基于张量表示的间歇性序列自适应区间预测

Adaptive Interval Prediction of Intermittent Series Based on Tensor Representation

在实际业务中,配件需求发生随机、需求量波动大,配件序列数据呈现明显的间歇性分布,同时由于人工报单失误或特殊事件等因素的影响,实际配件需求易发生异常变化,导致传统的时间序列预测方法难以捕捉配件需求量的演化规律,预测结果不确定性高、可靠性不足。为解决上述问题,提出了一种基于张量表示的间歇性序列自适应区间预测方法。首先,利用层次聚类,基于间歇性序列的平均需求间隔和平方变异系数指标筛选相似序列形成序列簇,用于提取簇内公共需求演化信息,增加可预测性;其次,通过张量分解重构原始需求序列,在最大限度保留序列核心信息的前提下平滑序列中的异常值;最后,构建一种自适应预测区间算法,通过动态更新机制得到配件需求量的预测值和预测区间,以确保结果的可靠性。利用某大型轨道交通制造企业实际售后数据进行验证,与现有典型时间序列预测方法相比,所提方法可有效挖掘不同特点间歇性序列的演化趋势,提高小样本间歇性序列的预测精度。实验结果表明:所提方法在间歇性特有指标均方根标准误差(RMSSE)上,相较于需求预测主流的深度学习方法平均降低了0.32,且当预测结果出现失真时,可提供一个可靠的弹性预测区间,为实际应用中企业智能备件计划决策提供了一种新的解决方案。

In the actual business,parts demand occured randomly and demand fluctuates,so the demand sequence for spare parts showed obvious intermittent distribution.At the same time,due to factors such as manual reporting errors or special events,the actual demand for spare parts was prone to abnormal changes,making it difficult for traditional time series prediction methods to capture the evolution of the demand for accessories,resulting in high uncertainty and insufficient reliability of prediction results.To solve this problem,an adaptive interval prediction method for intermittent series based on tensor representation was proposed.Firstly,hierarchical clustering was used to screen similar sequences based on the average demand interval and square of the coefficient of variation of acces-sory sequences,forming sequence clusters to increase predictability.Secondly,the original demand sequence was reconstructed by tensor decomposition.The outliers in the sequence were then corrected while retaining the core information of the original sequence to maximum extent.Finally,an adaptive prediction interval algorithm was constructed,which could obtain the predicted value and prediction interval of the parts demand through the dynamic update mechanism to ensure the reliability of the results.The proposed method was validated on the aftersales dataset from a large vehicle manufacturing enterprise.Compared with existing time series prediction methods,the proposed method could effectively extract the evolutionary trend of various types of intermittent series and improve the prediction accuracy on the intermittent time series with small size as well.Experiments showed that the average root mean square scaled error(RMSSE)of this method was 0.32 lower than that of the mainstream in-depth learning method of demand prediction.More importantly,when the prediction results were distorted,the proposed method could provide a reliable and flexible prediction interval,which could be helpful to provide a feasible solution for intelligent parts management.