A power distribution system real-time network topology and parameter identification method and system

Abstract

The application provides a power distribution system real-time network topology and parameter identification method and system, comprising: collecting data of smart meter nodes installed in a historical power distribution network to construct a data set; calculating an admittance matrix from the data set, deriving a network topology structure, and preliminarily estimating line parameters; modeling the power grid topology based on a graph convolution network (GCN), calculating the importance of each node, and then iteratively optimizing the placement strategy of a measurement device (SMD) through a loss function; on this basis, using SMD data as the input of a graph neural network (GNN) to identify the real-time network topology and parameters of the system; if the network topology is identified to have changed, the pre-trained GNN parameters are fine-tuned through a transfer learning method to adapt to the new topology structure, and the parameters are re-identified. Through the method, the measurement device placement strategy can be optimized and the real-time performance and accuracy of the power distribution network topology and parameter identification can be improved under the condition of limited measurement devices.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of power distribution system monitoring and analysis, and particularly relates to a power distribution system real-time network topology and parameter identification method and system. BACKGROUND

[0002] With the wide application of distributed energy sources such as solar photovoltaic (PV), electric vehicles (EV) and energy storage systems, the power distribution network is increasingly frequently connected and disconnected with distributed devices. The traditional power distribution network is experiencing a transition from a one-way power supply mode to a two-way interactive mode, which makes the power distribution system more complex and unstable. Therefore, real-time monitoring and parameter identification of the power distribution system are particularly important. Especially for the future smart grid and flexible configuration of the power distribution network, being able to quickly respond to topology changes and accurately identify parameters is the key to ensuring the reliable operation of the power distribution system.

[0003] The traditional power distribution system monitoring method relies on devices such as synchronous phasor measurement units (PMU) for real-time data collection, but due to the high cost and deployment density problems, the power distribution system cannot usually achieve complete observation. In many cases, some areas of the power distribution network do not have sufficient synchronous measuring devices (SMD), which makes it impossible for the system to accurately infer the system state under the conditions of topology changes and load fluctuations. This not only affects the real-time monitoring of power flow, but also limits the fault detection and response capabilities of the power distribution system.

[0004] Previous parameter identification methods rely on traditional measurement techniques such as smart meter (AMI) data, but due to the long measurement period of these devices (e.g. every 15 minutes or longer) and the lack of time synchronization function, they cannot be directly used for real-time parameter identification. In addition, some traditional methods also need to rely on load forecasting or pseudo-measurement to fill in the gaps in the observation data, but these methods will reduce the estimation accuracy and have a large error in actual application.

[0005] In order to solve these problems, researchers have begun to try to use machine learning-based methods for parameter identification and topology identification of power distribution systems. Among them, GCN as a powerful neural network has attracted widespread attention due to its performance in high-dimensional complex data. However, most existing neural network methods assume that the power distribution network topology is fixed and require a relatively dense measurement device in the network, which is not easy to achieve in real power distribution systems. Therefore, how to accurately identify parameters and topology in real time through neural networks under the condition of limited observation data and frequent switch state changes is still a technical problem that needs to be solved urgently. SUMMARY

[0006] The technical problem to be solved by the present application is how to optimize the placement strategy of the measuring device and improve the real-time performance and accuracy of the power distribution network topology and parameter identification under the condition of limited measuring devices.

[0007] The present application solves the above technical problems by the following technical means:

[0008] The present application provides a power distribution system real-time network topology and parameter identification method, comprising:

[0009] S1, collecting data of nodes installed with smart meters in the historical power distribution network, and constructing a data set;

[0010] S2, using the measurement history data of the smart meter to solve the admittance matrix and deduce the network topology structure and preliminarily estimate the line parameters;

[0011] S3, modeling the power grid topology based on the graph convolution network GCN, taking each power grid node as a node of the graph and the power grid line as an edge of the graph, calculating the importance of each node through the message passing mechanism, and then iteratively optimizing the graph convolution network GCN parameters through the total loss function, and obtaining the placement strategy of the measuring device SMD by using the graph convolution network GCN;

[0012] S4, on the basis of optimizing the placement of the real-time measuring device SMD, using the real-time measuring device SMD data as the input of the graph neural network GNN to identify the real-time network topology and parameters of the system;

[0013] S5, using the real-time current data collected by the SMD to identify whether the power distribution network topology state changes; if it changes, the pre-trained graph neural network GNN parameters are fine-tuned to adapt to the new topology structure, and the parameters are re-identified.

[0014] Further, the node data installed with the smart meter includes the voltage amplitude, active power and reactive power; the historical data of the nodes installed with the smart meter in the power distribution network are used as the input features; the node voltage amplitude matrix , is constructed for the voltage amplitude of the node , and the active power injection matrix and the reactive power injection matrix are constructed.

[0015] Further, the S2 comprises the following steps:

[0016] S21, constructing a power flow model equation as follows:

[0017]

[0018] ​

[0019] wherein, and are the active and reactive power injections of node , respectively, and are the voltage magnitudes of node and node , respectively, and are the conductance and susceptance in admittance matrix, respectively, and are the voltage phase angles of node and node , respectively;

[0020] S22, simplifying the power flow model equation into linearized equation as follows:

[0021]

[0022]

[0023] S23, based on equation (6) and equation (7), the conductance and the susceptance between each node are approximately calculated, and then the admittance matrix is obtained, the elements of the admittance matrix are as follows:

[0024]

[0025]

[0026] wherein, is the self-admittance, is the mutual admittance;

[0027] S24, the admittance matrix elements are sparsified, and a threshold is set to exclude unimportant connections, and the specific exclusion manner is as follows:

[0028]

[0029] wherein, is the admittance between node and node , is the threshold, and the expression is as follows:

[0030]

[0031] wherein, The nodes in the admittance matrix In phase The self-admittance at that node is the admittance value at that node. It is the number of nodes in the network. It is an empirical constant used to control the degree of sparsity;

[0032] S25. Based on the topology inferred in S24, the conductance of the line parameters of the distribution network is analyzed using a linear regression method. and susceptance Make a preliminary estimate.

[0033] Furthermore, the linear regression method described in S25 for analyzing the conductance of line parameters in the power distribution network... and susceptance The specific steps for making an estimate are as follows:

[0034] The objective function is defined using the least squares (OLS) method, as shown below:

[0035]

[0036] in, It is a node active power, These are elements of the admittance matrix. It is a node The voltage amplitude.

[0037] Further, S3 includes the following steps:

[0038] S31. Based on the topology derived in S2, construct the adjacency matrix. , It is an n-order square matrix, in the following form:

[0039]

[0040] for In distribution networks, the element is represented as... The electrical connection relationships are as follows:

[0041]

[0042] Set Node Initial state feature vector: , Data from smart meters includes voltage and power; the node's initial state feature vector is then used. and adjacent order matrices In the input graph convolutional network GCN, the graph convolutional network is stacked by multiple layers, and the iterative calculation formula of the node feature is as follows:

[0043]

[0044]

[0045] wherein, , is a degree matrix; is a feature of the i-th layer, is a to-be-trained weight parameter of the i-th layer, is a ReLU activation function, denotes information aggregation of a node and its neighbors; S32, the node feature transformed by formula (15) and

[0046] , the attention score of the neighbor node is calculated by introducing a LeakyReLU function , as follows:

[0047]

[0048] wherein, the weight matrix is a to-be-trained parameter, denotes the attention coefficient of a node to a neighbor , and is a to-be-trained attention weight parameter; denotes feature vector splicing; is a feature of the i-th node; and are linearly transformed features of the node and the neighbor , respectively; the normalized attention weight is calculated by a Softmax function , as follows:

[0049]

[0050] wherein, is a neighbor set of the node ; denotes the importance of the node at the node ;

[0051] S33, the score of the measurement device candidate point is utilized to screen the measurement device candidate point;

[0052] ​​​S34. Introduce the optimization objective of resource consumption during the training process and design a loss function.

[0053] Furthermore, S33 includes the following steps:

[0054] S331. An attention mechanism is used to aggregate neighborhood information to obtain new features of candidate points for the measurement device. As shown in the following formula:

[0055]

[0056] in, For nodes In the Features of the layer; They are neighbors Transformed features For the neighbors For nodes Contribution weight;

[0057] S332, the result from formula (19) Calculate the candidate point scores for the measurement devices at all nodes. As shown in the following formula:

[0058]

[0059] S333, Select The K highest-scoring measurement points are selected as candidate points for the optimal measurement device. In the graph attention network (GAT), these points are then... The score is passed to a Classify by layers, The function converts the output of each node into a probability distribution of whether SMD is installed, as shown below:

[0060]

[0061] in, The node predicted by the model The probability of needing SMD.

[0062] Further, S34 includes the following steps:

[0063] S341. The node classification loss function uses the binary cross-entropy loss function to calculate the difference between the predicted probability of whether each node needs SMD and the actual label, as shown in the following formula:

[0064]

[0065] in, It is a node The true label, is the node predicted by the model the probability of whether SMD is needed;

[0066] S342, introduce a resource consumption regularization term to limit the number of SMDs allocated, as follows:

[0067]

[0068] wherein, is a regularization coefficient;

[0069] S343, combine equations (22) and (23) to form the total loss function, as follows:

[0070]

[0071] By minimizing the total loss function, the graph convolutional network GCN parameters to be trained are iterated.

[0072] Further, the S4 comprises the following steps:

[0073] S41, according to the installation node of the S3 optimized SMD, collect the data of the installation SMD node in the three-phase power distribution network, including voltage phasor, current phasor, active power and reactive power, and construct a data set wherein, represents the number of nodes installed with SMD, represents the data of the th node, including , respectively, represents the data of the th node including voltage, current, active power, and reactive power, and each parameter is represented as follows: ; After normalizing the data set , the normalized data set

[0074] wherein, represents the three-phase in the power distribution network, represents the number of sampling points, respectively, represents the voltage of the th node in the three-phase at the th sampling point; respectively, represents the current of the th node in the three-phase at the th sampling point; respectively, represents the active power of the th node in the three-phase at the th sampling point; respectively, represents the reactive power of the th node in the three-phase at the Variables at each sampling point; yes The row normalization result represents the normalized value of the node across all sampled data points;

[0075] S42, will Compared with the admittance matrix estimated in step S2 As input to the graph neural network (GNN) model, a transition function is established based on the power flow equation, thereby constructing the GNN model; in a three-phase distribution network, nodes... The state is determined by its own voltage and adjacent node voltage The update is performed using the power flow equation, specifically calculated as follows:

[0076]

[0077] in, For nodes Three-phase complex power injection; For nodes and adjacent nodes Three-phase complex voltage; For nodes The self-admittance matrix, For nodes and adjacent nodes The mutual admittance matrix between them and They are and The complex conjugate transpose; for The set of neighboring nodes of a node; and Calculated from line impedance; This indicates that the Hadamard product is multiplied element by element;

[0078] The power flow equation is converted into a real matrix representation as follows:

[0079]

[0080] in, and They represent the three-phase complex voltages respectively. The real and imaginary parts; It is the real number representation of the nodal impedance matrix; It is the real number representation of the admittance matrix of adjacent nodes; and These are complex power injections. and voltage The complex conjugate, and are the real and imaginary parts of the complex conjugate product of the complex power injection and the complex voltage, respectively; in the graph neural network GNN, this physical constraint is used as a transition function for updating the voltage state at each layer; the global representation of the transition function is:

[0081]

[0082] where, the node state of the entire network; the feature information of the line; the physical transition function defining the state update;

[0083] The graph neural network GNN is used as the backbone network to learn the global topology of the power grid and perform parameter prediction; the node voltage is iteratively calculated using the transition function as follows:

[0084]

[0085] where, denotes the number of iteration steps;

[0086] The iterative calculation continues until the following convergence condition is met:

[0087]

[0088] where, is a small threshold value used to determine convergence;

[0089] S43, in network training, the mean square error MSE is used as the loss function, as follows:

[0090]

[0091] where, denotes the mean square error of the nodes at time t, is the true voltage measured by the SMD, is the predicted voltage of the graph neural network GNN;

[0092] By calculating the gradient of the loss function with respect to the line parameters, the graph neural network GNN line parameters are iteratively optimized;

[0093] Define the current line parameters as , the gradient of the loss function with respect to is calculated by the backpropagation method; for the line parameter of the th line, the gradient calculation is as follows:

[0094]

[0095] in, It is a set of time indexes. Let represent the sample loss function at all time points; then, stochastic gradient descent (SGD) is used to update the line parameters, as follows:

[0096]

[0097] in, It's the learning rate. It is the current iteration step number. In steps The next The line parameters of the line.

[0098] Further, S5 includes the following steps:

[0099] S51. Using real-time current data measured by SMD, determine whether the distribution network topology has changed; set the input data as follows: ,in For the first Current phasor data for each node;

[0100] When topology occurs, current flows between the new branches, and the new branch input... With output current Satisfying continuity:

[0101]

[0102] If the topology has not changed, no current flows in the new branch:

[0103]

[0104] Then the new access branch Total change in current As shown in the following formula:

[0105]

[0106] in, express Mutually( The current flowing into the new branch; express Mutually( The current flowing out of the new branch; express The larger of the current flowing into and out of the new branch; Indicates to Numerical summation, that is, summing of... Sum of the maximum currents in each phase;

[0107] To avoid the influence of noise on a single measurement, a time window method is used to perform multi-time statistics as follows:

[0108]

[0109] As shown in formula (36), if continuous At that moment The sum of currents is greater than the threshold. If the current is zero for an extended period, it is determined that a topological change has occurred; if the current remains zero for an extended period, it is determined that the topology has not changed.

[0110] To identify the topological connections of newly connected devices, compute nodes and Current correlation between As shown in the following formula:

[0111]

[0112] in, express point Current value and The covariance of the current value is shown in formula (38); express point Current variance; express point Current variance; Indicates consecutive moments The average value of the point current; Indicates consecutive moments The average value of the point current;

[0113]

[0114]

[0115]

[0116] Set threshold Adjust the adjacency matrix for As shown in the following formula:

[0117]

[0118] like , then it represents a node and Directly connected;

[0119] S52, after recognizing the topology switching through the change of current data, the trained graph neural network GNN model is fine-tuned through the transfer learning method, so that the model can adapt to the new topology structure, and the process of transfer learning is as follows :

[0120]

[0121] wherein, is the updated graph neural network GNN full connection layer weight parameter, is the full connection layer weight parameter of the original training model, is the parameter update amount obtained by fine-tuning;

[0122] S53, after recognizing the topology change, the fine-tuned graph neural network GNN is used for parameter identification of the system, and the parameter identification process of the system is as follows:

[0123]

[0124] wherein, is the estimated voltage phasor, is the updated GNN network full connection layer weight parameter, is the adjusted topology state, is the graph neural network GNN.

[0125] The application also provides a power distribution system real-time network topology and parameter identification system, which applies the above method during system operation, and comprises the following modules:

[0126] The intelligent meter data construction module is used for collecting the data of the intelligent meter nodes installed in the historical power distribution network, and constructing a data set.

[0127] The line parameter estimation module is used for calculating the admittance matrix by using the measurement historical data of the intelligent meter, deriving the network topology structure, and preliminarily estimating the line parameters.

[0128] The SMD placement optimization module is used for modeling the power grid topology based on the graph convolution network GCN, taking each power grid node as a graph node and the power grid line as a graph edge, calculating the importance of each node through the message passing mechanism, and then iteratively optimizing the graph convolution network GCN parameters through the total loss function, and obtaining the placement strategy of the measurement device SMD by using the graph convolution network GCN.

[0129] The line parameter output module is used for making real-time network topology and parameter identification of the system by using the real-time measurement device SMD data as the input of the graph neural network GNN on the basis of optimizing the placement of the real-time measurement device SMD node.

[0130] A model migration learning module is configured to identify whether the power distribution network topology state changes by using the real-time current data collected by the SMD; if the topology state changes, the pre-trained graph neural network (GNN) parameters are fine-tuned by using the migration learning method to adapt to the new topology structure, and the parameters are re-identified.

[0131] The present application has the advantages of:

[0132] (1) The present application effectively reduces the dependence on high-cost phasor measurement units (PMUs) by using only smart meters and a small number of SMDs, which not only reduces the installation and maintenance costs of equipment, but also makes the method widely applicable to power distribution systems, and even in the case of limited measurement resources, real-time high-precision topology estimation and parameter identification can be achieved.

[0133] (2) The present application optimizes the placement of the SMDs by using the graph convolutional neural network (GCN) and the stochastic gradient descent (SGD) method. By introducing a regularization term and a loss function, the position of the measurement device can be dynamically adjusted with the minimum number of measurement devices to ensure the accuracy and stability of topology estimation and parameter identification.

[0134] (3) The present application uses a deep neural network model based on the graph neural network (GNN) and the fully connected network, which significantly improves the accuracy of power distribution network topology identification. The graph neural network (GNN) is good at processing graph structure data, and the node iteration formula of the transition function of formula (28) is used to update the node voltage feature, while the fully connected network further extracts complex global features through multiple layers of nonlinear transformation. The graph neural network (GNN) is responsible for learning local structural features, and then these features are fused and nonlinearly regressed to obtain more accurate network topology and line parameter estimation.

[0135] (4) The present application combines the migration learning technology, so that the neural network model can quickly adapt to the new structure when the topology changes, avoiding the high computational cost of retraining the model in the traditional method. The parameters of the deep graph neural network (GNN) are saved, and only the fully connected layer weights are fine-tuned to avoid retraining the entire neural network. The present application can realize fast and efficient topology update, ensuring the real-time and accuracy of parameter identification.

[0136] (5) The present application uses the physical power flow equation as the transition function in the graph neural network (GNN), so that the node state (voltage value) update can follow the physical law of the power system. By embedding the known power flow physical equation, the physical constraint is directly integrated into the training process of the graph neural network, ensuring that the output voltage and power flow of the network comply with the actual physical behavior of the power system.

[0137] (6) The application can realize accurate parameter identification in the case of topology change by collecting data in real time and combining the measurement data of the intelligent electric meter and the synchronous measurement device. The application reduces the consumption of resources while improving the real-time performance of the system and effectively improving the accuracy of topology identification and parameter estimation through an optimized measurement point deployment strategy. BRIEF DESCRIPTION OF DRAWINGS

[0138] Figure 1 A flowchart of a power distribution system real-time network topology and parameter identification method according to an embodiment of the application. DETAILED DESCRIPTION

[0139] To make the objectives, technical solutions and advantages of the embodiments of the application clearer, the technical solutions in the embodiments of the application will be described below in a clear and complete manner in conjunction with the embodiments of the application. Obviously, the described embodiments are some but not all of the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the application.

[0140] Embodiment 1

[0141] An embodiment of the application provides a power distribution system real-time network topology and parameter identification method, as shown in Figure 1 , which comprises the following steps.

[0142] S1, collect data of nodes with installed intelligent electric meters in a historical power distribution network, and construct a data set. The specific implementation steps are as follows.

[0143] The node data with installed intelligent electric meters includes voltage amplitude, active power and reactive power. The historical data of the nodes with installed intelligent electric meters in the power distribution network are used as input features. A node voltage amplitude matrix , is constructed, where V is the voltage amplitude of node , and N is the total number of nodes; an active power injection matrix is constructed; and a reactive power injection matrix is constructed.

[0144] S2, use the measurement historical data of the intelligent electric meter to solve the admittance matrix, derive the network topology structure, and preliminarily estimate the line parameters. The specific implementation steps are as follows.

[0145] The power flow model usually includes two types of equations, namely, active power equation and reactive power equation. The power flow of each node is affected by voltage, phase angle and network admittance matrix. For a three-phase power distribution network, the basic equation of power flow is as follows:

[0146] S21, construct the power flow model equation, as follows: ​

[0147]

[0148]

[0149] where, and are the active and reactive power injections of node , and are the voltage magnitudes of node and node , and are the conductance and susceptance in admittance matrix, and are the voltage phase angles of node and node . Based on the power flow equations, the admittance values between nodes can be solved from the active and reactive power injections, which are used for subsequent topology and parameter identification.

[0150] The power flow equations are nonlinear. To simplify the calculation, and the difference of voltage phase angles is small (usually less than 5°), the first-order Taylor expansion approximation can be performed, thereby converting the nonlinear power flow equations into linear equations. The difference of voltage phase angles is small, which can be simplified by the first-order approximation formula (3):

[0151]

[0152] Based on the approximation of formula (3), the power flow equations can be linearized into formula (4) and formula (5):

[0153]

[0154]

[0155] S22, for a three-phase power distribution system, the linearization of the power equation and the voltage phase angle equation of the node can be further simplified to the linearization equation, as follows:

[0156]

[0157]

[0158] S23, based on the approximate calculation of formula (6) and formula (7) between each node conductance and susceptance , and then the admittance matrix is obtained, the elements of the admittance matrix are as follows:

[0159]

[0160]

[0161] wherein, is the self-admittance (diagonal element), is the mutual admittance (off-diagonal element);

[0162] Admittance matrix is the matrix describing the electrical connections of the power distribution network, whose elements represent the electrical connection relationship between node and node , the construction of the admittance matrix depends on the conductance and the susceptance .

[0163] S24, since the power distribution network is usually a sparse network, i.e. most nodes are not directly connected, therefore many elements of the admittance matrix are zero, only the nodes connected to the current node have admittance values. In order to avoid false connections caused by noise or measurement errors, the admittance matrix elements are sparsified, a threshold is set to exclude unimportant connections, the specific exclusion method is as follows:

[0164]

[0165] wherein, is the admittance between node and node , and is the threshold, the expression is as follows:

[0166]

[0167] wherein, is the self-admittance of node in phase , i.e. the admittance value of the node, is the number of nodes in the network, is an empirical constant used to control the degree of sparsification;

[0168] The threshold is used to determine which elements in the admittance matrix are considered to be valid connections. If the absolute value of the element in the admittance matrix is greater than or equal to the threshold , it is considered that there is an electrical connection between node and node . It is considered that there is an electrical connection between node and There is a connection between them; if the absolute value of the element in the admittance matrix is less than a threshold , it is considered that there is no direct connection between the node and the node . In this way of sparsification, unimportant connections can be effectively removed, so as to infer the actual topology of the power distribution network and further analyze and optimize it.

[0169] S25, based on the topology inferred in S24, the line parameters conductance and susceptance of the power distribution network are preliminarily estimated by using a linear regression method. The linear regression method is solved by the least squares method OLS, and the objective function is set as follows:

[0170]

[0171] Among them, is the active power of node , is the element of the admittance matrix, is the voltage amplitude of node .

[0172] S3, based on the graph convolution network GCN, the power grid topology is modeled, each node of the power grid is taken as a node of the graph, and the power grid line is taken as an edge of the graph. The importance of each node is calculated through the message passing mechanism, and then the graph convolution network GCN parameters are iteratively optimized through the total loss function. The placement strategy of the measurement device SMD is obtained by using the graph convolution network GCN. The specific implementation steps are as follows:

[0173] S31, based on the topology inferred in S2, an adjacency matrix is constructed, which is an n-order square matrix and has the following form:

[0174]

[0175] is the element in , which represents the electrical connection relationship of in the power distribution network, and the specific value is as follows:

[0176]

[0177] The initial state feature vector of node is set as: , which comes from the smart meter data including voltage and power; the initial state feature vector of node and the adjacency matrix In the input graph convolutional network (GCN), the graph convolutional network is composed of multiple stacked layers, and its node feature iterative calculation formula is as follows:

[0178]

[0179]

[0180] in, , It is a degree matrix; It is the first Features of the layer It is the first The training weight parameters of the layer, It is the activation function ReLU. This represents the aggregation of information about nodes and their neighbors;

[0181] S32, Node features after transformation using formula (15) and The LeakyReLU function is introduced to calculate the attention scores of neighboring nodes. As shown in the following formula:

[0182]

[0183] Among them, the weight matrix The parameters to be trained, Represents a node To the neighbors Attention coefficient These are the attention weight parameters to be trained; This indicates the concatenation of eigenvectors; and These are nodes and neighbors Features after linear transformation; for Node features; normalized attention weights are calculated using the Softmax function. As shown in the following formula:

[0184]

[0185] in, It is a node The set of neighbors; Candidate points for representative measuring devices At the node The importance of location;

[0186] S33. Scoring candidate points using a measuring device. Candidate points for the measuring device are screened. Specific steps include:

[0187] S331, aggregate neighborhood information using attention mechanism to obtain new features of measurement device candidate points As follows:

[0188]

[0189] Wherein, The node The features of the first layer; The neighbor Transformed features, The contribution weight of the neighbor To the node ;

[0190] S332, calculate the MeasurementImportance Score of all nodes from the results of formula (19) As follows:

[0191]

[0192] S333, select The top K measurement points with the highest scores as the optimal measurement device candidate points, and pass the Scores to a Layer for classification, The function converts the output of each node into a probability distribution of whether to install SMD, as follows:

[0193]

[0194] Wherein, The probability that the model predicts whether the node Needs SMD.

[0195] S34, in order to minimize SMD resource consumption, not only the accuracy of node classification needs to be concerned, but also the optimization goal of resource consumption needs to be introduced in the training process, and the loss function is designed. The specific steps include:

[0196] S341, the node classification loss function uses binary cross entropy loss function to calculate the difference between the predicted probability of whether each node needs SMD and the actual label, as follows:

[0197]

[0198] Wherein, The true label of the node , taking 0 or 1, indicating whether SMD is needed,​​ is the node predicted by the model the probability of whether the node needs SMD;

[0199] S342, in order to minimize the SMD resource consumption, that is, to encourage the nodes classified into the "need SMD" category as few as possible, a resource consumption regularization term is introduced to limit the number of SMD allocation, as follows:

[0200]

[0201] wherein, is the regularization coefficient; this term encourages the network to predict as few nodes as possible to need SMD, thereby reducing resource consumption.

[0202] S343, combine equations (22) and (23) to form the total loss function, as follows:

[0203]

[0204] By using the Adam optimization algorithm to minimize the total loss function, the graph convolutional network GCN parameters to be trained are iterated. The GCN network can not only optimize the accuracy of node classification, but also achieve the goal of minimizing SMD resource consumption. In the training process, in addition to minimizing the classification loss, the value of the SMD resource consumption regularization term needs to be reasonably adjusted to ensure that the model can not only classify efficiently, but also control the number of SMDs.

[0205] S4, on the basis of optimizing the placement of real-time measurement devices SMD nodes, use the SMD data of real-time measurement devices as the input of the graph neural network GNN to make real-time network topology and parameter identification of the system. The specific implementation steps are:

[0206] S41, according to the installation node of S3 confirmed SMD that has been optimized, collect the data of the SMD nodes installed in the three-phase distribution network, including voltage phasor, current phasor, active power and reactive power, and construct a data set wherein, represents the number of nodes installed with SMD, represents the data of the th node, including , respectively, represents the data of the th node including voltage, current, active power and reactive power, and each parameter is represented as follows: ; after normalizing the data set , the normalized data set

[0207] wherein, represents the three-phase in the distribution network, Indicates the number of sampling points. They represent the first The voltage at the node in the three-phase configuration is at the [number]th [node]. Variables at each sampling point; They represent the first The current at the node in the three-phase configuration is at the [number]th node. Variables at each sampling point; They represent the first The active power of each node under three-phase conditions is at the [number]th [node]. Variables at each sampling point; They represent the first The reactive power of the node in three-phase operation is at the [number]th node. Variables at each sampling point; yes The row normalization result represents the normalized value of the node across all sampled data points;

[0208] S42, will Compared with the admittance matrix estimated in step S2 As input to the graph neural network (GNN) model, a transition function is established based on the power flow equation, thereby constructing the GNN model; in a three-phase distribution network, nodes... The state is determined by its own voltage and adjacent node voltage The update is performed using the power flow equation, specifically calculated as follows:

[0209]

[0210] in, For nodes Three-phase complex power injection; For nodes and adjacent nodes Three-phase complex voltage; For nodes The self-admittance matrix, For nodes and adjacent nodes The mutual admittance matrix between them and They are and The complex conjugate transpose; for The set of neighboring nodes of a node; and Calculated from line impedance; This indicates that the Hadamard product is multiplied element by element;

[0211] The power flow equation is converted into a real matrix representation as follows:

[0212]

[0213] where, and are the real and imaginary parts of the three-phase complex voltage , respectively; is the real representation of the nodal impedance matrix; is the real representation of the adjacent nodal admittance matrix; and are the complex conjugates of the complex power injection and the voltage , respectively, and are the real and imaginary parts of the complex conjugate product of the complex power injection and the complex voltage, respectively; in the graph neural network GNN, this physical constraint is used as the transition function for updating the voltage state at each layer; the global representation of the transition function is:

[0214]

[0215] where, the node states (voltage values) of the entire network. the characteristic information (resistance, reactance) of the line; the physical transition function that defines the state update;

[0216] The graph neural network GNN is used as the backbone network to learn the global topology structure of the power grid and perform parameter prediction; the node voltage is calculated iteratively using the transition function as follows:

[0217]

[0218] where, denotes the number of iteration steps;

[0219] The iterative calculation continues until the following convergence condition is met:

[0220]

[0221] where, is a small threshold value used to determine convergence; when using the graph neural network GNN model for line parameter identification, this task is treated as a regression problem because the voltage parameters of each node are to be predicted; the graph neural network GNN model is used to aggregate the local features of the nodes and the information of the neighboring nodes, and the fully connected network is used to regress the voltage parameters of each node. The fully connected layer is a branch network, and two layers of fully connected neural networks are connected after each branch; the first layer contains 256 nodes, and the second layer contains 128 nodes.

[0222] S43. In network training, the mean squared error (MSE) is used as the loss function, as shown in the following formula:

[0223]

[0224] in, Indicates time t The mean square error of each node. It is the actual voltage measured by SMD. It is the predicted voltage of the graph neural network (GNN); by calculating the gradient of the loss function with respect to the line parameters, the line parameters of the graph neural network (GNN) are iteratively optimized.

[0225] Define the current line parameters as follows The loss function is calculated using the backpropagation method. The gradient; for the first Line parameters of the line The gradient is calculated as follows:

[0226]

[0227] in, It is a set of time indexes. Let represent the sample loss function at all time points; then, stochastic gradient descent (SGD) is used to update the line parameters, as follows:

[0228]

[0229] in, It's the learning rate. It is the current iteration step number. In steps The next The line parameters are determined using the Stochastic Gradient Descent (SGD) method. This method minimizes the loss function and updates the line parameters. After each update, the model adapts better to the data provided by SGD, resulting in more accurate line parameters.

[0230] S5. Use real-time current data acquired by SMD to identify whether the distribution network topology has changed; if it has changed, fine-tune the parameters of the pre-trained graph neural network (GNN) using transfer learning to adapt to the new topology and re-identify the parameters. The specific implementation steps are as follows:

[0231] S51. Using real-time current data measured by SMD, determine whether the distribution network topology has changed; set the input data as follows: ,in For the first Current phasor data for each node;

[0232] In the case of topology occurrence, current flows between new branches, new branches input and output current Satisfy continuity:

[0233]

[0234] In the case of topology not occurring, no current flows in new branches:

[0235]

[0236] Then the newly connected branch The total amount of current change As follows:

[0237]

[0238] Wherein, Indicates The current of phase Flow into the new branch; Indicates The current of phase Flow out of the new branch; Indicates The greater value of the current of phase flowing into and out of the new branch; Indicates The numerical sum, that is, the sum of the maximum current of phase;

[0239] At the same time, in order to avoid the influence of noise on single measurement, the time window method is used for multi-time statistics as follows:

[0240]

[0241] As shown in formula (36), if the current sum of Continuous times is greater than the threshold value , it is judged that there is a topology change; if the current is 0 for a long time, it is judged that the topology has not changed;

[0242] In order to identify the topology connection relationship of the newly connected device, the current correlation Between nodes Is calculated as follows:

[0243]

[0244] Wherein, Indicates The current value of point And ​​Covariance of current values, as shown in equation (38); represents point variance of current values; represents point variance of current values; represents consecutive time average value of point current; represents consecutive time average value of point current;

[0245]

[0246]

[0247]

[0248] Set threshold , adjust the adjacency matrix to , as follows:

[0249]

[0250] If , it represents that nodes and are directly connected;

[0251] S52, after identifying the topology switching through the change of current data, the trained graph neural network GNN is fine-tuned through the transfer learning method, so that the model can adapt to the new topology structure, and the process of transfer learning is as follows :

[0252]

[0253] wherein, is the updated graph neural network GNN full connection layer weight parameter, is the full connection layer weight parameter of the original training model, is the parameter update amount obtained by fine-tuning;

[0254] S53, after identifying the topology change, the fine-tuned graph neural network GNN is used for parameter identification of the system, and the parameter identification process of the system is as follows:

[0255]

[0256] wherein, is the estimated voltage phasor, is the updated GNN network full connection layer weight parameter, for the adjusted topological state, The graph neural network GNN is used for modeling the power grid topology. The graph neural network GNN model outputs the voltage phasor estimation of each node of the system, ensuring high accuracy in real-time estimation after the topology changes.

[0257] Finally, the parameter identification is performed through the real-time SMD data, and the prediction accuracy of the model is evaluated by using the mean square error MSE and the absolute error MAE of the voltage phase angle, as follows:

[0258]

[0259] wherein, the real voltage phasor, the predicted voltage phasor. By comparing the performance of the fine-tuned model and the baseline model (such as LSE) under different topologies, the effectiveness of transfer learning is verified.

[0260] Embodiment 2

[0261] It should be further noted that based on the same inventive concept, the present embodiment also provides a real-time network topology and parameter identification system for a power distribution system. The system runs the method described in Embodiment 1, and includes the following modules:

[0262] The smart meter data construction module is used to collect data of nodes with smart meters installed in the historical power distribution network, and to construct a data set.

[0263] The line parameter estimation module is used to calculate the admittance matrix using the measurement history data of the smart meter, to derive the network topology structure, and to preliminarily estimate the line parameters.

[0264] The SMD placement optimization module is used to model the power grid topology based on the graph convolution network GCN, to take each power grid node as a node of the graph and the power grid line as an edge of the graph, and to iteratively optimize the graph convolution network GCN parameters through the total loss function, and to obtain the placement strategy of the measurement device SMD using the graph convolution network GCN.

[0265] The line parameter output module is used to identify the real-time network topology and parameters of the system based on the optimized placement of the measurement device SMD node, using the real-time measurement device SMD data as the input of the graph neural network GNN.

[0266] The model transfer learning module is used to identify whether the power distribution network topology state has changed using the real-time current data collected by the SMD. If the topology state has changed, the graph neural network GNN parameters that have been pre-trained are fine-tuned to adapt to the new topology structure, and the parameters are re-identified.

[0267] The above examples are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that the technical solutions recorded in the foregoing examples can be modified, or some technical features can be replaced by equivalent features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for real-time network topology and parameter identification of a power distribution system, characterized in that, include: S1. Collect data from smart meter nodes installed in historical distribution networks and construct a dataset; S2. Using the historical measurement data of smart meters, calculate the admittance matrix, deduce the network topology, and make a preliminary estimate of the line parameters. S3. Power grid topology modeling based on graph convolutional network (GCN): each power grid node is treated as a node in the graph, and power grid lines are treated as edges in the graph. The importance of each node is calculated through a message passing mechanism. Then, the parameters of the graph convolutional network (GCN) are iteratively optimized through the total loss function. The placement strategy of measurement device (SMD) is obtained using the graph convolutional network (GCN). S4. Based on optimizing the placement of nodes of the real-time measurement device (SMD), the SMD data is used as the input of the graph neural network (GNN) to identify the real-time network topology and parameters of the system. S5. Use real-time current data collected by SMD to identify whether the distribution network topology has changed; if it has changed, use transfer learning to fine-tune the parameters of the pre-trained graph neural network (GNN) to adapt to the new topology and re-identify the parameters.

2. The method for real-time network topology and parameter identification of a power distribution system according to claim 1, characterized in that, Specifically, S1 is: Collect node data for smart meter installation, including voltage amplitude, active power, and reactive power; use historical data of smart meter installation nodes in the distribution network as input features; Construct the node voltage magnitude matrix respectively , For nodes voltage amplitude, Total number of nodes; Active power injection matrix Reactive power injection matrix .

3. The method for real-time network topology and parameter identification of a power distribution system according to claim 1, characterized in that, S2 includes the following steps: S21. Construct the power flow model equations as follows: in, and These are nodes The injection of active and reactive power, and These are nodes and nodes voltage amplitude, and These are the conductance and susceptance in the admittance matrix, respectively. and These are nodes and nodes The voltage phase angle; S22. Simplify the power flow model equations into linear equations, as follows: S23. The conductance between each node is approximately calculated based on equations (6) and (7). and susceptance Thus, the admittance matrix is ​​obtained. Elements of the admittance matrix As shown in the following formula: in, For self-guided admittance, For mutual admittance; S24. Elements of the admittance matrix Sparsity processing is employed, and a threshold is set. To exclude unimportant connections, the specific exclusion method is as follows: in, For nodes and nodes Admittance between The threshold is expressed as follows: in, The nodes in the admittance matrix In phase The self-admittance at that node is the admittance value at that node. It is the number of nodes in the network. It is an empirical constant used to control the degree of sparsity; S25. Based on the topology inferred in S24, the conductance of the line parameters of the distribution network is analyzed using a linear regression method. and susceptance Make a preliminary estimate.

4. The method for real-time network topology and parameter identification of a power distribution system according to claim 3, characterized in that, The linear regression method described in S25 for analyzing the conductance of line parameters in a power distribution network and susceptance The specific steps for making an estimate are as follows: The objective function is defined using the least squares (OLS) method, as shown below: in, It is a node active power, These are elements of the admittance matrix. It is a node The voltage amplitude.

5. The method for real-time network topology and parameter identification of a power distribution system according to claim 2, characterized in that, S3 includes the following steps: S31. Based on the topology derived in S2, construct the adjacency matrix. , It is an n-order square matrix, in the following form: for In distribution networks, the element is represented as... The electrical connection relationships are as follows: Set Node Initial state feature vector: , Data from smart meters includes voltage and power; the node's initial state feature vector is then used. and adjacent order matrices In the input graph convolutional network (GCN), the graph convolutional network is composed of multiple stacked layers, and its node feature iterative calculation formula is as follows: in, , It is a degree matrix; It is the first Features of the layer It is the first The training weight parameters of the layer, It is the activation function ReLU. This represents the aggregation of information about nodes and their neighbors; S32, Node features after transformation using formula (15) and The LeakyReLU function is introduced to calculate the attention scores of neighboring nodes. As shown in the following formula: Among them, the weight matrix The parameters to be trained, Represents a node To the neighbors Attention coefficient These are the attention weight parameters to be trained; This indicates the concatenation of eigenvectors; and These are nodes and neighbors Features after linear transformation; for Node features; normalized attention weights are calculated using the Softmax function. As shown in the following formula: in, It is a node The set of neighbors; Candidate points representing measurement devices At the node The importance of location; S33. Scoring candidate points using a measuring device. Screening of candidate points for the measuring device; S34. Introduce the optimization objective of resource consumption during the training process and design a loss function.

6. The method for real-time network topology and parameter identification of a power distribution system according to claim 5, characterized in that, S33 includes the following steps: S331. An attention mechanism is used to aggregate neighborhood information to obtain new features of candidate points for the measurement device. As shown in the following formula: in, For nodes In the Characteristics of the layer; Neighbor Transformed features For the neighbors For nodes Contribution weight; S332, the result from formula (19) Calculate the candidate point scores for the measurement devices at all nodes. As shown in the following formula: S333, Select The K highest-scoring measurement points are selected as candidate points for the optimal measurement device. In the graph attention network (GAT), these points are then... The score is passed to a Classify by layers, The function converts the output of each node into a probability distribution of whether SMD is installed, as shown below: in, The node predicted by the model The probability of needing SMD.

7. The method for real-time network topology and parameter identification of a power distribution system according to claim 6, characterized in that, S34 includes the following steps: S341. The node classification loss function uses the binary cross-entropy loss function to calculate the difference between the predicted probability of whether each node needs SMD and the actual label, as shown in the following formula: in, It is a node The true label, The node predicted by the model The probability of needing SMD; S342. Introduce a resource consumption regularization term to limit the number of SMDs allocated, as shown in the following formula: in, It is the regularization coefficient; S343. Combining equations (22) and (23), we obtain the total loss function, as shown below: The parameters of the graph convolutional network (GCN) to be trained are iterated by minimizing this total loss function.

8. The method for real-time network topology and parameter identification of a power distribution system according to claim 3, characterized in that, S4 includes the following steps: S41. Based on the SMD installation nodes optimized and confirmed in S3, collect data on the SMD installation nodes in the three-phase distribution network, including voltage phasors, current phasors, active power, and reactive power, and construct a dataset. ,in, This indicates the number of nodes where SMD is installed. Indicates the first Data from each node, including , respectively representing the first The data for each node includes voltage, current, active power, and reactive power, and each parameter is represented as follows: ; For dataset After normalization, we obtain the normalized dataset. in, Indicates three phases in a power distribution network. Indicates the number of sampling points. They represent the first The voltage at the node in the three-phase configuration is at the [number]th [node]. Variables at each sampling point; They represent the first The current at the node in the three-phase configuration is at the [number]th node. Variables at each sampling point; They represent the first The active power of each node under three-phase conditions is at the [number]th [node]. Variables at each sampling point; They represent the first The reactive power of the node in three-phase operation is at the [number]th node. Variables at each sampling point; yes The row normalization result represents the normalized value of the node across all sampled data points; S42, will Compared with the admittance matrix estimated in step S2 As input to the graph neural network (GNN) model, a transition function is established based on the power flow equation, thereby constructing the GNN model; in a three-phase distribution network, nodes... The state is determined by its own voltage and adjacent nodes The update is performed using the power flow equation, specifically calculated as follows: in, For nodes Three-phase complex power injection; For nodes and adjacent nodes Three-phase complex voltage; For nodes The self-admittance matrix, For nodes and adjacent nodes The mutual admittance matrix between them and They are and The complex conjugate transpose; for The set of neighboring nodes of a node; and Calculated from line impedance; This indicates that the Hadamard product is multiplied element by element; The power flow equation is converted into a real matrix representation as follows: in, and They represent the three-phase complex voltages respectively. The real and imaginary parts; It is the real number representation of the nodal impedance matrix; It is the real number representation of the admittance matrix of adjacent nodes; and These are complex power injections. and voltage The complex conjugate, and These are the real and imaginary parts of the complex conjugate product of complex power injection and complex voltage, respectively; in graph neural networks (GNNs), this physical constraint is used as a transition function. This is used to update the voltage state at each level; the global representation of the transition function is: in, The node status of the entire network; Characteristic information of the line; Define the physical transition function for state updates; A graph neural network (GNN) is used as the backbone to learn the global topology of the power grid and predict parameters; node voltages are predicted using transition functions. The following formula is calculated iteratively: in, Indicates the number of iterations; Iterative calculations continue until the following convergence condition is met: in, It is a small threshold used to determine convergence; S43. In network training, the mean squared error (MSE) is used as the loss function, as shown in the following formula: in, Indicates time t The mean square error of each node. It is the actual voltage measured by SMD. It is the predicted voltage of the graph neural network (GNN); The graph neural network (GNN) optimizes the line parameters iteratively by calculating the gradient of the loss function with respect to the line parameters. Define the current line parameters as follows The loss function is calculated using the backpropagation method. The gradient; for the first Line parameters of the line The gradient is calculated as follows: in, It is a set of time indexes. Let represent the sample loss function at all time points; then, stochastic gradient descent (SGD) is used to update the line parameters, as follows: in, It's the learning rate. It is the current iteration step number. In steps The next The line parameters of the line.

9. The method for real-time network topology and parameter identification of a power distribution system according to claim 8, characterized in that, S5 includes the following steps: S51. Using real-time current data measured by SMD, determine whether the distribution network topology has changed; set the input data as follows: ,in For the first Current phasor data for each node; When topology occurs, current flows between the new branches, and the new branch input... With output current Satisfying continuity: If the topology has not changed, no current flows in the new branch: Then the new access branch Total change in current As shown in the following formula: in, express Mutually( The current flowing into the new branch; express Mutually( The current flowing out of the new branch; express The larger of the current flowing into and out of the new branch; Indicates to Numerical summation, that is, summing of... Sum of the maximum currents in each phase; To avoid the influence of noise on a single measurement, a time window method is used to perform multi-time statistics as follows: As shown in formula (36), if continuous At that moment The sum of currents is greater than the threshold. If the current is zero for an extended period, it is determined that a topological change has occurred; if the current remains zero for an extended period, it is determined that the topology has not changed. To identify the topological connections of newly connected devices, compute nodes and Current correlation between As shown in the following formula: in, express point Current value and The covariance of the current value is shown in formula (38); express point Current variance; express point Current variance; Indicates consecutive moments The average value of the point current; Indicates consecutive moments The average value of the point current; Set threshold Adjust the adjacency matrix for As shown in the following formula: like , then it represents a node and Directly connected; S52. After identifying the topology switch through changes in current data, the trained Graph Neural Network (GNN) model is fine-tuned using transfer learning to adapt to the new topology. The transfer learning process is as follows: : in, For the updated weight parameters of the fully connected layers in the graph neural network (GNN), These are the weight parameters of the fully connected layers in the original trained model. This refers to the parameter update amount obtained through fine-tuning; S53. After identifying the topological changes, the system parameters are identified using a fine-tuned graph neural network (GNN). The parameter identification process is as follows: in, For the estimated voltage phasor, For the updated GNN network fully connected layer weight parameters, For the adjusted topology state, This is a graph neural network (GNN).

10. A real-time network topology and parameter identification system for a power distribution system, characterized in that, Includes the following modules: The smart meter data construction module is used to collect data from smart meter nodes installed in the historical distribution network and construct a dataset. The line parameter estimation module is used to calculate the admittance matrix, deduce the network topology, and make preliminary estimates of the line parameters using the historical measurement data of smart meters. The SMD placement optimization module is used to model the power grid topology based on the graph convolutional network (GCN). Each power grid node is treated as a node in the graph, and the power grid lines are treated as edges in the graph. The importance of each node is calculated through a message passing mechanism, and the parameters of the graph convolutional network (GCN) are iteratively optimized through the total loss function. The placement strategy of the measurement device (SMD) is obtained using the graph convolutional network (GCN). The line parameter output module is used to identify the real-time network topology and parameters of the system by using the real-time measurement device (SMD) data as the input of the graph neural network (GNN) based on the optimized placement of nodes of the real-time measurement device (SMD). The model transfer learning module is used to identify whether the distribution network topology has changed using real-time current data collected by SMD. If it has changed, the parameters of the pre-trained graph neural network (GNN) are fine-tuned to adapt to the new topology using transfer learning methods, and the parameters are re-identified.

Images