基于客户行为的能源零售销量预测

发布时间：2018-03-13 16:41

本文选题：销量预测　切入点：能源零售行业　出处：《山东大学》2017年硕士论文　论文类型：学位论文

【摘要】：能源行业的销量预测能够为企业内部的高效管理提供依据,帮助企业宏观调控来提高经济效益。现有方法大多基于历史销售数据进行时间序列分析,或是采用回归、神经网络等方法提取环境因素等外部数据进行预测。然而,销量受宏观经济因素、周期性因素、周边环境因素等多因素综合影响,直接体现在客户的细粒度消费行为上,现有工作在销量预测方面缺少对客户行为的深入分析,因此需要挖掘与销量相关的影响因素并量化为可解释的预测指标。本文引入细粒度的客户消费行为作为重要的分析依据,从客户行为上理解各影响因素对销量的影响,提出了基于客户行为分析的销量预测框架。通过聚合碎片化的客户行为数据,分析了代表性群体客户的购买行为与零售销量之间的内在关联性,既考虑了个体间消费习惯的差异性,又归纳出群体客户的行为模式一致性,同时为管理人员提供了可解释的依据。根据不同类型客户群体的数量分布规律,对群体客户行为进行预测,并进一步作为客户行为特征进行销量预测。针对能源零售管理模式,设计了三种零售销量预测目标:单站短期预测、单站长期预测与多站联合预测。基于销量的时间序列得到周期性特征,基于客户消费行为进行分析得到客户行为特征。单站预测使用单站相关特征,多站联合预测基于站间相关性筛选得到其他站与单站相关的特征,对单站预测特征集合进行扩展,实现联合预测,并采用包裹式特征筛选和预测模型进行销量预测。选用两个真实的油品交易数据集,从两个方面对预测方法进行验证,一方面对比了融入不同形式的客户行为特征对预测准确度的影响,另一方面对比了使用不同预测模型对预测准确度的影响,并评估预测模型的适用性。实验结果表明本文工作的有效性,基于客户行为的销量预测能够得到更好的预测结果。
[Abstract]:The prediction of sales volume in the energy industry can provide the basis for efficient management within the enterprise and help the enterprise to improve its economic benefits through macro-control. Most of the existing methods are based on historical sales data for time series analysis or regression. Neural networks and other methods extract external data such as environmental factors for prediction. However, sales volume is affected by macroeconomic factors, periodicity factors, peripheral environmental factors and other factors, which are directly reflected in the consumer's fine-grained consumption behavior. The existing work lacks in-depth analysis of customer behavior in sales forecasting, so it is necessary to mine the influencing factors related to sales volume and quantify them into interpretable predictors. In this paper, fine-grained customer consumption behavior is introduced as an important basis for analysis. In order to understand the influence of various factors on sales volume from customer behavior perspective, a sales forecasting framework based on customer behavior analysis is proposed. This paper analyzes the intrinsic relationship between the purchasing behavior of representative group customers and the retail sales volume, considers the differences of consumption habits among individuals, and concludes the consistency of behavior patterns of group customers. According to the quantity distribution law of different types of customer groups, the behavior of group customers is predicted, and the sales volume is predicted as the characteristics of customer behavior. Three kinds of retail sales forecasting objectives are designed: single station short term prediction, single station long term prediction and multi station combined forecasting. Based on the time series of sales volume, the periodic characteristics are obtained. Based on the analysis of consumer behavior, the characteristics of customer behavior are obtained. Single station prediction uses single station correlation feature, multi-station joint prediction based on inter-station correlation screening to obtain other stations and single-station correlation features, the single station prediction feature set is extended. The joint prediction is realized, and the package feature selection and forecasting model are used to predict the sales volume. Two real oil transaction data sets are used to verify the forecasting method from two aspects. On the one hand, it compares the influence of different forms of customer behavior characteristics on prediction accuracy, on the other hand, it compares the influence of different forecasting models on prediction accuracy. The experimental results show that the proposed method is effective, and the sales forecast based on customer behavior can get better prediction results.
【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：O212.1

【相似文献】