A method and system for online monitoring of power quality
Through deep learning and hierarchical map mapping technology, a power quality monitoring system is built to generate intelligent maintenance solutions, which solves the problem that existing systems are difficult to provide personalized maintenance suggestions, and improves the data utilization and processing efficiency of power quality monitoring.
Patent Information
- Application Number
- CN202411338025.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-25
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2044-09-25
AI Technical Summary
The existing power quality monitoring system is difficult to provide personalized and targeted maintenance suggestions, and relies on the experience of maintenance workers to determine the cause of the failure and the handling methods, and fails to make full use of system data.
Deep learning method is used to build a correlation matrix and duration prediction model, map nonlinear load equipment, abnormal data and test actions through hierarchical graphs, generate an abnormal event test plan in the target space, and update the hierarchical graph with user feedback to provide intelligent maintenance solutions. .
It realizes intelligent processing of nonlinear load equipment and power data, uses maintenance experience to provide personalized testing solutions, and improves the data utilization and processing efficiency of power quality monitoring.
Smart Images

Figure CN118858827B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of power quality monitoring and processing, and more specifically, to a power quality online monitoring method and system. Background Art
[0002] Most current power quality monitoring systems rely on automated data collection and algorithm analysis, mainly for real-time monitoring of power parameters such as voltage, current, harmonics, etc. in the power system, aiming to detect potential power system problems. However, the existing monitoring systems have the following shortcomings:
[0003] Although existing systems can automatically collect data and analyze anomalies, in actual operation, the experience of maintenance workers is still the main way to solve problems. Maintenance workers often use their own experience and on-site monitoring data to determine the cause of the fault and choose a treatment method. These behaviors are not fully utilized in existing monitoring systems, making it difficult for the system to provide personalized and targeted maintenance suggestions. Summary of the invention
[0004] The present invention provides a method and system for online monitoring of power quality, which solve the technical problems raised in the background technology.
[0005] In a first aspect, a method for online monitoring of power quality comprises:
[0006] Step S101, collecting abnormal power data of nonlinear load devices in a sample space; collecting abnormal test behaviors of maintenance personnel on nonlinear load devices and the duration of the abnormal test behaviors;
[0007] Nonlinear load devices and abnormal data are labeled and abnormal test behaviors are classified, and K types of nonlinear load devices, M types of abnormal data and N test actions are obtained;
[0008] Step S102, constructing a correlation matrix based on each group of abnormal power data and each group of abnormal test behaviors, and constructing training samples based on each correlation matrix and the duration of each group of abnormal test behaviors, and performing deep learning based on the training samples to obtain a duration prediction model;
[0009] Step S103, constructing a hierarchical graph, mapping K types of nonlinear load devices as first-level nodes of the hierarchical graph, mapping M types of abnormal data as second-level nodes of the hierarchical graph, and repeatedly mapping N test actions as third-level to N-level nodes of the hierarchical graph;
[0010] Extract the entity relationships between nonlinear load devices, abnormal data and test actions, and construct the edges of nodes at all levels based on the entity relationships;
[0011] Step S104, constructing a monitoring data processing platform based on the hierarchical graph and the duration prediction model, the monitoring data processing platform is used to process abnormal nonlinear load devices in the target space and their abnormal power data, and obtain a test plan for abnormal events in the target space.
[0012] Furthermore, a correlation matrix is constructed based on each set of abnormal power data and each set of abnormal test behavior, including:
[0013] The association matrix is an M×N matrix constructed by M types of abnormal data and N test actions. The element in the i-th row and j-th column of the association matrix represents the association between the i-th abnormal data and the j-th test action. If there is an association between the i-th abnormal data and the j-th test action, the element in the i-th row and j-th column of the association matrix is (i, j), otherwise it is (0, 0).
[0014] Furthermore, each correlation matrix and the duration of each group of abnormal test behaviors are used to construct training samples, and a duration prediction model is obtained by deep learning based on the training samples, including:
[0015] Normalize the duration data of each correlation matrix and each group of abnormal test behaviors;
[0016] The normalized correlation matrix is used as the input of the duration prediction model;
[0017] The duration prediction model includes M hidden layers, each hidden layer includes N processing units; the hidden state of the Nth processing unit of the Mth hidden layer is input into a preset classifier, and the classification space of the classifier represents the total test duration of the group of abnormal test behaviors for the corresponding abnormal power data; the duration prediction model uses the mean error variance loss function to update the weight parameters and bias parameters.
[0018] Furthermore, the entity relationships among nonlinear load devices, abnormal data and test actions are extracted, and the edges of nodes at all levels are constructed based on the entity relationships;
[0019] According to the nonlinear load devices, abnormal data and test actions, K first-level nodes mapping K types of nonlinear load devices, M second-level nodes mapping M types of abnormal data and N third-level nodes mapping N test actions are constructed respectively;
[0020] The nonlinear load equipment, abnormal data and test actions are connected by their correlation, and the correlation is used as the weight of the corresponding edge;
[0021] The association degree is used to represent the probability of the i-th node and the j-th node in the hierarchical graph co-occurring in the sample space data; the calculation formula for the co-occurrence probability is as follows:
[0022] ;in, represents the probability of the i-th node and the j-th node appearing together in the hierarchical graph, represents the total number of times the i-th node appears in the sample space data, represents the total number of times the jth node appears in the sample space data, Represents the total number of times the i-th node and the j-th node appear together in the sample space data, and 1 is an integer 1.
[0023] Furthermore, a monitoring data processing platform is constructed based on the hierarchical graph and the duration prediction model. The monitoring data processing platform is used to process the abnormal nonlinear load equipment and its abnormal power data in the target space, and obtain a test plan for abnormal events in the target space. The monitoring data processing platform includes:
[0024] A data acquisition unit, used for acquiring abnormal nonlinear load devices in the target space and abnormal data of the nonlinear load devices;
[0025] A test suggestion unit is used to determine a test plan for abnormal events in the target space based on abnormal nonlinear load devices in the target space, abnormal data of the corresponding nonlinear load devices, correlation between nonlinear load devices, correlation between abnormal data and test actions, and correlation between test actions, and traverse the hierarchical graph according to a preset threshold value, combined with a duration prediction model;
[0026] A feedback collection unit, used to collect users' opinions on the test plan of abnormal events in the target space, and pre-process the opinions on the test plan to obtain feedback data;
[0027] A feature extraction unit is used to process the feedback data, and the feature extraction unit includes:
[0028] The word segmentation module uses the Chinese word segmentation tool to segment the feedback data;
[0029] The part-of-speech tagging module uses the part-of-speech tagging tool to tag the feedback data after word segmentation;
[0030] The recognition module uses recognition tools to perform entity recognition and relational word recognition on the feedback data after part-of-speech tagging, extract node entities related to the hierarchical graph, and build connection relationships between node entities;
[0031] The entity linking module matches and links the node entities in the feedback data with the nodes in the hierarchical graph;
[0032] The hierarchical graph updating module updates the association degree of each node in the hierarchical graph according to the connection relationship between the node entities.
[0033] Furthermore, the hierarchical graph is traversed according to the preset threshold value, and the test scheme for abnormal events in the target space is determined in combination with the duration prediction model, including:
[0034] Obtaining nonlinear load devices in the target space and abnormal data of corresponding nonlinear load devices;
[0035] Determine the test sequence of the nonlinear load devices according to the correlation between each nonlinear load device and other nonlinear load devices in the target space;
[0036] If the correlation degree of the X nonlinear load devices to be selected is the same as that of the current nonlinear load device, the X nonlinear load devices to be selected are all used as the nonlinear load devices to be tested next, and X branches of the test scheme are obtained;
[0037] Determine the order of the test actions according to the correlation between the abnormal data of the nonlinear load device of the i-th test and the N test actions and the correlation between the N test actions, combined with the preset correlation threshold of the test actions;
[0038] If the correlation between the Y test actions to be selected and the current abnormal data is the same, then the Y test actions to be selected are all used as the initial test actions, and Y branches are obtained; if the correlation between the Z test actions to be selected and the current test action is the same, then the current Z abnormal data to be selected are all used as the abnormal data of the next test, and Z branches are obtained;
[0039] Wherein, when the correlation between the test action to be selected and the current test action is less than the correlation threshold of the test action, the selection is stopped;
[0040] Determine several test scenarios for abnormal events in the target space, and use a duration prediction model to evaluate the duration of each test scenario to determine the final test scenario.
[0041] Furthermore, the duration prediction model is used to evaluate the duration of each test plan to determine the final test plan, including:
[0042] Construct a correlation matrix for each test scenario;
[0043] The correlation matrix is input into the duration prediction model, and the duration prediction model outputs the test duration of the test plan;
[0044] Select several options as alternatives based on the test duration of each test option;
[0045] The screening includes comparing the test duration of each test solution and selecting a test solution whose duration is lower than a preset time threshold, inputting the test solutions whose duration is lower than the time threshold into a generative language tool and scoring them, selecting a test solution whose score is higher than the preset score threshold, and obtaining a number of alternative solutions;
[0046] Several alternative solutions are randomly pushed to the target space through a random algorithm.
[0047] Furthermore, according to the connection relationship between the node entities, the association degree of each node in the hierarchical graph is updated, including:
[0048] The connection relationship between node entities is divided into strong positive association, weak positive association, no association, weak negative association and strong negative association according to the part-of-speech tagging by the recognition tool; wherein each association state is respectively set with a corresponding update value;
[0049] The association degree of the hierarchical graph nodes is updated based on the connection relationship between the node entities. The calculation formula is as follows:
[0050] ;in, represents the updated association between node i and node j in the hierarchical graph, Represents the current association between node i and node j in the hierarchical graph, Represents the updated value of the association state between the node entities mapped by node i and node j in the hierarchical graph, and Respectively represent the upper and lower thresholds of the association degree of node i and node j in the hierarchical graph, Represents the weight parameter.
[0051] In a second aspect, an online power quality monitoring system includes:
[0052] The data acquisition module collects abnormal power data of nonlinear load devices in the sample space; collects abnormal test behaviors of maintenance personnel on nonlinear load devices and the duration of the abnormal test behaviors;
[0053] The prediction module constructs a correlation matrix based on each group of abnormal power data and each group of abnormal test behaviors, and constructs a training sample based on each correlation matrix and the duration of each group of abnormal test behaviors, and performs deep learning based on the training samples to obtain a duration prediction model;
[0054] A hierarchical graph construction module constructs a hierarchical graph, maps K types of nonlinear load devices to the first-level nodes of the hierarchical graph, maps M types of abnormal data to the second-level nodes of the hierarchical graph, and repeatedly maps N test actions to the third to N-level nodes of the hierarchical graph;
[0055] The monitoring and processing module builds a monitoring data processing platform based on the hierarchical graph and the duration prediction model. The monitoring data processing platform is used to process abnormal nonlinear load devices in the target space and their abnormal power data, and obtain a test plan for abnormal events in the target space.
[0056] In a third aspect, the present invention provides a computing device, comprising:
[0057] one or more processors;
[0058] The storage device is used to store one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the method described.
[0059] The beneficial effect of the present invention is that by decomposing the maintenance operation of the maintenance worker into several actions and then constructing them into a hierarchical graph, it is possible to intelligently process abnormal nonlinear load equipment and abnormal power data occurring in any target space and provide solutions, and update the hierarchical graph based on the feedback of the solutions, so that the power quality monitoring data can be processed in an artificial experience manner, so that the power quality monitoring data can be fully utilized. BRIEF DESCRIPTION OF THE DRAWINGS
[0060] Figure 1 It is a flow chart of a method for online monitoring of power quality of the present invention. DETAILED DESCRIPTION
[0061] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that the discussion of these embodiments is only to enable those skilled in the art to better understand and implement the subject matter described herein, and the functions and arrangements of the elements discussed may be changed without departing from the scope of protection of the contents of this specification. Each example may omit, replace or add various processes or components as needed. In addition, the features described relative to some examples may also be combined in other examples.
[0062] like Figure 1 As shown, a method for online monitoring of power quality includes:
[0063] Step S101, collecting abnormal power data of nonlinear load devices in a sample space; collecting abnormal test behaviors of maintenance personnel on nonlinear load devices and the duration of the abnormal test behaviors;
[0064] Nonlinear load devices and abnormal data are labeled and abnormal test behaviors are classified, and K types of nonlinear load devices, M types of abnormal data and N test actions are obtained;
[0065] Step S102, constructing a correlation matrix based on each group of abnormal power data and each group of abnormal test behaviors, and constructing training samples based on each correlation matrix and the duration of each group of abnormal test behaviors, and performing deep learning based on the training samples to obtain a duration prediction model;
[0066] Step S103, constructing a hierarchical graph, mapping K types of nonlinear load devices as first-level nodes of the hierarchical graph, mapping M types of abnormal data as second-level nodes of the hierarchical graph, and repeatedly mapping N test actions as third-level to N-level nodes of the hierarchical graph;
[0067] Extract the entity relationships between nonlinear load devices, abnormal data and test actions, and construct the edges of nodes at all levels based on the entity relationships;
[0068] Step S104, constructing a monitoring data processing platform based on the hierarchical graph and the duration prediction model, the monitoring data processing platform is used to process abnormal nonlinear load devices in the target space and their abnormal power data, and obtain a test plan for abnormal events in the target space.
[0069] In one embodiment of the present invention, a correlation matrix is constructed based on each group of abnormal power data and each group of abnormal test behavior, including:
[0070] The association matrix is an M×N matrix constructed by M types of abnormal data and N test actions. The element in the i-th row and j-th column of the association matrix represents the association between the i-th abnormal data and the j-th test action. If there is an association between the i-th abnormal data and the j-th test action, the element in the i-th row and j-th column of the association matrix is (i, j), otherwise it is (0, 0).
[0071] The sample space is the space of several power anomalies in historical data;
[0072] K types of nonlinear load devices means that nonlinear load devices are classified based on their types and powers, and a total of K types of nonlinear load devices are obtained; M types of abnormal data means that abnormal data are classified based on their types and numerical values, and a total of M types of abnormal data are obtained; N test actions means that each group of abnormal test behaviors is decomposed to obtain test actions, and the test actions are classified to obtain a total of N test actions;
[0073] For example, a factory has nonlinear load devices such as inverters, two 2-kilowatt UPSs, and one 100-watt UPS. Based on the type and power of the nonlinear load devices, these nonlinear load devices are classified, and K=3 types of nonlinear load devices are obtained, namely inverters, 2-kilowatt UPSs, and 100-watt UPSs. The inverter generates abnormal data, voltage fluctuations, harmonic distortion, and current overload. According to the type and value of the abnormal data, it is assumed that a total of M=4 types of abnormal data are obtained, namely, overvoltage, undervoltage, harmonic distortion, and current overload. The maintenance worker tests each group of abnormal power data. Each group of abnormal test behaviors can be decomposed into multiple specific test actions, testing the input and output voltage of the equipment, detecting the harmonic content, and adjusting the equipment load. After decomposition and classification, it is assumed that we get N=4 test actions, namely: voltage detection, harmonic testing, equipment power off and restart, and equipment load adjustment.
[0074] In one embodiment of the present invention, each correlation matrix and the duration of each group of abnormal test behaviors are used to construct a training sample, and a duration prediction model is obtained by deep learning based on the training samples, including:
[0075] Normalize the duration data of each correlation matrix and each group of abnormal test behaviors;
[0076] The normalized correlation matrix is used as the input of the duration prediction model;
[0077] The duration prediction model includes M hidden layers, each hidden layer includes N processing units; the hidden state of the Nth processing unit of the Mth hidden layer is input into a preset classifier, and the classification space of the classifier represents the total test duration of the group of abnormal test behaviors for the corresponding abnormal power data; the duration prediction model uses the mean error variance loss function to update the weight parameters and bias parameters.
[0078] Normalize each correlation matrix and each group of abnormal test behavior duration data; the calculation formula is as follows:
[0079] ;in, represents the normalized data, Represents the input data, Indicates the maximum input data. Represents the minimum input data;
[0080] It should be noted that the calculation formula of the hidden layer of the duration prediction model is as follows:
[0081] ;
[0082] ;
[0083] ;
[0084] ;
[0085] in, represents the output of the reset gate, represents the output of the reset gate, represents the output of the candidate hidden state, represents the input of the duration prediction model, Indicates updating the hidden state. and Represents the weight parameter of the reset gate, represents the bias of the reset gate, and Respectively represent the weight parameters of the reset gate, represents the bias of the reset gate, and represents the weight parameter of the candidate hidden state, represents the bias of the candidate hidden state, , represents the hidden state output by the N-1th unit of the Mth hidden layer, represents the hidden state output by the Mth unit in the N-1th hidden layer, express Activation function, represents the hyperbolic tangent function;
[0086] In one embodiment of the present invention, entity relationships between nonlinear load devices, abnormal data, and test actions are extracted, and edges of nodes at various levels are constructed based on the entity relationships;
[0087] According to the nonlinear load devices, abnormal data and test actions, K first-level nodes mapping K types of nonlinear load devices, M second-level nodes mapping M types of abnormal data and N third-level nodes mapping N test actions are constructed respectively;
[0088] The nonlinear load equipment, abnormal data and test actions are connected by their correlation, and the correlation is used as the weight of the corresponding edge;
[0089] The association degree is used to represent the probability of the i-th node and the j-th node in the hierarchical graph co-occurring in the sample space data; the calculation formula for the co-occurrence probability is as follows:
[0090] ;in, represents the probability of the i-th node and the j-th node appearing together in the hierarchical graph, represents the total number of times the i-th node appears in the sample space data, represents the total number of times the jth node appears in the sample space data, Represents the total number of times the i-th node and the j-th node appear together in the sample space data, and 1 is an integer 1.
[0091] It should be noted that edges are constructed between nonlinear load devices, between nonlinear load devices and abnormal data, between abnormal data and test actions, and between test actions.
[0092] In one embodiment of the present invention, a monitoring data processing platform is constructed based on a hierarchical graph and a duration prediction model. The monitoring data processing platform is used to process abnormal nonlinear load devices in a target space and abnormal power data thereof, and obtain a test plan for abnormal events in the target space. The monitoring data processing platform includes:
[0093] A data acquisition unit, used for acquiring abnormal nonlinear load devices in the target space and abnormal data of the nonlinear load devices;
[0094] The test suggestion unit determines the test plan for abnormal events in the target space according to the abnormal nonlinear load devices in the target space, the abnormal data of the corresponding nonlinear load devices, the correlation between the nonlinear load devices, the correlation between the abnormal data and the test actions, and the correlation between the test actions, and traverses the hierarchical graph according to the preset threshold value, combined with the duration prediction model;
[0095] A feedback collection unit, used to collect users' opinions on the test plan of abnormal events in the target space, and pre-process the opinions on the test plan to obtain feedback data;
[0096] A feature extraction unit is used to process the feedback data, and the feature extraction unit includes:
[0097] The word segmentation module uses the Chinese word segmentation tool to segment the feedback data;
[0098] The part-of-speech tagging module uses the part-of-speech tagging tool to tag the feedback data after word segmentation;
[0099] The recognition module uses recognition tools to perform entity recognition and relational word recognition on the feedback data after part-of-speech tagging, extract node entities related to the hierarchical graph, and build connection relationships between node entities;
[0100] The entity linking module matches and links the node entities in the feedback data with the nodes in the hierarchical graph;
[0101] The hierarchical graph updating module updates the association degree of each node in the hierarchical graph according to the connection relationship between the node entities.
[0102] For example, in a factory, the target space includes a frequency converter and a UPS system; the data acquisition unit monitors the voltage, current and power data of the frequency converter and the UPS system in real time. When the frequency converter has abnormal voltage fluctuations during operation, the data acquisition unit records the abnormal data that the voltage of the frequency converter rises to 380V and the current overload of the UPS system. These abnormal data and the information of the nonlinear load device are collected and transmitted to the monitoring data processing platform; the test suggestion unit determines to test the frequency converter first and then the UPS system based on the correlation between the monitoring frequency converter and the UPS system and their abnormal data and the nonlinear load devices; for example, the monitoring data processing platform recommends first performing voltage and current detection on the frequency converter, and then adjusting the load of the UPS system and performing harmonic testing; the factory maintenance personnel executed the test plan of the monitoring data processing platform and found that the problem was partially solved, but there were still some small voltage fluctuations. The maintenance personnel gave feedback on this test plan, pointing out that the plan was effective but did not completely solve the problem, and suggested adding a cable connection inspection step for the UPS system. The feedback collection unit collected the opinions of the maintenance personnel and preprocessed them to form structured feedback data; in the feedback data, the maintenance personnel gave the following feedback: "The plan is effective, but it is necessary to add an inspection of the cable connection of the UPS system after the harmonic test of the UPS system." The feature extraction unit first decomposes the feedback into several words through the word segmentation module, such as: "Plan / effective / need / add / UPS system / harmonic test / after / cable connection / check". Then, the part-of-speech tagging module will annotate these words and recognize that "plan" is a noun, "effective" is an adjective, and "check" is a verb; then, the recognition module further analyzes the data after part-of-speech tagging, extracts "cable connection" and "harmonic test" as two test action node entities, and recognizes that "need to add" represents an action relationship and "after" represents a sequence relationship; therefore, the association degree from "harmonic test" to "cable connection" will be increased;
[0103] It should be noted that the Chinese word segmentation tool may be Jieba word segmentation tool; the part-of-speech tagging tool may be LTP; the entity recognition tool may be SpaCy; and the recognition tool may be Standford NER;
[0104] In one embodiment of the present invention, a test scheme for abnormal events in a target space is determined by traversing a hierarchical graph according to a preset threshold value and combining a duration prediction model, including:
[0105] Obtaining nonlinear load devices in the target space and abnormal data of corresponding nonlinear load devices;
[0106] Determine the test sequence of the nonlinear load devices according to the correlation between each nonlinear load device and other nonlinear load devices in the target space;
[0107] If the correlation degree of the X nonlinear load devices to be selected is the same as that of the current nonlinear load device, the X nonlinear load devices to be selected are all used as the nonlinear load devices to be tested next, and X branches of the test scheme are obtained;
[0108] Determine the order of the test actions according to the correlation between the abnormal data of the nonlinear load device of the i-th test and the N test actions and the correlation between the N test actions, combined with the preset correlation threshold of the test actions;
[0109] If the correlation between the Y test actions to be selected and the current abnormal data is the same, then the Y test actions to be selected are all used as the initial test actions, and Y branches are obtained; if the correlation between the Z test actions to be selected and the current test action is the same, then the current Z abnormal data to be selected are all used as the abnormal data of the next test, and Z branches are obtained;
[0110] Wherein, when the correlation between the test action to be selected and the current test action is less than the correlation threshold of the test action, the selection is stopped;
[0111] Determine several test scenarios for abnormal events in the target space, and use a duration prediction model to evaluate the duration of each test scenario to determine the final test scenario.
[0112] For example, in a factory, four abnormal nonlinear load devices A1, A2, A3 and A4 are monitored; among them, the correlation between A1 and the other three nonlinear load devices is , the correlation between A2 and the other three nonlinear load devices are , the correlation between A3 and the other three nonlinear load devices are , the correlation between A4 and the other three nonlinear load devices are ; A4 has the highest correlation value, so A4 is selected as the first nonlinear load device to be tested. Based on the fact that A4 has the same attention to A1 and A3, the test plan will launch two test branches in parallel, and the test order of subsequent nonlinear load devices will be obtained based on A1 and A3; for example, the test order of nonlinear load devices is ; The abnormal data of A4 include four abnormal data: overvoltage, undervoltage, abnormal current and harmonic distortion. The correlation between these four abnormal data and N test actions is calculated respectively, and the test action with the highest comprehensive correlation value among the N test actions is selected as the first test action. The calculation formula of the comprehensive correlation value of each test action is as follows:
[0113] ;in, Indicates the comprehensive relevance value of the test action, , , and Respectively represent the correlation values of the four abnormal data of overvoltage, undervoltage, abnormal current and harmonic distortion with the test action. , , and The weight parameters respectively represent the correlation values of the four abnormal data of overvoltage, undervoltage, abnormal current and harmonic distortion to the test action;
[0114] According to the correlation between the first test action and N test actions, the test action with the highest correlation with the first test action is selected as the second test action; when the i+1th test action is selected based on the i-th test action, there are two test actions among the N test actions that have the same correlation with the i-th test action, then the test plan will launch two test branches in parallel; when the j+1th test action is selected based on the j-th test action, the correlation values of the N test actions and the j-th test action are all less than the correlation threshold of the test action, then the branch under the test plan ends;
[0115] In summary, the tests for each nonlinear load device include:
[0116]
[0117]
[0118]
[0119]
[0120] Assuming that there are 2 parallel branches for nonlinear load device A1, 4 parallel branches for nonlinear load device A2, 2 parallel branches for nonlinear load device A3, and 2 parallel branches for nonlinear load device A14, based on the existence of two parallel schemes among A1, A2, A3 and A4, it is determined that there are ultimately 64 parallel schemes.
[0121] It should be noted that the precision of the correlation value is taken to two decimal places.
[0122] Furthermore, the duration prediction model is used to evaluate the duration of each test plan to determine the final test plan, including:
[0123] Construct a correlation matrix for each test scenario;
[0124] The correlation matrix is input into the duration prediction model, and the duration prediction model outputs the test duration of the test plan;
[0125] Based on the test time of each test solution, a plurality of solutions are selected as alternative solutions; wherein the screening operation includes comparing the test time of each test solution and selecting a test solution whose test time is lower than a preset test time threshold, inputting the test solution whose test time is lower than the test time threshold into a generative language tool and scoring it, selecting a test solution whose score is higher than the preset score threshold, and obtaining a plurality of alternative solutions;
[0126] Several alternative solutions are randomly pushed to the target space through a random algorithm.
[0127] For example, 64 correlation matrices are constructed based on 64 parallel schemes, and the duration prediction model is used to evaluate the duration of each correlation matrix, and the test duration of each test scheme is obtained respectively; 20 test schemes whose test duration is lower than the duration threshold are selected, and then the generative language tool is used to evaluate the rationality of these 20 test schemes respectively, and the 10 test schemes with the highest evaluation scores are selected, and these 10 test schemes are used as alternatives, and these 10 alternatives are randomly pushed to the factory maintenance personnel;
[0128] Among them, the generative language tool can be chatgpt or Tongyi Qianwen;
[0129] The calculation formula of the random algorithm is as follows:
[0130] ;in, represents the probability that the ith alternative is randomly pushed, represents the evaluation score of each alternative by the generative language tool, represents the random perturbation parameter of the ith alternative;
[0131] In another embodiment of the present invention, the hierarchical diagram is pre-updated based on the evaluation of the test plan by the generative language tool; for example, the generative language tool evaluates "the cable should be checked before testing the voltage, rather than directly testing the harmonics", and the Chinese word segmentation tool, part-of-speech tagging tool and recognition tool are used to segment, tag and recognize the evaluation sentence, thereby enhancing the relevance of the test action "check the cable" to the test action "test the voltage", and reducing the relevance of the test action "test the harmonics" to the test action "test the voltage";
[0132] Furthermore, according to the connection relationship between the node entities, the association degree of each node in the hierarchical graph is updated, including:
[0133] The connection relationship between node entities is divided into strong positive association, weak positive association, no association, weak negative association and strong negative association according to the part-of-speech tagging by the recognition tool; wherein each association state is respectively set with a corresponding update value;
[0134] The association degree of the hierarchical graph nodes is updated based on the connection relationship between the node entities. The calculation formula is as follows:
[0135] ;in, represents the updated association between node i and node j in the hierarchical graph, Represents the current association between node i and node j in the hierarchical graph, Represents the updated value of the association state between the node entities mapped by node i and node j in the hierarchical graph, and Respectively represent the upper and lower thresholds of the association degree of node i and node j in the hierarchical graph, Represents the weight parameter.
[0136] For example, the correlation between the test action "check cable" and the test action "test voltage" is 0.6. Based on the feedback data or the evaluation of the generative language tool, the correlation between the test action "check cable" and the test action "test voltage" is updated, and the updated correlation between the test action "check cable" and the test action "test voltage" is 0.63, so that the next test plan is more reasonable;
[0137] Preferably, the upper threshold and lower threshold of adjacent nodes i and j in the hierarchical graph are 0.8 and 0.2 respectively.
[0138] The present invention provides an online power quality monitoring system, comprising:
[0139] The data acquisition module collects abnormal power data of nonlinear load devices in the sample space; collects abnormal test behaviors of maintenance personnel on nonlinear load devices and the duration of the abnormal test behaviors;
[0140] The prediction module constructs a correlation matrix based on each group of abnormal power data and each group of abnormal test behaviors, and constructs a training sample based on each correlation matrix and the duration of each group of abnormal test behaviors, and performs deep learning based on the training samples to obtain a duration prediction model;
[0141] A hierarchical graph construction module constructs a hierarchical graph, maps K types of nonlinear load devices to the first-level nodes of the hierarchical graph, maps M types of abnormal data to the second-level nodes of the hierarchical graph, and repeatedly maps N test actions to the third to N-level nodes of the hierarchical graph;
[0142] The monitoring and processing module builds a monitoring data processing platform based on the hierarchical graph and the duration prediction model. The monitoring data processing platform is used to process abnormal nonlinear load devices in the target space and their abnormal power data, and obtain a test plan for abnormal events in the target space.
[0143] A computing device comprising:
[0144] one or more processors;
[0145] The storage device is used to store one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the method described.
[0146] The above describes an embodiment of the present embodiment, but the present embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present embodiment, ordinary technicians in this field can also make many forms, all of which are within the protection of the present embodiment.
Claims
1. A method for online monitoring of power quality, characterized in that: include: Step S101, collecting abnormal power data of nonlinear load devices in a sample space; Collect abnormal test behaviors of maintenance personnel on nonlinear load equipment and the duration of the abnormal test behaviors; Nonlinear load devices and abnormal data are labeled and abnormal test behaviors are classified, and K types of nonlinear load devices, M types of abnormal data and N test actions are obtained; Step S102, constructing a correlation matrix based on each group of abnormal power data and each group of abnormal test behaviors, and constructing training samples based on each correlation matrix and the duration of each group of abnormal test behaviors, and performing deep learning based on the training samples to obtain a duration prediction model; Step S103, constructing a hierarchical graph, mapping K types of nonlinear load devices as first-level nodes of the hierarchical graph, mapping M types of abnormal data as second-level nodes of the hierarchical graph, and repeatedly mapping N test actions as third-level to N-level nodes of the hierarchical graph; Extract the entity relationships between nonlinear load devices, abnormal data and test actions, and construct the edges of nodes at all levels based on the entity relationships; Step S104, constructing a monitoring data processing platform based on the hierarchical graph and the duration prediction model, the monitoring data processing platform is used to process abnormal nonlinear load devices in the target space and their abnormal power data, and obtain a test plan for abnormal events in the target space.
2. The method for online monitoring of power quality according to claim 1, characterized in that: A correlation matrix is constructed based on each set of abnormal power data and each set of abnormal test behavior, including: The association matrix is an M×N matrix constructed by M types of abnormal data and N test actions. The element in the i-th row and j-th column of the association matrix represents the association between the i-th abnormal data and the j-th test action. If there is an association between the i-th abnormal data and the j-th test action, the element in the i-th row and j-th column of the association matrix is (i, j), otherwise it is (0, 0).
3. The method for online monitoring of power quality according to claim 1, characterized in that: A training sample is constructed by combining each correlation matrix and the duration of each group of abnormal test behaviors. A duration prediction model is obtained by deep learning based on the training samples, including: Normalize the duration data of each correlation matrix and each group of abnormal test behaviors; The normalized correlation matrix is used as the input of the duration prediction model; The duration prediction model includes M hidden layers, each hidden layer includes N processing units; the hidden state of the Nth processing unit of the Mth hidden layer is input into a preset classifier, and the classification space of the classifier represents the total test duration of the group of abnormal test behaviors for the corresponding abnormal power data; the duration prediction model uses the mean error variance loss function to update the weight parameters and bias parameters.
4. The method for online monitoring of power quality according to claim 1, characterized in that: Extract the entity relationships between nonlinear load devices, abnormal data and test actions, and construct the edges of nodes at all levels based on the entity relationships; According to the nonlinear load devices, abnormal data and test actions, K first-level nodes mapping K types of nonlinear load devices, M second-level nodes mapping M types of abnormal data and N third-level nodes mapping N test actions are constructed respectively; The nonlinear load equipment, abnormal data and test actions are connected through their correlation, and the correlation is used as the weight of the corresponding edge; The association degree is used to represent the probability of the i-th node and the j-th node in the hierarchical graph co-occurring in the sample space data; the calculation formula for the co-occurrence probability is as follows: ;in, represents the probability of the i-th node and the j-th node appearing together in the hierarchical graph, represents the total number of times the i-th node appears in the sample space data, represents the total number of times the jth node appears in the sample space data, Represents the total number of times the i-th node and the j-th node appear together in the sample space data.
5. The method for online monitoring of power quality according to claim 1, characterized in that: A monitoring data processing platform is constructed based on the hierarchical graph and the duration prediction model. The monitoring data processing platform is used to process abnormal nonlinear load devices in the target space and abnormal power data thereof, and obtain a test plan for abnormal events in the target space. The monitoring data processing platform includes: A data acquisition unit, used for acquiring abnormal nonlinear load devices in the target space and abnormal data of the nonlinear load devices; A test suggestion unit is used to determine a test plan for abnormal events in the target space based on abnormal nonlinear load devices in the target space, abnormal data of the corresponding nonlinear load devices, correlation between nonlinear load devices, correlation between abnormal data and test actions, and correlation between test actions, and traverse the hierarchical graph according to a preset threshold value, combined with a duration prediction model; A feedback collection unit, used to collect users' opinions on the test plan of abnormal events in the target space, and pre-process the opinions on the test plan to obtain feedback data; A feature extraction unit is used to process the feedback data, and the feature extraction unit includes: The word segmentation module uses the Chinese word segmentation tool to segment the feedback data; The part-of-speech tagging module uses the part-of-speech tagging tool to tag the feedback data after word segmentation; The recognition module uses recognition tools to perform entity recognition and relational word recognition on the feedback data after part-of-speech tagging, extract node entities related to the hierarchical graph, and build connection relationships between node entities; The entity linking module matches and links the node entities in the feedback data with the nodes in the hierarchical graph; The hierarchical graph updating module updates the association degree of each node in the hierarchical graph according to the connection relationship between the node entities.
6. The method for online monitoring of power quality according to claim 5, characterized in that: According to the preset threshold, the hierarchical graph is traversed and combined with the duration prediction model to determine the test plan for abnormal events in the target space, including: Obtaining nonlinear load devices in the target space and abnormal data of corresponding nonlinear load devices; Determine the test sequence of the nonlinear load devices according to the correlation between each nonlinear load device and other nonlinear load devices in the target space; If the correlation degree of the X nonlinear load devices to be selected is the same as that of the current nonlinear load device, the X nonlinear load devices to be selected are all used as the nonlinear load devices to be tested next, and X branches of the test scheme are obtained; Determine the order of the test actions according to the correlation between the abnormal data of the nonlinear load device of the i-th test and the N test actions and the correlation between the N test actions, combined with the preset correlation threshold of the test actions; If the correlation between the Y test actions to be selected and the current abnormal data is the same, then the Y test actions to be selected are all used as the initial test actions, and Y branches are obtained; if the correlation between the Z test actions to be selected and the current test action is the same, then the current Z abnormal data to be selected are all used as the abnormal data of the next test, and Z branches are obtained; Wherein, when the correlation between the test action to be selected and the current test action is less than the correlation threshold of the test action, the selection is stopped; Determine several test scenarios for abnormal events in the target space, and use a duration prediction model to evaluate the duration of each test scenario to determine the final test scenario.
7. The method for online monitoring of power quality according to claim 6, characterized in that: The duration prediction model is used to estimate the duration of each test plan to determine the final test plan, including: Construct a correlation matrix for each test scenario; The correlation matrix is input into the duration prediction model, and the duration prediction model outputs the test duration of the test plan; Select several options as alternatives based on the test duration of each test option; The screening includes comparing the test duration of each test solution and selecting a test solution whose duration is lower than a preset time threshold, inputting the test solutions whose duration is lower than the time threshold into a generative language tool and scoring them, selecting a test solution whose score is higher than the preset score threshold, and obtaining a number of alternative solutions; Several alternative solutions are randomly pushed to the target space through a random algorithm.
8. The method for online monitoring of power quality according to claim 5, characterized in that: According to the connection relationship between node entities, update the association degree of each node in the hierarchical graph, including: The connection relationship between node entities is divided into strong positive association, weak positive association, no association, weak negative association and strong negative association according to the part-of-speech tagging by the recognition tool; wherein each association state is respectively set with a corresponding update value; The association degree of the hierarchical graph nodes is updated based on the connection relationship between the node entities. The calculation formula is as follows: ;in, represents the updated association between node i and node j in the hierarchical graph, Represents the current association between node i and node j in the hierarchical graph, Represents the updated value of the association state between the node entities mapped by node i and node j in the hierarchical graph, and Respectively represent the upper and lower thresholds of the association degree of node i and node j in the hierarchical graph, Represents the weight parameter.
9. An online power quality monitoring system, characterized in that: Applied to the method according to any one of claims 1 to 8, comprising: The data acquisition module collects abnormal power data of nonlinear load devices in the sample space; collects abnormal test behaviors of maintenance personnel on nonlinear load devices and the duration of the abnormal test behaviors; The prediction module constructs a correlation matrix based on each group of abnormal power data and each group of abnormal test behaviors, and constructs a training sample based on each correlation matrix and the duration of each group of abnormal test behaviors, and performs deep learning based on the training samples to obtain a duration prediction model; A hierarchical graph construction module constructs a hierarchical graph, maps K types of nonlinear load devices to the first-level nodes of the hierarchical graph, maps M types of abnormal data to the second-level nodes of the hierarchical graph, and repeatedly maps N test actions to the third to N-level nodes of the hierarchical graph; The monitoring and processing module builds a monitoring data processing platform based on the hierarchical graph and the duration prediction model. The monitoring data processing platform is used to process abnormal nonlinear load devices in the target space and their abnormal power data, and obtain a test plan for abnormal events in the target space.
10. A computing device, characterized in that include: one or more processors; A storage device for storing one or more programs, when the one or more programs are executed by the one or more processors, the one or more processors implement the method as claimed in any one of claims 1 to 8.
