基于EEMD和GRU的市场需求预测研究——人工智能时代的供应链管理与决策支持体系探索  

Market Demand Forecasting Research Based on EEMD and GRU—Supply Chain Management and Decision Support System Exploration in the Era of Artificial Intelligence

作　　者：李博阳 徐志昊 

机构地区：[1]吉林大学,吉林 长春 [2]青岛大学,山东 青岛 [3]中国科学院青岛生物能源与过程研究所,山东 青岛

出　　处：《计算机科学与应用》2022年第2期290-303,共14页Computer Science and Application

摘　　要：市场需求预测在社会生产和提高管理决策水平等方面意义重大。虽然当前的研究集中于使用深度学习模型对市场需求进行预测,但是市场需求时序序列中存在着高频率的噪声和复杂的非线性波动模式,这些模型对市场需求的变化规律和对某些局部极值的拟合能力并不出色。因此,DeepMDF被引入用于预测非平稳、非线性的市场需求数据。DeepMDF主要由两部分组成,分别是用于分解市场需求数据的集成经验模态分解(EEMD)和门控循环单元(GRU)。EEMD可以降低GRU拟合数据非线性波动规律的难度;GRU可以针对长期的时序数据进行建模。实验表明,DeepMDF的性能出色。与基准模型相比,DeepMDF在预测市场需求方面的均方根误差(RMSE)平均降低了约69.06%。总的来说,DeepMDF能够很好地完成预测市场需求的工作。Market demand forecasting is significant in social production and improving management decisions. Although current studies have focused on forecasting market demand using deep learning models, there are high-frequency noise and complex nonlinear fluctuation patterns in the market demand time series. These models did not have excellent ability to fit the changing patterns of market demand and some local extremes. Therefore, DeepMDF was introduced for predicting non-stationary and nonlinear market demand data. DeepMDF consists of two main components, the Ensemble Empirical Mode Decomposition (EEMD) for decomposing market demand data and the Gated Recurrent Units (GRU). EEMD can reduce the difficulty of fitting the nonlinear fluctuation pattern of data by GRU. GRU can model for long-term time-series data modeling. Experiments showed that DeepMDF had excellent performance. Compared with the baselines, the Root Mean Square Error (RMSE) reduced by approximately 69.06% on average. Overall, DeepMDF can do the job of forecasting market demand very well.

关 键 词：市场需求预测 时序序列 非线性波动模式 集成经验模态分解 门控循环单元 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]