Dual moving average trick spots microgrid islanding in milliseconds

Engineers have unveiled a lightweight statistical algorithm that can tell when a fragment of the power grid has slipped into unintentional islanding in as little as seven milliseconds, using nothing more than voltage measurements and basic arithmetic. The method, developed by Elijah Gbenga Osunkentan, Ricardo Caneloi dos Santos and Adedeji Adeyemi and published in Discover Electronics, relies on a dual exponential moving average framework that watches how the voltage at the point of common coupling behaves statistically rather than how far it drifts from a fixed threshold. Because it needs only elementary operations, the researchers argue it can run comfortably on low-cost embedded industrial controllers, avoiding the expensive signal-processing hardware that many modern detection schemes demand.

Unintentional islanding occurs when a section of the distribution network becomes electrically separated from the main utility grid yet continues to be energized by local distributed generators, for example after a fault. The condition is more than a technical curiosity. Islanding can compromise personnel safety, cause out-of-phase reclosing, and drive voltage and frequency variations that damage connected equipment. IEEE Standard 1547-2018 therefore requires distributed energy resources to detect islanding and cease energization within two seconds. In microgrids rich with inverter-based resources such as photovoltaic arrays and battery energy storage systems, the task is harder still, because these systems have reduced inertia and limited fault current contribution compared with conventional synchronous generators.

The central difficulty is the non-detection zone. When generated power closely matches local load demand, the voltage and frequency at the point of common coupling barely move after disconnection, leaving conventional passive methods with too little signal to work with. At the same time, any detector must distinguish genuine islanding from everyday disturbances such as load switching, capacitor switching and distributed energy resource transitions without tripping falsely. Existing alternatives each carry drawbacks: active and hybrid techniques add coordination complexity and power quality concerns, while many signal-processing and machine-learning approaches demand computationally intensive algorithms or extensive offline training that rules out cheap, real-time embedded deployment.

The new algorithm attacks the problem by monitoring normalized voltage dynamics through two parallel exponential moving averages. A slow EMA, with a smoothing coefficient of 0.05, estimates the long-term voltage baseline, changing gradually and remaining relatively insensitive to transient disturbances. A fast EMA, with a coefficient of 0.10, tracks the short-term running mean and variance of the normalized voltage signal. The raw point-of-common-coupling voltage is first converted to a root mean square magnitude over a sliding window of 64 samples, roughly eight milliseconds at 60 Hz, sampled at 8 kHz. Dividing the RMS voltage by the slow baseline normalizes the signal, stripping away slow operating-point variations so the analysis focuses on relative deviations instead of absolute magnitude.

The key insight is that islanding changes the character of voltage fluctuations, not just their size. Under grid-connected operation the point of common coupling exhibits small statistical fluctuations, but once the utility grid’s support is lost, transient voltage oscillations appear and the estimated variance rises noticeably. Even under near power-balanced conditions where absolute voltage deviations stay small, the loss of grid stiffness increases short-term variance relative to the long-term baseline. The fast EMA responds to this variance increase within a few sampling intervals while the slow EMA remains anchored to the pre-islanding baseline, producing a detectable divergence in the detection ratio even when conventional threshold-based methods would see nothing to trip on.

An adaptive threshold built from the baseline plus a scaled variance term generates a normalized detection ratio, which is compared against decision limits of 0.98 and 1.02. If the ratio crosses either boundary, islanding is declared. The researchers selected these parameters through a systematic simulation-based sensitivity analysis. Raising the fast EMA coefficient shortened detection but invited false trips, while very small values caused missed detections through excessive smoothing; the chosen value of 0.10 achieved zero false trips, zero missed detections and a mean detection time of 17.2 milliseconds. The threshold pair of 0.98 and 1.02 was the only configuration tested that simultaneously delivered zero false trips, zero missed detections and an acceptable mean detection time across all evaluated scenarios. Narrower bands detected faster but tripped falsely, while wider bands reduced false trips at the cost of longer detection times and missed detections.

Validation proceeded in two stages. First, software simulations on an RTDS-based microgrid model comprising a photovoltaic array, a battery energy storage system and a synchronous diesel generator operating at 480 V and 60 Hz. Simulation islanding detection times ranged from roughly seven milliseconds under large power mismatch, where pronounced voltage deviations appear immediately after disconnection, to eighteen milliseconds under near power-balanced conditions, the most challenging case where the detection ratio needs longer to exceed the adaptive threshold. Small power mismatch conditions fell in between at approximately twelve milliseconds. Critically, all non-islanding disturbances, including load switching, capacitor switching and DER switching events, were correctly classified, and noise levels of 50 dB and 60 dB signal-to-noise ratio produced no incorrect islanding declarations.

The method was then implemented on an NI CompactRIO embedded controller with a 266 MHz real-time processor and FPGA-based input-output chassis, running a deterministic timed loop at 8 kHz in LabVIEW Real-Time. Measured CPU utilization stayed below five percent throughout all experimental scenarios. In the hardware-in-the-loop tests, the RTDS executed the electromagnetic transient simulation of the microgrid with a 125-microsecond time step and fed point-of-common-coupling voltage signals to the controller through analog interface modules, with a measured round-trip communication latency of approximately 250 microseconds included in all reported detection times. Hardware-in-the-loop detection times ranged from eight to twenty milliseconds, slightly longer than simulation due to communication latency and finite-precision effects. Near power-balanced islanding was detected in twenty milliseconds, small mismatch cases averaged thirteen milliseconds, and large mismatch cases averaged ten milliseconds, all comfortably within the IEEE 1547-2018 two-second clearing requirement.

Selectivity held up in practice. During hardware-in-the-loop testing the method differentiated between normal grid disturbances and islanding events in every case, maintaining high discrimination capability regardless of operating conditions, with no false trips observed under load switching, capacitor switching or DER switching during either simulation or hardware testing. The one weakness appeared in computational simulations under severe noise: at 40 dB signal-to-noise ratio, all four tested non-islanding scenarios produced incorrect islanding declarations, as noise-induced fluctuations distorted the extracted statistical features and pushed the detection ratio past the decision boundaries. The authors note such noise levels are significantly more severe than typically encountered in practical measurement systems, but the result shows performance degrades under highly noisy conditions, and they suggest additional filtering, threshold adaptation or a confirmation-based logic requiring several consecutive out-of-bounds samples as future remedies.

The researchers are candid about the limits of the validation. Testing used a single RTDS-based grid-following inverter microgrid configuration, and the reported parameter values are specific to the 480 V, 60 Hz test system, requiring re-tuning for different configurations. Generalization to other topologies, higher DER penetration levels and grid-forming inverter scenarios remains future work, as does characterization under distorted voltage conditions targeting total harmonic distortion levels of five to eight percent. A direct experimental benchmark against conventional passive methods under identical conditions was beyond the study’s scope, and the literature comparison in the paper is explicitly illustrative rather than a strict benchmark since the referenced methods were evaluated on different test systems. Even so, the combination of millisecond-scale detection, zero false trips under realistic disturbances, and a computational footprint light enough for a low-cost controller makes a compelling case that careful statistics, not heavy computation, may be the key to safer microgrids.

Subject of Research: A statistical voltage-based passive method for detecting unintentional islanding in microgrids

Article Title: An efficient and cost-effective statistical voltage-based method for passive islanding detection in microgrids

Article References: Osunkentan, E. G., dos Santos, R. C., & Adeyemi, A. (2026). An efficient and cost-effective statistical voltage-based method for passive islanding detection in microgrids. Discover Electronics, 3(1), Article 132. https://doi.org/10.1007/s44291-026-00284-x

Image Credits: AI Generated

DOI: 10.1007/s44291-026-00284-x

Keywords: islanding detection, microgrids, distributed energy resources, exponential moving average, voltage monitoring, hardware-in-the-loop, IEEE 1547-2018, non-detection zone, embedded controllers, power system protection, efficient, cost-effective