The safety assessment method of wind power system based on machine learning has become a hot spot at present, but the influence of sample noise is not fully considered. It is difficult to ensure the accuracy and reliability of system transient voltage stability assessment. In this paper, a noisy input multi-class Gaussian process （NIMGP） is constructed, which introduces a sparse Gaussian process and selects induction points instead of some original input points for training to reduce the complexity of model calculation. Secondly, additive Gaussian noise is introduced into the input data in the model to achieve anti-noise processing, and the Gaussian process with noisy input is approximated by Taylor series method, so that the input noise is converted into output noise and the model evaluation performance is improved. Finally, simulation is carried out in a New England 39-bus system with wind farms, including the stabilization of the transient voltage of the system, critical and unstable states and the stability margin of the stable sample for prediction. The comparison of simulation results under various working conditions show that NIMGP has strong generalization ability and good prediction accuracy under different working conditions.