Power grid fault online identification system and method based on big data

Abstract

The present invention discloses an online identification system and method for power grid faults based on big data, which relates to the technical field of power systems, including: collecting multi-source data of power grid systems through sensors; filtering data through a hierarchical storage strategy, building a data lake in combination with a multi-dimensional data index, and storing multi-source data in a classified manner; aggregating and converting data in the data lake using a real-time stream processing framework, extracting high-dimensional features related to faults through wavelet transform combined with principal component analysis; based on high-dimensional features, using a graph neural network combined with a deep learning model to identify the real-time state of the power grid; and locating faults through a time difference method combined with a graph theory algorithm based on the recognition result of the real-time state combined with the topology of the power grid. The online identification system for power grid faults based on big data provided by the present invention can efficiently manage and store a large amount of multi-source data in the power grid, and improve the speed of data processing and access.

Description

Technical Field

[0001] The present invention relates to the technical field of power systems, and in particular to a system and method for online identification of power grid faults based on big data. Background Art

[0002] With the continuous development and expansion of power grids, the diagnosis and location of power system faults have become increasingly complex. At present, online identification technology for power grid faults mainly relies on traditional monitoring systems and rule-based fault detection methods. These methods usually rely on the collection and manual analysis of sensor data, with relatively simple processing methods and difficulty in adapting to the complex operating conditions and large-scale grid topologies in the power grid.

[0003] In the prior art, fault diagnosis usually relies on model-based detection methods or statistical analysis methods. For example, by monitoring parameters such as current, voltage, and frequency, simple threshold judgment or rule-based judgment is used to perform fault diagnosis. However, the limitations of these methods are mainly manifested in the following aspects:

[0004] Traditional methods mostly rely on limited data sources and have weak real-time data processing capabilities, making it difficult to cope with the dynamic changes of emergencies and complex fault types in the power grid.

[0005] The power grid system involves a large amount of sensor data. Existing fault detection methods cannot efficiently process and extract high-dimensional features in the data, resulting in insufficient ability to identify complex faults.

[0006] Existing fault location methods are mostly based on static analysis of power grid topology, ignoring the dynamic characteristics of the power grid and unable to accurately locate the fault point in a complex power grid, resulting in slow power grid recovery and affecting the reliability of power supply.

[0007] Most existing technologies still rely on manual intervention or simple algorithm models, lack intelligent adaptability and optimization capabilities to the operating status of the power grid, and are unable to perform real-time adjustments and optimizations in complex environments. Summary of the invention

[0008] In view of the above-mentioned problems, the present invention is proposed.

[0009] Therefore, the technical problems solved by the present invention are: first, the existing technology cannot efficiently process and store multi-source, large-scale, high-dimensional data in the power grid, resulting in the accuracy and real-time performance of fault detection being limited. Secondly, the existing technology lacks in-depth mining of the complex topology and spatial characteristics of the power grid in terms of feature extraction and fault diagnosis, and cannot fully utilize the dynamic state information of the power grid for accurate diagnosis. Finally, the fault location accuracy is low, and the specific fault point cannot be quickly and accurately located, thus affecting the rapid recovery and stable operation of the power grid.

[0010] In order to solve the above technical problems, the present invention provides the following technical solutions: an online identification system for power grid faults based on big data, comprising:

[0011] The data acquisition module is used to collect multi-source data of the power grid system through sensors; the data storage module is used to filter data through a hierarchical storage strategy, build a data lake in combination with multi-dimensional data indexes, and classify and store multi-source data; the feature extraction module is used to aggregate and transform data in the data lake using a real-time stream processing framework, and extract high-dimensional features related to faults through wavelet transform combined with principal component analysis; the fault diagnosis module is used to identify the real-time status of the power grid based on high-dimensional features, using graph neural networks combined with deep learning models; the fault location module is used to locate faults based on the recognition results of the real-time status combined with the grid topology structure through the time difference method combined with graph theory algorithms.

[0012] As a preferred solution of the online identification system of power grid faults based on big data described in the present invention, the hierarchical storage strategy includes: using a hierarchical storage algorithm to calculate priority scores, real-time evaluation and allocation of multi-source data to the first storage layer and the second storage layer; With prediction priority threshold For comparison, when When , it is stored in the first storage layer. When the data is stored in the second storage layer, the multi-dimensional data index is set based on time, space and device type.

[0013] As a preferred solution of the online identification system for power grid faults based on big data described in the present invention, wherein: the hierarchical storage algorithm is used to calculate the priority score, including collecting data from multiple data sources of the power grid According to predefined categories Classification is performed, and the categories include voltage data, current data, frequency data, phase data and environmental parameter data; for each category , from the classified data Extract key feature vectors from , expressed as,

[0014] ,

[0015] in, Indicates the real-time rating. represents the volatility score, represents the frequency of outliers, Indicates the importance score of data sources;

[0016] Based on the data Historical access frequency data For each category , using the adaptive time series prediction model to predict the access frequency at future time t ;

[0017] According to the extracted key feature vector and connectivity indicators in power grid topology , calculate the data Importance score , expressed as,

[0018] ,

[0019] in, Representation data The importance score of represents the data category set, k is the category index, indicating the category to which the data belongs, and m is the feature index, indicating the extracted feature score. Representation characteristics The weight coefficient of Representation data The score of the mth key feature under category k, represents the weight coefficient of the connectivity index, Representation data Node Connectivity indicators;

[0020] By monitoring the operation status of the power grid and the external environment in real time, the clustering-based DBSCAN algorithm is used to detect and identify abnormal events. , based on the detected abnormal events, Importance score Perform dynamic adjustment, the adjustment formula is:

[0021] ,

[0022] in, represents the adjusted importance score, is the data for the jth event The adjustment factor for importance, is the jth detected abnormal event, and p is the number of detected events;

[0023] Combined with the predicted visit frequency and the adjusted importance score , calculate the data The overall priority score , expressed as,

[0024] ,

[0025] in, Representation data The priority score of and is the weight coefficient used to balance access frequency and importance, and is the exponential coefficient.

[0026] As a preferred solution of the online identification system for power grid faults based on big data described in the present invention, wherein: the adaptive time series prediction model includes using a moving average method to identify the long-term trend of access frequency and perform trend analysis;

[0027] Wavelet transform is used to extract the periodic pattern of access frequency and seasonal analysis is performed; the autoregressive integrated moving average model is used to predict future access frequency by combining the results of trend analysis and seasonal analysis. .

[0028] As a preferred solution of the online identification system for power grid faults based on big data described in the present invention, the trend analysis is expressed as:

[0029] ,

[0030] in, is the trend smoothing coefficient, , is the trend value at time t-1, is the access frequency at time t-1;

[0031] The seasonal analysis is expressed as,

[0032] ,

[0033] in, is the jth wavelet coefficient, is the jth wavelet basis function, is the number of decomposition levels of wavelet transform;

[0034] The future access frequency is expressed as,

[0035] ,

[0036] in, is the autoregression coefficient, is the moving average coefficient, is the predicted access frequency of category k at time t-p, is the prediction error of category k at time t-q, P and Q are the orders of the autoregressive and moving average models respectively.

[0037] As a preferred solution of the online identification system for power grid faults based on big data described in the present invention, the connectivity index is expressed as:

[0038] ,

[0039] in, represents the set of all nodes in the power grid, represents the total number of shortest paths between node s and node t, It represents the number of paths passing through node i in the shortest path between node s and node t.

[0040] As a preferred solution of the online identification system of power grid faults based on big data described in the present invention, the aggregation and conversion includes real-time aggregation of multi-source data from the first storage layer and the second storage layer according to the three dimensions of time, space and equipment type; the aggregated data are standardized respectively to eliminate the differences between different data sources and dimensions; an adaptive wavelet transform method is used to dynamically select the optimal wavelet basis function according to the data characteristics of each storage layer; multi-layer wavelet decomposition is performed on the data of each storage layer to extract approximate coefficients and detail coefficients, and seasonal components are reconstructed by accumulating detail coefficients; the reconstructed high-frequency and low-frequency seasonal components are respectively organized into high-frequency feature matrices and low-frequency feature matrices, and a comprehensive feature matrix is ​​formed by feature splicing; principal component analysis is used to reduce the dimension of the comprehensive feature matrix and extract key features; and features are selected and optimized by an optimization algorithm.

[0041] As a preferred solution of the big data-based online identification system for power grid faults described in the present invention, the identification of the real-time status of the power grid includes normalizing the high-dimensional feature vector to generate a feature set; using a multi-layer graph convolutional network including residual connections and a multi-head self-attention mechanism to extract spatial features in the grid topology structure, and generating global graph features through a graph pooling layer; and outputting the probability distribution of various types of faults through a fully connected layer and an improved Softmax function.

[0042] As a preferred solution of the big data-based online identification system for power grid faults described in the present invention, the fault location includes:

[0043] Receive the real-time fault status identification results of the fault diagnosis module, including the fault type and preliminary location; construct a power grid topology map containing node attributes; based on the fault signal propagation model, calculate the time difference of the fault signal reaching each measurement point to preliminarily locate the fault area; the fault status input interface provides preliminary positioning results, the power grid topology model provides spatial information, the time difference calculation module preliminarily locates the fault area, the graph theory algorithm module accurately locates the fault point, and the positioning result output interface feeds back the final result.

[0044] Another object of the present invention is to provide an online identification method for power grid faults based on big data, which solves the shortcomings of the prior art in terms of fault diagnosis accuracy, real-time performance and positioning accuracy, and effectively improves the intelligent diagnosis and precise positioning capabilities of power grid faults.

[0045] To solve the above technical problems, the present invention provides the following technical solutions: an online identification method for power grid faults based on big data, comprising: collecting multi-source data of the power grid system through sensors; building a data lake through a hierarchical storage strategy and multi-dimensional data index, and classifying and storing multi-source data; using a real-time stream processing framework to filter, aggregate and convert data, and extracting high-dimensional features related to faults through wavelet transform combined with principal component analysis; based on high-dimensional features, using a graph neural network combined with a deep learning model to identify the real-time status of the power grid; according to the recognition result of the real-time status combined with the grid topology, the fault is located through the time difference method combined with a graph theory algorithm.

[0046] Beneficial effects of the present invention: The online identification system for power grid faults based on big data provided by the present invention adopts a hierarchical storage strategy and multi-dimensional data index. The present invention can efficiently manage and store a large amount of multi-source data in the power grid, improve the speed of data processing and access, and solve the bottleneck problem of traditional methods in data processing and storage. By combining wavelet transform with principal component analysis, the present invention achieves efficient dimensionality reduction and optimization in feature extraction, improves the accuracy of fault diagnosis, and can extract key features of power grid faults.

[0047] In terms of fault diagnosis, the present invention combines graph neural networks with deep learning models to deeply explore the spatial characteristics of the power grid topology and operating status, automatically learn the complex relationships between nodes in the power grid, and thus achieve efficient fault identification. In terms of fault location, the present invention combines the time difference method and graph theory algorithm, and through accurate signal propagation models and power grid topology information, it greatly improves the accuracy of fault location, can accurately and quickly locate the fault point, and shorten the power grid recovery time.

[0048] In addition, the fault diagnosis and positioning system of the present invention has dynamic optimization capabilities and can make intelligent adjustments according to the real-time operating status of the power grid, thereby improving the robustness and adaptability of the system, especially when facing a complex power grid environment, and can handle faults more efficiently. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0050] Figure 1 An overall structural diagram of an online identification system for power grid faults based on big data provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0051] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0052] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0053] Embodiment 1:

[0054] Reference Figure 1 , which is an embodiment of the present invention, provides an online identification system for power grid faults based on big data, including:

[0055] A data acquisition module, used for collecting multi-source data of the power grid system through sensors;

[0056] The data storage module is used to filter data through a hierarchical storage strategy, build a data lake in combination with multi-dimensional data indexes, and classify and store multi-source data;

[0057] The feature extraction module is used to aggregate and transform the data in the data lake using a real-time stream processing framework, and extract high-dimensional features related to faults through wavelet transform combined with principal component analysis;

[0058] Fault diagnosis module, which is used to identify the real-time status of the power grid based on high-dimensional features using graph neural networks combined with deep learning models;

[0059] The fault location module is used to locate the fault based on the recognition result of the real-time status combined with the grid topology structure, through the time difference method combined with the graph theory algorithm.

[0060] The tiered storage strategy includes using a tiered storage algorithm to calculate priority scores, evaluate in real time, and allocate multi-source data to a first storage tier and a second storage tier;

[0061] Prioritize With prediction priority threshold For comparison, when When , it is stored in the first storage layer. When , store it to the second storage layer;

[0062] The multi-dimensional data index is set based on time, space and device type.

[0063] The hierarchical storage algorithm is used to calculate the priority score, including collecting data from multiple data sources of the power grid According to predefined categories Classification is performed, wherein the categories include voltage data, current data, frequency data, phase data and environmental parameter data;

[0064] For each category , from the classified data Extract key feature vectors from , expressed as,

[0065] ,

[0066] in, Indicates the real-time rating. represents the volatility score, represents the frequency of outliers, Indicates the importance score of data sources;

[0067] Based on the data Historical access frequency data For each category , using the adaptive time series prediction model to predict the access frequency at future time t ;

[0068] According to the extracted key feature vector and connectivity indicators in power grid topology , calculate the data Importance score , expressed as,

[0069] ,

[0070] in, Representation data The importance score of represents the data category set, k is the category index, indicating the category to which the data belongs, and m is the feature index, indicating the extracted feature score. Representation characteristics The weight coefficient of Representation data The score of the mth key feature under category k, represents the weight coefficient of the connectivity index, Representation data Node Connectivity indicators;

[0071] By monitoring the operation status of the power grid and the external environment in real time, the clustering-based DBSCAN algorithm is used to detect and identify abnormal events. , based on the detected abnormal events, Importance score Perform dynamic adjustment, the adjustment formula is:

[0072] ,

[0073] in, represents the adjusted importance score, is the data for the jth event The adjustment factor for importance, is the jth detected abnormal event, and p is the number of detected events;

[0074] Combined with the predicted visit frequency and the adjusted importance score , calculate the data The overall priority score , expressed as,

[0075] ,

[0076] in, Representation data The priority score of and is the weight coefficient used to balance access frequency and importance, and is the exponential coefficient.

[0077] The adaptive time series prediction model includes using a moving average method to identify the long-term trend of access frequency and perform trend analysis;

[0078] Wavelet transform was used to extract periodic patterns of visitation frequency and conduct seasonal analysis;

[0079] Combining the results of trend analysis and seasonal analysis, the autoregressive integrated moving average model is used to predict future visit frequency. .

[0080] The trend analysis is expressed as,

[0081] ,

[0082] in, is the trend smoothing coefficient, , is the trend value at time t-1, is the access frequency at time t-1;

[0083] The seasonal analysis is expressed as,

[0084] ,

[0085] in, is the jth wavelet coefficient, is the jth wavelet basis function, is the number of decomposition levels of wavelet transform;

[0086] The future access frequency is expressed as,

[0087] ,

[0088] in, is the autoregression coefficient, is the moving average coefficient, is the predicted access frequency of category k at time t-p, is the prediction error of category k at time t-q, P and Q are the orders of the autoregressive and moving average models respectively.

[0089] The connectivity index is expressed as,

[0090] ,

[0091] in, represents the set of all nodes in the power grid, represents the total number of shortest paths between node s and node t, It represents the number of paths passing through node i in the shortest path between node s and node t.

[0092] The aggregating and converting includes real-time aggregating multi-source data from the first storage layer and the second storage layer according to three dimensions of time, space and device type;

[0093] The aggregated data are standardized to eliminate the differences between different data sources and dimensions;

[0094] Adopting adaptive wavelet transform method, the optimal wavelet basis function is dynamically selected according to the data characteristics of each storage layer;

[0095] Perform multi-layer wavelet decomposition on the data of each storage layer, extract the approximate coefficients and detail coefficients, and reconstruct the seasonal components by accumulating the detail coefficients;

[0096] The reconstructed high-frequency and low-frequency seasonal components are organized into a high-frequency feature matrix and a low-frequency feature matrix, respectively, and a comprehensive feature matrix is ​​formed by feature splicing;

[0097] Apply principal component analysis to reduce the dimension of the comprehensive feature matrix and extract key features;

[0098] Features are selected and optimized through optimization algorithms.

[0099] The identifying the real-time state of the electric power grid includes normalizing the high-dimensional feature vector to generate a feature set;

[0100] A multi-layer graph convolutional network with residual connections and multi-head self-attention mechanism is used to extract spatial features in the power grid topology and generate global graph features through a graph pooling layer.

[0101] Through the fully connected layer and the improved Softmax function, the probability distribution of various faults is output.

[0102] The fault location includes receiving a real-time fault status identification result from a fault diagnosis module, including a fault type and a preliminary location;

[0103] Construct a power grid topology diagram including node attributes;

[0104] Based on the fault signal propagation model, the time difference of the fault signal reaching each measurement point is calculated to preliminarily locate the fault area;

[0105] The fault status input interface provides preliminary positioning results, the power grid topology model provides spatial information, the time difference calculation module preliminarily locates the fault area, the graph theory algorithm module accurately locates the fault point, and the positioning result output interface feeds back the final result.

[0106] Specifically, the collected data from multiple data sources of the power grid According to predefined categories Classification is performed, and the categories at least include voltage data, current data, frequency data, phase data and environmental parameter data.

[0107] Data category collection , defined as including all data types that need to be monitored and analyzed, such as: Voltage data: including the voltage amplitude and phase angle of each node. Current data: including the current intensity of each line. Frequency data: including the frequency fluctuation of the system. Phase data: including the phase information of each node of the power system. Environmental parameter data: including external environmental factors such as temperature and humidity.

[0108] For each category , from the classified data Extract key feature vectors from , including real-time score, volatility score, outlier frequency and data source importance score.

[0109] Real-time scoring Reflects the frequency of data collection and update rate. The real-time score is intended to measure the real-time update capability of data. The higher the score, the more frequent the data update and the stronger the real-time performance. The specific calculation method is as follows:

[0110] ,

[0111] in, is the time window length (e.g., 1 minute), It represents the data update frequency of category k in time window t, and the calculation formula is:

[0112] ,

[0113] Data collection frequency calculation statistics class k in the time window Number of data updates within .

[0114] Real-time score calculation:

[0115] ,

[0116] Volatility Score The standard deviation or variance of the data values ​​is used to measure the stability of the data. The volatility score reflects the stability of the data by calculating the standard deviation or variance of the data within a certain time window. The greater the volatility, the higher the score, which may indicate potential failures.

[0117] Calculation formula:

[0118] ,

[0119] in, It represents the standard deviation of the data of category k in time window t, and the calculation formula is:

[0120] ,

[0121] ,

[0122] Collect data points of category k in time window t .

[0123] Mean calculation: ,

[0124] Standard deviation calculation: ,

[0125] Volatility score assignment: ,

[0126] Outlier frequency Reflects the number of abnormal fluctuations or outliers detected, and evaluates abnormal patterns in the data. The outlier frequency score evaluates the frequency of abnormal fluctuations in the data by counting the number of outliers detected within a certain time window. The more outliers, the higher the score, which may indicate a risk of failure.

[0127] Calculation formula:

[0128] ,

[0129] in, is an indicator function, which is 1 if the condition is met, otherwise it is 0. is the threshold coefficient, usually with a value of 2, indicating that values ​​exceeding two standard deviations from the mean are considered outliers.

[0130] Collect data points of category k in time window t .

[0131] The mean and standard deviation are calculated in the same way as the volatility score.

[0132] Outlier Detection:

[0133] ,

[0134] Outlier frequency calculation:

[0135] ,

[0136] Data source importance score The weight score reflects the data coming from key equipment (such as synchronized phasor measurement units (PMUs)) to measure the criticality of the data source. The data source importance score assigns different weights to different data sources by evaluating the criticality of the data source in the power grid topology. The data provided by key data sources (such as PMUs) is more important in fault identification and has a higher score.

[0137] Calculation formula: ;

[0138] in, is the importance weight of data source i, which is calculated as follows:

[0139] ,

[0140] ;

[0141] in, is the betweenness centrality of node i, indicating the importance of node i in all shortest paths, is the total number of shortest paths between node s and node t, is the number of paths occupied by node i in the shortest path between node s and node t, is the set of all nodes in the power grid, and M is the total number of nodes.

[0142] Representing the power grid as a graph ,in is a node set, is the edge set. For every pair of nodes , calculate the total number of shortest paths . Compute node i in The proportion of the shortest path . Add up all To get the betweenness centrality of node i .

[0143] Weight normalization: ,

[0144] Data source importance score assignment: ;

[0145] Based on data Historical access frequency data For each category , using the adaptive time series prediction model to predict the access frequency at future time t The forecasting model includes three sub-steps: trend analysis, seasonal analysis and forecast integration.

[0146] The trend analysis sub-step aims to identify the long-term trend of the access frequency data. The exponential smoothing method is used to calculate the long-term trend of the historical access frequency. . Calculation formula:

[0147] ,

[0148] in, is the trend smoothing coefficient, , is the access frequency data of category k at the previous time t-1, is the trend value of the previous moment.

[0149] Set initial trend value is the access frequency at the first moment .

[0150] Trend Update: ;

[0151] Iterative calculation: At each time t, the trend value is updated according to the above formula .

[0152] The seasonal analysis sub-step aims to extract the periodic patterns in the visit frequency data. Wavelet transform is used to decompose the historical visit frequency data and extract the periodic components. , calculation formula:

[0153] ,

[0154] in, is the jth wavelet coefficient, is the jth wavelet basis function, is the number of decomposition levels of wavelet transform.

[0155] Select a wavelet basis function that is suitable for the characteristics of the power system data, such as Daubechies wavelet or Haar wavelet.

[0156] Historical access frequency data Perform multi-layer wavelet decomposition to obtain approximate coefficients and detail factor , decomposition formula:

[0157] ,

[0158] Extract seasonal ingredients:

[0159] The detail factor Accumulation as seasonal component ,

[0160] ,

[0161] Reconstruct the periodic pattern of access frequency using the accumulated detail coefficients .

[0162] The forecast integration sub-step combines the results of trend analysis and seasonal analysis and uses the autoregressive moving average model (ARMA) to predict the future visit frequency. To make a prediction, the calculation formula is:

[0163] ,

[0164] in, is the autoregression coefficient, is the moving average coefficient, is the prediction error of category k at time t-q, P and Q are the orders of the autoregressive and moving average models respectively.

[0165] Determine the model order P and Q, usually through the autocorrelation function (ACF) and partial autocorrelation function (PACF) plots. Estimate the autoregressive coefficients and the moving average coefficient .

[0166] Based on historical forecast errors , adjust the predicted value and calculate the predicted value:

[0167] ,

[0168] At each time t, based on the latest trend, seasonality, and historical forecast errors, calculate the visit frequency forecast for the future time t .

[0169] Optimize the parameters of the ARMA model using least squares or maximum likelihood estimation methods and .

[0170] Using rolling window technology, model parameters are updated regularly to adapt to dynamic changes in data and improve prediction accuracy.

[0171] According to the extracted key feature vector and connectivity indicators in power grid topology , calculate the data Importance score , and its calculation formula is:

[0172] ,

[0173] in, For data The importance score of is a set of data categories, k is a category index, indicating the category to which the data belongs, and m is a feature index, indicating the extracted feature score (1 to 4). Features The weight coefficient of is dynamically determined by the feature importance evaluation method (such as feature importance scoring based on random forest), is the weight coefficient of the connectivity index, which is determined by the system optimization algorithm to balance the impact of connectivity on importance. For data Node The connectivity index of the power grid is calculated using the betweenness centrality measurement method of the power grid nodes.

[0174] Betweenness Centrality To measure the node The importance of power grid topology. The calculation formula is:

[0175] ,

[0176] in, represents the set of all nodes in the power grid, represents the total number of shortest paths between node s and node t, It represents the number of paths passing through node i in the shortest path between node s and node t.

[0177] Representing the power grid as a graph ,in is a node set, is the edge set.

[0178] Use Dijkstra algorithm to calculate all node pairs The total number of shortest paths .

[0179] For each node , calculate its betweenness centrality :

[0180] ,

[0181] Weight normalization:

[0182] ;

[0183] Among them, M is the total number of nodes.

[0184] The relative importance of each feature in fault identification is evaluated through the feature importance scoring method based on random forest, and the weight coefficient of each feature is dynamically determined. .

[0185] It is determined through optimization algorithms (such as genetic algorithms and particle swarm optimization) to balance the influence of connectivity indicators in the importance score, ensuring that the importance score considers both the characteristics of the data itself and reflects its criticality in the power grid topology.

[0186] By monitoring the operation status of the power grid and the external environment in real time, the clustering-based DBSCAN algorithm is used to detect and identify abnormal events. , and dynamically adjust the importance score of the data based on the detected abnormal events.

[0187] DBSCAN algorithm: density clustering algorithm, used to detect abnormal events in the power system, such as sudden load changes, voltage fluctuations, etc. The calculation formula is:

[0188] ,

[0189] in, represents the adjusted importance score, is the data for the jth event The adjustment coefficient of importance is dynamically determined through system learning methods. is the jth detected abnormal event.

[0190] Collect power grid operation data and external environment data in real time. Use DBSCAN algorithm to cluster real-time data and identify abnormal events with low density. . Determine the number of abnormal events p.

[0191] Based on the abnormal events detected , calculate the adjustment factor.

[0192] Calculate the adjusted importance score:

[0193] ,

[0194] Dynamically determine the impact coefficient of each event on the importance score through learning methods (such as online learning algorithms, reinforcement learning) , ensuring flexibility and adaptability of adjustments.

[0195] Combined with the predicted visit frequency and the adjusted importance score , calculate the data The overall priority score , used for subsequent storage tier allocation, calculation formula:

[0196] ,

[0197] in, Representation data The overall priority score of and Represents the weight coefficient used to balance access frequency and importance, which is dynamically determined by the optimization algorithm. and Represents an exponential coefficient, which is used for nonlinear adjustment to improve the stability and accuracy of the score.

[0198] The comprehensive priority score comprehensively evaluates the storage priority of data by combining access frequency and importance score. The introduction of exponential coefficient makes the scoring system more flexible and adaptable, and can effectively cope with the diversity and variability of data characteristics.

[0199] Access frequency prediction result acquisition: According to the access frequency prediction step, obtain the future access frequency prediction value of each category k .

[0200] Importance score acquisition: Get the adjusted importance score , comprehensive score calculation:

[0201] ,

[0202] Adjust the weight coefficients through optimization algorithms (such as genetic algorithms, particle swarm optimization) and And the exponential coefficient and , to minimize the fault identification error rate and storage resource consumption.

[0203] Iterative optimization to ensure comprehensive scoring accuracy and rationality.

[0204] Furthermore, the real-time stream processing framework is used to collect multi-source data from the first storage layer (high priority data) and the second storage layer (low priority data) in real time from the data lake. The collected data includes but is not limited to sensor data, equipment status information and historical operation records of the power system.

[0205] The input data source is the multi-source data in the first storage layer and the second storage layer; the processing tool is the real-time stream processing framework; the collection process is that the framework reads the data stream in the data lake in real time through the connector to ensure the timeliness and continuity of the data.

[0206] In order to improve data quality and reduce the impact of noise and outliers on subsequent analysis, moving average filtering or median filtering is used to smooth the data curve and reduce high-frequency noise, which can be expressed as:

[0207] ,

[0208] in, is the original data point, and n is the filter window radius.

[0209] Use the Z-score method to detect outliers, define a threshold (such as an absolute value greater than 3), and replace outliers with adjacent normal values ​​or delete them, as shown in:

[0210] ,

[0211] in, is the data mean, is the standard deviation of the data. , it is considered as an outlier.

[0212] Multi-dimensional data aggregation

[0213] After data collection and filtering, the data is aggregated in real time according to the three dimensions of time, space and device type in order to extract features of different levels and characteristics.

[0214] In time dimension aggregation, the first storage layer time window: uses a minute-level time window (such as 60 seconds).

[0215] The aggregation indicator is to calculate the average value of the data in each time window ,sum , maximum value and minimum value .

[0216] ,

[0217] ,

[0218] ,

[0219] ,

[0220] Where N is the number of data points in the time window.

[0221] Time window in the second storage layer: use hourly time window (such as 3600 seconds).

[0222] Aggregate indicator: also calculate the average value of data ,sum , maximum value and minimum value .

[0223] In spatial dimension aggregation, high priority data is aggregated:

[0224] Regional division: Data is divided into several regions or nodes based on the geographical location or topological nodes of the power grid.

[0225] Aggregation method: Aggregate the data in each region or node and calculate high-frequency change statistical indicators.

[0226] ,

[0227] Where M is the number of data points in a region or node, It is the average value after spatial dimension aggregation.

[0228] Low priority data aggregation: Regional division: The same geographical area or node division as high priority data.

[0229] Aggregation methods: Calculate low-frequency changes and long-term average statistical indicators.

[0230] ,

[0231] In the device type dimension aggregation, the classification standard is: classification by device type (such as transformer, circuit breaker, synchronous phasor measurement unit PMU, etc.).

[0232] High-priority data: Calculates high-frequency data statistics for each device type.

[0233] ,

[0234] Where K is the number of data points within the device type, It is the average value after aggregation by device type.

[0235] Low-priority data: Calculates low-frequency data statistics for each device type, and also calculates the average, total, maximum, and minimum values.

[0236] The aggregated data were standardized to eliminate the differences between different data sources and dimensions and ensure the consistency and comparability of the data in subsequent analysis, expressed as:

[0237] ,

[0238] in, is the original data point, is the data mean, is the standard deviation of the data.

[0239] For normalized high priority data and low priority data Perform frequency domain and time domain analysis to determine the main frequency components and time characteristics (such as mutations, periodic fluctuations, etc.) in the data.

[0240] The spectrum of the data is analyzed by fast Fourier transform (FFT) to identify the main frequency components; the time-frequency characteristics of the data are analyzed by wavelet packet analysis to determine the appropriate time and frequency resolution.

[0241] Build a base function library containing a variety of wavelet base functions to cover different frequency and time resolution requirements. The base function library includes but is not limited to: Daubechies6 (db6); Symlets4 (sym4); Coiflets2 (coif2); Wavelet Packet (WaveletPacket);

[0242] Performance evaluation indicators: Energy concentration, reconstruction error and noise suppression ability are defined as evaluation indicators.

[0243] Energy concentration: measures the concentration of energy of the wavelet basis function in the frequency domain, expressed as:

[0244] ,

[0245] in, is the wavelet coefficient, is the energy of the original data.

[0246] Reconstruction error: measures the accuracy of data reconstruction after wavelet transform, expressed as:

[0247] ,

[0248] Where D is the original data, To reconstruct the data.

[0249] Noise suppression ability: measures the effect of wavelet basis function in the denoising process, expressed as:

[0250] ,

[0251] in, is the denoised data.

[0252] Genetic algorithm is used to dynamically select the optimal wavelet basis function from the basis function library.

[0253] Randomly generate multiple combinations of wavelet basis functions.

[0254] The EC, RE, and NS of each combination are calculated to comprehensively evaluate its performance.

[0255] The combinations with higher fitness are selected to enter the next generation.

[0256] Generate new combinations through crossover and mutation.

[0257] The fitness evaluation and selection operations are repeated until a termination condition (such as a maximum number of iterations or a fitness threshold) is met.

[0258] The wavelet basis function with the highest fitness is selected as the optimal basis function.

[0259] First storage layer: Select a wavelet basis function (such as db6) that can efficiently capture fast changes and high-frequency components.

[0260] Second storage layer: Select wavelet basis functions (such as sym4) suitable for capturing long-term trends and low-frequency components.

[0261] The first storage layer (high priority data) decomposition layer number setting: set 6 layers of wavelet decomposition layers , to obtain higher time and frequency resolution and capture fast-changing fault signals.

[0262] Second storage layer (low priority data) decomposition layer number setting: set 3 layers of wavelet decomposition layers , to focus on long-term trends and data characteristics in a stable state.

[0263] The first storage layer data decomposition is expressed as:

[0264] ,

[0265] in, is the j-th layer approximation coefficient, representing the low-frequency part, is the j-th layer detail coefficient, representing the high frequency part, WaveletDecompose is the wavelet decomposition function, and Wavelet1 is the wavelet basis function selected for the first storage layer (such as db6).

[0266] The second storage layer data is decomposed as follows:

[0267] ,

[0268] in, is the j-th layer approximation coefficient, is the detail coefficient of the jth layer, and Wavelet2 is the wavelet basis function (such as sym4) selected for the second storage layer.

[0269] In seasonal component extraction, the first storage layer is expressed as:

[0270] ,

[0271] in, It is the seasonal component of the first storage layer, representing the high-frequency variation part.

[0272] The second storage layer is represented as:

[0273] ,

[0274] in, It is the seasonal component of the second storage layer, representing the low-frequency variation part.

[0275] Seasonal component reconstruction:

[0276] The first storage layer is represented as:

[0277] ;

[0278] in, is the reconstructed high-frequency seasonal component, is the reconstruction function of the j-th layer wavelet basis function.

[0279] The second storage layer is represented as:

[0280] ,

[0281] in, For the reconstructed low-frequency seasonal components, the data matrix is ​​constructed;

[0282] The reconstructed seasonal component of the first storage layer Organized into high-frequency feature matrix , where each line Represents a feature vector within a time window:

[0283] ,

[0284] Where N is the number of time windows, is the high-frequency feature vector in the i-th time window.

[0285] The reconstructed seasonal component of the second storage layer Organized into low-frequency feature matrix , where each line Represents a feature vector within a time window:

[0286] ,

[0287] in, is the low-frequency feature vector in the i-th time window.

[0288] Through the feature concatenation strategy, the high-frequency feature matrix and the low-frequency feature matrix Splice by column to form a comprehensive feature matrix :

[0289] ,

[0290] in, Each row of contains the comprehensive feature vector of high-frequency and low-frequency features.

[0291] Calculate the comprehensive feature matrix The covariance matrix of , expressed as:

[0292] ,

[0293] Where N is the number of samples (number of time windows), is the comprehensive feature matrix.

[0294] Calculate the covariance matrix The eigenvalues ​​and corresponding eigenvectors of are selected, and the first five eigenvectors with the largest eigenvalues ​​are selected to form the principal component basis matrix W, which is expressed as:

[0295] ,

[0296] in, is the eigenvector of the mth principal component.

[0297] The comprehensive feature matrix Projecting onto the principal component basis matrix W, we get the high-dimensional feature matrix F after dimensionality reduction, expressed as:

[0298] ,

[0299] Among them, each row of F Represents the high-dimensional feature vector of a sample after dimensionality reduction.

[0300] Combined with the fault identification requirements, the contribution of each principal component in fault identification is evaluated, and the principal component with the greatest impact on fault identification is selected to form a key feature vector. The specific methods include:

[0301] Contribution calculation: Calculate the variance contribution of each principal component and select the principal components whose cumulative contribution rate reaches more than 95%.

[0302] Threshold setting: Set the minimum threshold of the principal component variance contribution rate (such as 10%) and select the principal components that exceed this threshold.

[0303] Optimization algorithm applications include:

[0304] Genetic method steps: Encode the feature selection problem into a chromosome representation, where each chromosome represents a feature subset. Use binary encoding, where 1 indicates that the feature is selected and 0 indicates that it is not selected.

[0305] Define the fitness function, and evaluate the fitness of each chromosome based on the fault identification accuracy. The fitness function can be expressed as: Fitness=Accuracy;

[0306] Among them, Accuracy is the recognition accuracy of the fault recognition model based on the selected features.

[0307] Selection operation: Use roulette wheel selection or tournament selection method to select chromosomes with higher fitness to enter the next generation.

[0308] Single-point or double-point crossover methods are used to exchange some genes of chromosomes to generate a new generation of chromosomes.

[0309] Randomly mutate some gene loci of chromosomes to increase the diversity of the population.

[0310] Repeat the fitness evaluation, selection, crossover and mutation operations to iteratively optimize the feature selection scheme until the preset number of iterations or fitness threshold is reached.

[0311] The chromosome with the highest fitness is selected as the optimal feature subset.

[0312] Lasso regression steps:

[0313] The L1 regularization term is introduced into the regression model to promote sparse solutions and automatically select important features. The loss function of Lasso regression is defined as: ;

[0314] Among them, y is the target variable, X is the feature matrix, W is the regression coefficient, is the regularization parameter.

[0315] Optimize the regularization parameters through cross-validation method , ensuring the generalization ability of the model and the effectiveness of feature selection.

[0316] The features corresponding to the non-zero regression coefficients are automatically selected through Lasso regression to form the final feature subset.

[0317] Combine the optimized key features to form the final high-dimensional feature vector , used for subsequent fault identification and diagnosis.

[0318] ,

[0319] in, It is the final high-dimensional feature vector, which contains the key features selected through optimization.

[0320] Based on the data characteristics of the first and second storage layers, the following parameters are dynamically adjusted to optimize the feature extraction process:

[0321] Wavelet basis function selection: Dynamically select the optimal wavelet basis function based on the characteristics of the storage level data.

[0322] Wavelet decomposition layers: The first storage layer is set to 6 decomposition layers , the second storage layer sets 3 decomposition layers .

[0323] Number of PCA principal components: Set the number of principal components M=5 to ensure the extraction of main features.

[0324] Fault correlation enhancement

[0325] By combining storage level information, we ensure that the extracted high-frequency features are highly correlated with fast fault signals, and the low-frequency features are related to long-term trends and potential hazards, thereby improving the accuracy and real-time performance of the overall fault identification system.

[0326] It should be noted that the real-time stream processing framework is used to aggregate multi-source data in real time according to the three dimensions of time, space and device type, and perform standardized processing to ensure the consistency and comparability of the data.

[0327] The optimal basis function is dynamically selected from a variety of wavelet basis functions through a genetic algorithm, and the wavelet basis functions suitable for extracting high-frequency and low-frequency components are selected respectively according to the data characteristics of different storage layers.

[0328] Multi-layer wavelet decomposition is performed on data from different storage layers to extract approximate coefficients and detail coefficients, and high-frequency and low-frequency seasonal components are reconstructed by accumulating detail coefficients.

[0329] The high-frequency and low-frequency feature matrices are concatenated column by column to form a comprehensive feature matrix, and the dimension is reduced through principal component analysis (PCA), and the main components are extracted as high-dimensional feature vectors.

[0330] Optimization algorithms such as genetic algorithm and Lasso regression are used to perform further feature selection and weight adjustment on the principal components to ensure that the extracted high-dimensional features are highly relevant to fault identification.

[0331] Based on the data characteristics of different storage layers, the wavelet basis function selection, the number of wavelet decomposition layers and the number of PCA principal components are dynamically adjusted to ensure that the extracted high-dimensional features contain both the real-time change information of high-priority data and the long-term trend characteristics of low-priority data, thereby improving the accuracy and real-time performance of the fault identification system.

[0332] The fault diagnosis module receives the high-dimensional feature vector from the feature extraction module. .

[0333] For high-dimensional feature vectors Improved dynamic Z-score standardization is performed, combined with time window sliding technology to adapt to real-time data changes, expressed as:

[0334] ,

[0335] in, is the high-dimensional feature vector at time point t, For time window The mean vector in , For time window The standard deviation vector within, A tiny constant vector to prevent the denominator from being zero.

[0336] The feature expression capability is enhanced through polynomial expansion and feature interaction generation, including first-order features, second-order polynomial features, and the product of feature pairs, expressed as:

[0337] ,

[0338] in, represents the element-wise square of the eigenvector, represents the outer product of eigenvectors, generating the product of eigenpairs.

[0339] Receive the enhanced feature vector after processing . Build a graph model based on the topological structure of the power grid ,in It is a collection of nodes, representing each device or area in the power grid. It is a set of edges, representing the connection relationship between devices.

[0340] For each node Assign the corresponding node feature vector , expressed as:

[0341] ,

[0342] The multi-layer graph convolutional network uses a multi-layer graph convolutional network with residual connections and variable weights to enhance feature extraction capabilities, expressed as:

[0343] ,

[0344] in, For the The node feature matrix of the layer, is the adjacency matrix with self-loops added, is the identity matrix.

[0345] for The degree matrix of is defined as:

[0346] ,

[0347] in, For the The weight matrix of the layer, For the The residual connection weight matrix of the layer, is the activation function.

[0348] The multi-head self-attention mechanism is introduced to improve the model's ability to perceive the relationship between different nodes, which can be expressed as:

[0349] ,

[0350] ,

[0351] in, is the attention weight between node i and node j, a is the trainable attention parameter vector, represents the vector concatenation operation, is the set of neighbor nodes of node i, Node features after attention mechanism.

[0352] The improved DiffPool layer is used to generate the global graph feature representation Z through the hierarchical graph pooling method, which is expressed as:

[0353] ,

[0354] in, is the final layer node feature matrix of the graph neural network, and S is the allocation matrix, which determines how the nodes are aggregated.

[0355] Receive the global graph feature representation Z. Further extract and integrate features through a series of fully connected layers, expressed as:

[0356] ,

[0357] ;

[0358] in, and is the weight matrix of the fully connected layer, and is the bias vector of the fully connected layer, is the activation function.

[0359] Using the improved Softmax activation function, combined with the temperature parameter , output the probability distribution of various types of failures:

[0360] ,

[0361] in, and is the weight matrix and bias vector of the i-th category of the output layer, C is the number of fault categories, It is a temperature parameter used to adjust the smoothness of the Softmax function.

[0362] The cross entropy loss function is combined with the entropy regularization term, which is defined as follows and expressed as:

[0363] ,

[0364] in, is the real label, using One-Hot encoding, is the predicted probability, is the regularization parameter, and K is the number of all weight matrices in the model.

[0365] Using the Adam optimizer, by combining the momentum term and the adaptive learning rate, the loss function is minimized and the model parameters are updated, which is expressed as:

[0366] ,

[0367] in, are model parameters, is the learning rate, m and v are the first-order and second-order momentum estimates, respectively. and is the momentum decay rate.

[0368] Model training and validation

[0369] The model parameters are iteratively optimized on the training set, and the early stopping method is combined with the learning rate decay strategy to prevent overfitting.

[0370] Evaluate model performance on the validation set and adjust hyperparameters (such as learning rate, regularization coefficient, temperature parameter ) to improve recognition accuracy and generalization ability.

[0371] Through the above technical features, the fault diagnosis module can accurately identify the real-time operating status and fault type of the power grid based on high-dimensional feature vectors, using improved graph neural networks combined with deep learning models. The introduction of residual connections, multi-head self-attention mechanisms and improved Softmax functions enhances the model's perception of the power grid topology and improves the accuracy and robustness of fault identification.

[0372] Receive real-time fault status identification results from the fault diagnosis module, including the fault type and the initial location of the fault .

[0373] Constructing the grid topology diagram: Constructing the grid topology based on the actual topology of the power system ,in It is a collection of nodes, representing each device or area in the power grid. It is a set of edges, representing the connection relationship between devices.

[0374] For each node Assign attribute information, such as equipment type, geographic location, load capacity , Equipment Status Score etc., to assist fault location, expressed as:

[0375] ,

[0376] in, For Node A collection of attributes.

[0377] Input signal synchronization: Determine the time when the fault occurs based on the fault status identification result .

[0378] Fault signal propagation model: Establish a fault signal propagation model in the power grid, taking into account the signal propagation speed v and path impedance And line length , defining the signal propagation time To each measuring point , expressed as:

[0379] ,

[0380] in, Fault point With measuring point The straight-line distance between is the set of paths from the fault point to the measurement point, is the impedance of edge e in the path, is the length of edge e in the path, is the attenuation factor or other impedance-related adjustment factor of edge e in the path.

[0381] Compare the actual measured signal arrival time and predicted arrival time The Difference , expressed as:

[0382] ,

[0383] According to the time difference The minimum value of , expressed as:

[0384] ,

[0385] in, is the time difference value of node v.

[0386] Based on the time difference calculation results, determine the candidate area or node set where the failure may occur .

[0387] Using the Dijkstra algorithm, multiple weight factors (such as impedance , Line length , Equipment load , Equipment Status Score , signal attenuation factor ), calculated from the initial positioning point To the candidate node The weighted shortest path , expressed as:

[0388] ,

[0389] in, , , , are weight coefficients, which respectively reflect the importance of impedance, line length, equipment load and status, and signal attenuation in positioning. is the impedance of edge e, is the length of side e, is the load capacity of node i, Score the device status of node i, is the signal attenuation factor of edge e.

[0390] Accurately locate the fault point: Comprehensive weighted shortest path results and time difference calculation, using weighted scoring method to determine the specific node or device location where the fault occurred , expressed as:

[0391] ,

[0392] ,

[0393] in is the comprehensive weight of edge e, Candidate Node The time difference value, is the weight coefficient for adjusting the impact of time difference.

[0394] The final fault location result (specific node or device location) ) is fed back to the system’s monitoring center or user terminal for subsequent troubleshooting and maintenance.

[0395] According to the actual fault handling results, the parameters of the time difference method and graph theory algorithm (such as propagation speed v, weight coefficient , , , , adjustment coefficient ) to dynamically adjust and optimize to improve the accuracy of fault location.

[0396] Utilize historical fault data to train and update fault location models, and enhance the system's adaptability to complex grid topologies and diverse fault types. Specific methods include incremental learning and parameter optimization of machine learning models:

[0397] ,

[0398] in, are model parameters, is the learning rate, is the loss function.

[0399] Through the above technical features, the fault location module can accurately locate the fault location in the power grid based on the real-time fault status identification results, combined with the grid topology structure and the improved time difference method and graph theory algorithm. The introduction of multiple weight factors and weighted scoring method enhances the accuracy and adaptability of the location algorithm and improves the efficiency and accuracy of fault handling.

[0400] Embodiment 2:

[0401] One embodiment of the present invention provides a method for online identification of power grid faults based on big data, comprising:

[0402] Collect multi-source data of the power grid system through sensors;

[0403] Build a data lake through tiered storage strategy and multi-dimensional data indexing to store multi-source data in a classified manner;

[0404] A real-time stream processing framework is used to filter, aggregate and transform data, and high-dimensional features related to faults are extracted through wavelet transform combined with principal component analysis;

[0405] Based on high-dimensional features, we use graph neural networks combined with deep learning models to identify the real-time status of the power grid;

[0406] According to the recognition results of real-time status and the topology of the power grid, fault location is carried out by combining the time difference method with graph theory algorithm.

[0407] Embodiment 3:

[0408] An embodiment of the present invention is different from the first two embodiments in that:

[0409] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program codes.

[0410] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in conjunction with such instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in conjunction with such instruction execution systems, devices or apparatuses.

[0411] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering or, if necessary, processing in another suitable manner, and then stored in a computer memory.

[0412] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0413] Embodiment 4:

[0414] As an embodiment of the present invention, an online identification system for power grid faults based on big data is provided. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through simulation experiments.

[0415] The experiment was conducted in a virtual power grid with 1,000 nodes and 5,000 edges, simulating the real-time operation status of the power grid and adding different types of fault events. These fault events include common short circuits, open circuits, overloads, and equipment failures. The occurrence time and duration of each fault event are set to be known. Grid operation data, including voltage, current, frequency, phase, and environmental parameters, are collected by sensors and transmitted to the data storage system in real time. The data collection frequency is set to once every 5 seconds, and a total of 1 hour of data is collected, with a total data volume of 72,000. The hardware environment used in the experiment is a standard data center architecture, using computing nodes that support large-scale data processing, and equipped with a GPU-accelerated graph neural network model training and inference platform.

[0416] Data preprocessing includes removing invalid data and outliers, and standardizing the data to eliminate differences between different data sources. After the data is aggregated and converted by the real-time stream processing framework, high-dimensional features related to the fault are extracted through wavelet transform and principal component analysis (PCA). Based on these features, the graph neural network (GNN) of the present invention is combined with a deep learning model for fault diagnosis, while the fault location is combined with the topological structure of the power grid, and the time difference method and graph theory algorithm are used to further accurately locate the fault point.

[0417] Compared with traditional methods, traditional methods perform fault diagnosis based on the threshold judgment model of current and voltage and use static topology analysis methods to locate faults. When the power grid topology is complex and the amount of data is large, the traditional method has low diagnostic accuracy and poor fault location accuracy. Specifically, the fault diagnosis of traditional methods relies on fixed thresholds and cannot adapt to dynamically changing power grid conditions, resulting in poor results when dealing with complex faults.

[0418] The experimental results show that the system of the present invention is superior to the traditional methods in terms of fault diagnosis, positioning accuracy, real-time performance and data processing capability. In terms of the accuracy of fault diagnosis, the method of the present invention reached 96.5%, while the accuracy of the traditional method was only 82.3%. In terms of fault location accuracy, the system of the present invention can accurately locate the fault point with an accuracy of 15.2 meters. In comparison, the positioning accuracy of the traditional method is 45.8 meters. In terms of real-time performance, the response time of the system of the present invention is 2.1 seconds, which is significantly lower than the 5.6 seconds of the traditional method. In addition, the data processing capability of the method of the present invention is also much higher than that of the traditional method, and can process 1,500 data per second, while the traditional method can only process 800 data per second.

[0419] The present invention combines graph neural networks with deep learning models to dynamically analyze and identify faults in high-dimensional data of power grids. Unlike traditional fault detection methods based on threshold setting, the present invention can automatically learn the characteristic patterns of the grid operation status, capture potential complex fault characteristics, and effectively eliminate the interference of environmental noise. Traditional methods usually rely on manually set thresholds, and there is a risk of detection errors due to changes in grid load or different equipment states. The present invention, through the graph neural network model, can process large-scale grid data and complex topological relationships, adapt to the dynamic changes in the grid operation status, and significantly improve the accuracy of fault diagnosis. In particular, it is more robust than traditional methods in high-frequency and instantaneous fault identification.

[0420] The present invention combines the grid topology information in the fault location process, and uses the time difference method and graph theory algorithm to accurately locate the fault area. Unlike the traditional method that relies on simple voltage and current signal differences for positioning, the present invention uses a fault signal model based on time difference propagation to accurately calculate the signal propagation time difference of each measurement point in the grid topology map. By accurately analyzing the fault signal propagation path through the graph theory algorithm, the fault source and fault location can be accurately identified in a complex grid environment.

[0421] The operating status of the power grid needs to be monitored in real time. Especially when a fault occurs, fault diagnosis and location must be completed as soon as possible to avoid further damage to the grid. The technical improvements in real-time performance of the present invention significantly improve the response speed of the system. A real-time stream processing framework is used to process multi-source data, and feature extraction and dimensionality reduction are performed in combination with wavelet transform and principal component analysis, which effectively reduces the delay time of data preprocessing in traditional methods. At the same time, the graph neural network model can complete the processing of high-dimensional features and fault identification in a relatively short time.

[0422] The present invention adopts a hierarchical storage strategy and multi-dimensional data index to build a data lake, which can effectively process large-scale, multi-source data from multiple power grid devices and sensors. Combined with a real-time stream processing framework, the system of the present invention can quickly filter and aggregate power grid data from different levels, automatically identify key information in the data, and effectively store and process it. Traditional methods usually rely on fixed rules and thresholds for data processing, which makes it difficult to cope with the real-time processing requirements of massive data in power grid systems.

[0423] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. The online identification system of power grid fault based on big data is characterized by: include: A data acquisition module, used for collecting multi-source data of the power grid system through sensors; The data storage module is used to filter data through a hierarchical storage strategy, build a data lake in combination with multi-dimensional data indexes, and classify and store multi-source data; The feature extraction module is used to aggregate and transform the data in the data lake using a real-time stream processing framework, and extract high-dimensional features related to faults through wavelet transform combined with principal component analysis; Fault diagnosis module, which is used to identify the real-time status of the power grid based on high-dimensional features using graph neural networks combined with deep learning models; The fault location module is used to locate the fault based on the recognition result of the real-time status and the topology of the power grid, using the time difference method combined with the graph theory algorithm; The fault location comprises: Receive real-time fault status identification results from the fault diagnosis module, including fault type and preliminary location; Construct a power grid topology diagram including node attributes; Based on the fault signal propagation model, the time difference of the fault signal reaching each measurement point is calculated to preliminarily locate the fault area; Based on the time difference calculation results, determine the candidate area or node set where the failure may occur; The Dijkstra algorithm is used to calculate the weighted shortest path from the preliminary positioning point to the candidate node by combining multiple weight factors, including impedance, line length, equipment load, equipment status score and signal attenuation factor; The weighted shortest path results and time difference calculation are combined, and a weighted scoring method is used to determine the specific node or device location where the fault occurred.

2. The online identification system for power grid faults based on big data according to claim 1, characterized in that: The tiered storage strategy, include, A tiered storage algorithm is used to calculate priority scores, evaluate and allocate multi-source data to the first and second storage tiers in real time; Prioritize With prediction priority threshold For comparison, when When , it is stored in the first storage layer. When , store it to the second storage layer; The multi-dimensional data index is set based on time, space and device type.

3. The online identification system for power grid faults based on big data according to claim 2, characterized in that: The priority score is calculated by using a hierarchical storage algorithm, including: Collect data from multiple data sources in the power grid According to predefined categories Classification is performed, wherein the categories include voltage data, current data, frequency data, phase data and environmental parameter data; For each category , from the classified data Extract key feature vectors from , expressed as, ， in, Indicates the real-time rating. represents the volatility score, represents the frequency of outliers, Indicates the importance score of data sources; Based on the data Historical access frequency data For each category , using the adaptive time series prediction model to predict the access frequency at future time t ; According to the extracted key feature vector and connectivity indicators in power grid topology , calculate the data Importance score , expressed as, ， in, Representation data The importance score of represents the data category set, k is the category index, indicating the category to which the data belongs, and m is the feature index, indicating the extracted feature score. Representation characteristics The weight coefficient of Representation data The score of the mth key feature under category k, represents the weight coefficient of the connectivity index, Representation data Node Connectivity indicators; By monitoring the operation status of the power grid and the external environment in real time, the clustering-based DBSCAN algorithm is used to detect and identify abnormal events. , based on the detected abnormal events, Importance score Perform dynamic adjustment, the adjustment formula is: ， in, represents the adjusted importance score, is the data for the jth event The adjustment factor for importance, is the jth detected abnormal event, and p is the number of detected events; Combined with the predicted visit frequency and the adjusted importance score , calculate the data The overall priority score , expressed as, ， in, Representation data The priority score of and is the weight coefficient used to balance access frequency and importance, and is the exponential coefficient.

4. The online identification system for power grid faults based on big data according to claim 3, characterized in that: The adaptive time series prediction model includes: Use the moving average method to identify the long-term trend of access frequency and conduct trend analysis; Wavelet transform was used to extract periodic patterns of visitation frequency and conduct seasonal analysis; Combining the results of trend analysis and seasonal analysis, the autoregressive integrated moving average model is used to predict future visit frequency. .

5. The online identification system for power grid faults based on big data according to claim 4, characterized in that: The trend analysis is expressed as, ， in, is the trend smoothing coefficient, , is the trend value at time t-1, is the access frequency at time t-1; The seasonal analysis is expressed as, ， in, is the jth wavelet coefficient, is the jth wavelet basis function, is the number of decomposition levels of wavelet transform; The future access frequency is expressed as, ， in, is the autoregression coefficient, is the moving average coefficient, is the predicted access frequency of category k at time t-p, is the prediction error of category k at time t-q, P and Q are the orders of the autoregressive and moving average models respectively.

6. The online identification system for power grid faults based on big data according to claim 5, characterized in that: The connectivity index is expressed as, ， in, represents the set of all nodes in the power grid, represents the total number of shortest paths between node s and node t, It represents the number of paths passing through node i in the shortest path between node s and node t.

7. The big data-based online identification system for power grid faults according to claim 6, characterized in that: The aggregation and conversion include: Real-time aggregation of multi-source data from the first storage layer and the second storage layer according to the three dimensions of time, space and device type; The aggregated data are standardized to eliminate the differences between different data sources and dimensions; Adopting adaptive wavelet transform method, the optimal wavelet basis function is dynamically selected according to the data characteristics of each storage layer; Perform multi-layer wavelet decomposition on the data of each storage layer, extract the approximate coefficients and detail coefficients, and reconstruct the seasonal components by accumulating the detail coefficients; The reconstructed high-frequency and low-frequency seasonal components are organized into a high-frequency feature matrix and a low-frequency feature matrix, respectively, and a comprehensive feature matrix is ​​formed by feature splicing; Apply principal component analysis to reduce the dimension of the comprehensive feature matrix and extract key features; Features are selected and optimized through optimization algorithms.

8. The online identification system for power grid faults based on big data according to claim 7, characterized in that: The identifying of the real-time status of the electric power grid includes: Normalize the high-dimensional feature vector to generate a feature set; A multi-layer graph convolutional network with residual connections and multi-head self-attention mechanism is used to extract spatial features in the power grid topology and generate global graph features through a graph pooling layer. Through the fully connected layer and the improved Softmax function, the probability distribution of various faults is output.

9. A method for applying to the big data-based online identification system for power grid faults as claimed in any one of claims 1 to 8, characterized in that: include: Collect multi-source data of the power grid system through sensors; Build a data lake through tiered storage strategy and multi-dimensional data indexing to store multi-source data in a classified manner; A real-time stream processing framework is used to filter, aggregate and transform data, and high-dimensional features related to faults are extracted through wavelet transform combined with principal component analysis; Based on high-dimensional features, we use graph neural networks combined with deep learning models to identify the real-time status of the power grid; According to the recognition results of real-time status and the topology of the power grid, fault location is carried out by combining the time difference method with graph theory algorithm.

Images