EURO 2025 Leeds
Abstract Submission

2301. Scenario Space Search for Two-Stage Stochastic Optimization Using Deep Generative Models

Invited abstract in session TD-38: Foundation Models and Optimization, stream Data Science meets Optimization.

Tuesday, 14:30-16:00
Room: Michael Sadler LG19

Authors (first author is the speaker)

1. Atefeh Hemmati Golsefidi
DTU Managment, Technical University of Denmark (DTU)
2. David Pisinger
Management, DTU

Abstract

Many important decisions are taken under uncertainty since we do not know the development of various parameters. In particular the ongoing green transition requires large and urgent societal investments in new energy modes, infrastructure and technology. The decisions are spanning over a very long time-horizon, and there are large uncertainty towards energy prices, demand of energy, and production from renewable sources. Such problems can be described as two-stage stochastic optimization problems. If the decision variables are discrete, such problems are extremely difficult to solve.
Instead of solving a complex stochastic optimization problem defined on a fixed set of forecasted scenarios, we propose to use an iterative process: We repeatedly generate new scenarios, solve them and find the corresponding investment solutions. Our novel way of optimization makes use of Deep Generative Models to generate small sets of scenarios matching the real distribution, and use a guided local search process to select scenarios that properly reflect properties of the full set of scenarios. The process can be seen as a local search in the scenario space. The framework is tested on three real-life problems: Facility location of charging stations, the capacity expansion problem, and vehicle routing problem using real-life historic scenarios.Computational results show that the framework is able to generate a small set of artificial scenarios that result in high-quality first-stage decisions.

Keywords

Status: accepted

Back to the list of papers