Multi-source data fusion medium and low voltage distribution network topology correctness verification and synchronous feedback method and system

Abstract

The invention provides a multi-source data fusion medium and low voltage power distribution network topology correctness verification and synchronous feedback method and system. According to the embodiment, an advanced data fusion technology is introduced, complex association can be established among the multi-source data, complementarity among the data sources is fully utilized, and the integrity and accuracy of the data are improved. The steps of data cleaning, preprocessing, feature extraction and the like are involved, so that the quality of fused data is ensured. In a topology verification algorithm, the embodiment of the invention can keep stability while considering multi-source data relevance, and identify potential modes and influence factors among different data sources, thereby more accurately performing topology calibration. According to the method, the low-rank approximation technology is introduced to reduce the time complexity of neural network reasoning, one-time parallel prediction of the power grid topology is achieved, and the method is suitable for real-time synchronous feedback. The low-rank approximation technology is innovatively adopted, compared with a traditional search mode based on iterative loop, the reasoning speed is increased, real-time synchronous feedback becomes possible, and important practicability and efficiency are achieved.

Description

Technical Field

[0001] The present invention relates to the fields of distribution network big data mining technology and artificial intelligence technology, and specifically to a multi-source data fusion medium and low voltage distribution network topology correctness verification and synchronous feedback method and system. Background Art

[0002] In the operation of today's power system, medium and low voltage distribution networks, as key components of energy transmission and distribution, play a vital role. However, with the continuous development and complexity of energy systems, medium and low voltage distribution networks face a series of challenges, such as variable load demand, increasing access to distributed energy, and dynamic changes in network topology.

[0003] In order to better control both the generation and demand aspects of the distribution network, monitoring tools (such as state estimation) are needed, among which topology information is a very important condition. Topology correctness verification in power grids refers to the process of verifying the topology structure in the power system. Its goal is to ensure that the power grid model is consistent with the actual power grid topology structure to avoid operational anomalies caused by problems such as incorrect database entries or incorrectly updated power grid status (such as circuit breaker status, line status after power outage). In the power system, the topology structure represents the connection relationship and power flow between various elements in the power grid (such as substations, transmission lines, transformers, etc.). Through topology correctness verification, power system operators can ensure that the model of the power grid is consistent with the actual operation, so as to better operate, plan and analyze the system and improve the reliability and safety of the power system. For the operation of the distribution network, a highly accurate topology calibration process is required. In the transmission network, the topology structure changes infrequently, and the topology of the power grid can be analyzed by installing special equipment to check connectivity. However, these methods are limited in secondary medium and low voltage distribution networks. For example, in many metropolitan areas, the secondary distribution network is huge and most of it is underground. This fact makes installing any topology-aware equipment time-consuming and expensive.

[0004] One way to solve this problem is to correct the currently running network topology through the collected power grid data and fix the topology errors in the distribution management system database. However, traditional topology verification methods only rely on the distribution network power data collected by the power grid, and these data may be incomplete and noisy, resulting in an incomplete understanding of the true power grid structure. In addition, the topology verification system may have time lags, which makes it impossible to detect and correct problems in the power grid in a timely manner, resulting in instability and unreliability of the power system. In recent years, due to the introduction of many advanced metering infrastructures and smart meters and the rapid development of the Internet, customers' electricity consumption sites can provide unprecedented amounts of data for improving and verifying distribution network models. Therefore, how to effectively use multi-source data to verify the correctness of the topology of the power grid is an urgent problem to be solved.

[0005] However, fusing multi-source data and designing corresponding algorithms is still a huge challenge. First, the fusion of multi-source data is a very technically challenging task. Different data sources may use different formats, units and acquisition frequencies, so the first task is to establish a unified data model. Data cleaning and preprocessing are inevitable steps to ensure that various data can be correctly interpreted and utilized. In addition, the spatiotemporal characteristics of the data need to be taken into account, because the operation of the distribution network is a dynamic process, and the timestamp and timing relationship of the data are crucial to the accuracy of the algorithm. Secondly, designing algorithms that can effectively process multi-source data is the key. Traditional topology verification algorithms may perform well when processing a single data source, but in the case of multi-source data, the correlation and mutual influence between different data need to be considered, which may involve the application of complex data mining, machine learning or deep learning techniques to extract implicit topological information from the data. The design of the algorithm should not only consider accuracy, but also focus on computational efficiency, because real-time performance is crucial to the stability of the power system. In addition, due to the complexity and criticality of the power system, the interpretability of the algorithm is also an important aspect. Operators and decision makers need to be able to understand how the algorithm works in order to make informed decisions. Therefore, algorithm design should not only pursue high accuracy, but also have good interpretability and understandability. Effectively using multi-source data for topology correctness verification requires overcoming the difficulties of data integration and processing, and ensuring that the designed algorithm can accurately and efficiently correct grid topology errors and provide synchronous feedback. Summary of the invention

[0006] The technical problem to be solved by the present invention is how to accurately and efficiently correct power grid topology errors to perform synchronous feedback.

[0007] The present invention solves the above technical problems through the following technical means:

[0008] The method for checking the correctness of the topology of a medium and low voltage distribution network and synchronous feedback based on multi-source data fusion includes the following steps:

[0009] Step 1: Preprocess and fuse the data from the user side and the power grid to obtain the power grid data feature X 1 、User-side data features X 2 , fusion data feature X 3 ; Extract the topological features of the user's network topology to obtain the topological feature X;

[0010] Step 2: The power grid data feature X 1 、User-side data features X 2 , fusion data feature X 3 , perform topological correctness check on topological feature X and obtain the check result. The specific process is as follows:

[0011] First, Transformer Block is used to transform the topological features X and the power grid data features X 1 、User-side data features X 2 , fusion data feature X 3 Interact and get feature Y 1 , Feature Y 2 , Feature Y 3 :

[0012] Y 1 = TransformerBlock(X, X 1 )

[0013] Y 2 = TransformerBlock(Y 1 , X 2 )

[0014] Y 3 = TransformerBlock(Y 2 , X 3 )

[0015] Secondly, for feature Y 1 , Feature Y 2 , Feature Y 3 Perform splicing operation to get Y f :

[0016] Y f =Concat(Y 1 ,Y 2 ,Y 3 )

[0017] Among them, Concat() is a concatenation operation;

[0018] Finally, Yf Send it to the multi-layer perceptron for prediction, get the topology verification result, and modify the topology structure according to the verification result;

[0019] For Transformer Block, the input feature is the topological feature X, and the feature to be interacted Y is the power grid data feature X 1 、User-side data features X 2 , fusion data feature X 3 One of them is the multi-head attention, assuming that the input feature is X, the feature to be interacted is Y, and H represents the number of attention heads;

[0020] First, perform a linear mapping on the input features:

[0021] Query i =XW i Q ,Key i =YW i K ,Value i =YW i V ,

[0022] Where i is the index of the attention head, i = 1, 2, ..., H, W i Q ,W i K ,W i V is the projection matrix corresponding to the i-th attention head, which is mapped using the input feature X to obtain the query Query i , use the feature Y to be interacted with to get the key Key i Value i ;

[0023] Normally, the attention score is calculated as:

[0024]

[0025] Where d is the feature dimension;

[0026] In the extreme case, in the original attention matrix Query i Key i T The low-rank matrix W is introduced and expressed as W = UV T , where U and V T is the matrix after low-rank decomposition, so the attention score calculation formula becomes: The attention score calculation formula becomes:

[0027]

[0028] in where r is the rank of the low-rank matrix.

[0029] The present invention proposes a multi-source data fusion method, which uses Bert to process multi-source heterogeneous data and can adaptively and effectively integrate information from different data sources. The multi-source data fusion method fine-tuned by Bert overcomes the problem that the traditional method relies on a single data source and may be incomplete and noisy. The present invention introduces a graph neural network to extract the topological features of the power grid and capture the complex relationship and topological structure between nodes in the power grid. By verifying the correctness of the power grid topology based on a deep learning method, the problem that the probability model in the traditional method is difficult to fully cover all aspects of the topological structure is overcome, making the network more suitable for complex power system structures and actual data scenarios, and providing a more advanced solution for the safety and reliability of the power system. The present invention introduces low-rank approximation technology to reduce the time complexity of neural network reasoning, realizes a one-time parallel prediction of the power grid topology, and is suitable for real-time synchronous feedback. The innovative use of low-rank approximation technology improves the reasoning speed compared to the traditional iterative cycle-based search method, making real-time synchronous feedback possible, and has important practicality and efficiency.

[0030] Furthermore, the data preprocessing method for the user side and the power grid in step 1 is as follows: first, data cleaning is performed to ensure the quality and consistency of the input data by deleting or repairing abnormal values, error values, and missing values; then, standardization and normalization are performed to normalize various data to similar scales for better comparison and fusion; finally, a pre-trained Bert model is used to extract features from the user side data and the power grid data to obtain the power grid data feature X 1 、User-side data features X 2 .

[0031] Furthermore, for the power grid data feature X 1 、User-side data features X 2 , use multi-layer perceptron to align and get the fusion data feature X 3 .

[0032] Furthermore, in step 1, the method for extracting the topological feature X is:

[0033] Assume that the network topology has N nodes, each node forms an N×D-dimensional matrix M, and then the relationship between each node forms an N×N-dimensional matrix A, that is, the adjacency matrix; M and A are the inputs of the graph convolutional network;

[0034]

[0035] Among them, H (l)is the feature of the lth layer, representing the feature matrix of the input node. For the input layer, H is M; W (l) is the weight matrix of the lth layer; σ is the nonlinear activation function, It is a modified version of the adjacency matrix A with self-loops added to ensure that each node takes its own characteristics into account; yes The diagonal degree matrix of ; l represents the number of layers of the network.

[0036] The present invention also provides a multi-source data fusion medium and low voltage distribution network topology correctness verification and synchronous feedback system, comprising:

[0037] The data preprocessing and fusion module is used to preprocess and fuse the data from the user side and the power grid to obtain the power grid data feature X 1 、User-side data features X 2 , fusion data feature X 3 ; Extract the topological features of the user's network topology to obtain the topological feature X;

[0038] Verification feedback module, used for the power grid data feature X 1 、User-side data features X 2 , fusion data feature X 3 , perform topological correctness check on topological feature X and obtain the check result. The specific process is as follows:

[0039] First, Transformer Block is used to transform the topological features X and the power grid data features X 1 、User-side data features X 2 , fusion data feature X 3 Interact and get feature Y 1 , Feature Y 2 , Feature Y 3 :

[0040] Y 1 = TransformerBlock(X, X 1 )

[0041] Y 2 = TransformerBlock(Y 1 , X 2 )

[0042] Y 3 = TransformerBlock(Y 2 , X 3 )

[0043] Secondly, for feature Y 1 , Feature Y 2 , Feature Y3 Perform splicing operation to get Y f :

[0044] Y f =Concat(Y 1 ,Y 2 ,Y 3 )

[0045] Among them, Concat() is a concatenation operation;

[0046] Finally, Y f Send it to the multi-layer perceptron for prediction, get the topology verification result, and modify the topology structure according to the verification result;

[0047] For Transformer Block, the input feature is the topological feature X, and the feature to be interacted Y is the power grid data feature X 1 、User-side data features X 2 , fusion data feature X 3 One of them is the multi-head attention, assuming that the input feature is X, the feature to be interacted is Y, and H represents the number of attention heads;

[0048] First, perform a linear mapping on the input features:

[0049] Query i =XW i Q ,Key i =YW i K ,Value i =YW i V ,

[0050] Where i is the index of the attention head, i = 1, 2, ..., H, W i Q ,W i K ,W i V is the projection matrix corresponding to the i-th attention head, which is mapped using the input feature X to obtain the query Query i , use the feature Y to be interacted with to get the key Key i Value i ;

[0051] Normally, the attention score is calculated as:

[0052]

[0053] Where d is the feature dimension;

[0054] In the extreme case, in the original attention matrix Query i Key i T The low-rank matrix W is introduced and expressed as W = UV V , where U and V T is the matrix after low-rank decomposition, so the attention score calculation formula becomes: The attention score calculation formula becomes:

[0055]

[0056] in where r is the rank of the low-rank matrix.

[0057] Furthermore, the data preprocessing method for the user side and power grid data in the data preprocessing and fusion module is as follows: first, data cleaning is performed to ensure the quality and consistency of the input data by deleting or repairing abnormal values, error values ​​and missing values; then, standardization and normalization are performed to normalize various data to similar scales for better comparison and fusion; finally, the pre-trained Bert model is used to extract features from the user side data and the power grid data to obtain the power grid data feature X 1 、User-side data features X 2 .

[0058] Furthermore, for the power grid data feature X 1 、User-side data features X 2 , use multi-layer perceptron to align and get the fusion data feature X 3 .

[0059] Furthermore, in the data preprocessing and fusion module, the method for extracting the topological feature X is:

[0060] Assume that the network topology has N nodes, each node forms an N×D-dimensional matrix M, and then the relationship between each node forms an N×N-dimensional matrix A, that is, the adjacency matrix; M and A are the inputs of the graph convolutional network;

[0061]

[0062] Among them, H (l) is the feature of the lth layer, representing the feature matrix of the input node. For the input layer, H is M; W (l) is the weight matrix of the lth layer; σ is the nonlinear activation function, It is a modified version of the adjacency matrix A with self-loops added to ensure that each node takes its own characteristics into account; yes The diagonal degree matrix of ; l represents the number of layers of the network.

[0063] The present invention also provides a processing device, characterized in that it includes at least one processor and at least one memory communicatively connected to the processor, wherein: the memory stores program instructions executable by the processor, and the processor calls the program instructions to execute the method as described above.

[0064] The present invention also provides a computer-readable storage medium, characterized in that the computer-readable storage medium stores computer instructions, and the computer instructions enable the computer to execute the above-mentioned method.

[0065] The advantages of the present invention are:

[0066] 1. The present invention proposes a multi-source data fusion method, which uses the large language model Bert for fine-tuning, so that information from different data sources can be adaptively and effectively integrated. The uniqueness of this technology is that it can not only make up for the defects of a single source of power data, but also establish a more complex relationship between multi-source data, giving full play to the complementarity between various data sources. In addition, this patent uses a graph neural network to extract the topological characteristics of the power grid, which can capture the complex relationships and topological structures between nodes in the power grid, and provide more meaningful representations for subsequent tasks.

[0067] 2. The present invention introduces a graph neural network to extract grid topology features and capture the complex relationships and topological structures between nodes in the power grid. By using a deep learning-based method to verify the correctness of the grid topology, the problem that the probability model in the traditional method is difficult to fully cover all aspects of the topological structure is overcome, making the network more suitable for complex power system structures and actual data scenarios, and providing a more advanced solution for the safety and reliability of the power system. The present invention proposes a new method based on deep learning, which adopts a data-driven approach and can more comprehensively consider and process multiple factors of the grid topology. In the neural network, data features from three different sources are used to interact with the grid topology, so that the network can make full use of the features of different domains to adaptively judge the correctness of the grid topology. This innovative method uses a deep learning model to determine the correctness of the grid topology in a more detailed and comprehensive manner, making the network more suitable for complex power system structures and actual data scenarios, and providing a more advanced solution for the safety and reliability of the power system.

[0068] 3. The present invention introduces low-rank approximation technology to reduce the time complexity of neural network reasoning, realizes one-time parallel prediction of power grid topology, and is suitable for real-time synchronous feedback. The innovative use of low-rank approximation technology improves the reasoning speed compared to the traditional search method based on iterative loops, making real-time synchronous feedback possible, which has important practicality and efficiency. This is of great significance for real-time monitoring of power systems and responding to potential problems, and provides more accurate and rapid support for real-time decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 It is a technical framework for topology correctness verification and synchronous feedback of medium and low voltage distribution networks based on multi-source data fusion of the present invention;

[0070] Figure 2 It is a multi-source data fusion framework based on neural network in an embodiment of the present invention;

[0071] Figure 3 This is an example of a topology verification network based on deep learning in an embodiment of the present invention;

[0072] Figure 4 This is an example of the Transformer Block structure in an embodiment of the present invention;

[0073] Figure 5 This is an example of a topology verification result predicted by a multilayer perceptron in an embodiment of the present invention. DETAILED DESCRIPTION

[0074] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in combination with the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0075] like Figure 1 As shown, the specific algorithm of this embodiment is divided into multi-source data fusion, topology verification based on deep learning and synchronous feedback technology. The following is a detailed description of each principle and actual algorithm process.

[0076] Step 1: Multi-source data feature extraction and fusion

[0077] When processing the given user-side data and power grid data, this embodiment takes a series of steps to ensure the high quality and applicability of the data. Considering that their formats may not be uniform, this embodiment first performs data cleaning to ensure the quality and consistency of the input data by removing or repairing outliers, error values, and missing values. Subsequently, this embodiment performs standardization and normalization to normalize various data to similar scales for better comparison and fusion. Finally, this embodiment uses a pre-trained Bert model to extract features from the user-side data and power grid data to capture key information and context in the data, and obtain the power grid data feature X 1 、User-side data features X 2Since these two features correspond to two different domains, their features are not aligned. Therefore, this embodiment uses a feature alignment module (multi-layer perceptron) to align the features and then obtain the final fusion feature. Obtain the fusion data feature X 3 .

[0078] For the user topological network structure data of the power grid, since the topological information has a graph structure, this embodiment uses a graph neural network (GNN) to extract its features. GCN (Graph Convolutional Network) is a classic model of graph neural network, which is mainly used to process graph structure data and learn the representation of nodes in order to perform various tasks on the graph. Its core idea is to iteratively learn nodes through local neighborhood information. It is based on the spectral theory of the graph Laplacian matrix and uses graph convolution operations to update the representation of nodes. By iteratively learning the information of nodes and edges, GNN can capture the complex relationship and topological structure between nodes in the power grid and provide more meaningful representations for subsequent tasks. Specific steps are as follows: Assume that this embodiment has a batch of graph data, in which there are N nodes, each node has its own features, and this embodiment assumes that the features of these nodes form an N×D-dimensional matrix M, and then the relationship between each node will also form an N×N-dimensional matrix A, that is, an adjacency matrix. M and A are the inputs of the model.

[0079]

[0080] Among them, H (l) is the feature of the lth layer, representing the feature matrix of the input node. For the input layer, H is M; W (l) is the weight matrix of the lth layer; σ is the nonlinear activation function, It is a modified version of the adjacency matrix A with a self-loop, ensuring that each node takes its own characteristics into account; yes The diagonal degree matrix of ; l represents the number of layers of the network.

[0081] The above operation process includes linear transformation of the input feature matrix, information aggregation through the adjacency matrix, and finally application of the activation function. By stacking multiple layers of GCN, the network can gradually capture more distant neighborhood information and improve the representation ability of the entire graph. It can handle graphs of varying sizes, and considers the topological structure of the graph in node representation learning, and finally obtains the topological feature X. The whole process is as follows Figure 2 shown.

[0082] Step 2: Topology verification based on deep learning

[0083] The topology correctness check in the power grid refers to the process of verifying the topology structure in the power system. Its goal is to ensure that the power grid model is consistent with the actual power grid topology structure to avoid operational anomalies caused by incorrect database entries or incorrectly updated power grid status. The verification process involves a lot of inspection work, including connectivity check, loop detection, node matching check, device status consistency check, etc. Traditional algorithms are difficult to cover all aspects, so this embodiment designs a method based on deep learning to consider multiple factors in a data-driven way. The network structure is as follows: Figure 3 As shown:

[0084] 1) In order to effectively process sequence data and capture long-range dependencies, this embodiment uses TransformerBlock for feature interaction. Its multi-head cross attention can effectively model the relationship between different source features, so that the topological features can fully learn information from other sources; multi-head self-attention can fully model the long-range dependencies of the topological features themselves, so that the features have global context information. The specific network structure of Transformer Block is as follows: Figure 4 :

[0085] For multi-head attention, assume that the input feature is X and the feature to be interacted is Y, which is the power grid data feature X 1 、User-side data features X 2 , fusion data feature X 3 One of them is to use H to represent the number of attention heads.

[0086] First, perform a linear mapping on the input features:

[0087] Query i =XW i Q ,Key i =YW i K ,Value i =YW i V ,

[0088] Where i is the index of the attention head, i = 1, 2, ..., H, W i Q ,W i K ,W i v is the projection matrix corresponding to the i-th attention head, which is mapped using the input feature X to obtain the query Query i , use the feature Y to be interacted with to get the key Key i Value i ;

[0089] Then calculate the attention score:

[0090]

[0091] Where d is the feature dimension;

[0092] Then use the attention score to value Value i Perform weighted summation to obtain the feature output of the i-th attention head:

[0093] Head i =Attention i Value i

[0094] Multi-head attention uses different attention heads to focus on different areas of the feature, and then fuses the outputs of multiple attention heads to get the final model output. The formula is as follows:

[0095] MultiHead=Concat(Head 1 ,...,Head H )W O

[0096] Where W O is the final output weight matrix. Overall, this mechanism allows the model to learn multiple relationships from the input sequence and assign different learning capabilities to different attention heads.

[0097] In multi-head cross attention, this embodiment maps the topological features to Query and maps the features of the source to be interacted to Key and Value; while in multi-head self-attention, Query, Key and Value are all obtained by mapping the current input features. After the multi-head self-attention sublayer, a feedforward neural network (FFN) is usually added to perform nonlinear transformations on the features. This feedforward network is fully connected and contains one or more hidden layers. This helps to further extract features and introduce nonlinearity, as shown in the following formula:

[0098] FFN(x)=ReLU(xW 1 +b 1 )W 2 +b 2

[0099] Where W 1 and b 1 are the weights and biases of the first fully connected layer, ReLU is the activation function, and W 2 and b 2 are the weights and biases of the second fully connected layer.

[0100] In addition, residual connections are added between the input and output of each sublayer (self-attention, feedforward neural network), and layer normalization is performed. Residual connections help alleviate the gradient vanishing problem, while layer normalization helps stabilize the training process. The formula is as follows:

[0101] Output=LayerNorm(x+Sublayer(x))

[0102] Where Sublayer(x) represents the output of a sublayer, such as the output of a self-attention mechanism or a feedforward neural network.

[0103]

[0104] where μ and σ are the mean and standard deviation, respectively, and γ and β are learnable scaling and offset parameters.

[0105] 1) In order to verify the correctness of the power system topology through multi-source data, the first consideration is how to effectively integrate the existing network topology features and information from other data sources. Only after the network topology fully understands the global context information of other data sources can the network determine whether the current topology structure is different from the actual physical circuit and make necessary corrections. In this process, this embodiment uses Transformer Block to realize the interaction between topological features and three different data source features. Considering the high correlation between the power grid data features and the distribution network topological features, this embodiment first interacts the topological features with the power grid data features. As mentioned in the previous article, the distribution network power data collected by the power grid may be incomplete and noisy, which may lead to the network's incomplete understanding of the power grid structure. Therefore, this embodiment then uses the data features on the user side to interact with the topological features to enhance the expression ability of the topological features. Finally, since the power grid data features and the user data features themselves have a certain correlation, but interacting with them separately may still have some shortcomings, this embodiment further interacts the fused features with the topological features. This step helps the network to understand the current power grid structure more comprehensively and improve the accuracy of the topological correctness. Through the comprehensive utilization and interaction of such multi-source data, this embodiment can more effectively verify the correctness of the power system topology and improve the reliability and safety of system operation. In summary, the whole process is:

[0106] Y 1 = TransformerBlock(X, X 1 )

[0107] Y 2 = TransformerBlock(Y 1 , X 2 )

[0108] Y 3 = TransformerBlock(Y 2 , X 3 )

[0109] Among them, X, X 1 , X 2 and X 3 They represent topological features, power grid data features, user-side data features, and financial data features respectively. That is to say, the topological features will interact with the features of other sources in a cascade manner to obtain the interactive feature Y 1 ,Y 2 ,Y 3 .

[0110] 2) After obtaining the features after the three interactions, this embodiment uses a multi-layer perceptron to predict the topology verification result. Specifically, this embodiment outputs a feature after each interaction, and finally this embodiment splices the three features:

[0111] Y f =Concat(Y 1 ,Y 2 ,Y 3 )

[0112] Concat() is a concatenation operation. After sufficient interaction of features from different sources, Y f It has the function of predicting the network topology verification, and then this embodiment sends it to the multi-layer perceptron for prediction. The multi-layer perceptron of this embodiment has enough capacity to predict the required topology verification results. The weight matrix of the first hidden layer is W (1) , with a bias of b (1) , the activation function is f is the sigmoid function, and the output is:

[0113] h (1) =f(W (1) ·Y f +b (1) )

[0114] Considering the comprehensive efficiency and accuracy, this embodiment uses a three-layer perceptron, that is, repeating the above steps three times, and outputting h (1)

[0115] As the input of the next layer. The flowchart is as follows Figure 5 As shown in the figure: The output of the multilayer perceptron is an N×N dimensional matrix, where the (i,j)th represents the correctness probability of the edge between the i-th node and the j-th node. When this probability is lower than a threshold, it is considered that there is no edge between the two nodes. The network topology is verified based on this correctness matrix. If the topological structure is incorrect, feedback modification can be performed. Figure 5 The topology diagram on the right is an example. The crossed lines in the diagram should not exist and can be used to correct the original topology.

[0116] This embodiment introduces this interactive mode to make the network more suitable for complex power system structures and actual data scenarios, which helps to improve the accuracy and reliability of power system monitoring and prediction. During the network training process, this embodiment inputs an erroneous topology structure that does not match the current physical power grid structure, combines the current power grid data and user-side data to predict the network topology part that should be modified, and then uses the correct topology as a supervisory signal to train the network end-to-end. After the training is completed, this embodiment can use the model to analyze the current data to determine whether the network topology is incorrect.

[0117] In this embodiment, real-time feedback under extreme conditions is more important. Generally speaking, the above design can meet the real-time synchronous feedback requirements of the network, but under some extreme requirements, such as peak power consumption, this embodiment needs to respond faster, so this embodiment additionally designs a network that can meet higher real-time requirements on the basis of the above. As mentioned above, the calculation formula for the attention score is:

[0118]

[0119] Its computational complexity is O(n 2 ), although the existing hardware foundation can complete reasoning and synchronous feedback very quickly, in order to ensure that the requirements can be met even when the real-time requirements are extremely high, this embodiment introduces low-rank approximation technology. Low-rank approximation reduces the rank of the attention matrix in the self-attention mechanism, reduces the number of parameters and the amount of calculation, thereby improving the calculation efficiency while maintaining the model performance. Specifically, in this embodiment, the original attention matrix Query i Key i T A low-rank matrix W is introduced and expressed as W = UV T , where U and V T is the matrix after low-rank decomposition, so the attention score calculation formula becomes:

[0120]

[0121] in Through such a low-rank approximation, the original matrix multiplication computational complexity is reduced from O(n 2 ) is reduced to O(n·r), where r is the rank of the low-rank matrix, which is much smaller than n. This greatly reduces the computational complexity and enables the topological verification structure to be inferred very quickly to satisfy synchronous feedback.

[0122] This embodiment proposes a method for multi-source data fusion, and realizes the effective integration of different data sources by introducing advanced data fusion technology. This technology can not only correct the defects of single source of power data, but also establish more complex associations between multi-source data, give full play to the complementarity between various data sources, and thus improve the overall integrity and accuracy of the data.

[0123] In terms of the topology verification algorithm, this embodiment not only considers the correlation between multi-source data, but also fully considers the stability of the algorithm in algorithm design. By identifying the potential patterns and influencing factors between different data sources, the algorithm can perform topology calibration more accurately, allowing the system to better adapt to the changing power grid structure.

[0124] In terms of real-time data processing and feedback mechanism, this embodiment designs an acceleration module of the cache technology attention mechanism based on the above algorithm. That is, the attention mechanism of low-rank approximation is used to ensure the speed of model reasoning so as to provide timely feedback. (Real-time data processing and feedback mechanism are introduced to ensure that the system can respond immediately when topology calibration information is available. Stream processing technology, distributed computing and efficient data transmission protocols are used to reduce processing delays and improve the response speed of the system. At the same time, consider introducing cache mechanisms and parallel computing to cope with data processing bottlenecks that may exist in large-scale and complex distribution network systems.)

[0125] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. Multi-source data fusion medium and low voltage distribution network topology correctness verification and synchronous feedback method, It is characterized in that The following steps are involved: Step 1: Preprocess and fuse the data from the user side and the power grid to obtain the power grid data feature X 1 、User-side data features X 2 , fusion data feature X 3 ; Extract the topological features of the user's network topology to obtain the topological feature X; Step 2: The power grid data feature X 1 、User-side data features X 2 , fusion data feature X 3 , perform topological correctness verification on topological feature X and obtain verification results. The specific process is as follows: First, Transformer Block is used to transform the topological features X and the power grid data features X 1 、User-side data features X 2 , fusion data feature X 3 Interact and get feature Y 1 , Feature Y 2 , Feature Y 3 : Y 1 =TransformerBlock(X,X 1 ) AND 2 =TransformerBlock(Y 1 ,X 2 ) AND 3 =TransformerBlock(Y 2 ,X 3 ) Secondly, for feature Y 1 , Feature Y 2 , Feature Y 3 Perform splicing operation to get Y f : AND f =Concat(And 1 ,AND 2 ,AND 3 ) Among them, Concat() is a concatenation operation; Finally, Y f Send it to the multi-layer perceptron for prediction, get the topology verification result, and modify the topology structure according to the verification result; For Transformer Block, the input feature is the topological feature X, and the feature to be interacted Y is the power grid data feature X 1 、User-side data features X 2 , fusion data feature X 3 One of them is the multi-head attention, assuming that the input feature is X, the feature to be interacted is Y, and H represents the number of attention heads; First, perform a linear mapping on the input features: Query i =XW i Q , where i is the index of the attention head, i = 1, 2, ..., H, is the projection matrix corresponding to the i-th attention head, which is mapped using the input feature X to obtain the query Query i , use the feature Y to be interacted with to get the key Key i Value i ; Normally, the attention score is calculated as: Where d is the feature dimension; In the extreme case, in the original attention matrix Queryi The low-rank matrix W is introduced and expressed as W = UV T , where U and V T is the matrix after low-rank decomposition, so the attention score calculation formula becomes: The attention score calculation formula becomes: in where r is the rank of the low-rank matrix.

2. According to the multi-source data fusion method for medium and low voltage distribution network topology correctness verification and synchronous feedback according to claim 1, It is characterized in that The method for preprocessing the data of the user side and the power grid in step 1 is as follows: first, data cleaning is performed to ensure the quality and consistency of the input data by deleting or repairing abnormal values, error values ​​and missing values; then, standardization and normalization are performed to normalize various data to similar scales for better comparison and fusion; finally, the pre-trained Bert model is used to extract features from the user side data and the power grid data to obtain the power grid data feature X 1 、User-side data features X 2 .

3. According to the multi-source data fusion method for medium and low voltage distribution network topology correctness verification and synchronous feedback according to claim 2, It is characterized in that For the power grid data feature X 1 、User-side data features X 2 , use multi-layer perceptron to align and get the fusion data feature X 3 .

4. The method for checking the correctness of the medium and low voltage distribution network topology and providing synchronous feedback based on multi-source data fusion according to any one of claims 1 to 3, It is characterized in that In step 1, the method for extracting the topological feature X is: Assume that the network topology has N nodes, each node forms an N×D-dimensional matrix M, and then the relationship between each node forms an N×N-dimensional matrix A, that is, the adjacency matrix; M and A are the inputs of the graph convolutional network; Where H( l ) is the feature of the lth layer, representing the feature matrix of the input node. For the input layer, H is M; W( l ) is the weight matrix of the lth layer; σ is the nonlinear activation function, It is a modified version of the adjacency matrix A with self-loops added to ensure that each node takes its own characteristics into account; yes The diagonal degree matrix of ; l represents the number of layers of the network.

5. Multi-source data fusion medium and low voltage distribution network topology correctness verification and synchronous feedback system, It is characterized in that include: The data preprocessing and fusion module is used to preprocess and fuse the data from the user side and the power grid to obtain the power grid data feature X 1 、User-side data features X 2 , fusion data feature X 3 ; Extract the topological features of the user's network topology to obtain the topological feature X; Verification feedback module, used for the power grid data feature X 1 、User-side data features X 2 , fusion data feature X 3 , perform topological correctness verification on topological feature X and obtain verification results. The specific process is as follows: First, Transformer Block is used to transform the topological features X and the power grid data features X 1 、User-side data features X 2 , fusion data feature X 3 Interact and get feature Y 1 , Feature Y 2 , Feature Y 3 : Y 1 =TransformerBlock(X,X 1 ) AND 2 =TransformerBlock(Y 1 ,X 2 ) AND 3 =TransformerBlock(Y 2 ,X 3 ) Secondly, for feature Y 1 , Feature Y 2 , Feature Y 3 Perform splicing operation to get Y f : AND f =Concat(And 1 ,AND 2 ,AND 3 ) Among them, Concat() is a concatenation operation; Finally, Y f Send it to the multi-layer perceptron for prediction, get the topology verification result, and modify the topology structure according to the verification result; For Transformer Block, the input feature is the topological feature X, and the feature to be interacted Y is the power grid data feature X 1 、User-side data features X 2 , fusion data feature X 3 One of them is the multi-head attention, assuming that the input feature is X, the feature to be interacted is Y, and H represents the number of attention heads; First, perform a linear mapping on the input features: Queryi=XWiQ, where i is the index of the attention head, i = 1, 2, ..., H, is the projection matrix corresponding to the i-th attention head, which is mapped using the input feature X to obtain the query Query i , use the feature Y to be interacted with to get the key Key i Value i ; Normally, the attention score is calculated as: Where d is the feature dimension; In the extreme case, in the original attention matrix Queryi The low-rank matrix W is introduced and expressed as W = UV T , where U and V T is the matrix after low-rank decomposition, so the attention score calculation formula becomes: The attention score calculation formula becomes: in where r is the rank of the low-rank matrix.

6. According to the multi-source data fusion medium and low voltage distribution network topology correctness verification and synchronous feedback system of claim 5, It is characterized in that The data preprocessing method for the user side and power grid data in the data preprocessing and fusion module is as follows: first, data cleaning is performed to ensure the quality and consistency of the input data by deleting or repairing abnormal values, error values ​​and missing values; then, standardization and normalization are performed to normalize various data to similar scales for better comparison and fusion; finally, the pre-trained Bert model is used to extract features from the user side data and the power grid data to obtain the power grid data feature X 1 、User-side data features X 2 .

7. According to the multi-source data fusion medium and low voltage distribution network topology correctness verification and synchronous feedback system of claim 6, It is characterized in that For the power grid data feature X 1 、User-side data features X 2 , use multi-layer perceptron to align and get the fusion data feature X 3 .

8. A medium and low voltage distribution network topology correctness verification and synchronous feedback system based on multi-source data fusion according to any one of claims 5 to 7, It is characterized in that In the data preprocessing and fusion module, the method for extracting the topological feature X is: Assume that the network topology has N nodes, each node forms an N×D-dimensional matrix M, and then the relationship between each node forms an N×N-dimensional matrix A, that is, the adjacency matrix; M and A are the inputs of the graph convolutional network; Where H( l ) is the feature of the lth layer, representing the feature matrix of the input node. For the input layer, H is M; W( l ) is the weight matrix of the lth layer; σ is the nonlinear activation function, It is a modified version of the adjacency matrix A with self-loops added to ensure that each node takes its own characteristics into account; yes The diagonal degree matrix of ; l represents the number of layers of the network.

9. A processing device, It is characterized in that It comprises at least one processor and at least one memory in communication with the processor, wherein the memory stores program instructions executable by the processor, and the processor calls the program instructions to execute any method according to claims 1 to 4.

10. A computer-readable storage medium, It is characterized in that The computer-readable storage medium stores computer instructions, and the computer instructions enable the computer to execute the method according to any one of claims 1 to 4.

Images