Intelligent Fault and Lightning Detection Algorithm for VSC-MTDC grids based on ResNet with Hybrid Attention Mechanism

• Fault and lightning detection must be accurate and rapid because VSC-MTDC faults can escalate within milliseconds, leaving little transient data before circuit breakers must act.Source: §I · p. 1
• Prior physical-model methods face sensitivity to fault conditions, system parameters, and manually set thresholds, while earlier deep-learning studies did not address internal lightning interference.Source: §I · pp. 1–2
• The proposed method converts double-ended ICTWs into S-transform time-frequency images, then uses RWHAM hybrid attention to emphasize key image information for fault and lightning detection.Source: §I · p. 2
• 21,620 PSCAD/EMTDC test cases support reliable, rapid fault and lightning detection with excellent generalization ability for VSC-MTDC grids.Source: §I · p. 2

II. RESEARCH SYSTEM

• The test system is a four-terminal bipolar VSC-HVDC grid with two topologies and 100-km frequency-dependent overhead lines.Current-limiting reactors are installed at line terminals or converter-station outlets.Source: §II · pp. 2–3

III. TIME-FREQUENCY FEATURE EXTRACTION USING S-TRANSFORM

• Phase-modal transformation decouples bipolar DC currents before the S-transform is applied to 1-mode current traveling waves at both line terminals.The S-transform provides higher frequency resolution at low frequencies and higher time resolution at high frequencies.Source: §III · p. 3
• A 0.2 ms ICTW window is rendered as a 224×224-pixel, 3-channel image by concatenating the two terminal matrices horizontally.The window avoids subsequent reflected and refracted waves while limiting detection delay; each single-terminal image spans 0–0.2 ms and 0–50 Hz on its time and frequency axes.Source: §III · p. 3
• Internal faults and lightning show similar two-terminal ICTW color trends, unlike external events; frequency patterns also distinguish faults from lightning.The images cover internal PGFs, external PPFs, and lightning interference; the tables show cases with varying locations and fault resistances or lightning positions.Source: §III · pp. 3–4, Table II · p. 4, Table III · p. 4

• The paper proposes RWHAM to balance protection speed and reliability in MTDC grids.The model is used to classify ICTW time-frequency feature maps.Source: §IV · p. 4
• RWHAM extracts complex ICTW time-frequency characteristics that traditional analytical methods are described as difficult to capture.Source: §IV · p. 4
• The feature maps are used to identify internal faults, internal lightning interference, and external faults or lightning interference accurately and quickly.Source: §IV · p. 4

A. Resnet structure

• ResNet learns the residual F(x)=H(x)-x, making deep networks easier to optimize than learning the original mapping directly.A typical residual block has two convolutional layers, each followed by batch normalization and ReLU.Source: A. Resnet structure · p. 5

B. Bottleneck Attention Module

• BAM channel attention pools each channel, compresses and restores channel depth, then assigns higher sigmoid weights to more informative channels.The channel representation is compressed to C/16 before being restored to C.Source: B. Bottleneck Attention Module · p. 5
• BAM spatial attention aggregates channel information and uses a 7×7 convolution to assign higher weights to critical spatial regions.Average- and max-pooled maps are concatenated before the convolution produces spatial weights.Source: B. Bottleneck Attention Module · p. 5
• Combining channel and spatial attention captures important features while suppressing irrelevant ones.The BAM output combines the input feature map with both attention maps.Source: B. Bottleneck Attention Module · p. 5

C. The proposed RWHAM model

• BAM within RB_blocks weights salient spatial regions and informative channels, with attention focused on terminal-to-terminal trends and low- versus high-frequency image patterns.The model is described as becoming more sensitive to these image patterns, improving detection accuracy.Source: C. The proposed RWHAM model · pp. 5–6, Figure 4 · p. 5
• RWHAM uses four cascaded residual layers containing 3, 4, 6, and 3 RB_blocks, followed by global average pooling and a fully connected classifier.Channel dimensions expand from 64 to 128, 256, and 512; Bayesian optimization selects hyperparameters using validation accuracy.Source: C. The proposed RWHAM model · p. 6
• Attention maps assign higher weights to high-frequency regions for internal lightning and low-frequency regions for faults, while left-terminal amplitude helps distinguish external from internal faults.The paper attributes the fault-versus-lightning distinction to differences in energy distribution across high and low frequencies.Source: C. The proposed RWHAM model · p. 6

D. The training process

• Training simulations span two topologies, multiple DC and AC fault types, fault locations, resistance ranges, and lightning waveforms.Lightning cases include 1.2/50 μs and 8/20 μs double-exponential waveforms; the simulation frequency is 100 kHz.Source: D. The training process · pp. 6–7
• The classifier uses three labels: internal faults, internal lightning interference, and external faults or lightning interference.External faults and lightning share a class because neither triggers a trip signal, whereas internal faults must be distinguished from internal lightning.Source: D. The training process · p. 7
• Training images are resized and cropped to 224×224 pixels, converted to tensors, normalized by channel, and paired with labels.The stated purposes of normalization include faster convergence and more stable gradient updates.Source: D. The training process · p. 7
• Training minimizes cross-entropy loss with L2 regularization using AdamW at a 0.001 learning rate for 30 epochs.The batch size is 128, and the reported loss after 30 epochs is 2.12×10^-5.Source: D. The training process · p. 7

V. SIMULATION VALIDATION AND ANALYSIS

• A detailed VSC-MTDC grid model was developed in PSCAD/EMTDC for algorithm validation.The model setup is described as the basis for validating the proposed algorithm.Source: §V · p. 7
• The study simulated 21,968 diverse fault cases described in Section IV to generate training datasets.Source: §V · p. 7
• Time-series simulations and numerical calculations were conducted on a computer with an Intel Core i7-1260P CPU.Source: §V · p. 7

A. Typical test

• 21,620 fault cases distinct from the training datasets were simulated to test performance across different fault and lightning locations and types.Source: A. Typical test · p. 7
• The test set contains 7,600 samples labeled 0, 1,368 labeled 1, and 12,652 labeled 2.The section reports calculating each label’s accuracy, precision, recall, and F1 score, plus accuracy standard deviation and ten-run averages.Source: A. Typical test · pp. 7–8
• At least 99.72% average accuracy, precision, recall, and F1-score was achieved for each label across 10 runs, with accuracy standard deviations below 0.01%.The reported test identifies internal faults, internal lightning, and external faults or lightning across locations and withstands fault resistances up to 1005 Ω.Source: Figure 8 · p. 8

B. Generalization ability

• Under voltage-level or transmitted-power scaling, normalization standardizes the time-frequency images, allowing the trained model to be applied without retraining.The passage describes scaling the signal by a factor k while grid impedance parameters remain constant.Source: B. Generalization ability · p. 8
• For 50 mH and 500 mH reactors, average accuracy, precision, recall, and F1-score are each at least 99.72% across 10 runs.Accuracy standard deviations are below 0.02%.Source: B. Generalization ability · p. 9, Table VIII · p. 9
• The algorithm is reported as nearly unaffected by reactor value and able to identify fault and lightning types across diverse topologies and almost all fault locations.Source: B. Generalization ability · p. 9

C. Detection time

• Detection time is defined from fault or lightning occurrence to detection completion and comprises T1, T2, and T3.T1 is the 0.2 ms extracted ICTW window; T3 averages 0.58 ms across test datasets.Source: C. Detection time · p. 9
• The transmission-delay components increase with DC line length, and the traveling-wave delay reaches its maximum when a fault occurs at the line terminal.Source: C. Detection time · p. 9
• For VSC-MTDC grids, when a DC line is no longer than 508 km, the algorithm can be used as main protection; when it exceeds 508 km, it can be used as backup protection.Source: C. Detection time · p. 9

D. Comparison with other algorithms

• For a 100 km DC line, maximum detection time is 2.52 ms; for an 800 km line, it rises to 8.33 ms and the method is suitable as backup protection.Source: D. Comparison with other algorithms · p. 10
• Among the compared algorithms, the proposed method achieves the highest detection accuracy, requires no manual threshold setting, and retains strong generalization when system parameters change.The comparison evaluates algorithms across the original system, an 800 km line, and a 50 mH current-limiting reactance.Source: D. Comparison with other algorithms · pp. 9–10, Table IX · p. 10

VI. CONCLUSION

• 99.97% detection accuracy is reported for nominal grid configurations, while accuracy remains above 99.94% under parameter changes without retraining.Source: §VI · p. 10
• For DC lines no longer than 508 km, detection time is within 5 ms for primary protection; in other scenarios the algorithm serves as backup protection.The conclusion attributes the other scenarios' longer detection times to traveling-wave and signal-propagation delays.Source: §VI · p. 10

APPENDIX A

• Tables A.I and A.II list hyperparameters for the core layers in BAM and RB_block.Source: Appendix A · p. 10
• Table A.III reports RWHAM core-layer input sizes, filter sizes, strides, and output sizes for conv1 and conv2.Source: Appendix A · p. 10