Data-driven Demand Forecasting And Intelligent Placement Research In Fast-moving Consumer Goods Supply Chain

As market competition intensifies,supply chain management for fast-moving consumer goods(FMCG)retail operations faces increasingly complex challenges.Diverse and personalized consumer demands,ever-shortening product lifecycles,and intensifying market competition create uncertainty in consumer demands.This uncertainty poses important challenges for business operations,and to mitigate the impact of demand uncertainty,it is crucial to have accurate demand forecasting and product delivery strategies for retail orders.However,traditional order demand forecasting models and product delivery strategies have some problems,such as the inability to use frequency data for prediction and a lack of coordination mechanisms.To address these issues,this paper proposes a method for predicting sales demand for FMCG and an intelligent product delivery strategy based on the prediction results.The main research contents are as follows:(1)Analysis and processing of FMCG retail business data.To address the massive,widely-sourced,and disordered nature of retail data,we use various data collection systems and other related systems to acquire relevant data.To address the problems with collected data,we integrate the data through cleaning,integration,and transformation methods,including data set selection,aggregation,cleaning,and transformation,to provide a preliminary feature set strongly associated with the forecasting results.Then,using variance filtering and mutual information filtering,we extract representative data features with strong associations to support subsequent time-series prediction and recommendation models.(2)We propose a Transfomer time-series prediction model based on order frequency segmentation,targeting the prediction of retail customer order demand.This paper notes that the high turnover rate and low shelf life of FMCG result in high order frequency,but traditional models focus more on the time sequence of order data and sales impact on sales cycles,neglecting the detailed information contained in order frequency.To address this issue,we propose a frequency sequence extraction method based on gated convolution in the improved self-attention mechanism time-series prediction model and improve the original Transformer input encoding.We propose a seq-fre multi-head selfattention structure based on improved encoding for modeling inter-sequence dependencies.We validate the model’s performance on a data set of orders from an FMCG company,and the results show that the model has some improvements over other baseline models.The results show that the improved SFTransformer model is more suitable for forecasting the demand for FMCG retail orders and lays the foundation for subsequent research on intelligent delivery models.(3)We propose an intelligent product delivery fusion model that integrates Bert4 REC and Deep FM.The original Bert4 REC recommendation algorithm focuses more on individual customers,making it unsuitable for the business perspective of enterprises,resulting in problems such as data sparsity and low accuracy in low-data-level retail order business data.To address this issue,we use marketing activity features and SFTranformer model prediction data as the weight module basis,adopt the Deep FM model for feature weight analysis,enabling the recommendation algorithm to focus on the needs of FMCG enterprises.We use user-product interaction data as basic data and adopt the Bert4 REC model for user-product behavior relationship modeling.We then use an MLP multi-layer perception to fuse the output into a hybrid model.The results show that the product delivery effect of the model has some improvement in three indicators compared to other models,indicating that the model is more suitable for FMCG enterprise scenarios.