Method and device for tracing commutation failure fault of high-voltage direct-current power transmission system

Abstract

The invention discloses a method and a device for tracing a commutation failure fault of a high-voltage direct-current power transmission system, and the method comprises the steps: taking three electrical quantity signals, namely an alternating-current voltage, a direct-current current and a direct-current voltage, of an inversion side of the high-voltage direct-current power transmission system as source data when the commutation failure occurs; extracting an optimal fault feature of the source data by adopting a fault feature extraction model and a feature screening algorithm based on feature density; and finally, in combination with a commutation fault detection model and the optimal fault characteristics, obtaining a fault type causing commutation failure of the high-voltage direct-current power transmission system. Through the above design, fault tracing of the commutation failure can be realized, and in the tracing process, the optimal fault features are extracted by using the machine learning model and the feature screening algorithm based on the feature density, so that interference of human subjective experience on feature extraction can be reduced, and the effectiveness of fault tracing is improved; therefore, the method is very suitable for large-scale application and popularization.

Description

Technical Field

[0001] The present invention relates to the technical field of commutation fault analysis of a high-voltage direct current (HVDC) transmission system, and in particular to a method and device for tracing a commutation failure fault of a high-voltage direct current (HVDC) transmission system. Background Art

[0002] Commutation failure is one of the most common fault types in HVDC transmission systems, typically occurring on the inverter side. Commutation failure can cause a sudden surge in DC current, impacting the converter valves and even triggering DC lockout, seriously threatening the safe and stable operation of the power grid. The AC bus voltage on the inverter side of the HVDC system, known as the commutation voltage, is the primary factor determining whether commutation failure occurs. The lower the commutation voltage, the more likely commutation failure will occur. In actual DC projects, the rapid drop in commutation voltage caused by inverter-side AC system faults is the primary cause of commutation failure. Various inverter-side AC system faults can cause commutation failure, including single-phase ground faults (AG, BG, CG), two-phase ground faults (AB-G, BC-G, CA-G), two-phase short-circuit faults (AB, BC, CA), and three-phase ground faults (ABCG).

[0003] Currently, in UHVDC projects, a commutation failure prediction control (CFPREV) link is usually configured to prevent commutation failure under external fault disturbance conditions and accelerate DC recovery after a fault. It consists of two parallel parts: one is based on zero-sequence voltage To detect single-phase faults; the other part is to transform the AC three-phase voltage to the stationary coordinate system through Clark transformation , then by To detect three-phase faults; when a single-phase fault occurs, the sum of the three-phase voltages is not zero, and the zero-sequence voltage Will increase; when three-phase fault occurs, The amplitude of the two detection indicators will decrease. Once the change in any one of the two detection indicators exceeds its corresponding threshold, it is determined that an AC fault has occurred. At this time, the commutation valve will be triggered in advance to prevent commutation failure. However, the CFPREV strategy only has the function of suppressing commutation failure and cannot achieve tracing of commutation failure. It cannot quickly determine the fault type that caused the commutation failure after it occurs.

[0004] Meanwhile, there are also some commutation failure tracing methods through neural networks, but these methods mostly need an additional human feature extraction step before realizing commutation failure tracing, so when the professional knowledge and experience are limited, the complexity of the fault mode of the HVDC system will make it challenging for humans to select the best features, that is, the feature extraction process will be affected by human experience, thereby affecting the accuracy of feature selection, and the quality of these selected features will directly affect the effectiveness of fault tracing. Therefore, based on the foregoing deficiencies, how to provide a high-accuracy high-voltage direct current power transmission system commutation failure fault tracing method has become a problem to be solved. SUMMARY

[0005] The technical problem to be solved by the present application is the commutation failure fault tracing problem of the high-voltage direct current power transmission system, and the purpose is to provide a high-voltage direct current power transmission system commutation failure fault tracing method and device, which solves the problem that the CFPREV strategy in the prior art only has the function of commutation failure suppression and cannot realize commutation failure fault tracing, and the problem that the traditional neural network is used for fault tracing and needs to extract features manually, thereby affecting the effectiveness of commutation failure fault tracing.

[0006] The present application is realized by the following technical scheme: In a first aspect, a high-voltage direct current power transmission system commutation failure fault tracing method is provided, comprising: obtaining operation data of an inverter side of the high-voltage direct current power transmission system at the time of commutation failure, wherein the operation data includes alternating current voltage, direct current voltage, and direct current of the inverter side; inputting the operation data into a fault feature extraction model to obtain first fault feature data of the direct current voltage, second fault feature data of the direct current, and third fault feature data of each phase voltage in the alternating current voltage; using a feature screening algorithm based on feature density to perform feature screening processing on the first fault feature data, the second fault feature data, and the third fault feature data of each phase voltage to obtain first optimal fault feature data of the direct current voltage, second optimal fault feature data of the direct current, and third optimal fault feature data of each phase voltage; inputting the first optimal fault feature data, the second optimal fault feature data, and the third optimal fault feature data of each phase voltage into a commutation failure detection model to obtain a commutation failure type of the high-voltage direct current power transmission system.

[0007] Based on the above disclosure, after obtaining the AC voltage, DC voltage, and DC current on the inverter side of the HVDC transmission system when a commutation failure occurs, the present invention first uses a fault feature extraction model to extract fault features from the aforementioned data, thereby obtaining fault feature data for the DC voltage, DC current, and each phase voltage in the AC voltage; then, the present invention uses a feature screening algorithm based on feature density to perform feature screening on the aforementioned fault feature data, thereby obtaining optimal fault feature data for the DC voltage, DC current, and each phase voltage; finally, the aforementioned optimal fault feature data is input into a commutation fault detection model to obtain the commutation fault type of the HVDC transmission system.

[0008] Through the above design, the present invention adopts three electrical quantity signals of AC voltage, DC current and DC voltage as source data for fault tracing, and then uses a machine learning model to extract and process the fault features to obtain corresponding fault feature data; then, a feature screening algorithm based on feature density is used to perform feature screening on the above fault features, and then the optimal fault feature data of each electrical quantity signal is obtained; finally, the optimal fault feature data of each electrical quantity signal is input into the commutation fault detection model, and the commutation fault type that causes the commutation failure of the high-voltage direct current transmission system can be obtained; based on this, compared with the traditional CFPREV strategy, the present invention can realize the fault tracing of commutation failure, and in the tracing process, a machine learning model and a feature screening algorithm based on feature density are used to extract the optimal fault features, so that the interference of human subjective experience on feature extraction can be reduced, thereby improving the effectiveness of fault tracing; therefore, the present invention is very suitable for large-scale application and promotion.

[0009] In one possible design, the operating data is input into a fault feature extraction model to obtain first fault feature data of the DC voltage, second fault feature data of the DC current, and third fault feature data of each phase voltage in the AC voltage, including: Obtaining a fault feature extraction model, wherein the fault feature extraction model includes five fault feature extraction networks, the DC voltage, the DC current, and each phase voltage of the AC voltage respectively corresponds to a fault feature extraction network, each fault feature extraction network includes a plurality of fault feature extraction units connected in sequence, and each fault feature extraction unit includes an encoder, a hidden layer, and a decoder connected in sequence; The DC voltage, the DC current, and each phase voltage of the AC voltage are respectively input into corresponding fault feature extraction networks to obtain first fault feature data of the DC voltage, second fault feature data of the DC current, and third fault feature data of each phase voltage of the AC voltage.

[0010] In a possible design, the fault feature extraction network corresponding to the DC voltage is trained in the following manner: Acquire a plurality of sample operating data of the inverter side of the high voltage direct current transmission system when commutation failure occurs, and extract a sample DC voltage from each sample operating data to form a first training set; Using the first training set to train the i-th fault feature extraction unit in the target network, and during the training process, calculating a loss function of the i-th fault feature extraction unit based on output data of the i-th fault feature extraction unit, and updating the model parameters of the i-th fault feature extraction unit based on the loss function until the loss function converges, thereby obtaining a trained i-th fault feature extraction unit, wherein the target network is any untrained fault feature extraction network; Inputting the first training set into a designated network and using the output data of the designated network to form a new training set, wherein the designated network includes the first i fault feature extraction units after training; Incrementing i by 1, updating the first training set to the new training set, and reusing the first training set to train the i-th fault feature extraction unit in the target network until i equals n, thereby obtaining an initial training feature extraction network, wherein the initial value of i is 1, and n is the total number of fault feature extraction units in the target network; The initially trained feature extraction network is subjected to supervised training to obtain a fault feature extraction network corresponding to the DC voltage after the supervised training.

[0011] In a possible design, the loss function of the i-th fault feature extraction unit is: (1) In formula (1), represents the loss function of the i-th fault feature extraction unit, represents the kth sample DC voltage in the first training set, represents the output data of the decoder in the i-th fault feature extraction unit after the k-th sample DC voltage is input to the i-th fault feature extraction unit, represents the weight matrix of the i-th fault feature extraction unit, represents a sparse constant, represents the weight decay coefficient, represents the sparse penalty coefficient, represents the total number of sample DC voltages, represents the average activation probability of the jth hidden neuron in the hidden layer of the i-th fault feature extraction unit, represents the total number of hidden neurons in the hidden layer, represents the norm operation; When the first training set is used to train the i-th fault feature extraction unit, the activation value of each hidden neuron in the hidden layer of each sample DC voltage is recorded, and the average of the activation values ​​of the j-th hidden neuron in the hidden layer of each sample DC voltage is used as the average activation probability of the j-th hidden neuron.

[0012] In one possible design, supervised training is performed on the initial training feature extraction network to obtain a fault feature extraction network corresponding to the DC voltage after the supervised training, including: randomly selecting a number of sample DC voltages from the first training set, and determining a fault category label corresponding to each of the selected sample DC voltages, so as to form a second training set using the selected number of sample DC voltages and the fault category label corresponding to each of the selected sample DC voltages, wherein the fault category label corresponding to any sample DC voltage is used to characterize a commutation fault type of the high voltage direct current transmission system; Based on the initial training feature extraction network, generating a pre-training feature extraction network, wherein the pre-training feature extraction network includes the initial training feature extraction network, the first fully connected layer and the Softmax classifier connected in sequence; The pre-trained feature extraction network is trained using each sample DC voltage in the second training set as input and the fault category probability corresponding to each sample DC voltage as output. During the training process, a cross-entropy loss function is calculated using the fault category probability and the fault category label corresponding to each sample DC voltage. The model parameters of the pre-trained feature extraction network are updated using the cross-entropy loss function, so as to obtain an optimal fault feature extraction network after the training is completed. The first fully connected layer and the Softmax classifier are removed from the optimal fault feature extraction network to obtain a fault feature extraction network corresponding to the DC voltage.

[0013] In one possible design, the first fault feature data includes a plurality of fault sub-feature data, wherein a feature screening algorithm based on feature density is used to perform feature screening on the first fault feature data to obtain first optimal fault feature data of the DC voltage, including: Calculating a local density index of each fault sub-feature data in the first fault feature data, wherein the local density index of any fault sub-feature data is used to characterize the local density of the any fault sub-feature data, and a smaller local density index indicates a greater local density; Calculate the characteristic density of each fault sub-characteristic data according to the local density index of each fault sub-characteristic data; By utilizing the characteristic density of each fault sub-characteristic data and adopting a greedy algorithm, a number of key fault sub-characteristic data are screened out from each fault sub-characteristic data to form the first optimal fault characteristic data.

[0014] In one possible design, calculating the local density index of each fault sub-feature data in the first fault feature data includes: For any fault sub-feature data, calculating a distance between the any fault sub-feature data and target data, wherein the target data is each fault sub-feature data in the first fault feature data except the any fault sub-feature data; Sort the target data in ascending order of distance to obtain a first sorting sequence; Filtering the first K target data from the first sorting sequence as the neighborhood data of any fault sub-feature data, where K is a positive integer; Calculating the distance between any fault sub-feature data and each neighboring data, and taking the mean of the sum of the distances between any fault sub-feature data and each neighboring data as the local density index; Accordingly, the characteristic density of each fault sub-characteristic data is calculated based on the local density index of each fault sub-characteristic data, which includes: For any fault sub-feature data, obtain density adjustment parameters; The product of the local density index of any fault sub-feature data and the density adjustment parameter is used as the feature density of any fault sub-feature data.

[0015] In one possible design, the characteristic density of each fault sub-characteristic data is utilized, and a greedy algorithm is adopted to screen out a number of key fault sub-characteristic data from each fault sub-characteristic data to form the first optimal fault characteristic data, including: Sorting the sub-feature data of each fault in descending order of feature density to obtain a second sorting sequence; removing the first fault sub-feature data from the second sorting sequence to obtain a third sorting sequence; Initialize the number of iterations s to 1, and put the first fault sub-feature data in the second sorting sequence into the initial sample representative set to obtain the representative sample set at the s-1th iteration; Filtering out designated data from the representative sample set at the s-1th iteration, wherein the designated data is the most recently added feature data in the representative sample set at the s-1th iteration; Selecting the sth fault sub-feature data from the third sorting sequence; Calculating the distance between the s-th fault sub-feature data and the specified data, and selecting the minimum feature density from the feature density of the s-th fault sub-feature data and the feature density of the specified data; Determine whether the distance between the s-th fault sub-feature data and the specified data is less than the minimum feature density; If not, the s-th fault sub-feature data is taken as a key fault sub-feature data and put into the representative sample set at the s-1-th iteration to obtain the representative sample set at the s-th iteration; Increment s by 1 and reselect specified data from the representative sample set at the s-1th iteration until s equals S, so as to utilize the key fault sub-feature data in the representative sample set at the sth iteration to form the first optimal fault feature data, where S is the total number of fault sub-feature data in the third sorting sequence.

[0016] In one possible design, the commutation fault detection model includes a feature extraction module, a feature enhancement module, and an output module connected in sequence; The feature extraction module is configured to perform secondary feature extraction processing on the input data using one-dimensional convolution to obtain a commutation fault feature vector, wherein the input data includes the first optimal fault feature data, the second optimal fault feature data, and the third optimal fault feature data of each phase voltage; A feature enhancement module is used to perform feature enhancement processing on the commutation fault feature vector using a global attention mechanism to obtain an enhanced fault feature vector; an output module, configured to perform feature mapping processing on the enhanced fault feature vector to obtain probability values ​​of the operating data belonging to various commutation fault types, and to use the commutation fault type corresponding to the maximum probability value as the commutation fault type of the high voltage direct current transmission system.

[0017] In a second aspect, a device for tracing a commutation failure fault in a high-voltage direct current transmission system is provided, comprising: an acquiring unit, configured to acquire operating data of the inverter side of the HVDC transmission system when a commutation failure occurs, wherein the operating data includes an AC voltage, a DC voltage, and a DC current of the inverter side; a feature extraction unit, configured to input the operating data into a fault feature extraction model to obtain first fault feature data of the DC voltage, second fault feature data of the DC current, and third fault feature data of each phase voltage in the AC voltage; a feature screening unit, configured to perform feature screening processing on the first fault feature data, the second fault feature data, and the third fault feature data of each phase voltage using a feature screening algorithm based on feature density, so as to obtain first optimal fault feature data of the DC voltage, second optimal fault feature data of the DC current, and third optimal fault feature data of each phase voltage; The fault tracing unit is used to input the first optimal fault characteristic data, the second optimal fault characteristic data and the third optimal fault characteristic data of each phase voltage into a commutation fault detection model to obtain the commutation fault type of the high voltage direct current transmission system.

[0018] Compared with the prior art, the present invention has the following advantages and beneficial effects: (1) Compared with the traditional CFPREV strategy, the present invention can realize the fault tracing of commutation failure. In the tracing process, a machine learning model and a feature screening algorithm based on feature density are used to extract the optimal fault features. In this way, the interference of human subjective experience on feature extraction can be reduced, thereby improving the effectiveness of fault tracing. Therefore, the present invention is very suitable for large-scale application and promotion. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the following briefly introduces the drawings required for use in the examples. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be considered as limiting the scope. A person of ordinary skill in the art can also derive other relevant drawings based on these drawings without inventive effort. In the drawings: Figure 1 A schematic flow chart of the steps of a method for tracing a commutation failure fault in a high-voltage direct current transmission system provided by an embodiment of the present invention; Figure 2 Flowchart of the model construction and offline training phase provided by an embodiment of the present invention; Figure 3 A flowchart of the online real-time tracing phase provided by an embodiment of the present invention; Figure 4 A network structure diagram of a fault characteristic network provided by an embodiment of the present invention; Figure 5 A schematic diagram of the structure of a commutation fault detection model provided by an embodiment of the present invention; Figure 6 A network structure diagram of a feature extraction module provided in an embodiment of the present invention; Figure 7 A network structure diagram of a feature extraction subunit provided in an embodiment of the present invention; Figure 8A network structure diagram of a feature enhancement module provided in an embodiment of the present invention; Figure 9 A network structure diagram of a channel attention unit provided in an embodiment of the present invention; Figure 10 A network structure diagram of a temporal attention unit provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0020] Example: See also Figure 1 As shown, the HVDC system commutation failure fault tracing method provided by this embodiment uses the three electrical quantity signals of AC voltage, DC current and DC voltage on the inverter side of the HVDC system when the commutation failure occurs as source data, and then adopts a fault feature extraction model and a feature screening algorithm based on feature density to extract the optimal fault features of the above source data; finally, combining the commutation fault detection model and the above optimal fault features, the fault type that causes the commutation failure of the HVDC system can be obtained; through the above design, this method can not only realize the commutation failure fault tracing, but also in the tracing process. In the method, a machine learning model and a feature screening algorithm based on feature density are used to extract the optimal fault features. In this way, the interference of human subjective experience on feature extraction can be reduced, thereby improving the effectiveness of fault tracing. Therefore, this method is very suitable for large-scale application and promotion. For example, this method can be but not limited to running on the power grid operation and maintenance end side. Optionally, the power grid operation and maintenance end side can be but not limited to using a server or computer. It can be understood that the aforementioned execution subject does not constitute a limitation on the embodiments of the present application. Accordingly, the operation steps of this method can be but not limited to the following steps S1 to S4.

[0021] S1. Obtaining operating data of the inverter side of the HVDC transmission system when a commutation failure occurs, wherein the operating data includes the AC voltage, DC voltage, and DC current of the inverter side. In this embodiment, for example, but not limited to, when a commutation failure occurs in the HVDC transmission system, a voltage sensor and a current sensor are used to collect the DC voltage, DC current, and AC voltage of the inverter side (collection can be set to a sampling period), and the AC voltage includes the a-phase voltage, the b-phase voltage, and the c-phase voltage. In this way, this embodiment uses the aforementioned AC voltage, DC voltage, and DC current as traceability data, extracts the optimal fault characteristics of the traceability data, and uses a neural network model to achieve commutation failure traceability.

[0022] Optionally, this embodiment first extracts the fault feature data of the aforementioned traceability data, and then uses a feature screening algorithm to screen out the optimal fault feature from the extracted fault features; wherein the fault feature extraction process is shown in the following step S2.

[0023] S2. Inputting the operating data into a fault feature extraction model to obtain first fault feature data of the DC voltage, second fault feature data of the DC current, and third fault feature data of each phase voltage of the AC voltage. In specific applications, this embodiment constructs a fault feature extraction model for each electrical quantity data, namely: the fault feature extraction model includes five fault feature extraction networks, see Figure 4 As shown, each fault feature extraction network includes several fault feature extraction units connected in sequence, and any fault feature extraction unit includes an encoder, a hidden layer and a decoder connected in sequence, and each phase voltage in the DC voltage, the DC current and the AC voltage corresponds to a fault feature extraction network respectively; in this way, it is equivalent to that the phase a voltage, the phase b voltage, the phase c voltage, the DC voltage and the DC current correspond to a fault feature extraction network respectively; based on this, in actual use, the DC voltage, the DC current and each phase voltage in the AC voltage are respectively input into the corresponding fault feature extraction network, and the corresponding fault feature data can be obtained.

[0024] In this embodiment, the aforementioned fault feature extraction network is equivalent to being composed of multiple sparse autoencoders stacked together to form a deep network structure, wherein the output of each fault feature extraction unit (i.e., sparse autoencoder) serves as the input of the next fault feature extraction unit, thereby performing preliminary feature extraction and data dimensionality reduction on the data input to the network; wherein, the connection diagram of the encoder, hidden layer, and decoder can be seen in Figure 4 shown.

[0025] Furthermore, before the aforementioned operating data is input into the corresponding fault feature extraction network, it can be normalized to unify the data. For example, the min-max normalization method can be used to normalize the DC voltage, DC current, and each phase voltage of the AC voltage collected during the aforementioned sampling period. For example, for a DC current (called a sampling point) within a sampling period, the minimum DC current within the sampling period is subtracted from the sampling point to obtain a first result. Then, the ratio of the first result to the difference between the maximum DC current and the minimum DC current within the sampling period is used as the normalization result. Of course, the normalization process of the remaining data is the same, and will not be repeated here.

[0026] Furthermore, in this embodiment, the number of layers of encoders, hidden layers and decoders in each fault feature extraction unit, and the number of neurons in each layer of encoders, decoders and hidden layers play a decisive role in the quality of feature extraction capabilities. Therefore, for example, but not limited to, the number of layers and the number of neurons can be used as parameters to be optimized; then, the Bayesian optimization method is used to perform parameter optimization to determine the optimal structure of the fault feature extraction unit. Of course, the Bayesian optimization method is a commonly used technology for model structure optimization, and its principle will not be repeated here.

[0027] After the network structure is determined, network training is required. In this embodiment, each fault feature extraction network needs to be trained with the corresponding traceability data, that is, the five fault feature extraction networks correspond to phase a voltage, phase b voltage, phase c voltage, DC voltage and DC current respectively. Therefore, it is necessary to use sample DC voltage, sample DC current, sample phase a voltage, sample phase b voltage and sample phase c voltage for separate training, so as to obtain a fault feature extraction network for each traceability data. The schematic diagram of the training stage can be seen in Figure 2 shown.

[0028] The following uses DC voltage as an example to illustrate the specific training process of the fault feature extraction network, as shown below: S21. Acquire several sample operating data of the inverter side of the HVDC transmission system during commutation failure, and extract a sample DC voltage from each sample operating data to form a first training set. In this embodiment, single-phase grounding faults (AG, BG, CG), two-phase grounding faults (AB-G, BC-G, CA-G), two-phase short-circuit faults (AB, BC, CA), and three-phase grounding faults (ABCG) can be set on the AC busbar on the inverter side of the HVDC transmission system simulation model. The above fault settings can be applied to each phase to obtain 10 types of commutation failures caused by AC faults, which are represented by labels 0-9. Simultaneously, different operating condition parameters, such as fault inductance and fault initial angle, are set for simulation, and three electrical quantity data, namely, AC voltage, DC voltage, and DC current, are collected on the inverter side. In this way, several sample operating data of the inverter side during commutation failure can be obtained, namely, the aforementioned sample DC voltage ( )、sample DC current( )、sample a phase voltage( )、sample b phase voltage( ) and sample c-phase voltage ( ).

[0029] At the same time, as mentioned above, each fault feature extraction network requires corresponding traceability data for training. Therefore, when training fault feature data for extracting DC voltage, it is necessary to extract sample DC voltages from a number of sample operating data, that is, the sample DC voltages corresponding to different commutation fault types of the high-voltage DC transmission system, that is, the DC voltages under 10 types of faults such as phase A grounding, phase B grounding, and phase AB grounding.

[0030] After obtaining the first training set, the first stage of training can be carried out, that is, unsupervised greedy layer-by-layer training, and the process is shown in the following steps S22 to S24.

[0031] S22. Use the first training set to train the i-th fault feature extraction unit in the target network. During the training process, calculate the loss function of the i-th fault feature extraction unit based on the output data of the i-th fault feature extraction unit, and update the model parameters of the i-th fault feature extraction unit based on the loss function until the loss function converges, thereby obtaining the trained i-th fault feature extraction unit, wherein the target network is any untrained fault feature extraction network. In this embodiment, it is also necessary to perform min-max normalization processing on the DC voltage of each sample in the first training set, and then input the normalized first training set into the i-th fault feature extraction unit in any untrained fault feature extraction network to perform feature extraction training.

[0032] During specific training, the data processing process of the encoder, hidden layer and decoder in the i-th fault feature extraction unit is as follows: First, the normalized sample DC voltage is used as the input of the encoder. After being encoded by the encoder, it is output to the hidden layer to obtain the encoding result of the hidden layer, which is expressed as: (2) In formula (2), represents the output of the hidden layer in the i-th fault feature extraction unit, represents the weight and bias of the encoder in the i-th fault feature extraction unit, represents the sigmoid activation function, represents the kth sample DC voltage in the first training set (normalized data).

[0033] Then, the output of the hidden layer is input into the corresponding decoder, and the decoding function is used to reconstruct the output of the hidden layer into the original data dimension. The output of the decoder is: (3) In formula (3), represents the output data of the decoder in the ith fault feature extraction unit after the kth sample direct current voltage is input into the ith fault feature extraction unit, represents the weight and bias of the decoder.

[0034] Thus, the input data can be mapped to a low-dimensional latent space (i.e. hidden layer) by the encoder to extract features by a nonlinear transformation; then, the original data is reconstructed from the hidden layer features; in this embodiment, by introducing a sparsity constraint in the hidden layer, most neurons are forced to output 0, and only a few neurons are activated, so that the key features of the input data are learned; based on this, the working principle of the fault feature extraction unit of this embodiment is that after the input data is nonlinearly transformed by the encoder, only a few key features (non-zero activation values) are retained in the hidden layer, realizing low-dimensional sparse representation of high-dimensional sparse data.

[0035] Thus, the training of the ith fault feature extraction unit can be realized by optimizing the reconstruction error through back propagation, that is, the weights and biases in the aforementioned encoder and decoder are updated based on the corresponding loss function and using the back propagation algorithm and gradient descent algorithm until the loss function converges.

[0036] Optionally, the following disclosure shows the loss function representation form of the ith fault feature extraction unit as follows: (1) In formula (1), represents the loss function of the ith fault feature extraction unit, represents the kth sample direct current voltage in the first training set, represents the output data of the decoder in the ith fault feature extraction unit after the kth sample direct current voltage is input into the ith fault feature extraction unit, represents the weight matrix of the ith fault feature extraction unit (i.e. the weights of the aforementioned encoder and decoder), represents a sparse constant, represents a weight decay coefficient, represents a sparse penalty coefficient, represents the total number of sample direct current voltages, represents the average activation probability of the jth hidden neuron in the hidden layer in the ith fault feature extraction unit, represents the total number of hidden neurons in the hidden layer, represents a norm operation.

[0037] Furthermore, when the first training set is used to train the i-th fault feature extraction unit, the model records the activation values ​​of each hidden neuron in the hidden layer for each sample DC voltage. Therefore, the mean of the activation values ​​of the j-th hidden neuron in the hidden layer for each sample DC voltage can be used as the average activation probability of the j-th hidden neuron.

[0038] Thus, it can be seen from the aforementioned loss function that this embodiment uses reconstruction error plus weight regularization and sparsity constraint to form the loss function of the i-th fault feature extraction unit, wherein the reconstruction error is used to evaluate the encoding-decoding reconstruction capability of the i-th fault feature extraction unit, and the weight regularization is used to impose constraints on the weights of the i-th fault feature extraction unit to prevent overfitting during the training process; and the sparsity constraint forces the codec to activate only a few neurons when representing data, so that the network learns more discriminative features.

[0039] Based on this, the i-th fault feature extraction unit can be used to learn the key features of the input sample DC voltage, that is, the fault features; and after completing the training of the i-th fault feature extraction unit, the training of the codec of the next layer can be carried out, and the process is shown in the following step S23.

[0040] S23. Input the first training set into the designated network, and use the output data of the designated network to form a new training set, where the designated network includes the first i fault feature extraction units after training. In this embodiment, this is equivalent to using the output of the trained first-layer codec (i.e., the i-th fault feature extraction unit) as the input, i.e., training data, of the next-layer codec (i.e., the second fault feature extraction unit). The aforementioned training process is then repeated until the last fault feature extraction unit is polled. At this point, unsupervised training is completed.

[0041] The greedy layer-by-layer unsupervised training process is shown in the following step S24.

[0042] S24. Increment i by 1, update the first training set to the new training set, and reuse the first training set to train the i-th fault feature extraction unit in the target network until i equals n, thereby obtaining an initial training feature extraction network, wherein the initial value of i is 1, and n is the total number of fault feature extraction units in the target network; in this embodiment, it is equivalent to using the output data of the trained codec of the previous layer as the input data of the codec of the next layer. For example, when i is equal to 1, the designated network includes the first fault feature extraction unit after training. Based on this, the first training set is input into the first fault feature extraction unit to obtain a new training set, and then the new training set is used to train the network. The second fault feature extraction unit; when the second fault feature extraction unit is trained, the specified network includes the first fault feature extraction unit and the second fault feature extraction unit; therefore, when training the third fault feature extraction unit, the first training set is input into the trained first fault feature extraction unit, and after encoding and decoding by the first fault feature extraction unit, the obtained features are then input into the trained second fault feature extraction unit for encoding and decoding; at this time, the output of the trained second fault feature extraction unit is used as the training data (i.e., input data) of the third fault feature extraction unit; thus, by continuously training according to this principle, the unsupervised training of the target network can be completed.

[0043] After completing the unsupervised training of the target network, in order to make the extracted fault features more conducive to the subsequent fault classification task, this embodiment also provides supervised training, that is, the input data is labeled data; wherein, the supervised training process is shown in the following step S25.

[0044] S25. Performing supervised training on the initial training feature extraction network to obtain a fault feature extraction network corresponding to the DC voltage after the supervised training. In a specific application, for example, but not limited to, the following steps S25a to S25d can be used to complete the supervised training of the initial training feature extraction network.

[0045] S25a. Randomly select a number of sample DC voltages from the first training set, and determine a fault category label corresponding to each selected sample DC voltage. A second training set is formed using the selected sample DC voltages and the fault category labels corresponding to each selected sample DC voltage. The fault category label corresponding to any sample DC voltage is used to characterize the type of commutation fault in the HVDC transmission system. In specific applications, the second training set can be formed by, but is not limited to, randomly selecting 1% of the data from the first training set and adding corresponding fault category labels. The fault category labels are the aforementioned 10 types of faults. For example, if a sample DC voltage is 100, the HVDC transmission system experienced a commutation failure due to a phase A ground fault; therefore, its fault category label is a phase A ground fault. The same process is used to determine the fault category labels for the remaining sample DC voltages in the first training set, and no further details are given here.

[0046] After the supervised training training set is constructed, the network structure can be updated, and the process is shown in the following step S25b.

[0047] S25b. Based on the initial training feature extraction network, a pre-trained feature extraction network is generated, wherein the pre-trained feature extraction network includes the initial training feature extraction network, the first fully connected layer, and the Softmax classifier connected in sequence. In a specific application, the aforementioned trained fault feature networks are connected in sequence to form the initial training feature extraction network. Then, the first fully connected layer is connected to the output end of the initial training feature extraction network, and the first fully connected layer is connected to the Softmax classifier, thereby facilitating fault classification.

[0048] In this way, after the pre-trained feature extraction network is generated, supervised training can be performed, as shown in the following step S25c.

[0049] S25c. The pre-trained feature extraction network is trained using the DC voltage samples in the second training set as input and the fault category probabilities corresponding to the DC voltage samples as output. During the training process, a cross-entropy loss function is calculated using the fault category probabilities and fault category labels corresponding to the DC voltage samples. The cross-entropy loss function is used to update the model parameters of the pre-trained feature extraction network, so as to obtain an optimal fault feature extraction network after the training is completed.

[0050] In this embodiment, each sample DC voltage in the second training set is input into the pre-trained feature extraction network to obtain the fault category probability predicted by each sample DC voltage. Then, the cross-entropy loss function is calculated in combination with the fault category label of each sample DC voltage. Finally, the cross-entropy loss is used to reversely adjust the weights and biases of each codec layer in the pre-trained feature extraction network until the cross-entropy loss function converges. At this point, each codec in the pre-trained feature extraction network can learn fault features that are more conducive to subsequent fault classification tasks.

[0051] In this way, after completing the supervised training, the aforementioned first fully connected layer and Softmax classifier can be removed to obtain a fault feature extraction network corresponding to the DC voltage, and the process is shown in the following step S25d.

[0052] S25d. Remove the first fully connected layer and the Softmax classifier from the optimal fault feature extraction network to obtain a fault feature extraction network corresponding to the DC voltage.

[0053] Through the above steps S21 to S25 and their sub-steps, the present invention adopts a combination of unsupervised and supervised training methods to train the fault feature extraction network. The role of the fault feature extraction network is to preliminarily extract fault features to serve the subsequent deep feature extraction and final classification tasks; the purpose of supervised training is to make the low-dimensional features extracted by the fault feature extraction network more discriminative, thereby being more conducive to subsequent fault classification tasks; therefore, supervised training uses the back propagation algorithm to transfer the cross-entropy loss gradient to the entire network, and jointly adjusts the weights and biases of each layer in the encoder until the cross-entropy loss function converges, then the fault feature extraction network corresponding to the DC voltage can be obtained; of course, the training process of the fault feature extraction network corresponding to the remaining traceability data is the same, and its schematic diagram can be seen in Figure 2 shown.

[0054] After the fault feature extraction model is trained, each electrical quantity data in the operating data is input into the corresponding fault feature extraction network, and the fault feature data corresponding to each electrical quantity data can be obtained. That is, the fault feature extraction model is used to perform preliminary feature extraction. The schematic diagram can be seen in Figure 3 As shown; at the same time, the fault feature data of each electrical quantity data adopts the low-dimensional encoding in the encoder hidden space (dimension is d) to represent the fault feature data, which contains several fault sub-feature data, and the dimension of each fault sub-feature data is d (a d-dimensional vector).

[0055] In this way, after the fault feature extraction is completed, feature screening can be performed, and the process is shown in the following step S3.

[0056] S3. Using a feature density-based feature screening algorithm, the first fault feature data, the second fault feature data, and the third fault feature data for each phase voltage are subjected to feature screening processing to obtain the first optimal fault feature data for DC voltage, the second optimal fault feature data for DC current, and the third optimal fault feature data for each phase voltage. In this embodiment, due to the complexity of HVDC system fault morphology, the low-dimensional representation after preliminary feature extraction by the fault feature extraction model may still contain some sample aliasing and high similarity, which is not conducive to subsequent in-depth feature extraction and classification. Therefore, this embodiment also provides a feature screening process before fault classification to select features with high resolution and distinct characteristics. A schematic diagram of this process can be seen in FIG. Figure 3 shown.

[0057] Since the optimal fault feature screening process for each electrical quantity data is the same, the following description will be made using the first fault feature data as an example, and the corresponding feature screening process may be, but is not limited to, steps S31 to S33 as shown below.

[0058] S31. Calculate a local density index for each fault sub-feature data item in the first fault feature data item, where the local density index for each fault sub-feature data item is used to characterize the local density of the fault sub-feature data item, and a smaller local density index indicates a greater local density. In this embodiment, the calculation process of the local density index is described using any fault sub-feature data item as an example, and may be, but is not limited to, steps S31a to S31d as described below.

[0059] S31a. For any fault sub-feature data, calculate the distance between the fault sub-feature data and target data, where the target data is each fault sub-feature data in the first fault feature data excluding the fault sub-feature data. In this embodiment, this is equivalent to calculating the distance between the fault sub-feature data and each of the remaining fault sub-feature data. As previously explained, the dimension of each fault sub-feature data is d. Therefore, the distance can be calculated using the following formula: (4) In formula (4), represents the distance between any fault sub-feature data and the cth target data, represents the data of the qth dimension in any fault sub-feature data, Represents the q-th dimension data in the c-th target data.

[0060] Thus, after calculating the distance between any fault sub-feature data and each target data based on the aforementioned formula (4), the local density index can be calculated based on the distance, and the process is shown in the following steps S31b to S31d.

[0061] S31b. Sort each target data in ascending order of distance to obtain a first sorted sequence. In this embodiment, after sorting each target data in descending order of distance, the neighborhood data of any fault sub-feature data can be filtered out. The process is shown in step S31c below.

[0062] S31c. Filter the first K target data from the first sorted sequence as the neighborhood data of any fault sub-feature data, where K is a positive integer. In specific applications, K may be specifically determined based on actual use and is not specifically limited here.

[0063] In this way, after obtaining the neighborhood data of any fault sub-feature data, the local density index can be calculated based on the neighborhood data. The calculation process is shown in the following step S31d.

[0064] S31d. Calculate the distance between any one of the fault sub-feature data and each of the neighboring data, and use the average of the sum of the distances between any one of the fault sub-feature data and each of the neighboring data as the local density index. In this embodiment, the average distance between any one of the fault sub-feature data and each of the neighboring data can reflect the density of the region in which any one of the fault sub-features is located in the latent space. That is, the smaller the average distance (referring to the local density index), the denser the surrounding samples are (i.e., the greater the local density); conversely, the larger the average distance, the sparser the samples in the region.

[0065] Therefore, through the aforementioned steps S31a to S31d, a local density index for characterizing the local density of the sample can be calculated, and then, the characteristic density of each fault sub-characteristic data can be calculated based on the local density index.

[0066] The calculation process of the characteristic density is shown in step S32 below: S32. Calculate the characteristic density of each fault sub-feature data item based on the local density index of each fault sub-feature data item. In this embodiment, for any fault sub-feature data item, a density adjustment parameter may be first obtained, but is not limited to being obtained. Then, the product of the local density index of the fault sub-feature data item and the density adjustment parameter is used as the characteristic density of the fault sub-feature data item. The density adjustment parameter is used to control the strictness of the similarity of the fault sub-feature data item. The smaller the value of the density adjustment parameter, the stricter the limitation on similar fault sub-feature data item.

[0067] Thus, based on the aforementioned step S32, after the characteristic density of each fault sub-feature data is calculated, the greedy strategy can be combined to select the optimal fault feature, and the process is shown in the following step S33.

[0068] S33. Using the characteristic density of each fault sub-feature data and a greedy algorithm, a number of key fault sub-feature data are screened from each fault sub-feature data to form the first optimal fault feature data. In a specific application, the screening process of key fault sub-feature data may be, but is not limited to, as shown in steps S33a to S33i below.

[0069] S33a. Sort the fault sub-feature data in descending order of feature density to obtain a second sorted sequence. In this embodiment, after sorting the fault sub-feature data in descending order of feature density, a greedy strategy is used to select data with distinct features, namely, key fault sub-feature data. The selection process is as follows.

[0070] S33b. Remove the first fault sub-signature data from the second sorted sequence to obtain a third sorted sequence. In this embodiment, removing the first fault sub-signature data from the second sorted sequence serves to use the data with the lowest feature density to initially represent the sample set. This allows subsequent determination of whether each fault sub-signature data in the third sorted sequence is critical data by using the feature data included in the representative sample set. The screening process is illustrated in steps S33c through S33i.

[0071] S33c. Initialize the number of iterations, s, to 1, and add the first fault sub-feature data in the second sorted sequence to the initial sample representative set, obtaining the representative sample set at the s-1th iteration. In this embodiment, this corresponds to the representative sample set at iteration 0 containing the fault sub-feature data with the lowest feature density (assuming this is fault sub-feature data H). Then, determine the most recently added data from the representative sample set at iteration s-1, so that this data can be used to filter out the key fault sub-feature data from the third sorted sequence. This process is illustrated in steps S33d through S33i.

[0072] S33d. Filter out designated data from the representative sample set at the s-1th iteration, where the designated data is the most recently added feature data in the representative sample set at the s-1th iteration. In a specific application, when s is 1, the most recently added feature data in the representative sample set at the 0th iteration is the aforementioned fault sub-feature data H. Then, based on the fault sub-feature data H, it can be determined whether the sth fault sub-feature data in the third sorting sequence is the critical fault sub-feature data. The determination process is shown in the following steps S33e to S33h.

[0073] S33e. Select the sth fault sub-feature data from the third sorting sequence.

[0074] S33f. Calculate the distance between the sth fault sub-feature data and the specified data, and select the smallest feature density from the feature density of the sth fault sub-feature data and the feature density of the specified data.

[0075] After calculating the distance between the two feature data to be compared and the minimum feature density between the two, it is possible to determine whether the sth fault sub-feature data is the critical fault sub-feature data based on this. The determination process is shown in the following steps.

[0076] S33g. Determine whether the distance between the s-th fault sub-feature data and the designated data is less than the minimum feature density. In a specific application, if the s-th fault sub-feature data and the designated data in the representative sample set at the s-1-th iteration meet the aforementioned conditions, then the s-th fault sub-feature data is too similar to the selected samples and the characteristics are not obvious. Therefore, it is necessary to skip the s-th fault sub-feature data, that is, the s-th fault sub-feature data should be discarded. Otherwise, it means that the two are not similar and the distinguishing characteristics are obvious. It can be selected as a key fault sub-feature data and included in the representative sample set. The aforementioned selection process is shown in the following step S33h.

[0077] S33h. If not, the s-th fault sub-feature data is taken as a key fault sub-feature data and put into the representative sample set at the s-1-th iteration to obtain the representative sample set at the s-th iteration.

[0078] In this way, through the above steps, the selection of key fault sub-feature data can be completed. Among them, if the s-th fault sub-feature data is determined to be the key fault sub-feature data and is placed in the representative sample set, the representative sample set of the first iteration can be obtained. Then, it is necessary to filter out the latest feature data from the representative sample set of the first iteration (at this time, it is the first fault sub-feature data in the third sorting sequence, assuming it is data L), then the designated data is the fault sub-feature data L; of course, if it is determined that the s-th fault sub-feature data does not meet the above conditions, the designated data in the representative sample set of the first iteration is still the fault sub-feature data H.

[0079] Thus, after completing the selection of key fault sub-feature data once, the next fault sub-feature data in the third sorting sequence can be selected, and the aforementioned process is repeated until all data in the third sorting sequence are polled, and the key fault sub-feature data in the first fault feature data can be screened out; wherein, the cyclic selection process is shown in the following steps.

[0080] S33i. Increment s by 1 and reselect specified data from the representative sample set at the s-1th iteration until s equals S, thereby utilizing the key fault sub-feature data in the representative sample set at the sth iteration to form the first optimal fault feature data, where S is the total number of fault sub-feature data in the third sorted sequence.

[0081] Therefore, through the aforementioned S31 to S33 and their sub-steps, the optimal fault feature data can be screened out from various fault feature data based on feature density and using a greedy strategy; thus, compared with traditional manual selection, this embodiment uses an autoencoder network and a feature screening algorithm based on feature density and a greedy strategy to perform feature screening, which can reduce the interference of human subjective experience on feature extraction, thereby improving the effectiveness of fault tracing.

[0082] After obtaining the optimal fault characteristics of the DC voltage, DC current, and each phase voltage of the AC voltage, they can be input into the commutation fault detection model to derive the fault type that causes the commutation failure of the HVDC transmission system. The process is shown in the following step S4.

[0083] S4. Input the first optimal fault feature data, the second optimal fault feature data, and the third optimal fault feature data of each phase voltage into the commutation fault detection model to obtain the commutation fault type of the HVDC transmission system. In this embodiment, after feature screening, the five optimal fault feature data are obtained. Then, the five optimal fault features can be spliced ​​to obtain a multi-source feature matrix. In this way, the matrix can be used as input data and input into the commutation fault detection model to obtain the commutation fault type. Optionally, see Figure 3 As shown in the figure, after the aforementioned input data is input into the commutation fault detection model, the DenseNet network integrated with the attention mechanism is first used to perform deep feature extraction. Then, the classifier is used to output the final commutation fault type, that is, the fault tracing result.

[0084] Optionally, one of the following network structures is provided for a commutation fault detection model: See also Figure 5As shown, the commutation fault detection model can include, but is not limited to, a feature extraction module, a feature enhancement module, and an output module connected in sequence; wherein, in actual application, the feature extraction module is used to use one-dimensional convolution to perform secondary feature extraction processing on the input data to obtain a commutation fault feature vector (the aforementioned input data includes the first optimal fault feature data, the second optimal fault feature data, and the third optimal fault feature data of each phase voltage); thus, the feature extraction module mainly performs in-depth feature extraction on the input data to provide an accurate data basis for subsequent fault classification.

[0085] In specific applications, after completing the deep feature extraction of the input data, this embodiment uses a global attention mechanism to improve feature selection and global perception capabilities, that is, through channel and time series weighting, to highlight key features, thereby weakening redundant or irrelevant information, and thus improving the classification accuracy of the subsequent output module, that is: the feature enhancement module is used to use the global attention mechanism to perform feature enhancement processing on the commutation fault feature vector to obtain an enhanced fault feature vector; finally, the output module is used to perform feature mapping processing on the enhanced fault feature vector to obtain the probability value of the operating data belonging to each type of commutation fault type, and the commutation fault type corresponding to the maximum probability value is used as the commutation fault type of the high-voltage direct current transmission system.

[0086] For further information, see Figure 6 and Figure 7 As shown, one of the network structures of the following public feature extraction modules: In this embodiment, the feature extraction module may include, but is not limited to, a first feature extraction unit, a second feature extraction unit, a third feature extraction unit, and a fourth feature extraction unit; wherein each feature extraction unit includes a number of feature extraction sub-units, and the feature sub-units in each feature extraction unit are densely connected, that is, each feature sub-unit is densely connected to all subsequent layers and serves as the input of the subsequent layers to achieve feature reuse.

[0087] Furthermore, for example, the first feature extraction unit includes six feature extraction subunits, the second feature extraction unit includes 12 feature extraction subunits, the third feature extraction unit includes 24 feature extraction subunits, and the fourth feature extraction unit includes 16 feature extraction subunits. Figure 6 shown.

[0088] For specific applications, see Figure 6 As shown, for example, the output end of the first feature extraction unit is connected to the first feature splicing unit ( Figure 6 middle, The input end of the first feature compression unit is connected with the output end of the first feature extraction unit, and the output end of the first feature compression unit is connected with the input end of the second feature extraction unit. The output end of the second feature extraction unit is connected with the input end of the second feature compression unit through the second feature splicing unit. Similarly, the output end of the second feature compression unit is connected with the input end of the third feature extraction unit, the output end of the third feature extraction unit is connected with the input end of the third feature compression unit through the third feature splicing unit, the output end of the third feature compression unit is connected with the input end of the fourth feature extraction unit, and the output end of the fourth feature extraction unit is connected with the input end of the feature enhancement module through the fourth feature splicing unit. In addition, referring to Figure 6 As shown in the figure, the input end of the first feature extraction unit and the input end of each feature splicing unit are used to receive the input data, and the output end of any feature compression unit is also connected with the input end of each feature splicing unit after the any feature compression unit. In this way, through the foregoing feature extraction module, deep feature extraction of the input data can be realized.

[0089] Optionally, referring to Figure 7 As shown in the figure, for example, any feature extraction subunit includes a first BN layer, a first ReLU layer, a first 1D convolution layer, a second BN layer, a second ReLU layer and a second 1D convolution layer connected in sequence, wherein the first 1D convolution layer adopts a 1*1 convolution kernel, and the second 1D convolution layer adopts a 3*1 convolution kernel. Meanwhile, referring to Figure 6 As shown in the figure, for example, any feature compression unit includes a third BN layer, a third ReLU layer, a third 1D convolution layer and an average pooling layer connected in sequence, wherein the third 1D convolution layer adopts a 1*1 convolution kernel, and the sampling step of the average pooling layer is 2.

[0090] In this way, through the foregoing detailed structure of the feature extraction module, this embodiment is equivalent to using four dense blocks (i.e., the foregoing first to fourth feature extraction units) + three transition layers (i.e., the first to third feature compression units) to form an improved DenseNet network, that is, the core of the feature extraction module is four dense blocks, each dense block contains a plurality of dense layers (i.e., the foregoing feature extraction subunit), each dense layer is connected to all subsequent layers in a dense manner and serves as the input of the next layer to realize feature reuse, and the output of each layer is as follows: (5) In formula (5), represents the input of the rth dense layer in a dense block, represents the feature mapping connection of the 0th, 1st,..., r-1th layer, represents a nonlinear transformation function.

[0091] Furthermore, the number of dense layers is used to control the depth of the entire feature extraction module, see Figure 7 As shown, each dense layer is divided into a bottleneck layer of 1×1Conv for dimensionality reduction, and is preceded and followed by a BN layer and a ReLU layer for normalization and nonlinear activation (i.e., the first BN layer, the second BN layer, the first ReLU layer, and the second ReLU layer). Finally, 3×1Conv (i.e., the second 1D convolution layer) is used for local feature extraction, while keeping the length of the input sequence unchanged. Finally, the current dense layer input features and the new features obtained by the above operation are spliced ​​in the channel dimension to achieve dense connection. In this way, this embodiment uses one-dimensional convolution to extract features from the input data to form a 1D-DenseNet to adapt to the dimension of the time series data. Based on this, the network can effectively capture the fine-grained information in the time series data, thereby improving the feature extraction capability.

[0092] Also, see Figure 6 As shown in the figure, two adjacent dense blocks (i.e., feature extraction units) are connected using a feature concatenation layer and a transition layer (i.e., feature compression unit); the transition layer has three layers, each of which consists of a BN layer (the third BN layer mentioned above), a ReLU layer (the third ReLU layer mentioned above), a 1DConv layer (the third 1D convolution layer mentioned above), and an average pooling layer AvgPool. AvgPool calculates the average value of all features in the sliding window as the output. The transition layer reduces the number of channels and halves the number of channels through a 1×1 Conv layer and an average pooling layer with a stride of 2, respectively, to compress the model and control the complexity of the model.

[0093] Therefore, the working process of the aforementioned feature extraction module is: The input data is input to the first feature extraction unit for feature extraction, and the output features of the last feature extraction subunit in the first feature extraction unit are spliced ​​with the input data to obtain feature vector 1. Then, feature vector 1 is input to the first feature compression unit for feature compression to obtain compressed feature vector 1, and the compressed feature vector 1 is simultaneously input to the second feature extraction unit, the second feature splicing unit, the third feature splicing unit and the fourth feature splicing unit; at this time, the second feature extraction unit uses the internal densely connected feature subunits to extract features from the compressed feature vector 1 to obtain feature vector 2, and input feature vector 2 to the second feature splicing unit; in this way, the second feature splicing unit performs feature splicing on feature vector 2, input data and compressed feature vector 1 to obtain feature vector 3; then, the second feature compression unit performs feature compression on feature vector 3 to obtain compressed feature vector 2; finally, the compressed feature vector 2 is input to the third feature extraction unit, the third feature splicing unit and the fourth feature splicing unit.

[0094] Next, the third feature extraction unit performs feature extraction on the compressed feature vector 2 to obtain feature vector 3; and the third feature splicing unit performs feature splicing on the feature vector 3, input data, compressed feature vector 1 and compressed feature vector 2 to obtain feature vector 4; then, the third feature compression unit performs feature compression on the feature vector 4 to obtain compressed feature vector 3; similarly, the third feature compression unit will also input the compressed feature vector 3 into the fourth feature extraction unit and the fourth feature splicing unit respectively.

[0095] Finally, the fourth feature extraction unit extracts features from the compressed feature vector 3 to obtain feature vector 4; and the fourth feature splicing unit performs feature splicing on feature vector 4, input data, compressed feature vector 1, compressed feature vector 2, and compressed feature vector 3 to obtain deep feature information, that is, the aforementioned commutation fault feature vector.

[0096] In this way, the feature extraction module can be used to extract the features of the input data and obtain a commutation fault feature vector, which can then be input into the feature enhancement module for feature enhancement.

[0097] Optionally, one of the following network structures of the public feature enhancement module: See also Figure 8 As shown, the example feature enhancement module may include, but is not limited to: a channel attention unit, a temporal attention unit, a first feature product unit and a second feature product unit ( Figure 8 The ⊙ in the figure represents a feature product unit), where the working process of each of the above units is as follows: The channel attention unit is used to reshape the commutation fault feature vector to obtain a reshaped feature vector and generate a channel attention weight of the reshaped feature vector; then, the first feature product unit is used to receive the commutation fault feature vector output by the feature extraction module, and perform a feature product operation on the channel attention weight and the commutation fault feature vector to obtain channel attention feature data; then, the timing attention unit is used to generate an attention map of the channel attention feature data in the timing dimension, and transmit the attention map in the timing dimension to the second feature product unit; finally, the second feature product unit is used to perform a feature product operation on the attention map in the timing dimension and the channel attention feature data, so as to obtain the enhanced fault feature vector after the feature product operation.

[0098] See also Figure 9 and Figure 10As shown, for example, the channel attention unit may include but is not limited to: a dimensionality reduction multilayer perceptron, a fourth ReLU layer, and a dimensionality increase multilayer perceptron connected in sequence, and the temporal attention unit may include but is not limited to a dimensionality reduction 1D convolution layer, a fourth BN layer, a fifth ReLU layer, a dimensionality increase 1D convolution layer, a fifth BN layer, and a nonlinear activation layer connected in sequence; wherein, the number of channels of the dimensionality increase multilayer perceptron is the same as the number of channels of the commutation fault feature vector, the dimensionality reduction 1D convolution layer and the dimensionality increase 1D convolution layer both use a 7×1 convolution kernel, and the number of channels of the dimensionality increase 1D convolution layer is the same as the number of channels of the commutation fault feature vector.

[0099] Thus, through the above explanation, the overall structure of the feature enhancement module based on the global attention mechanism is the channel-temporal attention mechanism, in which the features output from the feature extraction module are weighted by the channel attention unit and the temporal attention unit, and the output features can highlight more critical features. Based on this, redundant or irrelevant information can be weakened, thereby improving the classification accuracy of subsequent classifiers.

[0100] Based on this, the working process of the aforementioned feature enhancement module is as follows: The channel attention unit first reshapes the commutation fault feature vector, then inputs the reshaped feature vector into a dimensionality reduction MLP layer (i.e., the aforementioned dimensionality reduction multilayer perceptron) and ReLU (the aforementioned fourth ReLU layer) to extract key information; then, it passes through a dimension recovery MLP layer (i.e., dimensionality increase multilayer perceptron) to generate channel weights , and generate At the same time, the data shape is restored to its original dimension, where the channel attention weight The calculation method is: (6) In formula (6), represents the weights and biases of the dimensionality reduction multilayer perceptron, represents the weights and biases of the multi-layer perceptron with increased dimensionality, represents the ReLU function, Represents the reshaped feature vector.

[0101] Then, channel weighting can be performed, and the process is: ⊙ , where Represents channel attention feature data.

[0102] After channel weighting, the temporal attention unit uses a one-dimensional convolution operation to capture temporal dependencies, that is, first use a 7×1 ConvThe features are processed by dimensionality reduction (i.e., 1D convolution layer is used for construction), and the feature distribution is stabilized by BN layer (the fourth BN layer) and ReLU layer (the fifth ReLU layer mentioned above); then, 7×1 Conv The feature is mapped back to the original number of channels (i.e., the dimension-increasing 1D convolution layer), and further normalized through the BN layer (i.e., the fifth BN layer). Finally, the output is limited to the [0, 1] interval through the Sigmoid function (i.e., the nonlinear activation layer) to generate the temporal attention weight , where the temporal attention weight (i.e., the attention map on the aforementioned time series dimension) is calculated as follows: (7) In formula (7), represents the Sigmoid function, represents batch normalization processing, They represent the dimensionality increase convolution operation and dimensionality reduction convolution operation respectively. Represents the ReLU function; where, , where x represents the data input to the BN layer, Represents the average value of the data input to the BN layer, The variance of the data input to the BN layer, This represents the offset and scale parameters, Represents a constant.

[0103] ; In the formula, x1 represents the data input to the ReLU layer; , where x2 is the input of the convolutional layer, Represents the weights and biases of the convolutional layer.

[0104] In this way, each position of the time series dimension is weighted, that is, using ⊙ , the enhanced fault feature vector is obtained; wherein ⊙ represents element-by-element multiplication.

[0105] Therefore, through the aforementioned feature enhancement module, the channel-temporal attention mechanism can be used to highlight more critical features, thereby improving the classification accuracy of subsequent classifiers.

[0106] After obtaining the enhanced fault feature vector and inputting it into the output module, the commutation fault type of the HVDC transmission system can be obtained. In this embodiment, one of the network structures of the output module is provided: See also Figure 5As shown, the output module may include, but is not limited to, a global average pooling layer and a classifier, and the classifier includes a second fully connected layer, a sixth ReLU layer, and a third fully connected layer, wherein the classification process is: The global average pooling layer is used to perform global average pooling processing on the enhanced fault feature vector to obtain a pooled fault feature vector, and input the pooled fault feature vector into the second fully connected layer; the second fully connected layer is used to perform feature dimensionality reduction processing on the pooled fault feature vector to obtain a reduced dimensionality feature vector, and input the reduced dimensionality feature vector into the sixth ReLU layer; wherein, the sixth ReLU layer is used to perform nonlinear transformation processing on the reduced dimensionality feature vector to obtain a nonlinear feature vector; and the second fully connected layer is used to use the Softmax function to perform feature mapping processing on the nonlinear feature vector to obtain the probability value of the operating data belonging to each type of commutation fault type.

[0107] In this embodiment, the output module consists of a global average pooling layer and a classifier, wherein the global average pooling averages all values ​​of the entire sequence and completely compresses the last dimension of the input feature to a scalar value of 1. The classifier consists of two fully connected layers and ReLU. The sequence after the global average pooling operation is flattened and input into the second fully connected layer of the classifier for feature dimensionality reduction. The ReLU layer is used after the second fully connected layer to perform nonlinear transformation on the extracted high-level features, thereby helping the network to better map to the final classification results; finally, the third fully connected layer is used to map the reduced dimensionality features to the output space, that is, to obtain the ratio of each fault category as the standardized score logits; finally, the Softmax function is used to calculate the probability values ​​corresponding to the logit values ​​of different fault types; in this way, the fault type with the highest probability can be selected as the commutation fault type predicted by the model.

[0108] Therefore, through the detailed structural description of the aforementioned commutation fault detection model, the multi-source feature matrix composed of the optimal fault features of each traceability data can be used to obtain the fault type that causes the commutation failure of the HVDC transmission system, thereby completing the commutation failure fault tracing.

[0109] In addition, for the training of the commutation fault detection model, the sample optimal fault feature data of the aforementioned sample operation data are used to form a multi-source feature matrix. Then, the multi-source feature matrix of each sample operation data is used as input, and the commutation fault type corresponding to each sample operation data is used as output to train the initial commutation fault detection model, so that after the training is completed, the commutation fault detection model is obtained; at the same time, during training, the example classifier also includes a Dropout layer, which is used to randomly discard some neurons during the training process to prevent the model from overfitting.

[0110] See also Figure 2 As shown, in this embodiment, it is assumed that the fault feature data output by the fault feature extraction model is (corresponding to the aforementioned sample a phase voltage sample, b phase voltage sample, c phase voltage sample, DC voltage sample and DC current sample respectively), after feature screening, the optimal fault feature of the sample is obtained as follows: , then the above five optimal fault features can be spliced ​​to obtain the multi-source feature matrix , so it can be used as input data and input into the commutation fault detection model for model training.

[0111] In this embodiment, the loss function of the commutation fault detection model is a cross-entropy loss function. Therefore, the loss can be calculated by forward propagation, and the network parameters can be updated after the gradient is calculated by back propagation. In this way, the iteration is repeated until the cross-entropy loss function converges, and the training process can be completed and the network parameters can be saved.

[0112] Finally, when online traceability is required, see Figure 3 As shown in the figure, after the system receives the signal of commutation failure, the AC voltage, DC voltage, and DC current fault data on the AC bus on the inverter side are collected; then, they are normalized and input into the fault feature extraction model for preliminary feature extraction; then, feature screening is performed to obtain a multi-source feature matrix, and finally, the multi-source feature matrix is ​​input into the DenseNet network with the fusion attention mechanism for multi-level and deep feature extraction, and finally the fault tracing results are output through the classifier to determine the specific fault type that caused the commutation failure.

[0113] Therefore, through the HVDC system commutation failure fault tracing method described in detail in the aforementioned steps S1 to S6, the present invention can not only realize the commutation failure fault tracing, but also use the machine learning model and the feature screening algorithm based on feature density to extract the optimal fault features during the tracing process. In this way, the interference of human subjective experience on feature extraction can be reduced, thereby improving the effectiveness of fault tracing; therefore, the present invention is very suitable for large-scale application and promotion.

[0114] In one possible design, the second aspect of this embodiment provides a hardware device for implementing the method for tracing the source of a commutation failure in a high-voltage direct current transmission system described in the first aspect of the embodiment, including: An acquisition unit is used to acquire operating data of the inverter side of the high-voltage direct current transmission system when a commutation failure occurs, wherein the operating data includes an AC voltage, a DC voltage, and a DC current on the inverter side.

[0115] The feature extraction unit is used to input the operating data into the fault feature extraction model to obtain the first fault feature data of the DC voltage, the second fault feature data of the DC current and the third fault feature data of each phase voltage in the AC voltage.

[0116] A feature screening unit is used to perform feature screening processing on the first fault feature data, the second fault feature data and the third fault feature data of each phase voltage using a feature screening algorithm based on feature density to obtain the first optimal fault feature data of the DC voltage, the second optimal fault feature data of the DC current and the third optimal fault feature data of each phase voltage.

[0117] The fault tracing unit is used to input the first optimal fault characteristic data, the second optimal fault characteristic data and the third optimal fault characteristic data of each phase voltage into a commutation fault detection model to obtain the commutation fault type of the high voltage direct current transmission system.

[0118] The working process, working details and technical effects of the device provided in this embodiment can be found in the first aspect of the embodiment and will not be described in detail here.

[0119] In one possible design, the third aspect of this embodiment provides another high-voltage direct current transmission system commutation failure fault tracing device. Taking the device as an electronic device as an example, it includes: a memory, a processor and a transceiver that are communicatively connected in sequence, wherein the memory is used to store computer programs, the transceiver is used to send and receive messages, and the processor is used to read the computer program and execute the high-voltage direct current transmission system commutation failure fault tracing method as described in the first aspect of the embodiment.

[0120] The working process, working details and technical effects of the electronic device provided in this embodiment can be found in the first aspect of the embodiment and will not be described in detail here.

[0121] A fourth aspect of this embodiment provides a storage medium storing instructions including the method for tracing the source of a commutation failure in a high-voltage direct current transmission system as described in the first aspect of the embodiment, that is, the storage medium stores instructions that, when executed on a computer, execute the method for tracing the source of a commutation failure in a high-voltage direct current transmission system as described in the first aspect of the embodiment.

[0122] The working process, working details and technical effects of the storage medium provided in this embodiment can be found in the first aspect of the embodiment and will not be described in detail here.

[0123] A fifth aspect of this embodiment provides a computer program product comprising instructions, which, when executed on a computer, causes the computer to execute the method for tracing the source of a commutation failure in a high-voltage direct current transmission system as described in the first aspect of the embodiment, wherein the computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for tracing the source of commutation failure in a high-voltage direct current transmission system, characterized in that: include: Acquiring operating data of the inverter side of the high-voltage direct current transmission system when the commutation failure occurs, wherein the operating data includes an AC voltage, a DC voltage, and a DC current on the inverter side; Inputting the operating data into a fault feature extraction model to obtain first fault feature data of the DC voltage, second fault feature data of the DC current, and third fault feature data of each phase voltage in the AC voltage; Performing feature screening processing on the first fault feature data, the second fault feature data, and the third fault feature data of each phase voltage using a feature screening algorithm based on feature density to obtain first optimal fault feature data of the DC voltage, second optimal fault feature data of the DC current, and third optimal fault feature data of each phase voltage; The first optimal fault characteristic data, the second optimal fault characteristic data, and the third optimal fault characteristic data of each phase voltage are input into a commutation fault detection model to obtain a commutation fault type of the high voltage direct current transmission system.

2. The method according to claim 1, characterized in that Inputting the operating data into a fault feature extraction model to obtain first fault feature data of the DC voltage, second fault feature data of the DC current, and third fault feature data of each phase voltage in the AC voltage, including: Obtaining a fault feature extraction model, wherein the fault feature extraction model includes five fault feature extraction networks, the DC voltage, the DC current, and each phase voltage of the AC voltage respectively corresponds to a fault feature extraction network, each fault feature extraction network includes a plurality of fault feature extraction units connected in sequence, and each fault feature extraction unit includes an encoder, a hidden layer, and a decoder connected in sequence; The DC voltage, the DC current, and each phase voltage of the AC voltage are respectively input into corresponding fault feature extraction networks to obtain first fault feature data of the DC voltage, second fault feature data of the DC current, and third fault feature data of each phase voltage of the AC voltage.

3. The method according to claim 2, characterized in that The fault feature extraction network corresponding to the DC voltage is trained in the following way; Acquire a plurality of sample operating data of the inverter side of the high voltage direct current transmission system when commutation failure occurs, and extract a sample DC voltage from each sample operating data to form a first training set; Using the first training set to train the i-th fault feature extraction unit in the target network, and during the training process, calculating a loss function of the i-th fault feature extraction unit based on output data of the i-th fault feature extraction unit, and updating the model parameters of the i-th fault feature extraction unit based on the loss function until the loss function converges, thereby obtaining a trained i-th fault feature extraction unit, wherein the target network is any untrained fault feature extraction network; Inputting the first training set into a designated network and using the output data of the designated network to form a new training set, wherein the designated network includes the first i fault feature extraction units after training; Incrementing i by 1, updating the first training set to the new training set, and reusing the first training set to train the i-th fault feature extraction unit in the target network until i equals n, thereby obtaining an initial training feature extraction network, wherein the initial value of i is 1, and n is the total number of fault feature extraction units in the target network; The initially trained feature extraction network is subjected to supervised training to obtain a fault feature extraction network corresponding to the DC voltage after the supervised training.

4. The method according to claim 3, characterized in that The loss function of the i-th fault feature extraction unit is: （1） In formula (1), represents the loss function of the i-th fault feature extraction unit, represents the kth sample DC voltage in the first training set, represents the output data of the decoder in the i-th fault feature extraction unit after the k-th sample DC voltage is input to the i-th fault feature extraction unit, represents the weight matrix of the i-th fault feature extraction unit, represents a sparse constant, represents the weight decay coefficient, represents the sparse penalty coefficient, represents the total number of sample DC voltages, represents the average activation probability of the jth hidden neuron in the hidden layer of the i-th fault feature extraction unit, represents the total number of hidden neurons in the hidden layer, represents the norm operation; When the first training set is used to train the i-th fault feature extraction unit, the activation value of each hidden neuron in the hidden layer of each sample DC voltage is recorded, and the average of the activation values ​​of the j-th hidden neuron in the hidden layer of each sample DC voltage is used as the average activation probability of the j-th hidden neuron.

5. The method according to claim 3, characterized in that Performing supervised training on the initial training feature extraction network to obtain a fault feature extraction network corresponding to the DC voltage after the supervised training, including: randomly selecting a number of sample DC voltages from the first training set, and determining a fault category label corresponding to each of the selected sample DC voltages, so as to form a second training set using the selected number of sample DC voltages and the fault category label corresponding to each of the selected sample DC voltages, wherein the fault category label corresponding to any sample DC voltage is used to characterize a commutation fault type of the high voltage direct current transmission system; Based on the initial training feature extraction network, generating a pre-training feature extraction network, wherein the pre-training feature extraction network includes the initial training feature extraction network, the first fully connected layer and the Softmax classifier connected in sequence; The pre-trained feature extraction network is trained using each sample DC voltage in the second training set as input and the fault category probability corresponding to each sample DC voltage as output. During the training process, a cross-entropy loss function is calculated using the fault category probability and the fault category label corresponding to each sample DC voltage. The model parameters of the pre-trained feature extraction network are updated using the cross-entropy loss function, so as to obtain an optimal fault feature extraction network after the training is completed. The first fully connected layer and the Softmax classifier are removed from the optimal fault feature extraction network to obtain a fault feature extraction network corresponding to the DC voltage.

6. The method according to claim 1, characterized in that The first fault feature data includes a plurality of fault sub-feature data, wherein the first fault feature data is subjected to feature screening processing using a feature screening algorithm based on feature density to obtain first optimal fault feature data of the DC voltage, including: Calculating a local density index of each fault sub-feature data in the first fault feature data, wherein the local density index of any fault sub-feature data is used to characterize the local density of the any fault sub-feature data, and a smaller local density index indicates a greater local density; Calculate the characteristic density of each fault sub-characteristic data according to the local density index of each fault sub-characteristic data; By utilizing the characteristic density of each fault sub-characteristic data and adopting a greedy algorithm, a number of key fault sub-characteristic data are screened out from each fault sub-characteristic data to form the first optimal fault characteristic data.

7. The method according to claim 6, characterized in that Calculating the local density index of each fault sub-feature data in the first fault feature data includes: For any fault sub-feature data, calculating a distance between the any fault sub-feature data and target data, wherein the target data is each fault sub-feature data in the first fault feature data except the any fault sub-feature data; Sort the target data in ascending order of distance to obtain a first sorting sequence; Filtering the first K target data from the first sorting sequence as the neighborhood data of any fault sub-feature data, where K is a positive integer; Calculating the distance between any fault sub-feature data and each neighboring data, and taking the mean of the sum of the distances between any fault sub-feature data and each neighboring data as the local density index; Accordingly, the characteristic density of each fault sub-characteristic data is calculated based on the local density index of each fault sub-characteristic data, which includes: For any fault sub-feature data, obtain density adjustment parameters; The product of the local density index of any fault sub-feature data and the density adjustment parameter is used as the feature density of any fault sub-feature data.

8. The method according to claim 6, characterized in that By using the characteristic density of each fault sub-feature data and adopting a greedy algorithm, a number of key fault sub-feature data are screened out from each fault sub-feature data to form the first optimal fault feature data, including: Sorting the sub-feature data of each fault in descending order of feature density to obtain a second sorting sequence; removing the first fault sub-feature data from the second sorting sequence to obtain a third sorting sequence; Initialize the number of iterations s to 1, and put the first fault sub-feature data in the second sorting sequence into the initial sample representative set to obtain the representative sample set at the s-1th iteration; Filtering out designated data from the representative sample set at the s-1th iteration, wherein the designated data is the most recently added feature data in the representative sample set at the s-1th iteration; Selecting the sth fault sub-feature data from the third sorting sequence; Calculating the distance between the s-th fault sub-feature data and the specified data, and selecting the smallest feature density from the feature density of the s-th fault sub-feature data and the feature density of the specified data; Determine whether the distance between the s-th fault sub-feature data and the specified data is less than the minimum feature density; If not, the s-th fault sub-feature data is taken as a key fault sub-feature data and put into the representative sample set at the s-1-th iteration to obtain the representative sample set at the s-th iteration; Increment s by 1 and reselect specified data from the representative sample set at the s-1th iteration until s equals S, so as to utilize the key fault sub-feature data in the representative sample set at the sth iteration to form the first optimal fault feature data, where S is the total number of fault sub-feature data in the third sorting sequence.

9. The method according to claim 1, characterized in that The commutation fault detection model includes a feature extraction module, a feature enhancement module and an output module connected in sequence; The feature extraction module is configured to perform secondary feature extraction processing on the input data using one-dimensional convolution to obtain a commutation fault feature vector, wherein the input data includes the first optimal fault feature data, the second optimal fault feature data, and the third optimal fault feature data of each phase voltage; A feature enhancement module is used to perform feature enhancement processing on the commutation fault feature vector using a global attention mechanism to obtain an enhanced fault feature vector; an output module, configured to perform feature mapping processing on the enhanced fault feature vector to obtain probability values ​​of the operating data belonging to various commutation fault types, and to use the commutation fault type corresponding to the maximum probability value as the commutation fault type of the high voltage direct current transmission system.

10. A device for tracing the source of commutation failure in a high-voltage direct current transmission system, characterized in that: include: an acquiring unit, configured to acquire operating data of the inverter side of the HVDC transmission system when a commutation failure occurs, wherein the operating data includes an AC voltage, a DC voltage, and a DC current of the inverter side; a feature extraction unit, configured to input the operating data into a fault feature extraction model to obtain first fault feature data of the DC voltage, second fault feature data of the DC current, and third fault feature data of each phase voltage in the AC voltage; a feature screening unit, configured to perform feature screening processing on the first fault feature data, the second fault feature data, and the third fault feature data of each phase voltage using a feature screening algorithm based on feature density, so as to obtain first optimal fault feature data of the DC voltage, second optimal fault feature data of the DC current, and third optimal fault feature data of each phase voltage; The fault tracing unit is used to input the first optimal fault characteristic data, the second optimal fault characteristic data and the third optimal fault characteristic data of each phase voltage into a commutation fault detection model to obtain the commutation fault type of the high voltage direct current transmission system.

Images