Classification method and system for power negative list data

Abstract

The invention is suitable for the technical field of data processing, and provides a power negative list data classification method and system, and the method comprises the steps: obtaining historical power data related to a power system, and carrying out the preprocessing; inputting the preprocessed historical power data into a preset mixed data distribution learning network for learning training, and outputting to obtain distribution characteristics of power negative list categories; positive example samples are screened based on the distribution characteristics of the power negative list categories, sample data obtained through screening are input into a single-class support vector machine model for training, and a trained classification model of the power negative list data is obtained; inputting target power data acquired in real time into the mixed data distribution learning network, mapping the target power data to a high-dimensional space, and extracting distribution features; and determining an input sample according to the distribution characteristics, inputting the sample into a trained classification model of the power negative list data, generating a target category, and realizing an effect of accurately identifying the power grid negative list data.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and particularly relates to a classification method and system for power negative list data. Background Art

[0002] With the increasing scale and complexity of the power system, the management and security of power data have become a key challenge. The power system not only faces conventional operation and maintenance problems, but also needs to cope with the security risks brought by complex network environments, especially unauthorized access, abnormal operations, and potential malicious attacks. As an important data source for recording abnormal behaviors, illegal operations, and potential risks, the power negative list is an important means to ensure power security and monitor the operating status of the power system. The data in the power system includes both normal operation information and records of abnormal behaviors. These negative list data are characterized by large volume, diverse types, different features, and complex business semantics, lacking significant features.

[0003] In the prior art, traditional classification and grading methods are currently adopted for the management of power negative list data, such as rule matching technology, manual monitoring and classification, etc. These methods cannot achieve sufficient performance when dealing with complex data structures and are difficult to reuse the associated information contained in complex data.

[0004] In view of this, a classification method and system for power negative list data are needed. Summary of the Invention

[0005] Embodiments of the present application provide a classification method and system for power negative list data, which are used to solve the problem of inaccurate classification of power negative list data.

[0006] The first aspect of the embodiments of the present application provides a classification method for power negative list data, including:

[0007] Obtain historical power data related to the power system, and preprocess the obtained historical power data;

[0008] Input the preprocessed historical power data into a preset hybrid data distribution learning network, and output the distribution characteristics of power negative list categories, where the power negative list categories include power negative list-related data and non-power negative list-related data;

[0009] Based on the distribution characteristics of the power negative list categories, screen positive example samples, and input the screened sample data into a one-class support vector machine model for training to obtain a trained classification model for power negative list data;

[0010] Preprocess the target power data obtained in real time and input it into a preset hybrid data distribution learning network to map the target power data to a high-dimensional space and extract the distribution characteristics of the target power data;

[0011] Determine the input samples according to the extracted distribution characteristics, and input the input samples into a trained classification model of power negative list data to generate the target category.

[0012] Furthermore, obtain the historical power data related to the power system and preprocess the obtained historical power data, including:

[0013] Identify different types of lexical units and syntactic units of the text data in the historical power data through lexical analysis and syntactic analysis;

[0014] Determine keywords, identifiers, and operators in the text data based on lexical units, and determine the syntactic structure and lexical relationships in the text data based on syntactic units;

[0015] Obtain the structured data and text data in the historical power data under management according to the keywords, identifiers, and operators, as well as the syntactic structure and lexical relationships.

[0016] Furthermore, input the preprocessed historical power data into a preset hybrid data distribution learning network, and output the distribution characteristics of the power negative list categories. The power negative list categories include power negative list-related data and non-power negative list-related data, including:

[0017] Set the number of components in the hybrid data distribution learning network and the initialization parameters of each component;

[0018] Use the structured data and the text data obtained by preprocessing as input samples, and calculate the probability density function values of the input samples belonging to the distributions of each component and the probability of the input samples belonging to the mixed distribution according to the number of components and the initialization parameters of each component;

[0019] Calculate the probability of the input samples belonging to the power negative list-related data according to the probability density function values of the input samples belonging to the distributions of each component and the probability of the input samples belonging to the mixed distribution;

[0020] Select the cross-entropy loss function to optimize the probability of the input samples belonging to the power negative list-related data;

[0021] Use the backpropagation algorithm to calculate the gradients of each parameter in the hybrid data distribution learning network with respect to the loss function according to the calculated loss function values;

[0022] The network parameters are updated using an optimization algorithm until the distribution characteristics of the power negative list categories are output after the hybrid data distribution learning network converges.

[0023] Furthermore, calculating the probability that the input sample belongs to the power negative list-related data according to the probability density function value that the input sample belongs to each component distribution and the probability that the input sample belongs to the hybrid distribution includes:

[0024] The probability density function value p k (x i ) of the input sample belonging to each component distribution:

[0025]

[0026] where: d is the feature dimension of the data, |∑ k | is the determinant of the covariance matrix ∑ k , (x i ―μ k ) T is the transpose of the vector (x i ―μ k );

[0027] The probability p(x i ) that the input sample belongs to the hybrid distribution:

[0028]

[0029] where: π k is the initialized mixing coefficient;

[0030] The probability that the input sample belongs to the power negative list-related data

[0031]

[0032] where: π1 is the mixing coefficient of the power negative list-related data or the power negative list-unrelated data.

[0033] Furthermore, the distribution characteristics include word vector space characteristics and semantic relationship characteristics, the word vector space characteristics and semantic relationship characteristics are determined by the lexical units and the syntactic units, and the distribution of the word vector space characteristics and semantic relationship characteristics of the power negative list-related data is different from that of the non-power negative list-related data in the high-dimensional space.

[0034] Furthermore, screening positive example samples based on the distribution characteristics of the power negative list categories, and inputting the screened sample data into a one-class support vector machine model for training to obtain a trained classification model for the power negative list data, includes:

[0035] Set the parameters of the kernel function and the penalty coefficient respectively;

[0036] Input the filtered sample data into the set kernel function for model training, and obtain the classification model of the trained power negative list data by solving the set Lagrangian function.

[0037] Furthermore, the step of inputting the filtered sample data into the set kernel function for model training and obtaining the classification model of the trained power negative list data by solving the set Lagrangian function includes:

[0038] By the method of Lagrange multipliers, introduce the Lagrange multiplier to obtain the dual problem of the one-class support vector machine model. For each training sample x i There is:

[0039]

[0040] The constraint condition is:

[0041]

[0042] Where: α i Is a non-zero Lagrange multiplier, n is the number of training samples, and v is a hyperparameter.

[0043] Furthermore, the expression of the classification model of the trained power negative list data in the step of inputting the filtered sample data into the set kernel function for model training includes:

[0044]

[0045] Where: α k Is a non-zero Lagrange multiplier, K(x i , x) is the value of the kernel function, and ρ is the intercept obtained through training.

[0046] Furthermore, the step of determining the input sample according to the extracted distribution characteristics and inputting the input sample into the trained classification model of the power negative list data to generate the target category includes:

[0047] When the output value of the classification model is greater than the set threshold, it is determined that the target power data belongs to the power negative list related data category;

[0048] When the output value of the classification model is less than or equal to the set threshold, it is determined that the target power data belongs to the non-power negative list related data category.

[0049] The second aspect of the embodiments of the present application provides a classification system for power negative list data, including:

[0050] A data acquisition and processing unit, configured to acquire historical power data related to a power system and preprocess the acquired historical power data;

[0051] A hybrid data distribution learning model training unit, configured to input the preprocessed historical power data into a preset hybrid data distribution learning network, and output the distribution characteristics of power negative list categories, where the power negative list categories include power negative list related data and non-power negative list related data;

[0052] A classification model determination unit, configured to screen positive example samples based on the distribution characteristics of the power negative list categories, and input the screened sample data into a one-class support vector machine model for training to obtain a trained classification model for power negative list data;

[0053] A distribution characteristic extraction unit for target power data, configured to preprocess the real-time acquired target power data and input it into a preset hybrid data distribution learning network to map the target power data to a high-dimensional space, and extract the distribution characteristics of the target power data;

[0054] A target category generation unit, configured to determine an input sample according to the extracted distribution characteristics, input the input sample into the trained classification model for power negative list data, and generate a target category.

[0055] As can be seen from the above technical solutions, the embodiments of the present application have the following advantages:

[0056] In the present invention, by acquiring and preprocessing the historical power data of the power system, inputting it into a preset hybrid data distribution learning network for training, acquiring the distribution characteristics of power negative list related and non-related data, screening positive example samples based on this for one-class support vector machine model training to obtain a classification model, and then mapping the real-time acquired target power data to a high-dimensional space through a preset hybrid data distribution learning network to extract distribution characteristics, determining the input sample accordingly and then inputting it into the classification model to generate a target category, effectively integrating the complex data of the power system, accurately capturing the characteristics of power negative list data, accurately classifying target power data, and providing a precise and comprehensive data basis for the risk control and safe operation of the power system.

[0057] Other advantages, objectives, and features of the present invention will be described to some extent in the subsequent specification, and to some extent, will be obvious to those skilled in the art based on the study of the following text, or can be taught from the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1Schematic diagram of the implementation process of a classification method for power negative list data in the present invention;

[0059] Figure 2 Structural diagram of the hybrid data distribution learning network in the present invention. Detailed implementation manners

[0060] The terms "first", "second", "third", "fourth", etc. (if any) in the description and claims of this application and the above-mentioned drawings are used to distinguish similar objects and do not necessarily need to be used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances so that the embodiments of this application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "comprising" and "corresponding to" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0061] Embodiment 1

[0062] In this embodiment, the implementation method can be implemented in a system, on a server, or on a terminal, and no specific limitation is made. Below, from the perspective of system implementation, the photovoltaic power station multi-device fault ride-through coordination control method in this application will be introduced. Please refer to Figure 1 , the method provided by the embodiment of this application includes the following steps:

[0063] S11. Obtain historical power data related to the power system, and preprocess the obtained historical power data;

[0064] In this embodiment, obtaining historical power data related to the power system is to collect relevant partial data of the power system, including structured data and text data. For structured data, a manual or rule-based method is used for preliminary screening, the data is extracted and the content of the structured data table is concatenated, and finally all the data and the corresponding negative list labels are stored centrally.

[0065] Here, the preprocessing is based on the acquired data for feature selection. For example, some information in the power data includes the user's power consumption situation, equipment-related information, business process information, etc. There is a large amount of redundancy in these raw data. To improve the efficiency of subsequent learning and classification, key features are extracted from the raw data to better describe the characteristics of the power data. The goal of feature selection is to extract the key information that can represent the important aspects of the power data while reducing data redundancy. The selected features include: for numerical data such as power consumption and electricity price; for character data such as equipment numbers and user names; for text data such as the description of key activities in the business process.

[0066] Specifically, this step also includes the following:

[0067] Identify different types of lexical units and syntactic units of the text data in the historical power data through lexical analysis and syntactic analysis respectively;

[0068] Based on the lexical units, determine the keywords, identifiers, and operators in the text data, and based on the syntactic units, determine the syntactic structure and lexical relationships in the text data.

[0069] Here, lexical analysis decomposes the input text statements into individual lexical units. These text statements can be text contents with certain formats and meanings such as operation records and fault descriptions in the power system. The lexical units obtained from lexical analysis are the basis for subsequent syntactic analysis. It transforms the original text into more structured units, which helps further processing and analysis. Syntactic analysis is carried out by defining syntactic rules. The syntax tree or abstract syntax tree obtained from syntactic analysis can clearly represent the syntactic structure of the text statements. In the power system, this structure can help understand the operation intention, fault cause, etc. For example, by analyzing the syntax tree of the power system operation record, it can be found whether the operation conforms to the specification and whether there are potential incorrect operation patterns.

[0070] Furthermore, the purpose of lexical analysis LEX is to break down the input text statements into a series of lexical units. Each lexical unit represents a basic element in the text statement, including keywords, identifiers, and operators. In the syntax analysis YACC stage, based on the lexical analysis, the sequence of lexical units is combined into statements or expressions with syntactic structures, aiming to generate a syntax tree or an abstract syntax tree according to the sequence of lexical units. This tree structure represents the syntactic structure of the text statement. YACC and LEX can perform a more in-depth analysis of the text data, extracting more accurate and detailed syntactic and lexical features. These features can be used as supplementary information to further enrich the representation of the data. When entering the mixed data distribution learning module, the network can better learn the data distribution by utilizing these enhanced features. For example, for some device configuration files or user feedback texts in power data, after being analyzed by YACC and LEX, some specific syntactic structures and lexical relationships may be extracted, which helps the mixed data distribution learning network capture the internal patterns of the data more precisely.

[0071] S12. Input the preprocessed historical power data into a preset mixed data distribution learning network to generate the distribution features of the power negative list categories, where the power negative list categories include power negative list-related data and non-power negative list-related data;

[0072] In this embodiment, step S12 further includes the following:

[0073] 1. Set the number of components in the mixed data distribution learning network and the initialization parameters of each component;

[0074] The purpose of mixed data distribution learning is to simultaneously learn the distributions of multiple data types in order to better classify power-sensitive data. By learning the data distribution, the original data is mapped to a high-dimensional space, thereby integrating the distribution relationships between different data types and providing a basis for subsequent classification and grading operations.

[0075] Please refer to Figure 2 the mixed data distribution learning network diagram in. Set a learnable lookup layer between the input layer and the first hidden layer. This lookup layer maps the discrete power data into continuous high-dimensional vectors, and its parameter dimension is determined according to the size of the power data vocabulary and the word embedding size. During the training process, the data of each category (negative list-related data and non-negative list-related data) is arranged into a sequence according to certain rules, and then a data unit is randomly removed, and the remaining sequence is passed through the lookup layer and into the network.

[0076] First, it is necessary to determine the number of components in the mixture model, that is, to assume how many different distributions the data is composed of. For example, when processing data related to the power negative list, according to the prior knowledge of the power system and the characteristics of the data, it can be assumed that the data related to the power negative list and the data not related to the power negative list roughly conform to two different distributions respectively. For instance, one distribution corresponds to the illegal transaction category in the power negative list data, and the other distribution corresponds to the non-illegal normal transaction category, etc. Here, the number of components k is set to 2.

[0077] For each component, it is necessary to initialize its mean vector μ k , covariance matrix ∑ k . The mean vector μ k 's dimension depends on the characteristic dimension of the data. For example, if the historical power data has d characteristics (such as different indicators like electricity quantity, voltage, etc.) after preprocessing, then μ k is a d-dimensional vector. When initializing, values within a reasonable range can be randomly generated as the initial value of the mean vector. For example, for the electricity quantity characteristic, when the unit is megawatt, its value range may be [0, 1000], then a value can be randomly generated within this range as a component of the mean vector of the electricity quantity characteristic under a certain component.

[0078] The covariance matrix ∑ k is a d×d symmetric matrix, which is used to describe the correlation of data in each characteristic dimension. Initially, it can be set as the identity matrix or a diagonal matrix can be generated according to some simple rules. The elements on the diagonal can be some small positive numbers, representing the initial variances in each characteristic dimension. For example, if it is set as a diagonal matrix, the elements on the diagonal are all 0.1, indicating that initially it is considered that the data in each characteristic dimension is relatively concentrated and the variance is small.

[0079] At the same time, it is also necessary to initialize the mixing coefficient π k , satisfying and π k ≥0. Each π k can be simply initialized as 1 / k. In our example where k = 2, that is, π1 = π2 = 0.5, indicating that initially it is considered that the proportions of the two distributions in the mixed distribution are the same.

[0080] 2. Identify lexical units and syntactic units based on historical power data, use the data after identifying lexical units and syntactic units as input samples, and calculate the probability density function values of the input samples belonging to the distributions of each component and the probability of the input samples belonging to the mixed distribution according to the number of components and the initialization parameters of each component;

[0081] 3. Calculate the probability that the input sample belongs to the data related to the power negative list according to the probability density function values of the input sample belonging to each component distribution and the probability of the input sample belonging to the mixture distribution;

[0082] For each historical power data sample x after preprocessing i (where i = 1, 2,..., N, and N is the total number of samples), according to the parameters of the current network, that is, the initialized mean vector μ k , covariance matrix Σ k and mixing coefficient π k , calculate the probability density function values of the sample belonging to each component distribution.

[0083] The probability density function values p k (x i ) of the input sample belonging to each component distribution:

[0084]

[0085] where: d is the feature dimension of the data, |Σ k | is the determinant of the covariance matrix Σ k , (x i ―μ k ) T is the transpose of the vector (x i ―μ k ).

[0086] The probability p(x i ) of the input sample belonging to the mixture distribution:

[0087]

[0088] where: π k is the initialized mixing coefficient.

[0089] This step is actually to comprehensively consider the "contribution" of each component distribution to the sample and obtain the probability of the sample under the entire mixture distribution.

[0090] Set component 1 to correspond to the data related to the power negative list, and component 2 to correspond to the data not related to the power negative list. Then, according to the probability p(x i ) of the sample x belonging to the mixture distribution calculated above and the probability density function values p i (x k ), the probability that the sample x i belongs to the data related to the power negative list can be further calculated i , that is which represents the probability estimate value that the sample output by the network belongs to the data related to the power negative list:

[0091]

[0092] Wherein: π1 is the mixing coefficient of data related to the power negative list or data not related to the power negative list.

[0093] Similarly, the probability that the sample x i belongs to data not related to the power negative list is

[0094] 4. Select the cross-entropy loss function to optimize the probability that the input sample belongs to data related to the power negative list;

[0095] To measure the probability estimation value output by the network and the actual category of the sample (it is known whether each historical power data sample belongs to data related to the power negative list or data not related to the power negative list, denoted by y i represents the actual category, y i =1 indicates belonging to data related to the power negative list, y i =0 indicates belonging to data not related to the power negative list), the cross-entropy loss function is selected here.

[0096] 5. Use the backpropagation algorithm to calculate the gradient of each parameter in the mixed data distribution learning network with respect to the loss function according to the calculated loss function value;

[0097] 6. Adopt an optimization algorithm to update the network parameters until the mixed data distribution learning network converges and outputs the distribution characteristics of the power negative list category.

[0098] Specifically, according to the calculated loss function value, use the backpropagation algorithm to calculate the gradient of each parameter in the network with respect to the loss function. After obtaining the gradients of each parameter, adopt an optimization algorithm to update the network parameters. Here, stochastic gradient descent (SGD) is adopted, and its update formula is:

[0099]

[0100] Wherein: θ is the network parameter, θ new is the updated network parameter, η is the learning rate, is the gradient of the loss function with respect to the parameter θ.

[0101] Continuously repeat steps such as forward propagation, loss function calculation, backpropagation, and parameter update until the network converges. The convergence criterion can be that the value of the loss function no longer decreases significantly, and finally output the distribution characteristics of the categories of the power negative list. Here, the distribution characteristics include word vector space characteristics and semantic relationship characteristics, which are determined by the lexical units and the syntactic units identified above. Among them, the word vector space characteristics and semantic relationship characteristics of the data related to the power negative list are different from those of the data related to the non-power negative list in the distribution in the high-dimensional space.

[0102] S13. Screen positive example samples based on the distribution characteristics of the categories of the power negative list, and input the screened sample data into a one-class support vector machine model for training to obtain a trained classification model for the power negative list data.

[0103] In this embodiment, step S13 further includes the following:

[0104] 1. Set the parameters of the kernel function and the penalty coefficient respectively;

[0105] 2. Input the screened sample data into the set kernel function for model training, and obtain a trained classification model for the power negative list data by solving the set Lagrangian function.

[0106] Here, the radial basis kernel function is selected, and the expression is as follows:

[0107] K(x i ,x j )=exp(―γ||x i ―x j || 2 )

[0108] Where: x i and x j are two data samples, γ is the bandwidth of the kernel function, and ||x i ―x j || 2 is the square of the Euclidean distance between the two samples. When initializing, the value of γ needs to be determined. Usually, a suitable value can be initially selected through methods such as cross-validation. For example, first set a range, and then try different values within this range, and determine the optimal γ value according to the performance of the model on the validation set.

[0109] The penalty coefficient C is used to control the degree of penalty for misclassifying training samples. A larger value of C means a more severe penalty for misclassification, and the model will be more inclined to classify all training samples correctly as much as possible, but may lead to overfitting; a smaller value of C allows a certain degree of misclassification, which may make the model more generalizable, but may miss some samples. At initialization, the value of C can also be determined based on experience or through preliminary experiments. Generally, some common values can be tried first, such as C = 0.1, 1, 10, etc., and then adjusted according to the performance of the model on the validation set.

[0110] Here, the goal of OCSVM is to find a hyperplane that is as far away from the origin as possible and separates most (or all) normal data points. Its optimization objective function is expressed as:

[0111]

[0112] where: ξ i is the slack variable, which is used to handle inseparable cases. This objective function can be regarded as maximizing the margin between the hyperplane and the origin while minimizing the misclassified support vectors.

[0113] By using the Lagrange multiplier method, introducing the Lagrange multiplier α i , the dual problem of OCSVM is obtained. For each training sample x i there is:

[0114]

[0115] The constraint conditions are:

[0116]

[0117] where: n is the number of training samples, and v is the hyperparameter.

[0118] After training is completed, the decision function of OCSVM can be expressed as:

[0119]

[0120] where: α i is the non-zero Lagrange multiplier, K(x i , x) is the value of the kernel function, ρ is the intercept obtained through training. If f(x) > 0, then the data point x belongs to the data related to the power negative list; if f(x) ≤ 0, then the data point x does not belong to the data related to the power negative list.

[0121] S14. After preprocessing the target power data obtained in real time, input it into the hybrid data distribution learning network to map the target power data to a high-dimensional space and extract the distribution characteristics of the target power data;

[0122] S15. Determine the input samples according to the extracted distribution characteristics, and input the input samples into the trained classification model of the power negative list data to generate the target category.

[0123] In steps S14 - S15, after the above model training is completed, preprocess the target power data obtained in real time, that is, identify the lexical units and syntactic units, input the power data containing the lexical units and syntactic units into the preset hybrid data distribution learning network, map it to a high - dimensional space, extract the distribution characteristics from it, then screen the input samples according to these distribution characteristics, input the screened samples into the trained classification model of the power negative list data, and finally generate the target category. Specifically, the threshold set in the above steps is 0. When the output value of the classification model is greater than the set threshold, it is determined that the target power data belongs to the category of power negative list - related data. When the output value of the classification model is less than or equal to the set threshold, it is determined that the target power data belongs to the category of non - power negative list - related data.

[0124] The present invention can achieve an accurate judgment on whether the target power data belongs to the category of power negative list - related data, providing accurate data basis for the stable operation and risk monitoring of the power system.

[0125] Embodiment Two

[0126] An embodiment of a classification system for power negative list data in the present invention includes the following steps:

[0127] A data acquisition and processing unit, which is used to acquire historical power data related to the power system and preprocess the acquired historical power data;

[0128] A hybrid data distribution learning model training unit, which is used to input the preprocessed historical power data into the preset hybrid data distribution learning network and output the distribution characteristics of the power negative list categories, where the power negative list categories include power negative list - related data and non - power negative list - related data;

[0129] A classification model determination unit, which is used to screen positive example samples based on the distribution characteristics of the power negative list categories and input the screened sample data into a one - class support vector machine model for training to obtain the trained classification model of the power negative list data;

[0130] A distribution characteristic extraction unit for target power data, which is used to preprocess the target power data obtained in real time and input it into the preset hybrid data distribution learning network to map the target power data to a high - dimensional space and extract the distribution characteristics of the target power data;

[0131] A target category generation unit, configured to determine an input sample according to the extracted distribution features, input the input sample into a trained classification model of power negative list data, and generate a target category.

[0132] Those of ordinary skill in the art can realize that the units of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. To clearly illustrate the interchangeability of hardware and software, the components of each example have been generally described according to their functions in the above description. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present invention.

[0133] In the embodiments provided by the present invention, it should be understood that the division of units is only a logical function division, and there may be other division methods in actual implementation. For example, multiple units can be combined into one unit, one unit can be split into multiple units, or some features can be ignored, etc. In addition, the functional units in each embodiment of the present invention can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above-mentioned integrated units can be implemented in the form of hardware or in the form of software functional units.

[0134] If the above-mentioned integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program codes, such as USB flash drives, read-only memories (ROMs), random access memories (RAMs), mobile hard disks, magnetic disks, or optical discs.

[0135] It is understandable that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the various embodiments of the present invention, and they should all be covered within the scope of the claims and the description of the present invention.

Claims

1. A classification method for power negative list data, characterized in that: include: Acquire historical power data related to the power system and pre-process the acquired historical power data; Input the preprocessed historical power data into a preset hybrid data distribution learning network, and output the distribution characteristics of the power negative list category, wherein the power negative list category includes power negative list related data and non-power negative list related data; Based on the distribution characteristics of the electricity negative list categories, positive samples are screened, and the screened sample data are input into a single-class support vector machine model for training to obtain a trained classification model for the electricity negative list data; The target power data acquired in real time is pre-processed and then input into a preset hybrid data distribution learning network to map the target power data into a high-dimensional space and extract the distribution characteristics of the target power data; An input sample is determined according to the extracted distribution features, and the input sample is input into a classification model of the trained electricity negative list data to generate a target category.

2. The classification method of power negative list data according to claim 1 is characterized in that: The acquiring of historical power data related to the power system and preprocessing of the acquired historical power data includes: Identify different types of lexical units and grammatical units of text data in the historical power data through lexical analysis and grammatical analysis; Determine keywords, identifiers and operators in the text data based on lexical units, and determine grammatical structures and lexical relationships in the text data based on grammatical units; According to the keywords, identifiers and operators as well as the grammatical structure and vocabulary relationship, structured data and text data in the managed historical power data are obtained.

3. The classification method of power negative list data according to claim 2 is characterized in that: The pre-processed historical power data is input into a preset hybrid data distribution learning network, and the distribution characteristics of the power negative list category are output, and the power negative list category includes power negative list related data and non-power negative list related data, including: Set the number of components in the hybrid data distribution learning network and the initialization parameters of each component; The structured data obtained by preprocessing and the text data are used as input samples, and the probability density function value of the input sample belonging to the distribution of each component and the probability that the input sample belongs to the mixed distribution are calculated according to the number of components and the initialization parameters of each component; Calculate the probability that the input sample belongs to the data related to the electricity negative list according to the probability density function value that the input sample belongs to each component distribution and the probability that the input sample belongs to the mixed distribution; Selecting a cross entropy loss function to optimize the probability that the input sample belongs to data related to the electricity negative list; The back propagation algorithm is used to calculate the gradient of each parameter in the mixed data distribution learning network relative to the loss function according to the calculated loss function value; The optimization algorithm is used to update the network parameters until the mixed data distribution learning network converges and outputs the distribution characteristics of the electricity negative list categories.

4. The classification method of power negative list data according to claim 3 is characterized in that: The calculating the probability that the input sample belongs to the data related to the electricity negative list according to the probability density function value that the input sample belongs to each component distribution and the probability that the input sample belongs to the mixed distribution includes: The probability density function value p of the input sample belonging to each component distribution k (x i ): Where: d is the characteristic dimension of the data, |∑ k | is the covariance matrix∑ k The determinant of (x i ―μ k ) T is a vector (x i ―μ k ) The probability p(x) that the input sample belongs to the mixed distribution i ): Where: π k To initialize the mixing coefficient; The probability that the input sample belongs to the data related to the electricity negative list Where: π1 is the mixing coefficient of the electricity negative list related data or the electricity negative list non-related data.

5. The classification method of power negative list data according to any one of claims 1 to 4, characterized in that: The distribution features include word vector space features and semantic relationship features, which are determined by the lexical units and the grammatical units. The word vector space features and semantic relationship features of the data related to the electricity negative list are different from the word vector space features and semantic relationship features of the data related to the non-electricity negative list in terms of their distribution in high-dimensional space.

6. The classification method of power negative list data according to claim 1 is characterized in that: The method of screening positive samples based on the distribution characteristics of the electricity negative list categories, and inputting the screened sample data into a single-class support vector machine model for training to obtain a trained classification model for the electricity negative list data includes: Set the parameters and penalty coefficients of the kernel function respectively; The screened sample data is input into the set kernel function for model training, and the trained classification model of the electricity negative list data is obtained by solving the set Lagrangian function.

7. The classification method of power negative list data according to claim 6 is characterized in that: The sample data obtained by screening is input into the set kernel function for model training, and the trained classification model of the power negative list data is obtained by solving the set Lagrangian function, including: By introducing the Lagrange multiplier method, we can get the dual problem of the single-class support vector machine model. For each training sample x i have: The constraints are: Where: α i is a non-zero Lagrange multiplier, n is the number of training samples, and v is a hyperparameter.

8. The classification method of power negative list data according to claim 7 is characterized in that: The expression of the classification model of the trained electricity negative list data includes: Where: α i is a non-zero Lagrange multiplier, K(x i ,x) is the value of the kernel function, and ρ is the intercept obtained through training.

9. The classification method of power negative list data according to claim 8 is characterized in that: Determining input samples according to the extracted distribution features, inputting the input samples into the classification model of the trained power negative list data, and generating target categories, including: When the output value of the classification model is greater than a set threshold, it is determined that the target power data belongs to the data category related to the power negative list; When the output value of the classification model is less than or equal to the set threshold, it is determined that the target power data belongs to the category of data related to the non-power negative list.

10. A classification system for power negative list data, characterized in that: include: A data acquisition and processing unit, used to acquire historical power data related to the power system and pre-process the acquired historical power data; A hybrid data distribution learning model training unit, used for inputting the preprocessed historical power data into a preset hybrid data distribution learning network, and outputting distribution characteristics of power negative list categories, wherein the power negative list categories include power negative list related data and non-power negative list related data; A classification model determination unit, used for screening positive samples based on the distribution characteristics of the electricity negative list categories, and inputting the screened sample data into a single-class support vector machine model for training to obtain a trained classification model for the electricity negative list data; A target power data distribution feature extraction unit is used to pre-process the target power data acquired in real time and then input it into a preset hybrid data distribution learning network to map the target power data into a high-dimensional space and extract the distribution features of the target power data; The target category generating unit is used to determine the input sample according to the extracted distribution characteristics, input the input sample into the classification model of the trained power negative list data, and generate the target category.

Images