The present paper presents a new NIR analysis method with partial least square regression (PLS) and artificial neural network (ANN) to improve the prediction precision of the protein model for milk powder. First, an efficient method named region selecting by genetic algorithms (RS-GA) was used to select the calibration region, and then the GA-PLS model was made to predict the linear part of the protein content in milk powder. And then in the region selected by RS-GA method, principal component analysis (PCA) was calculated. The principal components were taken as the input of ANN model. The remnant values by subtracting the standard values and the GA-PLS validation values were regarded as the output of ANN. The ANN model was made to predict the nonlinear part of the protein content. The final result of the model was the addition of the two model's validation values, and the root mean squared error of prediction (RMSEP) was used to estimate the mixed model. A full region PLS model (Fr-PLS) was also made, and the RMSEP of the Fr-PLS, GA-PLS and GA-PLS+PC-ANN model was 0.511, 0.440 and 0.235, respectively. The results show that the prediction precision of the protein model for milk powder was largely improved when adding the nonlinear port in the NIR model, and this method can also be used for other complex material to improve the prediction precision.