Power system security using reinforcement learning

Sunyer Nestares, José María

Permalink

https://hdl.handle.net/2142/107777 Copy

Description

Title

Power system security using reinforcement learning

Author(s)

Sunyer Nestares, José María

Contributor(s)

Zhang, Richard Y.

Issue Date

2020-05

Keyword(s)

• power system security
• cascading failure
• reinforcement learning
• dynamic programming
• security analysis
• N-k criterion

Abstract

The challenge of controlling and guaranteeing the security of the power system is constantly evolving, particularly in light of significant predicted growth in the deployment of renewable energy, and increasing use of electric vehicles. Existing power systems are made secure, in part, using the N-1 criterion, in which the system is required to remain within operational limits with the loss of any individual component. This does not assess the risk of cascading failure, which is likely to become more commonplace with the large-scale, distributed integration of small, stochastic components, such as renewable generators or electric vehicles. In this thesis, we describe an AI-based advisory tool to verify the N − k criterion over transmission lines, meaning that the system is required to be secure with the disconnection of k out of N total transmission lines. The tool is designed to identify cascading mechanisms, in which the disconnection of one line overloads another, thereby resulting in a sequence of disconnections downstream. Our key insight is to formulate this cascading problem as an instance of the shortest path, a classic problem in dynamic programming and reinforcement learning with a number of standard solutions. The assumptions of the formulation are validated on small-scale IEEE test cases using exhaustive search. Finally, we investigate simulation-based techniques from the reinforcement learning literature, as a means of overcoming the curse of dimensionality for large-scale, real-world power systems.

Type of Resource

text

Language

en

Permalink

http://hdl.handle.net/2142/107777

Owning Collections