The Feedforward Neural Network (FNN) and the Proper Orthogonal Decomposition (POD) method are used to construct the nonlinear relationship between the POD coefficients of the anti-icing temperature field and the input parameters, which establishes a fast prediction model of the surface temperature in the design of anti-icing system. The K-fold cross-validation method is introduced to fully utilize the data samples and ensure the reliability and stability of the prediction model. The results show that the prediction accuracy of the wing 2D surface temperature is greatly improved compared with the results only predicted with the FNN. The average prediction error of 8 test cases is within 1.89 ℃, and the maximum prediction error is 8.64 ℃, which meets the simulation accuracy requirement for the anti-icing system performance.