A machine Learning approach to spare parts inventory sizing under rare demand conditions

The industry faces, among its many challenges, the need to appropriately size the inventory of strategic spare parts. These items typically exhibit a history of low consumption; however, their unavailability may result in delays in repair and maintenance activities, leading to operational unavailability. Such effects may occur at different scales. On the one hand, maintaining a large inventory of strategic items can ensure higher operational availability. On the other hand, it entails additional costs related to storage, preservation, and tied-up capital. Therefore, a trade-off solution must be achieved. The use of traditional or simpler techniques to determine the optimal inventory level for each spare part often faces limitations due to the lack of historical data, especially in facilities at the early stages of their operation and maintenance lifecycle. Moreover, the diversity of applications associated with certain materials further increases the complexity of this problem. In this context, the present work proposes a machine learning–based methodology using two well-established algorithms: Zero-Inflated Poisson (ZIP) and Zero-Inflated Negative Binomial (ZINB). This approach is applied to approximately 27,000 spare parts, enabling a faster definition of strategic inventory levels and generating value for the company.