In the demand forecast of air material spare parts, we must fully consider that the actual air material spare parts are affected by many factors, which have the characteristics of multi-component, complexity and non-stationary. In this paper, based on wavelet analysis, the non-stationary original data is decomposed to the appropriate number of layers, so as to realize the stabilization of signal. Because the low-frequency signal reconstructed after decomposition represents the trend term and the high-frequency signal represents the random term, the GM(1,1) model and AR(p) model are selected to process the low-frequency signal and the high-frequency signal respectively, and the forecast value of each component is synthesized to obtain the final forecast result. The effectiveness of the algorithm is verified by an example.