A Deep Learning-Based Method and System for Verifying Multi-Source Data from Power Grid Measurement Centers

Abstract

This invention discloses a deep learning-based method and system for verifying multi-source data in a power grid measurement center. The method includes: collecting remote signaling data and telemetry data from the circuit and constructing a dataset; preprocessing the data in the dataset using a preprocessing module to convert the remote signaling data and telemetry data into a first remote signaling embedding matrix and a first telemetry embedding matrix, respectively; obtaining the data correlation between the first remote signaling embedding matrix and the first telemetry embedding matrix through an attention interaction module, and converting them into a second remote signaling embedding matrix and a second telemetry embedding matrix, respectively; training a convolutional neural network based on the dataset to obtain a verification model; collecting real-time telemetry data and remote signaling data from the circuit and processing them through the preprocessing module and the attention interaction module, and verifying whether the data is abnormal using the verification model. Compared with traditional data verification, this invention is more efficient and accurate, and has a high degree of automation, which can reduce labor costs.

Description

Technical Field

[0001] This invention relates to the field of power grid measurement center data verification technology in power systems, and more specifically, to a deep learning-based method and system for verifying multi-source data from power grid measurement centers. Background Technology

[0002] Deep learning (DL) is a high-level abstract algorithm that uses multiple complex neural network structures to represent multi-layered nonlinear transformations. In recent years, it has demonstrated powerful capabilities in various fields of machine learning, including image classification, machine translation, face recognition, and object detection. The development of deep learning relies heavily on powerful computing power and massive amounts of data. Therefore, since the advent of the first CNN architecture—LeNet-5—in 1998, deep learning experienced a decade-long stagnation until AlexNet won the ImageNet challenge in 2012. After that, various deep learning-based neural networks, such as VGG Net, GoogLeNet, and ResNet, shone brightly in the field of computer vision (CV). Networks continue to evolve towards greater depths, with increasingly superior performance. Deep learning has also propelled the development of natural language processing (NLP). The Transformer emerged in 2017, achieving unprecedented success in various areas of NLP. Later models derived from Transformer transformations, such as BERT, ELMo, and GPT, have all demonstrated excellent performance.

[0003] A Convolutional Neural Network (CNN) is a special type of neural network primarily used in image recognition, speech recognition, and natural language processing. The core of a CNN is the convolutional layer, which extracts features from the input data through convolutional operations. Each neuron in a convolutional layer is connected to only a subset of neurons in the previous layer, enabling CNNs to handle large-scale datasets. CNNs also include pooling layers and fully connected layers. Pooling layers reduce network complexity and the risk of overfitting by decreasing the size and number of feature maps. Fully connected layers multiply the output of the pooling layers by the weights and add a bias term to obtain the final output. The training process for CNNs typically uses backpropagation and gradient descent algorithms to adjust the network parameters, enabling the network to accurately classify or predict input data.

[0004] In power systems, data verification is a crucial means of ensuring the accuracy and reliability of grid data. Existing grid data verification methods mainly include rule-based verification, logical verification, and statistical verification. However, these existing verification methods all have certain limitations. For example, rule-based verification can only verify data that conforms to preset rules and cannot verify abnormal data; logical verification requires manual writing of verification logic, which is inefficient; and statistical verification depends on the accuracy of the statistical model and the quality of the data, requiring high data quality. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a deep learning-based method for verifying multi-source data from power grid measurement centers, which solves the technical problems of low accuracy and low processing efficiency in existing verification methods.

[0006] The present invention adopts the following technical solution.

[0007] A deep learning-based method for verifying multi-source data from power grid measurement centers includes the following steps:

[0008] Step 1: Collect remote signaling data and telemetry data from the circuit and construct a dataset. Preprocess the data in the dataset through the preprocessing module, and convert the remote signaling data and telemetry data into the first remote signaling embedding matrix and the first telemetry embedding matrix, respectively.

[0009] Step 2: Obtain the data association between the first remote signaling embedding matrix and the first telemetry embedding matrix through the attention interaction module, and convert them into the second remote signaling embedding matrix and the second telemetry embedding matrix respectively.

[0010] Step 3: Train the convolutional neural network based on the dataset, and use the trained convolutional neural network as the verification model;

[0011] Step 4: Collect real-time telemetry data and remote signaling data from the circuit, process them through the preprocessing module and the attention interaction module, and verify the data for anomalies through the verification model.

[0012] Preferably, in step 1, remote signaling data is collected. Remote signaling data refers to the switch status in each circuit. Suppose there are M different switches in a certain circuit. The status of each switch is independent of each other. The status of each switch is collected as open or closed, and the status of each switch is obtained, which constitutes the remote signaling data of the circuit. A total of M remote signaling data are collected.

[0013] Telemetry data is collected, and the telemetry data of the power equipment in each circuit is collected to form the circuit's telemetry data. Each collected telemetry data is a three-dimensional data group, and each three-dimensional data group includes the power equipment ID, telemetry data type, and data value. The types of telemetry data include: active power, reactive power, current, and voltage of the power equipment. A total of N telemetry data are collected.

[0014] Preferably, preprocessing the data in the dataset using the preprocessing module further includes:

[0015] The first remote signaling embedding matrix and the first telemetry embedding matrix are preset to be l. tp ×l vec The matrix is ​​used to fill or remove data based on the number of remote signaling and telemetry data collected by each circuit, so that the number of remote signaling and telemetry data both reach l. tp indivual;

[0016] Each teleindication data point and each telemetry data point, after being filled or removed, is mapped to a fixed-length array. vec ;

[0017] The obtained l tp Each remote signaling data vector and l tp The telemetry data vectors are concatenated to obtain the first telemetry embedding matrix s and the first telemetry embedding matrix d, and both the first telemetry embedding matrix s and the first telemetry embedding matrix d are l tp ×l vec The matrix.

[0018] Preferably, in the preprocessing, if the number of remote signaling data or telemetry data is less than 1... tp If there are 1, fill in the special characters. <pad>Make it complete to l tp If the remote signaling data or telemetry data exceeds 1 tp If there are 1, then remove the excess data so that the number of remote signaling data and remote measurement data is 1. tp indivual.

[0019] Preferably, step 2 further includes:

[0020] Step 2-1: Use cosine similarity to measure the correlation between remote signaling data and telemetry data to obtain the similarity matrix M of the effective parts of the remote signaling data and telemetry data. s ;

[0021] Step 2-2, for the similarity matrix M of the effective part s Softmax processing is performed separately in the column and row directions to obtain the column correlation matrix M. ss The correlation matrix M of the row direction sd ;

[0022] Steps 2-3 involve converting the column direction correlation matrix M... ss The correlation matrix M of the row direction sd Fill the space to obtain a size of l tp ×l tp First correlation matrix matrix1 and second correlation matrix matrix2;

[0023] Steps 2-4: Multiply the first remote signaling embedding matrix s and the first telemetry embedding matrix d with the first correlation matrix matrix1 and the second correlation matrix matrix2 respectively to obtain the second remote signaling embedding matrix S and the second telemetry embedding matrix D.

[0024] Preferably, step 2-1 further includes:

[0025] Extracting the effective parts of the first remote signaling embedding matrix s and the first telemetry embedding matrix d, we obtain the effective part matrix s of the first remote signaling embedding matrix. v And the effective part matrix d in the first telemetry embedding matrix v The effective part is the matrix part formed by converting the collected remote signaling data and telemetry data into vectors in the first remote signaling embedding matrix s and the first telemetry embedding matrix d, rather than the vector content converted by the special symbols used for filling.

[0026] The cosine similarity is used to measure the correlation between remote signaling data vectors and telemetry data vectors, thus obtaining the effective part matrix s in the first remote signaling embedding matrix. v And the effective part matrix d in the first telemetry embedding matrix v The similarity matrix M between them s And M s satisfy:

[0027]

[0028] Among them, s v d represents the effective portion of the first remote signaling embedding matrix. v The effective portion of the first telemetry embedding matrix is ​​represented by the similarity matrix M. s The element in the i-th row and j-th column represents the remote signaling data s i With telemetry data d j degree of correlation, s i Let d represent the i-th remote signaling data. j This represents the similarity matrix obtained from the j-th telemetry data.

[0029] Preferably, in step 2-2, the column direction correlation matrix M ss The correlation matrix M of the row direction sd satisfy:

[0030] M ss =softmax line (M s )

[0031] M sd =softmax row (M s )

[0032] Among them, softmax line () indicates that softmax processing is applied to the column directions of the matrix. row () indicates that softmax processing is applied to the row direction of the matrix, resulting in the column direction correlation matrix. Row direction correlation matrix

[0033] Preferably, steps 2-3 further include: converting the column direction correlation matrix M ss The correlation matrix M of the row direction sd Fill all with 0 to 1 tp ×l tp From the matrix, we obtain two complete first correlation matrices matrix1 and second correlation matrices matrix2, which respectively satisfy...

[0034] Preferably, in steps 2-4, the second remote signaling embedding matrix S and the second telemetry embedding matrix D satisfy:

[0035] D = matrix1 T ·s

[0036] S = matrix2·d

[0037] Where s and d represent the first remote signaling embedding matrix and the first telemetry embedding matrix, respectively, and the resulting second remote signaling embedding matrix and second telemetry embedding matrix are...

[0038] The present invention also provides a deep learning-based multi-source data verification system for power grid measurement centers, comprising: a data acquisition module, a preprocessing module, an attention interaction module, and a neural network module;

[0039] The data acquisition module is used to collect power grid switch data and telemetry data;

[0040] The preprocessing module preprocesses the collected data to obtain the original embedding matrix;

[0041] The attention interaction module further processes the original embedding matrix to obtain a new embedding matrix;

[0042] The neural network module constructs and trains a convolutional neural network to obtain a verification model. The verification model is then used to verify the real-time data of the circuit to determine whether the data is abnormal.

[0043] The present invention also provides a terminal, including a processor and a storage medium;

[0044] The storage medium is used to store instructions;

[0045] The processor is used to operate according to the instructions to execute the steps of the verification method.

[0046] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the verification method.

[0047] The beneficial effect of this invention lies in its ability, compared with existing technologies, to automatically identify telemetry data with abnormal matching with remote signaling data such as switches, thereby achieving the purpose of data verification. A deep learning model is proposed, comprising a preprocessing module, an attention interaction module, and a neural network module. In actual operation, the switch combination patterns in the circuit and power grid data are input into the model. Through steps such as embedding, attention interaction, feature extraction, and matching degree calculation, and using the matching degree threshold obtained during model training, the model outputs whether the telemetry data is abnormal. Deep learning-based power grid data verification can automatically learn the potential connections between remote signaling data and telemetry data in the circuit, making it more accurate and reliable than traditional data verification, while also saving manpower and improving efficiency.

[0048] The present invention has at least the following beneficial effects:

[0049] 1. This invention aims at verifying the consistency of multi-source electrical data in the power grid. Using deep learning as a method, it can obtain more accurate data verification results by learning from historical data. Compared with traditional data verification methods that require manual analysis and comparison of data, it can improve the accuracy and efficiency of data verification.

[0050] 2. This invention can process large amounts of data in a short time, which greatly improves the speed of data verification and reduces the possible errors of manual operation. It has the advantages of high efficiency and high degree of automation. In addition, this invention can be adjusted and optimized according to actual needs, and can also adapt to different data types and data volumes, and has strong scalability.

[0051] 3. This invention can reduce labor costs. Traditional data verification methods require manual analysis and comparison of data, which consumes a lot of manpower and time. However, power grid data verification based on deep learning can reduce the dependence on manual labor and save manpower and time costs. It uses convolutional neural networks to extract features from the embedding matrices of switches and data, and uses max pooling to find the largest element of each feature map and concatenate them. This transforms the feature matrix obtained by CNN into a vector, which facilitates the calculation of the matching score between switches and data. Attached Figure Description

[0052] Figure 1 This is a flowchart illustrating the multi-source data verification method in this invention;

[0053] Figure 2 This is a schematic diagram of the attention interaction module in this invention;

[0054] Figure 3 This is an overall framework diagram of the multi-source data verification system in this invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, other embodiments obtained by those skilled in the art without creative effort are all within the protection scope of this invention.

[0056] like Figure 1 As shown, this invention proposes a deep learning-based multi-source data verification method for power grid measurement centers. It utilizes a deep learning model to perform rationality and consistency verification on telemetry data (active power, reactive power, current, voltage, etc.) and teleindication data (switch positions, etc.) collected from multiple systems, identifying model matching errors and data anomalies. By combining teleindication data (switch positions, etc.) with telemetry data (current, voltage, etc.) labeled as normal / abnormal, a matching score is calculated between the telemetry data and the teleindication data. This matching score is then used to determine whether the telemetry data is abnormal. The method specifically includes the following steps:

[0057] Step 1: Collect remote signaling data and telemetry data from the circuit and construct a dataset. Preprocess the data in the dataset through the preprocessing module, and convert the remote signaling data and telemetry data into the first remote signaling embedding matrix and the first telemetry embedding matrix, respectively.

[0058] Collect remote signaling data, which refers to the switch status in each circuit. Suppose there are M different switches in a certain circuit, and the status of each switch is independent of each other. Collect the status of each switch as open or closed to obtain the status of each switch, which constitutes the remote signaling data of the circuit. A total of M remote signaling data are collected.

[0059] Telemetry data is collected, and the telemetry data of the power equipment in each circuit is collected to form the circuit's telemetry data. Each collected telemetry data is a three-dimensional data group, and each three-dimensional data group includes the power equipment ID, telemetry data type, and data value. The types of telemetry data include: active power, reactive power, current, and voltage of the power equipment. A total of N telemetry data are collected.

[0060] A dataset is constructed based on the collected remote signaling and telemetry data of each circuit. The dataset consists of three parts: a training set, a development set, and a test set, which are used in subsequent steps to train the model, adjust parameters, and verify the model's performance.

[0061] The training set created 50 (s', d) pairs for each circuit. + d - The triplet, s' represents a set of remote signaling data corresponding to a set of switch states in the circuit, d + This represents the normal telemetry data matching this group of switches, d - Abnormal telemetry data representing mismatches with this group of switches is randomly generated from the data pool;

[0062] The development and test sets generate 100 (s', d) values ​​for each set of switch states. * (label) triples, s' and d * These represent remote signaling data and telemetry data, respectively, with label being the tag. If d * For d + Then the label is 1, d * For d - Then the label is 0.

[0063] Furthermore, such as Figure 2 As shown, the remote signaling data and telemetry data in the dataset are preprocessed by the preprocessing module to obtain the first remote signaling embedding matrix s and the first telemetry embedding matrix d. The preprocessing specifically includes:

[0064] The first remote signaling embedding matrix and the first telemetry embedding matrix are preset to be l. tp ×l vec The matrix is ​​used to fill or remove data based on the number of remote signaling and telemetry data collected by each circuit, so that the number of remote signaling and telemetry data both reach l. tp indivual;

[0065] Specifically, if the number of remote signaling data or telemetry data is less than 1 tp If there are 1, fill in the special characters. <pad>Make it complete to l tp In subsequent processing, for <pad>Specialized processing, disregarding its correlation with normal switching or data, facilitates batch data processing in the model without affecting its accuracy. If the remote signaling or telemetry data exceeds l tp If there are 1, then remove the excess data so that the number of remote signaling data and remote measurement data is 1. tp indivual;

[0066] Each teleindication data point and each telemetry data point, after being filled or removed, is mapped to a fixed-length array. vec The vector; it can be understood that for the collected remote signaling data, telemetry data, and special symbols filled in, vector mapping is required, and the mapped value is obtained as l. tp Each remote signaling data vector and l tp One telemetry data vector;

[0067] Preferably, the vector mapping rule is as follows: all remote signaling data and telemetry data are integrated into a corpus, and the gensim tool is used to train the corpus to obtain a word2vec model suitable for the dataset; the word2vec model maps each data in the remote signaling data and telemetry data to a fixed-length vector by looking up a table, and those data for which no corresponding relationship can be found in the word2vec model will be mapped to a zero vector of the same length;

[0068] The obtained l tp Each remote signaling data vector and l tp The telemetry data vectors are concatenated to obtain the first telemetry embedding matrix s and the first telemetry embedding matrix d, and both the first telemetry embedding matrix s and the first telemetry embedding matrix d are l tp ×l vec The matrix.

[0069] For example, in the preprocessing stage, each switch and data is padded or truncated to 200 dimensions, i.e., l tp =200, if less than 200, fill in the leading special character. <pad>Elements exceeding 200 are truncated and discarded. Each element is mapped to a vector of length 300, i.e., l vec =300.

[0070] Step 2: Obtain the data association between the first remote signaling embedding matrix and the first telemetry embedding matrix through the attention interaction module, and transform the first remote signaling embedding matrix and the first telemetry embedding matrix into the second remote signaling embedding matrix and the second telemetry embedding matrix, respectively.

[0071] The attention interaction module can connect two unrelated branches, the switch and the data, and fully consider the inherent relationship between the switch and the data. The resulting second remote signaling embedding matrix and second telemetry embedding matrix contain the correlation information between the remote signaling data and the telemetry data, which helps the subsequent neural network to extract more comprehensive features.

[0072] Specifically, the attention interaction module processes the intrinsic relationship between the preprocessed first remote signaling embedding matrix and the first telemetry embedding matrix to obtain a new second remote signaling embedding matrix and a second telemetry embedding matrix, which facilitates feature extraction by the subsequent neural network module.

[0073] Furthermore, step 2 also includes:

[0074] Step 2-1: Use cosine similarity to measure the correlation between remote signaling data and telemetry data to obtain the similarity matrix M of the effective parts of the remote signaling data and telemetry data. s ;

[0075] Specifically, the correlation obtained is the correlation between the effective data in the first telemetry embedding matrix s and the first telemetry embedding matrix d. Step 2-1 also includes:

[0076] Extracting the effective parts of the first remote signaling embedding matrix s and the first telemetry embedding matrix d, we obtain the effective part matrix s of the first remote signaling embedding matrix. v And the effective part matrix d in the first telemetry embedding matrix v The effective part is the matrix portion formed by converting the collected remote signaling data and telemetry data into vectors in the first remote signaling embedding matrix s and the first telemetry embedding matrix d, rather than the vector content converted by the special symbols used for filling.

[0077] The cosine similarity is used to measure the correlation between remote signaling data vectors and telemetry data vectors, thus obtaining the effective part matrix s in the first remote signaling embedding matrix. v And the effective part matrix d in the first telemetry embedding matrix v The similarity matrix M between them s And M s satisfy:

[0078]

[0079] Among them, s v d represents the effective portion of the first remote signaling embedding matrix. v The effective portion of the first telemetry embedding matrix is ​​represented by the similarity matrix M. s The element in the i-th row and j-th column represents the remote signaling data s i With telemetry data d j degree of correlation, s i Let d represent the i-th remote signaling data. j This represents the j-th telemetry data;

[0080] It is understandable that in step 1, M and N telemetry and teleindication data were collected respectively. In this invention, M and N do not exceed 1. tp Under the premise of , the effective part matrix s of the first remote signaling embedding matrix is ​​obtained. v And the effective part matrix d in the first telemetry embedding matrix v Each satisfies And similarity matrix

[0081] If M or N is greater than l tp Then the effective part matrix s in the first remote signaling embedding matrix v Similar to the first telemetry embedding matrix s, the effective part matrix d in the first telemetry embedding matrix... v The similarity matrix is ​​the same as the first telemetry embedding matrix d. At this point, the two column direction correlation matrices M obtained in subsequent step 2-2 are... ss The correlation matrix M of the row direction sd Size is also l tp ×l tp .

[0082] Step 2-2, for the similarity matrix M of the effective part s Softmax processing is performed separately in the column and row directions to obtain the column correlation matrix M. ss The correlation matrix M of the row direction sd ;

[0083] Column Directional Correlation Matrix M ss The correlation matrix M of the row direction sd satisfy:

[0084] M ss =softmax line (M s )

[0085] M sd =softmax row (M s )

[0086] Among them, softmax line () indicates that softmax processing is applied to the column directions of the matrix. row () indicates that softmax processing is applied to the row direction of the matrix, resulting in the column direction correlation matrix. Row direction correlation matrix

[0087] Steps 2-3 involve converting the column direction correlation matrix M... ss The correlation matrix M of the row direction sd Fill the space to obtain a size of l tp ×l tp First correlation matrix matrix1 and second correlation matrix matrix2;

[0088] Specifically, in step 1, filling the remote signaling data s and telemetry data d into a fixed-length vector is merely to facilitate batch data processing by the network; the special symbols used for filling are... <pad>It has no practical meaning, so it can be regarded as a special symbol. <pad>The correlation between the effective components and the column direction correlation matrix M is 0, meaning only the correlation between the effective components is considered. Therefore, in this step, the column direction correlation matrix M is... ss The correlation matrix M of the row direction sd Fill all 0 to l tp ×l tp From the matrix, we obtain two complete first correlation matrices matrix1 and second correlation matrices matrix2, which respectively satisfy...

[0089] Steps 2-4: Multiply the first remote signaling embedding matrix s and the first telemetry embedding matrix d with the first correlation matrix matrix1 and the second correlation matrix matrix2 respectively to obtain the second remote signaling embedding matrix S and the second telemetry embedding matrix D.

[0090] Specifically, the second remote signaling embedding matrix S and the second telemetry embedding matrix D satisfy:

[0091] D = matrix1 T .s

[0092] S = matrix2·d

[0093] Where s and d represent the first remote signaling embedding matrix and the first telemetry embedding matrix, respectively.

[0094] Since the size of the first correlation matrix matrix1 and the second correlation matrix matrix2 is l tp ×l tp The size of the first remote signaling embedding matrix s and the first telemetry embedding matrix d is l tp ×l vec Therefore, the second remote signaling embedding matrix and the second telemetry embedding matrix are obtained.

[0095] Step 3: Train the convolutional neural network and use the trained convolutional neural network as a verification model;

[0096] The role of convolutional neural networks is to extract features from switches and data and obtain their corresponding vector representations for the final matching score calculation. Specifically, convolutional neural networks are used to extract features from the embedding matrices of switches and data, and max pooling is used to find the largest element of each feature map and concatenate them. This transforms the feature matrix obtained by the CNN into a vector, which facilitates the calculation of the matching score between switches and data.

[0097] Let the convolution kernel of a convolutional neural network be... Where e is the width of the convolution kernel W, l vec The length of the convolution kernel W is the same as the length of the vector transformed in step 1, both being l. vec ;

[0098] Then, the second telemetry embedding matrix S and the second telemetry embedding matrix D after passing through the convolution kernel are respectively transformed into:

[0099]

[0100]

[0101] Among them, O S O D The feature maps of the second telemetry embedding matrix S and the second telemetry embedding matrix D after being transformed by convolution kernels, S i D represents the i-th row vector data of the second remote signaling embedding matrix S. j Let represent the j-th row vector data of the second telemetry embedding matrix D, and Represents the sum of positional products, where X∈{S, D}, that is, X represents the second teleindication embedding matrix S or the second telemetry embedding matrix D, W pq X represents the element in the p-th row and q-th column of the convolution kernel W. pq This represents the element in the p-th row and q-th column of X; W and b represent the convolution kernel and bias parameters, respectively; f(·) represents the activation function, such as the ReLU function, tanh function, etc.

[0102] Furthermore, assume that the CNN has a total of h c With convolution kernels, the second remote signaling embedding matrix S and the second telemetry embedding matrix D after passing through the CNN are respectively transformed into:

[0103]

[0104]

[0105] O here S O D The second remote signaling embedding matrix S and the second remote sensing embedding matrix D are respectively processed by h c Feature maps transformed by each convolutional kernel This represents the feature map of the second remote signaling embedding matrix S after transformation by the k-th convolutional kernel. This represents the feature map after the second telemetry embedding matrix D is transformed by the k-th convolution kernel.

[0106] After h c The result after the convolution kernel transformation The above conversion formula can be simplified as follows:

[0107] 0 S =CNN S (S)

[0108] 0 D =CNN D (D)

[0109] After feature extraction by CNN, the feature map O S and 0 D Max pooling is performed to obtain vector representations of the second teleindication embedding matrix S and the second telemetry embedding matrix D:

[0110]

[0111]

[0112] Where max_pooling() represents max pooling. Let S and D be the vector representations of the second teleindication embedding matrix and the second telemetry embedding matrix, respectively.

[0113] Furthermore, cosine similarity is used to measure the matching score between remote information data and telemetry data:

[0114]

[0115] Where ||·|| represents the magnitude of the vector, and Sim(s,d) represents the matching score between remote information data and telemetry data.

[0116] During the model training phase, the model output is Sim(s, d). Since the input data for training contains labels indicating whether the telemetry data is normal under this switch combination, the model will automatically train a threshold based on the matching score and training labels.

[0117] Top-1 accuracy (ACC@1) is used as an evaluation metric for training convolutional neural networks to assess the model's accuracy. This metric calculates the proportion of correctly identified data points out of the total number of data points. The model updates its parameters based on this metric. The formula for calculating top-k accuracy (ACC@k) is as follows:

[0118]

[0119] Here, 100 indicates that the development or test set contains 100 combinations of switch states. It's understandable that the number of combinations can be set according to actual needs. This represents one set of normal telemetry data corresponding to the i-th switch combination. This represents the top k telemetry data points with the highest matching degree corresponding to the i-th switch combination. If Include but =1, otherwise It is 0.

[0120] During the training phase, the model will automatically fit a threshold based on the final matching score and label. When the model is actually working, if the matching score between the input switch and the data is higher than the threshold, the data is considered normal; otherwise, the data is considered abnormal.

[0121] In actual operation, if the matching score between the switch and the data is greater than the threshold, the model outputs 1, indicating that the data is normal; otherwise, the model outputs 0, indicating that the data is abnormal.

[0122] Preferably, during the model training phase, a marginal loss function is used to guide the updating of parameters in the model. The formula for the loss function is as follows:

[0123]

[0124] Where P is a constant that controls the distance between normal and abnormal data; d + d represents the normal data under this switch, and d- represents the abnormal data under this switch. By defining this loss function, as training progresses, the matching score between a certain switch combination and its normal data can be made higher and higher, while the matching score between it and its abnormal data can be made lower and lower, until the matching score between normal data and abnormal data is Q higher than that between normal data and abnormal data.

[0125] As the model is trained, As large as possible Keep it as small as possible, so that the loss will continue to decrease and the model accuracy will become higher and higher.

[0126] The Adam optimization algorithm is used to train the network and update its parameters. Adam is a stochastic objective function optimization algorithm based on first-order gradients. Given an initial learning rate, it can dynamically adjust the learning rate during training by comprehensively considering the first-order and second-order moment estimates of the gradient.

[0127] The number of output channels of the CNN is set to 500, the parameter Q in the loss function can be set to 0.1, the initial learning rate of the Adam optimization algorithm is set to 0.001, and the batch size of the training dataset is set to 100.

[0128] The training set is used to train the parameters in the model, such as the content of the CNN convolution kernel; the development set is used to tune hyperparameters, such as the size of the CNN convolution kernel and the number of CNN output feature maps; and the test set is used to evaluate the quality of the model.

[0129] Step 4: Collect real-time telemetry data and teleindication data of the circuit, process them, and verify whether the data is abnormal through the verification model.

[0130] After the collected real-time telemetry and remote signaling data of the circuit are processed by the preprocessing module and the attention interaction module, the second remote signaling embedding matrix and the second telemetry embedding matrix corresponding to the real-time telemetry and remote signaling data are obtained. These are then input into the verification model to obtain the verification result of whether the real-time telemetry and remote signaling data of the circuit are abnormal.

[0131] like Figure 3 As shown, this invention also proposes a multi-source data verification system for power grid measurement centers based on deep learning. The above verification method can be implemented based on this system, which specifically includes: a data acquisition module, a preprocessing module, an attention interaction module, and a neural network module.

[0132] The data acquisition module is used to collect power grid switching data and telemetry data.

[0133] The preprocessing module preprocesses the collected data to obtain the original embedding matrix;

[0134] The attention interaction module further processes the original embedding matrix to obtain a new embedding matrix;

[0135] The neural network module constructs and trains a convolutional neural network to obtain a verification model. The verification model is then used to verify the real-time data of the circuit to determine whether the data is abnormal.

[0136] The beneficial effects of this invention are that, compared with the prior art, this invention can automatically learn the potential relationship between remote signaling data and telemetry data in the circuit, which is more accurate and reliable than traditional data verification, while saving manpower and improving efficiency.

[0137] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0138] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination thereof. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0139] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0140] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as "C" or similar languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0141] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.< / pad> < / pad> < / pad> < / pad> < / pad> < / pad>

Claims

1. A deep learning-based method for verifying multi-source data from a power grid measurement center, characterized in that, Includes the following steps: Step 1: Collect remote signaling data and telemetry data from the circuit and construct a dataset. Preprocess the data in the dataset through the preprocessing module, and convert the remote signaling data and telemetry data into the first remote signaling embedding matrix and the first telemetry embedding matrix, respectively. Step 2: Obtain the data association between the first remote signaling embedding matrix and the first telemetry embedding matrix through the attention interaction module, and convert them into the second remote signaling embedding matrix and the second telemetry embedding matrix respectively. Step 2 also includes: Step 2-1: Use cosine similarity to measure the correlation between remote signaling data and telemetry data to obtain the similarity matrix M of the effective parts of the remote signaling data and telemetry data. s ; Step 2-2, for the similarity matrix M of the effective part s Softmax processing is performed separately in the column and row directions to obtain the column correlation matrix M. ss The correlation matrix M of the row direction sd ; Steps 2-3 involve converting the column direction correlation matrix M... ss The correlation matrix M of the row direction sd Fill the space to obtain a size of l tp ×l tp The first correlation matrix matrix1 and the second correlation matrix matrix2; Steps 2-4: Multiply the first remote signaling embedding matrix s and the first telemetry embedding matrix d with the first correlation matrix matrix1 and the second correlation matrix matrix2 respectively to obtain the second remote signaling embedding matrix S and the second telemetry embedding matrix D. In step 2-2, the column direction correlation matrix M ss The correlation matrix M of the row direction sd satisfy: M ss =softmax line (M s ) M sd =softmax row (M s ) Among them, softmax line () indicates that softmax processing is applied to the column directions of the matrix. row () indicates that softmax processing is applied to the row direction of the matrix to obtain the column direction correlation matrix. Row direction correlation matrix M represents the number of remote signaling data collected, and N represents the number of telemetry data collected. Steps 2-3 further include: [The text abruptly ends here, likely due to an incomplete sentence or a formatting error.] ss The correlation matrix M of the row direction sd Fill all with 0 to l tp ×l tp From the matrix, we obtain two complete first correlation matrices matrix1 and second correlation matrices matrix2, which respectively satisfy... In steps 2-4, the second remote signaling embedding matrix S and the second telemetry embedding matrix D satisfy the following: D=matrix1 T ·s S = matrix2·d Where s and d represent the first remote signaling embedding matrix and the first telemetry embedding matrix, respectively, and the resulting second remote signaling embedding matrix and second telemetry embedding matrix S, Step 3: Train the convolutional neural network based on the dataset, and use the trained convolutional neural network as the verification model; Step 4: Collect real-time telemetry data and remote signaling data from the circuit, process them through the preprocessing module and the attention interaction module, and verify the data for anomalies through the verification model.

2. The method for verifying multi-source data of power grid measurement centers based on deep learning according to claim 1, characterized in that, In step 1, remote signaling data is collected. Remote signaling data refers to the switch status in each circuit. Suppose there are M different switches in a certain circuit. The status of each switch is independent of each other. The status of each switch is collected as open or closed. The status of each switch is obtained, which constitutes the remote signaling data of the circuit. A total of M remote signaling data are collected. Telemetry data is collected, and the telemetry data of the power equipment in each circuit is collected to form the circuit's telemetry data. Each collected telemetry data is a three-dimensional data group, and each three-dimensional data group includes the power equipment ID, telemetry data type, and data value. The types of telemetry data include: active power, reactive power, current, and voltage of the power equipment. A total of N telemetry data are collected.

3. The deep learning-based multi-source data verification method for power grid measurement centers according to claim 2, characterized in that, The preprocessing of the data in the dataset by the preprocessing module also includes: The first remote signaling embedding matrix and the first telemetry embedding matrix are preset to be l. tp ×l vec The matrix is ​​used to fill or remove data based on the number of remote signaling and telemetry data collected by each circuit, so that the number of remote signaling and telemetry data both reach l. tp indivual; Each teleindication data point and each telemetry data point, after being filled or removed, is mapped to a fixed-length array. vec ; The obtained l tp Each remote signaling data vector and l tp The telemetry data vectors are concatenated to obtain the first telemetry embedding matrix s and the first telemetry embedding matrix d, and both the first telemetry embedding matrix s and the first telemetry embedding matrix d are l tp ×l vec The matrix.

4. The deep learning-based multi-source data verification method for power grid measurement centers according to claim 3, characterized in that, In the preprocessing, if the number of remote signaling data or telemetry data is less than 1 tp If there are 1, fill in the special characters. <pad>Make it complete to l tp If the remote signaling data or telemetry data exceeds 1 tp If there are 1, then remove the excess data so that the number of remote signaling data and remote measurement data is 1. tp indivual.< / pad> 5. The method for verifying multi-source data of power grid measurement centers based on deep learning according to claim 1, characterized in that, Step 2-1 further includes: Extracting the effective parts of the first remote signaling embedding matrix s and the first telemetry embedding matrix d, we obtain the effective part matrix s of the first remote signaling embedding matrix. v And the effective part matrix d in the first telemetry embedding matrix v The effective part is the matrix part formed by converting the collected remote signaling data and telemetry data into vectors in the first remote signaling embedding matrix s and the first telemetry embedding matrix d, rather than the vector content converted by the special symbols used for filling. The cosine similarity is used to measure the correlation between remote signaling data vectors and telemetry data vectors, thus obtaining the effective part matrix s in the first remote signaling embedding matrix. v And the effective part matrix d in the first telemetry embedding matrix v The similarity matrix M between them s And M s satisfy: Among them, s v d represents the effective portion of the first remote signaling embedding matrix. v The effective portion of the first telemetry embedding matrix is ​​represented by the similarity matrix M. s The element in the i-th row and j-th column represents the remote signaling data s. i With telemetry data d j degree of relevance, s i Let d represent the i-th remote signaling data. j This represents the similarity matrix obtained from the j-th telemetry data.

6. A deep learning-based multi-source data verification system for power grid measurement centers, utilizing the verification method described in any one of claims 1-5, characterized in that, include: Data acquisition module, preprocessing module, attention interaction module, neural network module; The data acquisition module is used to collect power grid switch data and telemetry data; The preprocessing module preprocesses the collected data to obtain the original embedding matrix; The attention interaction module further processes the original embedding matrix to obtain a new embedding matrix; The neural network module constructs and trains a convolutional neural network to obtain a verification model. The verification model is then used to verify the real-time data of the circuit to determine whether the data is abnormal.

7. A terminal, comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1-5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1-5.

Images