Physics-informed learning detects islanding and discriminates power-quality disturbances in grid-forming inverter-dominated distribution networks

The proliferation of grid-forming (GFM) inverters in modern distribution networks has fundamentally altered the islanding-detection problem. Unlike grid-following converters, droop- and virtual-synchronous-generator (VSG)-controlled GFM units actively regulate voltage and frequency, suppressing the very deviations that conventional passive schemes rely upon and thereby enlarging the non-detection zone (NDZ) under near-balanced power conditions. This paper proposes a Physics-Informed Lightning-Search-Optimised Temporal Attention Network (PI-LSO-TAN) that jointly detects unintentional islanding and discriminates it from power-quality disturbances (PQDs) and non-islanding faults in GFM-dominated feeders. The framework fuses a physics-informed feature set rate-of-change-of-frequency (ROCOF) estimated from the analytic signal, total harmonic distortion, phase-jump and current-imbalance descriptors with a raw-window convolutional embedding processed by a self-attention pooling layer. A Lightning Search Algorithm (LSA) simultaneously optimises the network hyperparameters and a physics-informed decision threshold, removing manual tuning. Across 10 independent dataset realisations of a documented, fully reproducible synthetic GFM feeder operated in the near-balanced NDZ regime, PI-LSO-TAN attains a four-class test accuracy of 96.34% (SD 0.96 percentage points), a macro-F1 of 0.9636 and an islanding recall of 93.49% at an islanding false-positive rate of 0.70%, misdeclaring 2.40 of 175 power-quality windows as islanding. Two findings define what the method does and does not contribute. First, when ten baselines spanning kernel, ensemble, convolutional, recurrent, attention-based and graph architectures are given the identical physics-informed side-channel, all 11 models fall within 1.20 percentage points of one another; none of the 6 deep sequence architectures differs significantly from the proposed model under Holm-corrected Wilcoxon tests, and of the ten baselines only 1 differ significantly at all MLP-SVM hybrid, ahead by up to 0.73 percentage points these being among the simplest feature-space classifiers in the comparison. On this task the discriminative power resides in the physics-informed representation rather than in the temporal-modelling mechanism. Substituting an alternative fourteen-dimensional representation confirms this directly principal components of the raw window reach 52.37%, wavelet band energies 73.60%, generic statistics 62.94%, and a purely learned embedding 30.17%, against 96.20% for the physics-informed set. Second, noise resilience is a property of the training distribution, not of the architecture: training at a single SNR leaves accuracy at 52.18% when the measurement noise floor moves to 30dB, whereas sampling the training SNR across 15 dB to 50 dB restores 88.79% at 30dB and 84.54% at 20dB for a cost of − 6.14 percentage points at the nominal operating point. The practical contribution is protective rather than architectural. A security-calibrated passive over/under-frequency, ROCOF and over/under-voltage relay detects none of the 3872 scanned near-balanced islanding events, leaving a 100.0% non-detection area, whereas the proposed detector reduces it to 0.00%, with no cell of the power-mismatch plane falling below 87.5% detection. Measured sliding-window detection latency is 80 ms on average and 130 ms at the 95th percentile, inside the 2 s envelope of IEEE 1547-2018 and IEC 62116, at 337 μsof computation per window and 28.7 kB of parameters. The results indicate that reliable anti-islanding protection in inverter-dominated grids is obtained by pairing a physics-informed representation with noise-augmented training, and that the resulting detector covers a region of the operating plane in which conventional passive protection is blind by construction.