Research On Safety Stock Management In Large-Scale Inventory Network

The outbreak of the COVID-19 pandemic has caused disruptions to the supply chains of multiple industries,where any problem in any link can lead to the paralysis of the entire supply chain.How to cope with uncertainty is a crucial issue in supply chain management.Safety stock is considered an effective management method to prevent shortages and deal with uncertainty.The rapid development of big data and artificial intelligence technologies has prompted numerous companies to transform towards datadriven and intelligent directions.Making intelligent decisions based on real-time and refined market data has become a key factor in enhancing the competitiveness of enterprises.How to utilize multidimensional data to make better inventory decisions is a challenge that enterprises must face in the age of digitization.This paper proposes a data model dual-driven framework for safety stock management in large-scale inventory networks,which optimizes the ”position” and ”quantity”of safety stock from a systematic perspective,while considering the demand uncertainty of multiple nodes.The framework is based on the classic guaranteed service model(GSM).This paper explores three scientific problems in inventory network management in Chapters 3 to 5,namely data-driven demand function estimation,solving large-scale guaranteed service models,and simulating network inventory strategies.Before delving into these scientific problems,the paper introduces the basic knowledge required for network inventory management in Chapter 2.Chapter 3 discusses the demand function estimation problem in the guaranteed service model.Traditional methods assume that demand follows a normal distribution and that demand is independent over time,but these assumptions are difficult to satisfy in reality.In practice,many demand nodes have “intermittent demand” issues,which represents as a high proportion of zero-demand days and have strong temporal correlation.This paper proposes a data-driven estimation method that establishes a functional quantile regression model(FQRM),which directly estimates the upper-bound demand function of each node using multidimensional features,without assuming the form of the demand distribution and time independence.An algorithm based on the alternating direction method of multipliers(ADMM)called the proximal ADMM(p-ADMM)is proposed to solve the model quickly,and a method to estimate the demand function for multiple quantile levels simultaneously is provided.Compared to traditional methods,this method has better asymptotic convergence properties in practical scenarios,making it more valuable in practical applications.Additionally,this estimation method is also suitable for estimating other time-varying demands.Chapter 4 studies how to efficiently solve large-scale guaranteed service models.An iterative decomposition(ID)algorithm that combines nonconvex optimization techniques and the structure of inventory networks is designed.This method first uses sequential linear programming(SLP)to find the locally optimal feasible solution of the guaranteed service model.Based on the analysis of feasible solutions,the inventory network is decomposed,and suitable algorithms are selected to continue the decomposition or optimization of each subnetwork based on its structure.We also provide a Python library Inv Net-GSM(https://github.com/durianh96/Inv Net-GSM)to enable readers to easily reproduce the experiments and compare the performance of different methods.The experimental results show that the ID algorithm performs significantly better than existing methods in large-scale and complex networks.Chapter 5 discusses how to simulate and evaluate the effectiveness of different strategies for large-scale inventory networks.The interdependence of nodes in inventory networks makes inventory strategy simulation very complex.This paper proposes two inventory network strategy simulation methods suitable for large-scale order-up-to inventory systems,which analyze the characteristics of nodes in the inventory network and reduce the number of simulation samples required.These two methods can significantly reduce the simulation time and maintain simulation accuracy,providing an effective evaluation tool for different inventory strategies.Inv Net-simulation(https://github.com/durianh96/Inv Net-simulation)provides the corresponding Python code.