Deep Reinforcement Learning For Online Distribution Power System Cybersecurity Protection

Bailey, T.J.; Johnson, Jay; Levin, Drew

doi:10.1109/smartgridcomm51999.2021.9631991

Cited by 4 publications

(1 citation statement)

References 3 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…A deep reinforcement learning technique was structured to provide cybersecurity protection on distributed power systems. The performance of the system was experimented on the IEEE 13-bus model and the simulation results are not found satisfactory under the greedy attack conditions [28].…”

Section: Related Studiesmentioning

confidence: 99%

Cyber Security in Power Systems Using Meta-Heuristic and Deep Learning Algorithms

2023

View full text Add to dashboard Cite

Supervisory Control and Data Acquisition system linked to Intelligent Electronic Devices over a communication network keeps an eye on smart grids' performance and safety. The lack of algorithms protecting the power system communication protocols makes them vulnerable to cyberattacks, which can result in a hacker introducing false data into the operational network. This can result in delayed attack detection, which might harm the infrastructure, cause financial loss, or even result in fatalities. Similarly, attackers may be able to feed the system with fake information to hoax the operator and the algorithm into making bad decisions at crucial moments. This paper attempts to identify and classify such cyber-attacks by using numerous deep learning algorithms and optimizing the data features with a metaheuristic algorithm. We proposed a Restricted Boltzmann Machine-based nature-inspired artificial root foraging optimization algorithm. Using a publicly available dataset produced in Mississippi State University's Oak Ridge National Laboratory, simulations are run on the Jupiter Notebook. Traditional supervised machine learning algorithms like Artificial Neural Networks, Convolutional Neural Networks, and Support Vector Machines are measured with the proposed algorithm to demonstrate the effectiveness of the algorithms. Simulations show that the proposed algorithm produced superior results, with an accuracy of 97.8% for binary classification, 95.6% for three-class classification, and 94.3% for multi-class classification. Thereby outperforming its counterpart algorithms in terms of accuracy, precision, recall, and f1 score.

show abstract

Section: Related Studiesmentioning

confidence: 99%