To improve wind power forecasting (WPF) accuracy and ensure computational efficiency by fully and effectively using the spatio-temporal correlations between wind farms, a very short-term adaptive WPF method based on spatio-temporal correlation is proposed. Vector autoregression is applied as a basic model to characterize the spatio-temporal correlation. To avoid the over-fitting problem of a target wind farm caused by redundant spatial information, sparse modeling is adopted to optimize the weights of data from reference wind farms. The forecasting model is trained by a recursive estimation algorithm. It updates the forecasting model in real-time according to the latest wind power measurements. The model can adapt to varying environments and reduce the computational burden. A case study is carried out using real data from 100 wind farms over a region. Results show that, in comparison with a set of benchmark models, the proposed method can achieve much higher forecasting accuracy while reducing the requirement for intensive computational resources.