EUROPT 2025
Abstract Submission

248. Data-driven dynamic scenario generation using Benders decomposition with adaptive oracles

Invited abstract in session TB-12: Optimisation under uncertainty in the power sector, stream Applications: AI, uncertainty management and sustainability.

Tuesday, 10:30-12:30
Room: B100/8009

Authors (first author is the speaker)

1. Hongyu Zhang
School of Mathematical Sciences, University of Southampton
2. Ken McKinnon
School of Mathematics, Edinburgh University
3. Andrew Reeves
The University of Edinburgh

Abstract

Scenario generation for a multi-stage stochastic program often aims to find a good finite discrete approximation of random variables, which is represented by a scenario tree. Many existing scenario generation algorithms perform scenario generation and reduction once prior to solving the problem and the tree is fixed. However, the information gained when solving a multi-stage stochastic program may detect other parameters that should have been treated uncertain and included in an expanded scenario tree. This paper proposes a data-driven dynamic scenario generation algorithm using Benders decomposition with adaptive oracles to dynamically adjust the scenario tree in the course of solving the underlying problem. The proposed algorithm detects the parameters that provide the most added value by being treated uncertain and returns a scenario tree and an ∈-optimal solution. The proposed scenario generation algorithm concerns the class of multi-stage stochastic programs with block separable recourse. We apply the proposed scenario generation algorithm to a stochastic power system investment planning problem, and the results show that (1) the proposed rules can effectively identify which parameters to explore in an expanded tree, and (2) the proposed algorithm can estimate the impact of expanding a scenario tree to include additional parameters without much computational effort.

Keywords

Status: accepted

Back to the list of papers