In order to solve the problem of model instability caused by spectral data difference, the transfer method of nutrient content prediction model of different batches of Chinese dishes is studied in this paper. Taking the samples of scrambled tomatoes and eggs prepared at an interval of 3 months as an example, spectral data are collected and protein content is determined by physical and chemical methods (120 samples per batch). The model of the second batch with better prediction effect is selected as the main model. A combination of PDS, MP and S/B (PDS-MP-S/B) is applied in dish model transfer to analyze the influence of different PDS window numbers and standard set numbers on the predicted results. When the number of PDS Windows is 11 and the number of standard sets is 100, the prediction result of protein content by PDS-MP-S/B algorithm is significantly better than that by no model transfer and by using the three algorithms separately. The absolute coefficient of prediction set (R2(Pred)) of prediction model is 0.9628, the relative prediction deviation is 5.6731, and the root mean square error of prediction is 0.3157. The model transfer is realized from three aspects of spectrum, model and result, which improves the universality of the model, reduces the cost of modeling, and provides theoretical support for the fast inspection of Chinese dishes.