Multi-mode power distribution network single-phase earth fault positioning method based on dynamic topology perception

Abstract

The invention discloses a multi-mode power distribution network single-phase earth fault positioning method based on dynamic topology perception, and belongs to the technical field of power distribution network fault diagnosis. The method comprises the following steps: firstly, analyzing physical structure change of a power distribution network in real time to construct a dynamic adjacency matrix, and extracting node features by using a space-time diagram convolutional network; and then multi-source data such as electrical quantity, environment quantity and equipment state are further combined to carry out fusion analysis, so that fault positioning is realized. According to the invention, the accuracy and reliability of single-phase earth fault identification of the power distribution network can be significantly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of distribution network fault diagnosis, and more specifically, to a method for locating single-phase grounding faults in a multimodal distribution network based on dynamic topology perception. Background Art

[0002] In a distribution network, single-phase grounding faults are common fault types, posing a significant threat to personal safety and the stable operation of equipment. Existing fault location methods, such as transient signal methods, steady-state signal methods, and wavelet transform methods, although they can play a certain role under specific conditions, each has its limitations:

[0003] Although the transient signal method utilizes the high-frequency characteristics at the moment of fault, it is vulnerable to noise interference and has a relatively weak theoretical basis; the steady-state signal method is based on power frequency components, but it has poor performance in resonant grounding systems; the wavelet transform method is suitable for short-time mutation signals, but has a high computational complexity.

[0004] In addition, these methods also face challenges such as weak transient signals, insufficient measurement accuracy, and difficulty in identifying high-resistance faults in practical applications.

[0005] In recent years, researchers have begun to explore more advanced algorithms and technologies to improve the accuracy of fault location. Artificial intelligence algorithms, such as neural networks, fuzzy logic, and machine learning, have been applied to fault pattern recognition to improve the adaptability and robustness of location. Among them, the active injection signal method enhances the identifiability of fault characteristics by injecting specific frequency signals into the power grid, especially suitable for high-impedance fault scenarios.

[0006] However, despite the certain progress made in the existing technology, it still cannot fully meet the requirements in a complex power grid environment. Summary of the Invention

[0007] In view of this, to at least partially solve the above technical problems, the present invention provides a method for locating single-phase grounding faults in a multimodal distribution network based on dynamic topology perception, aiming to improve the accuracy and robustness of fault location by comprehensively using various modal information and dynamic topology perception technology.

[0008] To achieve the above object, the present invention adopts the following technical solutions:

[0009] A method for locating single-phase grounding faults in a multimodal distribution network based on dynamic topology perception, the steps include,

[0010] Determine the real-time adjacency matrix according to the topology map of the distribution network;

[0011] Combine the adjacency matrix and the corresponding time-series signals, and use a spatio-temporal graph convolutional network to extract spatio-temporal features;

[0012] Determine the likelihood probability of electrical quantities based on spatiotemporal characteristics;

[0013] The likelihood probability of electrical quantity, environment and equipment status is integrated to determine the posterior probability of fault and locate the fault based on the posterior probability of fault.

[0014] Furthermore, the adjacency matrix is determined in real time based on the topology of the distribution network, including:

[0015] According to the physical connection relationship of the distribution network, a topological map is constructed and the initial static adjacency matrix is determined;

[0016] Combined with the circuit breaker status, the real-time adjacency matrix is determined as follows:

[0017] A t =aA static +(1-a)·diag(S t )

[0018] Where A t represents the real-time adjacency matrix, A static is the initial static adjacency matrix, S t is the real-time circuit breaker state vector (elements are 0 or 1, indicating that the device is open / closed), α is the topology dynamic weight coefficient (0.6 to 0.8, empirical value), and diag() represents the diagonal matrix.

[0019] Furthermore, the real-time adjacency matrix is symmetrically normalized, and the formula is expressed as:

[0020]

[0021] Where, is the normalized adjacency matrix, D t is the degree matrix.

[0022] Furthermore, the formula for extracting spatiotemporal features using spatiotemporal graph convolutional network is expressed as:

[0023]

[0024] Where H (l+1) represents the node feature tensor of the l+1th layer, σ represents the activation function, The kth time convolution kernel of the lth layer, K represents the number of time convolution kernels, ⊙ Hadamard product (element-wise multiplication), H (l) ∈R N×T×C Represents the node feature tensor of the lth layer, N represents the number of nodes, T represents the time step, and C represents the number of channels. is the spatial convolution kernel.

[0025] Furthermore, the formula for extracting spatiotemporal features using spatiotemporal graph convolutional network is expressed as:

[0026]

[0027] Wherein, H (l+1) represents the node feature tensor of the (l + 1)-th layer, σ represents the activation function, the k-th temporal convolutional kernel of the l-th layer, K represents the number of temporal convolutional kernels, A t represents the real-time adjacency matrix, ⊙ is the Hadamard product (element-wise multiplication), H (l) ∈R N×T×C represents the node feature tensor of the l-th layer, N represents the number of nodes, T represents the number of time steps, C represents the number of channels, is the spatial convolutional kernel.

[0028] Furthermore, the corresponding time series signals include zero-sequence current response and voltage disturbance response.

[0029] Furthermore, a preset frequency voltage disturbance is injected through the active grounding device to amplify the zero-sequence current response; wherein, the amplified zero-sequence current increment is:

[0030]

[0031] Wherein, V inj is the amplitude of the injected disturbance voltage (usually 5 - 10V); R f is the grounding fault resistance; is the zero-sequence impedance of the capacitance and inductance to the ground of line k.

[0032] Furthermore, the likelihood probability of electrical quantities is determined according to the matching degree between the spatio-temporal features and the fault labels.

[0033] Furthermore, the fault posterior probability is determined by the following formula:

[0034]

[0035] Wherein, P(F k ) is the prior fault probability of line k;

[0036] When the data of each modality is independent under the given fault condition,

[0037] P(D∣F k ) = P(D elec ∣F k )·P(D env ∣F k )·P(D dev ∣F k )

[0038] Wherein, P(D elec ∣F k ) represents the likelihood probability of electrical quantities, P(Denv |F k ) represents the likelihood probability of environmental quantity, P(D dev |F k ) represents the likelihood probability of device status, D = [D elec , D env , D dev represents multi-modal data; D elec represents electrical quantity, D env represents environmental quantity, D dev represents device status.

[0039] Furthermore, the spatio-temporal convolutional network performs transfer learning with the goal of minimizing the domain adaptation loss function, where the domain adaptation loss function is:

[0040]

[0041] In the formula, h s , h s are the feature representations of the source domain (existing fault data) and the target domain (new power grid) respectively, k(·, ·) is the Gaussian kernel function, and

[0042]

[0043] where x is the data in the source domain and y is the data in the target domain, represents the key hyperparameter.

[0044] The present invention discloses and provides a single-phase grounding fault location method for a multi-modal distribution network based on dynamic topology perception. Compared with the prior art, it has the following effects:

[0045] (1) By real-time analyzing the changes in the physical structure of the distribution network, constructing a dynamic adjacency matrix, and using a spatio-temporal graph convolutional network (ST-GCN) to extract node features, high-precision fault location can be achieved;

[0046] (2) By combining multi-source data such as electrical quantity, environmental quantity, and device status for fusion analysis, the accuracy and reliability of fault identification are further improved;

[0047] (3) By introducing an active perturbation enhancement mechanism and a transfer learning framework, the robustness of fault location in complex scenarios is significantly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained according to the provided drawings without creative efforts.

[0049] Figure 1 Flow chart of the single-phase grounding fault location method for multimodal distribution network based on dynamic topology perception provided by the present invention. Detailed implementation manners

[0050] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0051] Many specific details are set forth in the following description in order to fully understand the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar extensions without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0052] The embodiment of the present invention breaks through the limitations of traditional methods and proposes an innovative solution combining dynamic topology perception and multimodal spatio-temporal fusion to achieve higher-precision and stronger-robustness fault location capabilities. This method realizes high-precision fault location by real-time parsing the physical structure changes of the distribution network, constructing a dynamic adjacency matrix, and using a spatio-temporal graph convolutional network (ST-GCN) to extract node features. It also combines multi-source data such as electrical quantities, environmental quantities, and equipment states for fusion analysis, further improving the accuracy and reliability of fault identification. The overall solution has the characteristics of strong innovation and wide application range, and has important application value in the construction of smart grids.

[0053] In one embodiment, the method for single-phase grounding fault location of multimodal distribution network based on dynamic topology perception in this application, the process refers to Figure 1 , and the steps include:

[0054] Determine the real-time adjacency matrix according to the topology graph of the distribution network;

[0055] Combine the adjacency matrix and the corresponding time series signals, and use a spatio-temporal graph convolutional network to extract spatio-temporal features;

[0056] Determine the likelihood probability of electrical quantities according to the spatio-temporal features;

[0057] Fuse the likelihood probability of electrical quantities, the likelihood probability of the environment, and the likelihood probability of equipment states to determine the posterior probability of the fault, and realize fault location according to the posterior probability of the fault.

[0058] In this embodiment,

[0059] First step, determine the real-time adjacency matrix through dynamic topology perception;

[0060] According to the physical connection relationship of the distribution network, this application includes real-time acquisition of the breaker status and the location of distributed power access points, constructs a dynamic topology graph, and determines the real-time adjacency matrix according to the following formula to solve the misjudgment problem caused by network reconstruction (such as switching of distributed power sources) in the traditional method;

[0061] A t = aA static +(1 - a)·diag(S t )

[0062] In the formula, A t represents the real-time adjacency matrix, A static is the initial static adjacency matrix, S t is the real-time breaker status vector (the elements are 0 or 1, indicating that the device is disconnected / closed), α is the topology dynamic weight coefficient (0.6 - 0.8, empirical value), and diag() represents the diagonal matrix.

[0063] Among them, the static topology modeling constructs the initial adjacency matrix A static through the grid connection relationship, where the element a ij = 1, indicating that nodes i and j are directly connected;

[0064] The dynamic injection refers to real-time acquisition of the breaker status vector S t . If a certain breaker is disconnected, the adjacency weight of its corresponding line is reduced, and the dynamic weight is adjusted through the diagonal matrix diag(S t ).

[0065] As a preferred implementation, the real-time adjacency matrix is symmetrically normalized, and the formula is expressed as:

[0066]

[0067] In the formula, is the normalized adjacency matrix, D t is the degree matrix, D t (i,i)= ∑ j A t (i,j), A t (i,j) represents the element in the i-th row and j-th column of the adjacency matrix A t at time t. If there is a connection between node i and node j, this value is usually 1; otherwise, it is 0. ∑ j A t (i,j) represents the sum of all elements in the i-th row of the matrix, and the result reflects the degree of node i at time t (i.e., the number of edges connected to it). D t (i,i) represents the degree matrix D tThe diagonal elements, where each element stores the degree of the corresponding node. The off-diagonal elements are usually 0. Therefore, the formula D t (i,i) = ∑ j A t (i,j) actually assigns the degree of node i to the i-th diagonal element of the degree matrix D t .

[0068] Step 2: Combine the adjacency matrix and the corresponding time series signals, and use a spatio-temporal graph convolutional network to extract spatio-temporal features;

[0069] This application combines time series signals such as zero-sequence current waveforms and voltage disturbance responses with the spatial relationship of the power grid topology, and extracts spatio-temporal features through node embedding and neighborhood aggregation to solve the problem of correlation analysis in complex scenarios such as arc faults and multi-point grounding.

[0070] This embodiment uses a spatio-temporal graph convolutional network (ST-GCN) to extract spatio-temporal features. ST-GCN can simultaneously capture spatial topological correlations and time evolution laws, and can identify the spatio-temporal propagation characteristics of arc faults.

[0071] Its spatio-temporal convolution operation can be expressed as:

[0072]

[0073] In the formula, H (l+1) represents the node feature tensor of the (l + 1)-th layer, σ represents the activation function, the k-th temporal convolution kernel of the l-th layer, K represents the number of temporal convolution kernels, ⊙ is the Hadamard product (element-wise multiplication), H (l) ∈R N×T×C represents the node feature tensor of the l-th layer, N represents the number of nodes, T represents the number of time steps, C represents the number of channels, is the spatial convolution kernel.

[0074] Specifically, when extracting spatio-temporal convolution features,

[0075] In the temporal dimension convolution, one-dimensional convolution is performed on the time series features (such as zero-sequence current waveforms) of each node to extract temporal correlations, that is:

[0076]

[0077] In the spatial dimension, the dynamic adjacency matrix is used to aggregate adjacent node features:

[0078]

[0079] Then, the expression ability of the model is enhanced through the activation function σ (such as ReLU).

[0080] To further optimize the above technical solution, when a suspected fault occurs, a specific frequency voltage disturbance is injected through the active grounding device to amplify the difference in zero-sequence current between the faulty line and the normal line, and improve the recognizable feature of high-resistance faults.

[0081] In this embodiment, the disturbance parameter optimization is guided by the following formula to improve the detection accuracy of high-resistance faults (such as R f > 10 kΩ) to 98%. Among them, the zero-sequence current response after disturbance injection is:

[0082]

[0083] In the formula, V inj Amplitude of the injected disturbance voltage (usually 5 - 10 V); R f Ground fault resistance; Zero-sequence impedance of the capacitance and inductance of line k to the ground.

[0084] In one embodiment, during execution, first, the disturbance signal is designed: a sinusoidal voltage with a specific frequency f inj (such as 150 Hz) is injected during the suspected fault period: v inj (t) = V inj sin(2πf inj t)

[0085] Then, the zero-sequence circuit analysis is carried out: according to the zero-sequence equivalent circuit of the distribution network, the difference in zero-sequence current between the faulty line and the non-faulty line is deduced. For the faulty line k, the zero-sequence current increment is:

[0086]

[0087] In this embodiment, the of the faulty line is significantly greater than that of the non-faulty line, and a differential feature for detection can be formed.

[0088] Thirdly, determine the likelihood probability of the electrical quantity according to the spatio-temporal characteristics;

[0089] Determine the likelihood probability of the electrical quantity according to the matching degree between the spatio-temporal characteristics and the fault label. In a preferred solution, it is calculated by the matching degree between the output features of ST-GCN and the fault label.

[0090] In this embodiment, during the process of determining the likelihood probability of the electrical quantity based on the spatio-temporal characteristics, a fault label database should be constructed in advance. The database provides a basis for matching and calculation, and can significantly improve the accuracy and efficiency of diagnosis. Among them, the database contains various fault types and their corresponding feature labels, providing a clear standard for model training and pattern recognition in actual applications. By comparing with the pre-stored fault labels, the most similar fault type can be quickly located, so as to efficiently determine the likelihood probability of the electrical quantity.

[0091] In the machine learning ST-GCN, the database serves as the training set, which can help the model learn the spatio-temporal characteristics of different faults and improve its generalization ability. By comprehensively collecting historical fault data and performing accurate annotation, various common and potential faults can be ensured to be covered.

[0092] Step 4: Integrate the likelihood probabilities of electrical quantities, environmental likelihood probabilities, and equipment status likelihood probabilities to determine the posterior fault probability, and achieve fault location based on the posterior fault probability.

[0093] In one embodiment, a Bayesian probability model of multi-modal data fusion is used to fuse multi-source information such as electrical quantities (zero-sequence current, voltage transient waveform, sudden change in active power), environmental quantities (temperature and humidity, thunderstorm warning for correcting the fault probability model), and equipment status (circuit breaker operation record, cable aging degree evaluation value), so as to achieve multi-dimensional joint decision-making and obtain a more complete fault criterion.

[0094] In this embodiment, the formula for calculating the posterior fault probability is:

[0095]

[0096] In the formula, P(F k ) is the prior fault probability of line k (based on historical statistical data), P(D / F k ), P(D / F i ), P(D / F k ), P(D / F i ) are conditional probability representations based on Bayesian statistics, and P(D / F k ) represents the probability of observing data D when it is known that line k has a fault (F k ). P(D / F i ) represents the probability of observing data D when it is known that line i has a fault (F i ). If P(D∣F k ) > P(D∣F i ), it indicates that under the current observation, the possibility of line k having a fault is higher. In the formula for calculating the posterior fault probability, it is preset that line k has a fault, and line i is all lines from 1 to N.

[0097] Conditional independence assumption: Assuming that each modal data is independent given the fault condition,

[0098] P(D∣F k ) = P(D elec ∣F k )·P(D env ∣F k )·P(D dev ∣F k )

[0099] Among them, P(D elec ∣F k ) represents the likelihood probability of electrical quantities, P(D env ∣F k ) represents the likelihood probability of environmental quantities, P(D dev ∣F k ) represents the likelihood probability of equipment status, D = [D elec , D env , D dev represents multi-modal data; D elec represents electrical quantities, D env represents environmental quantities, D dev represents equipment status.

[0100] In this embodiment, the likelihood probabilities of the environment and equipment status are empirical values and are key parameters for probability adjustment and evaluation in fault diagnosis; they are used to improve the accuracy of diagnosis results by reflecting the influence of environmental conditions and equipment conditions on the occurrence of faults.

[0101] P(D env ∣F k ) is the environmental quantity correction, which represents the degree of influence of the observed data D on the fault type Fk under specific environmental conditions. According to the influence of different environments (such as temperature, humidity, etc.) on equipment performance, the possibility of fault occurrence is adjusted to make the diagnosis closer to the actual operation situation. For example, the cable fault probability increases by 30% during thunderstorm weather.

[0102] P(D dev ∣F k ) is the equipment status weight, which is based on the current status of the equipment (such as aging degree, maintenance records, etc.) to calculate its influence on the occurrence probability of the fault type Fk. By quantifying the role of equipment status in the fault probability, it provides a basis for predictive maintenance and reduces unexpected shutdowns. The aging weight coefficient is an important part of the equipment status weight, which directly reflects the influence of the degradation of the equipment over time on its reliability. A coefficient determined according to the service life, operating status and historical data of the equipment is used to adjust the change in fault probability caused by aging. Incorporating this coefficient into the Bayesian formula can more accurately evaluate the health status of the equipment at different life cycle stages.

[0103] In this application, the aging equipment weight coefficient: γ = 1 + 0.2 × aging level.

[0104] In a preferred embodiment, the ST-GCN model is deeply analyzed in the cloud. By combining the transfer learning framework and using the historical fault database to optimize the model parameters, the rapid adaptation of the cross-regional power grid is achieved. The transfer learning steps include:

[0105] Feature alignment objective: Minimize the difference in feature distributions between the source domain and the target domain, enabling the ST-GCN model to be transferable across different power grids;

[0106] MMD calculation: Map to the Reproducing Kernel Hilbert Space (RKHS) through a kernel function to calculate the distribution distance between the two domains.

[0107] In this embodiment, the MMD loss is calculated using a Gaussian kernel function; the calculation formula is:

[0108]

[0109] In the formula, h s , h s are the feature representations of the source domain (existing fault data) and the target domain (new power grid) respectively, k(·,·) is the Gaussian kernel function, and

[0110]

[0111] where x is the data in the source domain, y is the data in the target domain, and ζ is a key hyperparameter that controls the similarity calculation between data points and the smoothness of the decision boundary. Reasonable setting can optimize the model performance and avoid overfitting or underfitting problems.

[0112] Furthermore, as a preference, the following total loss function is used for joint optimization:

[0113] L total = L CE + λL MMD

[0114] In the formula, L CE is the cross-entropy classification loss, and λ is the trade-off coefficient.

[0115] The edge-cloud collaborative computing architecture provided by this application can complete preliminary feature extraction and real-time alarm by deploying a lightweight CNN model at the substation terminal; at the same time, the ST-GCN model can achieve fast adaptation of cross-regional power grids through in-depth analysis in the cloud, combined with the transfer learning framework, and using the historical fault library to optimize the model parameters. Based on the MMD loss, the new power grid can reach an accuracy of 90%+ with only a small number of samples, reducing the communication load by 60% compared with traditional methods and saving computing power.

[0116] In this specification, each embodiment is described in a progressive manner. The key point of each embodiment is to illustrate the differences from other embodiments. The same or similar parts among the embodiments can be referred to each other. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method part.

[0117] The foregoing description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A single-phase grounding fault location method for multimodal distribution networks based on dynamic topology perception, characterized in that Determine the real-time adjacency matrix according to the topology diagram of the distribution network; Combine the adjacency matrix and the corresponding time-series signals, and use a spatio-temporal graph convolutional network to extract spatio-temporal features; Determine the likelihood probability of electrical quantities according to the spatio-temporal features; Fuse the likelihood probability of electrical quantities, the likelihood probability of the environment, and the likelihood probability of equipment status to determine the fault posterior probability, and realize fault location according to the fault posterior probability.

2. The grounding fault location method according to claim 1, characterized in that, Determine the adjacency matrix in real time according to the topology diagram of the distribution network, including: Construct a topology diagram and determine the initial static adjacency matrix according to the physical connection relationship of the distribution network; Combine the breaker status and determine the real-time adjacency matrix according to the following formula: A t = aA static +(1 - a)·diag(S t ) Wherein, A t represents the real-time adjacency matrix, and A static is the initial static adjacency matrix, S t is the real-time breaker status vector, α is the topological dynamic weight coefficient, and diag() represents the diagonal matrix.

3. The grounding fault location method according to claim 2, characterized in that, Perform symmetric normalization on the real-time adjacency matrix, and the formula is expressed as: In the formula, is the normalized adjacency matrix, and D t is the degree matrix.

4. The grounding fault location method according to claim 3, wherein, The formula for using a spatio-temporal graph convolutional network to extract spatio-temporal features is expressed as: where, H (l+1) represents the node feature tensor of the (l + 1)-th layer, σ represents the activation function, the k-th temporal convolutional kernel of the l-th layer, K represents the number of temporal convolutional kernels, ⊙ is the Hadamard product, H (l) ∈ R N×T×C represents the node feature tensor of the l-th layer, N represents the number of nodes, T represents the number of time steps, C represents the number of channels, is the spatial convolutional kernel.

5. The grounding fault location method according to claim 1, characterized in that The formula for using a spatio-temporal graph convolutional network to extract spatio-temporal features is expressed as: Where, H (l+1) represents the node feature tensor of the (l + 1)-th layer, σ represents the activation function, is the k-th temporal convolutional kernel of the l-th layer, K represents the number of temporal convolutional kernels, A t represents the real-time adjacency matrix, ⊙ is the Hadamard product, H (l) ∈ R N×T×C represents the node feature tensor of the l-th layer, N represents the number of nodes, T represents the time step, C represents the number of channels, is the spatial convolutional kernel.

6. The grounding fault location method according to claim 1, wherein The corresponding time-series signals include zero-sequence current response and voltage disturbance response.

7. The grounding fault location method according to claim 6, wherein Inject a preset frequency voltage disturbance through an active grounding device to amplify the zero-sequence current response; among them, the amplified zero-sequence current increment is: Where, V inj Amplitude of the injected disturbance voltage; R f Ground fault resistance; Zero-sequence impedance of the capacitance to ground and inductance of line k.

8. The grounding fault location method according to claim 1, characterized in that Determine the likelihood probability of electrical quantities according to the matching degree between the spatio-temporal features and the fault labels.

9. The grounding fault location method according to claim 1, characterized in that, Determine the fault posterior probability through the following formula: where P(F k ) is the prior fault probability of line k; When each modal data is independent under the given fault conditions, P(D|F k ) = P(D elec |F k ) · P(D env |F k ) · P(D dev |F k ) Among them, P(D elec ∣F k ) represents the likelihood probability of electrical quantities, P(D env ∣F k ) represents the likelihood probability of environmental quantities, P(D dev ∣F k ) represents the likelihood probability of equipment status, D = [D elec , D env , D dev represents multi-modal data; D elec represents electrical quantities, D env represents environmental quantities, D dev represents equipment status.

10. The grounding fault location method according to claim 1, wherein The spatio-temporal convolutional network performs transfer learning with the goal of minimizing the domain adaptation loss function, where the domain adaptation loss function is: where h s , h s are the feature representations of the source domain (existing fault data) and the target domain (new power grid) respectively, k(·,·) is the Gaussian kernel function, and where x is the data in the source domain and y is the data in the target domain, represents a key hyperparameter.

Images