Learning-Based Detection of Malicious Volt-VAr Control Parameters in Smart Inverters

TL;DR

This paper introduces a machine learning-based intrusion detection system that verifies the legitimacy of Volt-Var Control parameters in smart inverters to prevent malicious manipulation and maintain distribution system stability.

Contribution

It presents a novel MLP-based method that detects malicious VVC curves using only local measurements, enhancing cybersecurity in smart inverter control.

Findings

High detection accuracy demonstrated in simulations

Effective prevention of voltage oscillations caused by malicious curves

Applicable to real-world distribution systems

Abstract

Distributed Volt-Var Control (VVC) is a widely used control mode of smart inverters. However, necessary VVC curve parameters are remotely communicated to the smart inverter, which opens doors for cyberattacks. If the VVC curves of an inverter are maliciously manipulated, the attacked inverter's reactive power injection will oscillate, causing undesirable voltage oscillations to manifest in the distribution system, which, in turn, threatens the system's stability. In contrast with previous works that proposed methods to mitigate the oscillations after they are already present in the system, this paper presents an intrusion detection method to detect malicious VVC curves once they are communicated to the inverter. The proposed method utilizes a Multi-Layer Perceptron (MLP) that is trained on features extracted from only the local measurements of the inverter. After a smart inverter is equipped with the proposed method, any communicated VVC curve will be verified by the MLP once received. If the curve is found to be malicious, it will be rejected, thus preventing unwanted oscillations beforehand. Otherwise, legitimate curves will be permitted. The performance of the proposed scheme is verified using the 9-bus Canadian urban benchmark distribution system simulated in PSCAD/EMTDC environment. Our results show that the proposed solution can accurately detect malicious VVC curves.