Reinforcement Learning for Design and Operation of Distribution Networks

Graduate Thesis Or Dissertation

Citations

Citeable URL: https://scholar.colorado.edu/concern/graduate_thesis_or_dissertations/vd66w177f

Abstract

Recent changes to the distribution electrical network such as increased consumer demand and var- ious renewable resources have made the need to replace and upgrade electrical distribution equip- ment more dire than ever. This thesis introduces various frameworks for model-free, data-driven reinforcement learning to optimize energy system control across electrical and thermal energy do- mains. We consider control at various points of the distribution network including aggregator price setting, system operator generator control, and variations of demand response that address grid services beyond peak shaving used in conventional demand response programs (such as voltage regulation). In this thesis, we seek to address the primary concern of the combination of ther- mal and electrical models, ensuring the feasibility of these results for both the customer and the electric utility. Each control study that includes load shifting validates that the thermal comfort requirements of the end-user are satisfied and does so without adding extraneous hardware. Finally, we introduce a study that considers the use of reinforcement learning algorithms in distribution network design by placing distributed electrical resources at optimal points to support grid-level goals. Preliminary results indicate that reinforcement learning can be valuable for improving grid operation.

Creator