CN-122020534-A - Power grid data processing method and system

CN122020534ACN 122020534 ACN122020534 ACN 122020534ACN-122020534-A

Abstract

The invention discloses a power grid data processing method and system, which comprises the steps of constructing a noise matrix and a vacancy matrix based on power grid data, obtaining an input matrix based on the noise matrix and the vacancy matrix, obtaining space-time fusion characteristics of the power grid data based on the input matrix, adding the space-time fusion characteristics of the power grid data into the vacancy matrix, respectively separating electrical characteristics and environmental characteristics from the vacancy matrix, generating a high-frequency environmental quantity predicted value based on the environmental characteristics, aligning the high-frequency environmental quantity predicted value with the electrical characteristics on a time axis, fusing the high-frequency environmental quantity predicted value with the electrical characteristics to obtain fusion characteristics, explicitly embedding a physical rule loss function into a loss function of a space-time attention network, and processing the fusion characteristics based on the space-time attention network.

Inventors

LIU QIANG
ZHANG RUIFANG
ZHAO WENYI
LI JIAKANG
YANG YUNQI
ZHANG CHENG
REN RUIJIE
LUO XIAOHAN
LIANG YULONG
MA JIA

Assignees

国网陕西省电力有限公司铜川供电公司

Dates

Publication Date: 20260512
Application Date: 20260127

Claims (10)

1. A method for processing grid data, comprising: constructing a noise matrix and a vacancy matrix based on the power grid data, obtaining an input matrix based on the noise matrix and the vacancy matrix, and obtaining space-time fusion characteristics of the power grid data based on the input matrix; Adding the space-time fusion characteristic of the power grid data into the vacancy matrix, respectively separating an electrical characteristic and an environmental characteristic from the vacancy matrix, generating a high-frequency environmental quantity predicted value based on the environmental characteristic, and aligning the high-frequency environmental quantity predicted value with the electrical characteristic on a time axis; And explicitly embedding a physical rule loss function into a loss function of a space-time attention network, and processing the fusion characteristic based on the space-time attention network.
2. A power grid data processing method as defined in claim 1, wherein, The constructing a noise matrix and a vacancy matrix based on the grid data includes: Determining abnormal data of the power grid data based on physical constraints corresponding to different power grid data, and constructing a noise matrix based on the abnormal data of the power grid data; And constructing a power grid topological graph based on the equipment ID and the equipment physical connection, calculating the smoothness of graph signals of the power grid topological graph, obtaining a partial graph smoothness anomaly score, and obtaining a vacancy matrix based on the partial graph smoothness anomaly score.
3. A power grid data processing method as defined in claim 1, wherein, The method comprises the steps of obtaining space-time fusion characteristics of power grid data based on an input matrix, extracting space dimension characteristics of the power grid data based on the input matrix, extracting time dimension characteristics of the power grid data based on the space dimension characteristics of the power grid data, and fusing the space dimension characteristics and the time dimension characteristics of the power grid data into the space-time fusion characteristics of the power grid data.
4. A power grid data processing method as defined in claim 2, wherein, Determining abnormal data of the power grid data based on physical constraints corresponding to different power grid data, and constructing a noise matrix based on the abnormal data of the power grid data, wherein the noise matrix comprises the following components: determining tolerance coefficients corresponding to different power grid data based on the power grid data; Traversing the power grid data, and determining abnormal value data in the power grid data based on the tolerance coefficient; setting the abnormal data as NAN, and constructing a noise matrix; The noise matrix consists of Personal device on In time steps And (5) observing the sequence composition of the characteristics.
5. A power grid data processing method as defined in claim 2, wherein, The calculation formula of the smoothness anomaly score of the partial graph is as follows: ; Wherein, the Represent the first Is characterized in that Time of day, apparatus Local graph smoothness anomaly scores at that point, Representation device At the position of Time of day feature Is used to determine the actual observed value of (a), Representation device Is a set of first-order neighbor devices, Representing neighbor devices The observed values at the same time point are, Representation device With neighbor devices Normalized spatial weights between; the obtaining the vacancy matrix based on the partial graph smoothness anomaly score comprises the following steps: Setting an abnormality threshold If the partial graph smoothness anomaly score Then, the judgment is that the device is abnormal, Setting the matrix to be a null value to obtain a null matrix 。
6. A power grid data processing method as defined in claim 1, wherein, The obtaining an input matrix based on the noise matrix and the null matrix includes: constructing a binary mask matrix based on the noise matrix; null matrix with zero values Filling the blank value to obtain an initialization tensor ; The tensor will be initialized Splicing the binary mask matrix in the characteristic dimension to obtain an input matrix ; The obtaining the space-time fusion characteristic of the power grid data based on the input matrix comprises the following steps: computing device With neighbor devices The correlation coefficient between the two is calculated as: ; Wherein, the For neighbour devices To the equipment Is used for the importance of the coefficient of (a), Representing devices separately And neighbor device Is used for the feature vector of (a), As a matrix of learnable linear transformation weights, As a learnable attention vector, Is a nonlinear activation function; Normalizing the coefficients by a Softmax function to obtain a final attention weight ; The multi-head weighting aggregation has the following calculation formula: ; Wherein, the Is a device A high-dimensional feature representation of spatial information fused to a neighborhood, For the characteristic splicing operation, the method comprises the following steps, As a function of the non-linear activation, In order to pay attention to the number of heads, Is the first The normalized attention weights calculated by the individual attention heads, Is the first A weight matrix of the individual attention headers, For neighbour devices Is a feature vector of (1); All devices of the whole network At all times Is characterized by the aggregation of (3) Stacking to form spatial feature tensors ; Feature extraction is carried out based on Bi-LSTM network, and the calculation formula is as follows: ; ; ; Wherein, the Is a spatial feature tensor At the moment of time Is used for the slicing of the slice, To be at the moment The spatio-temporal fusion characteristics of the grid data, For the moment of time Is arranged to be in a forward hidden state, For the moment of time Is arranged to be in a forward hidden state, For the moment of time Is in a backward hidden state of (1), For the moment of time Is hidden in the backward direction.
7. A power grid data processing method as defined in claim 1, wherein, The generating a high-frequency environment quantity predicted value based on the environment characteristic includes: using MLP to observe low frequency environment Mapping to a potential state space to obtain a potential state of the environment observation value at a low-frequency moment; The ODE solver is utilized to integrate and deduce the potential state of the environment observation value at the high-frequency target moment by taking the potential state of the environment observation value at the low-frequency moment as an initial value, the potential state of the environment observation value at the high-frequency target moment is a high-frequency environment quantity predicted value, and a deduction formula is as follows: ; Wherein, the Is a neural network that is parameterized, Indicating an initial potential state of the device, As integral variable, representing slave To the point of Is used for the continuous time of (a), Is in an intermediate state and is represented at a moment in the integrating path Is a potential state of (c).
8. A power grid data processing method as defined in claim 1, wherein, The expression of the physical rule loss function is as follows: ; Wherein, the As a function of the loss of the physical rule, For the constraint of the flow equation, For the kirchhoff's current law constraint, In the constraint of the law of ohms, In order to run the boundary constraint(s), As weight parameters, carrying out self-adaptive determination through a dynamic weight adjustment mechanism; the expression of the constraint of the tide equation is as follows: ; Wherein, the The power is represented by a value representing the power, The voltage is represented by a voltage value, The current is represented by a value representing the current, The number of the devices is represented by the number, Represents the square of the L2 norm; the expression of the kirchhoff law constraint is: ; Wherein, the Representing slave devices Flow direction device Is set to the branch current value of (1), Representing slave devices Flow direction device Branch current values of (2); The expression of the ohm's law constraint is: ; Wherein, the Representing the number of bus lines in the electrical network, Representation device The voltage phase difference across the two ends, Representing slave devices Flow direction device Is set to the branch current value of (1), Representation device The physical impedance phase difference at the two ends; The expression of the operation boundary constraint is: ; Wherein, the The number of the devices is represented by the number, Apparatus and method for controlling the operation of a device Is used for the voltage of the (c) transformer, Representing the upper and lower physical security limits allowed by the device, respectively; The expression of the dynamic weight adjustment mechanism is as follows: ; Wherein, the Represent the first The degree of violation of the individual physical constraints, Indicating a tolerance threshold value, Representing the sensitivity coefficient.
9. A power grid data processing method as defined in claim 8, wherein, The expression of the loss function of the physical rule loss function explicitly embedded into the space-time attention network is as follows: ; Wherein, the In order to account for the total loss, To null matrix with zero values The task of performing the null value filling is lost, Is the weight of the physical loss.
10. A power grid data processing system, comprising: The data preprocessing module is used for constructing a noise matrix and a vacancy matrix based on the power grid data, obtaining an input matrix based on the noise matrix and the vacancy matrix, and obtaining space-time fusion characteristics of the power grid data based on the input matrix; the data fusion module is used for adding the space-time fusion characteristic of the power grid data into the vacancy matrix, respectively separating an electrical characteristic and an environmental characteristic from the vacancy matrix, generating a high-frequency environmental quantity predicted value based on the environmental characteristic, and aligning the high-frequency environmental quantity predicted value with the electrical characteristic on a time axis; And the data processing module is used for explicitly embedding the loss function of the physical rule into the loss function of the space-time attention network and processing the fusion characteristic based on the space-time attention network.

Description

Power grid data processing method and system Technical Field The invention relates to the technical field of power distribution networks, in particular to a power grid data processing method and system. Background With the rapid development of smart power grids, a power grid fusion terminal (such as a distribution network smart terminal, an internet of things sensing device and the like) is used as a core node for data acquisition, and is widely deployed in various links of power generation, power transmission, power transformation, power distribution and power utilization, so that real-time acquisition and transmission of mass data such as power grid running states, device working conditions, environment parameters and the like are realized. However, due to factors such as device performance difference, communication interference, environmental noise, human misoperation and the like, the data collected by the terminal generally has the problems of non-standard data and non-uniform format, abnormal value interference, data deletion and the like. The irregular data and the irregular format increase the data preprocessing cost, reduce the analysis efficiency, even cause erroneous judgment (such as temperature threshold overrun alarm error) due to unit confusion, cause model training deviation due to abnormal value interference and even trigger false alarm (such as equipment overload protection misoperation), and damage the data time sequence integrity due to data missing, thereby affecting the accuracy of trend analysis (such as load prediction) and state estimation (such as tidal current calculation). For the problems, because the prior art lacks a unified data quality evaluation standard, indexes such as data integrity, accuracy, consistency and the like are difficult to quantify, so that the data credibility cannot be ensured. Downstream applications (such as fault diagnosis, optimal scheduling) may make false decisions based on low quality data, increasing grid operational risk. Disclosure of Invention Aiming at the problems of the background technology, the invention provides a power grid data processing method and a system, which are used for solving the problems of non-standardization and non-uniform format, abnormal value interference, data deletion and the like of data acquired by a power grid fusion terminal and improving the accuracy of abnormal detection. The technical scheme disclosed by the invention is as follows: a method of grid data processing, comprising: constructing a noise matrix and a vacancy matrix based on the power grid data, obtaining an input matrix based on the noise matrix and the vacancy matrix, and obtaining space-time fusion characteristics of the power grid data based on the input matrix; Adding the space-time fusion characteristic of the power grid data into the vacancy matrix, respectively separating an electrical characteristic and an environmental characteristic from the vacancy matrix, generating a high-frequency environmental quantity predicted value based on the environmental characteristic, and aligning the high-frequency environmental quantity predicted value with the electrical characteristic on a time axis; And explicitly embedding a physical rule loss function into a loss function of a space-time attention network, and processing the fusion characteristic based on the space-time attention network. Further, the constructing a noise matrix and a vacancy matrix based on the grid data includes: Determining abnormal data of the power grid data based on physical constraints corresponding to different power grid data, and constructing a noise matrix based on the abnormal data of the power grid data; And constructing a power grid topological graph based on the equipment ID and the equipment physical connection, calculating the smoothness of graph signals of the power grid topological graph, obtaining a partial graph smoothness anomaly score, and obtaining a vacancy matrix based on the partial graph smoothness anomaly score. Further, the method for obtaining the space-time fusion feature of the power grid data based on the input matrix comprises the steps of extracting space dimension features of the power grid data based on the input matrix, extracting time dimension features of the power grid data based on the space dimension features of the power grid data, and fusing the space dimension features and the time dimension features of the power grid data into the space-time fusion feature of the power grid data. Further, the determining the abnormal data of the power grid data based on the physical constraints corresponding to different power grid data, and constructing the noise matrix based on the abnormal data of the power grid data includes: determining tolerance coefficients corresponding to different power grid data based on the power grid data; Traversing the power grid data, and determining abnormal value data in the power grid data based on the tolerance coefficien