Machine-Learning Driven Load Shedding to Mitigate Instability Attacks in Power Grids
Justin Tackett, Benjamin Francis, Luis Garcia, David Grimsman, Sean Warnick · 2025
Read Machine-Learning Driven Load Shedding to Mitigate Instability Attacks in Power Grids by Justin Tackett, Benjamin Francis, Luis Garcia, David Grimsman, Sean Warnick (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-26532.txt.
Introduction
Justin Tackett and co-authors defend power grids against attacks that shake them apart rather than switch them off. An instability attack injects wobbling dynamics that can cascade into blackouts. Operators answer by shedding load, cutting chosen customers until the grid calms, but no set rule picks which loads to drop.
Their method learns the choice from data. On the standard IEEE fourteen-bus test grid, run through a commercial analyzer, the system sizes up the grid's state and picks its cuts. A signal method called modified Prony analysis proved able to spot an attack's telltale oscillations and trigger the defense. The proof of concept points toward automated, targeted load shedding as a standing guard.
Catalog record
| CREATOR | Justin Tackett, Benjamin Francis, Luis Garcia, David Grimsman, Sean Warnick |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 33 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2509-26532
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Machine-Learning Driven Load Shedding to Mitigate Instability Attacks in Power Grids Thanks: Information and Decision Algorithms Laboratories (IDeA Labs), Brigham Young University, Provo, Utah University of Utah, Salt Lake City, Utah Corresponding author [email protected], [email protected] . All authors are also affiliated with Achilles Heel Technologies (AHT), and the work was supported through AHT by the US Department of Energy, contract DE-SC0021693. The views expressed here are entirely the authors and do not necessarily reflect any official government position. Justin Tackett, Benjamin Francis, Luis Garcia, David Grimsman, Sean Warnick Abstract Every year critical infrastructure becomes more complex and we grow to rely on it more and more. With this reliance, it becomes an attractive target for cyberattacks from sophisticated actors, with one of the most attractive targets being the power grid. One class of attacks, instability attacks, is a newer type of attack that has relatively few protections developed. We present a cost effective, data-driven approach to training a supervised machine learning model to retrofit load shedding decision systems in power grids with the capacity to defend against instability attacks. We show a proof of concept on the IEEE 14 Bus System using the Achilles Heel Technologies Power Grid Analyzer, and show through an implementation of modified Prony analysis (MPA) that MPA is a viable method for detecting instability attacks and triggering defense mechanisms. I Introduction Throughout the past two hundred years, the power grid has become a core part of the infrastructure of the world. Every modern facility relies on electricity to sustain the way of life that has become prevalent in first world countries, powering everything from life sustaining equipment to financial transaction infrastructure. Recent events like the mass power outage in Spain [ 1 ] , widescale blackouts and brownouts every summer [ 2 ] , and fears of future blackouts [ 3 ] serve as reminders that the power grid is delicate and without it, life as we know it comes apart at the seams. With this potential disaster and its consequences in mind, it is a troubling idea that wide scale power outages could be purposefully brought about by terrorists or nation-state sponsored attackers, and a yet further troubling idea given that attacks like these have already happened at large scales [ 4 ] . Because of this, it is critical to come to understand what kinds of attacks may be around the corner in the ever growing landscape of cyber-warfare targeting critical infrastructure, and that is what this paper sets out to do. One sophisticated class of next generation cyber-attacks is instability attacks. These attacks connect potentially unrelated parts of a system and introduce feedback between them that destabilizes the entire system. Because it affects the entire system and can, in theory, be introduced anywhere in the system, the contact points can be extremely hard to detect 1 1 1 In fact, even the causes of non-malicious power outages can be extremely difficult to establish [ 1 ] . , resulting in stubborn persistence with devastating impact [ 5 , 6 , 7 , 8 ] . Because this class of attacks is so new, few, if any, strategies have been developed to mitigate them. Mitigation is particularly challenging given that the attack surface (the set of potentially exposed contact points) can be extensive and hard to cover. Currently, there are some efforts to try to raise awareness and provide solutions for these kind of attacks, but the general understanding of how to respond in the event of instability attacks is still in its early stages [ 9 ] . This paper presents a novel retrofitting of current remedial action mechanisms, namely load shedding mechanisms, using a cost-effective, data-driven approach to be able to respond to attacks before they do significant damage and interrupt them with 92% overall accuracy. Methods for attack detection, as well as tools for creating training data of the attacks, are presented, which allow operators to train models on their own systems and establish software based logic to be able to respond to this new class of attack. II Related Work The foundational paper for this work is Tackett et al. [ 9 ] , which illustrates how instability attacks can be formed for a power grid, the control theory that supports their existence, and the effects it can have on a system. While similar works and case studies exist, [ 10 , 11 ] , to our knowledge it is the first of its kind describing instability attacks on power systems by introducing modified or artificial feedback into the system to drive it to be small-signal unstable. A much more extensive catalog exists regarding using load shedding with machine learning, the approach taken in this paper. Schweitzer et al. [ 12 ] , in conjunction with their development of modified Prony analysis, use load shedding to combat unstable inter-area oscillations. This is seen in other works as well; Toro et al. implemented automatic load shedding for the development of more robust remedial action schemes [ 13 ] , which gives precedence for using load shedding without an human operator in the loop to combat growing oscillations. Justin and Paternain [ 14 ] implement a reinforcement learning (RL) agent in conjunction with a classifier of system response given a certain input to do adaptive load shedding. This approach is very similar to the work presented here, and integrating the classifier here with a similar type of RL agent could prove useful in broad scale implementation with the hereafter mentioned ’operator-type’ entity. The principal difference between their paper and this one is that we focus on training a classifier on instability attacks (with infrastructure to detect these attacks)—system responses that would almost certainly not be captured in traditional system configurations used in that paper. There are many other things happening in this space, a systematic survey of which was done recently by Skrjanc et al. [ 15 ] . Cyber-physical anomaly and intrusion detection using machine learning is yet another application that has seen considerable work in recent years, both broadly for cyber-physical systems as well as specifically for power grids [ 16 , 17 , 18 , 19 ] . Modified Prony analysis (MPA) is a tool we have repurposed for objectives different than the original intent of inter-area oscillations [ 12 ] , however our application is conceptually close to theirs and so is presented as a secondary or minor contribution, with credit belonging to the Schweitzer Electronic Labs (SEL) for this tool. Others have also found and used this tool in similar areas to those used at SEL [ 13 ] . Fig. 1: (a) Top: An example of naively shedding according to under voltage load shedding schemes under an instability attack. y here is the voltage of the third bus in the IEEE 14 Bus System. The tenth load is shed at 10 seconds, and the attack worsens relative to what it was before. (b) Bottom: An example of a stable load shed with the same system initial conditions and topology as (a) under a different attack, shedding the same load, showing the same value in the y axis as in (a). The behavior of (a) and (b)