Method and system for identifying bad measurement data of high-voltage direct-current power transmission system

Abstract

The invention discloses a method and a system for identifying bad measurement data of a high-voltage direct-current power transmission system. Firstly, measurement data of an optical fiber current transformer, a direct-current voltage divider and a zero-flux current transformer of the high-voltage direct-current power transmission system are collected; then bad samples of the measurement data are extracted for generative adversarial training, and a sample set is balanced; carrying out Spark parallelization decision tree model training on the balanced sample set; summarizing the plurality of decision trees obtained by training to form a random forest bad data identification model; and utilizing the bad data model to identify bad measurement data of each optical fiber current transformer, the direct current voltage divider and the zero-flux current transformer.

Description

Technical Field

[0001] The present invention relates to the technical field of power system protection and control, and in particular to a method and system for identifying bad measurement data of a high-voltage direct current transmission system. Background Art

[0002] DC transmission has been widely used in the fields of regional power grid interconnection and new energy transmission due to its advantages of low loss, short-circuit current limitation, and no limitation on synchronous operation stability. The safe and stable operation of DC transmission systems depends on complex DC control and protection systems, and DC measurement equipment represented by optical fiber current transformers, DC voltage dividers, and zero-flux current transformers are the basis of DC transmission control and protection systems. The structure or principle of DC measurement equipment is different from that of AC measurement equipment, and due to its short application history, the stability, accuracy, and drift issues in actual projects still need to be resolved. Although self-check logic is configured in DC measurement equipment to determine whether measurement anomalies or failures have occurred, field incidents of DC locking due to errors such as data drift or sudden drops still occur from time to time.

[0003] In recent years, the reliability of DC measurement equipment has attracted attention. Existing research mainly focuses on the state assessment, fault diagnosis and reliability improvement of DC measurement equipment such as optical fiber current transformers. However, there is currently no technical method to fundamentally eliminate the defects of DC measurement equipment. Some researchers have paid attention to the DC lockout problem caused by abnormal DC measurement data in converter stations, but they mainly focus on the analysis of the causes of DC lockout accidents. There are few studies on identifying abnormal DC measurement data from the perspective of DC control and protection. UHV DC systems usually adopt a dual or three-out-of-two redundant configuration. However, the dual configuration fails to completely solve the problem of false operation caused by abnormal data of a single set of equipment, and three-out-of-two has problems such as high cost and high operation and maintenance pressure. Some studies construct causal networks based on the relationship between electrical quantities and causal rule associations to diagnose data anomalies of DC measurement equipment. However, due to the measurement accuracy problem, there is a certain error in the corresponding relationship between electrical quantities, and there is still a possibility of misdiagnosis, and the real-time nature of the diagnosis may be difficult to meet the requirements.

[0004] There are a series of bad data identification methods based on residual search, deep learning, K-means, dynamic reasoning chain, etc. in AC power system. However, the time constant of DC transmission system is small, and the control and protection system is usually required to start and respond within 5ms. Therefore, the bad data identification of DC transmission system needs to have fast detection capability, and existing methods such as neural network and expert system are difficult to meet the requirements of DC control and protection system in terms of real-time performance. Researchers used linearized residual equations to realize bad data identification, but the identification effect is not good and the computational efficiency is poor, so it cannot be directly applied to data identification of DC system. In addition, since bad data accounts for a minority in the measurement data set, the identification process belongs to an unbalanced classification problem. When the traditional random forest algorithm is used for unbalanced classification problems, it is difficult to extract minority class samples during the sampling process, resulting in poor classification effect of the trained decision tree.

[0005] Prior art document 1 (CN116756146, application publication date 2023.09.15) discloses an engineering management data processing system and method, including a data input device, a data cleaning module, a judgment module, a data labeling module, a data processing unit, a feature value management module, and a data storage unit. The present invention classifies data by forming a random forest through multiple decision trees, associates the data storage unit with the index, and establishes different regional blockchains. The system uses Spark as a computing engine to adopt distributed operations, and can simultaneously establish and analyze multiple models.

[0006] Prior art document 2 (CN 105678398 A, application publication date 2016.06.15) discloses a power load forecasting method based on big data technology and a research and application system based on the method. With the support of big data technology, it integrates a variety of mature open source products to form a data source, data integration, data storage, data calculation, data analysis, implementation of power load characteristics analysis and power load forecasting analysis. Prior art document 3 (CN 113570192 A, application publication date 2021.10.29) discloses an agricultural social intelligent service system based on big data, which is built on a cloud service PaaS platform, and includes a data acquisition system, a data storage and computing system, a data analysis system and an application system that all interact with data through the platform; the data acquisition system includes a real-time acquisition system for structured data and a real-time acquisition system for unstructured log data; the data storage and computing system includes a Hadoop distributed computing platform, and the data analysis system includes a machine learning model for predicting each link.

[0007] The existing technical documents 1-3 often only use Spark applications as computing engines for modeling and analysis. The proposed method also fully considers the combination of the "Map-Reduce" concept and algorithm of the Spark framework. The "Map" operation will first read the HDFS file and create a new RDD, and execute the Cache instruction locally to write it to the RDD cache. Subsequently, multiple samplings are performed to calculate the information gain rate at each split point based on feature b and threshold. The "Reduce" operation will summarize the identification results of each decision tree to complete the parallel training under the Spark framework.

[0008] By integrating the "Map-Reduce" concept with the algorithm, the present invention further explores the Spark computing efficiency and achieves the technical requirement that the process from fault to processing of bad data identification in the DC transmission system should be within 5ms.

[0009] Compared with technical document 3, the present invention adopts Wasserstein distance instead of JS divergence as the cost function, and adds a penalty term to realize Lipschitz constraint. Thus, sample expansion and data feature extraction are realized under the conditions of randomness, mutation and small number of samples of bad data of DC measurement equipment. At the same time, the problems of gradient vanishing, gradient explosion and non-convergence that often occur in the existing generative adversarial network training process are solved. Summary of the invention

[0010] The present invention introduces the Spark architecture to meet the requirements of rapid identification of bad data in the DC power transmission system, and optimizes the data distribution of the DC measurement data set through a generative adversarial network, changes the proportion of bad data in the data sample, solves the problem of unbalanced classification, and realizes the identification of bad data of DC measurement equipment through model parallel training. The method can quickly and accurately identify bad data in the DC power transmission system, and has good acceleration ratio and scalability, which helps to provide a safe and stable operation level of the DC power transmission system and its sending and receiving end AC power grid.

[0011] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0012] The present invention provides a method for identifying bad measurement data of a high voltage direct current transmission system, which is characterized by comprising the following steps:

[0013] The historical measurement data of the optical fiber current transformer, the DC voltage divider and the zero-flux current transformer of the high-voltage direct current transmission system are collected; the original samples of bad data of the historical measurement data are extracted for generative adversarial training, and the sample set is balanced so that the number of bad data samples accounts for 50% of the total sample set; the Spark parallelized decision tree model is trained on the balanced sample set; the multiple decision trees obtained by training are aggregated to form a random forest bad data identification model; the measurement data of the optical fiber current transformer, the DC voltage divider and the zero-flux current transformer are collected, and the measurement data are input into the bad data identification model to identify the bad measurement data.

[0014] Preferably, the bad data original sample is extracted according to the following steps:

[0015] The original data sets of the optical fiber current transformer, DC voltage divider, and zero flux current transformer are processed as follows:

[0016]

[0017] Where P foct,0 , P dcvd,0 , P zfct,0 They are the original data sets of optical fiber current transformer, DC voltage divider, and zero flux current transformer, respectively. foct,nor , P dcvd,nor , P zfct,nor They are normalized data sets of fiber optic current transformer, DC voltage divider, and zero flux current transformer respectively; data that differs from the average value by 3 standard deviations is defined as bad data, and data that differs from the average value by 1 standard deviation is defined as normal measurement data; the original samples of bad data that differ from the average value by 3 standard deviations are extracted to form the original bad data set for generative adversarial training.

[0018] Preferably, the bad data generation adversarial training method for the high voltage direct current transmission system is:

[0019] The Gaussian noise of the same dimension as the original sample of bad data is input into the generator to output the generated sample of bad data; the generator loss function is used to quantify the inconsistency between the generated sample of bad data and the original sample of bad data; the generated sample of bad data and the original sample of bad data are input into the discriminator to construct the discriminator loss function; game adversarial training is performed; the generated sample of bad data is merged with the normalized data set.

[0020] Preferably, the generator loss function is determined as follows:

[0021] L G =1-D(G(z))

[0022] Where: L Gis the generator loss function, G is the generator function, D is the discriminator function, and z is the input noise data vector.

[0023] Preferably, the discriminator loss function is determined as follows:

[0024]

[0025] Where: L D is the discriminator loss function, E is the expected function, P r is the distribution of the original data x input to the generator, P g is the noise distribution, and x is determined by the type of bad data device that needs to be trained. or They correspond to optical fiber current transformer, DC transformer and zero flux transformer respectively.

[0026] Preferably, the objective function of game adversarial training is determined as follows:

[0027]

[0028] In the formula, ||·|| p is the p-norm, is the gradient operator, λ is the penalty coefficient, For the generated data, ζ follows a uniform distribution between [0,1]; is the distribution along the bad data P r (·) and the generated bad data distribution P g (·) The straight lines between the sampling points are uniformly sampled.

[0029] Preferably, the Spark parallel decision tree model training method for the high voltage direct current transmission system is:

[0030] P foct,1 , P dcvd,1 , P zfct,1 Perform n random sampling to obtain C foct , C dcvd , C zfct ; Calculate the information entropy T(X) of the balanced DC measurement data m ), calculated as follows:

[0031]

[0032] In the formula, X m is the DC measurement data, m=1, 2, 3 represent the balanced DC measurement data of the optical fiber current transformer, DC voltage divider, and zero flux current transformer respectively; a k is the ratio of bad data or normal measurement data in the sample, k = 0 represents bad data, k = 1 represents normal measurement data;

[0033] Taking feature b as the split point, establish a set {b 1 ,b 2 ,b 3 ,...,b j The segmentation point set S after} = {s i =(b i +b i+1 ) / 2|1≤i≤j-1}; Calculate the information gain Φ(X m ,b,s i );Calculate the information gain rate of the measured data O(X m ,b,s i );

[0034] The data classification result of the decision tree is obtained. The classification of data by a single decision tree is based on the gain rate of feature b O(X m ,b,s i ) corresponds to the maximum information gain rate O(X m ,b,s b ) Split point s b Select;

[0035] Set the threshold ε and determine the accuracy of recognition. If the information gain rate is not greater than ε, the feature classification ability is determined to be unqualified; when the gain rate O(X m ,b,s i ) corresponds to the maximum information gain rate O(X m ,b,s b ) is greater than ε, b i ≤s b , b i >s b The data are classified into each category;

[0036] From b i ≤s b Select another measurement data with the largest information gain rate As a classification point, The data are regarded as bad data, otherwise they are normal measurement data;

[0037] The above process is repeated recursively until the data classified at a leaf node all belong to the same category.

[0038] Preferably, the information gain Φ(X m ,b,s i ) is calculated as follows:

[0039]

[0040] Where: |Xm | indicates X m The amount of data; Represents feature b i ≤s i , b i >s i The weight of

[0041] The information gain rate of the measured data is O(X m ,b,s i ) is calculated as follows:

[0042]

[0043] Where: I(b) represents the characteristic intrinsic value.

[0044] Preferably, in step S4, the method for constructing the random forest bad data identification model is:

[0045] C-based foct , C dcvd , C zfct It can be obtained by n foct , n dcvd , n zfct Decision Tree Random forest bad data identification model F foct , F dcvd , F zfct :

[0046]

[0047] Each decision tree is combined and aggregated to form a random forest. The final identification result of the measurement data is obtained based on the voting of each decision tree. When 50% of the decision tree identification results are bad data, the final identification result is defined as bad data.

[0048] The present invention also provides a system for identifying bad measurement data of a high-voltage direct current transmission system. The system is a system used by the above-mentioned method for identifying bad measurement data of a high-voltage direct current transmission system, and is characterized in that it comprises:

[0049] Preprocessing module, used for data normalization and construction of bad data sets;

[0050] Sample reconstruction module, used to reconstruct and generate bad data samples and balance the sample set;

[0051] Model training module, used to perform random forest parallel training and establish a bad data identification model;

[0052] The discrimination module is used to discriminate whether the DC measurement data is bad data, inputs the measurement data to be discriminated, and outputs the identification result voted by the bad data identification model.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] 1. The Spark architecture is introduced to reduce the access delay of disk reading and writing, and multiple nodes are automatically executed in parallel for decision tree training and result aggregation, overcoming the problem of insufficient speed in identifying bad data of DC measurement equipment.

[0055] 2. Compared with AC measuring equipment, the identification of bad data of DC measuring equipment has the problem that its data features are difficult to extract due to its random distribution and rapid mutation. The present invention combines the generative adversarial network to realize sample expansion and feature extraction of bad data, effectively solving the influence of the bad data characteristics of DC measuring equipment on the identification effect.

[0056] 3. The present invention can improve the rapidity of identifying bad data in DC measurement data while maintaining good identification effect, and has good acceleration ratio and scalability, which is of great significance to ensuring the safe operation of the DC system.

[0057] 4. The implementation method of the present invention is clear. It only needs to collect the real-time operation data of the current high-voltage direct current transmission system and process the data in real time through the random forest algorithm. It is easy to implement and has strong economy and practicality. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 The present invention discloses a method for identifying bad measurement data of a high-voltage direct current transmission system;

[0059] Figure 2 It is a comparative diagram of the effects of the present invention;

[0060] Figure 3 It is a flow chart of a bad data generation adversarial training method for a high voltage direct current transmission system in the present invention;

[0061] Figure 4 It is a schematic diagram of the classification process of bad data of the decision tree in the present invention;

[0062] Figure 5 It is a schematic diagram of the parallel training process based on the Spark framework of the present invention. DETAILED DESCRIPTION

[0063] The following will be combined with the drawings in the embodiments of 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.

[0064] like Figure 1 As shown, embodiment 1 of the present invention provides a method for identifying bad measurement data of a high voltage direct current transmission system, comprising the following steps:

[0065] S1, collect measurement data of optical fiber current transformer, DC voltage divider and zero flux current transformer of high voltage DC transmission system;

[0066] S2, extract the original samples of bad samples of the measurement data for generative adversarial training and balance the sample set;

[0067] S3, perform Spark parallel decision tree model training on the balanced sample set;

[0068] S4, aggregating the multiple decision trees obtained through training to form a random forest bad data identification model;

[0069] S5. Using the bad data model, identify the bad measurement data of each optical fiber current transformer, DC voltage divider, and zero flux current transformer.

[0070] In specific implementation, in step S2, bad data is extracted according to the following steps:

[0071] S2.1. Process the original data sets of the optical fiber current transformer, DC voltage divider, and zero flux current transformer as follows:

[0072]

[0073] Where P foct,0 , P dcvd,0 , P zfct,0 is the original data set of optical fiber current transformer, DC voltage divider, and zero flux current transformer, P foct,nor , P dcvd,nor , P zfct,nor Normalized data sets for fiber optic current transformer, DC voltage divider, and zero flux current transformer;

[0074] S2.2. Data that differs from the mean by 3 standard deviations is defined as bad data, data that differs from the mean by 1 standard deviation is defined as normal measurement data, and data that differs from the mean by 1 standard deviation to 3 standard deviations includes both normal measurement data and bad data and is not further classified;

[0075] S2.3. Extract the original samples of bad data that differ from the mean value by 3 standard deviations for generative adversarial training.

[0076] like Figure 3 As shown, in the specific implementation, in step S2.3, the bad data generation adversarial training method of the high voltage direct current transmission system is:

[0077] The Gaussian noise with the same dimension as the original sample of bad data of the DC measurement equipment is input into the generator, and the loss function is used to quantify the inconsistency between the bad data generated samples and the bad data original samples; the bad data generated samples and the bad data original samples are input into the discriminator to construct the loss function; game adversarial training is performed; and the bad data generated samples are merged with the normalized data set.

[0078] In specific implementation, the loss function is determined as follows:

[0079] L G =1-D(G(z))

[0080] Where: L G is the generator loss function, G is the generator function, D is the discriminator function, and z is the input noise data vector.

[0081] The generator function G is a parameterized neural network implementation. Its input is a random vector that obeys the same dimension of Gaussian white noise as the real data, and its output is a sample sampled from a certain distribution Pg. By giving real data for game training, Pg will gradually approach the function of the real data distribution.

[0082] The discriminator function D refers to the function used to determine whether the input data comes from real data or samples generated by G. After adversarial training, G and D reach a Nash equilibrium state, that is, D cannot determine whether its input comes from real samples or samples generated by G. At this time, it can be considered that G has learned the distribution of real data and extracted its distribution characteristics.

[0083] The damage function is determined as follows:

[0084]

[0085] Where: L D is the discriminator loss function, E is the expected function, P r is the distribution of the real data x input to the generator, P g is the noise distribution, and x is determined by the type of bad data device that needs to be trained. or Corresponding to optical fiber current transformer, DC transformer, and zero flux transformer in sequence;

[0086] The specific objective function of game adversarial training is determined by the following formula:

[0087]

[0088] In the formula, ||·|| p is the p-norm, is the gradient operator, λ is the penalty coefficient, For the generated data, ζ follows a uniform distribution in [0,1]. is the distribution along the bad data P r (·) and the generated bad data distribution P g (·) The straight lines between the sampling points are uniformly sampled.

[0089] like Figure 4 As shown, in the specific implementation, in step S3, the Spark parallel decision tree model training method of the high voltage direct current transmission system is:

[0090] S3.1. P foct,1 , P dcvd,1 , P zfct,1 Perform n random sampling to obtain C foct , C dcvd , C zfct ;

[0091] S3.2. Calculate the information entropy T(X m ), calculated as follows:

[0092]

[0093] In the formula, X m is the DC measurement data, m = 1, 2, 3 represent the training samples of the optical fiber current transformer, DC voltage divider, and zero flux current transformer respectively; a k is the ratio of bad data or normal measurement data in the sample, k = 0 represents bad data, k = 1 represents normal measurement data;

[0094] S3.3, taking feature b as the split point, establish a set {b 1 ,b 2 ,b 3 ,...,b j The segmentation point set S after} = {s i =(b i +b i+1 ) / 2|1≤i≤j-1};

[0095] S3.4, calculate the information gain Φ(X m ,b,s i );

[0096] S3.5. Calculate the information gain rate O(X m ,b,s i );

[0097] S3.6, the data classification results of the decision tree are obtained. The classification of data by a single decision tree is based on the gain rate O(X m ,b,si ) corresponds to the maximum information gain rate O(X m ,b,s b ) Split point s b Select.

[0098] In specific implementation, the information gain Φ(X m ,b,s i ) is calculated as follows:

[0099]

[0100] Where: |X m | indicates X m The amount of data; Represents feature b i ≤s i , b i >s i The weight of .

[0101] Preferably, the information gain rate of the measured data is O(X m ,b,s i ) is calculated as follows:

[0102]

[0103] Where: I(b) represents the characteristic intrinsic value.

[0104] In specific implementation, in step S4, the method for constructing the random forest bad data identification model is:

[0105] C-based foct , C dcvd , C zfct It can be obtained by n foct , n dcvd , n zfct Decision Tree Random forest bad data identification model F foct , F dcvd , F zfct :

[0106]

[0107] Each decision tree is combined and aggregated to form a random forest, and the final recognition result of the measurement data is obtained based on the voting of each decision tree. The proposed Spark architecture is applied to the DC data storage and calculation of parallel training. The parallel training process is as follows Figure 5As shown. Parallel training fully considers the combination of Spark's "Map-Reduce" idea and algorithm. The random forest parallel training based on the Spark framework is divided into two parts, corresponding to the "Map" and "Reduce" stages. In the first part, the HDFS file is read and a new RDD is created. The Cache instruction is executed locally to write to the RDD cache. Then, multiple samplings are performed to calculate each split point {b based on the feature b and the threshold ε. 1 ,b 2 ,b 3 ,...,b j The second part is to get C based on the calculation in the first part. foct , C dcvd , C zfct The corresponding decision tree summarizes the recognition results to complete the parallel training under the Spark framework.

[0108] Embodiment 2 of the present invention provides a system for identifying bad measurement data of a high-voltage direct current transmission system. The system is a system used by the above-mentioned method for identifying bad measurement data of a high-voltage direct current transmission system, and is characterized by comprising:

[0109] Preprocessing module, used for data normalization and construction of bad data sets;

[0110] Sample reconstruction module, used to reconstruct and generate bad data samples and balance the sample set;

[0111] Model training module, used to perform random forest parallel training and establish a bad data identification model;

[0112] The discrimination module is used to discriminate whether the DC measurement data is bad data, inputs the measurement data to be discriminated, and outputs the identification result voted by the bad data identification model.

[0113] Example 3

[0114] Taking a high-voltage direct current transmission system as an example, the DC measurement data samples of the system are shown in Table 1. The data sources include the historical operation database of the optical fiber current transformer, DC voltage divider, and zero-flux current transformer in the DC system, and the online operation monitoring data. The number of samples is the number of samples of the three types of DC measurement equipment after the acquisition sources are summarized. The number of training set samples is the number of samples with replacement sampling after normalization, which is used for decision tree training after 3 times standard deviation sample expansion. The test set samples are the number of samples that are not extracted by replacement sampling after normalization, which are used to evaluate the training effect of the model. The number of 3 times standard deviation samples is the number of samples in the original sample that exceeds 3 times the standard deviation of the mean. Such samples are bad data and are used for generative training to generate more bad data samples to balance the training sample set.

[0115] Table 1 DC measurement data sample information

[0116]

[0117] Based on Pytorch, a random forest training model for DC measurement equipment failure data was built. The number of decision trees was set to 100. The experimental platform used a dual-core, 4GB memory, 1 Master and 8 Slaves, a total of 9 nodes. During the training process, each node independently trained the decision tree and then aggregated it to form a random forest. The training process and principle of all decision trees were the same. The Spark kernel performed task scheduling to enable 9 nodes to perform training and result aggregation.

[0118] In order to more intuitively reflect the training effect of the model of the present invention, the present invention is compared with the fuzzy C-means algorithm and the support vector machine algorithm. foct,0 , P dcvd,0 , P zfct,0 The sample set consists of the measurement data that were not collected after each sampling. Test. Use the recall rate μ TP , precision μ TF , overall accuracy μ Acc And comprehensive evaluation index F score Evaluate the training effect of the model. Recall rate μ TP , precision μ TF , overall accuracy μ Acc They are:

[0119]

[0120] In the formula, The number of samples correctly identified as normal measurement data; The number of samples that could not be correctly identified as normal measurement data; The number of samples correctly identified as bad data; is the number of samples that were not correctly identified as bad data.

[0121] F score The original calculation expression is:

[0122]

[0123] In the formula, γ reflects μ TP and μ TF The larger the γ, the more important the precision rate is. Both the recall rate and the precision rate are very important, so in this example, F score Take γ=1.

[0124] like Figure 2As shown, the overall accuracy of the three algorithms is above 95%. The precision of the fuzzy C-means algorithm is slightly higher than that of the present invention, but the recall and overall accuracy are lower than those of the present invention; the recall of the support vector machine algorithm is slightly higher than that of the present invention, but the precision performance is poor. The fuzzy C-means algorithm divides normal data and bad data according to the size of similarity, and misjudges some normal data as bad data. Therefore, the recall rate of the present invention is superior to that of the fuzzy C-means algorithm. The support vector machine algorithm cannot learn the characteristics of all bad data, and there is a certain missed detection rate. Therefore, the precision rate of the present invention is obviously superior to that of the support vector machine algorithm.

[0125] The results show that the present invention can improve the rapidity of identifying bad DC measurement data while maintaining a good identification effect, which is of great significance for ensuring the safe operation of the DC system.

[0126] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for identifying bad measurement data in a high voltage direct current transmission system, characterized in that: The steps include: Collect historical measurement data of optical fiber current transformers, DC voltage dividers and zero-flux current transformers of high-voltage DC transmission systems; extract original samples of bad data from historical measurement data for generative adversarial training, balance the sample set so that the number of bad data samples accounts for 50% of the total sample set; train the balanced sample set with Spark parallelized decision tree model; aggregate multiple decision trees obtained from training to form a random forest bad data identification model; The measurement data of the optical fiber current transformer, the DC voltage divider and the zero flux current transformer are collected, and the measurement data are input into the bad data identification model to identify the bad measurement data.

2. The method for identifying bad data in a high voltage direct current transmission system of a power grid according to claim 1, characterized in that: The bad data original samples are extracted according to the following steps: The original data sets of the optical fiber current transformer, DC voltage divider, and zero flux current transformer are processed as follows: Where P foct,0 , P dcvd,0 , P zfct,0 They are the original data sets of optical fiber current transformer, DC voltage divider, and zero flux current transformer, respectively. foct,nor , P dcvd,nor , P zfct,nor They are normalized data sets of fiber optic current transformer, DC voltage divider, and zero flux current transformer respectively; data that differs from the average value by 3 standard deviations is defined as bad data, and data that differs from the average value by 1 standard deviation is defined as normal measurement data; the original samples of bad data that differ from the average value by 3 standard deviations are extracted to form the original bad data set for generative adversarial training.

3. A generative adversarial training method for identifying bad data in a high voltage direct current transmission system of a power grid according to claim 2, characterized in that: The bad data generation adversarial training method for the HVDC transmission system is: The Gaussian noise of the same dimension as the original sample of bad data is input into the generator to output the generated sample of bad data; the generator loss function is used to quantify the inconsistency between the generated sample of bad data and the original sample of bad data; the generated sample of bad data and the original sample of bad data are input into the discriminator to construct the discriminator loss function; game adversarial training is performed; the generated sample of bad data is merged with the normalized data set.

4. A generative adversarial training method for identifying bad data in a high voltage direct current transmission system of a power grid according to claim 3, characterized in that: The generator loss function is determined as follows: L G =1-D(G(z)) Where: L G is the generator loss function, G is the generator function, D is the discriminator function, and z is the input noise data vector.

5. A generative adversarial training method for identifying bad data in a high voltage direct current transmission system of a power grid according to claim 4, characterized in that: The discriminator loss function is determined as follows: Where: L D is the discriminator loss function, E is the expected function, P r is the distribution of the original data x input to the generator, P g is the noise distribution, and x is determined by the type of bad data device that needs to be trained. or They correspond to optical fiber current transformer, DC transformer and zero flux transformer respectively.

6. A generative adversarial training method for identifying bad data in a high voltage direct current transmission system of a power grid according to claim 5, characterized in that: The objective function of game adversarial training is determined by the following formula: In the formula, ||·|| p is the p-norm, is the gradient operator, λ is the penalty coefficient, For the generated data, ζ follows a uniform distribution between [0,1]; is the distribution along the bad data P r (·) and the generated bad data distribution P g The straight lines between the sampling points in (·) are uniformly sampled.

7. A method for identifying bad measurement data in a high voltage direct current transmission system according to claim 6, characterized in that: The Spark parallel decision tree model training method for the HVDC transmission system is: P foct,1 , P dcvd,1 , P zfct,1 Perform n random sampling to obtain C foct , C dcvd , C zfct ; Calculate the information entropy T(X) of the balanced DC measurement data m ), calculated as follows: Where, X m is the DC measurement data, m=1, 2, 3 represent the balanced DC measurement data of the optical fiber current transformer, DC voltage divider, and zero flux current transformer respectively; a k is the ratio of bad data or normal measurement data in the sample, k = 0 represents bad data, k = 1 represents normal measurement data; Taking feature b as the split point, establish a set {b 1 ,b 2 ,b 3 ,...,b j The segmentation point set S after} = {s i =(b i +b i+1 ) / 2|1≤i≤j-1}; Calculate the information gain Φ(X m ,b,s i );Calculate the information gain rate of the measured data O(X m ,b,s i ); The data classification result of the decision tree is obtained. The classification of data by a single decision tree is based on the gain rate of feature b O(X m ,b,s i ) corresponds to the maximum information gain rate O(X m ,b,s b ) Split point s b Select; Set the threshold ε and determine the accuracy of recognition. If the information gain rate is not greater than ε, the feature classification ability is determined to be unqualified; when the gain rate O(X m ,b,s i ) corresponds to the maximum information gain rate O(X m ,b,s b ) is greater than ε, b i ≤s b , b i >s b The data are classified into each category; From b i ≤s b Select another measurement data with the largest information gain rate As a classification point, The data are regarded as bad data, otherwise they are normal measurement data; The above process is repeated recursively until the data classified at a leaf node all belong to the same category.

8. A method for identifying bad measurement data in a high voltage direct current transmission system according to claim 7, characterized in that: The information gain Φ(X m ,b,s i ) is calculated as follows: Where: |X m | indicates X m The amount of data; Represents feature b i ≤s i , b i >s i The weight of The information gain rate of the measured data is O(X m ,b,s i ) is calculated as follows: Where: I(b) represents the characteristic intrinsic value.

9. A method for identifying bad measurement data in a high voltage direct current transmission system according to claim 1, characterized in that: In step S4, the method for constructing the random forest bad data identification model is: C-based foct , C dcvd , C zfct It can be obtained by n foct , n dcvd , n zfct Decision Tree Random forest bad data identification model F foct , F dcvd , F zfct : Each decision tree is combined and aggregated to form a random forest. The final identification result of the measurement data is obtained based on the voting of each decision tree. When 50% of the decision tree identification results are bad data, the final identification result is defined as bad data.

10. A system for identifying bad measurement data of a high-voltage direct current transmission system, the system being the system used in the method for identifying bad measurement data of a high-voltage direct current transmission system according to any one of claims 1 to 9, characterized in that: include: Preprocessing module, used for data normalization and construction of bad data sets; Sample reconstruction module, used to reconstruct and generate bad data samples and balance the sample set; Model training module, used to perform random forest parallel training and establish a bad data identification model; The discrimination module is used to discriminate whether the DC measurement data is bad data, inputs the measurement data to be discriminated, and outputs the identification result voted by the bad data identification model.

Images