A power grid digital resource modeling management method based on meta learning

Abstract

The application discloses a power grid digital resource modeling management method based on meta learning, comprising the following steps: obtaining power grid digital resources after data governance, performing vector conversion on structured data and unstructured data respectively to obtain initial instance data; obtaining context-aware instance data based on a Bi-LSTM model according to the initial instance data; calculating a prototype and classifying the instance data according to the context-aware instance data to construct a context-aware prototype network based on meta learning; obtaining power grid digital resource modeling based on meta learning based on a local-global binary research and judgment-based outlier identification and marking strategy based on the context-aware prototype network based on meta learning, and realizing power grid digital resource management based on meta learning. The application helps to meet the professional needs of dispatching, equipment, marketing, development and other business lines, and has important significance for the construction of a new power system.

Description

Technical Field

[0001] This invention relates to a method for digital resource modeling and management of power grids based on meta-learning, belonging to the field of power grid digital technology. Background Technology

[0002] With the development of new power systems and the increasing demands for digital transformation management of power grids, the problems of a large quantity, diverse types, wide distribution, and difficult management of digital power grid resources are becoming increasingly prominent, posing greater challenges to the overall management, dispatching, maintenance, operation monitoring, and collaborative processing of these resources. Managing digital power grid resources aligns with the trend of digital revolution and integration, requiring the construction of a more efficient equipment asset management system to comprehensively promote the digital transformation of equipment management. However, existing methods require a large number of labeled training samples for modeling and managing digital power grid resources, yielding less than ideal results. Given the extremely complex and expensive process of labeling massive amounts of digital power grid resources, and considering the limited amount of labeled data, more efficient and accurate meta-learning-based management of digital power grid resources is crucial for meeting the professional needs of dispatching, equipment, marketing, and development departments, and is of great significance for the construction of new power systems.

[0003] Currently, the following shortcomings still exist in the digital resource modeling and management of power grids:

[0004] First, the digital resources of the power grid are numerous, diverse, widely distributed, and have complex relationships. Due to equipment failures, data loss, and other reasons, some key information is missing or abnormal, requiring the identification, supplementation, and adjustment of relevant data.

[0005] Second, most existing methods fall under the category of supervised learning, and their performance largely depends on the quantity and quality of labeled samples. However, due to the workload and high cost of labeling, large, high-quality training sets are not easily obtained in practical applications. Furthermore, current methods for managing digital grid resources cannot address the problem of new categories emerging or a limited number of instances during training. Effective solutions for managing digital grid resources using a small number of instances are still lacking.

[0006] Third, the digital resources of the power grid still have problems such as inaccurate classification and outlier data. How to effectively identify and handle outliers and how to coordinate with users to analyze abnormal situations are huge challenges facing the management of digital resources of the power grid.

[0007] The information disclosed in this background section is intended only to enhance the understanding of the overall background of the invention and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a meta-learning-based modeling and management method for digital power grid resources. First, a missing value supplementation method based on first-order difference and an outlier adjustment method based on regression fitting and median absolute deviation effectively address the missing or anomaly of key information in digital power grid resources. Second, a context-aware prototype network based on meta-learning is constructed, realizing the application of the fusion of context awareness and prototype network in the power grid field. Finally, an outlier identification and labeling strategy based on local-global binary judgment further improves the accuracy of modeling and management.

[0009] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0010] This invention discloses a method for digital resource modeling and management of power grids based on meta-learning, comprising:

[0011] The process involves acquiring digitalized power grid resources after data governance, wherein the data governance includes missing value imputation based on first-order difference and outlier adjustment based on regression fitting and median absolute deviation; the digitalized power grid resources include structured data and unstructured data.

[0012] The structured and unstructured data are respectively transformed into vectors to obtain structured data vectors and unstructured data vectors;

[0013] By concatenating the structured data vector and the unstructured data vector, initial instance data is obtained;

[0014] Based on the initial instance data, context-aware instance data is obtained using the Bi-LSTM model;

[0015] Based on the context-aware instance data, a prototype is calculated and the instance data is classified to construct a context-aware prototype network based on meta-learning.

[0016] Based on the aforementioned meta-learning-based context-aware prototype network, and using a local-to-global binary judgment-based outlier identification and labeling strategy, a meta-learning-based digital resource modeling of the power grid is obtained, thereby realizing meta-learning-based digital resource management of the power grid.

[0017] Furthermore, the missing value imputation based on first-order difference includes:

[0018] Obtain structured data {a1,a2,...,a n};

[0019] Compute the first-order difference {a′1,a′2,...,a′} of structured data n-1}, where a′1=a2-a1,...,a' n-1 =an -a n-1 ;

[0020] Structured data a with missing values ​​to be filled j The supplements include the following:

[0021]

[0022] Where p is the selection of a j The number of structured values ​​before and after, u is in a j Select the number of nearby data points with the same trend, where 0 < u < p;

[0023] case 1: Response to a j The first-order differences should be kept consistent in sign, and the average of the p data points before and after the first-order differences should be taken as a. j Missing supplementary values;

[0024] case 2: Response to a j The first-order differences are opposite in sign, let a. j The average of two adjacent structured data points is used as a. j Missing supplementary values;

[0025] case 3: Response to a j The first-order differences before and after each u-1 remain consistent in sign, and there is only one consecutive positive-to-negative change outside this range. The average of the structured data before and after each u is taken as a. j Missing supplementary values;

[0026] case 4: Response to a j The positive and negative changes of the first-order difference are discontinuous, let a j The average of two adjacent structured data points is used as a. j Missing supplementary values.

[0027] Furthermore, the outlier adjustment based on regression fitting and median absolute deviation includes:

[0028] Obtain the regression fitting curve f(x) of the structured data;

[0029] Calculate the median of the absolute bias of the structured data, i.e., γ = median(|a|). i -f(a i )|), where γ represents the median, median represents the median value, and a i Let f(a) represent the i-th structured data. i ) represents the regression fitted curve f(x) on a i The corresponding value;

[0030] The adjustments to structured data are as follows:

[0031]

[0032] case 1: Response to a i The corresponding value f(a) on the regression fitting curve i When the absolute value of the difference between ) does not exceed β*γ, that is, when f(a) satisfies i )-β*γ<a i <f(a i When )+β*γ, then a is considered to be i This is not an outlier and requires no adjustment.

[0033] case 2: Response to a i The corresponding value f(a) on the regression fitting curve i When the absolute value of the difference between ) exceeds β*γ, and satisfies a i ≥f(a i When )+β*γ, then a is considered to be i It is an outlier, take f(a) i ) as an adjustment value;

[0034] case 3: Response to a i The corresponding value f(a) on the regression fitting curve i When the absolute value of the difference between ) exceeds β*γ, and satisfies a i ≤f(a i When )-β*γ, then a is considered to be i It is a low outlier, take f(a) i ) as an adjustment value;

[0035] Where β is a coefficient. n is the amount of structured data, and n0 is the frequency of the data representing the user's location.

[0036] Furthermore, the structured data is transformed into multiple m-dimensional vectors to obtain a structured data vector; the expression for the phasor dimension is as follows:

[0037]

[0038] Where, N i It is the sum of the number of structured data and the number of words in the unstructured data in the i-th power grid digital resource, and n2 represents the number of power grid digital resources;

[0039] The expression for the structured data vector is as follows:

[0040] b1 = [a1, a2, ..., a m ]

[0041] b2 = [a m+1 ,a m+2 ,…,a 2m ]

[0042] ,…,

[0043] b r =[a (r-1)m+1 ,…,a n ,0,…,0]

[0044] Where b1 represents the first m-dimensional vector; b2 represents the second m-dimensional vector; b r Let {a1, a2, ..., a} represent the r-th m-dimensional vector; m represents the dimension; r represents the number of m-dimensional vectors; n} represents structured data, where (r-1)m<n≤rm.

[0045] Furthermore, the unstructured data undergoes vector transformation, including:

[0046] An initial vector representation of unstructured data is obtained by concatenating word embeddings and part-of-speech tagging embeddings.

[0047] Based on the initial vector representation, an updated vector representation of the unstructured data is obtained using a graph convolutional neural network based on the syntax dependency tree.

[0048] Based on the updated vector representation, and using hypergraph aggregation, we obtain the local representation and the global representation that incorporates contextual information in the unstructured data.

[0049] Based on the local and global representations, interactive processing is performed to obtain the final unstructured data vector.

[0050] Furthermore, the final expression for the unstructured data vector is as follows:

[0051] H′ i,j =tanh(W·[g i H v ]+b)

[0052] Among them, H′ i,j Let g be the vector representation of the i-th sentence and j-th word in the unstructured data vector; tanh is the non-linear activation function; W is the parameter matrix; [·; ·] represents the concatenation of two vectors; g i H represents the i-th sentence; v b is a global representation that incorporates contextual information; b is a bias vector.

[0053] Furthermore, based on the initial instance data, context-aware instance data is obtained using a Bi-LSTM model; wherein the expression for the context-aware instance data is as follows:

[0054]

[0055] Among them, D i ' is the vector representation of the i-th instance data in context awareness, D i For the i-th initial instance data, The input sequence vector is the order of the feedforward LSTM model. This is the input sequence vector that is reversed in the backward LSTM model;

[0056] The expression for the initial instance data is as follows:

[0057] D i =[H′1,...,H' L ,b1,...,b r ] = [d1,d2,...,d L+r ]

[0058] Among them, D i Let {H′1,...,H′} represent the initial instance data of the i-th instance. L Let {b1,...,b2} be the vector representation of L sentences in unstructured data. r} represents r vectors transformed from structured data.

[0059] Furthermore, based on the context-aware instance data, a prototype is calculated and the instance data is classified to construct a meta-learning-based context-aware prototype network, wherein the expression of the prototype is as follows:

[0060]

[0061] Among them, c k S is the prototype of the k-th category; k For the set of instance vectors of the k-th class in the support set; D i For the i-th initial instance data; y i For the i-th initial instance data D i The mark; D′ i This is a vector representation of the i-th instance data for context awareness.

[0062] Furthermore, a meta-learning-based digital resource modeling of the power grid is obtained, including:

[0063] Based on the context-aware prototype network, the prototype change distance after removing instance vectors is calculated to obtain local outliers;

[0064] Based on the threshold of the prototype variation distance, the overall outlier marking threshold is determined, and the overall outlier is obtained;

[0065] By adjusting the local or global outliers, a meta-learning-based model of digital power grid resources is obtained, thereby realizing meta-learning-based digital power grid resource management.

[0066] Furthermore, the expression for calculating the prototype change distance after removing instance vectors is as follows:

[0067]

[0068] Among them, c k Represents the instance vector {D′1...D′} i ...D' n The corresponding prototype; S j Represents the prototype c k Remove instance vector D' j The set of corresponding instance vectors, i.e., {D′1,...,D′} j-1 ,D' j+1 ,...,D' n}, j∈[1, n]; ||S j || represents set S j The number of instance vectors in the ρ j This indicates the removal of instance vector D' j The distance of the prototype change afterward.

[0069] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0070] This invention discloses a meta-learning-based modeling and management method for digital power grid resources. First, it effectively addresses the missing or abnormal key information of digital power grid resources by employing a missing value supplementation method based on first-order difference and an outlier adjustment method based on regression fitting and median absolute deviation. Second, it constructs a context-aware prototype network based on meta-learning, realizing the application of the fusion of context awareness and prototype networks in the power grid field. Finally, it further improves the accuracy of modeling and management by employing an outlier identification and labeling strategy based on local-global binary judgment.

[0071] This invention proposes a formula for calculating vector dimension, thereby determining the most suitable vector dimension and improving the expressive power of vectors and the training speed of meta-learning. Based on the characteristics of digital power grid resources, it utilizes unstructured data modeling based on hypergraphs to fully model local and global contextual information, resulting in richer vector representations.

[0072] This invention obtains context-aware instance data through a Bi-LSTM model; then, the instance data is mapped to a vector space through a mapping function and the prototype is calculated, which improves the speed of learning the vector space representation and achieves better generalization ability.

[0073] This invention introduces a relational metric function into the construction of a context-aware prototype network based on meta-learning. It adopts a learnable deep distance metric, which is more effective than traditional fixed metrics and can perform more accurate classification in the process of modeling complex digital resources of the power grid. Attached Figure Description

[0074] Figure 1 This is a flowchart of a digital resource modeling and management method for power grids based on meta-learning;

[0075] Figure 2 This is a flowchart of vector transformation for unstructured data. Detailed Implementation

[0076] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0077] Example

[0078] This embodiment discloses a method for digital resource modeling and management of power grids based on meta-learning, including:

[0079] The process involves acquiring digitalized power grid resources after data governance, including missing value imputation based on first-order difference and outlier adjustment based on regression fitting and median absolute deviation; the digitalized power grid resources include structured and unstructured data.

[0080] Perform vector transformations on structured and unstructured data respectively to obtain structured data vectors and unstructured data vectors;

[0081] Concatenate structured and unstructured data vectors to obtain initial instance data;

[0082] Based on the initial instance data, context-aware instance data is obtained using the Bi-LSTM model;

[0083] Based on context-aware instance data, a prototype is calculated and the instance data is classified to construct a context-aware prototype network based on meta-learning.

[0084] Based on the context-aware prototype network based on meta-learning, and the outlier identification and labeling strategy based on local-global binary judgment, a meta-learning-based model of power grid digital resources is obtained, realizing meta-learning-based power grid digital resource management.

[0085] The technical concept of this invention is as follows: First, a missing value supplementation method based on first-order difference and an outlier adjustment method based on regression fitting and median absolute deviation effectively handle the missing or anomaly of key information in digital power grid resources; second, a context-aware prototype network based on meta-learning is constructed, realizing the application of the fusion of context awareness and prototype network in the power grid field; finally, an outlier identification and labeling strategy based on local-to-global binary judgment further improves the accuracy of modeling and management.

[0086] Specific steps are as follows: Figure 1 As shown:

[0087] S01 Data governance for digital resources of the power grid.

[0088] With the development of new power systems and the advancement of digital transformation of power grids, the problems of large quantity, diverse types, wide distribution, and difficult management of digital power grid resources are becoming increasingly prominent, bringing more challenges to the overall management, scheduling and maintenance, operation monitoring, and collaborative processing of digital power grid resources.

[0089] Digital power grid resources encompass both data resources and equipment resources. Data resources include power grid data resources and power digitization network data resources, while equipment resources include digital hardware, software, virtual resources, basic support resources, spare parts, etc. Effective modeling of these resources aligns with the trend of digital revolution and integration, facilitating the digital transformation of the power grid. During the modeling process, it is necessary to establish appropriate models for the digital power grid resources based on the diverse data contained in the data and equipment resources, including grid structure, operating status, and environmental information, thereby enabling better analysis and management of related power grid resources.

[0090] The diverse data of digital power grid resources includes both structured and unstructured data. Due to issues with communication transmission channels such as networks provided by operators, hardware failures, outdated software versions, and incorrect terminal encryption configurations—all problems inherent to data acquisition equipment—as well as the low frequency of collecting critical data such as load data, some key information is missing or abnormal. Therefore, data governance of digital power grid resources is necessary to supplement missing values ​​and adjust outliers.

[0091] In this embodiment, it is assumed that the digital resources of the power grid contain n structured data {a1, a2, ..., a...} n To calculate the first difference of structured data, subtract the data from the previous period from the data of the next period, and denote the data as {a′1, a′2, ..., a′}. n-1}, where a′1=a2-a1,...,a' n-1 =a n -a n-1 .

[0092] To determine whether there are missing values ​​in the structured data of digital power grid resources, let's assume a j For missing values, a missing value imputation method based on first-order difference is used. The specific steps are as follows:

[0093] (1) Select a j p structured values ​​before and after, let a j The first-order difference of the first p values ​​is denoted as A1 = [a' j-p ,…,a' j-2 ], will a j The first-order difference of the last p values ​​is denoted as A2 = [a' j+1 ,...,a' j+p-1 ].

[0094] (2) Missing value a j The supplements are generated according to the following rules:

[0095]

[0096] Where p is the selection of a j The number of structured values ​​before and after, u is in a j Select the number of nearby data points with the same trend, where 0 < u < p;

[0097] case 1: Response to a j If the first-order differences before and after the data points are of the same sign, it indicates that the data segment formed by the p data points before and after the data points has a consistent trend. Therefore, the average of the p data points before and after the data points is taken as a. j Missing supplementary values.

[0098] case 2: Response to a j If the first-order differences are opposite in sign, it indicates that the data segments formed by the preceding and following p data points have opposite trends. Let a... j The average of two adjacent structured data points is used as a. j Missing supplementary values.

[0099] case 3: Response to a j If the first-order differences of each of the preceding and following u-1 data points maintain a consistent sign and exhibit only one consecutive positive-to-negative change outside this range, it indicates that the data segment formed by the preceding and following p data points shows a consistent trend around the missing point in the 2u points. The average of the preceding and following u structured data points is taken as a. j Missing supplementary values.

[0100] case 4: Response to a j If the positive and negative changes of the first-order difference are discontinuous, it indicates that the data segment formed by the p data points before and after the difference exhibits an oscillating trend. Let a... j The average of two adjacent structured data points is used as a. jMissing supplementary values.

[0101] After imputing the missing values, it is necessary to determine whether there are any abrupt changes in the structured data of the digitized power grid resources. If so, an outlier adjustment method based on regression fitting and median absolute deviation is used to adjust the abrupt changes. The specific steps are as follows:

[0102] (1) The structured data of the digital power grid resources are subjected to a high-order polynomial fitting algorithm of least squares. The least squares method is used to measure the deviation between the data points and the fitted curve. When the sum of the squares of the difference between the ordinate of the fitted curve and the ordinate of the data points is the smallest, the fit is considered to be the best. At this time, the regression fitted curve f(x) is obtained.

[0103] (2) Calculate the median of the absolute bias of the structured data, i.e., γ = median(a i -f(a i ), where γ represents the median, and median represents finding a. i The median value, a i Let f(a) represent the i-th structured data. i ) represents the regression fitted curve f(x) on a i The corresponding value;

[0104] (3) The following algorithm is used to adjust all data to achieve the adjustment of the mutation point:

[0105]

[0106] The rules for adjusting outliers based on regression fitting and median absolute deviation are as follows:

[0107] case 1: Response to a i The corresponding value f(a) on the regression fitting curve i When the absolute value of the difference between ) does not exceed β*γ, that is, when f(a) satisfies i )-β*γ<a i <f(a i When )+β*γ, then a is considered to be i This is not an outlier and requires no adjustment.

[0108] case 2: Response to a i The corresponding value f(a) on the regression fitting curve i When the absolute value of the difference between ) exceeds β*γ, and satisfies a i ≥f(a i When )+β*γ, then a is considered to be i The outlier is too high and needs to be adjusted downwards. Take f(a) i ) as an adjustment value;

[0109] case 3: Response to a i The corresponding value f(a) on the regression fitting curve i When the absolute value of the difference between ) exceeds β*γ, and satisfies a i ≤f(a i When )-β*γ, then a is considered to be i The outlier is too low and needs to be adjusted upwards. Take f(a) i ) as an adjustment value;

[0110] Where β is a coefficient. n is the amount of structured data, and n0 is the frequency of data for the user's location, typically 96 (default), 288, or 1440 per day, which can be configured by the user.

[0111] S02 Vector Transformation of Structured Data

[0112] In neural networks, the parameter matrix needs to be continuously adjusted to make the prediction results more and more accurate. Meta-learning is no exception. The original structured and unstructured data cannot be directly used in the training process of meta-learning, so it is necessary to extract the feature vectors of entities as the input of the neural network.

[0113] Since the vectors converted from unstructured data are input into the neural network, it is crucial to ensure the consistency of vector dimensions to prevent data format mismatches and subsequent data loss during meta-learning. Therefore, this invention does not directly convert all data into a single vector, transforming an independent set of numbers into an n-dimensional vector. Instead, it chooses to convert the structured data into r m-dimensional vectors. That is, b1 = [a1, a2, ..., a...]. m ], b2 = [a m+1 ,a m+2 ,...,a 2m ],…,b r =[a (r -1)m+1,...,a n ,0,...,0], where (r-1)m<n≤rm.

[0114] The dimension of the vector is obtained using the vector dimension calculation formula. Subsequent steps need to maintain consistency in dimension, converting unstructured data into m-dimensional vectors as well. When the selected dimension is too low, the vector representation will have significant bias because a large amount of information is lost. Conversely, when the dimension is too high, the vector representation is prone to overfitting, incorporating a large amount of noise. A series of structured numerical values ​​can be directly converted into vectors, while unstructured data requires further processing.

[0115] Therefore, this method also provides a formula for calculating vector dimensions, as follows:

[0116]

[0117] Where, N i It is the sum of the number of structured data and the number of words in the unstructured data in the i-th power grid digital resource, and n2 represents the number of power grid digital resources.

[0118] Vector transformation of S03 unstructured data

[0119] 1. Obtain an initial vector representation of unstructured data through embedding and concatenation.

[0120] Unstructured data in digital power grid resources cannot be directly converted into vectors. This invention takes text as an example and uses the concatenation of word embedding and part-of-speech tagging embedding to obtain the initial input vector of unstructured data.

[0121] In the word feature representation layer, sentence s i Each word w in ij They will all be mapped to a vector v ij To obtain a more efficient feature representation, the final word vector v is generated by concatenating the following two types of vectors. ij .

[0122] (1) Word embedding: This is a method of converting words into numerical vectors. In order to analyze them using standard machine learning algorithms, these vectors, which have been converted into numbers, need to be used as input in numerical form. The word embedding process is to embed a high-dimensional space with the number of all words into a continuous vector space with a much lower dimension. Each word or phrase is mapped to a vector in the real number field. Words that frequently appear together are mapped to nearby positions in the vector space. The result of word embedding is the generation of word vectors.

[0123] Word2Vec is an efficient model for training word vectors. It encodes each word into a vector. These encoded vectors are not randomly generated, but rather reflect the relationships between the words. Here, pre-trained vectors are used to obtain word embeddings.

[0124] (2) Part-of-speech tagging embedding: Part-of-speech tagging is the process of determining the most appropriate part-of-speech tag for each word in a given sentence. The correctness of part-of-speech tagging directly affects subsequent syntactic and semantic analysis, and is one of the foundations of information processing. The words in the corpus are classified according to their parts of speech. The part of speech of a word is determined by its meaning, morphology, and grammatical function in its language. Part-of-speech tagging is the process of determining the grammatical category of each word in a given sentence, identifying its part of speech, and tagging it. In simple terms, it involves segmenting the sentence and then marking the nature of the words, such as noun (n), verb (v), etc., after each segment. Part-of-speech tags usually contain a large amount of syntactic information, which can provide important information for digital resource modeling of the power grid. The Stanford CoreNLP model is used to obtain the part-of-speech tag for each word in the sentence and convert it into a real-valued vector.

[0125] (3) Word embedding and part-of-speech tagging embedding concatenation

[0126] By concatenating the results of word embedding and part-of-speech tagging embedding, we obtain the word w in the sentence. ij The initial word represents v ij as follows:

[0127] v ij =[q ij ,p ij ]

[0128] Among them, the word w ij It is the j-th word of the i-th sentence in unstructured data, q ij The word w ij Word embedding vector, p ij The word w ij The part-of-speech tagging embedding vector.

[0129] 2. Modeling Unstructured Data Based on Hypergraphs

[0130] Considering that different unstructured data contain the same entities, this invention proposes a hypergraph-based modeling of unstructured data for power grid digitization resources, which models the interaction between local and global contexts and derives meaningful representations of the unstructured data for each power grid digitization resource.

[0131] Hypergraphs are a fundamental component of the hypergraph-based text modeling method proposed in this invention. Most existing methods operate on a simple graph where each edge connects only two vertices. However, in many practical applications, object relationships are far more complex than pairwise relationships. To describe such complex relationships, a useful graph model is a hypergraph, where a hyperedge can connect more than two vertices. The hypergraph in this method is defined as follows:

[0132] Given a hypergraph G = (V, E, W), where V is the finite set of vertices, E is the set of hyperedges, and W is the set of weights for the hyperedges, and any hyperedge e ∈ E is a subset of the set of nodes V. It's worth noting that if every hyperedge e contains only two nodes, the hypergraph degenerates into a simple graph. Therefore, hypergraphs can be used to represent relationships between two or more nodes and the objects represented by those nodes. For each node v ∈ V, The degree of a node is represented by d(e) = |e|, where e is the number of vertices contained in the hyperedge. Hypergraph-based representation learning methods can effectively model complex relationships between more than two objects. By combining hypergraphs and deep neural networks, the performance of deep graph neural networks in complex situations is improved.

[0133] Hypergraph-based text modeling uses stacked K layers of Hypergraph Aggregation Neural Network (HANN), where K is the model's hyperparameters. In the context of machine learning, hyperparameters are parameters whose values ​​are set before the learning process begins, rather than parameters obtained through training data. Each HANN layer consists of three parts:

[0134] (1) Graph Convolutional CN based on Syntax Dependency Tree: used to model local context information, that is, to consider the semantic and syntactic information in each sentence.

[0135] Sentences s for unstructured data in power grid digital resources i Let i ∈ {1, ..., L}, where L is the number of sentences contained in the unstructured data of this resource. Using the initial representation of each word in step 1 of S03, the i-th unstructured sentence s is... i Convert to matrix X i , where X i The j-th row of the matrix represents the vector representation of the j-th word in the i-th sentence. Each word is treated as a node in the graph, and the grammatical dependency tree obtained from the open-source NLP tool StanfordCoreNLP serves as the topological structure between nodes. The adjacency matrix A represents the relationships between nodes. If node v... i With node v j If there is a syntactic dependency between A and B, then A i,j =1, otherwise 0. On the topology described in this graph, a two-layer GCN is applied to s. i By modeling the local contextual information in the sentence, we can obtain the sentence s. i New vector representation H i The representation of each word (i.e., a node in the graph) takes into account local context information. The formal representation of the two-layer GCN is as follows:

[0136]

[0137] in, I is the identity matrix of the same order as A. The degree matrix of the graph (i.e. Furthermore, ReLU is a non-linear activation function, W (0) and W (1) These are the parameter matrices for the two layers of the GCN. This invention treats directed edges in the syntax dependency tree as undirected edges and constructs an undirected graph-based GCN to model local context information. This approach effectively prevents overfitting of the model.

[0138] (2) Hypergraph aggregation: used to model global context information, that is, to consider information in multiple sentences in a document.

[0139] Due to the complexity of digital power grid resource modeling, some resource entities often appear multiple times in unstructured data. Therefore, effectively utilizing the global context information in unstructured data for modeling is fundamental to the subsequent information extraction. Here, a hypergraph is used to model the global context information throughout the unstructured data. Each sentence is considered a hyperedge in the hypergraph, and resource entities are considered nodes in the hypergraph; thus, the unstructured data can be represented as a hypergraph. Since resource entities may appear multiple times in multiple sentences of the unstructured data, the H... (The sentence is incomplete and requires further context for accurate translation). i (The j-th row in the matrix is ​​sentence s) i The vector representation of the corresponding node in the hypergraph is extracted from the vector representation of the j-th word in the hypergraph. This representation takes into account global context information.

[0140] According to the definition of a hypergraph, V is the set of nodes, and E is the set of hyperedges. Resource entities are considered as nodes in the hypergraph, and each sentence is considered as a hyperedge. Based on the representation of each word in the sentence, a max-pooling operation (MAX) is applied. pool (H i The representation of the sentence is obtained and used as the representation of the corresponding hyperedge in the hypergraph, denoted as g. i For each node v in the hypergraph, the set of its adjacent hyperedges (i.e., the sentences in which the entity corresponding to that node appears) is denoted as . according to The representation of hyperedges is learned through an attention mechanism, which teaches the contribution weight of each hyperedge to the representation of node v in the hypergraph. Assume the hyperedges... The attention score is denoted as α. i By using an aggregation function AGGREGATE(·), a global representation H of an entity considering the global context can be obtained. v(That is, the representation of nodes in the hypergraph), which incorporates global contextual information from unstructured data, i.e., considers the semantic information of the same entity throughout the unstructured data. The formal representation is as follows:

[0141]

[0142] in This indicates that the entity corresponding to node v is in sentence S. i The representation of α. i =softmax(s(g) i H v )) is for the superedge e i Attention weights for node v in the hypergraph.

[0143] (3) Interaction between Local and GlobalContexts: used to obtain richer vector representations.

[0144] This approach learns richer representations of relevant concepts in each sentence by modeling the interaction between local and global contexts. Without loss of generality, we assume that node v in the hypergraph is in sentence s. i Let j be the j-th word. By modeling the interaction between local and global contexts, a new representation of the j-th word can be obtained, expressed by the following formula:

[0145] H′ i,j =tanh(W·[g i H v ]+b)

[0146] Among them, H′ i,j Let g be the vector representation of the i-th sentence and j-th word in the unstructured data vector; tanh is the non-linear activation function; W is the parameter matrix; [·; ·] represents the concatenation of two vectors; g i H represents the i-th sentence; v b is a global representation that incorporates contextual information; b is a bias vector.

[0147] Sentences i The new representation H′ i (H′ i,j H′ i The j-th row, i.e., s i The new representation of the j-th word takes into account the information of the nodes in the hypergraph (i.e., global context information) and the sentence s. i The representation information (i.e., g) i This means that the global context information in the hypergraph aggregation step is fed back into the local representation learning.

[0148] By constructing K stacked hypergraph aggregation layers (Stacked K HANN Layers), the new representation H′ of the sentence in the previous HANN layer is obtained. i This serves as the initial representation of the nodes in the subsequent HANN layer's GCN based on the syntax dependency tree; that is, the subsequent layer uses the H′ obtained from the previous layer. i Replace the initial representation X of the sentence at this level. i This method allows for the full modeling of local and global contextual information, resulting in the learning of richer vector representations for subsequent information extraction.

[0149] S04 Structured and Unstructured Data Concatenation

[0150] After steps S01-S03, both the structured and unstructured data of the power grid digital resources have been converted into vector representations. By concatenating the vector representations of the two, initial instance data is obtained, which prepares for building a context-aware prototype network in the following section to gather information at different granularities.

[0151] The expression for the initial instance data is as follows:

[0152] D i =[H′1,...,H' L ,b1,...,b r ] = [d1,d2,...,d L+r ]

[0153] Where, {H′1,...,H' L Let {b1,...,b2} be the vector representation of L sentences in unstructured data. r} represents r vectors transformed from structured data.

[0154] It should be noted that each piece of data in the initial instance data serves as instance data for the support set or query set in the construction plot.

[0155] S05 Constructing a Context-Aware Network Based on Meta-Learning

[0156] In digital modeling, most existing methods fall under the category of supervised learning, and their performance largely depends on the quantity and quality of labeled samples. The challenge is that the high cost of manual labeling makes it impossible to provide large-scale, high-quality training sets in practical applications. Secondly, existing methods struggle when handling new categories with few or no instances during training. The traditional approach in machine learning research is to acquire a large dataset for a specific task and train a model from scratch using that dataset. However, when the scenario changes, the model needs to be retrained. To address these issues, the machine learning community has proposed a method called "meta-learning," which aims to acquire the ability to "learn by learning," enabling it to quickly learn new tasks based on existing "knowledge." When labeled data is limited, meta-learning can effectively handle few-shot learning tasks, overcoming the problem of limited sample sizes for classification or regression tasks. Using few-shot learning methods based on meta-learning in studies with limited sample sizes yields better results, therefore, the choice depends on the sample situation in digital modeling.

[0157] Meta-learning, also known as learn-to-learn, is based on the idea of ​​constructing a large number of episodes (meta-tasks) during the meta-training phase to learn meta-knowledge. This meta-knowledge is then used to improve the model's generalization ability, allowing the model to learn the commonalities of meta-tasks across different episodes while ignoring task-specific parts of the meta-tasks. After training, the trained model is tested in a meta-test, which allows for the introduction of a small number of new classes—classes not present in the meta-training.

[0158] In few-shot meta-learning, each episode contains a support set and a query set. Each support set consists of C classes selected from the training set, with K samples drawn from each of the C classes, for a total of C×K samples. Each query set selects Q samples from the training set that do not belong to any of the C classes already selected in the training set. The meta-test dataset is constructed in the same way as the meta-training dataset, except that the classes selected in the meta-test phase cannot overlap with those selected in the meta-training phase. This few-shot meta-learning is then called a C-way K-shot problem.

[0159] Modeling sequence vectors using traditional Long Short-Term Memory (LSTM) models has a problem: it cannot encode information from back to front. In finer-grained classification, attention needs to be paid to the interactions between preceding and following sequence vectors. Therefore, this invention employs a bidirectional Long Short-Term Memory (Bi-LSTM) model, combining forward and backward LSTM models, which can better capture bidirectional semantic dependencies. This applies to the digitized power grid resource data D... i d j (1≤j≤L+r) are sequentially input into the forward LSTM model and the backward LSTM model, respectively. The hidden states are calculated in the LSTM models in both directions. The final hidden states generated by the forward and backward LSTMs are concatenated as the initial instance data D of the power grid digital resource data. i The final vector representation D i ':

[0160]

[0161] Among them, D i ' is the vector representation of the i-th instance data in context awareness, D i For the i-th initial instance data, The input sequence vector is the order of the feedforward LSTM model. This is the input sequence vector that is reversed in the backward LSTM model.

[0162] S06 Constructing a Prototype Network Based on Meta-Learning

[0163] This invention employs prototypical networks, a metric-based meta-learning method. The idea behind this model is to map instance data from each category in the support set to a vector space using a mapping function, calculate the mean of the instance data by category, and use this mean as the prototype for each category. Then, it classifies the instance data in the query set based on a relational metric function. Training ensures that instance data within the same category is closest to its own category's prototype vector while being farthest from the prototypes of other categories. First, after step S05, the initial instance data D has been... i Transform into a context-aware instance data vector representation D i This achieves information extraction from the original data. Next, the vector for each class in the support set (using S) is calculated. k The average value of (represented by c) is used to obtain the prototype (using c). k (Representation). It is formalized as follows:

[0164]

[0165] Among them, c k S is the prototype of the k-th category; k For the set of instance vectors of the k-th class in the support set; D i For the i-th initial instance data; y i For the i-th initial instance data D i The mark; D i 'A vector representation of context-aware instance data.

[0166] Steps S02-S05 effectively map instances to the vector space, thereby bringing digitized power grid resources of the same class closer together spatially, where they would otherwise be far apart. Therefore, c k This can be viewed as a representation of cluster centers. During training, the samples are labeled, meaning we know the label y of the sample. i And the prototype to which it belongs. Next, the representation f of each instance in the query set is calculated using a metric function (denoted as d(·,·)). θ (x i ) and c k The distance between them is calculated, and the query instance x is obtained by calculating SoftMax. i Probability of belonging to the prototype:

[0167]

[0168] In prototype networks, d(·,·) typically uses squared Euclidean distance as a traditional metric. However, studies such as relational networks have shown that learnable deep distance metrics are more effective than fixed metrics when assigning categories to instances in a query set. Therefore, this invention employs a relational metric.

[0169] In the relational metric function, the input is each pair of prototypes c in the query set obtained in the above steps. k and instance vector D i The output is a scalar between 0 and 1, representing the instance vector D. i 'and prototype c k The score between them. The main structure of the relation metric function contains a neural tensor layer, which has a significant advantage in measuring the relationship between two vectors. The k-th prototype c k and the i-th instance vector D′ i Relationship score r ik It is expressed as follows:

[0170] r ik =d(D′) i c k )

[0171]

[0172] Where M h ∈R D*D h∈[1,…,D] represents a slice of tensor parameters, M [1:D] ∈R D*D*D Represent a tensor, Represents vector e∈R Q , f represents the non-linear activation function ReLU, and W and b are the parameters of the fully connected layer. The sigmoid function is used as the activation function.

[0173] After each training round, a loss function is needed to estimate the degree of discrepancy between the model's predictions and the true values. This loss function is a non-negative real-valued function; the smaller the loss function, the better the model's robustness. A loss function suitable for the model and the scenario can accelerate the model's convergence speed and improve training effectiveness.

[0174] This invention uses the log-log loss function as the loss function to train the model to complete the regression task. The log-log loss function can very well represent the probability distribution and is very suitable for handling multiple scenarios, especially multi-class classification. The loss function is formalized as follows:

[0175] J(θ)=-log(P θ (y i =k|D i |))

[0176] During the training and testing phases, no parameter updates or tuning are performed. The entire model is trained on a plot-by-plot basis, and its minimum value is found using stochastic gradient descent, thus converging to learn a better result. At this point, it can be assumed that after training, this projection function can project samples of the same type to a location with a relatively close proximity. Based on the plot-based loss, the parameters in the framework are updated using the backpropagation algorithm.

[0177] S07 Outlier Identification and Tagging

[0178] For instance vectors belonging to the prototype, there are some cases of inaccurate classification. An outlier identification and labeling strategy based on local-global binary judgment will be used. The specific steps are as follows:

[0179] Using the prototype c of the kth category k For example, S k For the prototype c belonging to the k-th category k The set of instance vectors, i.e., {D′1…D′ i …D' n}, where n represents belonging to prototype c k The number of instance vectors.

[0180] In order, those belonging to c k After removing the instance vectors, we get a set of n removed instance vectors, denoted as S1…S2. i …S n Where S1 is represented as {D′2…D′ i …D' n}, S n Represented as {D′1…D′ i …D' n-1}; Calculate the prototype change distance after removing instance vectors.

[0181] Calculate the removal of an instance vector D' j The formula for the subsequent prototype change distance is as follows:

[0182]

[0183] Among them, c k Represents the instance vector {D′1…D′ i …D' n The corresponding prototype; S j Represents the prototype c k Remove instance vector D' j The set of corresponding instance vectors, i.e., {D′1,…,D′} j-1 ,D' j+1 ,…,D' n}, j∈[1, n]; ||S j || represents set S j The number of instance vectors in the ρ j This indicates the removal of instance vector D' j The distance of the prototype change afterward.

[0184] The calculated prototype variation distances {ρ1,ρ2,...,ρ} will be obtained. n After sorting in ascending order, we get {ρ′1,ρ′2,...,ρ′}. n} will be ranked in The distances are denoted as ρ″1, ρ″2, ρ″3, respectively, and n represents the prototype c. k The number of instance vectors.

[0185] This strategy will identify and label outliers from both local and global perspectives.

[0186] (1) For prototype c k The instance with the largest variation distance ρ' n To determine whether a value is an outlier, the instance vector corresponding to that value is considered a local outlier if it meets the following criteria:

[0187] ρ′n >θmax(ρ′1,ρ′2,...,ρ′ n-1 )

[0188] θ = 3 is the default value, but users can also configure it according to their actual needs.

[0189] (2) Determine the outlier marking threshold ρ″0 based on the prototype variation distance threshold calculation algorithm. The algorithm is as follows:

[0190]

[0191] Calculate the outlier labeling threshold for all prototypes sequentially and use the maximum value as the overall outlier labeling threshold ρ0. This will satisfy ρ... i All instance vectors greater than ρ0 are marked as global outliers.

[0192] Afterwards, the local and overall outliers will be submitted to the user for evaluation. The user can adjust the outliers to regular points according to the actual situation.

[0193] In summary, this invention first proposes a data governance strategy for digital power grid resources. It employs a missing value imputation method based on first-order difference and an outlier adjustment method based on regression fitting and median absolute deviation to handle missing or anomalies in key information. A formula for calculating vector dimension is proposed to determine the optimal vector dimension, thereby improving the expressive power of vectors and the training speed of meta-learning. Based on the characteristics of digital power grid resources, it uses unstructured data modeling based on hypergraphs to fully model local and global contextual information, resulting in richer vector representations.

[0194] Secondly, this invention realizes the application of a meta-learning-based fusion strategy of context-aware networks and prototype networks (including a meta-learning-based Bi-LSTM, a prototype network, and an introduced relation module) in the power grid field. A meta-learning-based context-aware network is constructed according to the characteristics of digital power grid resources. A Bi-LSTM is used to train a mapping function that converts instances into spatial vectors, obtaining the final vector representation of the digital power grid resource data. A meta-learning-based prototype network is constructed to map instances to the vector space through the mapping function and calculate the prototype for each category, improving the speed of learning the vector space representation and achieving better generalization ability. The introduction of a relation module trains a more meaningful distance function; learnable deep distance metrics are more effective than fixed metrics, enabling more accurate classification in the complex process of modeling digital power grid resources.

[0195] Finally, this invention proposes an outlier identification and labeling strategy based on a local-to-global binary analysis. For prototype-based digital power grid resources, some classifications may be inaccurate. The strategy involves calculating the prototype variation distance after removing instances based on both local and global conditions, and determining the outlier labeling threshold to handle instances with inaccurate classifications.

[0196] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0197] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0198] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0199] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0200] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for digital resource modeling and management of power grids based on meta-learning, characterized in that, include: The process involves acquiring digitalized power grid resources after data governance, wherein the data governance includes missing value imputation based on first-order difference and outlier adjustment based on regression fitting and median absolute deviation; the digitalized power grid resources include structured data and unstructured data. The missing value imputation based on first-order difference includes: Obtaining structured data ; Computing the first difference of structured data ,in ; Structured data with missing values ​​to be filled The supplements include the following: ； Where p is the selection The number of structured values ​​before and after, u is Select the number of nearby data points with the same trend and ; case 1: Response to The first-order differences should be kept consistent in sign, and the average of the p data points before and after should be taken as the mean. Missing supplementary values; case 2: Response to The first-order differences are opposite in sign, take The average of two adjacent structured data points is used as Missing supplementary values; case 3: Response to The first-order differences before and after each u-1 remain consistent in sign, and there is only one consecutive positive-to-negative change outside this range. The average of the structured data before and after each u is taken as... Missing supplementary values; case 4: Response to The positive and negative changes of the first-order difference are discontinuous, so we take... The average of two adjacent structured data points is used as Missing supplementary values; The outlier adjustment based on regression fitting and median absolute deviation includes: Obtaining regression fitting curves from structured data ; Calculate the median of the absolute deviation of the structured data, i.e. ,in, Indicates the median value. This indicates finding the median value. This represents the i-th structured data. For regression fitting curve superior The corresponding value; The adjustments to structured data are as follows: ； case 1: response to Corresponding value on the regression fitting curve The absolute value of the difference between them does not exceed When, that is, satisfied At that time, it is determined that This is not an outlier and requires no adjustment. case 2: Response to Corresponding value on the regression fitting curve The absolute value of the difference between them exceeds At that time, and satisfying At that time, it is determined that It is an outlier, take As an adjustment value; case 3: Response to Corresponding value on the regression fitting curve The absolute value of the difference between them exceeds At that time, and satisfying At that time, it is determined that It is a low outlier, take As an adjustment value; in, For coefficients, , It is the amount of structured data. It is the frequency of data taken from the user's location; The structured and unstructured data are respectively transformed into vectors to obtain structured data vectors and unstructured data vectors; By concatenating the structured data vector and the unstructured data vector, initial instance data is obtained; Based on the initial instance data, context-aware instance data is obtained using the Bi-LSTM model; Based on the context-aware instance data, a prototype is calculated and the instance data is classified to construct a context-aware prototype network based on meta-learning. Based on the aforementioned meta-learning-based context-aware prototype network, and using a local-to-global binary judgment-based outlier identification and labeling strategy, a meta-learning-based digital resource modeling of the power grid is obtained, thereby realizing meta-learning-based digital resource management of the power grid.

2. The method for digital resource modeling and management of power grids based on meta-learning according to claim 1, characterized in that, The structured data is transformed into multiple m-dimensional vectors to obtain a structured data vector; the expression for the phasor dimension is as follows: ； in, It is the sum of the number of structured data and the number of words in the unstructured data of the i-th digital power grid resource. This indicates the quantity of digital resources in the power grid; The expression for the structured data vector is as follows: ； in, This represents the first m-dimensional vector; This represents the second m-dimensional vector; This represents the r-th m-dimensional vector; m represents the dimension; r represents the number of m-dimensional vectors. Representing structured data, .

3. The method for digital resource modeling and management of power grids based on meta-learning according to claim 2, characterized in that, The unstructured data is transformed into a vector, including: An initial vector representation of unstructured data is obtained by concatenating word embeddings and part-of-speech tagging embeddings. Based on the initial vector representation, an updated vector representation of the unstructured data is obtained using a graph convolutional neural network based on the syntax dependency tree. Based on the updated vector representation, and using hypergraph aggregation, we obtain the local representation and the global representation that incorporates contextual information in the unstructured data. Based on the local and global representations, interactive processing is performed to obtain the final unstructured data vector.

4. The method for digital resource modeling and management of power grids based on meta-learning according to claim 3, characterized in that, The final expression for the unstructured data vector is as follows: ； in, This is the vector representation of the j-th word of the i-th sentence in an unstructured data vector. It is a non-linear activation function; For parameter matrices; This indicates the concatenation of two vectors; Let i represent the i-th sentence; A global representation that incorporates contextual information; This is the bias vector.

5. The method for digital resource modeling and management of power grids based on meta-learning according to claim 4, characterized in that, Based on the initial instance data, context-aware instance data is obtained using a Bi-LSTM model; wherein the expression for the context-aware instance data is as follows: ； in, For the context-aware vector representation of the i-th instance data, For the i-th initial instance data, The input sequence vector is the order of the feedforward LSTM model. This is the input sequence vector that is reversed in the backward LSTM model; The expression for the initial instance data is as follows: ； in, This represents the initial instance data of the i-th instance. It is a vector representation of L sentences in unstructured data. It is r vectors converted from structured data.

6. The method for digital resource modeling and management of power grids based on meta-learning according to claim 5, characterized in that, Based on the context-aware instance data, a prototype is calculated and the instance data is classified to construct a meta-learning-based context-aware prototype network, wherein the expression of the prototype is as follows: ； in, This is the prototype of the k-th category; To support the set of instance vectors of the k-th class in the set; This is the initial instance data for the i-th instance; For the i-th initial instance data The mark; This is a vector representation of the i-th instance data for context awareness.

7. The method for digital resource modeling and management of power grids based on meta-learning according to claim 6, characterized in that, The resulting meta-learning-based digital resource modeling of the power grid includes: Based on the context-aware prototype network, the prototype change distance after removing instance vectors is calculated to obtain local outliers; Based on the threshold of the prototype variation distance, the overall outlier marking threshold is determined, and the overall outlier is obtained; By adjusting the local or global outliers, a meta-learning-based model of digital power grid resources is obtained, thereby realizing meta-learning-based digital power grid resource management.

8. The method for digital resource modeling and management of power grids based on meta-learning according to claim 7, characterized in that, The expression for calculating the prototype change distance after removing instance vectors is as follows: ； in, Represents instance vectors The corresponding prototype; Representing the prototype Remove instance vectors The set of corresponding instance vectors, i.e. ; Represents a set The number of instance vectors in the middle; Represents the removal of instance vectors The distance of the prototype change afterward.

Images