Date: April 26, 2024
Time: 10-11am EST (Eastern Standard Time)
Speaker 1: Renata Pedrini
Talk 1 Title: Handling the Impact of Climate Change in the Long-Term Generation Scheduling Problem via Distributionally Robust SDDP
Speaker 2: Aras Selvi
Talk 2 Title: It's All in the Mix: Wasserstein Machine Learning with Mixed Features
Zoom Information: Sign up to our mailing list at:
https://mymaillists.usc.edu/sympa/subscribe/stoprog-webinar-l
————————————————————————————————————————————————————————————————
Talk 1 Abstract: The long-term generation scheduling (LTGS) problem presented important developments over the years. In this work, we analyze different risk aversion strategies for handling the impact of climate change in the LTGS problem in hydro-dominated power systems using stochastic dual dynamic programming (SDDP). Despite the advances in modeling and solution strategy, there are important issues related to the distributions of inflows (and other random variables), especially with climate change. First, even though time series or other statistical methods are used to model these random parameters, the true probability distribution is never fully known. Furthermore, historical values used to devise the statistical models may no longer be valid as processes shift. We focus on solutions to mitigate these issues protecting the system from deviations from the inflow scenario distribution, achieved by a Distributionally Robust Optimization (DRO) framework. Instead of a single distribution, DRO considers all distributions that are sufficiently close to this nominal distribution and optimizes a worst-case expected (or risk-averse) objective, where the expectations concern all the considered distributions. DRO is more realistic because it explicitly considers existing data while recognizing that forecasts may contain errors. The DRO policies are tested against risk-neutral (expected value minimization) and CVaR risk-averse approaches using data from the Brazilian power system. Policies are compared as well as the practical application of different algorithms. The results indicate that incorporating DRO improves the out-of-sample performance of policies.
(Authors: R. Pedrini, G. Bayraksan, E. C. Finardi, F. Beltrán)
Talk 2 Abstract: The recent advent of data-driven and end-to-end decision-making across different areas of operations management has led to an ever closer integration of prediction models from machine learning and optimization models from operations research. A key challenge in this context is the presence of estimation errors in the prediction models, which tend to be amplified by the subsequent optimization model – a phenomenon that is often referred to as the Optimizer's Curse or the Error-Maximization Effect of Optimization. A contemporary approach to combat such estimation errors is offered by distributionally robust problem formulations that consider all data-generating distributions close to the empirical distribution derived from historical samples, where 'closeness' is determined by the Wasserstein distance. While those techniques show significant promise in problems where all input features are continuous, they scale exponentially when binary and/or categorical features are present. This work demonstrates that such mixed-feature problems can indeed be solved in polynomial time. We present a practically efficient algorithm to solve mixed-feature problems, and we compare our method against alternative techniques both theoretically and empirically on standard benchmark instances.
(Authors: Reza Belbasi, Aras Selvi, Wolfram Wiesemann)
Speaker 1 Bio: Renata is a fifth-year Ph.D. candidate in Electrical Engineering at Brazil's Federal University of Santa Catarina (UFSC), where she works under the guidance of Prof. Erlon Finardi. Her research is centered on addressing challenges in power system operation and planning, with a particular emphasis on leveraging innovative algorithms and methodologies. Recently, Renata completed a 10-month collaboration with Prof. Guzin Bayraksan at The Ohio State University, focusing on developing strategies for power system operation in response to the impacts of climate change. Her collaborative efforts extend beyond academia, as she actively collaborates with energy companies to gain insights into the real-world challenges faced by the power system. She currently integrates CEPEL, the Center for Electrical Energy Research in Brazil working with inflow scenario generation.
Speaker 2 Bio: Aras is a doctoral candidate at Imperial College Business School, supervised by Professor Wolfram Wiesemann. He is affiliated with the Computational Optimization Group and the Data Science Institute of Imperial College London and has recently completed PhD research internships at The Alan Turing Institute and J.P. Morgan AI research. His research interests are the theory of data-driven decision making under uncertainty and its applications in machine learning, privacy, and fairness. In his recent works, he has been working on designing optimal privacy mechanisms, developing efficient algorithms for robust machine learning, as well as approximating hard decision making problems via robust optimization.
Posted on 2024-04-23 by Sarah Fores