Column-and-constraint generation approach to partition-based risk-averse two-stage stochastic programs

Abstract

Typically, two-stage stochastic programs have been modeled and solved based on the finite support assumption, but the large number of scenarios makes it hard to solve, and there also are potential risks of inaccurate estimation of underlying distribution. In this paper, to mitigate the drawbacks, we present a novel risk-averse two-stage stochastic program with finite support, which we call partition-based risk-averse two-stage stochastic program. In the program, a set of scenarios is partitioned into several groups, and the second-stage cost is defined as the expectation of risk levels for all of the groups. In particular, the conditional value-at-risk is considered as a risk measure for each group, and so the risk level of the model is affected by a quantile parameter or a partition of a given set of scenarios. In order to solve the model exactly for a given partition, a column-and-constraint generation algorithm is proposed. In addition, a scenario partitioning algorithm to enable the risk level of the model to be close to a given target is devised, and partitioning schemes for combining it with the proposed column-and-constraint generation algorithm are proposed. Extensive numerical experiments were performed that demonstrated the effectiveness of the proposed partitioning schemes and the efficiency of the proposed solution approach.