Enforcing Constraints in Learning-Augmented Online Optimization: Theory and Applications to Energy Systems

Author: Chen, James Yuxuan

Year: 2024

Degree: Senior thesis (Major)

Advisor: Wierman, Adam C.

Committee Member: None, None

Option: Electrical Engineering

DOI: 10.7907/fmhp-dg57

Abstract

Increasing renewable penetration into the power grid is critical for combating climate change. To implement this successfully, it is crucial to design real-time dispatch algorithms that are robust to the uncertainty that renewable sources present. It has proven difficult to produce effective large-scale dispatches on the fly using traditional methods; as such, this has motivated research into incorporating modern machine learning (ML) methods into economic dispatch. In order for ML-based dispatch algorithms to be effectively deployed, they must have the level of performance guarantees necessary for a safety-critical setting like the grid, and also be able to enforce strict operational constraints. In the first part of this work, we consider the problem of designing learning-augmented algorithms for online optimization in the presence of ramp and feasibility constraints, and provide some of the first results in this space to our knowledge. We use these insights to develop learning-augmented algorithms that adhere to these constraints, and demonstrate how they can effectively balance between algorithm performance and the potential for constraint violations. In the second part of this work, we consider the complementary problem of training an ML model to perform economic dispatch in the face of complex operational constraints. In particular, we utilize a plant model and historical data from a real-world co-generation plant, and develop methods to enforce constraints in our ML model. Our results demonstrate that ML models can simultaneously achieve good performance and minimize constraint violations in a real-world dispatch setting.

Files