基于LSTM-MHKAN的离散制造业物料需求预测方法研究  

Research on Material Demand Forecasting Method for Discrete Manufacturing Industry Based on LSTM-MHKAN

作　　者：吴小芳 成益盈 杨美怡 杨磊 WU Xiaofang;CHENG Yiying;YANG Meiyi;YANG Lei(School of Information Engineering,Henan University of Science and Technology,Luoyang 471023,China;CITIC Heavy Industries Co.,Ltd.,Luoyang 471003,China)

机构地区：[1]河南科技大学信息工程学院,河南洛阳471023 [2]中信重工机械股份有限公司,河南洛阳471003

出　　处：《河南科技大学学报(自然科学版)》2025年第2期48-58,M0005,共12页Journal of Henan University of Science And Technology:Natural Science

基　　金：国家自然科学基金项目(61976243);河南省重点研发专项项目(231111222600)。

摘　　要：在工业生产中,物料需求量受到众多因素的影响,表现出高度复杂且多变的动态特性。这些动态特性包括非线性关系、短期波动以及潜在的长期趋势,从而导致传统预测方法无法精确应对这些挑战。针对这一问题,设计了一种结合长短期记忆网络(LSTM)与Kolmogorov-Arnold网络(KAN)以及多头注意力机制(MHA)的LSTM-MHKAN物料需求预测方法。该方法通过3个步骤来优化物料需求预测过程。首先,利用LSTM捕捉物料需求数据中的时序依赖,识别短期变化并调整模型参数以适应动态波动。其次,引入MHA对LSTM输出结果进行加权,增强模型对关键需求波动的敏感性。最后,KAN算法对加权后的注意力输出进行建模,捕捉非线性关系并自适应预测未来需求。实验结果表明,与传统预测算法相比,LSTM-MHKAN能有效降低均方误差和平均绝对误差,并提高决定系数。这些结果验证了LSTM-MHKAN算法在离散制造业物料需求预测中的有效性,为减少制造业成本提供了有力的决策支持。In industrial production,material demand is influenced by a myriad of factors,exhibiting highly complex and dynamic characteristics.These dynamic features include nonlinear relationships,short-term fluctuations,and potential long-term trends,which pose significant challenges to traditional forecasting methods.To address this issue,this study proposes an innovative material demand forecasting method,LSTM-MHKAN,which integrates Long Short-Term Memory networks(LSTM),Kolmogorov-Arnold Networks(KAN),and Multi-Head Attention(MHA)mechanisms.The method optimizes the forecasting process in three key steps.First,LSTM is utilized to capture temporal dependencies within the material demand data,identifying short-term variations and adjusting model parameters to accommodate dynamic fluctuations.Second,MHA is introduced to weight the LSTM outputs,enhancing the model′s sensitivity to critical demand fluctuations.Finally,the KAN algorithm is applied to model the weighted attention outputs,capturing nonlinear relationships and adaptively forecasting future demand.Experimental results show that,compared to traditional forecasting algorithms,LSTM-MHKAN effectively reduces mean absolute error and mean absolute error,while improving R-Square.These results validate the effectiveness of LSTM-MHKAN in discrete manufacturing material demand prediction,providing strong decision support for cost reduction in the manufacturing industry.

关 键 词：物料预测 LSTM神经网络 Multi Head Attention Kolmogorov-Arnold网络 时间序列预测 

分 类 号：TP183.9[自动化与计算机技术—控制理论与控制工程]