HVDC system commutation failure fault tracing method based on double-flow gate fusion network

Abstract

The application discloses a kind of based on double-flow gate fusion network's HVDC system commutation failure fault tracing method, steps include: constructing TriFuseNet model for commutation failure fault tracing, obtaining normalized fault data by normalizing processing to fault signal, direct current signal and alternating voltage signal are converted into time-frequency diagram and MTF image respectively, each fault graph in fault data set is input into trained TriFuseNet model for classification one by one.The commutation failure fault tracing method constructs TriFuseNet network and utilizes the feature extraction module PAR Block therein as core feature extraction module, so as to realize the fine extraction of feature map in different scales and different directions;Cross-modal feature is realized by multi-scale feature extraction fusion module in TriFuseNet network Gradual interaction fusion.

Description

Technical Field

[0001] This invention relates to a fault tracing method, and more particularly to a fault tracing method for commutation failure in an HVDC system based on a dual-stream gated fusion network. Background Technology

[0002] High-voltage direct current (HVDC) transmission offers advantages such as high-power transmission, low energy loss, and effectiveness in ensuring network stability under dynamic load changes, asynchronous operation, and varying load conditions. It has become a key measure in my country to address the reverse distribution of energy and load centers. However, HVDC systems use thyristors without self-turn-off capability as commutation elements, making them prone to commutation failures when AC faults occur on the inverter side. Commutation failures lead to a sharp increase in DC current and a significant drop in DC voltage and power. Consecutive commutation failures can even trigger DC blocking, threatening the safe and stable operation of the power system.

[0003] In practical engineering, commutation failures cannot be completely avoided, and the influencing factors are numerous and complexly coupled. Currently, fault analysis and handling of DC systems heavily rely on the operation and maintenance experience and skills of on-site personnel, lacking effective intelligent analysis methods and decision support tools. It is necessary to intelligently classify the causes of commutation failures in HVDC systems to assist operation and maintenance personnel in conducting fault analysis and handling, thereby effectively grasping the system status, improving fault handling efficiency, and ensuring the safe and stable operation of the system.

[0004] Artificial intelligence (AI) technology possesses powerful feature extraction and data processing capabilities, enabling it to provide intelligent decision support for operations and maintenance personnel. Some scholars have also applied AI technology to the classification of commutation failures and conducted research and exploration in this area.

[0005] The core component of a conventional HVDC transmission system is the thyristor commutator valve. When current transfers between the two valves, the commutation process cannot be completed instantaneously due to inductance. After commutation between the two valve arms is completed, the valve expected to be turned off may fail to regain its blocking capability within a certain period under reverse voltage; or the commutation process may not be completed, and the valve expected to be turned off may reverse its phase to the valve expected to be turned on when the valve voltage changes from negative to positive. This phenomenon is called commutation failure. The essence of commutation failure is that the turn-off angle γ is less than the minimum turn-off angle γ0 required for the thyristor to regain its blocking capability. Therefore, a necessary condition for the commutation process to proceed normally is that the time corresponding to the turn-off angle γ is sufficient for the commutated valve to regain its blocking capability. The commonly used expression for the turn-off angle γ is: In the formula: L r I is the commutation inductance; ω is the system angular frequency; I d U is direct current; Lβ is the effective value of the AC bus line voltage on the inverter side; β is the leading trigger angle; φ is the zero-crossing offset angle of the commutation voltage caused by a system fault.

[0006] When an AC fault occurs on the inverter-side AC bus, causing a voltage drop and an increase in DC current on the converter bus, the switching angle γ of the converter valve continuously decreases, as shown by the expression for the switching angle γ. Commutation failure will occur when γ < γ0. Therefore, AC system faults on the inverter side are the main cause of commutation failure. Specific fault types can be categorized as: single-phase grounding (AG, BG, CG), two-phase short circuit (AB, AC, BC), two-phase grounding (ABG, ACG, BCG), and three-phase grounding (ABG) faults. Based on the above analysis, commutation failure is mainly affected by the combined influence of AC voltage and DC current on the inverter side. Considering only a single electrical quantity as input may lead to misjudgment or missed judgment. Summary of the Invention

[0007] The purpose of this invention is to provide a fault tracing method for commutation failure in HVDC systems based on a dual-stream gated fusion network. This method can automatically model the complex causal relationship between the two using neural networks, thereby accurately diagnosing the fault type that leads to commutation failure.

[0008] Technical solution: The present invention provides a method for tracing the source of commutation failure in an HVDC system based on a dual-stream gated fusion network, comprising the following steps:

[0009] Step 1: Construct a TriFuseNet model for commutation failure fault tracing and train the constructed TriFuseNet model;

[0010] Step 2: Obtain the fault signal of commutation failure in HVDC system, normalize the fault signal to obtain normalized fault data, and then perform data preprocessing on the normalized fault data. Convert the DC current signal into a time-frequency graph and the AC voltage signal into an MTF image. Combine the time-frequency graph and MTF image at the same time as a fault graph group, and use each fault graph group to construct a fault dataset.

[0011] Step 3: Input each fault map group in the fault dataset into the trained TriFuseNet model for classification, output the commutation failure fault tracing results, and obtain the specific fault type that caused the commutation failure.

[0012] Furthermore, in step 1, the constructed TriFuseNet model includes a classifier, a multi-scale feature extraction and fusion module, and two preliminary feature extraction modules;

[0013] Two preliminary feature extraction modules are used to receive the time-frequency map and MTF image in the fault map group respectively, and perform preliminary feature extraction on the time-frequency map and MTF image respectively to obtain coarse-grained features in the time-frequency map and MTF image;

[0014] The multi-scale feature extraction and fusion module is used to extract coarse-grained features at multiple scales to obtain fine-grained fault features, and to fuse the fine-grained fault features to obtain fault fusion features.

[0015] The classifier is used to classify the fault fusion features and obtain the fault source tracing results of commutation failure.

[0016] Furthermore, the preliminary feature extraction module includes a first Conv layer and a second Conv layer;

[0017] The output data of the first Conv layer is used as the input data of the second Conv layer; the first Conv layer of the two preliminary feature extraction modules is used to receive the time-frequency map and the MTF image respectively, and the second Conv layer is used to output the coarse-grained features of the time-frequency map and the MTF image; the first Conv layer is used to perform 5×5 convolution processing, and the second Conv layer is used to perform 3×3 convolution processing.

[0018] Furthermore, the multi-scale feature extraction and fusion module includes a BGF fusion module and two multi-scale feature extraction modules;

[0019] The multi-scale feature extraction module includes three sequentially connected feature extraction modules (PAR Blocks), which are used to perform fine-grained fault feature extraction for three stages.

[0020] The BGF fusion module is used to fuse the fine-grained fault features obtained in the same stage of two multi-scale feature extraction modules and output the fused fine-grained fault features.

[0021] Furthermore, the feature extraction module PAR Block includes a first Conv2d convolutional layer, a first PARU layer, a second PARU layer, and a first AvgPool layer. The first Conv2d convolutional layer, the first PARU layer, and the second PARU layer are connected in sequence. The input data of the first Conv2d convolutional layer is added to the output data of the second PARU layer and then input to the first AvgPool layer. The input of the first Conv2d convolutional layer serves as the input of the feature extraction module PAR Block, and the output of the first AvgPool layer serves as the output of the feature extraction module PAR Block.

[0022] Furthermore, the first PARU layer and the second PARU layer are PARU layers with the same structure; the PARU layer includes a pyramid convolutional layer, a second Conv2d convolutional layer, an ICBAM layer, a third Conv2d convolutional layer, and a second AvgPool layer; the pyramid convolutional layer, the second Conv2d convolutional layer, the ICBAM layer, and the third Conv2d convolutional layer are connected in sequence, and the input data of the pyramid convolutional layer is added to the output data of the third Conv2d convolutional layer and then input to the second AvgPool layer; the input of the pyramid convolutional layer is used as the input of the PARU layer, and the output of the second AvgPool layer is used as the output of the PARU layer.

[0023] Furthermore, the BGF fusion module includes a first BGF layer, a second BGF layer, a third BGF layer, and a fourth BGF layer connected in sequence. The first BGF layer is used to receive and fuse the coarse-grained features of the time-frequency map and the MTF image output by the two preliminary feature extraction modules, and output a first fused feature. The second BGF layer is used to receive and fuse the first fused feature and the fine-grained fault features output by the PAR Block feature extraction module of the first stage of the two multi-scale feature extraction modules, and output a second fused feature. The third BGF layer is used to receive the second fused feature and the fine-grained fault features output by the PAR Block feature extraction module of the second stage of the two multi-scale feature extraction modules, and output a third fused feature. The fourth BGF layer is used to receive the third fused feature and the fine-grained fault features output by the PAR Block feature extraction module of the third stage of the two multi-scale feature extraction modules, and output the fused fine-grained fault features.

[0024] Furthermore, the first BGF layer includes a BGF layer; the second and third BGF layers each include a third Conv layer, a BGF layer, and a fourth Conv layer connected in sequence; the fourth BGF layer includes a third Conv layer, a BGF layer, a SimAM layer, and a fourth Conv layer connected in sequence; the third Conv layer is used for 1×1 convolution processing, and the fourth Conv layer is used for 3×3 convolution processing; the BGF layer is used for multimodal feature fusion; the SimAM layer is used to mine the spatial importance within the fused feature map.

[0025] Furthermore, the specific steps for multimodal feature fusion in the BGF layer are as follows:

[0026] First, global average pooling is performed on the feature maps F1 and F2 from the two modalities. and The compressed representations of the features are obtained as S1 and S2, respectively. , The global average pooling is calculated as follows:

[0027]

[0028] In the formula: and These represent feature maps F1 and F2 respectively. BGF and F2 BGF The global average pooling result of the c-th channel at spatial location (H×W), where H represents the height of the feature map, W represents the width of the feature map, and C represents the number of feature channels. This represents the feature map on the c-th channel;

[0029] The two features are then concatenated and fed into a multilayer perceptron (MLP) to model the nonlinear dependencies between channels, generating a fusion gating weight F. Then merge the gating weights F BGF After dividing the data into two feature representations, W1 and W2, two gated vectors P1 and P2 are generated using a parameterized Softmax function. The adaptive acquisition of features for each channel is as follows:

[0030]

[0031]

[0032]

[0033] In the formula, τ is a parameter factor used to adjust the weight distribution of the two features during fusion. The smaller τ is, the more Softmax tends to select only one branch, while the larger τ is, the more evenly Softmax fuses the two branches. This indicates a splicing operation performed at the channel level;

[0034] Finally, the two gated vectors P1 and P2 are weighted onto the original modality feature map to obtain weighted fusion features, and finally, the fusion output is formed by element-wise addition:

[0035]

[0036] In the formula, This represents the dot product within the channel.

[0037] Furthermore, the classifier includes a fifth Conv layer, a GAP layer, a first FC layer, and a second FC layer connected in sequence; the fifth Conv layer is used for 3×3 convolution processing, the GAP layer is used for global average pooling processing, and the first FC layer and the second FC layer are used to convert the features after global average pooling processing into classification output.

[0038] Compared with existing technologies, the advantages of this invention are as follows: by constructing a TriFuseNet network and using its feature extraction module PAR Block as the core feature extraction module, fine extraction of feature maps at different scales and directions is achieved; the multi-scale feature extraction and fusion module in the TriFuseNet network realizes the stepwise interactive fusion of cross-modal features; before entering the classifier, the SimAM layer in the TriFuseNet network further refines the internal structure of the fused feature map through an attention mechanism. Attached Figure Description

[0039] Figure 1 This is a flowchart of the method of the present invention;

[0040] Figure 2 This is a schematic diagram of the TriFuseNet model of the present invention;

[0041] Figure 3 This is a schematic diagram of the feature extraction module PAR Block of the present invention;

[0042] Figure 4 This is a schematic diagram of the PARU layer of the present invention;

[0043] Figure 5 This is a schematic diagram of the BGF layer of the present invention;

[0044] Figure 6 This is a time-frequency diagram of a phase-A ground fault occurring at the AC busbar on the inverter side of this invention.

[0045] Figure 7 This is a time-frequency diagram of a phase-to-phase (AB) short circuit occurring at the AC bus on the inverter side of the present invention.

[0046] Figure 8 This is the time-frequency diagram of phase AB grounding at the AC busbar on the inverter side of the present invention;

[0047] Figure 9 This is a time-frequency diagram of a three-phase ground fault occurring at the AC busbar on the inverter side of the present invention.

[0048] Figure 10 This is the MTF image of a phase-A ground fault occurring at the AC busbar on the inverter side of this invention.

[0049] Figure 11 This is the MTF image of a phase-to-phase short-circuit fault occurring at the AC bus on the inverter side of this invention.

[0050] Figure 12 This is the MTF image of an AB-phase ground fault occurring at the AC busbar on the inverter side of this invention.

[0051] Figure 13This is the MTF image of a three-phase ground fault occurring at the AC busbar on the inverter side of the present invention.

[0052] Figure 14 This is a schematic diagram of the pyramid convolutional layer structure of the present invention;

[0053] Figure 15 This is a schematic diagram of the ICBAM layer structure of the present invention;

[0054] Figure 16 The structure of the CIGRE HVDC benchmark test model of the present invention;

[0055] Figure 17 This is a graph showing the change in loss value during the model training process of this invention;

[0056] Figure 18 This is a graph showing the accuracy change during the model training process of the present invention.

[0057] Figure 19 This is the confusion matrix of the test results of this invention;

[0058] Figure 20 This is a graph showing the effect of different noise levels on the model performance in this invention;

[0059] Figure 21 The results are experimental test results under data loss conditions according to the present invention. Detailed Implementation

[0060] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the embodiments described.

[0061] like Figure 1 As shown, the present invention discloses a method for tracing the source of commutation failure in an HVDC system based on a dual-stream gated fusion network, comprising the following steps:

[0062] Step 1: Construct a TriFuseNet model for commutation failure fault tracing and train the constructed TriFuseNet model;

[0063] Step 2: Obtain the fault signal of commutation failure in HVDC system, normalize the fault signal to obtain normalized fault data, and then perform data preprocessing on the normalized fault data. Convert the DC current signal into a time-frequency graph and the AC voltage signal into an MTF image. Combine the time-frequency graph and MTF image at the same time as a fault graph group, and use each fault graph group to construct a fault dataset.

[0064] Step 3: Input each fault map group in the fault dataset into the trained TriFuseNet model for classification, output the commutation failure fault tracing results, and obtain the specific fault type that caused the commutation failure.

[0065] By constructing the TriFuseNet network and utilizing its feature extraction module PAR Block as the core feature extraction module, fine extraction of feature maps at different scales and directions is achieved. The multi-scale feature extraction and fusion module in the TriFuseNet network enables stepwise interactive fusion of cross-modal features. Before entering the classifier, the SimAM layer in the TriFuseNet network further refines the internal structure of the fused feature maps through an attention mechanism.

[0066] Furthermore, such as Figure 2 As shown, in step 1, the constructed TriFuseNet model includes a classifier, a multi-scale feature extraction and fusion module, and two preliminary feature extraction modules;

[0067] Two preliminary feature extraction modules are used to receive the time-frequency map and MTF image in the fault map group respectively, and perform preliminary feature extraction on the time-frequency map and MTF image respectively to obtain coarse-grained features in the time-frequency map and MTF image;

[0068] The multi-scale feature extraction and fusion module is used to extract coarse-grained features at multiple scales to obtain fine-grained fault features, and to fuse the fine-grained fault features to obtain fault fusion features.

[0069] The classifier is used to classify the fault fusion features and obtain the fault source tracing results of commutation failure.

[0070] Furthermore, the preliminary feature extraction module includes a first Conv layer and a second Conv layer;

[0071] The output data of the first Conv layer is used as the input data of the second Conv layer; the first Conv layer of the two preliminary feature extraction modules is used to receive the time-frequency map and the MTF image respectively, and the second Conv layer is used to output the coarse-grained features of the time-frequency map and the MTF image; the first Conv layer is used to perform 5×5 convolution processing, and the second Conv layer is used to perform 3×3 convolution processing.

[0072] Furthermore, the multi-scale feature extraction and fusion module includes a BGF fusion module and two multi-scale feature extraction modules;

[0073] The multi-scale feature extraction module includes three sequentially connected feature extraction modules (PAR Blocks), which are used to perform fine-grained fault feature extraction for three stages.

[0074] The BGF fusion module is used to fuse the fine-grained fault features obtained in the same stage of two multi-scale feature extraction modules and output the fused fine-grained fault features.

[0075] Furthermore, such as Figure 3 As shown, the feature extraction module PAR Block includes a first Conv2d convolutional layer, a first PARU layer, a second PARU layer, and a first AvgPool layer. The first Conv2d convolutional layer, the first PARU layer, and the second PARU layer are connected in sequence. The input data of the first Conv2d convolutional layer is added to the output data of the second PARU layer and then input to the first AvgPool layer. The input of the first Conv2d convolutional layer is used as the input of the feature extraction module PAR Block, and the output of the first AvgPool layer is used as the output of the feature extraction module PAR Block.

[0076] Furthermore, such as Figure 4 As shown, the first PARU layer and the second PARU layer are PARU layers with the same structure; the PARU layer includes a pyramid convolutional layer, a second Conv2d convolutional layer, an ICBAM layer, a third Conv2d convolutional layer, and a second AvgPool layer; the pyramid convolutional layer, the second Conv2d convolutional layer, the ICBAM layer, and the third Conv2d convolutional layer are connected in sequence, and the input data of the pyramid convolutional layer is added to the output data of the third Conv2d convolutional layer and then input to the second AvgPool layer; the input of the pyramid convolutional layer is used as the input of the PARU layer, and the output of the second AvgPool layer is used as the output of the PARU layer.

[0077] Furthermore, the first, second, and third Conv2d convolutional layers are all two-dimensional convolutional layers used to perform cross-correlation operations between the input and the convolutional kernel; the first and second AvgPool layers are both used for average pooling, a pooling operation commonly used in convolutional neural networks (CNNs). The main purpose of pooling is to reduce the size of the feature map by downsampling, thereby reducing computational complexity and preventing overfitting. Average pooling achieves this purpose by calculating the average value of all pixels within the pooling window.

[0078] Furthermore, each layer of the pyramid convolutional layer uses different types of convolutional kernels, such as Figure 14As shown, the size and depth of its convolutional kernels are variable, thus effectively extracting detailed information at different scales. Furthermore, each convolutional kernel in the PyConv pyramid convolutional layer is replaced by a grouped convolution, which reduces computational cost, speeds up computation, and lowers model complexity. Simultaneously, when combined with multi-scale structures, it maintains good feature representation capabilities by limiting the channel range perceived by convolutional kernels at different scales. Finally, the output feature maps F of each branch are... i The input and output feature maps are concatenated along the channel dimension to form a new feature map of the same size. The entire computation process of the pyramid convolutional block is as follows:

[0079]

[0080]

[0081] In the formula: k i G represents the kernel size. i In this invention, the number of kernels n in the pyramid convolutional layer PyConv is set to 4, and the kernel size k represents the number of groups. i If the values ​​are set to 3, 5, 7, and 9, then each convolutional branch F i The output channel dimension is 1 / 4 of the input channel dimension, and the number of groups G in the grouped convolution is... i Set to 1, 2, 4, and 8.

[0082] Furthermore, the ICBAM layer introduces a spatial awareness mechanism to transform the input feature map... First, rearrange it in the spatial dimension as Nonlinear feature extraction is then performed using a multilayer perceptron (MLP) consisting of two fully connected layers. The output tensor is then restored to its original dimensions [H, W, C], thereby assigning an independent attention weight to each channel at each spatial location, achieving a more refined channel enhancement mechanism.

[0083] Regarding spatial attention, unlike existing CBAM layers which treat it as a whole two-dimensional region, the ICBAM layer further refines spatial attention into two directional branches, modeling along the horizontal (W) and vertical (H) directions of the feature map, respectively. Each direction uses average pooling and max pooling to extract features, and then channel aggregation and three convolutional layers are used to construct attention weight maps, capturing the response patterns of features along the horizontal (x) and vertical (y) axes, enabling the extraction of more comprehensive and richer information. Finally, the attention weights in the two directions are synergistically fused with the channel attention output to form... This results in a multi-dimensional enhanced expression. The specific calculation process of ICBAM can be represented as follows:

[0084]

[0085]

[0086]

[0087]

[0088] In the formula, M C (F1 AM () represents channel attention. For channel attention-weighted output, M W (F2 AM M represents the horizontal attention mechanism. H (F2 AM This refers to the vertical attention mechanism. σ represents element-wise multiplication, Reshape() represents shape reshaping operation, σ() is the Sigmoid activation function, AvgPool() is the average pooling operation, MaxPool() is the max pooling operation, Conv³() represents three consecutive convolutional layers, and MLP() is the multilayer perceptron.

[0089] The above M C (F1 AM M represents channel attention with spatial awareness. W (F2 AM ) and M H (F2 AM These correspond to the spatial attention branches in the W and H directions, respectively. The fused attention features... It not only possesses the ability to model differences across global channels, but also accurately mines local response patterns of spatial structures in the horizontal and vertical directions. This effectively enhances the network's feature selectivity and fine-grained perception capabilities, providing more stable and robust feature support for subsequent feature fusion and classification tasks.

[0090] Furthermore, the BGF fusion module includes a first BGF layer, a second BGF layer, a third BGF layer, and a fourth BGF layer connected in sequence. The first BGF layer is used to receive and fuse the coarse-grained features of the time-frequency map and the MTF image output by the two preliminary feature extraction modules, and output a first fused feature. The second BGF layer is used to receive and fuse the first fused feature and the fine-grained fault features output by the PAR Block feature extraction module of the first stage of the two multi-scale feature extraction modules, and output a second fused feature. The third BGF layer is used to receive the second fused feature and the fine-grained fault features output by the PAR Block feature extraction module of the second stage of the two multi-scale feature extraction modules, and output a third fused feature. The fourth BGF layer is used to receive the third fused feature and the fine-grained fault features output by the PAR Block feature extraction module of the third stage of the two multi-scale feature extraction modules, and output the fused fine-grained fault features.

[0091] Furthermore, the first BGF layer includes a BGF layer; the second and third BGF layers each include a third Conv layer, a BGF layer, and a fourth Conv layer connected in sequence; the fourth BGF layer includes a third Conv layer, a BGF layer, a SimAM layer, and a fourth Conv layer connected in sequence; the third Conv layer is used for 1×1 convolution processing, and the fourth Conv layer is used for 3×3 convolution processing; the BGF layer is used for multimodal feature fusion; the SimAM layer is used to mine the spatial importance within the fused feature map.

[0092] Furthermore, such as Figure 5 As shown, the specific steps for multimodal feature fusion in the BGF layer are as follows:

[0093] First, global average pooling is performed on the feature maps F1 and F2 from the two modalities. and The compressed representations of the features are obtained as S1 and S2, respectively. , Let R be the set of real numbers. The global average pooling is calculated as follows:

[0094]

[0095] In the formula: and These represent feature maps F1 and F2 respectively. BGF and F2 BGF The global average pooling result of the c-th channel at spatial location (H×W), F1 BGF and F2 BGFEach is a feature map of size H×W×C, where each element is a real number, H represents the height of the feature map, W represents the width of the feature map, and C represents the number of feature channels. This represents the feature map on the c-th channel;

[0096] The two features are then concatenated and fed into a multilayer perceptron (MLP) to model the nonlinear dependencies between channels, generating a fusion gating weight F. BGF , Then merge the gating weights F BGF After dividing the data into two feature representations, W1 and W2, two gated vectors P1 and P2 are generated using a parameterized Softmax function. , The adaptive acquisition of features for each channel is as follows:

[0097]

[0098]

[0099]

[0100] In the formula, τ is a parameter factor used to adjust the weight distribution of the two features during fusion. The smaller τ is, the more Softmax tends to select only one branch, while the larger τ is, the more evenly Softmax fuses the two branches. This indicates a splicing operation performed at the channel level;

[0101] Finally, the two gated vectors P1 and P2 are weighted onto the original modality feature map to obtain weighted fusion features, and finally, the fusion output is formed by element-wise addition:

[0102]

[0103] In the formula, This represents the dot product within the channel.

[0104] Furthermore, the SimAM layer does not involve intermodal weight allocation, but instead measures the "activation necessity" of each neuron on a channel-by-channel and position-by-position basis, improving the response of key regions and suppressing useless redundant information, thereby improving the subsequent classification performance and model robustness. Moreover, the SimAM layer does not require the introduction of additional parameters, and only models based on the consistency and discriminativeness of neuron responses, thus having better parameter efficiency and generalization ability.

[0105] Furthermore, the classifier includes a fifth Conv layer, a GAP layer, a first FC layer, and a second FC layer connected in sequence; the fifth Conv layer is used for 3×3 convolution processing, the GAP layer is used for global average pooling processing, and the first FC layer and the second FC layer are used to convert the features after global average pooling processing into classification output.

[0106] Furthermore, in step 2, when normalizing the fault signals to obtain normalized fault data, the Min-Max method is used to normalize the original data. This avoids the problem of different dimensions between data points, thereby accelerating the model training process and reducing the risk of overfitting. The normalization calculation formula is as follows:

[0107]

[0108] In the formula, The data is normalized, and x is the original data. max and x min These are the maximum and minimum values ​​of the scaling interval, respectively. The scaling interval of this invention is [-1, 1].

[0109] Furthermore, in step 2, when converting the DC current signal into a time-frequency diagram, the continuous wavelet transform (CWT) is used to transform it into a time-frequency diagram to highlight non-stationary transient characteristics and enhance fault sensitivity. For any signal f(t), the CWT transformation formula is:

[0110]

[0111]

[0112] In the formula, The wavelet basis functions are obtained by scaling and translating the mother wavelet, where a is the scaling factor, b is the translation factor, * denotes conjugate, and R represents a real number. The wavelet basis functions used in this invention are Morse wavelets, such as... Figures 6-9 The images shown are time-frequency images of size 224×224 after CWT transformation, representing the occurrence of A-phase grounding, AB-phase short circuit, AB-phase grounding, and three-phase grounding faults at the AC bus on the inverter side.

[0113] Furthermore, in step 2, when converting the AC voltage signal into an MTF image, since the MTF can only process one signal at a time, before performing the MTF conversion, the phase voltages are stitched together vertically and arranged into a long-term sequence in the order of phases a, b, and c before conversion; for this time sequence First, the continuous values ​​in the time series are discretized and divided into a finite number of Q states. Common discretization methods include equal-width binning, quantile method (also called equal-frequency binning), and clustering methods. To ensure a balanced number of samples for each state, thereby making the statistical results of the transition matrix more stable, this invention uses the quantile method to discretize the AC voltage sequence. Subsequently, based on the Markov process concept, the transitions between adjacent points in the time series are statistically analyzed, and the state transition matrix is ​​constructed as follows:

[0114]

[0115] In the formula: elements Indicates from state e i Transition to state e j Based on the above formula, extending the transition probability to the entire time dimension yields the following MTF matrix:

[0116]

[0117] Each element of matrix M is determined by the transition probability between the states at time points i and j, thus characterizing the state transition pattern of the time series on a global scale in a two-dimensional plane;

[0118] The final matrix M can be viewed as a two-dimensional image, where each pixel value reflects the state transition relationship between corresponding time points, such as... Figures 10-13 The images shown are MTF images for A-phase grounding, AB-phase short circuit, AB-phase grounding, and three-phase grounding faults at the AC bus.

[0119] This invention uses the CIGRE Benchmark HVDC simulation model established in the PSCAD / EMTDC platform for fault simulation, and its basic structure is as follows: Figure 16 As shown, this DC system uses a single-pole 12-pulse converter with a rated DC voltage of 500kV and a capacity of 1000MW. The three-phase AC voltage U on the inverter side is collected. a U b U c and DC current I dc The signal was used as fault sample data. The sampling frequency was set to 10 kHz, the simulation step size was set to 50 μs, and the plotting step size was set to 100 μs. All faults were located on the converter bus on the inverter side, and the fault duration was 0.1 s (from 1.5 s to 1.6 s). To establish the fault dataset, fault parameters with different fault types, fault initial phase angles, transition resistances, and fault locations were set on the inverter side to simulate different fault conditions. The fault location was specified as occurring at the AC line, and the distance from the AC bus was specified in kilometers. Specific fault parameters are shown in Table 1.

[0120] Based on the above simulation parameter settings, data within one cycle after the fault was collected as samples. These samples were then categorized and statistically analyzed according to the fault type causing the commutation failure. Each fault type had 720 data sets, resulting in a total dataset of 7200 fault data sets, with each set containing 200 sample points. After data normalization, AC voltage and DC current were processed according to step 2, yielding 7200 images each. These images were then labeled and divided into training, validation, and test sets in a 7:2:1 ratio.

[0121] Table 1 shows the parameter settings for different operating conditions on the inverter side.

[0122]

[0123] The training iterations were set to 100, the batch size to 32, and the initial learning rate to 0.001. The experiment was conducted using Python within the PyTorch framework. The system ran on a deep learning workstation with an NVIDIA RTX A6000 GPU and 48GB of VRAM. The curves showing the changes in loss and accuracy during model training are shown below. Figure 17 and 18 As shown in the curve, the loss curve fluctuates greatly in the early stages due to the instability of forward and backward propagation. However, with the increase of iterations, the network loss generally shows a steady downward trend, while the training accuracy shows a steady upward trend, indicating that the model is continuously learning the features of the training data and its fitting ability is gradually improving. After about 70 iterations, the training accuracy reached over 99%, and the validation accuracy reached over 98%, and then remained stable, indicating that the network had converged. After 100 iterations, the highest classification accuracy on the validation set was 98.89%. The network parameters of the model at this point were saved and imported into the test set for testing, with an accuracy of 98.47%. The confusion matrix of the test results is shown below. Figure 19 As shown, the numbers on the diagonal represent the number of samples correctly classified for each fault type. It can be seen that the vast majority of samples were correctly classified, with only 1-3 samples misclassified for a few fault types. Training results indicate that the network fits well, without overfitting or underfitting, and the proposed model structure has good identification ability for commutation failure faults in HVDC systems.

[0124] To verify the effectiveness of each module in model classification, the ablation experiments shown in Table 2 were designed in this invention. The experimental results in the table show that each module plays a positive role in the performance of commutation failure tracing. For example, after introducing the feature fusion module BGF on the basis of the basic two-stream convolutional network in Combination 2, the accuracy and precision both improved by 2.77%, the recall improved by 2.92%, and the F1 score increased from 96.78% to 93.9%, an improvement of 2.88 percentage points. This proves that the designed gate-based stepwise feature fusion strategy can significantly improve diagnostic performance and is more robust and discriminative than the simple feature concatenation fusion method of the original two-stream convolutional network. Combinations 3, 4, and 6 confirm that the feature extraction module PAR, composed of pyramidal convolutional layers PyConv and ICBAM layers, is used in this invention. Block performs better than using any module alone because the combination of the PyConv convolutional layer and the ICBAM layer enables the network to extract features at different scales and in different directions. This allows the network to discover subtle differences in feature maps of different fault types through this feature extraction module. The comparison between combinations 6 and 7 shows that the SimAM attention mechanism introduced in the last stage after fusion, before the classifier, can strengthen the final fused features and improve the model's generalization ability and robustness.

[0125] Table 2 shows the results of the ablation experiment.

[0126]

[0127] Considering the interference of various noises in real signals, to test the noise resistance of the source tracing method proposed in this invention, Gaussian white noise with different signal-to-noise ratios was added to the test data. A lower signal-to-noise ratio indicates more noise in the original signal, more severe waveform distortion, and is less conducive to the model making correct judgments. The test results are as follows: Figure 20 As shown in the results, the performance indicators gradually decrease with increasing noise interference, and the performance degradation is significant when the signal-to-noise ratio is less than 40dB. Overall, the performance indicators are at a high level, and the network has good noise immunity, capable of handling the noise issues that arise during the operation of the HVDC system.

[0128] In practical HVDC systems, when a fault in the AC grid on the inverter side causes commutation failure in the converter, data loss may occur during data acquisition due to unavoidable factors such as voltage / current transformer failure or saturation, and electromagnetic / harmonic interference. This means that data may not be completely recorded during acquisition. This situation can affect system reliability assessment, leading to errors in analyzing and judging the operating status of the power system, thus impacting the safe and stable operation of the power system. Data loss can be categorized into two types: random discrete loss and random continuous loss. Random discrete loss refers to randomly selecting several discrete sampling points from the AC voltage and DC current waveform data in the test set and setting them to zero; random continuous loss refers to randomly selecting several consecutive sampling points and setting them to zero. Both scenarios can comprehensively evaluate the ability of the method of this invention to trace the source of commutation failure under data loss conditions.

[0129] Experimental test results under data loss are as follows Figure 21 As shown, the model's accuracy in tracing the source of data gradually decreases with the increase in the number of lost sampling points. Particularly when only a small number of sampling points are lost, the model is more resistant to consecutive loss than to discrete loss. However, when a large number of sampling points are lost, the model's resistance to discrete loss surpasses that to consecutive loss. Overall, regardless of whether there is discrete or consecutive loss, the model's classification accuracy remains above 75% even when more than 35 sampling points are lost. This demonstrates that the proposed method can effectively resist the impact of data loss.

[0130] In summary, the HVDC system commutation failure fault tracing method proposed in this invention, based on the dual-stream gated fusion network TriFuseNet, uses the original AC voltage and DC current fault signals as raw data. These are then processed into two-dimensional feature images as fault feature datasets. Based on the characteristics of these datasets, the dual-stream gated fusion network TriFuseNet is designed for feature extraction and fusion, achieving intelligent fault tracing for commutation failure. The method proposed in this invention can provide useful reference for maintenance personnel to quickly determine the fault type causing commutation failure. Through experimental comparative analysis, the following conclusions can be drawn from this invention:

[0131] 1) Compared with the prior art, the present invention determines the two signals of inverter side AC voltage and DC current as the data source required for commutation failure tracing based on the theoretical calculation formula of the turn-off angle, avoiding the potential information deviation caused by using a single source signal. Furthermore, according to the different fault characteristics presented by the signals, MTF and CWT are selected as signal processing methods, providing favorable data support for subsequent network learning.

[0132] 2) The source tracing model of this invention uses the PAR Block feature extraction module, which is composed of a pyramid convolutional layer PyConv and an ICBAM layer, as the core feature extraction module. The pyramid convolutional layer PyConv extracts at different scales, while the ICBAM layer calculates from different dimensions of the feature map. The combination of the two can achieve comprehensive and detailed extraction of fault features.

[0133] 3) The BGF fusion module designed in this invention can perform step-by-step fusion of different modal features. Compared with the direct splicing or weighted average fusion method in the prior art, the BGF fusion module adaptively determines different weights of different features through a gating mechanism, which solves the inconsistency and redundancy problems in the fusion process of multi-source heterogeneous features.

[0134] As described above, although the invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the invention itself. Various changes in form and detail may be made without departing from the spirit and scope of the invention as defined in the appended claims.

Claims

1. A method for tracing the source of commutation failure in an HVDC system based on a dual-stream gated fusion network, characterized in that, Includes the following steps: Step 1: Construct a TriFuseNet model for commutation failure fault tracing and train the constructed TriFuseNet model; Step 2: Obtain the fault signal of commutation failure in HVDC system, normalize the fault signal to obtain normalized fault data, and then perform data preprocessing on the normalized fault data. Convert the DC current signal into a time-frequency graph and the AC voltage signal into an MTF image. Combine the time-frequency graph and MTF image at the same time as a fault graph group, and use each fault graph group to construct a fault dataset. Step 3: Input each fault map group in the fault dataset into the trained TriFuseNet model for classification, output the commutation failure fault tracing results, and obtain the specific fault type that caused the commutation failure. In step 1, the constructed TriFuseNet model includes a classifier, a multi-scale feature extraction and fusion module, and two preliminary feature extraction modules; Two preliminary feature extraction modules are used to receive the time-frequency map and MTF image in the fault map group respectively, and perform preliminary feature extraction on the time-frequency map and MTF image respectively to obtain coarse-grained features in the time-frequency map and MTF image; The multi-scale feature extraction and fusion module is used to extract coarse-grained features at multiple scales to obtain fine-grained fault features, and to fuse the fine-grained fault features to obtain fault fusion features. The classifier is used to classify the fault fusion features and obtain the fault source tracing results of commutation failure.

2. The method for tracing the source of commutation failure in an HVDC system based on a dual-stream gated fusion network according to claim 1, characterized in that, The preliminary feature extraction module includes a first Conv layer and a second Conv layer; The output data of the first Conv layer is used as the input data of the second Conv layer; the first Conv layer of the two preliminary feature extraction modules is used to receive the time-frequency map and the MTF image respectively, and the second Conv layer is used to output the coarse-grained features of the time-frequency map and the MTF image. The first Conv layer is used for 5×5 convolution processing, and the second Conv layer is used for 3×3 convolution processing.

3. The method for tracing the source of commutation failure in an HVDC system based on a dual-stream gated fusion network according to claim 1, characterized in that, The multi-scale feature extraction and fusion module includes a BGF fusion module and two multi-scale feature extraction modules; The multi-scale feature extraction module includes three sequentially connected feature extraction modules (PAR Blocks), which are used to perform fine-grained fault feature extraction for three stages. The BGF fusion module is used to fuse the fine-grained fault features obtained in the same stage of two multi-scale feature extraction modules and output the fused fine-grained fault features.

4. The method for tracing the source of commutation failure in an HVDC system based on a dual-stream gated fusion network according to claim 3, characterized in that, The feature extraction module PAR Block includes a first Conv2d convolutional layer, a first PARU layer, a second PARU layer, and a first AvgPool layer. The first Conv2d convolutional layer, the first PARU layer, and the second PARU layer are connected in sequence. The input data of the first Conv2d convolutional layer is added to the output data of the second PARU layer and then input to the first AvgPool layer. The input of the first Conv2d convolutional layer serves as the input of the feature extraction module PAR Block, and the output of the first AvgPool layer serves as the output of the feature extraction module PAR Block.

5. The method for tracing the source of commutation failure in an HVDC system based on a dual-stream gated fusion network according to claim 4, characterized in that, The first and second PARU layers are PARU layers with the same structure. The PARU layer includes a pyramid convolutional layer, a second Conv2d convolutional layer, an ICBAM layer, a third Conv2d convolutional layer, and a second AvgPool layer. The pyramid convolutional layer, the second Conv2d convolutional layer, the ICBAM layer, and the third Conv2d convolutional layer are connected in sequence. The input data of the pyramid convolutional layer is added to the output data of the third Conv2d convolutional layer and then input to the second AvgPool layer. The input of the pyramid convolutional layer is used as the input of the PARU layer, and the output of the second AvgPool layer is used as the output of the PARU layer.

6. The method for tracing the source of commutation failure in an HVDC system based on a dual-stream gated fusion network according to claim 3, characterized in that, The BGF fusion module includes a first BGF layer, a second BGF layer, a third BGF layer, and a fourth BGF layer connected in sequence. The first BGF layer is used to receive the time-frequency map and the coarse-grained features of the MTF image output by the two preliminary feature extraction modules, perform fusion processing, and output the first fused feature. The second BGF layer is used to receive the first fused feature and the fine-grained fault features output by the first-stage feature extraction module PAR Block of the two multi-scale feature extraction modules, perform fusion processing, and output the second fused feature. The third BGF layer is used to receive the second fused feature and the fine-grained fault features output by the feature extraction module PAR Block of the second stage of the two multi-scale feature extraction modules, perform fusion processing, and output the third fused feature; the fourth BGF layer is used to receive the third fused feature and the fine-grained fault features output by the feature extraction module PAR Block of the third stage of the two multi-scale feature extraction modules, perform fusion processing, and output the fused fine-grained fault features.

7. The method for tracing the source of commutation failure in an HVDC system based on a dual-stream gated fusion network according to claim 6, characterized in that, The first BGF layer includes a BGF layer; the second BGF layer and the third BGF layer each include a third Conv layer, a BGF layer and a fourth Conv layer connected in sequence; the fourth BGF layer includes a third Conv layer, a BGF layer, a SimAM layer and a fourth Conv layer connected in sequence. The third Conv layer is used for 1×1 convolution processing, the fourth Conv layer is used for 3×3 convolution processing; the BGF layer is used for multimodal feature fusion; and the SimAM layer is used to mine the spatial importance within the fused feature map.

8. The method for tracing the source of commutation failure in an HVDC system based on a dual-stream gated fusion network according to claim 7, characterized in that, The specific steps for multimodal feature fusion in the BGF layer are as follows: First, global average pooling is performed on the feature maps F1 and F2 from the two modalities. and The compressed representations of the features are obtained as S1 and S2, respectively. , Let R be the set of real numbers. The global average pooling is calculated as follows: In the formula: and Let F1 and F2 represent the global average pooling results of the c-th channel at spatial location (H×W), respectively, where H represents the height of the feature map, W represents the width of the feature map, and C represents the number of feature channels. This represents the feature map on the c-th channel; The two features are then concatenated and fed into a multilayer perceptron (MLP) to model the nonlinear dependencies between channels, generating a fusion gating weight F. BGF , Then merge the gating weights F BGF After dividing the data into two feature representations, W1 and W2, two gated vectors P1 and P2 are generated using a parameterized Softmax function. The adaptive acquisition of features for each channel is as follows: , , In the formula, τ is a parameter factor used to adjust the weight distribution of the two features during fusion. The smaller τ is, the more Softmax tends to select only one branch, while the larger τ is, the more evenly Softmax fuses the two branches. This indicates a splicing operation performed at the channel level; Finally, the two gated vectors P1 and P2 are weighted onto the original modality feature map to obtain weighted fusion features, and finally, the fusion output is formed by element-wise addition: In the formula, This represents the dot product within the channel.

9. The method for tracing the source of commutation failure in an HVDC system based on a dual-stream gated fusion network according to claim 1, characterized in that, The classifier consists of a fifth Conv layer, a GAP layer, a first FC layer, and a second FC layer connected in sequence. The fifth Conv layer is used for 3×3 convolution processing, the GAP layer is used for global average pooling processing, and the first FC layer and the second FC layer are used to convert the features after global average pooling processing into classification output.

Images