Safety stock is a critical component of inventory management for core manufacturing enterprise. Nevertheless, current safety stock models struggle to effectively address the intermittent distribution characteristics of spare parts work order data, which makes it fail to obtain reliable safety stock decision when facing uncertain cases of significant demand volatility and random failure. To address this concern, this paper proposes a dynamic safety stock model with intermittent spare parts demand (SPD) forecasting. First, this paper proposes a multivariate intermittent time series forecasting method based on a Bayesian graph neural network. By leveraging the graph structure, this method captures structured information among series, while the incorporation of the Bayesian network enables the evaluation of the uncertainty within SPD sequences, thereby achieving confidence interval forecasting for intermittent time series. Then, this paper obtains a basic setting of reorder point and maximum stock on a three level warehousing architecture by minimizing excess inventory cost and shortage cost simultaneously. And a dynamic safety stock is obtained by fusing the forecasting interval into the basic safety stock setting. Finally, an actual spare parts dataset from a Chinese large rail transit manufacturing enterprise is employed for validation. The experimental results show that the proposed model not only gets higher SPD prediction accuracy, but also simultaneously improves inventory turnover and coverage, Furthermore, it verifies that more accurate demand forecasting is key to enhancing performance of safety stock.