By adding 0.3% (volume fraction) polyvinyl alcohol (PVA) fiber into C30 concrete, the full immersion-drying tests under the action of different concentration of solution were carried out respectively, so as to explore the performance of anti deterioration performance of PVA fiber concrete. Taking deterioration test data as original sample value, GM (1,1) model, BP neural network model and GM (1,1)-BP neural network combination model were established respectively to compare the fitting accuracy of sample data. The relative dynamic elastic modulus after 35~50 cycles was predicted, and the overall change trend was analyzed. The results show that the evaluation indexes of concrete specimens change most stably in 10 times of the reference concentration solution, indicating that the specimens with 0.3% (volume fraction) PVA have better anti deterioration performance in high concentration solution. GM (1,1) model can accurately predict the overall trend change of sample. BP neural network model is more accurate in predicting the change trend of single sample point, with the highest overall accuracy. The combination model overcomes the shortcomings of two single models and has the best prediction effect. The predicted value of combination model is consistent with the change trend of test value.