| With the development of society and the enhancement of public health awareness,retail pharmacies are playing an important role in the daily life of the public,and the scale of pharmaceutical chain enterprises is also growing and expanding.As a result,the number of retail pharmacies has surged in recent years,and the competition among drug retailers has gradually increased.Making full use of the massive data generated in the process of sales management through data mining is of great significance to the rational allocation of resources and efficient operation of enterprises.Reasonable sales forecast has a profound impact on the procurement,distribution and inventory management of enterprises,at the same time,it’s conducive to helping enterprises make sales decisions and optimize supply chain management.In order to improve the overstock or shortage of retail chain drugstores,this paper conducted data mining and demand forecasting research on the massive data generated in the process of sales management of H enterprise.The research is mainly divided into three steps: feature extraction,influence factor exploration and sales prediction model building.In the feature extraction stage,the data files from different sources of H enterprise are integrated,and the data are preprocessed.In view of the small amount of information reflected by some original features,the feature enhancement is carried out by constructing features.On the basis of the original basic features,the promotion methods and historical sales conditions are mainly refined.Finally,we extract the big promotion sales data set and daily sales data set.In the exploration stage of influencing factors,the influence of commodity attributes,store attributes and external seasonal factors on drug sales was explored by means of seasonal index,linear regression and the importance of integrated learning model features.The research showed that store information,commodity information,seasonal information and recent sales had a significant impact on sales.In the model building stage,which is the research focus of this paper,based on the characteristics of H enterprise data including static information,historical observation information and known future information,this paper improves the multivariate time series prediction model based on time fusion converters with cyclic neural network and attention mechanism as the main components,aiming at the noise impact caused by the input of known future information to the original model.The attention diagram is modified by adding a branch of future knowledge information guidance on the basis of multiple attention modules of the model.The experimental data were used in the large sales promotion data set and daily sales data set of H enterprise extracted in the first stage of this paper.The experimental results show that the prediction error of the improved algorithm is significantly reduced compared with the original model.Finally,this paper describes the significance of this model in the application of retail chain drugstores,and discusses the limitations of this model and the future research direction. |
