Power grid data checking method, system and equipment based on multi-interleaving deep learning model and storage medium

Through the multi-interleaved deep learning model, the intrinsic and timing correlation of power grid data is extracted using CNN and GRU to dynamically generate thresholds, which solves the problem of low efficiency in detecting abnormal data in traditional distribution networks, and realizes rapid response and high-accuracy power grid data calibration nuclear, ensures safe operation of the power grid.

CN118940053APending Publication Date: 2024-11-12NARI INFORMATION & COMM TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411071221.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-06
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

Traditional distribution networks need to set thresholds artificially when detecting abnormal data, and the missing data filling efficiency is low, making it difficult to achieve real-time monitoring and ensure safe grid operation.

Method used

The power grid data verification method based on the multi-interleaved deep learning model is adopted. By pre-processing and feature extraction of the power grid data, the intrinsic relationship and timing correlation between the data are extracted by CNN and GRU, and the determination threshold is dynamically generated to achieve real-time monitoring of abnormal data.

Benefits of technology

It realizes rapid response and accurate identification of grid data abnormalities, improves data verification efficiency and accuracy, reduces labor costs, and ensures safe operation of the grid.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118940053A_ABST
    Figure CN118940053A_ABST
Patent Text Reader

Abstract

The invention discloses a power grid data checking method, system and equipment based on a multi-interleaving deep learning model and a storage medium, and the method comprises the steps: taking power grid data at i sampling moments as i groups of historical data, taking a group of power grid data at the (i + 1) th sampling moment as a group of prediction data, and copying the group of prediction data for i copies as i groups of prediction data; preprocessing each group of historical data and each group of prediction data to obtain respective embedded matrixes with fixed sizes; performing feature extraction and maximum pooling processing on the embedded matrix corresponding to each group of historical data and each group of prediction data by adopting a trained multi-interleaving deep learning model to obtain a vector representation form corresponding to each group of historical data and each group of prediction data; and then a similarity calculation formula is adopted to calculate a matching degree score between the historical data and the prediction data to judge whether the power grid data is abnormal, so that the purpose of data checking is achieved, and the method has the advantages of high response speed, high recognition accuracy and the like.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of power grid data verification in power systems, and specifically relates to a power grid data verification method, system, equipment and storage medium based on multiple interwoven deep learning models. Background Art

[0002] Convolutional Neural Network (CNN) is a deep learning technology that is mainly used to process images and other two-dimensional data. It has achieved great success in computer vision, image recognition, and pattern classification. CNN consists of convolutional layers, pooling layers, and fully connected layers. The convolutional layer is the core of CNN, which uses convolution kernels (also called filters) to scan the input image and extract features from it. The convolution operation effectively reduces the number of parameters that need to be trained, making CNN suitable for large-scale data; the pooling layer is used to reduce the size of the feature map while retaining important information. The most common pooling operation is maximum pooling, which selects the maximum value in each region as the output, which helps to reduce the amount of calculation and improve the robustness of the model; the fully connected layer is used after convolution and pooling to convert the feature map into the final classification or regression result. The neurons in the fully connected layer are connected to all the neurons in the previous layer to form a densely connected network. CNN has a wide range of applications, including face recognition, autonomous driving, medical image analysis, etc. Its success is partly due to its ability to effectively capture local features in images and to train large-scale data sets.

[0003] The Gated Recurrent Unit (GRU) is a simplified and powerful Recurrent Neural Network (RNN) structure designed to better capture the larger dependencies between time steps in a time series. GRU introduces the concept of gates, and its structure mainly consists of four parts: reset gate, update gate, candidate hidden state, and hidden state. The role of the reset gate is to help capture short-term dependencies in a time series; the role of the update gate is to capture long-term dependencies in a time series; the candidate hidden state is calculated based on the current input and the previous hidden state; the hidden state is the final output state. Each hidden unit of the GRU has an independent reset gate and update gate, so it can learn to capture dependencies in different time ranges. As a variant of RNN, GRU has shown amazing capabilities in natural language processing, speech recognition, time series prediction and other fields.

[0004] In the power grid system, improving the utilization rate of the grid and reducing the loss of the grid are one of the problems that each power company needs to solve urgently. Due to the large number of electricity users, the heavy load in the distribution network continues to increase. Affected by various uncertain factors in the network, the voltage in the power system is prone to instability. If the voltage is too high, serious accidents such as voltage collapse and fire may occur. Therefore, how to detect abnormal data in the power grid in a timely manner has become the key to protecting the safety of the power grid.

[0005] A large amount of grid data is generated during the operation of the power grid system, such as temperature, humidity, pressure, ripple, vacuum, magnetic field, thermocouple, thermal resistor, voltage, current, resistance, frequency and other operating data. These data are not only related to each other, but also have certain connections in their dynamic changes within a certain time range. When detecting abnormal data, traditional distribution networks need to manually set the threshold of abnormal data, and the filling of missing data is also inefficient. Summary of the invention

[0006] Purpose of the invention: In order to solve the problems that traditional distribution networks need to manually set thresholds for abnormal data when detecting abnormal data, and the filling of missing data is also inefficient, the present invention proposes a power grid data verification method, system, equipment and storage medium based on a multiple interleaved deep learning model based on the characteristics of power grid data, which utilizes the inherent connection between power grid data and the connection between data in time series to dynamically generate a judgment threshold for each data, thereby achieving the purpose of real-time monitoring of abnormal data in the power grid and ensuring the safe operation of the power grid.

[0007] Technical solution: A power grid data verification method based on multiple interwoven deep learning models, comprising the following steps:

[0008] Step 1: All power grid data at the same sampling time is called a group of data. The power grid data at the i sampling time is taken as the i-group historical data, and the group of power grid data at the i+1 sampling time is taken as a group of predicted data. This group of predicted data is copied i times as the i-group predicted data;

[0009] Step 2: Preprocess each set of historical data and each set of predicted data to obtain their respective fixed-size embedding matrices;

[0010] Step 3: Use the trained multiple interleaved deep learning model to perform feature extraction and maximum pooling processing on the embedding matrix corresponding to each group of historical data and each group of predicted data to obtain the vector representation corresponding to each group of historical data and each group of predicted data; then use the similarity calculation formula to calculate the degree of matching between each group of historical data and its corresponding group of predicted data, and output the power grid data verification result according to the degree of matching.

[0011] Furthermore, the preprocessing of each set of historical data and each set of predicted data to obtain respective embedding matrices of fixed size includes:

[0012] First, clean each set of historical data and each set of forecast data;

[0013] Secondly, use the gensim tool to get a suitable embedding model;

[0014] Finally, the embedding model maps each set of historical data and each set of predicted data into an embedding matrix of fixed length through table lookup. For the power grid data that cannot find a corresponding relationship in the embedding model, it will be mapped into a zero vector of the same length.

[0015] Furthermore, the multiple interwoven deep learning model is composed of the superposition of multiple single interwoven networks; the input of each single interwoven network is the embedding matrix obtained in step 2, and the input embedding matrix is ​​passed through a bidirectional gated recurrent unit Bi-GRU to obtain a new embedding matrix of the same size as the original embedding matrix that pays attention to the correlation between each sampling moment within the embedding matrix; the new embedding matrix is ​​added to the input embedding matrix and sent to the convolutional neural network CNN to extract the correlation features between each power grid data within a single sampling moment.

[0016] Furthermore, the input and output shapes of each single interleaved network are the same.

[0017] Furthermore, the length of the convolution kernel used in the convolutional neural network (CNN) remains the same as the length of the embedding matrix.

[0018] Furthermore, the feature extraction and maximum pooling processing specifically include: a final feature matrix generated by superposition of all single interleaved networks is subjected to maximum pooling processing to obtain a vector representation corresponding to each set of historical data and each set of predicted data.

[0019] Furthermore, the feature extraction and maximum pooling processing specifically include: the feature matrix generated by each layer of the single interleaved network is subjected to maximum pooling processing to obtain a vector; and then all the vectors are stacked to obtain a vector representation corresponding to each group of historical data and each group of predicted data.

[0020] Further, the trained multiple interleaved deep learning model is obtained by following the steps below:

[0021] The embedding matrices corresponding to all historical data and all predicted data, as well as labels with matching / mismatching between historical data and predicted data, where the label is 1 when matching and 0 when mismatching, are input into the multi-interleaved deep learning model, and the matching score between historical data and predicted data is calculated using the multi-interleaved deep learning model;

[0022] During the training phase, the following marginal loss function is used to guide the update of parameters in the multi-interleaved deep learning model, and the Adam optimization algorithm is used to train the multi-interleaved deep learning model and update parameters:

[0023]

[0024] Where N is (h, p + ,p - ) The number of triplets, h represents a set of historical data, p + represents a set of predicted data that matches the set of historical data, p - Represents a set of predicted data that does not match the set of historical data; M is a constant used to control the margin between the matching predicted data and the unmatched predicted data; during the training process, the marginal loss function makes the matching score between a historical data and its matching predicted data higher and higher, and the matching score between the unmatched predicted data lower and lower, until the matching score between the matching predicted data and the unmatched predicted data is higher than that between the matching predicted data and the unmatched predicted data by M.

[0025] Furthermore, the network parameters of each single interleaved network are independent of each other.

[0026] The present invention discloses a power grid data verification system based on multiple interwoven deep learning models, comprising:

[0027] The data preprocessing module is used to take the power grid data at the i sampling moments as i groups of historical data, take a group of power grid data at the i+1 sampling moment as a group of prediction data, and copy the group of prediction data i times as i groups of prediction data; preprocess each group of historical data and each group of prediction data to obtain their respective fixed-size embedding matrices;

[0028] The power grid data verification module is used to use the trained multiple interleaved deep learning model to perform feature extraction and maximum pooling processing on the embedding matrix corresponding to each group of historical data and each group of predicted data, and obtain the vector representation corresponding to each group of historical data and each group of predicted data; then use the similarity calculation formula to calculate the degree of matching between each group of historical data and its corresponding group of predicted data, and output the power grid data verification result according to the degree of matching.

[0029] Furthermore, the preprocessing of each set of historical data and each set of predicted data to obtain respective embedding matrices of fixed size includes:

[0030] First, clean each set of historical data and each set of forecast data;

[0031] Secondly, use the gensim tool to get a suitable embedding model;

[0032] Finally, the embedding model maps each set of historical data and each set of predicted data into an embedding matrix of fixed length through table lookup. For the power grid data that cannot find a corresponding relationship in the embedding model, it will be mapped into a zero vector of the same length.

[0033] Furthermore, the multiple interwoven deep learning model is composed of the superposition of multiple single interwoven networks; the input of each single interwoven network is the embedding matrix obtained in step 2, and the input embedding matrix is ​​passed through a bidirectional gated recurrent unit Bi-GRU to obtain a new embedding matrix of the same size as the original embedding matrix that pays attention to the correlation between each sampling moment within the embedding matrix; the new embedding matrix is ​​added to the input embedding matrix and sent to the convolutional neural network CNN to extract the correlation features between each power grid data within a single sampling moment.

[0034] Furthermore, the input and output shapes of each single interleaved network are the same, and the network parameters of each single interleaved network are independent of each other.

[0035] Furthermore, the length of the convolution kernel used in the convolutional neural network (CNN) remains the same as the length of the embedding matrix.

[0036] Furthermore, the feature extraction and maximum pooling processing specifically include: a final feature matrix generated by superposition of all single interleaved networks is subjected to maximum pooling processing to obtain a vector representation corresponding to each set of historical data and each set of predicted data.

[0037] Furthermore, the feature extraction and maximum pooling processing specifically include: the feature matrix generated by each layer of the single interleaved network is subjected to maximum pooling processing to obtain a vector; and then all the vectors are stacked to obtain a vector representation corresponding to each group of historical data and each group of predicted data.

[0038] Further, the trained multiple interleaved deep learning model is obtained by following the steps below:

[0039] The embedding matrices corresponding to all historical data and all predicted data, as well as labels with matching / mismatching between historical data and predicted data, where the label is 1 when matching and 0 when mismatching, are input into the multi-interleaved deep learning model, and the matching score between historical data and predicted data is calculated using the multi-interleaved deep learning model;

[0040] During the training phase, the following marginal loss function is used to guide the update of parameters in the multi-interleaved deep learning model, and the Adam optimization algorithm is used to train the multi-interleaved deep learning model and update parameters:

[0041]

[0042] Where N is (h, p + ,p- ) The number of triplets, h represents a set of historical data, p + represents a set of predicted data that matches the set of historical data, p - Represents a set of predicted data that does not match the set of historical data; M is a constant used to control the margin between the matching predicted data and the unmatched predicted data; during the training process, the marginal loss function makes the matching score between a historical data and its matching predicted data higher and higher, and the matching score between the unmatched predicted data lower and lower, until the matching score between the matching predicted data and the unmatched predicted data is higher than that between the matching predicted data and the unmatched predicted data by M.

[0043] The present invention discloses a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the processor implements the steps of a power grid data verification method based on a multiple interleaved deep learning model.

[0044] The present invention discloses a storage medium, which stores a power grid data verification program. When the power grid data verification program is executed by at least one processor, the steps of a power grid data verification method based on a multiple interleaved deep learning model are implemented.

[0045] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0046] (1) The present invention uses historical data at multiple different sampling moments to predict the data at the next sampling moment, and determines whether the power grid data is abnormal by calculating the matching score between the historical data and the predicted data, thereby achieving the purpose of data verification, and has the advantages of fast response speed and high recognition accuracy;

[0047] (2) The present invention uses CNN to focus on the intrinsic connection between different power grid data at a single sampling moment, and uses GRU to focus on the intrinsic connection of the same power grid data in time series at different sampling moments, and performs data verification from different angles and dimensions, which can improve the verification efficiency of the model;

[0048] (3) The present invention uses a specially designed CNN convolution kernel for feature extraction, which can ensure the integrity of a set of sampled data and improve the accuracy of the model; at the same time, the parameters of CNN and GRU are also specially designed so that the input and output shapes of the single interleaved network are the same and can be superimposed for use, thereby increasing the flexibility of the model;

[0049] (4) Compared with traditional data verification, the present invention is more efficient and accurate, and has a high degree of automation, which can reduce labor costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a schematic diagram of the overall framework of the deep learning model;

[0051] Figure 2 It is a single interlaced network structure diagram;

[0052] Figure 3 This is a diagram of a multiple interwoven network structure. DETAILED DESCRIPTION

[0053] The technical solution of the present invention is now further described in conjunction with the accompanying drawings and embodiments.

[0054] Embodiment 1:

[0055] This embodiment proposes a power grid data verification method based on a multiple interleaved deep learning model, which mainly includes the following steps:

[0056] Step 1: All power grid data at the same sampling time is called a group of power grid data. Multiple groups of power grid data at different sampling times are obtained to generate multiple groups of historical data and multiple groups of predicted data, and map them into vectors of fixed length to obtain the embedding matrix corresponding to the historical data and the predicted data. The power grid data in this step refers to the current, voltage, resistance, etc. in the power grid. The specific operations include:

[0057] All historical data and all predicted data are recorded as h and p respectively. Assuming that h and p are composed of all power grid data at 48 sampling moments, we have:

[0058] h=[h1,h2,…,h 48 ]

[0059] p=[p1,p2,…,P 48 ]

[0060] Among them, h i , i = 1, 2, ..., 48 represents all power grid data at 48 sampling moments, that is, 48 ​​sets of historical data, and p j =h 49 , j = 1, 2, ..., 48, that is, the predicted data is all the power grid data at the 49th sampling moment. In order to keep consistent with the historical data in form, h 49 48 copies were made. The forecast data is copied from the historical data at the next sampling moment, so that the historical data and the forecast data remain the same in shape, which is convenient for subsequent processing.

[0061] h and , are preprocessed by dimension increase to obtain a fixed-size embedding matrix for subsequent processing of the model. The main operations include: first, cleaning each set of historical data and each set of predicted data to ensure data integrity; second, using the gensim tool to train the data set to obtain an embedding model suitable for the data set; finally, the embedding model maps each set of power grid data to a vector of fixed length through table lookup. For power grid data that cannot find a corresponding relationship in the embedding model, it will be mapped to a zero vector of the same length. At this point, each set of historical data and predicted data will be mapped to an embedding matrix of fixed size.

[0062] Step 2: Use the multiple interleaved deep learning model to extract features from the embedding matrix corresponding to the historical data and the predicted data, and obtain the vector representation corresponding to the historical data and the predicted data after maximum pooling. Use the following cosine similarity calculation formula to calculate the matching degree between the historical data and the predicted data, so as to realize the data verification of the power grid.

[0063] The cosine similarity calculation formula is expressed as:

[0064]

[0065] Among them, a(x1, y1) and b(x1, y1) represent two vectors, and ||·|| represents the modulus of the vector.

[0066] The multiple interleaved deep learning model is now further explained.

[0067] Convolutional neural network (CNN) can extract features from a data set at a single sampling moment, while gated recurrent unit (GRU), as a variant of recurrent neural network (RNN), can focus on the correlation between data sets at different sampling moments. CNN and GRU complement each other and can be used together to better extract features from historical and predicted data sets. The multi-interleaved deep learning model uses CNN and GRU alternately multiple times, that is, a single interleaved network is mainly composed of GRU and CNN, and multiple single interleaved networks are superimposed and used, which is equivalent to using GRU and CNN alternately multiple times. It can extract features from two dimensions: a single sampling moment and different sampling moments.

[0068] The multi-interleaved deep learning model of this embodiment is a superposition of multiple single interleaved networks. Therefore, it is necessary to ensure that the input and output shapes of the single interleaved networks are the same, so that they can be superimposed to form a multi-interleaved network. By alternately using CNN and Bi-GRU, not only can the features within a single sampling moment power grid data group be extracted, but also the correlation between power grid data groups at different sampling moments can be focused on. Figure 2As shown in the figure, the input of the single interleaved network is an embedding matrix. First, the input embedding matrix is ​​passed through the bidirectional gated recurrent unit Bi-GRU to obtain a new embedding matrix of the same size as the original embedding matrix, which pays attention to the correlation between each sampling moment within the embedding matrix; the new embedding matrix is ​​added to the original embedding matrix and sent to CNN to extract the correlation features between each power grid data within a single sampling moment. In order not to cut off the feature association between all data within a single sampling moment, the convolution kernel size of CNN in the multi-interleaved deep learning model is not equal in length and width in the traditional sense. The length of the convolution kernel of CNN is the same as the length of the embedding vector, and the width is set to 1, that is, the features of the data group at a single sampling moment are extracted.

[0069] According to different ways of generating results, the multiple intertwined deep learning model of this embodiment has two different forms, namely, multiple intertwined deep learning model (Multiple Intertwined Networks), denoted as MIN; and multiple intertwined deep learning model-stack (Multiple Intertwined Networks-stack), denoted as MIN-s. Figure 3 As shown in the figure, both MIN and MIN-s are formed by superimposing single interleaved networks. The only difference is that the final vector representation of MIN is generated by the maximum pooling of the final feature matrix generated by the superposition of all single interleaved networks, while the final vector representation of MIN-s is formed by stacking the vectors obtained by the maximum pooling of the feature matrix generated by each layer of single interleaved networks.

[0070] After the multiple interleaved deep learning models extract the feature matrix, they are processed by maximum pooling to represent the historical data group and the predicted data group in the form of a one-dimensional vector.

[0071] During the training phase of the multiple interleaved deep learning models, the input of the multiple interleaved deep learning models is the embedding matrix corresponding to the historical data group and the predicted data group, as well as the labels with matching / non-matching between the two. The multiple interleaved deep learning models are used to calculate the matching scores between the historical data and the predicted data. During the training phase, the multiple interleaved deep learning models will automatically fit a threshold based on the final matching scores and labels. In actual work, if the matching score between the input historical data and the predicted data is higher than the threshold, the data is considered normal, otherwise the data is considered abnormal.

[0072] During the multi-interleaved deep learning model training phase, the following marginal loss function is used to guide the update of parameters in the model.

[0073]

[0074] Where N is (h, p + , p -) The number of triplets; M is a constant that controls the margin between the matching prediction data and the mismatching prediction data. By defining this loss function, as the training progresses, the matching score between a historical data and its matching prediction data becomes higher and higher, and the matching score between its mismatching prediction data becomes lower and lower, until the matching score between the matching prediction data and the mismatching prediction data is higher than that between the matching prediction data and the mismatching prediction data by M. We hope that as the model is trained, As big as possible, As small as possible, so that the loss will continue to decrease and the model accuracy will become higher and higher.

[0075] The Adam optimization algorithm is used to train the network and update parameters. Adam is a stochastic objective function optimization algorithm based on first-order gradient. Given an initial learning rate, the learning rate can be dynamically adjusted during training by comprehensively considering the first-order moment estimation and second-order moment estimation of the gradient.

[0076] In this embodiment, the data set is divided into three parts: a training data set (Training), a development data set (Development) and a test data set (Test), which are used to train the model, adjust the hyperparameters and verify the quality of the model respectively.

[0077] For each circuit, 50 (h, p + , p - ) triple, h represents a set of historical data, p + represents a set of prediction data that matches the set of historical data, and p represents a set of prediction data that does not match the set of historical data, which are randomly generated in the data pool. Therefore, the training data set has a total of 5,000,000 (h, p + , p - ) triplet.

[0078] The development data set and the test data set generate 100 (h, p, label) triplets for each set of historical data and predicted data, where h and p represent historical data and predicted data respectively, and label is the label. + , then label is 1 and p is p - , then label is 0.

[0079] The following top-1 accuracy (ACC@1) is used as the evaluation metric:

[0080]

[0081] Where N represents the number of historical data and predicted data combinations contained in the development set or test set. represents one of the normal prediction data corresponding to the i-th combination, represents the predicted data with the highest matching degree for the first k combinations. Include but is 1, otherwise is 0.

[0082] Specifically, in the embedding stage, this embodiment maps each set of data to a vector of length 300; the number of layers of the multi-interleaved network is set to 3, the number of output channels of CNN is set to 300, the number of hidden units of Bi-GRU is set to 150, the parameter M of the loss function is set to 0.1, the initial learning rate of the Adam optimization algorithm is set to 0.001, and batch_size is set to 100. It is worth mentioning that the network parameters of each single interleaved network are independent of each other, otherwise the parameters of each network cannot be optimized during the back propagation process, resulting in a situation where it is difficult to satisfy everyone, thereby affecting the overall performance of the network.

[0083] This embodiment creates historical data groups and forecast data groups for sampling data such as current and voltage at different times in the power grid. According to the intrinsic correlation between the data at the same sampling time and the temporal correlation between the data at different sampling times, a deep learning model is used to perform data verification, monitor abnormal data, and ensure the safety of power grid operation.

[0084] Embodiment 2:

[0085] Based on Example 1, this embodiment proposes a power grid data verification system based on multiple interwoven deep learning models, including:

[0086] The data preprocessing module is used to take the power grid data at the i sampling moments as i groups of historical data, take a group of power grid data at the i+1 sampling moment as a group of prediction data, and copy the group of prediction data i times as i groups of prediction data; preprocess each group of historical data and each group of prediction data to obtain their respective fixed-size embedding matrices;

[0087] The power grid data verification module is used to use the trained multiple interleaved deep learning model to perform feature extraction and maximum pooling processing on the embedding matrix corresponding to each group of historical data and each group of predicted data, and obtain the vector representation corresponding to each group of historical data and each group of predicted data; then use the similarity calculation formula to calculate the degree of matching between each group of historical data and its corresponding group of predicted data, and output the power grid data verification result according to the degree of matching.

[0088] Embodiment 3:

[0089] This embodiment discloses a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps disclosed in any of the above embodiments are implemented.

[0090] Embodiment 4:

[0091] This embodiment discloses a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps disclosed in any of the above embodiments are implemented.

[0092] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0093] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0094] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A power grid data verification method based on multiple interwoven deep learning models, Features: The following steps are involved: Step 1: All power grid data at the same sampling time is called a group of data. The power grid data at the i sampling time is taken as the i-group historical data, and the group of power grid data at the i+1 sampling time is taken as a group of predicted data. This group of predicted data is copied i times as the i-group predicted data; Step 2: Preprocess each set of historical data and each set of predicted data to obtain their respective fixed-size embedding matrices; Step 3: Use the trained multiple interleaved deep learning model to perform feature extraction and maximum pooling processing on the embedding matrix corresponding to each group of historical data and each group of predicted data to obtain the vector representation corresponding to each group of historical data and each group of predicted data; then use the similarity calculation formula to calculate the degree of matching between each group of historical data and its corresponding group of predicted data, and output the power grid data verification result according to the degree of matching.

2. According to the method for verifying power grid data based on multiple interwoven deep learning models according to claim 1, Features: The preprocessing of each set of historical data and each set of predicted data to obtain respective fixed-size embedding matrices includes: First, clean each set of historical data and each set of forecast data; Secondly, use the gensim tool to get a suitable embedding model; Finally, the embedding model maps each set of historical data and each set of predicted data into an embedding matrix of fixed length through table lookup. For the power grid data that cannot find a corresponding relationship in the embedding model, it will be mapped into a zero vector of the same length.

3. According to the method for verifying power grid data based on multiple interwoven deep learning models as described in claim 1, Features: The multi-interleaved deep learning model is composed of the superposition of multiple single interleaved networks; the input of each single interleaved network is the embedding matrix obtained in step 2, and the input embedding matrix is ​​passed through a bidirectional gated recurrent unit Bi-GRU to obtain a new embedding matrix of the same size as the original embedding matrix that pays attention to the correlation between each sampling moment within the embedding matrix; the new embedding matrix is ​​added to the input embedding matrix and sent to the convolutional neural network CNN to extract the correlation features between each power grid data within a single sampling moment.

4. According to claim 3, a power grid data verification method based on multiple interwoven deep learning models, Features: The input and output shapes of each single interleaved network are the same.

5. According to claim 3, a power grid data verification method based on multiple interwoven deep learning models, Features: The length of the convolution kernel used in the convolutional neural network CNN is kept the same as the length of the embedding matrix.

6. A power grid data verification method based on multiple interwoven deep learning models according to claim 3, Features: The feature extraction and maximum pooling processing specifically include: a final feature matrix generated by superposition of all single interleaved networks is subjected to maximum pooling processing to obtain a vector representation corresponding to each set of historical data and each set of predicted data.

7. The power grid data verification method based on multiple interwoven deep learning models according to claim 3, Features: The feature extraction and maximum pooling processing specifically include: the feature matrix generated by each layer of the single interleaved network is subjected to maximum pooling processing to obtain a vector; and then all the vectors are stacked to obtain a vector representation corresponding to each group of historical data and each group of predicted data.

8. The power grid data verification method based on multiple interwoven deep learning models according to claim 1, Features: Follow the steps below to get the trained multi-interleaved deep learning model: The embedding matrices corresponding to all historical data and all predicted data, as well as labels with matching / mismatching between historical data and predicted data, where the label is 1 when matching and 0 when mismatching, are input into the multi-interleaved deep learning model, and the matching score between historical data and predicted data is calculated using the multi-interleaved deep learning model; During the training phase, the following marginal loss function is used to guide the update of parameters in the multi-interleaved deep learning model, and the Adam optimization algorithm is used to train the multi-interleaved deep learning model and update parameters: Where N is (h, p + ,p - ) The number of triplets, h represents a set of historical data, p + represents a set of predicted data that matches the set of historical data, p - Represents a set of predicted data that does not match the set of historical data; M is a constant used to control the margin between the matching predicted data and the unmatched predicted data; during the training process, the marginal loss function makes the matching score between a historical data and its matching predicted data higher and higher, and the matching score between the unmatched predicted data lower and lower, until the matching score between the matching predicted data and the unmatched predicted data is higher than that between the matching predicted data and the unmatched predicted data by M.

9. The power grid data verification method based on multiple interwoven deep learning models according to claim 3, Features: The network parameters of each single interleaved network are independent of each other.

10. A power grid data verification system based on multiple interwoven deep learning models, Features: include: A data preprocessing module is used to take the power grid data at i sampling moments as i groups of historical data, take a group of power grid data at the i+1th sampling moment as a group of prediction data, and copy the group of prediction data i times as i groups of prediction data; Preprocess each set of historical data and each set of forecast data to obtain their respective fixed-size embedding matrices; The power grid data verification module is used to use the trained multiple interleaved deep learning model to perform feature extraction and maximum pooling processing on the embedding matrix corresponding to each group of historical data and each group of predicted data, and obtain the vector representation corresponding to each group of historical data and each group of predicted data; then use the similarity calculation formula to calculate the degree of matching between each group of historical data and its corresponding group of predicted data, and output the power grid data verification result according to the degree of matching.

11. A power grid data verification system based on multiple interwoven deep learning models according to claim 10, Features: The preprocessing of each set of historical data and each set of predicted data to obtain respective fixed-size embedding matrices includes: First, clean each set of historical data and each set of forecast data; Secondly, use the gensim tool to get a suitable embedding model; Finally, the embedding model maps each set of historical data and each set of predicted data into an embedding matrix of fixed length through table lookup. For the power grid data that cannot find a corresponding relationship in the embedding model, it will be mapped into a zero vector of the same length.

12. A power grid data verification system based on multiple interwoven deep learning models according to claim 10, Features: The multi-interleaved deep learning model is composed of the superposition of multiple single interleaved networks; the input of each single interleaved network is the embedding matrix obtained in step 2, and the input embedding matrix is ​​passed through a bidirectional gated recurrent unit Bi-GRU to obtain a new embedding matrix of the same size as the original embedding matrix that pays attention to the correlation between each sampling moment within the embedding matrix; the new embedding matrix is ​​added to the input embedding matrix and sent to the convolutional neural network CNN to extract the correlation features between each power grid data within a single sampling moment.

13. A power grid data verification system based on multiple interwoven deep learning models according to claim 12, Features: The input and output shapes of each single interleaved network are the same, and the network parameters of each single interleaved network are independent of each other.

14. A power grid data verification system based on multiple interwoven deep learning models according to claim 12, Features: The length of the convolution kernel used in the convolutional neural network CNN is kept the same as the length of the embedding matrix.

15. A power grid data verification system based on multiple interwoven deep learning models according to claim 12, Features: The feature extraction and maximum pooling processing specifically include: a final feature matrix generated by superposition of all single interleaved networks is subjected to maximum pooling processing to obtain a vector representation corresponding to each set of historical data and each set of predicted data.

16. A power grid data verification system based on multiple interwoven deep learning models according to claim 12, Features: The feature extraction and maximum pooling processing specifically include: the feature matrix generated by each layer of the single interleaved network is subjected to maximum pooling processing to obtain a vector; and then all the vectors are stacked to obtain a vector representation corresponding to each group of historical data and each group of predicted data.

17. A power grid data verification system based on multiple interwoven deep learning models according to claim 10, Features: Follow the steps below to get the trained multi-interleaved deep learning model: The embedding matrices corresponding to all historical data and all predicted data, as well as labels with matching / mismatching between historical data and predicted data, where the label is 1 when matching and 0 when mismatching, are input into the multi-interleaved deep learning model, and the matching score between historical data and predicted data is calculated using the multi-interleaved deep learning model; During the training phase, the following marginal loss function is used to guide the update of parameters in the multi-interleaved deep learning model, and the Adam optimization algorithm is used to train the multi-interleaved deep learning model and update parameters: Where N is (h, p + ,p - ) The number of triplets, h represents a set of historical data, p + represents a set of predicted data that matches the set of historical data, p - Represents a set of predicted data that does not match the set of historical data; M is a constant used to control the margin between the matching predicted data and the unmatched predicted data; during the training process, the marginal loss function makes the matching score between a historical data and its matching predicted data higher and higher, and the matching score between the unmatched predicted data lower and lower, until the matching score between the matching predicted data and the unmatched predicted data is higher than that between the matching predicted data and the unmatched predicted data by M.

18. A computer device, It is characterized in that It includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of a power grid data verification method based on multiple interwoven deep learning models as described in any one of claims 1 to 9 are implemented.

19. A storage medium, It is characterized in that The storage medium stores a power grid data verification program, which, when executed by at least one processor, implements the steps of a power grid data verification method based on multiple interwoven deep learning models as described in any one of claims 1 to 9.