|
" Reinforcement Learning and Game Theory for Smart Grid Security "
Paul, Shuva
Ni, Zhen
| Document Type
|
:
|
Latin Dissertation
|
| Language of Document
|
:
|
English
|
| Record Number
|
:
|
1105921
|
| Doc. No
|
:
|
TLpq2348815481
|
| Main Entry
|
:
|
Ni, Zhen
|
|
|
:
|
Paul, Shuva
|
| Title & Author
|
:
|
Reinforcement Learning and Game Theory for Smart Grid Security\ Paul, ShuvaNi, Zhen
|
| College
|
:
|
South Dakota State University
|
| Date
|
:
|
2019
|
| student score
|
:
|
2019
|
| Degree
|
:
|
Ph.D.
|
| Page No
|
:
|
206
|
| Abstract
|
:
|
This dissertation focuses on one of the most critical and complicated challenges facing electric power transmission and distribution systems which is their vulnerability against failure and attacks. Large scale power outages in Australia (2016), Ukraine (2015), India (2013), Nigeria (2018), and the United States (2011, 2003) have demonstrated the vulnerability of power grids to cyber and physical attacks and failures. These incidents clearly indicate the necessity of extensive research efforts to protect the power system from external intrusion and to reduce the damages from post-attack effects. We analyze the vulnerability of smart power grids to cyber and physical attacks and failures, design different game-theoretic approaches to identify the critical components vulnerable to attack and propose their associated defense strategy, and utilizes machine learning techniques to solve the game-theoretic problems in adversarial and collaborative adversarial power grid environment. Our contributions can be divided into three major parts: Vulnerability identification: Power grid outages have disastrous impacts on almost every aspect of modern life. Despite their inevitability, the effects of failures on power grids' performance can be limited if the system operator can predict and identify the vulnerable elements of power grids. To enable these capabilities we study machine learning algorithms to identify critical power system elements adopting a cascaded failure simulator as a threat and attack model. We use generation loss, time to reach a certain percentage of line outage/generation loss, number of line outages, etc. as evaluation metrics to evaluate the consequences of threat and attacks on the smart power grid. Adversarial gaming in power system: With the advancement of the technologies, the smart attackers are deploying different techniques to supersede the existing protection scheme. In order to defend the power grid from these smart attackers, we introduce an adversarial gaming environment using machine learning techniques which is capable of replicating the complex interaction between the attacker and the power system operators. The numerical results show that a learned defender successfully narrows down the attackers' attack window and reduce damages. The results also show that considering some crucial factors, the players can independently execute actions without detailed information about each other. Deep learning for adversarial gaming: The learning and gaming techniques to identify vulnerable components in the power grid become computationally expensive for large scale power systems. The power system operator needs to have the advanced skills to deal with the large dimensionality of the problem. In order to aid the power system operator in finding and analyzing vulnerability for large scale power systems, we study a deep learning technique for adversary game which is capable of dealing with high dimensional power system state space with less computational time and increased computational efficiency. Overall, the results provided in this dissertation advance power grids' resilience and security by providing a better understanding of the systems' vulnerability and by developing efficient algorithms to identify vulnerable components and appropriate defensive strategies to reduce the damages of the attack.
|
| Subject
|
:
|
Artificial intelligence
|
|
|
:
|
Electrical engineering
|
| |