Abstract: This paper is concerned with distribution and inventory replenishment of a global consumer electronics manufacturing supply chain, which consists of a central warehouse, oversea central distribution centers, and retailers. Direct shipments from the central distribution centers to retailers leads to faster response but more risk of obsolete inventory, compared to shipment from central warehouse to the retailers by air transportation. The silent features of the problem such as lifecycle and sales pattern are analyzed in details, then demand forecasting model using real-time sales data, and the inventory replenishment optimization models are developed so as to tradeoff the service level and risks in the supply. Specifically, the demand forecast by machine learning algorithms is proposed, and considering the shipment paths, service levels, supply capacity and inventory levels, a nonlinear programming model to minimize the total quantity in the inventory system is developed. Moreover, the nonlinear programming model is then linearized as a linear programming model and the sample average approximation is employed to solve the problem. Also, a heuristic algorithm is proposed based on a modified base-stock policy. Numerical studies show that the heuristic algorithm can obtain results close to that of mathematical programming with much less computation effort. Compared with the real operations data, the proposed methods can effectively reduce inventory and increase the service level.

Key words: distribution supply chain, multi-echelon inventory optimization, demand forecasting, machine learning, sample average approximation

摘要： 以某大型消费电子品制造商的产品跨境配送和补货为背景，研究跨境分销供应链中的补货管理和库存优化问题。该分销供应链由国内中央仓、海外地区分销仓（或国家分销仓，简称本地仓）和本地零售商组成。其中：通过地区分销仓可以直接给零售商发货，因而具有更快的服务响应速度，但其中的库存有较大的呆滞风险；国内中央仓可直接向零售商发货，但其服务响应速度则大幅降低。分析了该问题的重要特征，提出了实时销量数据驱动、基于机器学习的需求预测；建立了同时考虑本地仓直发比例、服务水平、产能限制与库存限制，以最小化系统库存水平的非线性规划模型，并通过抽样平均逼近将模型转化为线性规划，设计了基于基本库存水平的启发式优化算法。基于实际数据的计算表明，该模型及其求解算法可有效地降低系统的库存水平，并提高系统的服务水平。

关键词: 分销供应链, 多级库存优化, 需求预测, 机器学习, 抽样平均逼近