Machine Learning Based Multi-Agent System for Detecting and Neutralizing Unseen Cyber-Attacks in AGC and HVDC Systems

This paper presents a multi-agent based approach to secure the automatic generation control (AGC) loop and the power modulation controller of a high voltage direct current system against cyber-attacks. To achieve this, we propose to embed a multi-agent system (MAS) to the existing AGC layer comprising a Master Agent (MA) and several Slave Agents (SAs). The MA is equipped with a one-class classifier (OCC) to scrutinize the wide-area signals for intrusion detection. For the SAs, Support Vector Regression (SVR) models are utilized that exploit local information for attack mitigation. In the proposed approach, a new algorithm for training the OCCs is adopted, which uses only
secured
data, without utilizing the
attacked
instances. Next, we present a framework for setting up the SVR models for online operation in the context of the AGC system. For performance evaluation, various test cases corresponding to the
secured case
, which includes nominal and stressed system operation, and
attacked case
, which includes multiple types of data integrity and availability attacks, grid states are considered. It is observed that the OCC, namely, Support Vector Data Description model performs reasonably well in identifying the attacks. Moreover, the performance of the SVR models during attack neutralization is found quite efficient and acceptable.