Machine Learning Shields Microgrid Relays From Four Classes of Cyber-Attack

Protective relays are the sentinels of modern power grids, and in microgrids their job is especially delicate. Directional overcurrent relays, or DOCRs, decide whether to trip a circuit breaker by weighing two things at once: how large the fault current is, and which direction it flows. Both judgments rest on current and voltage measurements delivered by instrument transformers at the relay’s own location. Researchers publishing in Results in Engineering now argue that this very dependence creates an underappreciated attack surface, because an adversary who can distort those local measurements or the relay’s own settings can trick the relay into tripping on perfectly healthy conditions. Their response is a machine learning framework, built around a support vector machine, that screens every trip decision and blocks the relay when the incoming data pattern looks manufactured rather than physical.

Legitimate faults form one class; cyber-attacks that mimic faults form the other. What distinguishes their threat model from much of the prior literature is its breadth. Many earlier detection schemes for differential and multi-agent protection assume that locally acquired measurements, instrument-transformer parameters, and relay settings are inherently trustworthy, concentrating instead on attacks against remotely communicated signals. The new work deliberately drops that assumption and considers four attack categories that strike different links in the relay’s measurement and decision chain: false data injection attacks that manipulate measured current and voltage magnitudes, time-synchronization attacks that distort the angular relationship between operating and polarizing quantities, transformation-ratio attacks that alter the effective current transformer ratio, and relay-setting attacks that tamper with the pickup current.

Each attack is modeled with the constraints that would make it genuinely effective against a real relay. A false data injection attack, for instance, must push the measured phase voltage below 0.9 per unit while raising the current above the pickup threshold but below the maximum fault current the relay could ever see, so the manipulated data satisfies the relay’s own tripping logic. Time-synchronization attacks shift current and voltage phase angles so that the directional unit reads a reverse fault as forward, or a forward fault as reverse; the paper shows an example where an eight-millisecond shift moves the apparent phase angle from 38.8 degrees to 211.7 degrees, while stressing that the required shift depends on the specific voltage-current relationship. Transformation-ratio attacks reduce the current transformer ratio by a calculated modification factor so that normal load current appears large enough to trip, bounded so the transformer neither saturates at load nor leaves the relay unresponsive. Relay-setting attacks lower the pickup current so ordinary load current is interpreted as a fault.

To train the classifier, the team simulated a modified radial IEEE 33-bus microgrid containing four synchronous distributed energy resources and eleven DOCRs, using DigSILENT and MATLAB. Detailed analysis focused on relay R1, the primary relay for faults between buses 2 and 3, because a false trip of R1 would island the entire microgrid by disconnecting all four DERs from the main grid. The training corpus comprised 660 distinct fault cases on line 2-3, spanning six fault locations from 15 to 90 percent of the line length, eleven fault resistances from one milliohm to fifty ohms, and ten fault types covering all single-line-to-ground, line-to-line, line-to-line-to-ground, and three-line-to-ground combinations. Against these, the researchers generated 800 attack scenarios from normal operating conditions: 350 false data injection cases, 350 time-synchronization cases, 50 transformation-ratio cases, and 50 relay-setting cases.

From each event, 24 features were extracted at the relay location: the magnitudes and angles of positive, negative, and zero sequence voltages and currents, plus the same quantities for the three phase voltages and currents. Their relative importance was ranked with the minimum-redundancy maximum-relevance algorithm. To make the classifier robust to imperfect sensing, three Gaussian noise levels of 10, 20, and 30 percent were folded into the data, and both noisy and noiseless measurements were used during training. The final dataset contained 5,840 samples with 24 features each, split 70 percent for training and 30 percent held out for testing. Hyperparameters were tuned with Bayesian optimization on the training subset alone, using two-fold cross-validation, with the Gaussian kernel, a kernel scale of 563.807, a box constraint of 16.1794, and data standardization emerging as the winning configuration.

On the unseen test set, the trained SVM achieved 99.4 percent overall accuracy. Class-specific results were balanced rather than lopsided: faults were classified with 99.5 percent precision and 99.1 percent recall, while attacks reached 99.3 percent precision and 99.6 percent recall, with F1-scores of 0.993 and 0.994 respectively. Because the attack categories were unevenly represented, the authors also examined detection at the subtype level. Single-phase false data injection, transformation-ratio, and relay-setting attacks were all detected at 100 percent in the test set, double-phase false data injection at 99.44 percent, three-phase false data injection at 98.25 percent, and time-synchronization attacks at 99.52 percent, indicating the overall metrics were not masking weak performance on the rarer attack types.

Speed matters in protection, and here the numbers are striking but need careful reading. The trained SVM required roughly nine microseconds to classify a feature vector in the MATLAB environment, but that figure covers classifier inference only. When measurement acquisition, sampling, feature extraction, memory access, and inference were all included, the total processing time came to about six milliseconds, which the authors note stays within the typical operating range of digital protective relays and below a 1.5-cycle operating interval. Training took approximately 43 seconds, the shortest among the classifiers benchmarked. In head-to-head comparisons on the same dataset against random forest, Naive Bayes, decision tree, k-nearest neighbors, and neural network models, the SVM’s accuracy, precision, recall, and F1-score all exceeded 99.1 percent.

The team also benchmarked against the PCA-Isolation Forest anomaly detection approach used in recent cyber-resilient protection schemes, retrained on the identical dataset. Both methods reached 100 percent classification accuracy on the considered cases, but the PCA-IF model needed about 1.6 times the SVM’s training time and about 11.5 milliseconds for total classification, roughly twice the SVM’s six milliseconds. An additional evaluation on twenty unseen fault and attack cases, generated with parameter combinations absent from model development and without any retraining, provided further evidence that the classifier generalizes to new fault types, locations, resistances, and unseen manipulation levels across all four attack categories.

The authors are explicit about the boundaries of these results. Everything reported is simulation-based, on a single radial test system with synchronous DERs, with detailed analysis confined to one relay; transfer to other network structures, inverter-dominated systems, meshed microgrids, or different relay locations requires independent validation. Current transformer saturation was assumed negligible, and the compensation technique proposed to mitigate it was not implemented or evaluated, so the results do not constitute validation under saturated conditions. Harmonic distortion, quantization effects, and other sensor non-idealities were not modeled, communication packet loss and delays fall outside the input architecture, and the threat model covers only false-tripping scenarios, excluding attacks coinciding with genuine faults and adversarial machine-learning attacks. No hardware-in-the-loop testing, real-time digital simulation, or commercial relay implementation was performed, so the timing figures remain simulation-based. Future work, the authors say, will extend validation to broader network conditions, additional relays and attack types, explicit CT saturation, and embedded-platform testing to establish end-to-end real-time feasibility.

Subject of Research: An SVM-based cyber-attack detection framework for directional overcurrent relay protection in microgrids

Article Title: A machine learning framework for cyber-resilient directional overcurrent protection in microgrids

Article References: Sheta, A. N., Elgamal, M. M., Eladl, A. A., Bureš, V., & Sedhom, B. E. (2026). A machine learning framework for cyber-resilient directional overcurrent protection in microgrids. Results in Engineering, 32, Article 113735. https://doi.org/10.1016/j.rineng.2026.113735

Image Credits: AI Generated

DOI: 10.1016/j.rineng.2026.113735

Keywords: microgrids, directional overcurrent relays, cyber-attacks, machine learning, support vector machine, false data injection, time-synchronization attacks, CT transformation ratio, relay settings, IEEE 33-bus, protection coordination, simulation