City Research Online

Abstract

The introduction of new technologies and increased penetration of renewable resources is altering the power distribution landscape which now includes a larger numbers of micro-generators. The centralized strategies currently employed for performing frequency control in a cost efficient way need to be revisited and decentralized to conform with the increase of distributed generation in the grid. In this paper, the use of Multi-Agent and Multi-Objective Reinforcement Learning techniques to train models to perform cost efficient frequency control through decentralized decision making is proposed. More specifically, we cast the frequency control problem as a Markov Decision Process and propose the use of reward composition and action composition multi-objective techniques and compare the results between the two. Reward composition is achieved by increasing the dimensionality of the reward function, while action composition is achieved through linear combination of actions produced by multiple single objective models. The proposed framework is validated through comparing the observed dynamics with the acceptable limits enforced in the industry and the cost optimal setups.

Publication Type:	Conference or Workshop Item (Paper)
Publisher Keywords:	Multi-Agent Reinforcement Learning, Multi-Objective Reinforcement Learning, Frequency Control, Economic Dispatch, Deep Deterministic Policy Gradient
Subjects:	G Geography. Anthropology. Recreation > GE Environmental Sciences T Technology > TK Electrical engineering. Electronics Nuclear engineering
Departments:	School of Science & Technology > Engineering

Export

Downloads