Direct current distribution line fault identification method based on improved VMD and SAE

Abstract

The invention discloses a DC distribution line fault identification method based on improved VMD and SAE. Firstly, a grey wolf optimization algorithm is adopted to improve a VMD method for single-pole grounding fault current, direct current during single-pole grounding fault is analyzed, and the optimal number of decomposition layers and penalty factors are selected as VMD decomposition parameters of all fault types. And then, calculating the intrinsic mode component IMF1 under each fault type by using Hilbert transform to obtain an amplitude, identifying an energy maximum value point by constructing an energy index calculation formula, and setting an energy index threshold as a starting criterion of fault detection. And finally, inputting the data of the optimal intrinsic mode component and the positive voltage data into SAE for dimensionality reduction and learning of feature data, and establishing a classification model between fault data and fault types, thereby completing discrimination of the fault types. The method has relatively high reliability and robustness, can realize identification and classification of various faults, and is of great significance for quickly formulating an electric power overhaul scheme and improving power supply reliability.

Description

Technical Field

[0001] The invention relates to a DC distribution line fault identification method based on improved VMD and SAE. Background Art

[0002] The development of DC distribution network in the future is an important way to achieve high proportion of new energy access and consumption, so how to achieve safe and stable operation of DC distribution network has always been the focus of people's attention. In actual operation, the occurrence probability of single-pole grounding fault in DC distribution network is relatively high. Under this fault, the voltage of non-fault pole will suddenly rise, posing a threat to line insulation. If the fault is not isolated in time, the line insulation will be destroyed, which will evolve into a more serious fault and even cause system collapse. At the same time, the widespread application of DC distribution network is restricted by the immaturity of protection technology. The low system impedance leads to a rapid increase in fault current, which may endanger the safety of power electronic equipment. Therefore, it is of great significance to study DC distribution line fault detection technology, especially the rapid detection of single-pole grounding fault.

[0003] At present, most of the existing fault detection methods only detect the voltage and current signals of DC side faults. However, the AC side fault also has a great impact on the voltage and current measurement of the DC side. The setting of the DC side threshold alone cannot effectively distinguish between AC and DC side faults, which can easily cause false operation of the DC side protection.

[0004] At present, traditional DC distribution line fault identification methods mainly rely on manual experience and threshold setting, and usually judge the fault status by simple feature analysis of current and voltage signals. The main problems of these methods are: (1) Only analyzing a single signal feature makes it difficult to fully characterize the complex characteristics of nonlinear and non-stationary signals, resulting in insufficient fault feature extraction; (2) The selection of parameter thresholds is highly dependent on manual experience, lacks automatic adjustment capabilities, and cannot adapt to different scenarios and complex working conditions; (3) Insufficient consideration of AC-DC coupling characteristics and multi-variable dependencies. Traditional methods are prone to misjudgment or missed judgment in multiple fault types and coupling scenarios, especially when there is severe noise interference or large dynamic changes in system characteristics, and the reliability and accuracy are low.

[0005] Therefore, in view of the shortcomings of the existing technology, it is necessary to design a new method to meet the actual usage needs. Summary of the invention

[0006] Purpose of the invention: In order to accurately analyze the electrical quantities on the DC side, identify the impact of AC side faults on the DC side, and realize DC distribution line fault identification with accurate distinction of fault types, the present invention proposes a DC distribution line fault identification method based on improved VMD and SAE.

[0007] To achieve the above object, the technical solution adopted by the present invention is:

[0008] A DC distribution line fault identification method based on improved VMD and SAE includes the following steps:

[0009] (1) Real-time extraction of the DC side current of the line, using the VMD algorithm improved by the Grey Wolf algorithm, the DC side current signal is decomposed by VMD through the optimized parameters, and the IMF is selected. 1 Modal components as response fault characteristics;

[0010] (2) Based on the IMF extracted in step (1) 1 Characteristic mode components, using Hilbert transform to identify IMF 1 Solve the amplitude and calculate the instantaneous energy of the signal component; use the common characteristics of instantaneous energy of different faults to establish system fault judgment criteria;

[0011] (3) According to the system fault judgment criteria established in step (2), the faults are divided into: single-pole grounding fault and bipolar short circuit fault on the DC side; single-phase grounding fault, two-phase grounding fault and three-phase grounding fault on the AC side; the kurtosis and correlation coefficient of the inherent modal components of each fault are calculated respectively, and the comprehensive indicators of different faults are obtained;

[0012] (4) According to the different comprehensive fault indicators calculated in step (3), the two inherent modal components with the largest comprehensive indicators and the positive and negative voltage data are selected and input into SAE for fault feature extraction, and the extracted features are used for fault classification through softmax, thereby realizing fault identification of the DC distribution network.

[0013] Specifically, the VMD algorithm improved by the grey wolf algorithm in step (1) is:

[0014] (1.1) Initialize the parameters of the gray wolf algorithm, including the number of wolves, the number of iterations, and the solution parameters. Use the penalty factor and modal decomposition number of VMD as optimization variables, and minimize the sample entropy as the fitness function. Use the gray wolf algorithm for optimization. According to the update rule, update the position and speed of each wolf, and calculate the fitness value of each wolf. According to the fitness value, update the optimal solution, suboptimal solution, and general solution. Select sample entropy as the fitness function to determine the best solution. Stop the iteration when the local minimum entropy is reached during the search. Output the penalty factor and modal decomposition number corresponding to the optimal solution and suboptimal solution, and perform VMD decomposition on the signal according to the optimized parameters.

[0015] (1.2) Extract the DC side current of the line in real time. According to the penalty factor and modal decomposition number of VMD obtained in step (1), the DC side current signal is decomposed by VMD with optimized parameters. After the DC side current signal is decomposed by VMD to obtain various modes, the intrinsic modal component IMF is selected. 1Modal components that characterize the fault response.

[0016] Specifically, the system fault criterion described in step (2) is:

[0017] The DC side current intrinsic modal component IMF obtained according to step (2) 1 , to the IMF 1 The modal components are Hilbert transformed and the instantaneous amplitude A of the signal is analyzed n (t) and instantaneous phase θ n (t), construct the instantaneous energy E n (t) is used to describe the energy state of the signal at a certain instant, that is:

[0018]

[0019]

[0020] In the formula, x n (t) represents the nth modal signal of the original signal; H[x n (t)] represents the output signal after Hilbert transform; X n (t) is the analytical signal, x n (t) and H[x n (t)] is its complex conjugate pair.

[0021] Fault detection is performed by identifying the maximum value point in the energy time series diagram, and a threshold η is set. The common characteristics of instantaneous energy of different faults are used to establish the system fault criterion. The system fault criterion is:

[0022] E n (t)>η

[0023] In the formula, the threshold η is set according to the maximum AC side power disturbance and IMF under load switching conditions. 1 The maximum value of the instantaneous energy is considered.

[0024] The system fault judgment criteria established in step (3) classifies the faults into: single-pole grounding fault and bipolar short circuit fault on the DC side; single-phase grounding fault, two-phase grounding fault and three-phase grounding fault on the AC side.

[0025] According to the fault classification method in step (3), the kurtosis and correlation coefficient of each modal component are calculated to calculate the comprehensive index of different faults. The calculation formulas of kurtosis, correlation coefficient and comprehensive index are as follows:

[0026] Kurtosis (KI): The kurtosis of each modal component is calculated to assess the sharpness of its distribution.

[0027]

[0028] Among them, K i is the kurtosis of the ith modal component, μ i It is the IMF i The average value, N is the number of samples

[0029] Correlation Coefficient (CI): Calculate the Pearson correlation coefficient between each modal component and the original signal.

[0030]

[0031] Among them, R i is the modal component MFI i Pearson correlation coefficient between S and the original signal S; i Represents the original signal value of the i-th data point; is the modal component MFI i The average value of all data points; is the average value of all data points of the original signal S; n is the total number of data points.

[0032] Comprehensive index (KCI): the product of kurtosis and the absolute value of the correlation coefficient.

[0033] KCI=|KI|*|CI|

[0034] According to the calculated comprehensive fault indicators, the two inherent modal components with the largest comprehensive indicators and the positive and negative voltage data are selected and input into SAE for fault feature extraction.

[0035] Specifically, a fault identification method described in step (4) is:

[0036] By considering the Deep SAE encoder input, hidden layer output, and reconstructed decoder output, the weight W is introduced e , bias b e And the encoding function f e (.) to map out L encoded outputs E. In the pre-training stage, the encoder is represented by the following formula:

[0037]

[0038] Where, X n is the data of the nth input sample, is the weight parameter of the encoder, connecting the nth input unit and the lth encoder hidden layer unit; is the bias parameter of the lth hidden layer unit of the encoder; f e (·) is an activation function that applies a nonlinear transformation to the sum of weighted inputs; E nlis the output value of the lth hidden layer neuron after being processed by the encoder; L is each hidden layer neuron in the encoder network.

[0039] The error is measured by the cross entropy loss function, and the goal is to reconstruct the input without loss. The cross entropy error as the cost function is minimized by the gradient descent technique, and the updated parameters and new parameters are calculated using the learning rate and the Hadamard product. The decoder layer is removed from the encoder architecture, and the last fully connected layer is named the softmax layer. The softmax activation function is used to classify the fault type according to the highest value of the probability distribution, thereby realizing the fault identification of the DC distribution network.

[0040] An electronic device includes one or more processors for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors execute a DC distribution line fault identification method based on improved VMD and SAE.

[0041] A storage medium stores a computer program, wherein the computer program is configured to execute a DC distribution line fault identification method based on improved VMD and SAE when running.

[0042] Beneficial effects: The present invention can effectively realize complex fault identification of DC distribution lines by combining improved variational mode decomposition (VMD) and gray wolf optimization algorithm (GWO). This method automatically optimizes VMD parameters through gray wolf optimization, which solves the limitations of traditional manual parameter adjustment and improves the decomposition accuracy of nonlinear and non-stationary signals. By utilizing the multimodal characteristics of VMD decomposition, combined with instantaneous energy, kurtosis and other characteristics, this method realizes the accurate extraction of complex fault features and automatic classification of multiple types of faults. Further introduction of deep learning models for classification and identification effectively improves the accuracy and robustness of fault identification. It has good universality, reliability and objectivity, and can be widely used in various engineering practices. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a flow chart of the present invention.

[0044] Figure 2 This is the topology diagram of the two-terminal power supply DC distribution network built in PSCAD / EMTDC.

[0045] Figure 3 This is the algorithm flow chart for improving VMD based on the Grey Wolf Optimization Algorithm.

[0046] Figure 4 is a stacked autoencoder model.

[0047] Figure 5 This is the flow chart of DC line fault detection and classification.

[0048] Figure 6 This is the classification confusion matrix obtained for the experimental test. DETAILED DESCRIPTION

[0049] The present invention will be further described below in conjunction with the accompanying drawings.

[0050] A two-terminal DC distribution network topology diagram built in PSCAD / EMTDC is shown below: Figure 2 As shown, the present invention proposes a DC distribution line fault identification method based on improved VMD and SAE, comprising the following steps:

[0051] To achieve the above object, the technical solution adopted by the present invention is:

[0052] A DC distribution line fault identification method based on improved VMD and SAE includes the following steps:

[0053] (1) Real-time extraction of the DC side current of the line, using the VMD algorithm improved by the Grey Wolf algorithm, the DC side current signal is decomposed by VMD through the optimized parameters, and the IMF is selected. 1 Modal components as response fault characteristics;

[0054] (2) Based on the IMF extracted in step (1) 1 Characteristic mode components, using Hilbert transform to identify IMF 1 Solve the amplitude and calculate the instantaneous energy of the signal component; use the common characteristics of instantaneous energy of different faults to establish system fault judgment criteria;

[0055] (3) According to the system fault judgment criteria established in step (2), the faults are divided into: single-pole grounding fault and bipolar short circuit fault on the DC side; single-phase grounding fault, two-phase grounding fault and three-phase grounding fault on the AC side; the kurtosis and correlation coefficient of the inherent modal components of each fault are calculated respectively, and the comprehensive indicators of different faults are obtained;

[0056] (4) According to the different comprehensive fault indicators calculated in step (3), the two inherent modal components with the largest comprehensive indicators and the positive and negative voltage data are selected and input into SAE for fault feature extraction, and the extracted features are used for fault classification through softmax, thereby realizing fault identification of the DC distribution network.

[0057] Specifically, the VMD algorithm improved by the grey wolf algorithm in step (1) is:

[0058] (1.1) Initialize the parameters of the gray wolf algorithm, including the number of wolves, the number of iterations, and the solution parameters. Use the penalty factor and modal decomposition number of VMD as optimization variables, and minimize the sample entropy as the fitness function. Use the gray wolf algorithm for optimization. According to the update rule, update the position and speed of each wolf, and calculate the fitness value of each wolf. According to the fitness value, update the optimal solution, suboptimal solution, and general solution. Select sample entropy as the fitness function to determine the best solution, stop iteration when the local minimum entropy is reached during the search, output the penalty factor and modal decomposition number corresponding to the optimal solution and suboptimal solution, and perform VMD decomposition on the signal according to the optimized parameters. The algorithm flow of improving VMD based on the gray wolf optimization algorithm is as follows: Figure 3 shown.

[0059] (1.2) Extract the DC side current of the line in real time. According to the penalty factor and modal decomposition number of VMD obtained in step (1), the DC side current signal is decomposed by VMD with optimized parameters. After the DC side current signal is decomposed by VMD to obtain various modes, the intrinsic modal component IMF is selected. 1 Modal components that characterize the fault response.

[0060] Specifically, the system fault criterion described in step (2) is:

[0061] The obtained DC side current intrinsic modal component IMF 1 , to the IMF 1 The modal components are Hilbert transformed and the instantaneous amplitude A of the signal is analyzed n (t) and instantaneous phase θ n (t), construct the instantaneous energy E n (t) is used to describe the energy state of the signal at a certain instant, that is:

[0062]

[0063] In the formula, x n (t) represents the nth modal signal of the original signal; H[x n (t)] represents the output signal after Hilbert transform; X n (t) is the analytical signal, x n (t) and H[x n (t)] is its complex conjugate pair.

[0064] Fault detection is performed by identifying the maximum value point in the energy time series diagram, and a threshold η is set. The common characteristics of instantaneous energy of different faults are used to establish the system fault criterion. The system fault criterion is:

[0065] E n (t)>η

[0066] In the formula, the threshold η is set according to the maximum AC side power disturbance and IMF under load switching conditions. 1 The maximum value of the instantaneous energy is considered.

[0067] The system fault judgment criteria established in step (3) classifies the faults into: single-pole grounding fault and bipolar short circuit fault on the DC side; single-phase grounding fault, two-phase grounding fault and three-phase grounding fault on the AC side.

[0068] According to the fault classification method in step (3), the kurtosis and correlation coefficient of each modal component are calculated to calculate the comprehensive index of different faults. The calculation formulas of kurtosis, correlation coefficient and comprehensive index are as follows:

[0069] Kurtosis (KI): The kurtosis of each modal component is calculated to assess the sharpness of its distribution.

[0070]

[0071] Among them, K i is the kurtosis of the ith modal component, μ i It is the IMF i The average value, N is the number of samples

[0072] Correlation Coefficient (CI): Calculate the Pearson correlation coefficient between each modal component and the original signal.

[0073]

[0074] Among them, R i is the modal component MFI i Pearson correlation coefficient between S and the original signal S; i Represents the original signal value of the i-th data point; is the modal component MFI i The average value of all data points; is the average value of all data points of the original signal S; n is the total number of data points.

[0075] Comprehensive index (KCI): the product of kurtosis and the absolute value of the correlation coefficient.

[0076] KCI=|KI|*|CI|

[0077] The different comprehensive fault indicators are calculated, and the two inherent modal components with the largest comprehensive indicators and the positive and negative voltage data are selected and input into SAE for fault feature extraction.

[0078] Specifically, a fault identification method described in step (4) is:

[0079] By considering the Deep SAE encoder input, hidden layer output, and reconstructed decoder output, the weight W is introduced e , bias b e And the encoding function f e (.) to map out L encoded outputs E. In the pre-training stage, the encoder is represented by the following formula:

[0080]

[0081] Where, X n is the data of the nth input sample, is the weight parameter of the encoder, connecting the nth input unit and the lth encoder hidden layer unit; is the bias parameter of the lth hidden layer unit of the encoder; f e (·) is an activation function that applies a nonlinear transformation to the sum of weighted inputs; E nl is the output value of the lth hidden layer neuron after being processed by the encoder; L is each hidden layer neuron in the encoder network. The stacked autoencoder model is as follows Figure 4 shown.

[0082] The error is measured by the cross entropy loss function, and the goal is to reconstruct the input without loss. The cross entropy error as the cost function is minimized by the gradient descent technique, and the updated parameters and new parameters are calculated using the learning rate and Hadamard product. The decoder layer is removed from the encoder architecture, and the last fully connected layer is named the softmax layer. The softmax activation function is used to classify the fault type according to the highest value of the probability distribution, thereby realizing the fault identification of the DC distribution network. The DC side line fault detection and classification process is as follows Figure 5 shown.

[0083] Simulation verification:

[0084] In order to test the effectiveness of the invention, a two-terminal power supply DC distribution network simulation model was built in PSCAD / EMTDC. Figure 2 As shown in the figure. Different positions and transition resistances are set in the system, and simulation experiments are carried out on the DC side and AC side faults. 1 For example, for the DC side single-pole grounding fault, the transition resistances are considered to be 0.1Ω, 50Ω, and 100Ω respectively; for the bipolar short circuit, the transition resistances are considered to be 0.1Ω and 100Ω respectively. At the same time, the near end means that the fault point is 0.5km away from the head end of the line, and the far end means that the fault point is 0.5km away from the end of the line. For the AC side grounding fault, because only the fault near the converter valve side is considered, the transition resistance is set to a maximum of 100Ω considering the less favorable situation. The results of the fault detection scheme proposed in this article under different fault scenarios are shown in the table.

[0085]

[0086] From the simulation results, it can be seen that for DC side faults, whether they are near-end or far-end faults, the calculated energy value is greatly affected by the transition resistance and decreases significantly with the increase of the transition resistance. This is because the larger the transition resistance, the smaller the fault current amplitude, and the corresponding component amplitude will also decrease, and the corresponding energy will also decrease. However, for all AC and DC side faults, and the transition resistance is as high as 100Ω, the instantaneous energy of this component can be greater than the set threshold η=5.0×10 -8 kA2. In summary, the fault detection method proposed in this paper is effective for detecting all fault types.

[0087] This paper trains a SAE network for the two best natural mode components of the measured current at the head end of the DC distribution line. The network architecture of SAE is 200-100-50-4, Sigmoid is used as the activation function, and cross entropy is used as the cost function. Through simulation, fault current data containing different transition resistances and fault distances are obtained, including 50 groups of positive grounding fault data, 50 groups of negative grounding fault data, 90 groups of inter-pole short-circuit fault data, 90 groups of AC side single-phase grounding fault data, 50 groups of AC side two-phase grounding short-circuit data, and 50 groups of AC side three-phase grounding short-circuit data, a total of 380 groups of data. Each fault type is divided into 70% training set and 30% test set, and the data set is divided into 266 training sets and 114 test sets. Considering that the current collected by fault detection is generally the positive and negative current of the converter outlet, due to the large capacitor in parallel on the DC side of the converter, the positive and negative current waveforms are not much different; the current waveforms of the AC side two-phase grounding short-circuit fault and the AC side three-phase grounding short-circuit fault are also similar. Therefore, this paper considers to classify the positive and negative pole faults on the DC side as single-pole grounding faults; and classify the two-phase grounding short circuit and three-phase grounding short circuit on the AC side as AC side phase-to-phase short circuit faults. In summary, this paper considers four types of faults, namely, DC side single-pole grounding fault, represented by PG; bipolar short circuit fault, represented by PP; AC side single-phase grounding fault, represented by LG; and phase-to-phase short circuit grounding fault, represented by LL.

[0088] Figure 6is the confusion matrix obtained from the test set. The figure shows the number and accuracy of each fault type correctly classified: for PG, 30 samples are accurately classified, accounting for 100% of the category; for PP, 27 samples are accurately classified, also accounting for 100%. LG has 26 samples accurately classified, accounting for 96.3%. LL has 29 samples accurately classified, accounting for 96.7%. The figure shows the number and error rate of each fault type misclassified: LG has 1 sample misclassified as PG, accounting for 3.7%; LL has 1 sample misclassified as LG, accounting for 3.3%. The far right of the row shows the recall, which is the ratio of the diagonal elements (the number of correct predictions) to the total number of actual categories (the sum of the rows). In this chart, the recall of PG and PP is 100%, while the recall of LG and LL is slightly lower, at 96.3% and 96.7%, respectively. The bottom of the column shows the precision, which is the proportion of correct predictions, and is the ratio of the diagonal elements (the number of correct predictions) to the total number of predicted classes (the sum of the columns). In this graph, the precision of PP and LL is 100%, and the precision of PG and LG is 96.8% and 96.3%, respectively. The bottom right corner of the chart shows the overall accuracy, which is 98.2%, indicating that the model is accurate in 98.2% of all predictions. In summary, the chart shows that this classification model has high precision and high recall.

[0089] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A DC distribution line fault identification method based on improved VMD and SAE, characterized by: The following steps are involved: (1) The DC side current of the line is extracted in real time. The VMD algorithm improved by the Grey Wolf algorithm is used to decompose the DC side current signal by VMD with optimized parameters, and IMF1 is selected as the modal component reflecting the fault characteristics. (2) Based on the IMF1 characteristic mode component extracted in step (1), the Hilbert transform is used to identify the IMF1 to solve the amplitude and calculate the instantaneous energy of the signal component; the common characteristics of the instantaneous energy of different faults are used to establish the system fault judgment criteria; (3) According to the system fault judgment criteria established in step (2), the faults are divided into: single-pole grounding fault and bipolar short circuit fault on the DC side; single-phase grounding fault, two-phase grounding fault and three-phase grounding fault on the AC side; the kurtosis and correlation coefficient of the inherent modal components of each fault are calculated respectively, and the comprehensive indicators of different faults are obtained; (4) According to the different comprehensive fault indicators calculated in step (3), the two inherent modal components with the largest comprehensive indicators and the positive and negative voltage data are selected and input into SAE for fault feature extraction, and the extracted features are used for fault classification through softmax, thereby realizing fault identification of the DC distribution network.

2. A DC distribution line fault identification method based on improved VMD and SAE as claimed in claim 1, characterized in that: A VMD algorithm improved by the Grey Wolf Algorithm described in step (1): (1.1) Initialize the parameters of the gray wolf algorithm, including the number of wolves, the number of iterations, and the solution parameters; use the penalty factor and modal decomposition number of VMD as optimization variables, minimize the sample entropy as the fitness function, use the gray wolf algorithm for optimization, update the position and speed of each wolf according to the update rule, and calculate the fitness value of each wolf. According to the fitness value, update the optimal solution, suboptimal solution, and general solution; (1.2) According to the solution obtained in step (1), select sample entropy as the fitness function to determine the optimal solution, stop iteration when the local minimum entropy is reached during the search, output the penalty factor and modal decomposition number corresponding to the optimal solution and the suboptimal solution, and perform VMD decomposition on the signal according to the optimized parameters.

3. A DC distribution line fault identification method based on improved VMD and SAE as claimed in claim 1, characterized in that: System failure criteria described in step (2): After decomposing the DC side current signal by VMD to obtain each mode, the IMF1 modal component is Hilbert transformed and the instantaneous amplitude A of the signal is analyzed. n (t) and instantaneous phase θ n (t), construct the instantaneous energy E n (t) is used to describe the energy state of the signal at a certain instant, that is: In the formula, x n (t) represents the nth modal signal of the original signal; H[x n (t)] represents the output signal after Hilbert transform; X n (t) is the analytical signal, x n (t) and H[x n (t)] is its complex conjugate pair; Fault detection is performed by identifying the maximum value point in the energy time series diagram, and a threshold η is set to determine whether the system fails; E n (t)>η In the formula, the threshold η is set according to the maximum value of the instantaneous energy of IMF1 under the conditions of avoiding the maximum AC side power disturbance and load switching.

4. A DC distribution line fault identification method based on improved VMD and SAE as claimed in claim 1, characterized in that: The different comprehensive fault indicators described in step (3) include: kurtosis (KI), coefficient (CI) and comprehensive index (KCI).

5. A DC distribution line fault identification method based on improved VMD and SAE as claimed in claim 4, characterized in that: In the kurtosis (KI): the kurtosis of each modal component is calculated to assess the sharpness of its distribution; Among them, K i is the kurtosis of the ith modal component, μ i It is the IMF i is the mean value of , and N is the number of samples.

6. A DC distribution line fault identification method based on improved VMD and SAE as claimed in claim 4, characterized in that: In the correlation coefficient (CI): the Pearson correlation coefficient between each modal component and the original signal is calculated; Among them, R i is the modal component MFI i Pearson correlation coefficient between S and the original signal S; i Represents the original signal value of the i-th data point; is the modal component MFI i The average value of all data points; is the average value of all data points of the original signal S; n is the total number of data points.

7. A DC distribution line fault identification method based on improved VMD and SAE as claimed in claim 4, characterized in that: In the comprehensive index (KCI): the comprehensive index (KCI) is the product of the kurtosis and the absolute value of the correlation coefficient; KCI=|KI|*|CI| By calculating the comprehensive index of each intrinsic modal component, the two intrinsic modal components with the largest comprehensive index and the positive electrode voltage data are selected as the input data of SAE.

8. A DC distribution line fault identification method based on improved VMD and SAE as claimed in claim 1, characterized in that: A fault identification method described in step (4): By considering the Deep SAE encoder input, hidden layer output, and reconstructed decoder output, the weight W is introduced e , bias b e And the encoding function f e (.) to map out L encoded outputs E. In the pre-training stage, the encoder is represented by the following formula: Where, X n is the data of the nth input sample, is the weight parameter of the encoder, connecting the nth input unit and the lth encoder hidden layer unit; is the bias parameter of the lth hidden layer unit of the encoder; f e (·) is an activation function that applies a nonlinear transformation to the sum of weighted inputs; E nl is the output value of the lth hidden layer neuron after being processed by the encoder; L is each hidden layer neuron in the encoder network; The error is measured by the cross entropy loss function, with the lossless reconstruction input as the goal. The cross entropy error as the cost function is minimized by the gradient descent technique, and the updated parameters and new parameters are calculated using the learning rate and Hadamard product. The decoder layer is removed from the encoder architecture, and the last fully connected layer is named the softmax layer, which uses the softmax activation function and has the same number of nodes as the number of fault types, and classifies the fault type according to the highest value of the probability distribution.

9. An electronic device, characterized in that: It includes one or more processors for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors execute the DC distribution line fault identification method based on improved VMD and SAE as described in any one of claims 1 to 8.

10. A storage medium storing a computer program, wherein: in, The computer program is configured to execute the DC distribution line fault identification method based on improved VMD and SAE as described in any one of claims 1 to 8 when running.

Images