Abstract: The problem of designing a production-distribution network which consists of plants, distribution centers and terminal markets is studied under the uncertainty of upstream production and downstream demand. Two statements of normality and abnormality are taken into consideration for the uncertainty of upstream production, and three statements of low, medium and high are taken into consideration for the uncertainty of downstream demand. Due to the abnormality in production can lead to the defective products, whether implementating the products monitoring in the upstream plant is considered. By considering both the cost of network operation and the performance risk caused by the uncertainty, three two-stage stochastic programming models for designing a production-distribtion network are developed. The first one is based on the expected cost minimization model which ignores the risk caused by the uncertainty; The second one is then presented by using condition value-at-rick (CVaR) to measure the cost performance of the production-distribution network. However, the CVaR criterion focus exclusively on the down-side risk which will lead to a too conservative solution. To overcome this weakness, both the expected cost and the corresponding CVaR measurement are considered to develop a Mean-CVaR-based model which is characterized by the risk aversion level and the pessimistic coefficient. Specially, the uncertainties in production and demand are described with a series of discrete scenarios which are generated by scenario tree approach. For the large-scale numbers of uncertain scenarios caused by the numerous potential nodes in the network, the scenario reduction technology is used to filtrate the scenarios, which significantly reduces the difficulty of solving the presented models. At last, some numerical calculations are executed to analyze the influence of the relevant parameters on the network performance, and the Pareto Effective Frontier evaluated by the expected cost and the conditional risk value is given. Furthermore, the impacts of the risk aversion level and the pessimistic coefficient on the performance of the production-distribution network are examined by the regression experimental design. The results show that the pessimistic coefficient has a greater impact on the network performance than the risk aversion level.In theory, the developed models in this paper can be easily expanded by considering the supplier selection or multi-period operations. In practice, the proposed models provideflexible options for the enterprise to build the production-distribution network. Moreover, by considering the risk-aversion attitude of the decision maker, the CVaR-based models can also provide effective operational decision support for the enterprise to avoid the potential loss induced by the uncertainty and risk.

Key words: production-distribution network design, risk aversion, mean-CVaR, uncertainty, experimental design