Spare parts management is the key to improving the reliability of production lines and achieving cost reduction and production efficiency. A spare parts demand forecasting model based on the improved Croston method was proposed for spare parts demand with intermittent and random characteristics. Based on the Syntetos criterion, the spare parts demand was classified into four categories by intermittent and fluctuating characteristics. For demands with fluctuating characteristics, spare parts demand prediction was decomposed into two types of problems: demand occurrence state prediction and demand quantity prediction, and an ensemble empirical mode decomposition (EEMD)-long short-term memory (LSTM) prediction model was designed for it. The EEMD method was applied to decompose the fluctuating sequence into several relatively smooth components, and then the LSTM method was used to forecast each component. For demands with intermittent characteristics, a signal modulation technique was introduced to serialize the binary sequence of occurrence states. The proposed method solves the problems of strong fluctuation and intermittency of spare parts demand. It has been applied to China Tobacco Hubei Industrial Co., Ltd., which proved the superiority and feasibility of the method.