An intelligent completion method and device for power grid missing data

Abstract

The present application belongs to the technical field of electrical engineering, and particularly relates to an intelligent completion method and device for power grid missing data. The method comprises the following steps: extracting complete sample data to perform random missing processing and marking a missing area, and constructing a global missing data set and a complete data set in pairs; obtaining a local missing data set based on the global missing data set and the marking range of the missing area; performing global completion on the missing area based on the global missing data set to obtain global completion data; performing local completion on the missing area based on the local missing data set to obtain local completion data; inputting the global completion data and the local completion data into a discriminator for identification, and alternately training a global generator and a local generator and the discriminator to generate a global-local generative adversarial network; and completing the missing area of the power grid based on the global-local generative adversarial network. The present application adds a local completion branch in a traditional generative adversarial network, thereby improving the accuracy and continuity of completion.

Description

Technical Field

[0001] This invention belongs to the field of electrical engineering technology, specifically relating to an intelligent method and device for completing missing data in the power grid. Background Technology

[0002] The completeness of power grid data is a prerequisite for subsequent intelligent decision-making and precise management of the power grid. Currently, traditional statistical methods are commonly used to address the problem of missing power grid data, such as discarding incomplete data, zeroing out missing values, linear interpolation, curve fitting, and nearest neighbor mean. Discarding incomplete data often reduces the data volume, which is detrimental to subsequent analysis and modeling. Zeroing out missing values ​​can easily lead to outliers, becoming a potential factor restricting model robustness. A series of interpolation-based methods typically only consider the correlation under a single attribute, ignoring the holistic nature of multi-attribute data. Nearest neighbor mean, on the other hand, selects the mean of samples closest in spatial distance to complete the data, treating power grid data as discrete samples and ignoring the continuity of data over time. Traditional statistical methods are inefficient at data completion and cannot achieve batch completion.

[0003] Existing methods for power grid data completion based on deep learning algorithms exist. Application No. CN202011115656.X proposes a distribution network big data repair method based on adversarial game theory. Application No. CN202111425843.2 proposes a method and system for power metering data completion based on generative adversarial networks (GANs). Wang Shouxiang et al., in their paper "A Method for Reconstructing Missing Measurement Data in Power Systems Using Improved Generative Adversarial Networks," proposed a method based on an improved generative adversarial network (WGAN). This method automatically learns complex spatiotemporal relationships that are difficult to explicitly model, such as correlations between measurements and load fluctuation patterns, through unsupervised training, and uses realism constraints and context similarity constraints to enable the generator to generate high-precision reconstructed data. Yang Yulian et al., in their paper "Reconstruction of Missing Measurement Data in Distribution Networks Based on Generative Adversarial and Dual Semantic Awareness," proposed a method that autonomously extracts data features using unsupervised generative adversarial training and combines this with dual semantic awareness reconstruction constraints to achieve missing data reconstruction. These power grid data completion methods rely on generative adversarial network methods to complete missing power grid data. These methods are more efficient than traditional statistical-based data completion methods, enabling batch completion of multi-attribute data. However, power grid data is often a continuous process, requiring temporal continuity in data completion. The aforementioned methods still have shortcomings; data completion using neural networks often fails to consider the local continuity of power grid data. Intelligent methods for batch completion based on the proximity principle are needed. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an intelligent data completion method and apparatus for missing power grid data. By adding a local completion branch to the generative adversarial network, the missing data is filled in using neighboring data, thereby improving the accuracy and continuity of the completion.

[0005] To address the shortcomings of existing technologies, the technical solution provided by this invention is as follows:

[0006] A method for intelligent data completion for missing power grid data includes,

[0007] S1: Construct a global-local generative adversarial network, including;

[0008] Extract complete sample data, perform random missing data processing, and label the missing regions to construct a pair of global missing datasets and complete datasets; obtain local missing datasets based on the global missing datasets and the labeled ranges of the missing regions;

[0009] The global generator performs global completion on the missing regions based on the global missing dataset, resulting in globally completed data; the local generator performs local completion on the missing regions based on the local missing dataset, resulting in locally completed data.

[0010] Global and local completion data are input into the discriminator for recognition. The global generator, local generator, and discriminator are trained alternately to generate a global-local generative adversarial network.

[0011] S2: Complete the missing regions of the power grid based on global-local generative adversarial networks.

[0012] Preferably, the local missing dataset obtained based on the global missing dataset and the labeled range of the missing regions includes,

[0013] The target detection network is trained based on the global missing dataset and the labeled range of the missing regions to obtain the predicted coordinates of the missing bounding boxes of the missing regions.

[0014] The object detection network obtains the mask X of the missing region based on the predicted bounding box coordinates of the missing region. mask ;

[0015] Mask X based on missing region mask and the globally missing dataset X m Generate a locally missing dataset X mask ⊙X m .

[0016] Preferably, the object detection network is constructed based on the fast R-CNN neural network, and the objective function is as follows:

[0017]

[0018] Wherein L(p,u,t) u (v) represents the loss function, u represents the number of categories, u = 1, 2, ..., K, where K is the total number of categories, p u λ represents the predicted missing probability of power grid data of category u, λ represents the regularization parameter, [u≥1] is the Iverson bracket, and smooth_L1 is the loss function. Let v be the coordinates of the missing border of the missing region of category u, i∈{a,b,c,d}, where a represents the x-coordinate of the top-left vertex of the missing border, b represents the y-coordinate of the top-left vertex of the missing border, c represents the width of the missing border, d represents the height of the missing border, and v i These are the coordinates of the missing borders for the predicted missing region of category u;

[0019] L(p,u,t u When v reaches its minimum value, the object detection network training is complete, and it outputs the predicted bounding box coordinates v of the missing region. i .

[0020] Preferably, the global generator performs global completion on the missing regions based on the global missing dataset to obtain globally completed data, including:

[0021] A global generator G is constructed using a residual network. G , globally missing dataset X m Input global generator G G After passing through the global generator G G The convolutional encoding yields the global completion data G. G (X m );

[0022] The local generator performs local completion on the missing regions based on the local missing dataset, obtaining locally completed data, including:

[0023] A local generator G is constructed using a residual network. L , to partially missing dataset X mask ⊙X m Input local generator G L After passing through the local generator G L The convolutional encoding yields the locally completed data G. L (X mask ⊙X m ).

[0024] Preferably, the objective functions of the global generator and the local generator are as follows:

[0025]

[0026] Among them, X mask ⊙G G(X m ) represents the global completion data G G (X m In the mask X of the missing area mask Complete the data within.

[0027] Preferably, the objective function of the discriminator is as follows:

[0028]

[0029] Among them, X mask ⊙G G (X m ) represents the global completion data G G (X m In the mask X of the missing area mask Complete the data within; X mask ⊙X c Represents the complete dataset X c Mask X in the missing area mask The actual data within.

[0030] An intelligent data completion device for missing power grid data includes a data preprocessing module and a global-local generative adversarial network; the global-local generative adversarial network includes a target detection network, a global generator, a local generator, and a discriminator;

[0031] The data preprocessing module is used to extract complete sample data, perform random missing data processing, and label missing regions to construct a pair of global missing datasets and complete datasets.

[0032] The target detection network is used to obtain a local missing dataset based on the global missing dataset and the labeled range of the missing regions;

[0033] The global generator is used to perform global completion on the missing regions based on the global missing dataset, obtain global completion data, and input the global completion data into the discriminator;

[0034] The local generator is used to locally complete the missing regions based on the local missing dataset, obtain the local completed data, and input the local completed data into the discriminator;

[0035] The discriminator is used to identify global and local completion data based on the complete dataset;

[0036] The global generator, local generator, and discriminator are trained alternately to generate the global-local generative adversarial network;

[0037] Global-local generative adversarial networks are used to complete missing areas of the power grid.

[0038] Preferably, the target detection network is specifically used for,

[0039] Training was completed based on the global missing dataset and the labeled range of the missing regions, and the coordinates of the missing bounding boxes of the predicted missing regions were obtained.

[0040] The mask X of the missing region is obtained based on the predicted missing region bounding box coordinates. mask ;

[0041] Mask X based on missing region mask and the globally missing dataset X m Generate a locally missing dataset X mask ⊙X m ;

[0042] The object detection network is built based on the fast R-CNN neural network, and its objective function is as follows:

[0043]

[0044] Wherein L(p,u,t) u (v) represents the loss function, u represents the number of categories, u = 1, 2, ..., K, where K is the total number of categories, p u λ represents the predicted missing probability of power grid data of category u, λ represents the regularization parameter, [u≥1] is the Iverson bracket, and smooth_L1 is the loss function. Let v be the coordinates of the missing border of the missing region of category u, i∈{a,b,c,d}, where a represents the x-coordinate of the top-left vertex of the missing border, b represents the y-coordinate of the top-left vertex of the missing border, c represents the width of the missing border, d represents the height of the missing border, and v i These are the coordinates of the missing borders for the predicted missing region of category u;

[0045] L(p,u,t u When v reaches its minimum value, the object detection network training is complete, and it outputs the predicted bounding box coordinates v of the missing region. i .

[0046] Preferably, the global generator is constructed using a residual network, specifically used for:

[0047] Input the globally missing dataset X m After convolutional encoding, the globally completed data G is obtained. G (X m );

[0048] The local generator is constructed based on a residual network and is specifically used for:

[0049] Input the partially missing dataset X mask ⊙Xm After convolutional encoding, the locally completed data G is obtained. L (X mask ⊙X m ).

[0050] Preferably, the objective functions of the global generator and the local generator are as follows:

[0051]

[0052] Among them, X mask ⊙G G (X m ) represents the global completion data G G (X m In the mask X of the missing area mask Complete the data within;

[0053] The objective function of the discriminator is as follows:

[0054]

[0055] Among them, X mask ⊙X c Represents the complete dataset X c Mask X in the missing area mask The actual data within.

[0056] The beneficial effects of this invention are:

[0057] This invention adds a local completion branch to the traditional generative adversarial network. It identifies missing data through an object detection network and uses neighboring data of the missing data to fill in the gaps. Global completion and local completion are trained together to generate a global-local generative adversarial completion network, which improves the accuracy and continuity of completion. Attached Figure Description

[0058] Figure 1 The flowchart for constructing a global-local generative adversarial network provided by this invention. Detailed Implementation

[0059] The present invention will be further described below with reference to the embodiments. 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.

[0060] A method for intelligent data completion for missing power grid data includes the following steps:

[0061] S1: Construct a global-local generative adversarial network, including;

[0062] 11) Extract the complete sample data, perform random missing data processing, and label the missing regions to construct paired global missing datasets and complete datasets:

[0063] A single data point in the power grid at a given moment is considered a sample. The power grid data can be represented as X∈R n×d Where n represents the number of samples and d represents the number of attributes of the power grid data, i.e., the feature dimension.

[0064] By constructing a one-to-one correspondence between globally missing and fully complete datasets, we can support the ability of global-local generative adversarial networks to learn to complete incomplete data into realistic, complete data. The original power grid data can be divided into complete and incomplete sample data, with the complete sample data used as training data and the incomplete samples used as real data to be completed.

[0065] This invention labels complete sample data as a complete dataset X. c Meanwhile, for the complete dataset X c The global missing dataset X is obtained by randomly selecting missing values. m The proportion of discarded values ​​is the same as the missing rate in the actual imputation data, and all missing values ​​are set to zero to simulate missing values ​​in a real-world environment. The global missing dataset X is then used. m and the complete dataset X c By matching them one-to-one, pairs of globally missing and complete datasets are formed, which serve as the training set (X) for the global-local generative adversarial network. m ,X c ).

[0066] This invention also addresses the globally missing dataset X. m The missing region is labeled by marking the missing region with a rectangular missing box, and the coordinates of the missing box are marked. The coordinates of the missing box are used by the subsequent object detection network to identify the missing range.

[0067] To eliminate the adverse effects caused by excessive differences in values ​​under different characteristics, it is necessary to normalize the power grid data to unify the data distribution.

[0068] 12) The global generator performs global completion on the missing regions based on the global missing dataset, resulting in globally completed data:

[0069] A global generator G is constructed using a residual network. G Global missing dataset X m Input global generator G G After passing through the global generator G G The convolutional encoding yields the global completion data G. G (X m ).

[0070] 13) A local missing dataset is obtained based on the global missing dataset and the labeled range of the missing regions; the local generator performs local completion on the missing regions based on the local missing dataset to obtain the local completed data:

[0071] Fast R-CNN neural network was used as the object detection network to identify locally missing data sets.

[0072] Using the globally missing dataset X m The object detection network is trained using the labeled range of the missing region to obtain the predicted bounding box coordinates of the missing region. The objective function of the object detection network is expressed as follows:

[0073]

[0074] L(p,u,t u (v) represents the loss function, where u is the category, u = 1, 2, ..., K, K is the total number of categories, p refers to the probability of missing data in the predicted power grid data, and t u is the coordinates of the missing border of the missing region of category u, v is the predicted coordinates of the missing border of the missing region of category u, and p u λ represents the predicted missing probability of power grid data of category u, λ represents the regularization parameter, [u≥1] is the Iverson bracket, smooth_L1 is the loss function, and t i u Let v be the coordinates of the missing border of the missing region of category u, i∈{a,b,c,d}, where a represents the x-coordinate of the top-left vertex of the missing border, b represents the y-coordinate of the top-left vertex of the missing border, c represents the width of the missing border, d represents the height of the missing border, and v i These are the coordinates of the missing bounding boxes for the predicted category u. The objective function is L(p,u,t). u When v reaches its minimum value, the object detection network training is complete, and it outputs the predicted bounding box coordinates v of the missing region. i .

[0075] Object detection networks detect missing bounding box coordinates v of the predicted missing regions. i Further obtain the mask X of the missing area mask This is used to narrow down the scope of the missing region.

[0076] The mask X of the missing region obtained by the object detection network mask and the globally missing dataset X m Combine to generate a locally missing dataset X mask ⊙X m A local generator G is constructed using a residual network. L X mask ⊙X mInput to local generator G L In the middle, generate locally completed data G L (X mask ⊙X m ).

[0077] Local generators fill in gaps in locally missing datasets by using adjacent complete data based on the principle of proximity. Compared to global completion, local completion increases the accuracy of the completion.

[0078] 14) Input the global completion data and local completion data into the discriminator for recognition. The global generator, local generator, and discriminator are trained alternately to generate a global-local generative adversarial network.

[0079] Similar to local completion, global completion also requires dividing the missing regions and using the mask X of the missing regions obtained by the object detection network. mask The global completion data is divided into parts, denoted as X. mask ⊙G G (X m ), representing global completion data G G (X m In the mask X of the missing area mask Complete the data within.

[0080] The global completion result and the local completion result are superimposed to form the comprehensive completion result X. mask ⊙G G (X m )+

[0081] G L (X mask ⊙X m The data is input into the discriminator D. The global and local generators need to maximize their completion performance to "deceive" the discriminator D into recognizing it as real, complete data. The objective function for training the global and local generators is as follows:

[0082]

[0083] The objective function of the global generator and the local generator is D(X) mask ⊙G G (X m )+G L (X mask ⊙X m The formula is )) = 1 to ensure that the completion error is minimized. Where, D(X) = 1. mask ⊙G G (X m )+G L (X mask ⊙Xm )) indicates that the discriminator evaluates the comprehensive completion result X. mask ⊙G G (X m )+G L (X mask ⊙X m The prediction of X when the discriminator selects X mask ⊙G G (X m )+G L (X mask ⊙X m When ) is identified as real data, D(X) mask ⊙G G (X m )+G L (X mask ⊙X m ))=1.

[0084] For the discriminator D, this invention employs a 1D convolutional classification network. The discriminator needs to improve its judgment ability, accurately distinguishing whether the data is forged. Its task mainly includes two aspects: firstly, accurately identifying real and complete data; and secondly, distinguishing whether the data is incomplete. Therefore, its objective function consists of two parts, expressed by the following formula:

[0085]

[0086] Among them, X mask ⊙X c Represents the complete dataset X c Mask X in the missing area mask The discriminator expects to process the real, complete data X. mask ⊙X c , D(X mask ⊙X c The value of D(X) can approach 1, while for completed data, D(X) can approach 1. mask ⊙G G (X m )+G L (X mask ⊙X m It can approach 0 to maximize the discrimination ability.

[0087] Considering that the process of completing and judging data integrity is a continuous adversarial one, the global generator, local generator, and discriminator need to be trained alternately to train the generative adversarial network. When, with the help of the discriminator, the global and local generators have been trained to the point that the discriminator cannot correctly determine whether the data has been completed, that is, when the objective functions of the global generator, local generator, and discriminator have all converged, the model iteration can be stopped. At this point, the network enters a state of "Nash equilibrium".

[0088] When the global-local generative adversarial completion network converges, the final completion result of the missing data can be represented as:

[0089] X mask ⊙G G (X m )+G L (X mask ⊙X m )+(1-X mask )⊙X m

[0090] Power grid data is a continuously recorded, regular dataset with a certain degree of continuity over time. The likelihood of sudden, extreme values ​​is low, and missing values ​​are often similar to those in nearby periods. Therefore, power grid data completion needs to consider the principle of proximity. This invention adds a local completion branch to a traditional generative adversarial network (GAN). It identifies missing data through a target detection network and uses neighboring data to fill in the gaps. Global and local completion are jointly trained to generate a global-local generative adversarial completion network, improving the accuracy and continuity of the completion process.

[0091] S2: Complete the missing regions of the power grid using a global-local generative adversarial network. When completing the missing regions, the actual data to be completed is used as the global missing dataset. This global missing dataset is input into a global generator to obtain the global completion result. An object detection network is used to identify the coordinates of missing borders in the global missing dataset and obtain the local missing dataset. This local missing dataset is then input into a local generator for local completion to obtain the local completion result. Finally, the combined global and local completion results are used to obtain the final completion result.

[0092] X mask ⊙G G (X m )+G L (X mask ⊙X m )+(1-X mask )⊙X m

[0093] The processes of global and local completion are similar to those of training a global-local generative adversarial completion network, and will not be described in detail here.

[0094] Existing technologies for using neural networks to complete power grid data all perform batch data completion. This is disadvantageous when only a few missing values ​​need to be completed in massive datasets, as redundant noise in massive datasets has a significant impact on small numbers of missing values. For small numbers of missing values, directly using data surrounding the missing data for completion is more accurate and efficient. This invention combines global and local completion methods, resulting in greater accuracy and efficiency compared to existing technologies.

[0095] This invention also provides an intelligent data completion device for missing power grid data, including a data preprocessing module and a global-local generative adversarial network; the global-local generative adversarial network includes a target detection network, a global generator, a local generator, and a discriminator;

[0096] The data preprocessing module is used to extract complete sample data, perform random missing data processing, and label missing regions to construct pairs of global missing datasets and complete datasets.

[0097] The object detection network is used to obtain a local missing dataset based on the global missing dataset and the labeled range of the missing regions;

[0098] The global generator is used to perform global completion on the missing regions based on the global missing dataset, obtain global completion data, and input the global completion data into the discriminator;

[0099] The local generator is used to locally complete the missing regions based on the local missing dataset, obtain the local completed data, and input the local completed data into the discriminator;

[0100] The discriminator is used to identify global and local completion data based on the complete dataset;

[0101] The global generator, local generator, and discriminator are trained alternately to generate the global-local generative adversarial network;

[0102] Global-local generative adversarial networks are used to complete missing areas of the power grid.

[0103] The object detection network is specifically used for,

[0104] Training was completed based on the global missing dataset and the labeled range of the missing regions, and the coordinates of the missing bounding boxes of the predicted missing regions were obtained.

[0105] The mask X of the missing region is obtained based on the predicted missing region bounding box coordinates. mask ;

[0106] Mask X based on missing region mask and the globally missing dataset X m Generate a locally missing dataset X mask ⊙X m ;

[0107] The object detection network is built based on the fast R-CNN neural network, and its objective function is as follows:

[0108]

[0109] Wherein L(p,u,t) u (v) represents the loss function, u represents the number of categories, u = 1, 2, ..., K, where K is the total number of categories, p u λ represents the predicted missing probability of power grid data of category u, λ represents the regularization parameter, [u≥1] is the Iverson bracket, and smooth_L1 is the loss function. Let v be the coordinates of the missing border of the missing region of category u, i∈{a,b,c,d}, where a represents the x-coordinate of the top-left vertex of the missing border, b represents the y-coordinate of the top-left vertex of the missing border, c represents the width of the missing border, d represents the height of the missing border, and v i These are the coordinates of the missing borders for the predicted missing region of category u;

[0110] L(p,u,t u When v reaches its minimum value, the object detection network training is complete, and it outputs the predicted bounding box coordinates v of the missing region. i .

[0111] The global generator is constructed using a residual network, specifically for...

[0112] Input the globally missing dataset X m After convolutional encoding, the globally completed data G is obtained. G (X m );

[0113] The local generator is constructed based on a residual network and is specifically used for:

[0114] Input the partially missing dataset X mask ⊙X m After convolutional encoding, the locally completed data G is obtained. L (X mask ⊙X m ).

[0115] The objective functions for the global generator and the local generator are as follows:

[0116]

[0117] Among them, Xmask ⊙G G (X m ) represents the global completion data G G (X m In the mask X of the missing area mask Complete the data within;

[0118] The objective function of the discriminator is as follows:

[0119]

[0120] Among them, X mask ⊙X c Represents the complete dataset X c Mask X in the missing area mask The actual data within.

[0121] 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.

[0122] 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.

[0123] 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.

[0124] 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.

[0125] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

[0126] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent data completion for missing data in power grids, characterized in that, include, S1: Construct a global-local generative adversarial network, including; Extract complete sample data, perform random missing data processing and label missing regions to construct paired global missing datasets and complete datasets; A local missing dataset is obtained based on the global missing dataset and the labeled range of the missing regions; The global generator performs global completion on the missing regions based on the global missing dataset, resulting in globally completed data; the local generator performs local completion on the missing regions based on the local missing dataset, resulting in locally completed data. Global and local completion data are input into the discriminator for recognition. The global generator, local generator, and discriminator are trained alternately to generate a global-local generative adversarial network. S2: Complete missing regions of the power grid based on a global-local generative adversarial network; The local missing dataset obtained based on the global missing dataset and the labeled range of the missing regions includes, The target detection network is trained based on the global missing dataset and the labeled range of the missing regions to obtain the predicted coordinates of the missing bounding boxes of the missing regions. The object detection network obtains a mask for the missing regions based on the predicted bounding box coordinates of the missing regions. ; Mask based on missing region and global missing dataset Generate locally missing datasets ; The object detection network is built based on the fast R-CNN neural network, and its objective function is as follows: in, Represents the loss function. It is a category. , The total number of categories, The predicted category is The probability of missing power grid data, Represents the regular expression parameter. It's Iverson (in brackets). For loss function, It is a category The coordinates of the missing border of the missing region. ,in This indicates the missing top-left corner vertex of the border. coordinate, This indicates the missing top-left corner vertex of the border. coordinate, Indicates the width of the missing border. Indicates the height of the missing border. The predicted category is The coordinates of the missing border of the missing region; The object detection network completes training when it reaches the minimum value, and simultaneously outputs the coordinates of the predicted missing bounding boxes of the missing regions. .

2. The intelligent data completion method for missing power grid data according to claim 1, characterized in that, The global generator performs global completion on the missing regions based on the global missing dataset, obtaining globally completed data, including: A global generator is constructed using a residual network. , global missing dataset Input global generator After passing through the global generator Convolutional encoding yields globally completed data. ; The local generator performs local completion on the missing regions based on the local missing dataset, obtaining locally completed data, including: A local generator is constructed using a residual network. To partially missing datasets Input local generator After local generator Convolutional encoding yields locally completed data. .

3. The intelligent data completion method for missing power grid data according to claim 2, characterized in that, The objective functions of the global generator and the local generator are as follows: in, Indicates global completion of data. Mask in the missing area Complete the data within.

4. The intelligent data completion method for missing power grid data according to claim 2, characterized in that, The objective function of the discriminator is as follows: in, Indicates global completion of data. Mask in the missing area Complete the data within; Represents the complete dataset Mask in the missing area The actual data within.

5. An intelligent data completion device for missing power grid data, characterized in that, It includes a data preprocessing module and a global-local generative adversarial network; the global-local generative adversarial network includes a target detection network, a global generator, a local generator, and a discriminator; The data preprocessing module is used to extract complete sample data, perform random missing data processing, and label missing regions to construct a pair of global missing datasets and complete datasets. The target detection network is used to obtain a local missing dataset based on the global missing dataset and the labeled range of the missing regions; The global generator is used to perform global completion on the missing regions based on the global missing dataset, obtain global completion data, and input the global completion data into the discriminator; The local generator is used to locally complete the missing regions based on the local missing dataset, obtain the local completed data, and input the local completed data into the discriminator; The discriminator is used to identify global and local completion data based on the complete dataset; The global generator, local generator, and discriminator are trained alternately to generate the global-local generative adversarial network; Global-local generative adversarial networks are used to complete missing areas of the power grid; The target detection network is specifically used for, Training was completed based on the global missing dataset and the labeled range of the missing regions, and the coordinates of the missing bounding boxes of the predicted missing regions were obtained. The mask for the missing region is obtained based on the predicted coordinates of the missing border data. ; Mask based on missing region and global missing dataset Generate locally missing datasets ; The object detection network is built based on the fast R-CNN neural network, and its objective function is as follows: in, Represents the loss function. It is a category. , The total number of categories, The predicted category is The probability of missing power grid data, Represents the regular expression parameter. It's Iverson (in brackets). For loss function, It is a category The coordinates of the missing border of the missing region. ,in This indicates the missing top-left corner vertex of the border. coordinate, This indicates the missing top-left corner vertex of the border. coordinate, Indicates the width of the missing border. Indicates the height of the missing border. The predicted category is The coordinates of the missing border of the missing region; The object detection network completes training when it reaches the minimum value, and simultaneously outputs the coordinates of the predicted missing bounding boxes of the missing regions. .

6. The intelligent data completion device for missing power grid data according to claim 5, characterized in that, The global generator is constructed using a residual network, specifically for... Input global missing dataset After convolutional encoding, global completion data is obtained. ; The local generator is constructed based on a residual network and is specifically used for: Input partially missing dataset After convolutional encoding, locally completed data is obtained. .

7. The intelligent data completion device for missing power grid data according to claim 6, characterized in that, The objective functions of the global generator and the local generator are as follows: in, Indicates global completion of data. Mask in the missing area Complete the data within; The objective function of the discriminator is as follows: in, Represents the complete dataset Mask in the missing area The actual data within.

Images