Wind power prediction is of great importance for the safe and stable operation of power systems. A sparse vector autoregressive prediction method based on Bagging hybrid strategy and kernel density estimation (KDE) is proposed for the very-short-term probability prediction problem of multiple wind farms. First, a number of bootstrap time series are generated by time series decomposition and residual term bootstrap. For each time series, a vector autoregressive (VAR) model is used for forecasting. To address the overfitting problem that tends to occur in traditional models when the number of wind farms is large, a sparse vector autoregressive model is used to filter the most effective regression coefficients and obtain a sparse coefficient matrix. The prediction model trained by each time series produces point prediction results separately. For multiple prediction results, the KDE method is used to produce the prediction results of probability density. The effectiveness of the proposed method for multiple wind farms is verified on real wind power cluster data, and the quantile score is used to evaluate the probability prediction accuracy. The results show that the method can effectively improve probability prediction accuracy.