Online identification system of power grid faults based on big data

Abstract

The present invention relates to the technical field of power grid fault identification, and specifically to an online identification system for power grid faults based on big data, comprising a data collection module, a feature extraction module and a feature analysis module, wherein: the data collection module collects real-time power grid operation data; the feature extraction module obtains the topological structure of the power grid according to the real-time power grid operation data; and uses wavelet transform technology to perform wavelet transform on the real-time power grid operation data to generate power time-frequency features; the feature analysis module uses a random forest algorithm to predict whether there is a fault according to the power time-frequency features, wherein: if the prediction result is normal, the system is operated normally; if the prediction result is a fault, the key fault location and key fault features are determined according to the power time-frequency features, the power grid topological structure and the feature importance evaluation of the random forest algorithm.

Description

Technical Field

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

[0002] Traditional power grid fault identification systems have several significant technical problems. First, traditional systems usually rely on a single fault feature and fixed threshold judgment. This method is difficult to deal with multiple fault types in complex power grid environments, resulting in insufficient detection accuracy and response speed.

[0003] Secondly, traditional systems lack comprehensive analysis of multi-dimensional features when predicting faults and are unable to fully utilize the time-frequency information in power grid operation data, which limits the system's sensitivity and accuracy in capturing power grid anomalies.

[0004] In addition, traditional fault location methods mostly rely on simple algorithms and cannot effectively utilize the complex relationships of the power grid topology, resulting in low efficiency and accuracy of fault location. Traditional systems also face the problem of overfitting. Due to deficiencies in feature selection and model training, the system has poor robustness and flexibility and is difficult to adapt to the dynamic changes in the power grid's operating status. These technical problems cause traditional power grid fault detection systems to show high false alarm and missed alarm rates in practical applications, and cannot effectively ensure the stable and safe operation of the power grid.

[0005] In summary, traditional power grid fault detection systems have technical bottlenecks in feature extraction, fault prediction, fault location, and model robustness, and need to use more advanced methods and algorithms to improve their performance. Summary of the invention

[0006] The purpose of the present invention is to provide an online identification system for power grid faults based on big data to solve the problems raised in the above-mentioned background technology.

[0007] To achieve the above object, the present invention provides the following technical solution: an online identification system for power grid faults based on big data, characterized in that it includes a data collection module, a feature extraction module and a feature analysis module, wherein:

[0008] The data collection module collects real-time power grid operation data;

[0009] The feature extraction module uses a graph theory algorithm to obtain the topological structure of the power grid based on the real-time power grid operation data; and uses wavelet transform technology to perform wavelet transform on the real-time power grid operation data to generate power time-frequency features;

[0010] The feature analysis module uses the random forest algorithm to predict whether there is a fault based on the power time-frequency characteristics, where:

[0011] If the prediction result is a fault, the overall fault location is determined based on the power time-frequency characteristics and the grid topology. According to the feature importance evaluation in the random forest, the importance of each feature to fault judgment is analyzed, and the feature with the highest importance is selected as the key fault feature. The key fault location is determined by combining the key fault features with the grid topology, and the overall fault location and the key fault location are determined to contain the relationship, where:

[0012] If the overall fault location includes the key fault location, the key fault location and key fault characteristics are determined as the final fault identification result;

[0013] If the overall fault location does not include the critical fault location, then select the feature with the second highest importance as the critical fault feature according to the importance, and re-determine the critical fault location and the inclusion relationship until it is determined that the critical fault location is within the overall fault location;

[0014] If the prediction result is normal, the system will run normally.

[0015] As a further improvement of the present technical solution, the data collection module collects real-time grid operation data, which includes but is not limited to current, voltage and frequency, as well as static structural data of the grid.

[0016] As a further improvement of the technical solution, the feature extraction module includes a topology analysis unit, which uses a graph theory algorithm to obtain the topology structure of the power grid according to the real-time power grid operation data, specifically including:

[0017] According to the static structure data in the real-time power grid operation data, the power grid is represented as a graph, denoted as G, which contains nodes V and edges E, where nodes V represent devices in the power grid and edges E represent the connection lines between devices;

[0018] For each edge E in the graph G, a weight is assigned based on electrical parameters, including but not limited to resistance and voltage level of the line;

[0019] Use the priority search algorithm in graph theory to traverse the entire graph G, and obtain a connection matrix C through traversal, which represents the connection relationship between nodes: if node i and node j are directly connected, then C[i,j]=1; otherwise, C[i,j]=0;

[0020] Different features of the graph are calculated to generate the topological structure of the power grid. The degree distribution D(k) represents the proportion of nodes with a specific number of connections in the entire graph, where D(k) = the number of nodes with degree k / total number of nodes.

[0021] As a further improvement of the technical solution, the feature extraction module includes a time-frequency analysis unit, which uses wavelet transform technology to perform wavelet transform on real-time power grid operation data to generate power time-frequency features, specifically including:

[0022] For each real-time power grid operation data x(t), a continuous wavelet transform is performed, and the wavelet basis function ψ(t) is adjusted by the scale parameter a and the translation parameter b. The wavelet coefficient W(a, b) is obtained by convolution calculation with x(t). The energy characteristics at different scales are extracted from the wavelet coefficients as the power time-frequency characteristics.

[0023] As a further improvement of the technical solution, the feature analysis module includes a prediction unit, which uses a random forest algorithm to predict whether a fault occurs according to the power time-frequency characteristics, specifically including:

[0024] Training the random forest model: Use a historical data set that contains label data of known faults and normal operation, and perform feature extraction using wavelet transform technology on the historical data set, and use it to train the random forest model; the random forest consists of multiple decision trees. During the training process, subsamples are randomly extracted from the training data set and each decision tree is constructed; for each decision tree, some features are randomly selected during the construction process to split nodes;

[0025] Fault prediction: The power time-frequency features are input into the trained random forest model. Each decision tree in the random forest will independently predict the input feature vector and output a prediction result. The prediction results of all decision trees are aggregated and the final fault prediction result is obtained using a majority voting mechanism.

[0026] As a further improvement of this technical solution, the construction process of the decision number is as follows:

[0027] A subset is randomly selected from the historical data set to build a single decision tree. For each node split, a feature subset is randomly selected. At each node, a feature and a split point are selected so that the split point can minimize the impurity before and after the split. The split conditions for measuring impurity include but are not limited to the Gini coefficient and information entropy.

[0028] As a further improvement of the technical solution, the feature analysis module includes a decision unit, and the decision unit makes decisions on different paths according to the fault prediction results, specifically including:

[0029] If the prediction result is normal, the system will run normally;

[0030] If the prediction result is a fault, the overall fault location is determined based on the power time-frequency characteristics and the grid topology. According to the feature importance evaluation in the random forest, the importance of each feature to fault judgment is analyzed, and the feature with the highest importance is selected as the key fault feature. The key fault location is determined by combining the key fault features with the grid topology, and the overall fault location is determined to have an inclusion relationship with the key fault location, where:

[0031] If the overall fault location includes the key fault location, the key fault location and key fault characteristics are determined as the final fault identification result;

[0032] If the overall fault location does not include the critical fault location, the feature with the second highest importance is selected as the critical fault feature according to the importance, and the critical fault location is re-determined and the inclusion relationship is determined again until the critical fault location is determined to be within the overall fault location.

[0033] As a further improvement of the technical solution, the method of determining the overall fault location according to the power time-frequency characteristics and the grid topology structure specifically includes:

[0034] Collect the power time-frequency characteristics during normal operation from the prediction unit, calculate its mean μ and standard deviation σ, and determine the threshold of the power time-frequency characteristics according to the formula E=μ+kσ, where E is the power time-frequency characteristic threshold and k is a multiple factor used to reflect the selection of abnormal sensitivity;

[0035] The power time-frequency characteristics corresponding to all nodes in the power grid topology are compared with the power time-frequency characteristic threshold, and the nodes with a value greater than the power time-frequency characteristic threshold are regarded as abnormal nodes, where: for each pair of abnormal nodes vm and vn, if C[m,n]=1, they are directly connected; according to the edge weights of the graph G, the Dijkstra algorithm is used to find the shortest path between all abnormal nodes;

[0036] Calculate the average shortest path length Li from each abnormal node vi to all other abnormal nodes, the fault source candidate vs = argmin (Li), vi∈A; where A is the set of abnormal nodes; argmax is used to find the independent variable that achieves the minimum value of the function;

[0037] With vs as the center, construct the minimum connected subgraph Gs so that Gs contains all abnormal nodes. The node set in Gs is the location where the overall fault occurs.

[0038] As a further improvement of the technical solution, the feature importance evaluation in the random forest is used to analyze the importance of each feature to fault judgment, and the feature with the highest importance is selected as the key fault feature, which specifically includes:

[0039] In each decision tree, the split of each node is based on a certain feature; for each split, the reduction in impurity caused by the split is obtained, and the reduction in impurity is regarded as the importance of the feature at this node;

[0040] For each tree, the importance of each feature at all nodes is accumulated to obtain the total importance of each feature in a single tree;

[0041] Calculate the average importance of each feature in all decision trees. For each feature, average its importance in all N trees. Specifically, divide the total importance of each feature in a single tree by the total number of decision trees N.

[0042] All features are sorted according to their average importance, an importance-ranked list is generated, and the feature with the largest average importance is selected as the key fault feature.

[0043] As a further improvement of the technical solution, the process of re-determining the location where the critical fault occurs is as follows:

[0044] According to the importance, the feature with the second highest importance is selected as the key fault feature again. Specifically, according to the importance sorting list, the next feature of the current key fault feature is selected as the key fault feature.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] 1. The random forest algorithm in the online identification system of power grid faults based on big data can not only predict faults, but also determine key fault features based on feature importance evaluation. In the fault prediction process, the random forest model integrates multiple decision trees to independently predict the power time-frequency features, and adopts a majority voting mechanism to obtain the final fault prediction results, which improves the reliability and accuracy of the prediction; further, the random forest model also evaluates the importance of each feature and selects the most critical features for fault judgment as key fault features; if the overall fault location does not include the key fault location, the system will select the second most important feature according to the size of the feature importance to re-locate the fault and determine the inclusion relationship until it is determined that the key fault location is within the overall fault location. This process not only improves the accuracy of fault identification through multiple applications of the random forest algorithm, but also makes the system more flexible and robust when facing complex power grid faults.

[0047] 2. The online identification system of power grid fault based on big data determines the overall fault location according to the power time-frequency characteristics and the grid topology structure, and determines the key fault location according to the key fault characteristics and the grid topology structure. Through multiple applications of the grid topology structure, the fault can be located and analyzed more accurately; the grid topology structure uses graph theory algorithms to represent the various devices and connection lines in the grid as a graph, and generates a connection matrix and degree distribution through traversal and analysis; these topological features not only help identify the key nodes and connection relationships in the grid, but also provide data support for fault path analysis and shortest path calculation through the calculation of edge weights; combined with the threshold judgment of power time-frequency characteristics, abnormal nodes and normal nodes can be effectively distinguished. On this basis, by comparing the key fault feature nodes with the overall fault location, the key fault location can be accurately determined, and the inclusion relationship between the two can be further analyzed; this method of multiple application of the grid topology structure makes the system more comprehensive and accurate in fault location and analysis, and effectively improves the efficiency and reliability of grid fault handling.

[0048] 3. The online identification system for power grid faults based on big data predicts faults based on power time-frequency characteristics, and determines the overall fault location based on power time-frequency characteristics and grid topology. Through multiple applications of power time-frequency characteristics, the accuracy and response speed of grid fault detection can be significantly improved. Power time-frequency characteristics are extracted through wavelet transform technology, which can carefully reflect the changes in grid operation data at different times and frequencies. This enables the system to quickly determine the occurrence of faults based on subtle frequency and time domain characteristics when capturing grid anomalies. By calculating the threshold of power time-frequency characteristics and comparing the actual operation data with the threshold, abnormal nodes can be effectively identified, thereby providing accurate data support for subsequent fault location and processing. The application of power time-frequency characteristics not only enhances the accuracy of fault detection, but also improves the system's comprehensive perception of the grid operation status, ensuring the safe and stable operation of the grid. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a schematic diagram of the overall module of the present invention;

[0050] Figure 2 It is a schematic diagram of a feature extraction module unit of the present invention;

[0051] Figure 3 It is a schematic diagram of the feature analysis module unit of the present invention.

[0052] In the figure: 100, data collection module; 200, feature extraction module; 201, topological analysis unit; 202, time-frequency analysis unit; 300, feature analysis module; 301, prediction unit; 302, decision unit. DETAILED DESCRIPTION

[0053] The following will be combined with the accompanying drawings in the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0054] See also Figure 1-2 The present invention provides a technical solution: an online identification system for power grid faults based on big data, including a data collection module 100, a feature extraction module 200 and a feature analysis module 300.

[0055] The data collection module 100 collects real-time grid operation data by installing various sensors, such as current transformers and voltage transformers, on key nodes and lines of the grid. These sensors monitor the operation status of the grid in real time. The real-time grid operation data includes parameters such as current, voltage, and frequency, as well as static structural data of the grid, such as the location and connection relationship of substations, generators, and loads.

[0056] The topology analysis unit 201 in the feature extraction module 200 uses a graph theory algorithm to obtain the topology structure of the power grid according to the real-time power grid operation data, specifically including:

[0057] According to the static structure data in the real-time power grid operation data, the power grid is represented as a graph, denoted as G, which contains nodes V and edges E. The nodes V represent the equipment in the power grid, such as substations, generators and loads, and the edges E represent the connection lines between the equipment.

[0058] For each edge E in the graph G, a weight is assigned based on electrical parameters, including but not limited to resistance and voltage level of the line;

[0059] Use the priority search algorithm in graph theory to traverse the entire graph G, and obtain a connection matrix C through traversal, which represents the connection relationship between nodes: if node i and node j are directly connected, then C[i,j]=1; otherwise, C[i,j]=0;

[0060] Different features of the graph are calculated to generate the topological structure of the power grid. The degree distribution D(k) represents the proportion of nodes with a specific number of connections (degree) in the entire graph, where D(k) = number of nodes with degree k / total number of nodes. For example, D(3) represents the proportion of nodes with three connections in the graph.

[0061] The time-frequency analysis unit 202 in the feature extraction module 200 uses wavelet transform technology to perform wavelet transform on the real-time power grid operation data to generate power time-frequency features, specifically including:

[0062] For each real-time power grid operation data x(t), a continuous wavelet transform is performed, and the wavelet basis function ψ(t) is adjusted by the scale parameter a and the translation parameter b. Then, the wavelet coefficient W(a, b) is obtained by convolution calculation with x(t). The energy characteristics at different scales are extracted from the wavelet coefficients as the power time-frequency characteristics. The energy characteristics reflect the changes of the signal at different times and frequencies.

[0063] The prediction unit 301 in the feature analysis module 300 uses the random forest algorithm to predict whether there is a fault according to the power time-frequency characteristics, specifically including:

[0064] Training the random forest model: Use a historical data set that contains label data of known faults and normal operation, and perform feature extraction on the historical data set using wavelet transform technology to ensure the consistency of the input data type. This is used to train the random forest model. The random forest consists of multiple decision trees. During the training process, subsamples are randomly extracted from the training data set and each decision tree is constructed. For each decision tree, some features are randomly selected during the construction process to split nodes to reduce overfitting. The construction process of the decision tree is as follows:

[0065] From the historical data set, a subset is randomly extracted (with replacement sampling), and this subset is used to build a single decision tree; for each node split, a feature subset is randomly selected instead of using all features, which increases the diversity and robustness of the model; at each node, a feature and a split point are selected so that the split point can minimize the impurity before and after the split, where the split conditions for measuring impurity include but are not limited to the Gini coefficient and information entropy;

[0066] Fault prediction: The power time-frequency features are input into the trained random forest model. Each decision tree in the random forest will independently predict the input feature vector and output a prediction result (such as fault or normal). The prediction results of all decision trees are aggregated and the final fault prediction result is obtained by majority voting.

[0067] The decision unit 302 in the feature analysis module 300 will make decisions on different paths according to the fault prediction results, specifically including:

[0068] If the prediction result is normal, the system will run normally;

[0069] If the prediction result is a fault, the overall fault location is determined based on the power time-frequency characteristics and the grid topology, including:

[0070] Collect the power time-frequency characteristics during normal operation from the prediction unit 301, calculate its mean μ and standard deviation σ, and determine the threshold of the power time-frequency characteristics according to the formula E=μ+kσ, where E is the power time-frequency characteristic threshold and k is a multiple factor used to reflect the selection of abnormal sensitivity;

[0071] The power time-frequency characteristics corresponding to all nodes in the power grid topology are compared with the power time-frequency characteristic threshold, and the nodes with a value greater than the power time-frequency characteristic threshold are regarded as abnormal nodes, where: for each pair of abnormal nodes vm and vn, if C[m,n]=1, they are directly connected; according to the edge weights of the graph G (based on electrical parameters such as resistance and voltage level), the Dijkstra algorithm is used to find the shortest path between all abnormal nodes;

[0072] Calculate the average shortest path length Li from each abnormal node vi to all other abnormal nodes, the fault source candidate vs = argmin (Li), vi∈A; where A is the set of abnormal nodes; argmax is used to find the independent variable that achieves the minimum value of the function;

[0073] With vs as the center, construct the minimum connected subgraph Gs so that Gs contains all abnormal nodes. The node set in Gs is the location where the overall fault occurs.

[0074] The decision unit 302 analyzes the importance of each feature to fault judgment based on the feature importance evaluation in the random forest, and selects the most important feature as the key fault feature, which specifically includes:

[0075] In each decision tree, the split of each node is based on a certain feature; for each split, the reduction in impurity caused by the split is obtained, and the reduction in impurity is regarded as the importance of the feature at this node;

[0076] For each tree, the importance of each feature at all nodes is accumulated, which gives the total importance of each feature in a single tree;

[0077] Calculate the average importance of each feature in all decision trees. For each feature, average its importance in all N trees. Specifically, divide the total importance of each feature in a single tree by the total number of decision trees N.

[0078] All features are sorted according to their average importance, an importance-ranked list is generated, and the feature with the largest average importance is selected as the key fault feature.

[0079] The decision unit 302 determines the critical fault location according to the critical fault characteristics and the grid topology, and determines whether the critical fault characteristics belong to the overall fault location, specifically including:

[0080] Compare the key fault features with the threshold of the power time-frequency features, find the key nodes that are less than the threshold of the power time-frequency features, and compare the key nodes with the nodes in the overall fault location one by one, where:

[0081] If the key node is among the nodes in the overall fault location, it means that the overall fault location contains the key fault location, and the key fault location and key fault characteristics are determined as the final fault identification result;

[0082] If the critical node is not among the nodes in the overall fault location, it means that the overall fault location does not contain the critical fault location. Then, according to the importance, the feature with the second highest importance is selected as the critical fault feature again. Specifically, according to the importance sorting list, the next feature of the current critical fault feature is selected as the critical fault feature; and the critical fault location is determined again and the inclusion relationship is determined again until it is determined that the critical fault location is within the overall fault location.

[0083] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and descriptions are only preferred examples of the present invention and are not intended to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. The online identification system of power grid fault based on big data is characterized by: The method comprises a data collection module (100), a feature extraction module (200) and a feature analysis module (300), wherein: The data collection module (100) collects real-time power grid operation data; The feature extraction module (200) uses a graph theory algorithm to obtain a topological structure of the power grid according to the real-time power grid operation data; and uses a wavelet transformation technology to perform a wavelet transformation on the real-time power grid operation data to generate a power time-frequency feature; The feature analysis module (300) uses a random forest algorithm to predict whether a fault occurs based on the power time-frequency features, wherein: If the prediction result is a fault, the overall fault location is determined based on the power time-frequency characteristics and the grid topology. According to the feature importance evaluation in the random forest, the importance of each feature to fault judgment is analyzed, and the feature with the highest importance is selected as the key fault feature. The key fault location is determined by combining the key fault features with the grid topology, and the overall fault location and the key fault location are determined to contain the relationship, where: If the overall fault location includes the key fault location, the key fault location and key fault characteristics are determined as the final fault identification result; If the overall fault location does not include the critical fault location, then select the feature with the second highest importance as the critical fault feature according to the importance, and re-determine the critical fault location and the inclusion relationship until it is determined that the critical fault location is within the overall fault location; If the prediction result is normal, the system will run normally.

2. The online identification system for power grid faults based on big data according to claim 1 is characterized in that: The data collection module (100) collects real-time power grid operation data, which includes but is not limited to current, voltage and frequency, as well as static structural data of the power grid.

3. The online identification system for power grid faults based on big data according to claim 1 is characterized in that: The feature extraction module (200) comprises a topology analysis unit (201), wherein the topology analysis unit (201) uses a graph theory algorithm to obtain a topology structure of a power grid according to real-time power grid operation data, specifically comprising: According to the static structure data in the real-time power grid operation data, the power grid is represented as a graph, denoted as G, which contains nodes V and edges E, where nodes V represent devices in the power grid and edges E represent the connection lines between devices; For each edge E in the graph G, a weight is assigned based on electrical parameters, including but not limited to resistance and voltage level of the line; Use the priority search algorithm in graph theory to traverse the entire graph G, and obtain a connection matrix C through traversal, which represents the connection relationship between nodes: if node i and node j are directly connected, then C[i,j]=1; otherwise, C[i,j]=0; Different features of the graph are calculated to generate the topological structure of the power grid. The degree distribution D(k) represents the proportion of nodes with a specific number of connections in the entire graph, where D(k) = the number of nodes with degree k / total number of nodes.

4. The online identification system for power grid faults based on big data according to claim 1 is characterized in that: The feature extraction module (200) comprises a time-frequency analysis unit (202), wherein the time-frequency analysis unit (202) performs wavelet transformation on real-time power grid operation data using wavelet transformation technology to generate power time-frequency features, specifically comprising: For each real-time power grid operation data x(t), a continuous wavelet transform is performed, and the wavelet basis function ψ(t) is adjusted by the scale parameter a and the translation parameter b. The wavelet coefficient W(a, b) is obtained by convolution calculation with x(t). The energy characteristics at different scales are extracted from the wavelet coefficients as the power time-frequency characteristics.

5. The online identification system for power grid faults based on big data according to claim 1 is characterized in that: The feature analysis module (300) comprises a prediction unit (301), wherein the prediction unit (301) uses a random forest algorithm to predict whether a fault occurs according to the power time-frequency feature, specifically comprising: Training the random forest model: Use a historical data set that contains label data of known faults and normal operation, and perform feature extraction using wavelet transform technology on the historical data set, and use it to train the random forest model; the random forest consists of multiple decision trees. During the training process, subsamples are randomly extracted from the training data set and each decision tree is constructed; for each decision tree, some features are randomly selected during the construction process to split nodes; Fault prediction: The power time-frequency features are input into the trained random forest model. Each decision tree in the random forest will independently predict the input feature vector and output a prediction result. The prediction results of all decision trees are aggregated and the final fault prediction result is obtained using a majority voting mechanism.

6. The online identification system for power grid faults based on big data according to claim 5 is characterized in that: The construction process of the decision number is as follows: A subset is randomly selected from the historical data set to build a single decision tree. For each node split, a feature subset is randomly selected. At each node, a feature and a split point are selected so that the split point can minimize the impurity before and after the split. The split conditions for measuring impurity include but are not limited to the Gini coefficient and information entropy.

7. The online identification system for power grid faults based on big data according to claim 1 is characterized in that: The feature analysis module (300) includes a decision unit (302), and the decision unit (302) makes decisions on different paths according to the fault prediction result, specifically including: If the prediction result is normal, the system will run normally; If the prediction result is a fault, the overall fault location is determined based on the power time-frequency characteristics and the grid topology. According to the feature importance evaluation in the random forest, the importance of each feature to fault judgment is analyzed, and the feature with the highest importance is selected as the key fault feature. The key fault location is determined by combining the key fault features with the grid topology, and the overall fault location is determined to have an inclusion relationship with the key fault location, where: If the overall fault location includes the key fault location, the key fault location and key fault characteristics are determined as the final fault identification result; If the overall fault location does not include the critical fault location, the feature with the second highest importance is selected as the critical fault feature according to the importance, and the critical fault location is re-determined and the inclusion relationship is determined again until the critical fault location is determined to be within the overall fault location.

8. The online identification system for power grid faults based on big data according to claim 7 is characterized in that: Determining the overall fault location according to the power time-frequency characteristics and the grid topology structure specifically includes: Collecting the power time-frequency characteristics during normal operation from the prediction unit (301), calculating the mean μ and standard deviation σ thereof, and determining the threshold of the power time-frequency characteristics according to the formula E=μ+kσ, where E is the power time-frequency characteristics threshold, and k is a multiple factor used to reflect the selection of abnormal sensitivity; The power time-frequency characteristics corresponding to all nodes in the power grid topology are compared with the power time-frequency characteristic threshold, and the nodes with a value greater than the power time-frequency characteristic threshold are regarded as abnormal nodes, where: for each pair of abnormal nodes vm and vn, if C[m,n]=1, they are directly connected; according to the edge weights of the graph G, the Dijkstra algorithm is used to find the shortest path between all abnormal nodes; Calculate the average shortest path length Li from each abnormal node vi to all other abnormal nodes, the fault source candidate vs = argmin (Li), vi∈A; where A is the set of abnormal nodes; argmax is used to find the independent variable that achieves the minimum value of the function; With vs as the center, construct the minimum connected subgraph Gs so that Gs contains all abnormal nodes. The node set in Gs is the location where the overall fault occurs.

9. The online identification system for power grid faults based on big data according to claim 8 is characterized in that: According to the feature importance evaluation in the random forest, the importance of each feature to fault judgment is analyzed, and the most important feature is selected as the key fault feature, which specifically includes: In each decision tree, the split of each node is based on a certain feature; for each split, the reduction in impurity caused by the split is obtained, and the reduction in impurity is regarded as the importance of the feature at this node; For each tree, the importance of each feature at all nodes is accumulated to obtain the total importance of each feature in a single tree; Calculate the average importance of each feature in all decision trees. For each feature, average its importance in all N trees. Specifically, divide the total importance of each feature in a single tree by the total number of decision trees N. All features are sorted according to their average importance, an importance-ranked list is generated, and the feature with the largest average importance is selected as the key fault feature.

10. The online identification system for power grid faults based on big data according to claim 8 is characterized in that: The process of re-determining the critical fault occurrence location is as follows: According to the importance, the feature with the second highest importance is selected as the key fault feature again. Specifically, according to the importance sorting list, the next feature of the current key fault feature is selected as the key fault feature.

Images