In order to fully reduce the error of ultra-short term power prediction of decentralized wind farms, a model of ultra short-term power prediction of decentralized wind farms based on self-organizing competitive neural network virtual wind measurement is proposed. The architecture of decentralized wind farm power prediction system is analyzed from the aspects of software, hardware and data flow, which provides the basis for the establishment of the model. The self-organizing competitive neural network theory is used to establish the virtual wind measurement model of large-scale wind farm. The virtual wind measurement point measurement data, temperature, weather type, wind direction, humidity and other prediction feature sets are used as input data, and the ultra-short term fan wind speed is used as output data to build the network. Thus, the wind speed environment of different fans in the decentralized wind farm is obtained based on the measurement data of several virtual wind measurement towers. Furthermore, the generalized error distribution optimization algorithm is used to fit the parameter data of each fan, and then, the ultra-short term predicted output of each fan is calculated based on the virtual wind measurement results. The validity of the proposed model is verified by an ultra-short term wind power prediction example of a regional distributed wind farm.