CN-115964953-B - Meta-learning-based power grid digital resource modeling management method

CN115964953BCN 115964953 BCN115964953 BCN 115964953BCN-115964953-B

Abstract

The invention discloses a power grid digital resource modeling management method based on meta-learning, which comprises the steps of obtaining power grid digital resources after data management, respectively carrying out vector conversion on structural data and unstructured data to obtain initial instance data, obtaining context-aware instance data based on a Bi-LSTM model according to the initial instance data, calculating prototypes and classifying the instance data according to the context-aware instance data to construct a context-aware prototyping network based on meta-learning, and obtaining power grid digital resource modeling based on the meta-learning according to an outlier identification and marking strategy based on local integral binary research and judgment based on the context-aware prototyping network. The invention is beneficial to meeting the professional demands of business lines such as dispatching, equipment, marketing, development and the like, and has important significance for the construction of novel power systems.

Inventors

CHEN LONG
DENG XIAO
LIN FENG
WANG XIAOYAN
LI SHENGSHENG
SHE YUNBO
SHI KANG
CHEN GANG
GE CHAOHUI
TU JINJIN

Assignees

南瑞集团有限公司
南京南瑞信息通信科技有限公司

Dates

Publication Date: 20260508
Application Date: 20230110

Claims (8)

1. A power grid digital resource modeling management method based on meta learning is characterized by comprising the following steps: acquiring a power grid digital resource after data management, wherein the data management comprises deficiency value supplementation based on first order difference and abnormal point adjustment based on regression fitting and median absolute deviation; the first-order difference-based missing value supplement includes: Obtaining structured data ; Computing first order differences of structured data Wherein ; Structured data to be complemented with missing values The supplement of (2) comprises the following steps: ; wherein p is selected from The number of the front and back structuring values, u, is Selecting the number of the data with consistent trend nearby ; Case1 response to The positive and negative of the front and back first-order differences are kept consistent, taking the average value of p data before and after Is a missing supplemental value of (2); case 2 response to The front and back first-order differences are reversed in positive and negative directions and taken Average of two adjacent structured data as Is a missing supplemental value of (2); case3 response to The positive and negative of each u-1 first-order difference are consistent, and continuous positive and negative changes only occur once outside the range, and the average value of each u structural data is taken as the average value Is a missing supplemental value of (2); case4 response to Discontinuous positive and negative changes of front and back first-order difference are taken Average of two adjacent structured data as Is a missing supplemental value of (2); The abnormal point adjustment based on regression fit and median absolute deviation comprises the following steps: Regression fit curve for structured data acquisition ; Calculating the median of the absolute deviation of the structured data, i.e , wherein, The median value is represented by a median value, The representation is a median value of the values, Representing the i-th structured data of the data, Fitting a curve to regression Upper part Corresponding values of (2); The structured data is adjusted as follows: ; case 1 response to Corresponding values on regression fit curve The absolute value of the difference between them does not exceed When it is satisfied that When it is determined that Is not an abnormal point and does not need to be adjusted; case 2 response to Corresponding values on regression fit curve The absolute value of the difference between them exceeds When and satisfy When it is determined that Is a point of high abnormality and is taken As an adjustment value; case 3 response to Corresponding values on regression fit curve The absolute value of the difference between them exceeds When and satisfy When it is determined that Is a low abnormal point, take As an adjustment value; Wherein, the As the coefficient of the light-emitting diode, , Is the amount of structured data that is to be structured, The value frequency of the data of the user location; vector conversion is carried out on the structured data and the unstructured data respectively, so that structured data vectors and unstructured data vectors are obtained; Splicing the structured data vector and the unstructured data vector to obtain initial instance data; obtaining context-aware instance data based on a Bi-LSTM model according to the initial instance data; According to the context-aware instance data, calculating a prototype and classifying the instance data, and constructing a context-aware prototype network based on meta-learning; And according to the context awareness prototype network based on meta learning, the outlier identification and marking strategy based on local whole binary research and judgment is used for obtaining the power grid digital resource modeling based on meta learning, so that the power grid digital resource management based on meta learning is realized.
2. The method for modeling and managing the digital resources of the power grid based on the meta-learning according to claim 1, wherein the method is characterized in that the structured data is subjected to vector conversion, the structured data is converted into a plurality of m-dimensional vectors, and the structured data vectors are obtained, and the expression of phasor dimensions is as follows: ; Wherein, the Is the sum of the number of structured data in the ith grid digitized resource and the number of words in unstructured data, Representing the number of digital resources of the power grid; the expression of the structured data vector is as follows: ; Wherein, the Represents the 1 st m-dimensional vector; Representing the 2 nd m-dimensional vector; represents the r-th m-dimensional vector, m represents the dimension, r represents the number of m-dimensional vectors; the data is represented in a structured form such that, 。
3. The meta-learning-based power grid digital resource modeling management method according to claim 2, wherein performing vector conversion on the unstructured data comprises: Splicing word embedding and part-of-speech tagging embedding is adopted to obtain an initial vector representation of unstructured data; according to the initial vector representation, a graph convolution neural network based on a grammar dependence tree obtains updated vector representation of unstructured data; obtaining local representation in unstructured data and global representation fusing context information based on hypergraph aggregation according to the updated vector representation; And carrying out interaction processing according to the local representation and the global representation to obtain a final unstructured data vector.
4. A method of managing digital resources of a power grid based on meta-learning as set forth in claim 3, wherein the expression of the final unstructured data vector is as follows: ; Wherein, the A vector representation of a jth word of an ith sentence in the unstructured data vector; Is a nonlinear activation function; is a parameter matrix; Representing a concatenation of two vectors; Is a representation of the ith sentence; a global representation for fusing context information; is a bias vector.
5. The method for modeling and managing power grid digital resources based on meta learning according to claim 4, wherein the context-aware instance data is obtained based on a Bi-LSTM model according to the initial instance data, and the expression of the context-aware instance data is as follows: ; Wherein, the For vector representation of the i-th instance data of context awareness, For the i-th initial instance data, For the forward LSTM model sequential input sequence vectors, An input sequence vector reverse to the backward LSTM model; The expression of the initial instance data is as follows: ; Wherein, the Representing the data of the i-th initial instance, Is a vector representation of L sentences in unstructured data, Is the r vectors into which the structured data is converted.
6. The method for managing power grid digital resource modeling based on meta learning according to claim 5, wherein a prototype is calculated and classified according to the context-aware instance data, and a context-aware prototype network based on meta learning is constructed, wherein an expression of the prototype is as follows: ; Wherein, the Prototype of the kth category; a set of instance vectors for a kth category in the support set; Is the ith initial instance data; for the ith initial instance data Is a mark of (2); is a vector representation of the i-th instance data of context awareness.
7. The meta-learning-based power grid digital resource modeling management method as claimed in claim 6, wherein obtaining the meta-learning-based power grid digital resource modeling comprises: According to the context-aware prototype network, calculating a prototype variation distance after eliminating the instance vector to obtain a local outlier; Determining a marking threshold value of the integral outlier according to the threshold value of the prototype variation distance to obtain the integral outlier; and adjusting the local outlier or the integral outlier to obtain the power grid digital resource modeling based on meta-learning, and realizing power grid digital resource management based on meta-learning.
8. The method for managing power grid digital resource modeling based on meta-learning according to claim 7, wherein the expression of the prototype variation distance after the instance vector is removed is calculated as follows: ; Wherein, the Representing instance vectors A corresponding prototype; Representing prototypes Eliminating instance vectors The set of post-corresponding instance vectors, i.e ; Representing a collection The number of instance vectors; representing culled instance vectors The prototype after which varies the distance.

Description

Meta-learning-based power grid digital resource modeling management method Technical Field The invention relates to a power grid digital resource modeling management method based on meta learning, and belongs to the technical field of power grid digitization. Background With the development of a novel power system and the continuous improvement of the digital transformation management requirement of a power grid, the problems of multiple quantity, various types, wide distribution, difficult management and the like of the digital resources of the power grid are more remarkable at present, and more challenges are brought to the aspects of overall management, scheduling maintenance, operation monitoring, collaborative processing and the like of the digital resources of the power grid. The management of the digital resources of the power grid conforms to the development trend of the digital revolution fusion, a more efficient equipment asset management system is constructed, and the digital transformation of the equipment management is comprehensively promoted. However, the existing method needs a large number of marked training samples to perform modeling management on digital resources of the power grid, and the obtained result is not ideal. In view of the fact that the marking of massive power grid digital resources is extremely complicated and expensive, under the condition that marking data are limited, the power grid digital resource management based on meta-learning is more efficient and accurate, is helpful for meeting professional requirements of business lines such as scheduling, equipment, marketing and development, and has important significance for the construction of novel power systems. At present, the following defects exist in the aspect of power grid digital resource modeling management: 1. the digital resources of the power grid are large in quantity, heterogeneous, wide in distribution and complex in association relation, and due to the reasons of equipment body faults, transmission data loss and the like, part of critical information is lost or abnormal, and related data need to be identified, supplemented and adjusted. 2. Most of the existing methods fall into the category of supervised learning, the performance of which depends to a large extent on the number and quality of the marked samples. However, due to the workload and high cost caused by marking, a large high-quality training set is not easy to obtain in practical application, and the existing management method for the power grid digital resources at present cannot solve the problem that new categories or examples are few in the training process, and an effective solution is not available for managing the power grid digital resources by using a small number of examples. 3. The problems of inaccurate classification, data outliers and the like of the power grid digital resources exist, and how to effectively identify and process outliers and how to cooperate with users to study and judge abnormal conditions is a great challenge for power grid digital resource management. The information disclosed in this background section is only for enhancement of understanding of the general background of the invention and should not be taken as an acknowledgement or any form of suggestion that this information forms the prior art already known to a person of ordinary skill in the art. Disclosure of Invention The invention aims to overcome the defects in the prior art, and provides a power grid digital resource modeling management method based on meta-learning, wherein firstly, a missing value supplementing method based on first order difference and an outlier adjusting method based on regression fitting and median absolute deviation are used for effectively processing the missing or abnormality of key information of power grid digital resources, secondly, a context perception prototype network based on meta-learning is constructed, the application of fusion of context perception and prototype network in the power grid field is realized, and finally, the accuracy of modeling management is further improved based on an outlier identification and marking strategy of local integral binary research and judgment. In order to achieve the above purpose, the invention is realized by adopting the following technical scheme: The invention discloses a power grid digital resource modeling management method based on meta learning, which comprises the following steps: acquiring a power grid digital resource after data management, wherein the data management comprises deficiency value supplementation based on first order difference and abnormal point adjustment based on regression fitting and median absolute deviation; vector conversion is carried out on the structured data and the unstructured data respectively, so that structured data vectors and unstructured data vectors are obtained; Splicing the structured data vector and the unstructured data vector to obt