Demand Forecasting for Server Manufacturing Using Neural Networks

Within the global market, server manufacturers face the challenges associated with competition, customer service costs, and a large variety of components that are expensive and carry a high inventory cost. Additional challenges arise from the extensive testing of components required for quality assurance, the complex product configurations, and the potential long lead time from suppliers. The unique nature of server manufacturing makes an accurate forecasting model imperative to developing a quick response to face these challenges. Currently at the operational level, demand forecasting in server systems is reliant on a discrete event simulation approach. This model relies on the demand data obtained from the planning department. Under certain conditions the model may be able to forecast inventory costs with up to an 80% accuracy level. This research proposes a neural network approach to forecasting the demands of server systems. Trends and seasonality patterns in historical demand data are extracted and modeled. Then a design of experiment is used to create the appropriate parameters for the network. The developed forecasting model is trained, tested, and validated using 52 weeks of demand data. The neural-network-based model can predict future demand with an average of 89% and 84% accuracy level for server models 1 and 2, respectively.