EURO 2024 Copenhagen
Abstract Submission

EURO-Online login

4255. Decision-focused learning with machine learning proxies for energy storage system optimization in energy markets

Invited abstract in session TB-3: Machine Learning in Applied Optimization, stream Data Science Meets Optimization.

Tuesday, 10:30-12:00
Room: 1005 (building: 202)

Authors (first author is the speaker)

1. Ruben Smets
2. Mathieu Tanneau
Polytechnique Montreal
3. Jean-François Toubeau
University of Mons
4. Kenneth Bruninx
TU Delft
5. Pascal Van Hentenryck
Georgia Institute of Technology
6. Erik Delarue
Applied Mechanics and Energy Conversion Section, KU Leuven

Abstract

In various sectors like energy, finance, and supply chain management, decision-makers optimize their decisions using forecast-informed optimization. In particular, energy market participants base their decisions on price forecasts. In this context, Decision-Focused Learning (DFL) has emerged, where the downstream decision problem is explicitly considered in the training procedure of the forecaster. This forecasting model is typically a neural network, which is trained using a gradient descent-based procedure. A pivotal challenge in DFL training is ensuring that the optimal decisions are differentiable with respect to the input parameters, being the price forecasts. A promising solution involves implicit differentiation of KKT conditions, albeit this significantly slows down training since it requires solving many optimization problems with every forward pass during training. We propose an innovative workaround using a Machine Learning (ML) proxy to replace the optimization problem in the forward pass. This approach is applied to Energy Storage Systems (ESS) participating in day-ahead and real-time balancing markets. Applying duality theory to the ESS optimization program allows us to simplify the ML proxy model, limiting its output to the dual variable of the energy balance constraint. We observe similar out-of-sample profit performance as traditional DFL methods, while reducing training time by 70% to 95%, illustrating the improved scalability of the proposed method.

Keywords

Status: accepted

Back to the list of papers