To address the issue that wind power forecasting typically requires substantial historical data and support from complex models, a wind power forecasting model under small-sample conditions is proposed. First, aiming at the problem of insufficient samples, a data augmentation method based on time-Markov chain Monte Carlo (Time-MCMC) is developed. Statistical analysis, interpolation, fitting, and adaptive recognition methods are employed for data cleaning to improve the diversity and quality of the samples. Second, by integrating multiple optimization strategies, a wind power forecasting model is constructed, where the backpropagation (BP) neural network is optimized using the improved hybrid aquila optimization and African vulture optimization algorithm (IHAOAVOA). Finally, the effectiveness of the data augmentation method is verified using actual case studies. Meanwhile, power forecasting is conducted on the augmented sample set using four models: the BP neural network, the BP neural network optimized by the aquila optimizer (AO), the BP neural network optimized by the African vulture optimization algorithm (AVOA), and the IHAOAVOA-BP neural network. The forecasting results show that, compared with the AVOA-BP model, the mean absolute error (MAE) and mean squared error (MSE) of the proposed model are reduced by at least 0.45 MW and 21.48% respectively.