基于K-Shape和LSTM两阶段优化的船舶备件深度学习需求预测

Deep Learning Enabled Demand Forecasting for Ship Spare Parts Based on K-Shape Clustering and Two-Stage LSTM Optimization

摘要: 船舶备件多为突发性需求，同时具有价值高、缺件损失大等特点，准确的需求预测可以帮助船舶企业优化库存，减少缺货导致的停航损失和紧急采购的高成本。因此，本文提出一种结合K-Shape聚类与两阶段LSTM的船舶备件深度学习预测方法。首先利用K-Shape算法对船舶备件产品的历史需求数据进行聚类；其次针对具有间歇性特征的备件构建两阶段LSTM模型，以期达到更好的预测效果；最后，以W公司为案例对所提方法进行验证。结果表明，两阶段LSTM在间歇性类别中预测效果最优，各评价指标优化率均明显优于其他类别，且相较于单阶段LSTM，RMSE、MAE、R²及MASE的优化率分别可达10.99%、15.39%、9.24%及16.10%。因此，所提出的两阶段LSTM方法适用于间歇性备件的需求预测，且预测效果更好。

Abstract: The demand for ship spare parts is often sudden and characterized by high value and significant losses due to shortages. Accurate demand forecasting can help enterprises optimize inventory, reduce losses from downtime caused by shortages, and lower the high costs of emergency procurement. Therefore, this paper proposes a deep learning enabled forecasting method that combines K-Shape clustering with a two-stage LSTM model. First, the K-Shape algorithm is used to cluster historical demand data for ship spare parts. Then, a two-stage LSTM model is constructed for spare parts with intermittent demand characteristics to achieve better forecasting performance. Finally, the proposed method is validated using Company W as a case study. The results show that the two-stage LSTM model performs best in the intermittent demand category, with all evaluation metrics significantly outperforming other categories. Compared to the single-stage LSTM, the optimization rates for RMSE, MAE, R², and MASE are 10.99%, 15.39%, 9.24%, and 16.10%, respectively. Thus, the proposed two-stage LSTM method is suitable for demand forecasting of intermittent spare parts and offers superior forecasting performance.