基于CNN-LSTM网络的在线多任务销售预测模型  

Online multitask sales forecasting model based on CNN-LSTM network

作　　者：王旭 廖涛[1] 张顺香[1] WANG Xu;LIAO Tao;ZHANG Shunxiang(School of Computer Science and Engineering,Anhui University of Science and Technology,Huainan Anhui 232001,China)

机构地区：[1]安徽理工大学计算机科学与工程学院,安徽淮南232001

出　　处：《阜阳师范大学学报（自然科学版）》2021年第2期85-91,共7页Journal of Fuyang Normal University:Natural Science

基　　金：国家自然科学基金面上项目(62076006);安徽省高校优秀青年人才支持计划项目(gxyq2017007);安徽省高等学校自然研究重点项目(KJ2016A202)资助。

摘　　要：有效的销售预测利于企业制定正确的营销策略,针对当前销售预测研究中存在无法进行实时在线多任务销售预测、稳定获取序列数据中的时序特征等问题,提出了一种基于CNN-LSTM网络的在线多任务销售预测模型。该模型上层CNN网络抽取得到重要的时序数据特征,进而输入到下层LSTM网络中进一步抽取复杂的不规则特征进行建模,最终得到预测结果。实验结果表明:CNN-LSTM模型不仅能在测试集上取得12.61%的最小平均绝对百分比误差,同时在长时间销售预测中,在线模型预测效果优于离线模型。Effective sales forecast is conducive to the formulation of the correct marketing strategy.Aiming at the problems in current sales prediction research,such as the inability to make real-time online multitask sales prediction and the stable acqui-sition of timing characteristics in sequential data,putting forward an online multitask sales forecasting model based on CNN-LSTM network.The preprocessed data are extracted from the upper CNN network to obtain important time-series data features,which are then input into the lower LSTM network,and complex irregular features are further extracted for modeling,and final-ly the prediction results are obtained.Experiments results show that the CNN-LSTM model can not only achieve the minimum average absolute percentage error of 12.61%on the test set,but also the online model is better than the offline model in the long-term sales forecast.

关 键 词：在线学习 多时间序列 卷积神经网络 长短期记忆网络 销售预测 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]