Power transmission line fault diagnosis and operation scheduling method and system based on network learning

Abstract

The application discloses a power transmission line fault diagnosis and operation and maintenance scheduling method and system based on network learning, relates to the technical field of power transmission line fault diagnosis and operation and maintenance scheduling, and comprises the following steps: acquiring electrical quantity data, equipment operation state data, ring network topology and time sequence data and the like, extracting relevant data to construct a high-risk equipment area and a visual high-risk area heat map, combining the electrical quantity data to construct a visual risk classification map, combining a convolutional neural network-long short-term memory network hybrid model to construct an intelligent recognition storage network and a fault type classification recognition model, optimizing the model through network learning and an attention mechanism, combining an operation and maintenance management system to autonomously allocate a maintenance team and resources, and then realizing autonomous optimization and closed-loop operation, so that the application realizes full-process coverage from fault perception, intelligent decision and efficient response, significantly shortens the reaction time after line fault occurrence, and comprehensively enhances the fault disposal and operation and maintenance capability of the power grid system.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of power transmission line fault diagnosis operation and maintenance scheduling, in particular to a power transmission line fault diagnosis and operation and maintenance scheduling method and system based on network learning. BACKGROUND

[0002] With the rapid development of modern industry and the continuous improvement of urbanization level, as an important infrastructure for power transmission, the application range and importance of power cables are increasingly prominent. Under the current economic and social development background, people's dependence on power cables is deepening, so the reliability and safety of the operation of power cables have attracted widespread attention. The stable operation of the cable is not only related to the normal work of the power system, but also directly affects the power supply guarantee of industrial production and daily life. Therefore, improving the operation reliability of power cables has become one of the key factors to ensure the stable operation of the power system.

[0003] Traditional power transmission systems mostly use single-line power transmission mode, and line faults have long relied on manual troubleshooting, that is, professional personnel are sent to conduct on-site patrol, and maintenance personnel need to carry a large number of equipment to deal with various faults. With the expansion of social electricity demand, the power grid structure is becoming more and more complex, forming a looped network power supply system, and the operation mode is more diverse and dynamic. Under this background, the traditional manual fault diagnosis and line patrol mode is not up to the task, not only requiring a large amount of manpower, but also having low response and processing efficiency in the face of complex faults, which is difficult to meet the modern power grid high reliability operation demand. At the same time, the operation and maintenance decision depends on the experience of personnel, the fault diagnosis and operation scheduling information sharing is not smooth, and there is a lack of collaborative mechanism, resulting in slow response and unreasonable resource allocation. In addition, the system lacks intelligent mechanism, and the new power system needs to accumulate operation and maintenance knowledge from zero, which affects the industry technology improvement and operation and maintenance mode innovation. SUMMARY

[0004] To solve the above technical problems, the power transmission line fault diagnosis and operation scheduling method and system based on network learning are provided, which solves the problem that the traditional power transmission system mainly adopts single line transmission mode, and the power transmission line fault is mainly relied on manual investigation and repair, that is, professional personnel are dispatched to conduct on-site patrol along the line, and the fault position is located and diagnosed, with the continuous expansion of the power consumption scale, the increasing requirement of power supply reliability, and the increasingly complex power network structure, a large number of newly added power transmission lines are connected to form a ring network power supply system, under this background, the traditional manual fault diagnosis and line patrol method is increasingly difficult to cope with the line fault, the operation decision still highly depends on manual experience, there is a lack of effective cooperation between fault diagnosis and operation scheduling, which leads to slow overall response speed, unreasonable resource allocation, lack of continuous learning and self-improvement mechanism of the system, and the operation experience of each region is difficult to realize safe and efficient sharing and inheritance, so that each new system needs to accumulate knowledge from zero, and the learning cost is repeated.

[0005] To achieve the above purposes, the technical scheme adopted by the present application is:

[0006] The power transmission line fault diagnosis and operation scheduling method and system based on network learning comprises:

[0007] Obtain electrical quantity data, equipment operation state data, operation management system, ring network topology and time sequence data, and historical fault data and alarm log data corresponding to the power transmission line ring network;

[0008] According to the PMU phasor data in the electrical quantity data, the power grid fault dynamic characteristics are obtained, and according to the power grid fault dynamic characteristics, the equipment operation state data and the ring network topology and time sequence data, a high-risk equipment area is established;

[0009] According to the space-time characteristics of the GIS map and the PMU phasor data in the electrical quantity data, a visual high-risk area heat map is generated in combination with the high-risk equipment area;

[0010] According to the PMU phasor data and the visual high-risk area heat map, a hierarchical index is established, and a visual risk classification map is established;

[0011] According to the visual risk classification map, the ring network topology and time sequence data, and the convolutional neural network-long short-term memory network hybrid model, an intelligent recognition storage network is established;

[0012] According to the historical fault data and alarm log data, a fault type classification recognition model is established in combination with the convolutional neural network-long short-term memory network hybrid model;

[0013] According to the recognition results of the intelligent recognition storage network and the fault type classification recognition model, and the operation management system, a repair team and resources are autonomously allocated;

[0014] Taking the intelligent identification storage network and fault type classification identification model as the local node, model parameters of the intelligent identification storage network and fault type classification identification model are extracted, and are respectively marked as local network parameters and local identification parameters, and are uploaded to the central server;

[0015] According to the attention mechanism, the local network parameters and the local identification parameters are fused, and corresponding models are respectively established according to the global convolutional neural network-long short-term memory network hybrid model, and the local node is updated according to the central server data.

[0016] Preferably, the power grid fault dynamic characteristics are obtained according to the PMU phasor data in the electrical quantity data, and a high-risk equipment area is established according to the power grid fault dynamic characteristics, the equipment operating state and the ring network topology and time sequence data, specifically including:

[0017] The electrical impact parameters of each node in the power grid fault dynamic characteristics are extracted, and the electrical impact parameters are normalized and weighted calculated according to the safe operation standard of the power system, to obtain the impact severity;

[0018] The real-time operation data of the equipment in the equipment operating state data are extracted and the equipment health degree is evaluated through a parameter fusion algorithm, and the power transmission network topology graph is extracted from the ring network topology and time sequence data;

[0019] A risk decision matrix is established according to the impact severity of each node and the equipment health degree, the matrix elements are cross-operated and the comprehensive risk level of each node is quantified, and a risk quantization data set is constructed;

[0020] According to the risk quantization data set output by the risk decision matrix and the power transmission network topology graph, a DBSCAN algorithm based on density clustering is used to establish a high-risk equipment area.

[0021] Preferably, the visual high-risk area heat map is generated according to the GIS map in the electrical quantity data and the spatio-temporal characteristics of the PMU phasor data combined with the high-risk equipment area, specifically including:

[0022] The terrain and the tower line direction in the GIS map are obtained, and a ring network space base map is constructed;

[0023] The ring network space base map is finely divided according to the ring network space base map and the ring network operation precision requirement, and is marked as a spatial risk unit;

[0024] The time dimension, space dimension and data dimension in the PMU phasor data are obtained and mapped, a PMU risk matrix is generated, and the PMU risk matrix is mapped to the ring network space base map to establish a primary visual map;

[0025] Obtain the power transmission line map and the equipment operation state in the real-time high-risk equipment area, and map the power transmission line map and the equipment operation state to the primary visualization map to establish a visualization risk area heat map.

[0026] Preferably, the hierarchical index is established according to the PMU phasor data and the visualization high-risk area heat map, and a visualization risk classification map is established, specifically including:

[0027] Obtain the multi-dimensional fault feature in the PMU phasor data, establish a fault feature severity classification index according to the power system safe operation standard and historical fault data, and map the standardized multi-dimensional fault feature to an interval according to range normalization.

[0028] Classify the standardized multi-dimensional fault feature according to the fault feature severity classification index and the K-means clustering algorithm, set the number of clusters to 3, assign the standardized multi-dimensional fault feature to the corresponding risk level cluster, and mark it as a fault feature classification matrix;

[0029] Obtain the equipment operation state data in the visualization high-risk area heat map, classify the equipment risk data according to the three quantile method, and construct an equipment risk data classification matrix;

[0030] Align the fault feature classification matrix and the equipment risk data classification matrix, generate a multi-dimensional fusion data set, and map the multi-dimensional fusion data set to the loop network space base map, and establish a visualization risk classification map according to color coding.

[0031] Preferably, the intelligent recognition storage network is established according to the visualization risk classification map, the loop network topology and the time series data, and a convolutional neural network-long short-term memory network hybrid model, specifically including:

[0032] Obtain the spatial distribution characteristics of each node in the visualization risk classification map, convert the visualization risk classification map into a three-dimensional feature tensor, input the three-dimensional feature tensor into a convolutional neural network for multi-level spatial feature extraction and output a one-dimensional spatial feature vector containing global spatial risk distribution rules;

[0033] Obtain the risk feature evolution law in the loop network topology and time series data, organize the dynamic topology feature into graph sequence data, input the graph sequence data into a long short-term memory network for modeling and output a one-dimensional time series vector for capturing long-term dependence and mutation characteristics of risk evolution;

[0034] According to the attention mechanism, the one-dimensional spatial feature vector of the convolutional neural network and the one-dimensional time series vector of the long short-term memory network are calculated by weight association, fused and an intelligent recognition storage network is established.

[0035] Preferably, the convolutional neural network-long short-term memory network hybrid model is established according to historical fault data and alarm log data, and specifically comprises:

[0036] The fault static features and environmental parameters in the historical fault data are acquired, a fault static feature set is constructed, and the fault static feature set is input into a convolutional neural network for multi-level static feature extraction,

[0037] The alarm log data is acquired, sorted according to timestamps, and an alarm time sequence is constructed, and the alarm time sequence is input into a long short-term memory network for time sequence correlation analysis;

[0038] The attention weights of the fault static features and the alarm time sequence correlation are calculated according to the mutual attention mechanism, and are fused into multi-dimensional joint features, and a fault type classification and recognition model is established according to the multi-dimensional joint features.

[0039] Preferably, the intelligent recognition storage network, the fault type classification and recognition model recognition result, and the operation and maintenance management system are used to autonomously allocate a maintenance team and resources, and specifically comprise:

[0040] The power transmission line loop network is identified according to the intelligent recognition storage network, and risk core information and risk evolution trends are acquired;

[0041] The fault core information and fault disposal requirements in the risk core information and risk evolution trends are identified according to the fault type classification and recognition model;

[0042] The states and load priorities of manpower and equipment in the operation and maintenance management system are acquired;

[0043] Unified quantitative indicators are constructed according to the fault disposal requirements, the states and load priorities of manpower and equipment, and a risk processing matrix is established;

[0044] Differential allocation indicators are formulated according to the risk processing matrix and the operation and maintenance management system, an optimal allocation scheme is solved according to a genetic algorithm, and manpower and equipment are autonomously allocated.

[0045] Preferably, the intelligent recognition storage network and the fault type classification and recognition model are taken as local nodes, model parameters of the intelligent recognition storage network and the fault type classification and recognition model are extracted, and are respectively marked as local network parameters and local recognition parameters, and are uploaded to a central server, and specifically comprise:

[0046] Core recognition data and feature fusion labeling rules in the intelligent recognition storage network are extracted, and are marked as local network parameters, and the core recognition data specifically comprise: spatial feature coding parameters, time sequence evolution modeling parameters, and memory storage mechanism parameters;

[0047] The core fault data and classification decision marking rules in the fault type classification and identification model are extracted and marked as local identification parameters, and the core fault data specifically includes: static feature extraction parameters, dynamic sequence modeling parameters, and knowledge evolution mechanism parameters.

[0048] The local network parameters and local identification parameters are preprocessed, a parameter verification mechanism is established, parameter abnormal distribution points are identified and processed according to an outlier detection algorithm, and the processed parameters are marked as upload parameters and uploaded to the central server.

[0049] Preferably, the multiple sets of local network parameters and local identification parameters are fused according to the attention mechanism, and corresponding models are established according to the global convolutional neural network-long short-term memory network hybrid model, and the local nodes are updated according to the central server data, specifically including:

[0050] The attention weight calculation is performed on the multiple sets of upload parameters according to the attention mechanism, and the results are marked as initial training model parameters and unified classification parameters, respectively;

[0051] The global intelligent identification storage network and the global fault type classification and identification model are established according to the global convolutional neural network-long short-term memory network hybrid model and the initial training model parameters and the unified classification parameters;

[0052] The global parameters after passing the verification of the global intelligent identification storage network and the global fault type classification and identification model are issued to each local node.

[0053] Preferably, the data acquisition module is used to acquire electrical quantity data, equipment operating state data, ring network topology and time sequence data, and historical fault data and alarm log data corresponding to the power transmission line ring network;

[0054] The chart establishment module is used to extract features corresponding to different data in the power transmission line ring network, and to establish high-risk equipment areas, visual high-risk area heat maps, and visual risk classification maps according to the corresponding features;

[0055] The model training module is used to establish an intelligent identification storage network according to the visual risk classification map, the ring network topology and time sequence data, and the convolutional neural network-long short-term memory network hybrid model, and to establish a fault type classification and identification model according to the historical fault data and alarm log data combined with the convolutional neural network-long short-term memory network hybrid model;

[0056] The operation and maintenance scheduling module is used to identify the output parameters of the intelligent identification storage network and the fault type classification and identification model, and to construct a risk processing matrix and a differentiated allocation index according to the output parameters and the operation and maintenance management system, and to autonomously allocate operation and maintenance teams and resources.

[0057] The networking learning module is configured to extract local network parameters and local identification parameters in the intelligent identification storage network and the fault type classification identification model in the local node, pre-process the local network parameters and the local identification parameters, and upload the local network parameters and the local identification parameters to the central server.

[0058] The global cooperation module is configured to perform attention weight calculation on the uploaded parameters of the multiple local nodes according to an attention mechanism, respectively input initial training model parameters and unified classification parameters into a global convolutional neural network-long short-term memory network hybrid model, and then establish a global intelligent identification storage network and a global fault type classification identification model, and distribute the global intelligent identification storage network and the global fault type classification identification model parameters that pass the verification to the local nodes.

[0059] Compared with the prior art, the beneficial effects of the present application are as follows:

[0060] In the present application, by adopting a multi-source data fusion technology, combining a convolutional neural network and a long short-term memory network hybrid model, the precision identification and intelligent judgment ability of complex fault types in the power grid are significantly improved, and with the help of a networking learning mechanism, information can be continuously obtained from new data and parameters can be autonomously optimized, so that the ability of continuous iteration and self-updating is possessed, the dynamic changes in the power grid operation can be effectively adapted, and the adjustment requirements of various new fault types can be flexibly responded to, realizing full-process coverage from fault perception, intelligent decision and efficient response, significantly shortening the reaction time after the occurrence of line fault, and comprehensively enhancing the fault disposal and operation and maintenance ability of the power grid system. BRIEF DESCRIPTION OF DRAWINGS

[0061] Figure 1 The figure is a whole process schematic diagram of the power transmission line fault diagnosis and operation and maintenance scheduling method based on networking learning proposed in the present application.

[0062] Figure 2 The figure is a process schematic diagram of autonomous operation and maintenance scheduling decision of the present application.

[0063] Figure 3 The figure is a process schematic diagram of distributed federated learning of the present application.

[0064] Figure 4 The figure is a process schematic diagram of the power transmission line fault diagnosis and operation and maintenance scheduling system based on networking learning proposed in the present application. DETAILED DESCRIPTION

[0065] The following description is used to disclose the present application so that those skilled in the art can implement the present application. The preferred embodiments in the following description are only as examples, and other obvious modifications can be thought of by those skilled in the art.

[0066] Referring to Figures 1-3 The power transmission line fault diagnosis and operation scheduling method and system based on network learning are shown, and the method comprises:

[0067] Obtain electrical quantity data, device operating state data, operation and maintenance management system, ring network topology and timing data, and historical fault data and alarm log data corresponding to the power transmission line ring network;

[0068] According to the PMU phasor data in the electrical quantity data, obtain the power grid fault dynamic characteristics, and according to the power grid fault dynamic characteristics, the device operating state data and the ring network topology and timing data, establish a high-risk device area;

[0069] According to the GIS map and the spatiotemporal characteristics of the PMU phasor data in the electrical quantity data, a visual high-risk area heat map is generated in combination with the high-risk device area;

[0070] According to the PMU phasor data and the visual high-risk area heat map, a hierarchical index is established, and a visual risk classification map is established;

[0071] According to the visual risk classification map, the ring network topology and timing data, and the convolutional neural network-long short-term memory network hybrid model, an intelligent recognition storage network is established;

[0072] According to the historical fault data and alarm log data, a fault type classification recognition model is established in combination with the convolutional neural network-long short-term memory network hybrid model;

[0073] According to the recognition results of the intelligent recognition storage network and the fault type classification recognition model, and the operation and maintenance management system, a repair team and resources are autonomously allocated;

[0074] Taking the intelligent recognition storage network and the fault type classification recognition model as local nodes, the model parameters of the intelligent recognition storage network and the fault type classification recognition model are extracted, respectively marked as local network parameters and local recognition parameters, and uploaded to a central server;

[0075] According to the attention mechanism, the local network parameters and the local recognition parameters are fused, and corresponding models are respectively established according to the global convolutional neural network-long short-term memory network hybrid model, and the local nodes are updated according to the central server data.

[0076] Further, according to the PMU phasor data in the electrical quantity data, the power grid fault dynamic characteristics are obtained, and according to the power grid fault dynamic characteristics, the device operating state and the ring network topology and timing data, a high-risk device area is established, which specifically comprises:

[0077] Extract the node electrical impact parameter in the power grid fault dynamic characteristics, and according to the electrical impact parameter, the electrical impact parameter is normalized and weighted according to the power system safety operation standard, and the node impact severity is obtained;

[0078] wherein the phasor measurement units deployed in the looped network collect real-time three-phase voltage and current phasor data of the power grid at a sampling rate of no less than 1 kHz, and the data is mapped to interval by range normalization, and the electrical shock parameters that cause system instability and equipment damage are taken as the criterion layer according to the safe operation standards of the power system, the accurate assessment of the severity of the electrical node shock is taken as the target layer, the electrical shock parameters of the criterion layer are compared according to the safe operation standards of the power system, the degree of the influence of the equipment damage on the system stability is graded by the scaling method, and the weight of the electrical shock parameters is calculated by the square root method and normalization processing, and the electrical shock parameters are calculated to obtain the severity of the node shock;

[0079] extracting real-time operation data of the equipment from the equipment operation state data, evaluating the equipment health degree, and extracting the power transmission network topology graph from the looped network topology and time sequence data;

[0080] integrating the equipment state monitoring system, obtaining the operation parameters of key equipment such as transformers and circuit breakers, and the power transmission network topology graph in the looped network topology and time sequence data, obtaining the dynamic characteristics of the power grid fault including extracting time domain characteristics, frequency domain characteristics and transient mutation characteristics, normalizing the obtained data into equipment health state indicators, eliminating the dimensional differences, comparing the obtained data by the analytic hierarchy process, calculating the weight, and then evaluating the equipment health degree by the weighted scoring method );

[0081] establishing a risk decision matrix according to the severity of the shock of each node and the equipment health degree, cross-operating the matrix elements and quantifying the comprehensive risk level of each node, and constructing a risk quantification data set;

[0082] wherein the risk decision matrix is constructed, the horizontal axis is the severity of the node shock, the vertical axis is the equipment health degree, the cell in the risk decision matrix is located, the pre-defined comprehensive risk level is determined by cross-operation of the matrix elements, and the comprehensive risk level calculation formula is:

[0083] ;

[0084] wherein, is the normalized shock severity score, is the normalized equipment health degree score, and = 0.6, = 0.4, according to the calculated comprehensive risk level, the risk level of the node is divided into four levels, wherein: low risk is set, medium risk is set, high risk is set, The device health degree is obtained by multi-parameter fusion, wherein the transformer state is judged based on oil temperature, load rate and oil chromatographic data, the circuit breaker state is judged according to the operation number and SF6 pressure, and the line state is judged by considering the conductor temperature and leakage current, and finally the node impact degree score and the device health degree score are cross-mapped in the risk decision matrix;

[0085] According to the risk quantification data set output by the risk decision matrix and the power transmission network topology map, a high-risk equipment area is established by using the DBSCAN algorithm based on density clustering;

[0086] Each power grid equipment operation state data in the risk quantification data set is taken as sample data, the electrical connection relationship is analyzed in combination with the power transmission network topology map, the field radius and the minimum sample number of the core point are identified by the DBSCAN algorithm, and all device nodes are traversed by the DBSCAN algorithm, the sample data with the same density and the node risk level of high risk and above are divided into a cluster, and the accurate division of the high-risk equipment area is formed, and the high-risk equipment area is superimposed with the GIS map to construct the high-risk equipment area, and provide spatial positioning basis for subsequent operation.

[0087] Further, according to the time and space characteristics of the GIS map and the PMU phasor data in the electrical quantity data, a visual high-risk area heat map is generated in combination with the high-risk equipment area, specifically including:

[0088] The trend of topography and the trend of tower line of the GIS map are obtained, and a loop network space base map is constructed;

[0089] The spatial coordinates of the trend of topography and the trend of tower line of the GIS map are obtained, and the loop network space base map with geographical reference is constructed;

[0090] According to the loop network space base map and the loop network operation precision requirement, the loop network space base map is finely divided, and is marked as a spatial risk unit;

[0091] According to the loop network operation precision requirement, the space base map is divided into uniform grid units, each grid is defined as a spatial risk unit, the size of which is determined by the operation precision requirement, the space map is divided by the equal-area grid division method, and the area of the grid unit is ensured to be consistent;

[0092] The time dimension, space dimension and data dimension in the PMU phasor data are obtained and mapped, a PMU risk matrix is generated, and the PMU risk matrix is mapped to the loop network space base map to establish a primary visual map;

[0093] Multi-dimensional feature extraction is performed on the PMU phasor data, a fault dynamic evolution sequence is extracted from a time dimension, geographical position distribution of monitoring points is obtained from a space dimension, and a risk quantization value of each node is calculated from a data dimension, a PMU risk matrix is generated through feature fusion, the risk matrix is mapped to a looped network space base map through a space interpolation algorithm, and a primary visualized map including a basic risk distribution is established;

[0094] A real-time high-risk equipment area transmission line map and equipment operation state are obtained, and the transmission line map and the equipment operation state are mapped to the primary visualized map, and a visualized risk area heat map is established;

[0095] The transmission line topology and the equipment operation state data in the real-time identified high-risk equipment area are superimposed to the primary visualized map, the comprehensive risk value of each space risk unit is converted into a corresponding color intensity through a risk value to color mapping function, and a final visualized risk area heat map is generated, in which a warm color tone represents a high-risk area and a cold color tone represents a low-risk area, and an intuitive space risk distribution visualized display is formed.

[0096] Further, a hierarchical index is established according to the PMU phasor data and the visualized high-risk area heat map, and a visualized risk classification map is established, specifically including:

[0097] Multi-dimensional fault features in the PMU phasor data are obtained, a fault feature severity classification index is established according to power system safe operation standards and historical fault data, and the standardized multi-dimensional fault features are mapped to an interval according to range normalization;

[0098] The multi-dimensional fault features in the PMU phasor data include time domain voltage and current rise and fall amplitudes, harmonic distortion rates, and transient feature phase mutation angles and rise slope parameters, extreme values caused by sensor interference and data transmission error codes are identified and removed by using a box plot method, at the same time, through Pearson correlation coefficient analysis, redundant features with a correlation greater than 0.85 are removed, range normalization method is selected to eliminate dimensional differences of multi-dimensional fault features, the fault feature severity classification index is established according to power system safe operation standards and historical fault statistical data, the fault severity is divided into four risk levels according to voltage sag amplitude, current impact multiple, harmonic distortion rate and phase mutation angle, high risk (feature parameter seriously exceeds the standard, fault risk is extremely high), medium risk (feature parameter is obviously abnormal, there is a fault risk), low risk (feature parameter appears slight abnormality, in a warning state) and normal operation (feature parameter is completely within the safe operation range), a fixed feature classification matrix is formed, and the highest fault level is marked when superimposing multiple features according to the principle of not lowering the high level;

[0099] According to the fault feature severity grading index and the K-means clustering algorithm, the standardized multi-dimensional fault features are graded, the number of clusters is set to 3, the standardized multi-dimensional fault features are assigned to the corresponding risk level cluster, and are marked as the fault feature grading matrix;

[0100] The standardized multi-dimensional fault features are checked, the missing values are filled in by the feature mean value of the risk level sample, then the historical fault data is divided, the abnormal samples deviating too much are removed by the 3σ criterion, the purified data is marked as sample data, the weight vector is calculated by the square root method and is normalized, and then the weight is obtained, the maximum eigenvalue and the corresponding eigenvector are calculated by the eigenvalue method, and the corresponding eigenvector is normalized, the obtained result is set as the initial weight vector of the fault feature, the weight vector is formed, the weighted Euclidean distance is defined according to the weight vector, which is used as the distance measurement standard of the K-means clustering algorithm, the K-means++ initialization strategy is used, the sample points with the farthest distance from each other are selected as the initial cluster centers according to the probability distribution of the fault feature severity grading index, the cluster centers are initialized, the weighted Euclidean distance between the sample data and the cluster centers is calculated, the sample data is assigned, and then the initial cluster is formed, then the weighted mean of each feature of all sample data in the initial cluster is calculated, the cluster center is updated, and the assignment and update are repeated until the cluster center moving distance ≤0.0001 or the repetition number reaches 1000 times, after the iteration is completed, the proportion of the historical fault risk level of the samples in the cluster is calculated, if the proportion of a single risk level in the cluster is more than 90%, it is qualified, otherwise, the initial cluster center is updated and the update iteration is performed, the risk level and the cluster center are mapped according to the average standardized feature mean value of the samples in the cluster, and the mapping result is marked as the fault feature grading matrix;

[0101] The device running state data in the visual high-risk area heat map is obtained, the device risk data is graded according to the three quantile method, and the device risk data grading matrix is constructed;

[0102] The device risk data in the visual high-risk area heat map is obtained, including the device risk probability calculated by fusing the PMU space-time features and the device abnormal data (the PMU space-time features and the device abnormal data are processed by the range normalization method and the weight vector is calculated by the square root method, and the weighted calculation), the abnormal parameter amplitude of the device abnormal parameter relative to the threshold value of the rated value, and the spatial correlation degree of the abnormal device and the surrounding high-risk device (the coordinates are unified and the distance is calculated by the Haversine formula, then the influence of the distance and the electrical correlation on the correlation degree is calculated by the inverse distance weighting and the electrical correlation correction, and the formula is: , in the formula, is the geographical distance between devices, For the electrical correlation between devices, For the minimum positive extreme value, For the distance attenuation factor (value 2), then measure the dispersion degree of all high-risk devices in space through the dispersion degree correction coefficient, and the formula is: , wherein, For the weighted average distance, For the distance standard deviation, when high-risk devices are clustered Tends to 1, the weighted contribution of the device is calculated, and finally the spatial correlation is obtained through normalization processing), while eliminating the fuzzy area of the visual high-risk area heat map and the abnormal data caused by sensor failure, calculating the full data quantile of the extracted standardized device risk parameters, dividing the three risk standards: emergency failure, early warning hidden danger and potential risk, and corresponding with the failure severity classification index, while the unclassified device parameters are marked as normal operation; At the same time, the device individual is taken as the row, and the fan core dimension is taken as the column to construct the device risk data classification matrix;

[0103] Align the fault feature classification matrix and the device risk data classification matrix, generate a multi-dimensional fusion data set, and map the multi-dimensional fusion data set to the ring network space base map, and establish a visual risk classification map according to the color coding;

[0104] Pretreatment is performed on the fault feature classification matrix and the device risk data classification matrix, the device ID is associated, the fault feature classification matrix and the device risk data classification matrix are associated through JOIN, a fusion data set mainly composed of device ID-fault feature level-device risk level-two-dimensional quantitative data is formed, the electrical characteristic risk and the physical characteristic risk of the device are ensured to correspond, the severity of the core fault feature of the fault feature level matrix is dimensionally aligned with the risk level and the spatial correlation degree in the device risk data classification matrix through feature dimension mapping, the fusion feature vector of the device is generated through the electrical characteristic risk and the physical characteristic risk, the fusion feature vector of the electrical characteristic risk and the physical characteristic risk is obtained through linear transformation layer, then the cross attention score between the fusion feature vectors of the electrical characteristic risk and the physical characteristic risk is calculated, the attention score is normalized using Softmax function to obtain the attention weight of the feature dimension, and finally the fusion calculation result is obtained through weighted feature fusion. The fusion calculation result is mapped to the GIS ring network space base map through color coding and dimension interaction function.

[0105] Further, according to the visual risk classification map, the ring network topology, the time series data and the convolutional neural network-long short-term memory network hybrid model, an intelligent identification storage network is established, which specifically includes:

[0106] The spatial distribution characteristics of each node in the visual risk classification diagram are obtained, the visual risk classification diagram is converted into a three-dimensional feature tensor, and the three-dimensional feature tensor is input into a convolutional neural network for multi-level spatial feature extraction and output of a one-dimensional spatial feature vector containing global spatial risk distribution rules;

[0107] In the grid data of the visual risk classification diagram, the spatial distribution characteristics of each node are extracted, including the geographic coordinate position of the node, the risk level, and the spatial distance from the adjacent node and the risk density of the region where the node is located. The visual risk classification diagram is converted into a three-dimensional feature tensor. The height and width are represented by geographical space grid division, and the channel encodes the multi-dimensional features fused in the grid unit. The three-dimensional feature tensor is deeply extracted by the convolutional layer of the convolutional neural network. Multi-layer convolution kernels are used to scan the visual risk classification diagram. The first convolutional layer mainly identifies the local risk distribution rule, the second convolutional layer mainly identifies the spatial correlation characteristics of a large area across regions, and the third convolutional layer mainly identifies the global risk distribution rule. At the same time of identifying the third convolutional layer, the attention mechanism combining channels and space is introduced, so that the convolutional neural network can adaptively focus on high-risk dense areas. The three-dimensional high-level feature map of the convolutional neural network enhanced by the attention mechanism is aggregated into a fixed-length one-dimensional spatial feature vector. After the model training is completed, the one-dimensional vector feature is projected to a two-dimensional space for visualization by t-SNE dimension reduction, and the feature effectiveness of the model is verified. The sample points derived from the same global risk rule show high cohesion in the two-dimensional space. If the clustering purity is greater than 90%, the feature is effective, otherwise the model parameters are adjusted and retrained;

[0108] The risk feature evolution rules in the ring network topology and time sequence data are obtained, the dynamic topology features are organized into graph sequence data, the graph sequence data is input into a long short-term memory network for modeling, and a one-dimensional time sequence vector is output for capturing the long-term dependence and mutation characteristics of risk evolution;

[0109] The static topology features including node set, edge set and topology matrix, dynamic topology features including topology structure time sequence change events and risk feature evolution rules are obtained. The dynamic topology features are organized into graph sequence data, the electrical connection relationship between nodes and the time causal dependence are retained, and the time steps are adaptively processed. The missing data is completed by time series interpolation method. The risk feature evolution rules are converted into time sequence input recognizable by long short-term memory network. The topology neighborhood information is integrated into the gating mechanism of long short-term memory network, and the risk evolution features of adjacent nodes are weighted and aggregated in the state update of the node itself.

[0110] According to the mutual attention mechanism, the one-dimensional spatial feature vector of the convolutional neural network and the one-dimensional time sequence vector of the long short-term memory network are calculated by weight association, fused and an intelligent recognition storage network is established.

[0111] According to the convolutional neural network and long short-term memory network fusion architecture, the one-dimensional spatial feature vector is reshaped in dimension, restored to a two-dimensional spatial feature map, and converted into a feature sequence input attention layer to obtain spatial feature vectors and time sequence feature vectors, and the spatial feature vector is calculated on the time sequence feature vector (Qs, Qs) , wherein, Qs is a query matrix generated by the spatial feature vector, Qk is a key matrix generated by the time sequence feature vector, is a scaling factor) and the time sequence feature vector on the spatial feature vector (Qk, Qk) , wherein, Qk is a query matrix generated by the time sequence feature vector, Qs is a key matrix generated by the spatial feature vector, is a scaling factor) attention distribution, and then the calculated time sequence dynamic information is injected into the static spatial feature, the spatial structure information is injected into the time sequence feature, and then the cross-attention weight between the spatial feature of the convolutional neural network and the state of the long short-term memory network is calculated. The convolutional neural network and the long short-term memory network can dynamically focus on the spatial region related to the current time sequence state, establish a hierarchical attention fusion mechanism, fuse the fine granularity of the bottom layer local spatial feature and the short-term time sequence mode, analyze the correlation between the middle layer complete area spatial mode and the medium-term evolution trend, and deeply integrate the global spatial distribution and the long-term risk law at the top layer. At the same time, a residual attention connection is designed to ensure effective gradient propagation and enhance feature reuse capability. At the same time, a memory enhancement module is constructed in the output layer, a storage mode is used to store the pattern through a memory matrix, and an attention mechanism is used to match and gradually shrink the current feature and the similar feature of the historical model. At the same time, according to the attention weight, the memory matrix storage is dynamically read and written to realize continuous accumulation and learning, so as to establish an intelligent recognition storage network.

[0112] Further, according to the historical fault data and alarm log data combined with the convolutional neural network-long short-term memory network hybrid model, a fault type classification recognition model is established, which specifically includes:

[0113] The fault static features and environmental parameters in the historical fault data are obtained, a fault static feature set is constructed, and the fault static feature set is input into the convolutional neural network for multi-level static feature extraction;

[0114] Among them, the fault waveform features, equipment parameters, steady-state comparison of electrical quantities before and after the fault, and environmental parameters in the extraction of fault static features are taken as sample data, and invalid sample data is removed, and according to the multi-dimensional design of the convolutional neural network, the adjacent features are reflected directly associated, and the multi-layer convolution and 1*1 pooling are used to prevent the loss of feature parameters of small matrices, the first layer of convolution kernel extracts the local mutation mode in the fixed waveform, the second layer of convolution kernel identifies the feature difference between different fault types, and the third layer of convolution kernel fuses the environmental parameters and electrical quantity features to form a comprehensive fault feature representation, extracts the static association in the sample data, and finally quantizes the mutual information of the feature vector and the fault type;

[0115] Acquire alarm log data, sort according to timestamp, construct alarm time sequence, input alarm time sequence into long short-term memory network for time sequence association analysis;

[0116] At the same time, the alarm log is reconstructed in time sequence, sorted according to timestamp, and the alarm type sequence, alarm interval length, alarm change trend and alarm propagation path are extracted, the long-term dependence relationship of the alarm event is learned through the long short-term memory network, the alarm evolution path is captured, and the abnormal mutation point and periodic change in the alarm sequence are identified;

[0117] According to the mutual attention mechanism, the attention weights of the fault static features and the alarm time sequence association are calculated, and are fused into multi-dimensional joint features, and a fault type classification recognition model is established according to the multi-dimensional joint features;

[0118] Through the mutual attention mechanism, the bidirectional attention weights between the fault static features and the alarm time sequence are calculated, the fault features are dynamically focused, the contribution proportion of static features and dynamic features is controlled through the gate vector, the fault feature representation containing historical early warning and real-time situation is formed, thereby the fusion of multi-dimensional joint features is completed, a hierarchical classification network is constructed according to the fused multi-dimensional joint features, the bottom classifier identifies the basic fault type, the middle classifier analyzes the composite fault mode, and the top classifier evaluates the fault severity level. At the same time, an uncertainty estimation mechanism is introduced, the confidence score is added according to the output result, and an online learning module is introduced, the learning is carried out according to the newly occurred low confidence fault case, and the classification boundary and feature weight are optimized, thereby the establishment of the fault type classification recognition model and the realization of autonomous optimization learning are completed.

[0119] Further, according to the intelligent identification storage network, the identification result, the fault type classification recognition model identification result and the operation and maintenance management system are autonomously allocated to the maintenance team and resources, specifically including:

[0120] According to the intelligent identification storage network, the power transmission line loop network is identified, the risk core information and the risk evolution trend are obtained;

[0121] According to the intelligent identification network, the power transmission line loop network is monitored and data is collected in real time, and the risk distribution in the loop network is identified through a space-time joint inference mechanism, and the spatial risk characteristics of each node in the loop network topology are extracted, the risk area and its spatial distribution pattern are identified, the risk influence range and node position are determined, the time sequence evolution law of the risk characteristics is analyzed through the long short-term memory network, the risk intensity development trend is captured, the risk propagation path and turning point are identified, and the short-term evolution direction and development trend of the risk are predicted;

[0122] According to the fault type classification identification model, the fault core information and the fault disposal demand in the risk evolution trend are identified;

[0123] According to the fault type classification identification model, the risk core information output by the intelligent identification storage network is analyzed, and the historical fault characteristics are matched, the risk spatial distribution characteristics, key node parameters and historical fault characteristics are calculated in multiple dimensions, the potential fault is matched, the specific type of fault, the root cause of fault and the fixed influence range are identified through the multi-layer classification identification network, according to the identification results of the specific type of fault and the root cause of fault, combined with the risk evolution trend, the targeted disposal strategy is generated, including: maintenance team skill requirement, maintenance equipment type, expected disposal time and safety measures;

[0124] The state and load priority of manpower and equipment in the operation and maintenance management system are obtained;

[0125] The real-time manpower resource state information in the operation and maintenance management system is obtained, including: on-duty state of maintenance team, professional skill qualification, current task state, location and expected response time, and the maintenance equipment state is obtained, and the load importance level and fault influence range of the loop network route area are calculated according to the multi-objective algorithm, and the power supply recovery urgency of different loads in the fault disposal process is evaluated;

[0126] According to the fault disposal demand, the state and load priority of manpower and equipment, a unified quantitative index is constructed, and a risk processing matrix is established;

[0127] According to the fault disposal demand, the time urgency coefficient is quantified, the time urgency coefficient is determined through the fault occurrence stage and the fault type, the state of manpower and equipment is quantified as technical complexity, the required professional skill and the special requirement of equipment are determined, the technical complexity and the time urgency coefficient are standardized evaluated, the availability index is established, the professional energy saving matching degree, the load priority and the power supply recovery urgency are calculated, the parameter weight is determined through the analytic hierarchy process, the priority score is generated, and then the risk processing matrix is constructed, wherein the horizontal axis represents the fault disposal urgency, and the vertical axis represents the resource optimization allocation level;

[0128] Based on the risk management matrix and operation and maintenance management system, differentiated allocation indicators are formulated, and the optimal allocation scheme is solved by genetic algorithm to autonomously allocate manpower and equipment;

[0129] The system obtains fault level information from the risk handling matrix and the available maintenance teams from the maintenance management system. Based on the fault level information, a differentiated allocation indicator system is established, and rapid response indicators are set. A response team is quickly assembled based on the fault level information, with complete operational equipment. Simultaneously, the maintenance team's professional skills must match at least 90%, and equipment availability must reach 100%. Subsequently, using real-number coding, the maintenance team, equipment resources, and task allocation are edited into individuals, and a fitness function is defined (…). In the formula, The optimization response efficiency, resource cost, and reliability of each individual are evaluated by setting the target weight coefficients and summing them to 1. Multi-objective optimization is introduced (where the initial weights are consistent with the target weights and the weights are adjusted through the adaptive strategy of Multi-ArmedBandit). The weight coefficients of each objective are calculated, and a detailed allocation scheme is generated to complete the work order autonomously.

[0130] Furthermore, using the intelligent identification storage network and fault type classification model as a local node, the model parameters of the intelligent identification storage network and fault type classification model are extracted, labeled as local network parameters and local identification parameters respectively, and uploaded to the central server. Specifically, this includes:

[0131] Extract the core identification data and feature fusion annotation rules from the intelligent identification storage network and mark them as local network parameters. The core identification data specifically includes: spatial feature encoding parameters, temporal evolution modeling parameters, and memory storage mechanism parameters.

[0132] Among them, the spatial feature encoding parameters include the weight matrix and bias vector of each convolutional layer in the convolutional neural network and the attention weight matrix used for spatial feature weighting in the attention mechanism; the temporal evolution modeling parameters include the weight matrix and bias term in the long short-term memory network, the hidden state initialization parameters of the recurrent neural unit, and the time window size and step size used to capture temporal dependencies; the memory storage mechanism parameters include the memory matrix used to store risk patterns and the weight coefficients and thresholds of memory updates and previous mechanisms; and the feature fusion annotation rules include the fusion layer weights and biases for fusing spatial features and temporal features, the normalization parameters used to achieve feature alignment and standardization, and the decision threshold and confidence interval.

[0133] Extract the core fault data and classification decision labeling rules from the fault type classification and identification model, and mark them as local identification parameters. The core fault data specifically includes: static feature extraction parameters, dynamic sequence modeling parameters, and knowledge evolution mechanism parameters.

[0134] The static feature extraction parameters include the convolution kernel weight and bias term for extracting features from the fault waveform and device parameters, the scaling factor and offset in the normalization layer, and the channel attention weight matrix for feature channel weighting. The dynamic sequence modeling parameters include the gating unit weight and state parameter for analyzing the alarm timing correlation, the graph neural network neighborhood aggregation weight for capturing the alarm propagation path, and the position encoding parameter of the time series feature encoder. The knowledge evolution mechanism parameters include the decision hyperplane offset vector for adjusting the classification boundary, the forgetting gate and memory reinforcement coefficient for fusing new and old knowledge, and the confidence calibration parameter for model uncertainty estimation. The classification decision labeling rules include the classification decision boundary function for different fault types, the threshold vector and confidence interval of the multi-level classifier, the logic rule and priority weight for compound fault judgment, and the scoring basis and weight allocation for fault severity level evaluation.

[0135] The local network parameters and local recognition parameters are preprocessed, a parameter verification mechanism is established, and the parameter abnormal distribution points are identified and processed according to the outlier detection algorithm. After processing, the uploaded parameters are marked and uploaded to the central server.

[0136] A parameter verification mechanism is established, and according to the statistical distribution of the outlier detection algorithm, the abnormal points in the parameter distribution are identified and removed, and the consistency of the parameter dimension is determined to ensure the shape matching of the parameter level. At the same time, the parameters are detected to prevent parameter distortion. Then, the importance of the parameters is evaluated, and the parameters are compressed. The uploaded parameters are encrypted by the SM4 algorithm, and the uploaded parameters are uploaded through a virtual private network.

[0137] Further, according to the attention mechanism, multiple sets of local network parameters and local recognition parameters are fused, and corresponding models are established according to the global convolutional neural network-long short-term memory network hybrid model. The local nodes are updated according to the central server data, which specifically includes:

[0138] According to the attention mechanism, the attention weight of multiple sets of uploaded parameters is calculated, and the results are marked as initial training model parameters and unified classification parameters, respectively.

[0139] The uploaded parameters of different local nodes are uniformly encoded, a parameter feature matrix is constructed, and the spatial distribution features, time sequence evolution patterns, and importance scores of the uploaded parameters are extracted as multi-dimensional spatial features. The relevance of different uploaded parameters is quantified, and according to the head attention mechanism, spatial, time sequence, importance, and consistency attention heads are calculated. According to the calculation results, a unified classification parameter system is constructed, the classification decision core rules are extracted, and then the data is divided into initial training model parameters and unified classification parameters according to the uploaded parameters of different models.

[0140] According to the global convolutional neural network-long short-term memory network hybrid model and the initial training model parameters, the unified classification parameters, a global intelligent recognition storage network and a global fault type classification recognition model are established.

[0141] According to the initial training model parameters, a hierarchical global network is constructed, spatial features and time sequence evolution features are extracted, and a regional parameter and global parameter mapping relationship is established, wherein the mapping function is determined by least squares fitting of historical data, the global standard coding rules are specified according to the unified classification parameters, the regional self-defined coding gap is eliminated, and the global standard distribution is converted through Z-score standardization to ensure the consistency of the classification features. Then, a global unified parameter library is established: a basic feature library for storing global standardized risk static features, a classification coding library for storing global unified fault type codes, and a resource association library for storing global standardized operation and maintenance resource parameters. According to the global convolutional neural network-long short-term memory network hybrid model and the initial training model parameters, a global intelligent recognition storage network is constructed for realizing global ring network risk unified recognition and global call of historical features. According to the global convolutional neural network-long short-term memory network hybrid model and the unified classification parameters, a global fault type classification recognition model is established, which is compatible with different regional fault feature differences and realizes global unified fault classification and disposal requirements.

[0142] According to the global intelligent recognition storage network and the global fault type classification recognition model, the global parameters are issued to each local node after verification.

[0143] The global intelligent recognition storage network and the global fault type classification recognition model are verified by new cases. When the global intelligent recognition storage network recognition accuracy is above 97%, the core parameters of the global intelligent recognition storage network are extracted. When the classification accuracy of the global fault type classification recognition model is above 98%, the core parameters of the global fault type classification recognition model are extracted. The core parameters are regionally adapted according to the environmental parameters and equipment type parameters. The local adjustment interface is reserved according to the environmental parameters and equipment type parameters. Different local nodes are adapted. Then, the core parameters are quantitatively compressed. The global parameter triggering conditions are: global model verification, global parameter version update, and local node active request. The information is stored through encryption storage and access control, and the operation log is reserved.

[0144] Further, referring to Figure 4 the power transmission line fault diagnosis and operation and maintenance scheduling system based on networking learning is proposed, which is used to realize the above method, including:

[0145] The data acquisition module is used to acquire electrical quantity data, equipment operation state data, ring network topology and time sequence data, and historical fault data and alarm log data corresponding to the power transmission line ring network.

[0146] The chart establishment module is used for extracting features corresponding to different data in the power transmission line loop network, and establishing a high-risk equipment area, visualizing a high-risk area heat map and visualizing a risk classification map according to the corresponding features;

[0147] The model training module is used for establishing an intelligent identification storage network according to the visualized risk classification map, the loop network topology, the time sequence data and the convolutional neural network-long short-term memory network hybrid model, and establishing a fault type classification identification model according to the historical fault data and the alarm log data combined with the convolutional neural network-long short-term memory network hybrid model;

[0148] The operation and maintenance scheduling module is used for identifying output parameters of the intelligent identification storage network and the fault type classification identification model, and constructing a risk processing matrix and a differentiated allocation index according to the output parameters and an operation and maintenance management system, and autonomously allocating operation and maintenance teams and resources;

[0149] The networking learning module is used for extracting local network parameters and local identification parameters in the intelligent identification storage network and the fault type classification identification model in the local node, preprocessing the local network parameters and the local identification parameters, and uploading the local network parameters and the local identification parameters to a central server;

[0150] The global collaboration module is used for performing attention weight calculation on uploaded parameters of multiple local nodes according to an attention mechanism, respectively substituting initial training model parameters and unified classification parameters into a global convolutional neural network-long short-term memory network hybrid model, and then establishing a global intelligent identification storage network and a global fault type classification identification model, and downloading the global intelligent identification storage network and the global fault type classification identification model parameters after verification to the local node.

[0151] The advantages of the present application are that: through deep fusion of PMU phasor data, GIS maps and historical data multi-source information, through a convolutional neural network-long short-term memory network hybrid model and an attention mechanism, precise extraction and joint analysis of fault space-time features are realized, complex fault types are accurately identified and classified, early warning and high-speed accurate diagnosis of loop network faults are realized, the timeliness of fault discovery and the reliability of diagnosis are improved, and through a networking learning mechanism, each local node can share knowledge and evolve collaboratively, through aggregation and downloading of global parameters, the model has a continuous self-optimization ability, combined with an intelligent risk processing matrix and a differentiated resource allocation strategy, accurate scheduling of repair teams and operation and maintenance resources is realized, operation and maintenance efficiency and resource utilization are significantly improved, the accuracy and timeliness of loop network fault diagnosis are greatly improved, and high-speed response and automatic closed loop of fault discovery, identification and processing are realized.

[0152] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above-mentioned embodiments, and the above-mentioned embodiments and descriptions in the specification are only the principles of the present application. Various changes and improvements can be made without departing from the spirit and scope of the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A power transmission line fault diagnosis and operation scheduling method based on networked learning, characterized in that, The method comprises the following steps: Obtain electrical quantity data, device operation state data, operation and maintenance management system, ring network topology and timing data, historical fault data and alarm log data corresponding to the ring network of the power transmission line; According to the PMU phasor data in the electrical quantity data, obtain the power grid fault dynamic characteristics, and according to the power grid fault dynamic characteristics, the device operation state data and the ring network topology and timing data, establish a high-risk device area; According to the spatial and temporal characteristics of the GIS map and the PMU phasor data in the electrical quantity data, combine the high-risk device area to generate a visual high-risk area heat map; According to the PMU phasor data and the visual high-risk area heat map, establish a hierarchical index, and establish a visual risk classification map; According to the visual risk classification map, the ring network topology and timing data, and the convolutional neural network-long short-term memory network hybrid model, establish an intelligent identification and storage network; According to the historical fault data and alarm log data, combine the convolutional neural network-long short-term memory network hybrid model to establish a fault type classification identification model; According to the output results of the intelligent identification and storage network and the fault type classification identification model, and the operation and maintenance management system, autonomously allocate a maintenance team and resources; Taking the intelligent identification and storage network and the fault type classification identification model as local nodes, extract the model parameters of the intelligent identification and storage network and the fault type classification identification model, respectively mark as local network parameters and local identification parameters, and upload to a central server; According to the attention mechanism, fuse the local network parameters and the local identification parameters, establish a corresponding global model according to the global convolutional neural network-long short-term memory network hybrid model, and update the local nodes according to the central server data.

2. The method of claim 1, wherein, According to the PMU phasor data in the electrical quantity data, obtain the power grid fault dynamic characteristics, and according to the power grid fault dynamic characteristics, the device operation state and the ring network topology and timing data, establish a high-risk device area, which specifically comprises: Extract the node electrical impact parameters in the power grid fault dynamic characteristics, normalize and weight the electrical impact parameters according to the safe operation standard of the power system, and obtain the node impact severity; Extract the device real-time operation data in the device operation state data, evaluate the device health degree, and extract the power transmission network topology graph in the ring network topology and timing data; According to the risk decision matrix of the node impact severity and the device health degree, cross operate the matrix elements and quantify the comprehensive risk level of each node, and construct a risk quantization data set; According to the risk quantization data set output by the risk decision matrix and the power transmission network topology graph, use the DBSCAN algorithm based on density clustering to establish a high-risk device area.

3. The method of claim 2, wherein, According to the spatial and temporal characteristics of the GIS map and the PMU phasor data in the electrical quantity data, combine the high-risk device area to generate a visual high-risk area heat map, which specifically comprises: Obtain the terrain trend and tower line trend in the GIS map, and construct a ring network space base map; According to the ring network space base map and the ring network operation and maintenance precision requirement, finely divide the ring network space base map, and mark as a spatial risk unit; The time dimension, the space dimension and the data dimension in the PMU phasor data are acquired, mapping processing is performed, a PMU risk matrix is generated, the PMU risk matrix is mapped to a looped network space base map, and a primary visualized map is established; A power transmission line map and a device operation state in a real-time high-risk device area are acquired, the power transmission line map and the device operation state are mapped to the primary visualized map, and a visualized risk area heat map is established.

4. The method of claim 3, wherein, The hierarchical indexes are established according to the PMU phasor data and the visualized high-risk area heat map, and a visualized risk classification map is established, specifically including: Obtaining multi-dimensional fault features in the PMU phasor data, establishing a fault feature severity grading index according to a power system safe operation standard and historical fault data, and mapping the multi-dimensional fault features processed by standardization to the fault feature severity grading index according to range normalization Interval; According to the fault feature severity hierarchical index and the K-means clustering algorithm, the multi-dimensional fault features after standardization processing are classified, the number of clusters is set to 3, the multi-dimensional fault features after standardization processing are distributed to the corresponding risk level clusters, and are marked as a fault feature hierarchical matrix; The device operation state data in the visualized high-risk area heat map are acquired, the device risk data are classified according to the three quantile method, and a device risk data hierarchical matrix is constructed; The fault feature hierarchical matrix and the device risk data hierarchical matrix are data-aligned, a multi-dimensional fusion data set is generated, the multi-dimensional fusion data set is mapped to the looped network space base map, and a visualized risk classification map is established according to color coding.

5. The method of claim 4, wherein, The intelligent recognition storage network is established according to the visualized risk classification map, the looped network topology and the time sequence data and the convolutional neural network-long short-term memory network hybrid model, specifically including: The spatial distribution features of each node in the visualized risk classification map are acquired, the visualized risk classification map is converted into a three-dimensional feature tensor, the three-dimensional feature tensor is input into the convolutional neural network for multi-level spatial feature extraction and output of a one-dimensional spatial feature vector containing global spatial risk distribution rules; The risk feature evolution rules in the looped network topology and the time sequence data are acquired, the dynamic topology features are organized into graph sequence data, the graph sequence data are input into the long short-term memory network for modeling and output of a one-dimensional time sequence vector for capturing long-term dependence and mutation features of risk evolution; According to the mutual attention mechanism, the one-dimensional spatial feature vector of the convolutional neural network and the one-dimensional time sequence vector of the long short-term memory network are calculated by weight association, and an intelligent recognition storage network is fused and established.

6. The method of claim 5, wherein, The fault type classification recognition model is established according to the historical fault data and the alarm log data in combination with the convolutional neural network-long short-term memory network hybrid model, specifically including: The fault static features and environmental parameters in the historical fault data are acquired, a fault static feature set is constructed, and the fault static feature set is input into the convolutional neural network for multi-level static feature extraction; The alarm log data are acquired, sorted according to timestamps, an alarm time sequence is constructed, and the alarm time sequence is input into the long short-term memory network for time sequence correlation analysis; According to the mutual attention mechanism, the attention weights of the fault static features and the alarm time sequence correlation are calculated, and are fused into multi-dimensional joint features, and a fault type classification recognition model is established according to the multi-dimensional joint features.

7. The method of claim 6, wherein, The intelligent identification storage network and the fault type classification identification model are taken as local nodes, model parameters of the intelligent identification storage network and the fault type classification identification model are extracted, and are respectively marked as local network parameters and local identification parameters, and are uploaded to a central server, specifically including: According to the intelligent identification storage network, the power transmission line loop network is identified, the risk core information and the risk evolution trend are obtained; According to the fault type classification identification model, the fault core information and the fault disposal demand in the risk core information and the risk evolution trend are identified; The state and load priority of manpower and equipment in the operation and maintenance management system are obtained; According to the fault disposal demand, the state and load priority of manpower and equipment, a unified quantitative index is constructed, and a risk processing matrix is established; According to the risk processing matrix and the operation and maintenance management system, a differentiated allocation index is formulated, an optimal allocation scheme is solved according to a genetic algorithm, and manpower and equipment are autonomously allocated.

8. The method of claim 7, wherein, The intelligent identification storage network and the fault type classification identification model are taken as local nodes, model parameters of the intelligent identification storage network and the fault type classification identification model are extracted, and are respectively marked as local network parameters and local identification parameters, and are uploaded to a central server, specifically including: The core identification data and feature fusion annotation rules in the intelligent identification storage network are extracted, and are marked as local network parameters, the core identification data specifically includes: spatial feature coding parameters, time sequence evolution modeling parameters and memory storage mechanism parameters; The core fault data and classification decision annotation rules in the fault type classification identification model are extracted, and are marked as local identification parameters, the core fault data specifically includes: static feature extraction parameters, dynamic sequence modeling parameters and knowledge evolution mechanism parameters; The local network parameters and the local identification parameters are preprocessed, a parameter verification mechanism is established, parameter abnormal distribution points are identified and processed according to an outlier detection algorithm, and the processed parameters are marked as uploaded parameters and uploaded to the central server.

9. The method of claim 8, wherein, According to the attention mechanism, a plurality of sets of local network parameters and local identification parameters are fused, corresponding models are established according to a global convolutional neural network-long short-term memory network hybrid model, and the local nodes are updated according to the data of the central server, specifically including: According to the attention mechanism, attention weight calculation is performed on a plurality of sets of uploaded parameters, and the results are respectively marked as initial training model parameters and unified classification parameters; According to the global convolutional neural network-long short-term memory network hybrid model and the initial training model parameters and the unified classification parameters, a global intelligent identification storage network and a global fault type classification identification model are established; According to the global parameters of the global intelligent identification storage network and the global fault type classification identification model that pass the verification, the global parameters are distributed to each local node.

10. A power transmission line fault diagnosis and operation scheduling system based on networked learning, for implementing the method of any one of claims 1-9, characterized in that, It includes: The data acquisition module is used for acquiring the corresponding electrical quantity data, equipment operation state data, loop network topology and time sequence data, historical fault data and alarm log data of the power transmission line loop network; The chart establishing module is used for extracting the features corresponding to different data in the power transmission line loop network, and establishing a high-risk equipment area, a visual high-risk area heat map and a visual risk classification map according to the corresponding features; The model training module is configured to establish an intelligent identification storage network according to a visual risk grading diagram, a ring network topology, time sequence data and a convolutional neural network-long short-term memory network hybrid model, and establish a fault type classification and identification model according to historical fault data and alarm log data combined with the convolutional neural network-long short-term memory network hybrid model. The operation and maintenance scheduling module is configured to identify output parameters of the intelligent identification storage network and the fault type classification and identification model, construct a risk processing matrix and a differentiated allocation index according to the output parameters and an operation and maintenance management system, and autonomously allocate operation and maintenance teams and resources. The networked learning module is configured to extract local network parameters and local identification parameters in the intelligent identification storage network and the fault type classification and identification model in the local node, pre-process the local network parameters and the local identification parameters, and upload them to a central server. The global collaboration module is configured to perform attention weight calculation on uploaded parameters of multiple local nodes according to an attention mechanism, respectively input initial training model parameters and unified classification parameters calculated into a global convolutional neural network-long short-term memory network hybrid model, and then establish a global intelligent identification storage network and a global fault type classification and identification model, and finally issue the global intelligent identification storage network and the global fault type classification and identification model parameters that pass the verification to the local node.

Images