Power grid fault rapid diagnosis system fused with AI algorithm

Abstract

The invention provides a power grid fault rapid diagnosis system fused with an AI algorithm, and relates to the technical field of power grid fault diagnosis. Three-phase voltage and current data of a power distribution network and distributed photovoltaic real-time output data are collected, and data noise is suppressed by using an adaptive morphological filter; extracting fault feature data under photovoltaic disturbance; further constructing a fault feature encoder through a comparative learning algorithm, and extracting fault feature representation which is not influenced by photovoltaic fluctuation; and a fault propagation relationship among nodes in the topological structure of the power distribution network is analyzed by combining a causal reasoning enhanced graph neural network, and a fault tracing path is generated. Thus, the accuracy of fault feature extraction in the photovoltaic disturbance environment can be effectively improved, the accuracy and efficiency of fault tracing and positioning are improved, rapid and accurate power grid fault diagnosis can be achieved, and therefore the safety and reliability of power distribution network operation are guaranteed.

Description

Technical Field

[0001] This invention relates to the field of power grid fault diagnosis technology, and more specifically, to a rapid power grid fault diagnosis system that integrates AI algorithms. Background Technology

[0002] With the transformation of the global energy structure and the rapid growth of electricity demand, the operating environment of power distribution networks is becoming increasingly complex. In particular, the large-scale grid connection of new energy sources such as distributed photovoltaics has greatly improved the flexibility and efficiency of power distribution network operation. However, the volatility and intermittency brought about by the integration of new energy sources have increased the uncertainty of the power distribution network's operating status, posing new challenges to fault diagnosis and location. Traditional power grid fault diagnosis technologies mainly rely on rule-based expert systems or simple signal processing methods. These technologies perform relatively stably in power grid environments with low new energy penetration. However, when faced with large-scale new energy integration scenarios, traditional methods are gradually revealing limitations in accuracy and efficiency.

[0003] Existing power grid fault diagnosis technologies mainly include model-based methods, data-driven methods, and hybrid methods combining both. Model-based methods rely on precise modeling of the distribution network's physical characteristics and topology; however, due to photovoltaic fluctuations and noise interference, these methods often struggle to accurately capture fault characteristics in complex environments. Data-driven methods, on the other hand, utilize massive amounts of historical data and machine learning algorithms to mine fault patterns, exhibiting strong adaptability and robustness; however, their diagnostic results lack interpretability and consistency to some extent. Furthermore, existing methods often fail to effectively extract key information accurately reflecting fault characteristics when facing high-frequency disturbances and nonlinear characteristics introduced by new energy sources such as photovoltaics, leading to a significant decrease in the accuracy of fault location and classification. Simultaneously, existing technologies have a relatively superficial understanding of power grid fault propagation patterns, often neglecting causal relationships between nodes, thus limiting the accuracy and efficiency of fault tracing. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a rapid fault diagnosis system for power grids that integrates AI algorithms. This system can, to some extent, solve the problem of false alarms caused by the similarity between partial discharge and lightning interference signal waveforms in GIS equipment within substations.

[0005] According to one aspect of the present invention, a rapid fault diagnosis system for power grids integrating AI algorithms is provided, comprising:

[0006] The fault feature acquisition and processing module is used to acquire three-phase voltage and current data and real-time output data of distributed photovoltaic power at each monitoring point of the distribution network, and to use an adaptive morphological filter to suppress noise in the three-phase voltage and current data to obtain fault feature data under photovoltaic disturbance.

[0007] The fault feature encoding module is used to map the fault feature data and the photovoltaic output data to a high-dimensional feature space and extract fault feature representations that are not affected by photovoltaic fluctuations.

[0008] The graph neural network module is used to analyze the fault propagation relationship between nodes in the distribution network topology based on the fault feature representation.

[0009] The causal reasoning module is used to identify fault sources based on the direction of energy transmission and generate fault tracing paths.

[0010] The fault diagnosis module is used to update the diagnostic rule base based on the fault tracing path and the fault feature representation, and output the fault equipment location result and fault type.

[0011] Furthermore, the adaptive morphological filter performs adaptive segmentation processing on the acquired three-phase voltage and current data, dynamically selects the size of the structural element based on the local variance and peak factor, and utilizes cascaded morphological opening and closing operations to effectively suppress measurement noise and photovoltaic fluctuations.

[0012] Furthermore, the dynamic selection of structuring element size includes:

[0013] When the local variance is greater than the preset threshold A and the peak factor is greater than the preset threshold B, a small-sized one-dimensional structuring element with a length of 3 sampling points is selected.

[0014] When the local variance is less than the preset threshold A or the peak factor is less than the preset threshold B, a large-size one-dimensional structuring element with a length of 7 sampling points is selected.

[0015] Furthermore, the fault feature encoding module uses a four-layer fully connected neural network as the basic encoder and uses the InfoNCE loss function to make the feature representations of the same fault under different photovoltaic outputs closer to each other, while the feature representations of different faults are further apart, thereby extracting the intrinsic fault features that are not affected by photovoltaic fluctuations.

[0016] Furthermore, the InfoNCE loss function calculates sample similarity as shown in the following formula:

[0017]

[0018] in, For sample similarity, This refers to the number of samples in the batch. The feature vector of the i-th anchor point sample after L2 normalization; Let be the positive sample feature vector of the i-th sample under the k-th photovoltaic fluctuation. Let j be the feature vector of the j-th negative sample; is the temperature parameter; · indicates the vector dot product operation.

[0019] Furthermore, the causal reasoning module, based on the node features of the graph neural network, uses a counterfactual reasoning framework to conduct intervention experiments on the nodes, and analyzes the propagation path and scope of the fault by calculating the causal effect score and constructing a causal relationship matrix.

[0020] Furthermore, the calculation of the causal effect score includes both direct causal effects and indirect causal effects;

[0021] The calculation of both the direct and indirect causal effects uses Euclidean distance to quantify the degree of feature change, but the indirect causal effect analyzes the feature changes of all intermediate nodes and uses a time decay term to reduce the propagation effect.

[0022] According to another aspect of the present invention, a method for rapid diagnosis of power grid faults integrating AI algorithms is provided, comprising:

[0023] Collect three-phase voltage and current data and real-time output data of distributed photovoltaic power at each monitoring point of the distribution network, and use an adaptive morphological filter to suppress noise in the three-phase voltage and current data to obtain fault characteristic data under photovoltaic disturbance.

[0024] A fault feature encoder was constructed using a contrastive learning algorithm to extract fault feature representations that are unaffected by photovoltaic fluctuations;

[0025] Based on the fault characteristics, a graph neural network enhanced with causal reasoning is used to analyze the fault propagation relationship between nodes in the distribution network topology and generate fault source tracing paths.

[0026] Based on the fault tracing path and combined with the fault feature representation, the diagnostic rule base is dynamically updated, and the fault equipment location result and fault type are output.

[0027] Compared with existing technologies, the power grid fault rapid diagnosis system integrating AI algorithms provided by this invention collects three-phase voltage and current data of the distribution network and real-time output data of distributed photovoltaic power generation. It uses an adaptive morphological filter to suppress data noise, thereby extracting fault feature data under photovoltaic disturbances. Furthermore, it constructs a fault feature encoder through a contrastive learning algorithm to extract fault feature representations unaffected by photovoltaic fluctuations. Finally, it combines causal reasoning-enhanced graph neural network analysis to analyze the fault propagation relationships between nodes in the distribution network topology, generating fault tracing paths. This effectively improves the accuracy of fault feature extraction under photovoltaic disturbance conditions, while also enhancing the precision and efficiency of fault tracing and location, contributing to rapid and accurate power grid fault diagnosis and ensuring the safety and reliability of distribution network operation. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0029] Figure 1 This is a system block diagram of a power grid fault rapid diagnosis system that integrates AI algorithms according to an embodiment of the present invention.

[0030] Figure 2 This is a flowchart of a method for rapid diagnosis of power grid faults that integrates AI algorithms according to an embodiment of the present invention. Detailed Implementation

[0031] Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. It is obvious that the described embodiments are merely some embodiments of the present invention, and not all embodiments of the present invention, and it should be understood that the present invention is not limited to the exemplary embodiments described herein.

[0032] Figure 1 This is a system block diagram of a power grid fault rapid diagnosis system integrating AI algorithms according to an embodiment of the present invention. Figure 1 As shown, the power grid fault rapid diagnosis system integrating AI algorithms includes:

[0033] The fault feature acquisition and processing module is used to acquire three-phase voltage and current data and real-time output data of distributed photovoltaic power at each monitoring point of the distribution network, and to use an adaptive morphological filter to suppress noise in the three-phase voltage and current data to obtain fault feature data under photovoltaic disturbance.

[0034] The fault feature encoding module is used to map the fault feature data and the photovoltaic output data to a high-dimensional feature space and extract fault feature representations that are not affected by photovoltaic fluctuations.

[0035] The graph neural network module is used to analyze the fault propagation relationship between nodes in the distribution network topology based on the fault feature representation.

[0036] The causal reasoning module is used to identify fault sources based on the direction of energy transmission and generate fault tracing paths.

[0037] The fault diagnosis module is used to update the diagnostic rule base based on the fault tracing path and the fault feature representation, and output the fault equipment location result and fault type.

[0038] The fault feature acquisition and processing module collects three-phase voltage data and three-phase current data through intelligent measurement devices at various monitoring points in the distribution network. The three-phase voltage data includes phase A voltage, phase B voltage, and phase C voltage, and the three-phase current data includes phase A current, phase B current, and phase C current. At the same time, it collects real-time photovoltaic output data from the power electronic equipment at the distributed photovoltaic grid-connected point. The real-time photovoltaic output data includes photovoltaic active power and reactive power output.

[0039] Secondly, an adaptive morphological filter is used to suppress noise for the three-phase voltage and current data. Specifically, the collected three-phase voltage and current data are first segmented at a sampling interval of 500 μs, with each segment being 2 ms long. For each segment, the local variance and peak factor are calculated. When the local variance is greater than a preset threshold A and the peak factor is greater than a preset threshold B, it is identified as a region of severe signal fluctuation, and a small one-dimensional structure element with a length of 3 sampling points is selected. When the local variance is less than the preset threshold A or the peak factor is less than the preset threshold B, it is identified as a region of stable signal, and a large one-dimensional structure element with a length of 7 sampling points is selected. Then, a cascaded opening and closing operation is performed on the selected structure element: first, the opening operation is performed to remove positive impulse noise, and then the closing operation is performed to remove negative impulse noise. After processing a data segment, the structure element is linearly interpolated at the boundary of adjacent data segments to smooth the transition and avoid abrupt changes in the filtering result between segments. After performing the above processing on the voltage and current of each phase, the filtered data are recombined to obtain fault characteristic data free from measurement noise and photovoltaic fluctuation interference. This filtering method, through adaptive adjustment of the structural element size, effectively suppresses different types of noise while maintaining the transient characteristics of the fault.

[0040] It should be noted that the process for determining thresholds A and B is as follows:

[0041] First, 1000 sets of sample data, including both fault and normal operating conditions, were selected from historical data. Each set of sample data contained complete three-phase voltage and current data. For each set of sample data, the local variance and peak factor within a sliding window were calculated, with the sliding window length set to 2 ms. When a fault occurred, data was extracted for 50 ms before and after the fault. The mean local variance and mean peak factor during the fault transient were statistically analyzed, denoted as Va and Fa. Simultaneously, the mean local variance and mean peak factor under normal operating conditions were statistically analyzed, denoted as Vn and Fn. Based on the statistical results, threshold A was set to (Va+Vn) / 2, and threshold B was set to (Fa+Fn) / 2. The thresholds determined through this data-driven approach effectively distinguish between fault transients and normal fluctuations, ensuring the adaptive morphological filter maintains good filtering performance under different operating conditions. To ensure the adaptability of the thresholds, they are updated every 3 months to adapt to dynamic changes in system operating characteristics.

[0042] Then, the noise-suppressed three-phase voltage and current data are used to construct a time-series data matrix with a sampling interval of 500 μs. For each monitoring point, data segments of 50 ms before and after the fault occurrence are extracted to obtain fault characteristic data. The fault characteristic data includes transient and steady-state components at the moment of fault occurrence, as well as voltage and current fluctuation characteristics caused by photovoltaic disturbance.

[0043] The fault feature encoding module constructs a four-layer fully connected neural network as the basic encoder. The input layer dimension is consistent with the fault feature data dimension, the number of nodes in the three hidden layers are 512, 256, and 128 respectively, and the output layer dimension is 64. The activation function is ReLU. When a fault sample is obtained, four positive samples are constructed by superimposing different levels of photovoltaic power output fluctuations (fluctuation features at 25%, 50%, 75%, and 100% of rated power) onto the sample. The original sample and these four positive samples then form a positive sample pair. Samples randomly selected from other fault types form negative samples. Once a batch of positive and negative samples is ready, they are simultaneously input into two identical basic encoders (parameter sharing) to obtain their representation vectors in the feature space. Then, the similarity between positive sample pairs and between negative samples is calculated using the InfoNCE loss function, where the similarity is measured using cosine similarity and the temperature parameter is set to 0.07. If the similarity of a positive sample pair is greater than that of a negative sample pair, the loss value is reduced; if the similarity of a positive sample pair is less than that of a negative sample pair, the loss value is increased. The network parameters are continuously adjusted using the Adam optimizer, with an initial learning rate set to 0.001, and dynamically adjusted during training based on a cosine annealing strategy. When training reaches a preset number of epochs or the contrastive loss on the validation set no longer decreases for five consecutive epochs, the best-performing base encoder is selected as the final fault feature encoder for subsequent feature extraction tasks.

[0044] When calculating sample similarity using the InfoNCE loss function, the 64-dimensional feature vector output by the encoder is first L2 normalized to ensure that the magnitude of all feature vectors is 1. When calculating the similarity of positive sample pairs, the feature vector of the original fault sample is taken as the anchor vector. This anchor vector is then multiplied by the feature vectors of four positive samples obtained under different photovoltaic fluctuations for the same fault, and divided by the temperature parameter of 0.07 to obtain four positive sample similarity scores. When calculating negative sample similarity, the anchor vector is multiplied by the feature vectors of other fault types in the batch, and again divided by the temperature parameter of 0.07 to obtain several negative sample similarity scores. Then, a softmax operation is performed on all similarity scores. The logarithm of the sum of positive sample similarity scores is used as the numerator, and the logarithm of the sum of all sample similarity scores is used as the denominator. The two are divided and the result is negative to obtain the InfoNCE loss value of the current anchor sample. If the current batch contains N samples, the loss values ​​of the N anchor samples need to be calculated, and the average of these N loss values ​​is used to obtain the final batch loss. During model training, minimizing this batch loss ensures that the feature representations of the same fault under different photovoltaic fluctuations are close to each other, while the feature representations of different faults are far apart, thus achieving effective extraction of the intrinsic features of the fault. More specifically, the infoNCE loss function is used to calculate sample similarity. It can be expressed by the following formula:

[0045]

[0046] in, This refers to the number of samples in the batch. The feature vector of the i-th anchor point sample after L2 normalization; The positive sample feature vector of the i-th sample under the k-th photovoltaic fluctuation (k=1,2,3,4 corresponds to the feature vector when the photovoltaic output is 25%, 50%, 75%, 100%). Let j be the feature vector of the j-th negative sample; is the temperature parameter; · indicates the vector dot product operation.

[0047] When tuning network parameters using the Adam optimizer, the initial learning rate is set to 0.001, the exponential decay rate β1 of the first-order moment estimate is 0.9, the exponential decay rate β2 of the second-order moment estimate is 0.999, and the numerical stability constant ε is 1e-8. After each training batch completes forward propagation, the InfoNCE loss value of the current batch is calculated, and the gradient corresponding to each network parameter is calculated using the backpropagation algorithm. For each parameter in the network, the Adam optimizer maintains its first-order moment estimate mt and second-order moment estimate vt, where mt is calculated by the weighted average of the first-order moment estimate from the previous time step and the current gradient, and vt is calculated by the weighted average of the second-order moment estimate from the previous time step and the square of the current gradient. After obtaining the bias-corrected first-order moment estimates and second-order moment estimates, the learning rate is multiplied by the first-order moment estimate and divided by the sum of the square root of the second-order moment estimate and ε to obtain the update step size for each parameter. If the current training epoch exceeds half of the total epochs, the learning rate is adjusted according to the cosine annealing strategy to the initial learning rate multiplied by the cosine function value, where the period of the cosine function is the remaining training epochs. This adaptive parameter update method allows the network to converge quickly in the early stages of training and be finely adjusted in the later stages, ultimately resulting in optimized network parameters. More specifically, the learning rate and parameter values ​​can be calculated using the following formula:

[0048]

[0049]

[0050] in, This represents the current iteration number; Total number of training rounds; This is a first-order moment estimate; It is a second-order moment estimate; This is the current learning rate; This is the current parameter value.

[0051] It is worth noting that although the above description outlines the general process for constructing a fault feature encoder, specific implementation details may vary depending on the type of distribution network fault and photovoltaic fluctuations. For example, let's take a common A-phase ground fault in a distribution network as a specific example to illustrate the encoder's operation: Assume an A-phase ground fault occurs on a distribution line connected to a photovoltaic power station. The fault waveform collected when the photovoltaic output is at a moderate level serves as the anchor sample. Its characteristics include a significant decrease in A-phase voltage approaching zero, while the voltages of phases B and C rise above their rated values. The A-phase current increases significantly, while the currents of phases B and C remain relatively stable. The encoder first converts this raw data into a feature vector and performs L2 normalization. Simultaneously, it collects fault data at other photovoltaic output levels for this fault as positive samples, and other types of fault data occurring within the same time period, such as B-phase ground faults, C-phase ground faults, and phase-to-phase short circuits, as negative samples. When calculating similarity, if the original A-phase ground fault feature vector and its positive sample feature vector under higher photovoltaic output yield a high similarity score, while the similarity score with the B-phase ground fault feature vector is low, this indicates that the encoder has effectively extracted the intrinsic features of the fault. By continuously optimizing the network parameters, the feature representation similarity of the A-phase ground fault under different photovoltaic outputs is maintained at a high level, while the feature representation similarity with other types of faults remains at a low level, thus achieving effective identification of fault features.

[0052] The graph neural network module models the distribution network topology as an undirected graph G=(V,E), where V represents the set of all nodes in the distribution network, and E represents the set of connections between nodes. After obtaining the fault feature representation, it is used as the input node features to the graph neural network. Feature propagation is performed through multiple layers of graph convolution, where each layer collects feature information from neighboring nodes and updates the state of the central node. If the fault features of a node are significantly correlated with the features of its neighboring nodes, a fault propagation relationship is considered to exist between these nodes. After completing graph convolution feature propagation, a causal reasoning module is introduced to analyze the causal relationships between nodes. By calculating the causal effect strength between node pairs, the direction of fault propagation is identified. If a state change in node A leads to a significant change in the state of node B, while the state change in node B has a relatively small impact on node A, it is inferred that the fault is likely propagating from node A to node B. After obtaining the causal relationships between all node pairs, the maximum spanning tree algorithm is used, with the causal effect strength as the edge weight, to generate a tracing path starting from the fault source node and passing through the main affected nodes. If, during path generation, a node is found to have a strong causal relationship with multiple nodes, the connection with the strongest causal effect is retained, and the other connections are considered as alternative paths. This graph structure analysis method based on causal reasoning not only identifies the propagation path of a fault but also reflects the attenuation characteristics of the fault during propagation, providing a reliable basis for fault tracing and localization.

[0053] The causal inference module constructs a structural causal model based on node features extracted from a graph neural network, using the feature vector of each node as an observed variable. When analyzing the causal relationship between node pairs, a counterfactual inference framework is employed. This involves intervening in the features of the source node while keeping the states of other nodes constant, and observing the degree of change in the features of the target node. If the features of the target node B change significantly after intervening in the source node A, it indicates that node A has a strong causal effect on node B. After analyzing the causal relationship between a pair of nodes, its causal effect score is calculated. This score consists of direct and indirect causal effects. The direct causal effect is obtained by calculating the Euclidean distance between the features of the target node before and after the intervention, while the indirect causal effect is estimated by analyzing the impact of the intervention on other intermediate nodes. If a node is found to have a strong causal effect on multiple nodes, further analysis of the temporal characteristics of these effects is needed. By comparing the propagation delay of the causal effects, the actual propagation order of the fault is determined. After completing the causal analysis of all node pairs, a causal relationship matrix is ​​constructed, where the matrix elements represent the strength of the causal effect between corresponding node pairs. This matrix visually displays the propagation path and scope of the fault in the network.

[0054] The causal effect score can be calculated using the following formula:

[0055] in, The score represents the causal effect of node A on node B. This is the feature vector of node B after intervention in node A; The feature vector of node B before intervention; The direct causal effect weighting coefficient has a value range of [0,1]. α is the indirect causal effect weighting coefficient, with a value range of [0,1] and α+β=1; This represents the number of intermediate nodes. This is the feature vector of the m-th intermediate node after intervention in node A; The feature vector of the m-th intermediate node before intervention; The propagation delay through the m-th intermediate node; is the time decay coefficient, which is a positive real number.

[0056] It should be noted that the calculation of the causal effect score mainly consists of two parts: direct causal effect and indirect causal effect. When calculating the direct causal effect, the feature vectors of the target node B before and after the intervention of the source node A are first obtained. The degree of feature change is measured by calculating the Euclidean distance between these two feature vectors, and normalization is performed using the maximum value of the original feature vector of the target node. When calculating the indirect causal effect, for each intermediate node, the normalized Euclidean distance of the intermediate node's feature vector before and after the intervention of the source node is also calculated, and a time decay term is introduced so that intermediate nodes with longer propagation delays contribute less to the causal effect. The time decay coefficient λ is used to control the decay rate. The final indirect causal effect score is obtained by averaging the contributions of all intermediate nodes. Due to the use of normalization and weight constraints, the value range of this causal effect score is limited to between 0 and 1, where a score closer to 1 indicates a stronger causal relationship, and a score closer to 0 indicates a weaker causal relationship.

[0057] The fault diagnosis module, based on the fault tracing path and combined with fault feature representation, dynamically updates the diagnostic rule base and outputs results according to the following steps: First, it performs fault propagation feature analysis on the identified fault tracing path, constructing a complete fault scenario description by extracting the abnormal state features of each node, the strength of causal relationships between nodes, and fault propagation time sequence information. After obtaining the fault feature representation, it performs similarity matching with existing fault patterns in the diagnostic rule base, using an attention-based feature similarity calculation method, focusing on the matching degree of key features such as fault location, fault propagation sequence, and equipment operating parameter deviations. If the similarity between the current fault scenario and a certain fault pattern in the rule base exceeds a preset threshold, the diagnostic rule corresponding to that pattern is directly used for fault location and classification. When a new fault pattern is found or existing rules cannot accurately describe the current fault scenario, the diagnostic rule base needs to be dynamically updated. The specific update process includes extracting typical feature patterns of the current fault scenario, establishing causal association rules for fault propagation, and setting threshold conditions for fault discrimination. When updating rules, the boundary conditions for fault occurrence, the time constraints of fault propagation, and the evolution law of equipment state are carefully considered to ensure that the new rules have strong universality and robustness. Once the rule base is updated, a multi-layered fault diagnosis strategy is adopted based on the latest diagnostic rules, combined with real-time monitoring data and historical operation records of the equipment. First, at the macro level, the area and key equipment where the fault occurs are located. Then, at the micro level, the specific fault type and severity are analyzed. If multiple devices are found to be potentially faulty simultaneously, the priority, scope of impact, and urgency of the faults are comprehensively considered, and the fault location results are output in descending order of importance. When outputting the fault type, not only is the specific classification of the fault given, but also the possible causes, development trends, and potential risks of the fault are provided to support decision-making in subsequent fault handling. After the fault diagnosis is completed, the diagnostic results are compared and verified with the actual situation. Based on the verification results, the diagnostic rule base is continuously optimized and improved to enhance the accuracy and reliability of fault diagnosis.

[0058] In summary, the power grid fault rapid diagnosis system based on the AI-integrated algorithm of this invention is explained. It collects three-phase voltage and current data from the distribution network and real-time output data from distributed photovoltaic (PV) power plants, uses an adaptive morphological filter to suppress data noise, and thus extracts fault feature data under PV disturbances. Furthermore, it constructs a fault feature encoder through a contrastive learning algorithm to extract fault feature representations unaffected by PV fluctuations. Finally, it combines causal reasoning-enhanced graph neural network analysis to analyze the fault propagation relationships between nodes in the distribution network topology and generate fault tracing paths. This effectively improves the accuracy of fault feature extraction under PV disturbance conditions, while also enhancing the precision and efficiency of fault tracing and location, contributing to rapid and accurate power grid fault diagnosis and ensuring the safety and reliability of distribution network operation.

[0059] According to another aspect of the present invention, a method for rapid diagnosis of power grid faults integrating AI algorithms is provided, such as... Figure 2 As shown, it includes:

[0060] Collect three-phase voltage and current data and real-time output data of distributed photovoltaic power at each monitoring point of the distribution network, and use an adaptive morphological filter to suppress noise in the three-phase voltage and current data to obtain fault characteristic data under photovoltaic disturbance.

[0061] A fault feature encoder was constructed using a contrastive learning algorithm to extract fault feature representations that are unaffected by photovoltaic fluctuations;

[0062] Based on the fault characteristics, a graph neural network enhanced with causal reasoning is used to analyze the fault propagation relationship between nodes in the distribution network topology and generate fault source tracing paths.

[0063] Based on the fault tracing path and combined with the fault feature representation, the diagnostic rule base is dynamically updated, and the fault equipment location result and fault type are output.

[0064] Here, those skilled in the art will understand that the specific operations of each step in the above-described method for rapid diagnosis of power grid faults integrating AI algorithms have been referenced above. Figure 1 and Figure 2 The description of the power grid fault rapid diagnosis system integrating AI algorithms is detailed here, and therefore, its repeated description will be omitted.

[0065] In summary, the rapid fault diagnosis method for power grids based on the fusion AI algorithm of this invention has been clarified. It collects three-phase voltage and current data from the distribution network and real-time output data from distributed photovoltaic (PV) power plants, and uses an adaptive morphological filter to suppress data noise, thereby extracting fault feature data under PV disturbances. Furthermore, a fault feature encoder is constructed using a contrastive learning algorithm to extract fault feature representations unaffected by PV fluctuations. Finally, a graph neural network enhanced with causal reasoning is combined to analyze the fault propagation relationships between nodes in the distribution network topology, generating fault tracing paths. This effectively improves the accuracy of fault feature extraction under PV disturbance conditions, while also enhancing the precision and efficiency of fault tracing and location, contributing to rapid and accurate power grid fault diagnosis and ensuring the safety and reliability of distribution network operation.

[0066] This embodiment also provides a computer device suitable for a power grid fault rapid diagnosis system integrating AI algorithms, including a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the power grid fault rapid diagnosis system integrating AI algorithms as proposed in the above embodiment.

[0067] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0068] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the power grid fault rapid diagnosis system integrating AI algorithms as proposed in the above embodiments.

[0069] The storage medium proposed in this embodiment and the data storage method proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0070] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A power grid fault rapid diagnosis system fused with AI algorithm, characterized in that, Comprise: a fault feature acquisition and processing module, for acquiring three-phase voltage and current data of each monitoring point of the power distribution network and real-time output data of the distributed photovoltaic, and using an adaptive morphological filter to suppress noise of the three-phase voltage and current data to obtain fault feature data under photovoltaic disturbance; a fault feature encoding module, for mapping the fault feature data and the photovoltaic output data to a high-dimensional feature space, and extracting fault feature representations not affected by photovoltaic fluctuations; a graph neural network module, for analyzing fault propagation relationships between nodes in the power distribution network topology according to the fault feature representations; a causal reasoning module, for identifying fault sources based on energy transmission directions, and generating fault tracing paths; a fault diagnosis module, for updating a diagnosis rule base based on the fault tracing paths and the fault feature representations, and outputting fault equipment positioning results and fault types.

2. The fused AI algorithm based power grid fault fast diagnostic system of claim 1, wherein, The adaptive morphological filter performs adaptive segmented processing on the acquired three-phase voltage and current data, dynamically selects the size of the structure element according to the local variance and the peak value factor, and uses morphological opening and closing operation cascade operation to effectively suppress measurement noise and photovoltaic fluctuations.

3. The fused AI algorithm based power grid fault fast diagnostic system of claim 2, wherein, The dynamic selection of the size of the structure element includes: when the local variance is greater than a preset threshold A and the peak value factor is greater than a preset threshold B, a small-size one-dimensional structure element with a length of 3 sampling points is selected; when the local variance is less than the preset threshold A or the peak value factor is less than the preset threshold B, a large-size one-dimensional structure element with a length of 7 sampling points is selected.

4. The fused AI algorithm based power grid fault fast diagnostic system of claim 3, wherein, The fault feature encoding module uses a four-layer fully connected neural network as a basic encoder, uses an InfoNCE loss function to make feature representations of the same fault under different photovoltaic outputs close to each other, and feature representations of different faults far away from each other, so as to extract fault intrinsic features not affected by photovoltaic fluctuations.

5. The fused AI algorithm based power grid fault fast diagnostic system of claim 4, wherein, The InfoNCE loss function calculates the similarity of samples as shown in the following formula: ， wherein, is a sample similarity, is a number of samples in a batch; is a feature vector of the i-th anchor sample after L2 normalization; is a positive sample feature vector of the i-th sample under the k-th photovoltaic fluctuation; is a j-th negative sample feature vector; is a temperature parameter; and · denotes a vector dot product operation.

6. The fused AI algorithm based power grid fault fast diagnostic system of claim 5, wherein, The causal reasoning module uses an counterfactual reasoning framework to perform intervention experiments on nodes based on the node features of the graph neural network, analyzes the propagation path and influence range of the fault by calculating the causal effect score and constructing the causal relationship matrix.

7. The fused AI algorithm based power grid fault fast diagnostic system of claim 6, wherein, The calculation of the causal effect score includes direct causal effect and indirect causal effect; Both the direct causal effect and the indirect causal effect use the Euclidean distance to quantify the degree of feature change, but the indirect causal effect analyzes the feature changes of all intermediate nodes and uses a time decay term to reduce the influence of propagation.

8. A power grid fault rapid diagnosis method fused with an AI algorithm, characterized in that, Comprise: acquiring three-phase voltage and current data of each monitoring point of the power distribution network and real-time output data of the distributed photovoltaic, and using an adaptive morphological filter to suppress noise of the three-phase voltage and current data to obtain fault feature data under photovoltaic disturbance; using a contrastive learning algorithm to construct a fault feature encoder to extract fault feature representations not affected by photovoltaic fluctuations; using a causal reasoning enhanced graph neural network to analyze fault propagation relationships between nodes in the power distribution network topology according to the fault feature representations, and generating fault tracing paths; based on the fault tracing paths and the fault feature representations, dynamically updating a diagnosis rule base to output fault equipment positioning results and fault types. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is characterized in that: The processor implements the steps of the power grid fault rapid diagnosis system of the fusion AI algorithm according to any one of claims 1-7 when executing the computer program.

10. A computer readable storage medium having stored thereon a computer program, characterized in that: The computer program is executed by the processor to implement the steps of the power grid fault rapid diagnosis system of the fusion AI algorithm according to any one of claims 1-7.

Images