An enhanced method and system for power distribution network ground fault simulation waveform data

Abstract

This invention discloses a method and system for enhancing simulated waveform data of grounding faults in distribution networks, relating to the field of distribution network fault diagnosis and data processing technology. The method includes: acquiring real fault waveform data from a distribution network project site; building a distribution network simulation model based on the network topology and operating parameters, and generating an ideal simulated waveform sequence corresponding to the real fault waveform data; extracting features from the real fault waveform data and the ideal simulated waveform sequence to obtain difference features and constructing a real disturbance feature set; dynamically training a pre-constructed reconstructed network to learn the disturbance distribution patterns in real operating conditions under electrical and physical constraints; generating new ideal simulated waveform sequences based on various distribution network fault scenarios and topology parameters using the distribution network simulation model; and inputting these sequences into the trained reconstructed network for disturbance injection and waveform reconstruction to generate enhanced simulated waveform data with real disturbance characteristics.

Description

Technical Field

[0001] This invention relates to a method and system for enhancing simulated waveform data of grounding faults in power distribution networks, belonging to the field of power distribution network fault diagnosis and data processing technology. Background Technology

[0002] Single-phase grounding faults in distribution networks, especially high-resistance grounding faults, exhibit weak transient characteristics, making accurate detection and fault location a persistent challenge in the field of relay protection. In recent years, data-driven AI-based fault diagnosis models have become a research hotspot. However, real-world high-resistance grounding fault recording data from engineering sites is extremely scarce, making it difficult to meet the demand for massive training samples required for deep learning.

[0003] Currently, the industry commonly uses simulation software to generate massive amounts of fault data to replace real data for model training. However, existing simulation models are too idealized and cannot accurately depict the complex electromagnetic transient noise, nonlinear transmission errors of instrument transformers, and severe waveform distortion caused by load fluctuations, resulting in a huge "domain offset" between simulation data and actual operating conditions. Traditional simulation data augmentation methods rely solely on simple mathematical techniques such as superimposing Gaussian white noise, which cannot reproduce the underlying physical distortion characteristics of real electrical equipment.

[0004] The aforementioned shortcomings directly result in severely limited generalization capabilities of AI recognition models trained on purely ideal simulation data in actual power distribution network engineering applications; when faced with complex on-site conditions, they are highly susceptible to maloperation or failure to operate by relay protection devices. Therefore, there is an urgent need for a method that can effectively inject the characteristics of real-world physical disturbances into the simulation waveforms to bridge the gap between simulation and reality. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for enhancing simulation waveform data of grounding faults in power distribution networks. By learning the disturbance distribution law from a small number of real waveform recordings and injecting it into a massive number of simulation waveforms under the constraints of electrical physical mechanisms, the "domain offset" problem between simulation data and real operating conditions is solved, and the consistency and diversity of training samples are enhanced.

[0006] To achieve the above objectives, the present invention is implemented using the following technical solution.

[0007] On one hand, the present invention provides a method for enhancing simulation waveform data of grounding faults in distribution networks, the method comprising:

[0008] Obtain real fault recording data from power distribution network engineering sites;

[0009] Based on the power grid topology and operating parameters at the power distribution network site, a power distribution network simulation model is built, and an ideal simulation waveform sequence corresponding to the actual fault recording data is generated.

[0010] Feature extraction is performed on the real fault recording data and the ideal simulation waveform sequence to obtain difference features, and a real disturbance feature set is constructed based on the difference features;

[0011] Based on the real disturbance feature set, the pre-constructed reconstruction network is dynamically trained; wherein, the trained reconstruction network, under the constraints of electrical and physical mechanisms, is able to learn the disturbance distribution law in real working conditions;

[0012] Based on the aforementioned power distribution network simulation model, new ideal simulation waveform sequences are generated using various power distribution network fault scenarios and topology parameters.

[0013] The newly added ideal simulation waveform sequence is input into the trained reconstruction network to perform perturbation injection and waveform reconstruction, generating enhanced simulation waveform data with real perturbation characteristics.

[0014] Furthermore, the method also includes continuously extracting a data buffer window containing a preset number of power frequency cycles before and after the fault transient process, wherein the total number of discrete sampling points contained in the data buffer window is as shown in the following formula:

[0015] (1),

[0016] In the formula, K represents the total number of discrete sampling points; K represents the number of power frequency cycles. This indicates the actual sampling frequency of the waveform recording device or relay protection device at the power distribution network construction site; Indicates the rated power frequency of the power distribution network system;

[0017] The time series set included in the data buffer window is shown in the following formula:

[0018] (2),

[0019] In the formula, W represents the time series set; This represents the pre-fault steady-state data segment used to provide a reference for the system background before the fault. This represents a data segment that records the high-frequency characteristics and abrupt changes in the fault transient process at the instant of the fault. This represents the post-fault steady-state data segment that characterizes the trend of fault evolution.

[0020] Furthermore, before extracting features from the real fault waveform data and the ideal simulation waveform sequence, the method further includes aligning the real fault waveform data and the ideal simulation waveform sequence along the time axis, specifically including:

[0021] The time-frequency domain analysis algorithm is used to locate the initial fault moment of the actual fault recording data.

[0022] Using the initial moment of the fault as the reference alignment point, the simulated waveform data at the corresponding moment in the ideal simulated waveform sequence is extracted, so that the real fault waveform data is synchronized with the ideal simulated waveform sequence.

[0023] Furthermore, the step of using time-frequency domain analysis algorithms to locate the initial fault time in the actual fault recording data specifically includes:

[0024] Obtain the zero-sequence voltage signal sequence or zero-sequence current signal sequence from the real fault recording data, remove the DC component, and then perform multi-scale discrete wavelet transform on the zero-sequence voltage signal sequence or zero-sequence current signal sequence.

[0025] Based on the Mallat algorithm, high-frequency detail coefficients reflecting the transient change characteristics of faults are extracted from the zero-sequence voltage signal sequence or zero-sequence current signal sequence after wavelet transform.

[0026] Based on the high-frequency detail coefficients, a wavelet transform modulus sequence is constructed, the energy of the detail coefficients is calculated and smoothed, and the sampling time corresponding to the maximum value of the smoothed energy sequence is set as the fault initial time.

[0027] Furthermore, the true perturbation feature set includes one or more of the following:

[0028] High-frequency noise at the construction site, amplitude distortion caused by transformer transmission and line voltage drop, voltage and current jitter caused by load fluctuations, unbalanced voltage and current, and transient effects at the moment of fault.

[0029] Furthermore, the pre-constructed reconstructed network includes an ideal waveform feature encoding module, a perturbation parameter latent space mapping module, and a realistic waveform decoding module, wherein:

[0030] The waveform feature encoding module employs a one-dimensional deep convolutional neural network to extract time-series dependencies through multi-layer causal convolution. It receives the ideal simulated waveform sequence and extracts the pure transient and steady-state features of the sequence to obtain a pure time-series feature vector. The discrete feature extraction of the k-th convolutional layer of the waveform feature encoding feature module is shown in the following expression:

[0031] (3),

[0032] In the formula, Indicates the first The output of the first layer of convolution The feature values ​​of each sampling point; Indicates the kernel size; Indicates the first The first layer of convolution kernel Each weighting coefficient; Indicates the first The bias of the convolutional kernel;

[0033] The perturbation parameter latent space mapping module employs a latent variable generator based on a variational autoencoder architecture, which maps features to a continuous latent space using a multilayer perceptron. It receives the real perturbation feature set and maps the features in the real perturbation feature set to learnable perturbation parameters, as shown in the following expression:

[0034] (4),

[0035] In the formula, Z represents the learnable perturbation parameter; Z represents the latent space variable; The vector representing the mean of the latent space distribution; The variance vector representing the latent space distribution; This represents a random noise vector sampled from a standard normal distribution;

[0036] The realistic waveform decoding module includes a feature fusion layer and a deconvolutional reconstruction network, used to input the perturbation parameters into the clean time series feature vector to generate enhanced simulation waveform data with realistic perturbation features, as shown in the following expression:

[0037] (5),

[0038] In the formula, This represents the enhanced simulation waveform data after fusion; Represents the pure time series feature vector Mean value along the channel dimension; Represents the pure time series feature vector Standard deviation in the channel dimension; Indicates the scaling factor; This represents the translation bias factor.

[0039] Furthermore, during the dynamic training of the pre-constructed reconstruction network based on the real perturbation feature set, the total loss function used includes distribution alignment loss, electrophysical mechanism constraint loss, and adaptive anti-malfunction penalty loss, as shown in the following expression:

[0040] (6),

[0041] In the formula, Represents the total loss function; This represents the distribution alignment loss; Indicates electrical and physical mechanism constraint loss; This indicates an adaptive penalty loss for preventing accidental activation; , as well as These represent the dynamic weighting coefficients for alignment loss, electrophysical mechanism constraint loss, and adaptive anti-maloperation penalty loss, respectively.

[0042] The distribution alignment loss This is used to align the distribution characteristics of the generated enhanced simulation waveform data with those of the real fault recording data;

[0043] The electrical physical mechanism constraint loss Transient zero-sequence energy in enhanced simulation waveform data used for constraint extraction The polarity and theoretical energy of the ideal simulated waveform sequence The polarity remains consistent, as shown in the following expression:

[0044] (7),

[0045] In the formula, Indicates the initial moment of the fault; Indicates a short time window following the occurrence of the fault; This represents the zero-sequence voltage transient signal in the enhanced simulation waveform data at time t; This represents the zero-sequence current transient signal in the enhanced simulation waveform data at time t.

[0046] The adaptive anti-misoperation penalty loss This is used to constrain the transient zero-sequence voltage peak value in the enhanced simulation waveform data generated when the input is a non-grounded fault. Less than the adaptive threshold The adaptive threshold is used to dynamically adjust the amplitude distortion parameter in the latent space, as shown in the following expression:

[0047] (8),

[0048] In the formula, , Represents the reliability coefficient; This represents the maximum background noise amplitude extracted from the power distribution network project; This represents the unbalanced mean of the steady-state zero-sequence voltage before the fault. This is the activation function.

[0049] Furthermore, while generating enhanced simulation waveform data with realistic perturbation characteristics, the method also includes:

[0050] Generate multi-dimensional structured classification label vectors for the enhanced simulation waveform data. The specific structure is defined as shown in the following expression:

[0051] (9),

[0052] In the formula, This indicates the label of the initial fault time determined through time-frequency domain analysis; Indicates the initial phase angle label of the fault; Label indicating fault grounding resistance; Labels indicating high-frequency noise signal-to-noise ratio levels; Labels indicating the degree of distortion;

[0053] The time series matrix of the enhanced simulation waveform data is compared with the corresponding structured classification label vector. Corresponding binding and storage are used to construct a high-fidelity augmented dataset.

[0054] Furthermore, the method also includes using the high-fidelity augmented dataset to perform supervised training on the downstream distribution network single-phase grounding fault identification model, so as to improve the anti-interference capability and generalization accuracy of the downstream distribution network single-phase grounding fault identification model.

[0055] On the other hand, the present invention provides an enhancement system for simulation waveform data of grounding faults in distribution networks, comprising:

[0056] The data acquisition module is configured to: acquire real fault waveform data from the power distribution network project site; build a power distribution network simulation model based on the power grid topology and operating parameters of the power distribution network project site, and generate an ideal simulation waveform sequence corresponding to the operating conditions of the real fault waveform data;

[0057] The feature extraction module is configured to: extract features from the real fault recording data and the ideal simulation waveform sequence to obtain difference features, and construct a real disturbance feature set based on the difference features;

[0058] The network training module is configured to: dynamically train the pre-constructed reconstructed network based on the real disturbance feature set; wherein the trained reconstructed network is able to learn the disturbance distribution pattern in real working conditions;

[0059] The waveform generation module is configured to generate a new ideal simulation waveform sequence based on various distribution network fault scenarios and topology parameters using the distribution network simulation model.

[0060] The waveform enhancement module is configured to: input the newly added ideal simulation waveform sequence into the trained reconstruction network, perform perturbation injection and waveform reconstruction, and generate enhanced simulation waveform data with real perturbation characteristics.

[0061] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0062] This invention extracts differences between the acquired real fault waveform data and the ideal simulation waveforms of the corresponding operating conditions to construct a real disturbance feature set, solving the problem of difficulty in directly modeling complex disturbances in the field. Furthermore, under the constraints of electrical and physical mechanisms, a reconstruction network is trained to learn the disturbance distribution patterns in real operating conditions, ensuring that the generated enhanced waveforms maintain consistency with real faults in core physical laws such as transient energy directionality, avoiding the defects of traditional data augmentation methods that destroy the physical characteristics of faults. In addition, after generating diverse ideal simulation waveforms in batches, the trained network is used for disturbance injection and waveform reconstruction, ultimately generating enhanced simulation waveform data that contains both real disturbance characteristics and strictly follows electrical and physical laws. This solves the problem of extreme scarcity of real fault samples at extremely low cost, while significantly bridging the "domain offset" between simulation data and real operating conditions. Attached Figure Description

[0063] Figure 1 The diagram shows the architecture of an enhancement method for simulation waveform data of grounding faults in a power distribution network provided by the present invention. Detailed Implementation

[0064] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0065] The term "and / or" simply describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0066] Example 1

[0067] See Figure 1 This embodiment provides a method for enhancing simulated waveform data of grounding faults in distribution networks. It does not rely on massive amounts of field waveform data, but only requires a small number of real samples to deeply mine and extract complex electromagnetic transient noise and nonlinear transmission characteristics of transformers. The method specifically includes:

[0068] Step S1: Obtain real fault waveform data from the power distribution network project site; build a power distribution network simulation model based on the power grid topology and operating parameters at the power distribution network project site, and generate an ideal simulation waveform sequence corresponding to the operating conditions of the real fault waveform data;

[0069] When a fault occurs in the distribution network, real fault waveform data is obtained from the distribution network project site. As a specific implementation method, the real fault waveform data is the real single-phase grounding fault waveform data.

[0070] Meanwhile, in practical engineering, the field waveform recording files generated by waveform recording devices or relay protection devices are usually stored in CSV or COMTRADE format, containing a massive number of long-time-series sampling points. To accurately capture the entire transient process of a single-phase ground fault while also considering the hardware processing bottleneck of the relay protection device, this embodiment does not perform indiscriminate learning on the entire waveform recording. Instead, it continuously extracts a data buffer window containing a preset number of cycles before and after the fault transient process, achieving efficient and complete capture of the fault's key information. The data buffer window completely includes the three processes before, during, and after the fault, and the time series set W of the data buffer window is shown in the following formula:

[0071] (2),

[0072] Steady-state data segment before the fault It can provide a reference benchmark for the enhancement method. By using the steady-state data of the first few cycles before the fault, the reference values ​​of background noise, unbalanced voltage and unbalanced current during normal system operation can be calculated. The data in the preceding process provides a basis for the model to distinguish between "normal" and "fault" states.

[0073] Fault transient process data segment It is mainly used to capture the core features when a fault occurs, such as recording the high-frequency features and sudden changes in distribution at the moment of the fault. The key to identifying single-phase grounding faults in distribution networks lies in the high-frequency transient process at the moment the fault occurs. The transient process data segment contains rich fault information, such as high-frequency oscillations and sudden changes in polarity.

[0074] Fault transient process data segment The main purpose is to characterize the fault evolution trend. Data from several cycles after the fault can reflect the waveform shape after the fault enters a steady state, and can be used to characterize the persistence and evolution trend of the fault.

[0075] In this embodiment, the data buffer window is preferably a data window of 10 consecutive power frequency cycles, wherein the first three cycles are divided into the steady-state data segment before the fault. The middle two cycles are divided into fault transient process data segments. The last five cycles are divided into fault transient process data segments. .

[0076] The total number of discrete sampling points contained in the data buffer window is shown in the following formula (1):

[0077] (1),

[0078] In the formula, K represents the total number of discrete sampling points; K represents the number of power frequency cycles, preferably K=10; The frequency of the waveform recording device or relay protection device at the power distribution network construction site is preferably the actual sampling frequency. ; Indicates the rated power frequency of the power distribution network system, preferably or The data length of the data buffer window is matched with the size of the ring data buffer array of the relay protection device, ensuring that the data generated by the method provided in this embodiment is strictly consistent with the data that the protection device can actually collect and process in terms of physical characteristics.

[0079] Meanwhile, based on the power grid topology and operating parameters at the distribution network engineering site, a distribution network simulation model is built in simulation software. The power grid topology and operating parameters include, but are not limited to, the actual line impedance, transformer capacity, and grounding parameters on site. After the distribution network simulation model is built, the same operating conditions as the actual fault recording data are set, and the distribution network simulation model is run to generate the corresponding ideal simulation waveform sequence, wherein the ideal simulation waveform sequence is consistent with the actual fault recording data in length.

[0080] A paired set of "real waveform recording data - ideal simulation waveform sequence" was obtained. The real waveform recording data contains real disturbance characteristics such as complex electromagnetic transient noise and transformer transmission errors, while the ideal simulation waveform sequence is a clean, disturbance-free ideal waveform. This paired data will serve as the basis for subsequent disturbance feature extraction and reconstruction network training.

[0081] Step S2: Extract features from the real fault recording data and the ideal simulation waveform sequence to obtain difference features, and construct a real disturbance feature set based on the difference features;

[0082] Before extracting features from the actual fault waveform data and the ideal simulation waveform sequence, the method proposed in this embodiment also includes time axis alignment processing of the actual fault waveform data and the ideal simulation waveform sequence, specifically including:

[0083] Step S2A1: Use a time-frequency domain analysis algorithm to locate the initial fault time of the actual fault recording data;

[0084] Step S2A1B1: Obtain the zero-sequence voltage signal sequence or zero-sequence current signal sequence from the real fault waveform data, remove the DC component, and then perform multi-scale discrete wavelet transform on the zero-sequence voltage or zero-sequence current signal sequence.

[0085] First, the DC component is removed from the zero-sequence voltage and zero-sequence current signals, as shown in the following expression:

[0086] (10)

[0087] In the formula, This represents the zero-sequence voltage signal at the nth sampling point after removing the DC component. This represents the zero-sequence voltage signal at the original nth sampling point. This represents the mean of the zero-sequence voltage signal sequence; This represents the zero-sequence current signal at the nth sampling point after removing the DC component. This represents the zero-sequence current signal at the original nth sampling point. This represents the mean of the zero-sequence current signal sequence.

[0088] As a specific implementation method, this embodiment selects the db4 wavelet as the mother wavelet function to perform multi-scale discrete wavelet transform on the signal sequence. When a single-phase ground fault occurs in the distribution network, the transient signal exhibits instantaneous abrupt changes, such as high-frequency oscillations or voltage drops. The db4 wavelet, with its good compact support characteristics, can accurately capture the precise location of this abrupt change without causing ambiguity due to the "tailing" of the wavelet function. Simultaneously, since the db4 wavelet supports the fast discrete wavelet transform of the Mallat algorithm, it can decompose the signal into approximation coefficients and detail coefficients at different scales, enabling full-band analysis from low to high frequencies. This adapts to variations in the characteristic frequencies of different faults, allowing the method to be adapted to different types of ground faults (such as metallic grounding, high-resistance grounding, and arcing grounding), thus improving the robustness of the algorithm.

[0089] Step S2A1B2: Based on the Mallat algorithm, extract the high-frequency detail coefficients reflecting the transient change characteristics of the fault in the zero-sequence voltage signal sequence or zero-sequence current signal sequence after wavelet transform;

[0090] As a specific implementation method, this embodiment is based on the Mallat algorithm. It uses an orthogonal filter bank of db4 wavelets to perform layer-by-layer discrete convolution and downsampling on the wavelet-transformed zero-sequence voltage signal sequence or zero-sequence current signal sequence to extract the first-order wavelet that reflects the transient change characteristics of the fault. High-frequency detail factor of the layer Specifically, as shown in the following expression:

[0091] (11),

[0092] In the formula, Represents the low-frequency approximation coefficients of the j-th layer; Indicates the translation parameter; This represents the coefficients of a high-pass filter of length 8 corresponding to the db4 wavelet, where, Indicates the position index of the high-pass filter; This represents the coefficients of a low-pass filter with a length of 8 corresponding to the db4 wavelet; This represents the low-frequency approximation coefficients of the (j-1)th layer, where n represents the nth sampling point; when j=0, This represents the original input zero-sequence voltage signal sequence or zero-sequence current signal sequence.

[0093] Step S2A1B3: Construct a wavelet transform modulus sequence based on the high-frequency detail coefficients, calculate the energy of the detail coefficients and perform smoothing, and set the sampling time corresponding to the maximum value of the smoothed energy sequence as the fault initial time.

[0094] As a specific implementation method, this embodiment uses a time-frequency domain analysis algorithm to locate the initial moment of the fault, specifically including:

[0095] The first High-frequency detail factor Arrange them in chronological order to construct a wavelet transform modulus sequence. And calculate the detailed system energy corresponding to each time point, as shown in the following expression:

[0096] (12),

[0097] In the formula, Indicates the signal at the sampling time The magnitude of the energy at that location, where, .

[0098] Since background noise and various irregular disturbances are unavoidable in the field signal, in order to effectively overcome the misjudgment of single-point abrupt changes caused by random white noise in the field, a sliding smoothing process is performed on the energy sequence. Assuming a smoothing algorithm of size [value missing] is used... If a sliding window is used to smooth the energy sequence, the smoothed energy obtained through convolution is shown in the following expression:

[0099] (13)

[0100] In the formula, i represents the summation index variable. The smoothed energy sequence is... It can effectively filter out burr interference, in The maximum value is found through global optimization, and the sampling index corresponding to this maximum value is the precise fault initiation time. .

[0101] Therefore, the final failure moment can be determined by finding a smoothed energy sequence. The maximum value in the range is used to locate the sampling time corresponding to the maximum value. That is, set as the initial time of the fault. .

[0102] Step S2A2: At the initial time of the fault Using the reference alignment point, the simulated waveform data at the corresponding time in the ideal simulated waveform sequence is extracted, so that the real fault recording data and the ideal simulated waveform sequence are synchronized at the start time of the transient change.

[0103] Since real fault recording data is usually based on the trigger time of the recording device, and the trigger time may be later than the actual time the fault occurred, the initial fault time is used as the starting point. Using the simulation waveform sequence as a reference alignment point, the simulation waveform sequence is trimmed or shifted to ensure that the real waveform and the simulation waveform are synchronized at the starting point of the transient change, so that subsequent work can be carried out under a unified time reference.

[0104] As a specific implementation method, the actual waveform recording data after time axis alignment is subtracted point by point from the ideal simulated waveform sequence to obtain the difference signal sequence. Frequency domain analysis is performed on this difference sequence to extract the energy distribution characteristics of the transient frequency band; simultaneously, statistical characteristics such as amplitude deviation and unbalance are calculated for the steady-state segment before the fault. These features are combined to construct a true disturbance feature set characterizing the electromagnetic interference in the field and the transmission error of the instrument transformer in the protection device. The true disturbance feature set includes one or more of the following: high-frequency noise in the engineering field, amplitude distortion caused by instrument transformer transmission and line voltage drop, voltage and current jitter caused by load fluctuations, unbalanced voltage and current, and transient effects at the moment of the fault.

[0105] Step S3: Based on the real disturbance feature set, dynamically train the pre-constructed reconstruction network; wherein, the trained reconstruction network, under the constraints of electrical and physical mechanisms, is able to learn the disturbance distribution law in real working conditions;

[0106] The pre-built reconstruction network in this embodiment includes an ideal waveform feature encoding module, a perturbation parameter latent space mapping module, and a realistic waveform decoding module, wherein:

[0107] Ideal waveform feature encoding module: Employs a one-dimensional deep convolutional neural network to receive the ideal simulated waveform sequence. As input, time series dependencies are extracted through multi-layer causal convolution, and the output is a pure transient and steady-state multi-dimensional time series feature vector representing the ideal simulation waveform at different time scales. The discrete feature extraction of the k-th convolutional layer of the waveform feature encoding feature module is shown in the following expression:

[0108] (3),

[0109] In the formula, The feature value of the nth sampling point in the output of the k-th convolutional layer is represented by M; M represents the kernel size. This represents the m-th weight coefficient of the k-th convolutional kernel; This represents the bias of the k-th convolutional kernel.

[0110] After nonlinear rectification using the ReLU activation function, negative invalid responses in the feature map are filtered out. Then, the difference features extracted from the paired real waveform data, i.e. the real disturbance feature set, are sent to the disturbance parameter latent space mapping module.

[0111] The disturbance parameter latent space mapping module employs a latent variable generator based on a variational autoencoder (VAE) architecture, receiving the real disturbance feature set as prior condition input. Through a multilayer perceptron, it maps the high-frequency noise, amplitude distortion, and load fluctuation features from the engineering site into a continuous latent space, outputting a mean vector of a multivariate Gaussian distribution. With variance vector Furthermore, a reparameterization technique is employed for sampling to generate a learnable disturbance parameter vector that incorporates the characteristics of the on-site physical disturbance, as shown in the following expression:

[0112] (4),

[0113] In the formula, Z represents the learnable perturbation parameter; Z represents the latent space variable; This represents the mean vector of the multivariate Gaussian distribution in the latent space; Let represent the variance vector of the multivariate Gaussian distribution in the latent space. This represents the Hadamard product, ensuring differentiability and allowing the perturbation parameter model to participate in end-to-end gradient descent training. This represents a random noise vector sampled from a standard normal distribution.

[0114] The realistic waveform decoding module includes a feature fusion layer and a deconvolutional reconstruction network. It employs an adaptive instance normalization mechanism and uses an affine transformation network to convert the learnable perturbation parameter vector. Convert to scale factor With translation bias factor And inject it channel by channel into the pure time series feature vector. In this process, the deep fusion of physical disturbance characteristics and ideal waveform content is completed, generating enhanced simulation waveform data with real disturbance characteristics, as shown in the following expression:

[0115] (5),

[0116] In the formula, This represents the enhanced simulation waveform data after fusion; Represents the pure time series feature vector Mean value along the channel dimension; Represents the pure time series feature vector Standard deviation in the channel dimension.

[0117] Features after fusion Subsequently, decoding and temporal reconstruction are performed using one-dimensional deconvolution and nonlinear activation functions, ultimately outputting enhanced simulation waveform data superimposed with the characteristics of on-site engineering physical disturbances.

[0118] To prevent the reconstructed network from generating erroneous waveforms that violate common electrical knowledge during the pursuit of realism, this embodiment constructs a total loss function that integrates prior knowledge from the domain during the model optimization stage. It is composed of distribution alignment loss Electrical and physical mechanism constraint loss and adaptive anti-misoperation penalty loss Weighted composition, specifically:

[0119] (6),

[0120] In the formula, , as well as These represent the dynamic weighting systems for alignment loss, electrophysical mechanism constraint loss, and adaptive anti-malfunction penalty loss, respectively.

[0121] This is used to align the distribution characteristics of the generated enhanced simulation waveform data with those of the real fault recording data;

[0122] This is the core physical defense line of this invention. At the instant a high-resistance grounding fault occurs, extremely weak transient signals are easily drowned out or reversed by injected high-frequency noise. Therefore, Transient zero-sequence energy in enhanced simulation waveform data used for constraint extraction The polarity and theoretical energy of the ideal simulated waveform sequence The polarity remains consistent, as shown in the following expression:

[0123] (7),

[0124] In the formula, Indicates the initial moment of the fault; Indicates a short time window following the occurrence of the fault; This represents the zero-sequence voltage transient signal in the enhanced simulation waveform data at time t; This represents the zero-sequence current transient signal in the enhanced simulation waveform data at time t.

[0125] Once the noise generated by the network causes the energy direction to reverse, i.e. when the polarities are opposite, an exponential penalty is applied, as shown in the following expression:

[0126] (14)

[0127] In the formula, For symbolic functions, The symbol representing theoretical energy takes the value of 1 or -1; this constraint forces the model to still follow the transient energy conservation and directionality mechanism of single-phase grounding faults when injecting the distortion characteristics of the current transformer.

[0128] Furthermore, this embodiment innovatively designs a three-phase voltage drop or load fluctuation condition that frequently causes protection devices to malfunction in actual engineering projects. The algorithm dynamically generates an adaptive anti-maloperation threshold based on the unbalanced mean of the steady-state zero-sequence voltage before the fault. It is used to dynamically adjust the amplitude distortion parameters in the hidden space, and to prevent the model from generating false zero-sequence over-limit features that could cause the protection device to malfunction.

[0129] When the system input is a non-fault simulation waveform explicitly identified as voltage jitter, i.e., when the input ideal simulation waveform is labeled as a pure three-phase voltage drop or load change (non-ground fault), if the network outputs a peak value greater than [a certain value] in order to forcibly fit the noise, If a false zero-sequence transient voltage is detected, the ReLU penalty function is triggered to constrain the peak value of the transient zero-sequence voltage in the generated enhanced simulation waveform data. Less than the adaptive threshold The penalty function is shown in the following expression:

[0130] (8),

[0131] In the formula, , Represents the reliability coefficient; This represents the maximum background noise amplitude extracted from the power distribution network project; It represents the unbalanced mean of the steady-state zero-sequence voltage before the fault.

[0132] This embodiment introduces electrical physical mechanism constraints into the loss function and designs an adaptive threshold logic with anti-voltage drop malfunction capability (i.e., innovative). Based on the high-frequency noise, amplitude distortion, and transient effect characteristics extracted from field waveforms, this method can calculate and dynamically adjust the disturbance parameter model within the latent space mapping module in real time, thereby enabling the reconstruction network to accurately fit and adaptively learn the nonlinear disturbance distribution law of field waveforms during transient processes.

[0133] Step S4: Using the aforementioned power distribution network simulation model, based on various power distribution network fault scenarios and topology parameters, generate a new ideal simulation waveform sequence;

[0134] First, the traversal space of topology and physical fault parameters is set in the distribution network simulation model. The automated script will drive the simulation software to set the neutral grounding method of the distribution network, the fault initial phase angle distribution range, the transition resistance range covering high resistance grounding conditions, the line types including overhead lines, cable lines and mixed lines, the fault types such as arc grounding, metallic grounding or intermittent grounding, and the fault distances from the bus at different lengths. The automated script drives the distribution network simulation model to traverse the parameters and generate a set of new ideal simulation waveform sequences in batches.

[0135] Step S5: The newly added ideal simulation waveform sequence is input into the trained reconstruction network to perform perturbation injection and waveform reconstruction, generating enhanced simulation waveform data with real perturbation characteristics.

[0136] The newly added set of ideal simulation waveform sequences is input one by one into the trained waveform realignment reconstruction network. During the forward propagation process, the reconstruction network adaptively injects high-frequency noise, amplitude distortion, and load fluctuation disturbance characteristics of corresponding intensities through its latent space mapping module, based on the set severity level of the on-site environment, and outputs enhanced simulation waveform data that approximates the actual working conditions of the engineering site. Simultaneously, the system automatically generates multi-dimensional structured classification label vectors for each set of enhanced simulation waveform data sequences. The specific structure is defined as shown in the following expression:

[0137] (9),

[0138] In the formula, This indicates the label of the initial fault time determined through time-frequency domain analysis; This represents the fault initial phase angle label extracted from the simulation model; This indicates the fault grounding resistance label extracted from the simulation model; A label indicating the signal-to-noise ratio level of high-frequency noise injected into the network; A label indicating the degree of distortion characterizing the saturation or transmission error of the current transformer;

[0139] The time series matrix of the enhanced simulation waveform data is compared with the corresponding structured classification label vector. One-to-one binding and serialization storage are used to construct a high-fidelity enhanced dataset.

[0140] Step S6: After constructing the high-fidelity augmented dataset, use it as the standard training sample set for supervised training of the single-phase grounding fault identification model in the downstream distribution network. Specific steps include:

[0141] Step S6A1: Obtain the enhanced simulation waveform data sequence from the high-fidelity enhanced dataset and feed it as input features into the preset downstream artificial intelligence fault identification model;

[0142] Step S6A2: Synchronously extract and enhance the multi-dimensional structured classification label vector corresponding to the simulation waveform data. Use it as a supervision signal for model training;

[0143] Step S6A3: During the model training process, the real physical disturbance features such as high-frequency noise, amplitude distortion and load fluctuation contained in the enhanced simulation waveform data are used to drive the downstream artificial intelligence fault identification model to learn anti-interference features, thereby improving the generalization recognition accuracy and anti-distortion ability of the model under complex engineering conditions.

[0144] Step S7: In this embodiment, after the training of the single-phase ground fault identification model in the downstream distribution network is completed, the waveforms of the test set that were not used in the training in the above-mentioned high-fidelity enhanced dataset are imported into a real-time digital simulator or relay protection tester. Through its internal high-precision digital-to-analog conversion module, the discrete digital time series matrix is ​​restored into real analog voltage and analog current signals. Subsequently, through physical wiring, the analog signals are injected in real time into the AC sampling plug-in of the distribution network relay protection device under test. During the hardware injection test, the test system synchronously monitors the actual action behavior of the relay protection device and retrieves the structured classification label vector corresponding to the injected waveform. As a benchmark, the results are compared with the actual results of the relay protection device. If the device malfunctions when a pure load fluctuation or three-phase voltage drop waveform simulated by the network is injected, or fails to operate when a single-phase grounding waveform with extremely high transition resistance and strong noise characteristics is injected, the system will trigger the feature backtracking mechanism. Based on the data recorded in the tag vector... and The parameters are used to accurately pinpoint the typical disturbance characteristics that cause device failure. These typical characteristics are then used as feedback verification signals to inversely optimize the single-phase ground fault identification model.

[0145] Example 2

[0146] Based on Example 1, this embodiment proposes an enhancement system for simulating waveform data of grounding faults in distribution networks. The system includes:

[0147] The data acquisition module is configured to: acquire real fault waveform data from the power distribution network project site; build a power distribution network simulation model based on the power grid topology and operating parameters of the power distribution network project site, and generate an ideal simulation waveform sequence corresponding to the operating conditions of the real fault waveform data;

[0148] The feature extraction module is configured to: extract features from the real fault recording data and the ideal simulation waveform sequence to obtain difference features, and construct a real disturbance feature set based on the difference features;

[0149] The network training module is configured to: dynamically train the pre-constructed reconstructed network based on the real disturbance feature set; wherein the trained reconstructed network is able to learn the disturbance distribution pattern in real working conditions;

[0150] The waveform generation module is configured to generate a new ideal simulation waveform sequence based on various distribution network fault scenarios and topology parameters using the distribution network simulation model.

[0151] The waveform enhancement module is configured to: input the newly added ideal simulation waveform sequence into the trained reconstruction network, perform perturbation injection and waveform reconstruction, and generate enhanced simulation waveform data with real perturbation characteristics.

[0152] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0153] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0154] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0155] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0156] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for enhancing simulation waveform data of grounding faults in distribution networks, characterized in that, The method includes: Obtain real fault recording data from power distribution network engineering sites; Based on the power grid topology and operating parameters at the power distribution network site, a power distribution network simulation model is built, and an ideal simulation waveform sequence corresponding to the actual fault recording data is generated. Feature extraction is performed on the real fault recording data and the ideal simulation waveform sequence to obtain difference features, and a real disturbance feature set is constructed based on the difference features; Based on the real disturbance feature set, the pre-constructed reconstruction network is dynamically trained; wherein, the trained reconstruction network, under the constraints of electrical and physical mechanisms, is able to learn the disturbance distribution law in real working conditions; Based on the aforementioned power distribution network simulation model, new ideal simulation waveform sequences are generated using various power distribution network fault scenarios and topology parameters. The newly added ideal simulation waveform sequence is input into the trained reconstruction network to perform perturbation injection and waveform reconstruction, generating enhanced simulation waveform data with real perturbation characteristics.

2. The method for enhancing simulation waveform data of grounding faults in distribution networks according to claim 1, characterized in that, The method further includes continuously extracting a data buffer window containing a preset number of power frequency cycles before and after the fault transient process. The total number of discrete sampling points contained in the data buffer window is shown in the following formula: （1）， In the formula, K represents the total number of discrete sampling points; K represents the number of power frequency cycles. This indicates the actual sampling frequency of the waveform recording device or relay protection device at the power distribution network construction site; Indicates the rated power frequency of the power distribution network system; The time series set included in the data buffer window is shown in the following formula: （2）， In the formula, W represents the time series set; This represents the pre-fault steady-state data segment used to provide a reference for the system background before the fault. This represents a data segment that records the high-frequency characteristics and abrupt changes in the fault transient process at the instant of the fault. This represents the post-fault steady-state data segment that characterizes the trend of fault evolution.

3. The method for enhancing simulation waveform data of grounding faults in distribution networks according to claim 1, characterized in that, Before performing feature extraction on the real fault waveform data and the ideal simulation waveform sequence, the method further includes aligning the real fault waveform data and the ideal simulation waveform sequence along the time axis, specifically including: The time-frequency domain analysis algorithm is used to locate the initial fault moment of the actual fault recording data. Using the initial moment of the fault as the reference alignment point, the simulated waveform data at the corresponding moment in the ideal simulated waveform sequence is extracted, so that the real fault waveform data is synchronized with the ideal simulated waveform sequence.

4. The method for enhancing simulation waveform data of grounding faults in distribution networks according to claim 3, characterized in that, The step of using time-frequency domain analysis algorithms to locate the initial fault moment in the actual fault recording data specifically includes: Obtain the zero-sequence voltage signal sequence or zero-sequence current signal sequence from the real fault recording data, remove the DC component, and then perform multi-scale discrete wavelet transform on the zero-sequence voltage signal sequence or zero-sequence current signal sequence. Based on the Mallat algorithm, high-frequency detail coefficients reflecting the transient change characteristics of faults are extracted from the zero-sequence voltage signal sequence or zero-sequence current signal sequence after wavelet transform. Based on the high-frequency detail coefficients, a wavelet transform modulus sequence is constructed, the energy of the detail coefficients is calculated and smoothed, and the sampling time corresponding to the maximum value of the smoothed energy sequence is set as the fault initial time.

5. The method for enhancing simulation waveform data of grounding faults in distribution networks according to claim 1, characterized in that, The true perturbation feature set includes one or more of the following: High-frequency noise at the construction site, amplitude distortion caused by transformer transmission and line voltage drop, voltage and current jitter caused by load fluctuations, unbalanced voltage and current, and transient effects at the moment of fault.

6. The method for enhancing simulation waveform data of grounding faults in distribution networks according to claim 1, characterized in that, The pre-constructed reconstructed network includes an ideal waveform feature encoding module, a perturbation parameter latent space mapping module, and a realistic waveform decoding module, wherein: The waveform feature encoding module employs a one-dimensional deep convolutional neural network to extract time-series dependencies through multi-layer causal convolution. It receives the ideal simulated waveform sequence and extracts the pure transient and steady-state features of the sequence to obtain a pure time-series feature vector. The discrete feature extraction of the k-th convolutional layer of the waveform feature encoding feature module is shown in the following expression: （3）， In the formula, Indicates the first The output of the first layer of convolution The feature values ​​of each sampling point; Indicates the kernel size; Indicates the first The first layer of convolution kernel Each weighting coefficient; Indicates the first The bias of the convolutional kernel; The perturbation parameter latent space mapping module employs a latent variable generator based on a variational autoencoder architecture, which maps features to a continuous latent space using a multilayer perceptron. It receives the real perturbation feature set and maps the features in the real perturbation feature set to learnable perturbation parameters, as shown in the following expression: （4）， In the formula, Z represents the learnable perturbation parameters; Z represents the latent space variables. The vector representing the mean of the latent space distribution; The variance vector representing the latent space distribution; This represents a random noise vector sampled from a standard normal distribution; The realistic waveform decoding module includes a feature fusion layer and a deconvolutional reconstruction network, used to input the perturbation parameters into the clean time series feature vector to generate enhanced simulation waveform data with realistic perturbation features, as shown in the following expression: （5）， In the formula, This represents the enhanced simulation waveform data after fusion; Represents the pure time series feature vector Mean value along the channel dimension; Represents the pure time series feature vector Standard deviation in the channel dimension; Indicates the scaling factor; This represents the translation bias factor.

7. The method for enhancing simulation waveform data of grounding faults in distribution networks according to claim 1, characterized in that, Based on the real perturbation feature set, during the dynamic training of the pre-constructed reconstruction network, the total loss function used includes distribution alignment loss, electrophysical mechanism constraint loss, and adaptive anti-malfunction penalty loss, as shown in the following expression: （6）， In the formula, Represents the total loss function; This represents the distribution alignment loss; Indicates electrical and physical mechanism constraint loss; This indicates an adaptive penalty loss for preventing accidental activation; , as well as These represent the dynamic weighting coefficients for alignment loss, electrophysical mechanism constraint loss, and adaptive anti-maloperation penalty loss, respectively. The distribution alignment loss This is used to align the distribution characteristics of the generated enhanced simulation waveform data with those of the real fault recording data; The electrical physical mechanism constraint loss Transient zero-sequence energy in enhanced simulation waveform data used for constraint extraction The polarity and theoretical energy of the ideal simulated waveform sequence The polarity remains consistent, as shown in the following expression: （7）， In the formula, Indicates the initial moment of the fault; Indicates a short time window following the occurrence of a fault; This represents the zero-sequence voltage transient signal in the enhanced simulation waveform data at time t; This represents the zero-sequence current transient signal in the enhanced simulation waveform data at time t. The adaptive anti-misoperation penalty loss This is used to constrain the transient zero-sequence voltage peak value in the enhanced simulation waveform data generated when the input is a non-grounded fault. Less than the adaptive threshold The adaptive threshold is used to dynamically adjust the amplitude distortion parameter in the latent space, as shown in the following expression: （8）， In the formula, , Represents the reliability coefficient; This represents the maximum background noise amplitude extracted from the power distribution network project; This represents the unbalanced mean of the steady-state zero-sequence voltage before the fault. This is the activation function.

8. The method for enhancing simulation waveform data of grounding faults in distribution networks according to claim 1, characterized in that, While generating enhanced simulation waveform data with realistic perturbation characteristics, the method also includes: Generate multi-dimensional structured classification label vectors for the enhanced simulation waveform data. The specific structure is defined as shown in the following expression: （9）， In the formula, This indicates the label of the initial fault time determined through time-frequency domain analysis; Indicates the initial phase angle label of the fault; Label indicating fault grounding resistance; Labels indicating high-frequency noise signal-to-noise ratio levels; Labels indicating the degree of distortion; The time series matrix of the enhanced simulation waveform data is compared with the corresponding structured classification label vector. Corresponding binding and storage are used to construct a high-fidelity augmented dataset.

9. The method for enhancing simulation waveform data of grounding faults in distribution networks according to claim 8, characterized in that, The method also includes using the high-fidelity augmented dataset to perform supervised training on the single-phase grounding fault identification model of the downstream distribution network.

10. An enhancement system for simulation waveform data of grounding faults in distribution networks, characterized in that, include: The data acquisition module is configured to: acquire real fault waveform data from the power distribution network project site; build a power distribution network simulation model based on the power grid topology and operating parameters of the power distribution network project site, and generate an ideal simulation waveform sequence corresponding to the operating conditions of the real fault waveform data; The feature extraction module is configured to: extract features from the real fault recording data and the ideal simulation waveform sequence to obtain difference features, and construct a real disturbance feature set based on the difference features; The network training module is configured to: dynamically train the pre-constructed reconstructed network based on the real disturbance feature set; wherein the trained reconstructed network is able to learn the disturbance distribution pattern in real working conditions; The waveform generation module is configured to generate a new ideal simulation waveform sequence based on various distribution network fault scenarios and topology parameters through the distribution network simulation model. The waveform enhancement module is configured to: input the newly added ideal simulation waveform sequence into the trained reconstruction network, perform perturbation injection and waveform reconstruction, and generate enhanced simulation waveform data with real perturbation characteristics.

Images