AI Learns to Untangle Overlapping Power Grid Disturbances with Confidence Correction

Modern electricity grids are noisier than ever. As solar panels, wind farms, power electronics, and nonlinear loads multiply across distribution networks, voltage waveforms are increasingly distorted by not just one problem at a time but by several disturbances overlapping in the same signal. A voltage sag may ride on top of persistent harmonics while a switching transient flashes through, creating composite power quality disturbances whose features blur into one another. Researchers in China now report a deep learning method that can identify these tangled signals with remarkable accuracy, even when the model encounters disturbance combinations it has never seen during training. The work, published in Frontiers in Energy Research, combines a fully convolutional network, a gated recurrent unit, and an attention mechanism with a novel confidence correction strategy.

The team, led by Biaolong Su of NARI Technology Co., Ltd. together with colleagues from State Grid Henan Electric Power Company, set out to solve a problem that has long frustrated grid engineers. Composite power quality disturbances, formed by the simultaneous coupling of transient events such as voltage sags and continuous distortions such as flicker, exhibit severe feature overlap and a masking effect in which key discriminative signatures are submerged beneath the dominant waveform. Traditional identification pipelines, which first extract features using tools like the Fourier transform, S-transform, wavelet transform, or empirical mode decomposition and then feed hand-picked features into classifiers such as decision trees or support vector machines, work reasonably well for single disturbances. But they struggle when several disturbances coexist, and their performance degrades sharply as background noise increases.

Deep learning has already been applied to this challenge, with convolutional neural networks, LSTM hybrids, and image-based approaches that convert time series into two-dimensional representations. Yet these methods share a common weakness: when confronted with disturbance types or combinations absent from the training data, they must be retrained with new examples incorporated, which severely limits practical deployment. The new approach takes a different path. It processes raw voltage waveforms directly, without any manual feature engineering or explicit signal transformation, and it is explicitly designed to generalize to unseen composite disturbances through a mechanism the authors call confidence correction.

The architecture is built on a three-dimensional synergistic framework. First, a dual-branch fully convolutional network extracts multi-scale spatial and temporal-frequency representations from the raw time-series signal. The two branches use different depths and kernel sizes: a shallower branch captures local, abrupt features that characterize transient disturbances, while a deeper branch expands the receptive field to capture global variations. Their outputs are concatenated and fused into multi-scale feature maps. Second, a gated recurrent unit models the long-term dynamic evolution of the disturbance across adjacent fundamental cycles, using update and reset gates to control what historical information is retained, which also helps alleviate the vanishing gradient problem that plagues conventional recurrent networks. Third, an attention mechanism computes dynamic weights across the hidden states, concentrating probability mass on the narrow temporal windows where transient anomalies occur while assigning near-zero weight to the long stretches of ordinary sinusoidal waveform that would otherwise dilute the signal.

This structural synergy addresses a fundamental imbalance in disturbance data. Transient events are highly localized, their anomalous signatures confined to brief windows, while continuous disturbances such as harmonics and flicker persist across the entire record. Conventional deep architectures processing such heterogeneous signals tend to let the dominant normal cycles wash out the weak representations of localized transients. By letting the attention layer dynamically highlight the informative segments, the framework neutralizes this information sparsity and learns balanced representations across complex composite patterns.

The most distinctive element, however, is the disturbance recognition enhancement mechanism built on confidence correction. The model treats disturbance identification as a multi-label problem: each of seven disturbance types, including voltage sags, swells, interruptions, flicker, harmonics, impulsive transients, and oscillatory transients, is a label that can be present or absent, and a single signal can carry several labels at once. After the network produces initial confidence scores for each category, an enhanced label distinguishes single from compound disturbances, and calibrated confidence enhancement factors adjust the prediction scores. Transient disturbances with concentrated, localized energy receive an explicit factor of 1.25, while steady-state periodic distortions receive an implicit factor of 1.10. These factors were determined through a constrained grid search on a validation set, with a strict safeguard: an inactive disturbance whose initial confidence falls below the 0.5 decision threshold must never be boosted across that boundary, eliminating false-positive pseudo-disturbances.

To test the system, the researchers synthesized an extensive dataset conforming to IEEE Std 1159 and IEC 61000-4-30 standards, generating 48,000 samples across 48 categories: seven single disturbances, seventeen dual composites, seventeen triple composites, six quadruple composites, and normal signals. Each signal comprised 1,280 sampling points covering ten cycles at 50 hertz, with Gaussian white noise added at signal-to-noise ratios uniformly distributed between 10 and 50 decibels. The results were striking. The model achieved 100 percent accuracy on normal and single disturbances, 99.75 percent on dual composites, 99.71 percent on triple composites, and 98.60 percent on the highly coupled quadruple cases, with Hamming losses as low as 0.0004.

Even more impressive was the zero-shot generalization. In one experiment, the model was trained without ever seeing nine triple disturbance categories and all six quadruple categories, yet it identified the completely unseen quadruple composites with 99.45 percent accuracy. When the training set was expanded to include all triple disturbances, accuracy on unseen quadruple cases rose to 99.83 percent. The authors attribute this combinatorial generalization to the fact that the network learns disentangled, noise-invariant representations of each constituent disturbance signature from lower-order examples, so a quadruple combination is not treated as an unlearnable monolithic pattern. The confidence correction mechanism then calibrates the marginal detection probabilities to produce reliable multi-label outputs for novel configurations.

The method also proved its mettle against noise and rival approaches. Benchmarked against classical pipelines such as wavelet transform with decision trees, S-transform with support vector machines, and empirical mode decomposition with support vector machines, the proposed model maintained 95.3 percent accuracy at a harsh 10 dB signal-to-noise ratio, where the traditional methods collapsed to between 68.5 and 72.1 percent. Against deep learning competitors including pure CNNs, CNN-LSTM hybrids, ResNet with Gramian angular fields, and Transformers, it likewise dominated under heavy noise, while requiring only 4.3 million floating-point operations, 4.5 minutes of training time, and 10 seconds of testing time, a fraction of the computational cost of its rivals. Validation on 1,000 field-recorded waveforms from actual distribution network monitoring terminals yielded an overall average accuracy of 99.7 percent in noise-free conditions and 96.1 percent at 10 dB, including successful identification of triple compound events in real data.

Ablation experiments confirmed that each component earns its place: removing attention dropped accuracy to 97.82 percent, removing temporal modeling to 96.45 percent, and removing the convolutional feature extraction to 95.96 percent, with all differences statistically significant at p below 0.01. The researchers envision next steps that include deploying the lightweight framework on embedded edge hardware such as DSP and FPGA-based monitoring devices for real-time inference, incorporating open-set learning to detect emerging unmodeled disturbance patterns, and extending the single-channel model into wide-area spatiotemporal networks capable of concurrent disturbance identification and fault source localization. For grids racing toward high renewable penetration, a compact model that reads raw waveforms, untangles overlapping disturbances, and generalizes to combinations it has never encountered could become an essential guardian of power quality.

Subject of Research: Deep learning identification of composite power quality disturbances in distribution networks

Article Title: Composite power quality disturbance identification method considering confidence correction

Article References: Composite power quality disturbance identification method considering confidence correction. (n.d.). https://doi.org/10.3389/fenrg.2026.1936537

Image Credits: AI Generated

DOI: 10.3389/fenrg.2026.1936537

Keywords: power quality disturbances, deep learning, FCN-GRU-Attention, confidence correction, smart grids, distribution networks, renewable energy integration, attention mechanism, multi-label classification, noise robustness, voltage sags, harmonics