A power grid fault diagnosis method and system based on SE-Inception network
Through the grid fault diagnosis method based on SE-Inception network, the alarm information of grid faults is extracted and classified, which solves the problems of complex modeling and relying on manual experience in the existing technology, and achieves efficient and accurate grid fault diagnosis.
Patent Information
- Application Number
- CN202210346959.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-04-03
- Publication Date
- 2025-02-07
- Estimated Expiration
- 2042-04-03
AI Technical Summary
The existing power grid fault diagnosis technology requires the introduction of a large number of protection and equipment action rules in the early stage of modeling. The modeling process and model maintenance process are difficult, the versatility is low, and relying on manual experience is difficult to meet the growing requirements for safe and stable operation of the power grid.
The grid fault diagnosis method based on SE-Inception network is adopted. By pre-processing, feature dimensionality reduction and sorting of alarm information text during power grid faults, it is converted into a grayscale diagram of alarm information, and input the SE-Inception network model for feature extraction and classification. , realize intelligent diagnosis of power grid faults.
It improves the speed and accuracy of grid fault diagnosis, reduces labor costs, and achieves more convenient and high-accurate grid fault diagnosis without introducing a lot of knowledge describing the behavior of the protection system.
Smart Images
Figure CN114638318B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of power grid fault diagnosis, and relates to a power grid fault diagnosis method and system based on an SE-Inception network. Background Art
[0002] As one of the most important subsystems in the energy management system EMS, the Supervisory Control And Data Acquisition (SCADA) system can monitor and control the operating equipment on site to realize various functions such as data acquisition, equipment control, measurement, parameter adjustment, and various signal alarms. It has the advantages of complete information, improved efficiency, correct grasp of system operation status, accelerated decision-making, and the ability to help quickly diagnose system faults. It has now become an indispensable tool for power dispatching. The SCADA system plays an irreplaceable role in improving the reliability, safety and economic benefits of power grid operation, reducing the burden on dispatchers, realizing the automation and modernization of power dispatching, and improving the efficiency and level of dispatching. When a fault occurs in the power grid, the protection device related to the faulty component will cause the protection to operate due to the detection of fault signs, drive the corresponding circuit breaker to trip, and report the alarm information of the protection action and circuit breaker tripping to the SCADA system of the dispatching center in real time.
[0003] Power grid fault diagnosis is the basis for accident analysis and handling. It is an important application for realizing the self-healing function of smart grid and is of great significance to improving the stability of power grid.
[0004] Existing fault diagnosis technologies need to introduce a large number of protection and equipment action rules in the early stage of modeling, and the modeling process and model maintenance process are difficult and have low versatility. When using the alarm information generated during a fault to diagnose a fault, a large amount of alarm information includes both switch quantity information such as protection action and circuit breaker tripping, as well as non-fault critical information such as device abnormality, voltage exceeding the limit, control circuit disconnection, and energy storage in place. Only the dispatching and operation personnel can diagnose power grid faults based on experience, which cannot meet the growing requirements for safe and stable operation of the power grid. How to use relevant technologies to achieve efficient and accurate power grid fault diagnosis has gradually become a research hotspot.
[0005] With the rapid development of computer technology and artificial intelligence technologies such as deep learning, effective use of data for power grid fault diagnosis has become an increasingly hot research topic due to the excellent data feature extraction and learning capabilities of deep learning. Summary of the invention
[0006] In order to solve the deficiencies in the prior art, the purpose of the present invention is to provide a power grid fault diagnosis method and system based on SE-Inception network, which can effectively utilize alarm information data, improve the speed and accuracy of power grid fault diagnosis, and realize the intelligentization of power grid fault diagnosis.
[0007] In order to achieve the above object, the present invention adopts the following technical solution:
[0008] A power grid fault diagnosis method based on SE-Inception network includes the following steps:
[0009] S1. Collect the alarm information text when the power grid fails, and pre-process the alarm information text;
[0010] S2, performing feature dimension reduction and sorting on the pre-processed alarm information text to obtain feature information text, encoding and normalizing the feature information text corresponding to the alarm event to form an alarm event dictionary;
[0011] S3, converting the alarm information of different fault types into an alarm event coding matrix according to the alarm event dictionary, and converting the alarm event coding matrix into an alarm information grayscale map;
[0012] S4. Construct a power grid fault diagnosis model based on SE-Inception network;
[0013] S5. Input the grayscale image of the alarm information into the power grid fault diagnosis model, obtain the power grid fault type judgment result, and realize the power grid fault diagnosis.
[0014] The present invention further includes the following preferred embodiments:
[0015] Preferably, the step S1 comprises:
[0016] S1.1. The SCADA system obtains in real time the sequence of events generated by the monitoring equipment when an accident occurs in the power grid, the operating status exceeds the limit, or the equipment is abnormal;
[0017] The event sequence record includes alarm information of protection action, circuit breaker tripping, voltage over-limit and equipment abnormality;
[0018] In the event sequence record, the alarm information is recorded in chronological order to generate an alarm information text when a power grid failure occurs;
[0019] S1.2. Collect the alarm information text from the SCADA system and simulation software, and perform deduplication and standardization processing on the alarm information text.
[0020] Preferably, step S2 comprises:
[0021] S2.1. Perform feature dimension reduction processing on the alarm information pre-processed in S1: Perform feature initial screening on the alarm information to remove irrelevant features and equivalent features;
[0022] S2.2, sort the remaining candidate features according to the principle of maximum relevance and minimum redundancy;
[0023] S2.3. Encode and normalize the characteristic information texts sorted in S2.2 corresponding to the alarm events to form an alarm event dictionary.
[0024] Preferably, the remaining candidate features are sorted according to the principle of maximum relevance and minimum redundancy in S2.2, and the specific steps are as follows:
[0025] 1) Sort by the maximum correlation between features and events. The formula is as follows:
[0026] maxD(x i ,y),x i ∈S
[0027] Where S is the remaining candidate feature set; x i is the i-th feature information; y is the line fault event; D is the normalized correlation calculation result;
[0028] 2) Calculate the redundancy of the feature information sorted by the maximum correlation, and based on the maximum correlation in 1)
[0029] Sorting results are sorted by maximum correlation and minimum redundancy. The formula is as follows:
[0030]
[0031] In the formula, is feature x i The difference between the correlation and redundancy calculations used for sorting; R(x i ,y) is the normalized redundancy calculation result; It is the result of sorting the features with maximum relevance and minimum redundancy.
[0032] Preferably, the step S3 specifically includes:
[0033] S3.1. Screen and sort the characteristic information texts in the alarm information of different fault types according to step S2;
[0034] S3.2. Take the top 20 feature information texts and convert them into a 4×5 order alarm event coding matrix according to the alarm event dictionary:
[0035] D(A)=(a ij ) 4×5 ;
[0036] Among them, a ij Indicates the corresponding codes of different alarm events;
[0037] The priority of the feature information corresponding to the position of each element in the encoding matrix decreases from left to right and from top to bottom;
[0038] S3.3. Convert the alarm event coding matrix into an alarm information grayscale image as the input sample of the power grid fault diagnosis model based on the SE-Inception network.
[0039] Preferably, the SE-Inception network in step S4 includes two convolutional layers, two pooling layers, one SE-Inception layer and one fully connected layer;
[0040] The constructed power grid fault diagnosis model based on SE-Inception network takes the grayscale image of alarm information as input, extracts features through convolutional layer and SE-Inception layer to output feature map, compresses the feature map dimension through pooling layer, combines the fault category features output by pooling layer through fully connected layer, and then sends them to Softmax classifier to classify different fault types.
[0041] Preferably, the SE-Inception layer includes several SE-Inception modules;
[0042] The SE-Inception module includes an SE module and an Inception module;
[0043] The SE module learns different weight coefficients for each feature channel, thereby optimizing network performance; the convolution kernels of each feature channel are different, resulting in different learned features;
[0044] The Inception module widens the network by horizontally arranging convolutional layers and performs multi-feature fusion;
[0045] Preferably, the SE module obtains the importance of each feature channel through learning, finds the channel with important features and assigns a larger weight to it, and assigns a smaller weight to the channel with less impact on the classification result;
[0046] The specific operations of the SE module are divided into the Squeeze phase and the Excitation phase;
[0047] In the squeeze stage, the global feature information of the image is obtained through global average pooling, and the output of the entire channel is obtained by expanding the receptive field;
[0048] In the Excitation stage, the importance of each feature channel is learned autonomously through the FC fully connected layer, different weights are assigned according to the importance, and then these weights are multiplied with the original feature map to obtain the final feature display.
[0049] Preferably, the step S5 specifically includes the following steps:
[0050] S5.1. The converted grayscale image of the alarm information is randomly divided into a training set and a test set according to a certain proportion;
[0051] The samples in the training set and the test set are provided with corresponding labels;
[0052] The labels include: single-phase ground short circuit, two-phase short circuit, two-phase ground short circuit, three-phase short circuit;
[0053] S5.2. Input the training set into the power grid fault diagnosis model for training, extract the grayscale features of the alarm information, and obtain the trained fault type discrimination model after repeated model parameter optimization and hyperparameter adjustment;
[0054] S5.3, input the test set samples into the discriminant model obtained after training with the training set and repeated parameter adjustment, and finally output the confusion matrix;
[0055] The confusion matrix includes the accuracy, precision, recall and F1 value evaluation indicators of fault type discrimination and the fault type discrimination results.
[0056] A power grid fault diagnosis system based on an SE-Inception network, the system is used to implement the power grid fault diagnosis method based on an SE-Inception network.
[0057] The beneficial effects of the present invention are as follows:
[0058] The present invention utilizes a deep learning network to transform power grid fault diagnosis into a graphic classification problem. By performing digitization and graphic processing on the alarm information text, the SE-Inception network with excellent performance in processing image classification problems can be used to perform power grid fault diagnosis. The constructed model can autonomously learn the fault features in the alarm information and directly output the fault diagnosis results, which greatly saves labor costs.
[0059] The present invention does not need to introduce a large amount of knowledge describing the behavior of the protection system. It only needs to extract the graphic features corresponding to different fault types from the alarm information text to achieve more convenient and more accurate power grid fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS
[0060] Figure 1This is a flow chart of a power grid fault diagnosis method based on a SE-Inception network of the present invention;
[0061] Figure 2 This is a flowchart of the implementation of the power grid fault diagnosis method based on the SE-Inception network of the present invention;
[0062] Figure 3 This is a structural diagram of the implementation of the power grid fault diagnosis model of the present invention;
[0063] Figure 4 This is an implementation structure diagram of the SE-Inception module of the present invention;
[0064] Figure 5 This is a schematic diagram of the Softmax classification principle of the present invention. DETAILED DESCRIPTION
[0065] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.
[0066] like Figure 1-2 As shown, embodiment 1 of the present invention provides a power grid fault diagnosis method based on the SE-Inception network, which is applicable to a fault diagnosis system whose fault information source is power grid fault alarm information. In a preferred but non-limiting embodiment of the present invention, the method includes the following steps:
[0067] S1. Collect the alarm information text when the power grid fails from the SCADA system and pre-process the alarm information text, including:
[0068] S1.1. The SCADA system obtains in real time the sequence of events (SOE) generated by the monitoring equipment when an accident occurs in the power grid, the operating status exceeds the limit, or the equipment is abnormal;
[0069] Further preferably, the event sequence record includes alarm information such as protection action, circuit breaker tripping, voltage exceeding limit and equipment abnormality.
[0070] In the event sequence record, the alarm information is recorded in chronological order to generate an alarm information text when a power grid failure occurs, which is an alarm information set.
[0071] Generate alarm information samples in batches through simulation software;
[0072] S1.2. Collect the alarm information text from the SCADA system, remove duplicates and standardize the alarm information text, and construct a data sample set for the power grid fault diagnosis model.
[0073] S1.2. Collect the alarm information text from the SCADA system and the simulation software TS2000, and perform deduplication and standardization processing on the alarm information text.
[0074] It also includes data obtained from the SCADA system and the simulation software TS2000, including both simulation-generated and actual collected data.
[0075] S2. Perform feature dimension reduction and sorting on the preprocessed alarm information text to obtain feature information text, encode and normalize the feature information text corresponding to the alarm event to form an alarm event dictionary, including:
[0076] S2.1. Perform feature dimension reduction on the alarm information pre-processed in S1:
[0077] Perform a preliminary feature screening on the alarm information to remove irrelevant features and some equivalent features;
[0078] S2.2, sort the remaining candidate features according to the principle of maximum relevance and minimum redundancy;
[0079] Further preferably, the specific steps are as follows:
[0080] 1) Sort by the maximum correlation between features and events, that is, calculate maxD r , then based on maxD r Sort by
[0081] The formula is as follows:
[0082] maxD r (x i ,y),x i ∈S
[0083] Where S is the remaining candidate feature set; x i is the i-th feature information; y is the line fault event; D is the normalized correlation calculation result;
[0084] 2) Calculate the redundancy of the feature information sorted by the maximum correlation, and based on the maximum correlation in 1)
[0085] Sorting results are sorted by maximum correlation and minimum redundancy. The formula is as follows:
[0086]
[0087] In the formula, is feature x i The difference between the correlation and redundancy calculations used for sorting; R(x i ,y) is the normalized redundancy calculation result; It is the result of sorting the features with maximum relevance and minimum redundancy.
[0088] S2.3. Encode and normalize the characteristic information texts sorted in S2.2 corresponding to the alarm events to form an alarm event dictionary.
[0089] S3, converting the alarm information of different fault types into an alarm event coding matrix, and converting the alarm event coding matrix into an alarm information grayscale map, including:
[0090] S3.1. Screen and sort the characteristic information texts in the alarm information of different fault types according to step S2;
[0091] S3.2. Through the analysis of the alarm information, it is found that the top 20 alarm information with the highest correlation with the fault event are sufficient to reflect the fault type characteristics contained in the alarm information, so the top 20 are selected, and the 4*5 matrix is convenient for generating pictures.
[0092] Take the top 20 feature information texts and convert them into a 4×5-order alarm event coding matrix according to the alarm event dictionary:
[0093] D(A)=(a ij ) 4×5 ;
[0094] Among them, a ij Indicates the corresponding codes of different alarm events;
[0095] The priority of the feature information corresponding to the position of each element in the encoding matrix decreases from left to right and from top to bottom.
[0096] S3.3. Convert the alarm event coding matrix into an alarm information grayscale image as the input sample of the power grid fault diagnosis model based on the SE-Inception network.
[0097] S3.3 can be implemented using Matlab or Python.
[0098] S4. Build a power grid fault diagnosis model based on SE-Inception network:
[0099] More preferably, if Figure 3 As shown, the SE-Inception network includes two convolutional layers, two pooling layers, one SE-Inception layer and one fully connected layer;
[0100] The power grid fault diagnosis model based on SE-Inception network inputs the grayscale image of alarm information, extracts features through the convolution layer and SE-Inception layer to output the feature map, compresses the feature map dimension through the pooling layer, and combines the fault category features output by the pooling layer in the fully connected layer, and then sends them to the Softmax classifier to classify different fault types.
[0101] The SE-Inception layer includes several SE-Inception modules;
[0102] like Figure 4 As shown, the SE-Inception module is composed of a Squeeze-and-Excitation (SE) module and an Inception module;
[0103] The SE module is similar to the attention mechanism in the time series model, which learns different weight coefficients for each channel to optimize the network performance.
[0104] The convolution kernels of each channel are different, resulting in different learned features;
[0105] The SE module obtains the importance of each feature channel through learning, finds the channels with important features and assigns them a larger weight, and assigns a smaller weight to the channels with less impact on the classification results;
[0106] The Inception module makes the network wider by horizontally arranging the convolutional layers, and can use a shallower network to obtain better model capabilities and perform multi-feature fusion, while being easier to train.
[0107] The specific operations of the SE module are mainly divided into the Squeeze stage and the Excitation stage;
[0108] In the squeeze stage, the global feature information of the image is obtained mainly through global average pooling, and the output of the entire channel is obtained by expanding the receptive field;
[0109] In the Excitation stage, the importance of each feature channel is learned autonomously through the FC fully connected layer, different weights are assigned according to the importance, and then these weights are multiplied with the original feature map to obtain the final feature display.
[0110] The classification principle of Softmax classifier is as follows Figure 5 As shown, Figure 5 Medium X i Represents the one-dimensional feature vector output by the fully connected layer, r i Represents the output feature category corresponding vector, Softmax(r i) represents the probability corresponding to the output feature category, where the feature category corresponding to the element with the largest output probability is the discrimination result.
[0111] S5, inputting the alarm information grayscale image into the power grid fault diagnosis model, obtaining the power grid fault type discrimination result, and realizing power grid fault diagnosis, specifically including the following steps:
[0112] S5.1. The converted large amount of alarm information grayscale images are randomly divided into a training set and a test set in a ratio of 8:2;
[0113] Further preferably, the training set and test set samples are both labeled accordingly;
[0114] The labels include: single-phase ground short circuit, two-phase short circuit, two-phase ground short circuit, and three-phase short circuit.
[0115] S5.2. Input the training set into the power grid fault diagnosis model for training, extract the grayscale features of the alarm information, and obtain the trained fault type discrimination model after repeated model parameter optimization and hyperparameter adjustment;
[0116] S5.3, input the test set samples into the discriminant model obtained after training with the training set and repeated parameter adjustment, and finally output the confusion matrix;
[0117] The confusion matrix includes the accuracy, precision, recall and F1 value evaluation indicators of fault type discrimination and the fault type discrimination results.
[0118] The present invention provides a power grid fault diagnosis system based on SE-Inception network, which is used to implement the above-mentioned power grid fault diagnosis method based on SE-Inception network.
[0119] The beneficial effects of the present invention are as follows:
[0120] The present invention utilizes a deep learning network to transform power grid fault diagnosis into a graphic classification problem. By performing digitization and graphic processing on the alarm information text, the SE-Inception network with excellent performance in processing image classification problems can be used to perform power grid fault diagnosis. The constructed model can autonomously learn the fault features in the alarm information and directly output the fault diagnosis results, which greatly saves labor costs.
[0121] The present invention does not need to introduce a large amount of knowledge describing the behavior of the protection system. It only needs to extract the graphic features corresponding to different fault types from the alarm information text to achieve more convenient and more accurate power grid fault diagnosis.
[0122] The applicant of the present invention has made a detailed explanation and description of the implementation examples of the present invention in conjunction with the drawings in the specification. However, those skilled in the art should understand that the above implementation examples are only preferred implementation schemes of the present invention, and the detailed description is only to help readers better understand the spirit of the present invention, but not to limit the scope of protection of the present invention. On the contrary, any improvements or modifications based on the inventive spirit of the present invention should fall within the scope of protection of the present invention.
Claims
1. A power grid fault diagnosis method based on SE-Inception network, characterized by: The method comprises the following steps: S1. Collect the alarm information text when the power grid fails, and pre-process the alarm information text; S2, performing feature dimension reduction and sorting on the pre-processed alarm information text to obtain feature information text, encoding and normalizing the feature information text corresponding to the alarm event to form an alarm event dictionary; The S2 includes: S2.
1. Perform feature dimension reduction processing on the alarm information pre-processed in S1: Perform feature initial screening on the alarm information to remove irrelevant features and equivalent features; S2.
2. Sort the remaining candidate features according to the principle of maximum relevance and minimum redundancy. The specific steps are as follows: 1) Sort by the maximum correlation between features and events. The formula is as follows: maxD(x i ,y),x i ∈S Where S is the remaining candidate feature set; x i is the i-th feature information; y is the line fault event; D is the normalized correlation calculation result; 2) Calculate the redundancy of the feature information sorted by maximum relevance, and perform maximum relevance minimum redundancy sorting based on the maximum relevance sorting result in 1). The formula is as follows: In the formula, is feature x i The difference between the correlation and redundancy calculations used for sorting; R(x i ,y) is the normalized redundancy calculation result; It is the result of sorting the features with maximum relevance and minimum redundancy; S2.3, encoding and normalizing the characteristic information texts sorted in S2.2 corresponding to the alarm events to form an alarm event dictionary; S3, converting the alarm information of different fault types into an alarm event coding matrix according to the alarm event dictionary, and converting the alarm event coding matrix into an alarm information grayscale map; The S3 specifically includes: S3.
1. Screen and sort the characteristic information texts in the alarm information of different fault types according to step S2; S3.
2. Take the top 20 feature information texts and convert them into a 4×5 order alarm event coding matrix according to the alarm event dictionary: D(A)=(a ij ) 4×5 ; Among them, a ij Indicates the corresponding codes of different alarm events; The priority of the feature information corresponding to the position of each element in the encoding matrix decreases from left to right and from top to bottom; S3.3, converting the alarm event coding matrix into an alarm information grayscale image as an input sample of the power grid fault diagnosis model based on the SE-Inception network; S4. Construct a power grid fault diagnosis model based on SE-Inception network; The SE-Inception network described in S4 includes two convolutional layers, two pooling layers, one SE-Inception layer and one fully connected layer; The constructed power grid fault diagnosis model based on SE-Inception network takes the grayscale image of alarm information as input, extracts features through convolutional layer and SE-Inception layer to output feature map, compresses the feature map dimension through pooling layer, combines the fault category features output by pooling layer through fully connected layer, and then sends them to Softmax classifier to classify different fault types. S5. Input the grayscale image of the alarm information into the power grid fault diagnosis model, obtain the power grid fault type judgment result, and realize the power grid fault diagnosis.
2. The power grid fault diagnosis method based on SE-Inception network according to claim 1 is characterized in that: The step S1 comprises: S1.
1. The SCADA system obtains in real time the sequence of events generated by the monitoring equipment when an accident occurs in the power grid, the operating status exceeds the limit, or the equipment is abnormal; The event sequence record includes alarm information of protection action, circuit breaker tripping, voltage over-limit and equipment abnormality; In the event sequence record, the alarm information is recorded in chronological order to generate an alarm information text when a power grid failure occurs; S1.
2. Collect the alarm information text from the SCADA system and simulation software, and perform deduplication and standardization processing on the alarm information text.
3. The power grid fault diagnosis method based on SE-Inception network according to claim 1 is characterized in that: The SE-Inception layer includes several SE-Inception modules; The SE-Inception module includes an SE module and an Inception module; The SE module learns different weight coefficients for each feature channel, thereby optimizing network performance; The convolution kernels of each feature channel are different, resulting in different learned features; The Inception module widens the network by horizontally arranging convolutional layers and performs multi-feature fusion.
4. The power grid fault diagnosis method based on SE-Inception network according to claim 3 is characterized in that: The SE module obtains the importance of each feature channel through learning, finds the channels with important features and assigns them a larger weight, and assigns a smaller weight to the channels with less impact on the classification results; The specific operations of the SE module are divided into the Squeeze phase and the Excitation phase; In the squeeze stage, the global feature information of the image is obtained through global average pooling, and the output of the entire channel is obtained by expanding the receptive field; In the Excitation stage, the importance of each feature channel is learned autonomously through the FC fully connected layer, different weights are assigned according to the importance, and then these weights are multiplied with the original feature map to obtain the final feature display.
5. The power grid fault diagnosis method based on SE-Inception network according to claim 1 is characterized in that: The step S5 specifically comprises the following steps: S5.
1. The converted grayscale image of the alarm information is randomly divided into a training set and a test set according to a certain proportion; The samples in the training set and the test set are provided with corresponding labels; The labels include: single-phase ground short circuit, two-phase short circuit, two-phase ground short circuit, three-phase short circuit; S5.
2. Input the training set into the power grid fault diagnosis model for training, extract the grayscale features of the alarm information, and obtain the trained fault type discrimination model after repeated model parameter optimization and hyperparameter adjustment; S5.3, input the test set samples into the discriminant model obtained after training with the training set and repeated parameter adjustment, and finally output the confusion matrix; The confusion matrix includes the accuracy, precision, recall and F1 value evaluation indicators of fault type discrimination and the fault type discrimination results.
6. A power grid fault diagnosis system based on SE-Inception network, characterized by: The system is used to implement the power grid fault diagnosis method based on the SE-Inception network as described in any one of claims 1 to 5.