Scenario Generation and Reduction

This paper investigates the optimisation of a two-stage stochastic program model for both an original dataset and for reduced datasets which have been obtained with the use of scenario reduction. The purpose of the scenario reduction in this research is to see whether a dataset with fewer but significantly representative scenarios can provide a similar optimal solution to the one obtained with the original dataset. The scenario reduction uses distance-based methods, particularly, a Kantorovich distance-mathematical model for which several distance metrics are applied, specifically, Manhattan distance, Euclidean distance, Canberra distance, Squared Euclidean distance and Chebyshev distance; hence, each of the found reduced datasets are based on each of these distances; together with other features such as the probabilities of each scenario in the original dataset and a tolerance parameter. After finding the reduced datasets and optimising the model, some distance metrics work better than others for some fixed features, and it is noticed that the values of those features have a significant effect on finding the optimal solution for each distance metric.