Data and model combined driving power distribution network cable early fault identification method

Abstract

The invention discloses a data and model combined driven early fault identification method for a power distribution network cable. The method comprises the following steps: establishing a power distribution network cable mathematical model considering the time-frequency characteristic and the volt-ampere characteristic of the cable; s2, according to the mathematical model of the power distribution network cable established in the step S1, under different insulation degradation degrees, calculating cable frequency-variable impedance, analyzing an impedance track, and extracting data characteristics of weak frequency-variable characteristics in an early fault state; s2, according to the power distribution network cable mathematical model established in the step S1, performing comparison improvement on the power distribution network cable mathematical model and a frequency-dependent cable model through the established equivalent circuit model, and constructing a parameter correction mechanism to establish a cable equivalent circuit model; according to the method, effective identification of the early fault of the power distribution network can be effectively realized, more accurate early fault feature extraction and effective identification can be realized by utilizing a data and model combined driving mode, and the method has relatively good universality, reliability and objectivity and can be widely applied to various engineering practices.

Description

Technical Field

[0001] The present invention relates to the technical field of distribution network fault identification, and in particular to a method for identifying early-stage faults of distribution network cables driven jointly by data and models. Background Art

[0002] With the accelerated cabling of urban distribution networks, early-stage faults (such as partial discharge and insulation carbonization) in underground cables due to insulation aging, mechanical damage, or environmental corrosion are increasing. The signals that cause these early-stage faults are weak and complex. Fault feature extraction must ensure both the extraction of abnormal signals from cable insulation damage models and the accurate extraction of fault signals under different distribution network operating conditions. Therefore, establishing a fault feature extraction method that is jointly driven by data and models, and utilizing lightweight convolutional neural networks (CNNs) for early-stage fault identification, is crucial for the safe and stable operation of power systems.

[0003] Currently, early fault identification in distribution networks primarily involves analyzing post-fault grid recording signals, utilizing electrical thresholds or artificial intelligence, manually setting thresholds, or using a single data-driven approach. Both approaches can achieve fault identification to a certain extent, but due to the nonlinear characteristics of fault signals during early-stage distribution network faults, the manually setting threshold approach cannot guarantee the reliability of early-stage fault identification for different types and degrees of severity. Data-driven approaches also fail to account for the differences in fault signals under varying degrees of insulation degradation. Relying solely on fault data recordings, they offer low robustness for identifying early-stage faults of varying severity. The characteristics of early-stage distribution network cable faults should include multidimensional features from both the model and the data. However, the current utilization of these multidimensional features is limited, making effective early-stage distribution network cable fault identification impossible and lacking a more specific description of the characteristics of early-stage cable faults.

[0004] At present, the main problems of distribution network fault identification methods are:

[0005] 1. The traditional frequency-dependent cable model requires strict setting of conductor geometric parameters. In actual working conditions, slight deviations in the parameters will lead to mismatch between the calculation of the frequency-dependent impedance matrix and the admittance matrix.

[0006] 2. The existing machine learning-based fault identification method only considers the data-driven approach and does not perform feature analysis on the impact caused by the different degrees of degradation of the cable model itself, resulting in low identification reliability.

[0007] Therefore, studying the data and model-driven early fault feature extraction method for distribution network and realizing early fault identification of distribution network is not only of theoretical research value, but also of important practical significance for engineering practice. To this end, a data and model-driven early fault identification method for distribution network cables is proposed. Summary of the Invention

[0008] The purpose of the present invention is to provide a method for early fault identification of distribution network cables driven jointly by data and models, so as to solve the problems raised in the above background technology.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a data and model-driven early fault identification method for distribution network cables, comprising the following steps:

[0010] S1. Establish a mathematical model of distribution network cables that takes into account the time-frequency characteristics and volt-ampere characteristics of the cables;

[0011] S2. Based on the mathematical model of the distribution network cable established in step S1, calculate the frequency-dependent impedance of the cable under different degrees of insulation degradation, analyze the impedance trajectory, and extract data features of the weak frequency-dependent characteristics under early fault conditions;

[0012] S3. Based on the mathematical model of the distribution network cable established in step S1, the equivalent circuit model is compared and improved with the frequency-dependent cable model, and a parameter correction mechanism is constructed to establish a cable equivalent circuit model;

[0013] S4. Based on the cable equivalent circuit model established in step S3, simulations are performed under different noise conditions, load conditions, and fault conditions to extract multidimensional model features of early faults under different operating and fault conditions;

[0014] S5. Based on the data and multi-dimensional features of the model under early fault conditions extracted in steps S2 and S4, a lightweight CNN decision-making system is established to realize early fault identification of distribution network cables.

[0015] Preferably, the mathematical model parameters of the distribution network cable in step S1 are:

[0016] Based on the time-frequency characteristics of the cable line transmission signal, the frequency-dependent resistance R(f) per unit length and the frequency-dependent inductance L(f) per unit length of the cable are designed respectively.

[0017]

[0018] Among them, R dc is the DC resistance, μ is the magnetic permeability, σ is the electrical conductivity, r is the conductor radius, D is the outer diameter of the insulation layer, and f is the frequency;

[0019] Based on the volt-ampere characteristics of the cable line transmission signal, the cable unit length voltage-related capacitance C (V) and unit length voltage-related conductance G (V) are designed respectively.

[0020] C(V)=C0(1+βV 2 )

[0021]

[0022] Where C0 and G0 are unit capacitance and conductance, β is the capacitance nonlinear coefficient, γ is the material characteristic parameter, and V is the voltage;

[0023] The mathematical model of the distribution network cable that was finally established considering the time-frequency characteristics and volt-ampere characteristics is:

[0024]

[0025] Preferably: the data feature extraction method for early stage faults in step S2 is:

[0026] By establishing a mathematical model for the cable, the frequency-dependent impedance Z(f) of the cable is calculated, the curvature k(f) of the frequency-dependent impedance trajectory is calculated, and the conductance and capacitance under the corresponding experimental voltage are calculated to form multi-dimensional data features;

[0027] Z(f)=R(f)+jωL(f)=R(f)+jX(f)

[0028]

[0029] Where X(f) is the reactance per unit length, R′ and X′ are the first-order derivatives, and X″ and R″ are the second-order derivatives.

[0030] Preferably, the cable equivalent circuit model in step S3 is:

[0031] The impedance and admittance matrices generated by the cable mathematical model are compared and analyzed with those of the simulation model. The multi-dimensional characteristic differences between the theoretical values ​​and the simulation values ​​are calculated. An evaluation function is established. The frequency-dependent cable model in PSCAD is used to modify the mathematical model parameters and establish a cable equivalent circuit model. The evaluation function is:

[0032]

[0033] in, is the maximum allowable error; Z theory 、Y theory is the impedance and admittance matrix of the theoretical model, Z sim 、Y sim is the impedance and admittance matrix of the simulation model; Z rated 、Y rated is the rated impedance and admittance matrix; PgP F is the Frobenius norm, and PgP2 is the spectral norm.

[0034] Preferably: the method for extracting the multidimensional model features of the early stage fault in step S4 is:

[0035] The following experiments are conducted in the early fault simulation of distribution network cable lines:

[0036] Samples with a signal-to-noise ratio of 10 to 20 dB of superimposed noise in the system;

[0037] Generate fault signatures for load rates between 30% and 120%;

[0038] Fault characteristics at different fault locations and initial phase angles;

[0039] The fault characteristics include fault three-phase current, zero-sequence current and three voltage waveforms.

[0040] Preferably, the lightweight CNN decision-making system in step S5:

[0041] Model feature branch: Deep separable convolution is used to process the differential features of the impedance spectrum;

[0042] Data feature branch: Combined with the Ghost module to achieve low-cost extraction of wavelet packet-sample entropy composite features;

[0043] A two-stage training strategy is performed for the above two features:

[0044] The first stage: training the model feature extractor, whose loss function L phy for:

[0045]

[0046] Among them, Z pred is the cable impedance frequency response curve predicted by CNN; Z meas Impedance data calculated by the cable mathematical model; is the total variation regularization term of the second-order derivative of impedance; λ is the physical regularization weight, which is a constant.

[0047] The data feature extractor is trained, and its loss function L data for:

[0048] L data =FocalLoss(y pred ,y label )

[0049] FocalLoss(p t )=-α t (1-p t ) γ log(p t )

[0050]

[0051] Among them, y predy is the classification / regression output of CNN for cable status; label is the measured label; FocalLoss is to improve the cross entropy loss to solve the problem of category imbalance; α t is the category weight factor, a constant, used to balance the ratio of positive and negative samples; γ is the modulation factor, a constant, which controls the weight attenuation of difficult and easy samples.

[0052] Preferred: Second stage: introduce distribution consistency loss term and use KL divergence to constrain the distribution consistency of dual-stream features:

[0053] L joint =αL phy +βL data +γ·D KL (p phy Pp data )

[0054] Among them, p phy is the hidden layer feature distribution of the physical feature extractor; p data is the hidden layer feature distribution of the data feature extractor; D KL is the KL divergence, which constrains the consistency of the dual-stream feature distribution; α is the physical loss weight, β is the data loss weight, and γ is the KL divergence weight.

[0055] Compared with the existing technology, the beneficial effects of the present invention are: the present invention can effectively realize the effective identification of early faults in the distribution network, and by using the joint driving method of data and model, it can realize more accurate early fault feature extraction and effective identification, and has good universality, reliability and objectivity, and can be widely used in various engineering practices. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 is a flow chart of the present invention;

[0057] Figure 2 It is the fault identification confusion matrix of the present invention. DETAILED DESCRIPTION

[0058] The technical solutions in the embodiments of the present invention will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present invention.

[0059] See also Figure 1-2 The present invention provides a technical solution: a method for early fault identification of distribution network cables driven by data and models, the flow chart is as follows Figure 1 As shown, the following steps are included:

[0060] 1. Establish a mathematical model of distribution network cables that takes into account the time-frequency characteristics and volt-ampere characteristics of the cables;

[0061] Based on the time-frequency characteristics of the cable line transmission signal, the frequency-dependent resistance R(f) per unit length and the frequency-dependent inductance L(f) per unit length of the cable are designed respectively;

[0062]

[0063] Among them, R dc is the DC resistance, μ is the magnetic permeability, σ is the electrical conductivity, r is the conductor radius, D is the outer diameter of the insulation layer, and f is the frequency;

[0064] Based on the volt-ampere characteristics of the cable line transmission signal, the cable unit length voltage-related capacitance C (V) and unit length voltage-related conductance G (V) are designed respectively.

[0065] C(V)=C0(1+βV 2 )

[0066]

[0067] Where C0 and G0 are unit capacitance and conductance, β is the capacitance nonlinear coefficient, γ is the material characteristic parameter, and V is the voltage.

[0068] The mathematical model of the distribution network cable that was finally established considering the time-frequency characteristics and volt-ampere characteristics is:

[0069]

[0070] 2. Based on the mathematical model of the distribution network cable established in step 1, calculate the cable frequency-dependent impedance Z(f) and the curvature k(f) of the frequency-dependent impedance trajectory under different degrees of insulation degradation, and calculate the conductance and capacitance under the corresponding experimental voltage to form multidimensional data features.

[0071] Z(f)=R(f)+jωL(f)=R(f)+jX(f)

[0072]

[0073] Where X(f) is the reactance per unit length, R′ and X′ are first-order derivatives, and X″ and R″ are second-order derivatives. This allows for analysis of the impedance trajectory and extraction of data features of weak frequency-varying characteristics under early fault conditions.

[0074] 3. Based on the distribution network cable mathematical model established in step 1, compare and analyze the impedance and admittance matrices generated by the cable mathematical model with the impedance and admittance matrices of the simulation model, calculate the multi-dimensional feature differences between the theoretical values ​​and the simulation values, establish an evaluation function, use the frequency-dependent cable model in PSCAD to modify the mathematical model parameters, and establish a cable equivalent circuit model. The evaluation function is:

[0075]

[0076] in, is the maximum allowable error; Z theory 、Y theory is the impedance and admittance matrix of the theoretical model, Z sim 、Y sim is the impedance and admittance matrix of the simulation model; Z rated 、Y rated is the rated impedance and admittance matrix; PgP F is the Frobenius norm, and PgP2 is the spectral norm.

[0077] 4. Based on the cable equivalent circuit model established in step 3, simulations are performed under different noise conditions, load conditions, and fault conditions. The following experiments are performed in the early fault simulation of the distribution network cable line:

[0078] 1. Samples with a signal-to-noise ratio of 10-20dB superimposed on the system;

[0079] 2. Generate fault signatures at different load rates (30%-120%);

[0080] 3. Fault characteristics of different fault locations and initial phase angles.

[0081] The fault characteristics include the three-phase current, zero-sequence current, and three voltage waveforms. Multi-dimensional model features of early-stage faults under different operating and fault conditions are extracted.

[0082] 5. Based on the early fault data and model multi-dimensional features extracted in steps 2 and 4, a lightweight CNN decision-making system is established:

[0083] Model feature branch: Deep separable convolution is used to process the differential features of the impedance spectrum;

[0084] Data feature branch: Combined with the Ghost module to achieve low-cost extraction of wavelet packet-sample entropy composite features.

[0085] A two-stage training strategy is performed for the above two features:

[0086] The first stage: training the model feature extractor, whose loss function L phy for:

[0087]

[0088] Among them, Z pred is the cable impedance frequency response curve predicted by CNN; Z meas Impedance data calculated by the cable mathematical model; is the total variation regularization term of the second-order derivative of impedance; λ is the physical regularization weight, which is a constant.

[0089] The data feature extractor is trained, and its loss function L data for:

[0090] L data =FocalLoss(y pred ,y label )

[0091] FocalLoss(p t )=-α t (1-p t ) γ log(p t )

[0092]

[0093] Among them, y pred y is the classification / regression output of CNN on cable status (such as aging degree, fault type); label is the measured label (such as insulation degradation level); FocalLoss is an improved cross entropy loss to solve the category imbalance problem; α t is the category weight factor, a constant, used to balance the ratio of positive and negative samples; γ is the modulation factor, a constant, which controls the weight attenuation of difficult and easy samples.

[0094] The second stage: introduce the distribution consistency loss term and use KL divergence to constrain the distribution consistency of dual-stream features

[0095] L joint =αL phy +βL data +γ·D KL (p phy Pp data )

[0096] Among them, p phy is the hidden layer feature distribution of the physical feature extractor; p data is the hidden layer feature distribution of the data feature extractor (such as the CNN feature map of the time domain waveform); D KL is the KL divergence, which constrains the consistency of the dual-stream feature distribution; α is the physical loss weight, β is the data loss weight, and γ is the KL divergence weight.

[0097] Ultimately, early fault identification of the distribution network is achieved.

[0098] Simulation Verification

[0099] To test the effectiveness of this invention, a lightweight CNN was trained for early-stage cable faults, operational disturbances, and other distribution network faults. The CNN architecture had an input dimension of 192×128 and an output dimension of 3, with the outputs being early-stage faults, other faults, and system disturbances, respectively. The lightweight CNN employed a Ghost-DSC hybrid architecture with heterogeneous feature processing. Ten sets of model feature data were obtained through MATLAB experiments. For the ten different cable insulation degradation scenarios, distribution network simulations were used to generate 30 sets of fault current data for different transition resistances, 30 sets of fault current data for different fault distances, 40 sets of fault current data for different load factors, and 20 sets of fault current data for different noise levels. Furthermore, 100 sets of single-phase-to-ground fault current data, 50 sets of two-phase short-circuit fault current data, and 50 sets of distribution network source-load disturbance current data were generated, for a total of 320 data sets. Each identification type was split into a 70% training set and a 30% test set, resulting in 224 training sets and 96 test sets. This paper considers classifying single-phase ground faults and two-phase short circuit faults as other faults, and classifying early faults under different operating conditions and fault conditions as cable early faults. In summary, this paper considers three identification types, namely cable early faults, represented by IF; other faults, represented by OF; distribution network source and load disturbances, represented by SCD. The classification results are shown in Figure 2. Figure 2 shown.

[0100] As shown in the figure, the rightmost side of the row shows the proportion of correct predictions, which is the ratio of the diagonal elements to the total number of predicted categories; the lowest end of the column shows the recall rate, which is the ratio of the diagonal elements to the total number of predicted categories. This method achieves high recall and precision for both IF and OF, exceeding 97%. For SCD, both recall and precision are 100%, indicating that this method is unaffected by power grid source and load fluctuations and can accurately identify faults. It also distinguishes early-stage distribution network faults from other faults with high accuracy, achieving early-stage distribution network fault identification.

[0101] To demonstrate the reliability and accuracy of this invention, this paper compares the proposed method with other methods and compares the early fault identification performance using a combined model and data approach with that driven by a single model or data approach. The specific comparison results are shown in Table 1.

[0102] As shown in Table 1, the proposed method for early fault identification of distribution network cables based on a lightweight CNN, jointly driven by data and models, has high accuracy in the distribution network cable early fault identification scenario. Compared with raw data driven only by the model or only by data, the method used in this paper, jointly driven by data and models, has a significantly improved early fault detection rate of 97.2%. Compared with the other two neural network identification models, this method may be less effective when driven by raw data alone, such as the LSTM with data-driven alone, has a higher early fault identification success rate than the lightweight CNN. However, when jointly driven by data and model, the lightweight CNN has a higher early fault detection rate and is simpler and lighter than the other two models, meeting the actual needs of the distribution network.

[0103] Table 1 Comparison of fault identification accuracy of different features or methods

[0104]

[0105] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A data and model-driven early fault identification method for distribution network cables, characterized by: The following steps are involved: S1. Establish a mathematical model of distribution network cables that takes into account the time-frequency characteristics and volt-ampere characteristics of the cables; S2. Based on the mathematical model of the distribution network cable established in step S1, calculate the frequency-dependent impedance of the cable under different degrees of insulation degradation, analyze the impedance trajectory, and extract data features of the weak frequency-dependent characteristics under early fault conditions; S3. Based on the mathematical model of the distribution network cable established in step S1, the equivalent circuit model is compared and improved with the frequency-dependent cable model, and a parameter correction mechanism is constructed to establish a cable equivalent circuit model; S4. Based on the cable equivalent circuit model established in step S3, simulations are performed under different noise conditions, load conditions, and fault conditions to extract multidimensional model features of early faults under different operating and fault conditions; S5. Based on the data and multi-dimensional features of the model under early fault conditions extracted in steps S2 and S4, a lightweight CNN decision-making system is established to realize early fault identification of distribution network cables.

2. The method for early fault identification of distribution network cables driven by data and model according to claim 1 is characterized by: The mathematical model parameters of the distribution network cable in step S1 are: Based on the time-frequency characteristics of the cable line transmission signal, the frequency-dependent resistance R(f) per unit length and the frequency-dependent inductance L(f) per unit length of the cable are designed respectively. Among them, R dc is the DC resistance, μ is the magnetic permeability, σ is the electrical conductivity, r is the conductor radius, D is the outer diameter of the insulation layer, and f is the frequency; Based on the volt-ampere characteristics of the cable line transmission signal, the cable unit length voltage-related capacitance C (V) and unit length voltage-related conductance G (V) are designed respectively. C(V)=C0(1+βV 2 ) Where C0 and G0 are unit capacitance and conductance, β is the capacitance nonlinear coefficient, γ is the material characteristic parameter, and V is the voltage; The mathematical model of the distribution network cable that was finally established considering the time-frequency characteristics and volt-ampere characteristics is:

3. The method for early fault identification of distribution network cables driven by data and model according to claim 1 is characterized by: In step S2, the data feature extraction method for early stage faults is as follows: By establishing a mathematical model for the cable, the frequency-dependent impedance Z(f) of the cable is calculated, the curvature k(f) of the frequency-dependent impedance trajectory is calculated, and the conductance and capacitance under the corresponding experimental voltage are calculated to form multi-dimensional data features; Z(f)=R(f)+jωL(f)=R(f)+jX(f) Where X(f) is the reactance per unit length, R′ and X′ are the first-order derivatives, and X″ and R″ are the second-order derivatives.

4. The method for early fault identification of distribution network cables driven by data and model according to claim 1 is characterized by: The cable equivalent circuit model in step S3 is: The impedance and admittance matrices generated by the cable mathematical model are compared and analyzed with the impedance and admittance matrices of the simulation model. The multi-dimensional characteristic differences between the theoretical values ​​and the simulation values ​​are calculated. An evaluation function is established. The frequency-dependent cable model in PSCAD is used to modify the mathematical model parameters and establish a cable equivalent circuit model. The evaluation function is: in, is the maximum allowable error; Z theory 、Y theory is the impedance and admittance matrix of the theoretical model, Z sim 、Y sim is the impedance and admittance matrix of the simulation model; Z rated 、Y rated is the rated impedance and admittance matrix; PgP F is the Frobenius norm, and PgP2 is the spectral norm.

5. The method for early fault identification of distribution network cables driven by data and model according to claim 1 is characterized by: In step S4, the method for extracting the multi-dimensional model features of the early stage fault is as follows: The following experiments are conducted in the early fault simulation of distribution network cable lines: Samples with a signal-to-noise ratio of 10 to 20 dB of superimposed noise in the system; Generate fault signatures for load rates between 30% and 120%; Fault characteristics at different fault locations and initial phase angles; The fault characteristics include fault three-phase current, zero-sequence current and three voltage waveforms.

6. The method for early fault identification of distribution network cables driven by data and model according to claim 1 is characterized by: The lightweight CNN decision-making system in step S5: Model feature branch: Deep separable convolution is used to process the differential features of impedance spectrum; Data feature branch: Combined with the Ghost module to achieve low-cost extraction of wavelet packet-sample entropy composite features; A two-stage training strategy is performed for the above two features: The first stage: training the model feature extractor, whose loss function L phy for: Among them, Z pred is the cable impedance frequency response curve predicted by CNN; Z meas Impedance data calculated by the cable mathematical model; is the total variation regularization term of the second-order derivative of impedance; λ is the physical regularization weight, which is a constant; The data feature extractor is trained, and its loss function L data for: L data =FocalLoss(y pred ,the label ) FocalLoss(p t )=-α t (1-p t ) γ log(p t ) Among them, y pred y is the classification / regression output of CNN for cable status; label is the measured label; FocalLoss is to improve the cross entropy loss and solve the problem of category imbalance; α t is the category weight factor, a constant, used to balance the ratio of positive and negative samples; γ is the modulation factor, a constant, which controls the weight attenuation of difficult and easy samples.

7. The method for early fault identification of distribution network cables driven by data and model according to claim 6 is characterized by: The second stage: introduce the distribution consistency loss term and use KL divergence to constrain the distribution consistency of dual-stream features: L joint =αL phy +βL data +γ·D KL (p phy P p data ) Among them, p phy is the hidden layer feature distribution of the physical feature extractor; p data is the hidden layer feature distribution of the data feature extractor; D KL is the KL divergence, which constrains the consistency of the dual-stream feature distribution; α is the physical loss weight, β is the data loss weight, and γ is the KL divergence weight.

Images