effectively addressing the uncertainty of synthesized sample quantity under few-shot conditions.