Spares have many kinds and complex specifications, its prediction is difficult, for the problem, the paper proposes the use of nonlinear characteristics of BP neural networks and self-learning ability, based on historical data of spares consumption trains the network of all spares to determine its network model, and used for the future consumption forecast for next year. Through the predictive value and actual value correction, combined with the fill rate of the spares, and ultimately determine the future consumption of next year. The example shows that the model has a greater accuracy and practicality.