基于VMD-EWT-QWLSTM-TPE深度学习模型的超短时物流需求多步预测  

Multi-step prediction of ultra-short-term logistics demand based on VMDEWT-QWLSTM-TPE deep learning model

作　　者：杨新彪 陈彦如[1,2] 秦娟 冉茂亮[1] YANG Xin-biao;CHEN Yan-ru;QIN Juan;RAN Mao-liang(School of Economics and Management,Southwest Jiaotong University,Chengdu 610031,China;Key Laboratory of Service Science and Innovation of Sichuan Province,Chengdu 610031,China)

机构地区：[1]西南交通大学经济管理学院,成都610031 [2]服务科学与创新四川省重点实验室,成都610031

出　　处：《控制与决策》2024年第6期1859-1868,共10页Control and Decision

基　　金：国家自然科学基金项目(71771190)。

摘　　要：超短时物流需求预测是企业物流资源智能调度的重要基础,然而,超短时物流需求数据具有强随机性、高波动性、非平稳性等特征,进行多步精确预测较为困难.基于此,构建基于串行数据分解和量子加权深度网络的超短时物流需求多步预测模型.首先,通过变分模态分解(VMD)和经验小波变换(EWT)的串行分解方法对超短时物流需求数据的时序特征进行有效提取,以剥离噪声信号,降低原始数据的非平稳性和随机性;然后,构建量子加权长短期记忆神经网络(QWLSTM)深度学习模型,设计多输入多输出策略对分解后的模态分量进行多步预测,并基于树形Parzen评估器(TPE)对QWLSTM的超参数组进行优化;最后,对各模态分量的预测结果进行重构.实验结果表明,所提出模型在平均绝对值误差(MAE)、均方误差(MSE)、加权平均绝对百分比误差(WMAPE)、校正决定系数(R2)方面,均优于其他15种对比模型.Ultra-short-term logistics demand forecasting is important for intelligent scheduling of enterprise logistics resources.As ultra-short-term logistics demand data is random,highly volatile,and nonstationary,it is difficult to accurately predict them for multi-steps.Considering such characteristics,this research proposes a combination model for ultra-short-term logistics demand forecasting based on the serial data decomposition and quantum-weighted deep nerual network.Firstly,the time series features of the ultra-short-term logistics demand data are extracted with the decomposition method of serializing variational mode decomposition(VMD)and empirical wavelet transform(EWT)to strip the noise signal and reduce the non-stationarity and randomness of the original data.Secondly,a quantumweighted long short-term memory neural network(QWLSTM)deep learning model is developed and a multi-input multi-output strategy is designed to predict the decomposed mode components in multi-steps,also the hyper-parameters of the QWLSTM are optimized by the Tree-Parzen-Estimator(TPE).Finally,the prediction results for each mode component are reconstructed.Numerous experiments are conducted,and the results show that the proposed model performs better than other 15 comparison models.

关 键 词：超短时物流需求 多步预测 串行数据分解 量子加权 深度学习模型 TPE参数优化 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]