Machine-Learning Driven Load Shedding to Mitigate Instability Attacks in Power Grids

TL;DR

Instability attacks threaten power-grid stability by injecting destabilizing dynamics. The authors retrofit load shedding with a data-driven ML classifier, triggered by an MPA-based alarm, to predict which loads to shed to stabilize the system, demonstrated on the IEEE 14 Bus System using the AHT Power Grid Analyzer. The work contributes a full pipeline—from data generation and labeling to a time-series and load-index encoder feeding a 3-layer MLP—that achieves a $F1$-score of $0.92$, showing that informed, system-aware shedding outperforms naive approaches. It offers a software-based defense with practical deployment potential over existing safety hardware, while highlighting limitations related to data latency, integrity, and the complexities of scalable, decentralized deployment for future robustness.

Abstract

Critical infrastructures are becoming increasingly complex as our society becomes increasingly dependent on them. This complexity opens the door to new possibilities for attacks and a need for new defense strategies. Our work focuses on instability attacks on the power grid, wherein an attacker causes cascading outages by introducing unstable dynamics into the system. When stress is place on the power grid, a standard mitigation approach is load-shedding: the system operator chooses a set of loads to shut off until the situation is resolved. While this technique is standard, there is no systematic approach to choosing which loads will stop an instability attack. This paper addresses this problem using a data-driven methodology for load shedding decisions. 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.

Figures (4)

• Figure 1: An example of naively shedding according to under voltage load shedding schemes under an instability attack. On the left is the IEEE 14 Bus system, with generators as smaller red circles and buses as the white nodes. Two attacks are shown; a red $\Delta$ and a yellow $\Delta$, where each $\Delta$ reads from a state in a given node (Bus 5 and Generator 3 for yellow and red $\Delta$'s, respectively) and both write to the same state in Bus 9; these are all highlighted accordingly. These attacks happen separately, with the same response: shedding the load on Bus 6. The resulting effect on the system, shown by the voltage for Bus 3 (in normalized units), is shown in the two graphs to the right. In the instance of the red $\Delta$, the attack gets worse, and in the instance of the yellow $\Delta$, the attack is interrupted, thereby showing that using pre-defined load-shedding schemes can yield stable or unstable outcomes, and a system-aware decision process is needed.
• Figure 2: The block diagram for the classifier used to determine if a load shed results in stable conditions or not. The classifier is made up of principally three parts: a time-series encoder, a load-index encoder, and a fully connected three layer MLP for classification. The output is passed through soft-max logits, which provide probabilities of classification as stable or unstable. For the IEEE 14 Bus System, this results in about 600,000 weights, relatively light-weight and able to be trained on a laptop in several hours.
• Figure 3: A distribution of stable sheds for all possible read/ write attack configurations, per load. For example, load 2 here represents how many times in the 12,808 samples that picking load 2 resulted in stability. This graph shows that, barring some small variance, there isn't one load that picking every time would result in stable outcomes, showing the need for a data-driven approach over a priori load shedding schemes.
• Figure 4: A comparison of the distributions of stable shed cases across different variable categories. These are taken from the 12,808 viable data samples used for training, testing and validation. They are shown to show that each read/ write combination is represented in the stable case training sets, and also showing the relative number of cases present. For a given bar, say reading Voltage for bus 3 (V 3), all possible write combinations and all load shedding options were summed together to a bar, where the colored bar is the count of stable load sheds and the grey bar is the total amount of samples for the given category, providing some insight into sample distribution. This is likewise for the right concerning write variables. The labels here represent different physical variables such as frequency $\omega$, rotor angle $\delta$, voltage $V$, load angle $\theta$, reactive and active loads and power generation $PL,QL,PG,QG$, and so on for generators/ loads $i$ for the IEEE 14 Bus System.