A power distribution network topology identification and state estimation method based on message passing neural network and online deep learning

Abstract

The application discloses a power distribution network topology identification and state estimation method based on a message passing neural network and online deep learning. The method firstly trains a neural network by using historical measurement data to establish a power distribution network topology identification model, then combines the obtained topology information with corresponding power grid measurement data to construct a graph data set, trains a message passing neural network by taking the obtained graph data set as a training set, and obtains a power distribution network state estimation model. Finally, real-time measurement data is input into the state estimation model of the topology identification model to obtain real-time state estimation results under the section. The application can accurately identify the power distribution network topology under limited measurement data and provide accurate state estimation for the power distribution system, thereby providing technical support for subsequent safety evaluation and optimal operation.

Description

Technical Field

[0001] This invention proposes a method for distribution network topology identification and state estimation based on message-passing neural networks and online deep learning, belonging to the field of AC power systems. Background Technology

[0002] With the increasing penetration of distributed renewable energy (DRE) in power systems, distribution networks face new challenges. Traditional distribution systems are based on unidirectional power flow, but with the integration of DRE, power flow becomes bidirectional, leading to frequent changes in network topology. To ensure the safe, stable, and economical operation of the distribution network, real-time and accurate acquisition of the distribution system's state information and topology has become crucial. However, existing distribution network state estimation techniques face numerous limitations due to limited measurement data, strong reliance on model-driven approaches, and inability to adapt to topology changes.

[0003] To address these issues, researchers have gradually shifted towards dynamic state estimation techniques and data-driven methods. Dynamic state estimation methods, such as the Extended Kalman Filter (EKF) and the Unscented Kalman Filter (UKF), can capture the temporal dynamic characteristics of distribution networks, but still rely on accurate physical models and sufficient measurement data. In contrast, data-driven methods utilize historical measurement data for training, constructing a mapping relationship between power grid measurement data and state variables. However, these methods often ignore topology changes or require large amounts of measurement data to ensure estimation accuracy. Summary of the Invention

[0004] The purpose of this invention is to address the above problems by providing a method for topology identification and state estimation of distribution networks based on message-passing neural networks and online deep learning, which can accurately obtain real-time state information in distribution networks.

[0005] Technical Solution: To achieve the aforementioned objectives, this invention proposes a method for distribution network topology identification and state estimation based on message-passing neural networks and online deep learning, comprising the following steps:

[0006] Step 1: Establish a distribution network topology identification model. Use smart meters or SCADA systems in the distribution network to collect voltage amplitude and power flow data in real time. Input the real-time collected measurement data into an online neural network model. After training with historical data, the model identifies the current topology based on the input voltage and power data. Introduce a hedging backpropagation mechanism into the neural network. By weighted summarizing the output results of each layer, the network depth is dynamically adjusted to ensure convergence when facing different topology changes and to identify the accurate topology.

[0007] Step 2: Construct a graph dataset, representing nodes in the distribution network as nodes in the graph and lines in the distribution network as edges in the graph; construct a graph dataset using node voltage amplitude, power injection features, and line power as edge features and the topology information obtained from the topology identification model in Step 1, and perform data normalization processing. This dataset will be used as training data for the message passing neural network.

[0008] Step 3: Establish a message-passing neural network power distribution network state estimation model. Use the constructed graph dataset to train the message-passing neural network offline. During the training process, compare the state estimates of the system node voltage amplitude and branch phase angle difference output by the message-passing graph neural network model with the true values, calculate the loss function, update the weights and parameters of the message-passing graph neural network model through the backpropagation algorithm, and obtain a message-passing graph neural network model adapted to different topologies through multiple iterations.

[0009] Step 4: Use the message-passing neural network distribution network state estimation model for estimation, including: constructing graph data of the current section based on the normalized current section measurement information and distribution network topology information, and inputting the graph data of the current section into the message-passing graph neural network model trained offline in Step 3 to obtain the state estimates of the system node voltage amplitude and branch phase angle difference of the current section.

[0010] Furthermore, the output of the distribution network topology identification model established in step 1 is the on / off status of the lines, as detailed below:

[0011] (1.1) Switching circuit status

[0012]

[0013] In the formula, a i y represents the state of the i-th switch circuit, where 0 indicates the circuit is open and 1 indicates the circuit is closed; y represents the output of the online neural network, with λ1 and λ2 set to constant values; when the output of the neural network exceeds λ2, the circuit switch is considered closed; otherwise, the circuit switch is considered open; in this scenario, λ1 = λ2 = 0.5;

[0014] (1.2) Measurement Equation

[0015] The measurement equations for historical measurement information can be expressed in the following form:

[0016] zh(x)=e

[0017] In the formula, z represents the measurement vector, x represents the state variable vector, h(·) represents the AC or DC power flow equation, and e represents the measurement residual vector.

[0018] For AC systems, the measurements include node voltage magnitude measurement, node active power injection measurement, node reactive power injection measurement, branch active power measurement, and branch reactive power measurement. The power flow equations are as follows:

[0019]

[0020] θ ii' =θ i -θ i' {i,i'}∈N AC

[0021] In the formula, P AC,ii’ and Q AC,ii’ V represents the active and reactive power of the branch connected to the i-th AC node and the i'-th AC node, respectively. i and V i’ θ represents the voltage magnitudes of the i-th AC node and the i'-th AC node, respectively. ii’ θ represents the phase angle difference between the branches connected to the i-th and i'-th AC nodes. i and θ i’ P represents the phase angles of the i-th and i'-th communication nodes, respectively. AC,i and Q AC,i Let represent the active power injected and the reactive power injected into the i-th AC node, respectively. and These represent the measurement errors of active power and reactive power in the branches connected to the i-th and i'-th AC nodes, respectively. and N represents the measurement error of the active power injection and the measurement error of the reactive power injection at the i-th AC node, respectively. AC G represents the set of nodes in the communication system. ii’ and B ii’ These represent the real and imaginary parts of the admittance matrix of the branch connected to the i-th and i'-th AC nodes, respectively;

[0022] (1.3) The specific details of the hedging back propagation mechanism are as follows:

[0023]

[0024] h 0 =z=[V,P node Q node ,P line Q line ]

[0025] In the formula, L represents the total number of layers in the neural network, and w l and b l Let Θ be the weight matrix and bias vector of hidden layer l. l and α l Here are the weight matrix and weight coefficients for the l-th layer. h represents the output of the l-th layer. l h represents the activation value of the l-th layer. 0 This represents the initial activation values ​​of the model, including the voltage amplitude measurement V and the node active power injection measurement P. node Reactive power injection measurement Q node Active power and flow rate measurement of the line P line Reactive power flow measurement Q line F(z) represents the output of the neural network.

[0026] Furthermore, in step 2, the graph dataset is constructed by abstracting the electrical components of the power distribution network system into a graph structure, and the data is normalized. The specific process is as follows:

[0027] (2.1) Node definition and feature extraction

[0028] In a power distribution network, components such as power buses, load nodes, and distributed generation sources are defined as nodes in the diagram. The voltage amplitude and power state information of each node are used as feature vectors and input into the neural network. The specific extraction steps are as follows:

[0029] Each physical node of a busbar and load node in the distribution network corresponds to a node in the diagram;

[0030] Based on data obtained from SCADA systems or smart meters, the voltage amplitude, active power, and reactive power of each node are extracted, and the node features are represented as a vector:

[0031] x i =[V i ,P i Q i ]

[0032] V i ∈V

[0033] P i ∈P node

[0034] Q i ∈Q node

[0035] In the formula, x i V represents the characteristics of node i. i P represents the voltage amplitude measurement at node i. i Q i These represent the active and reactive power measurements at node i, respectively.

[0036] (2.2) Definition and Feature Extraction of Edges

[0037] e.g. ij =[P ij Q ij ]

[0038] P ij ∈P line

[0039] Q ij ∈Q line

[0040] In the formula, eg ij P represents the feature of edge ij. ij Q represents the active power measurement of side ij. ij This indicates the measurement of reactive power on side ij;

[0041] (2.3) Constructing a graph dataset

[0042] Graph data construction: Based on the network structure obtained from topology identification, a graph G(V,E) is generated, where V is the set of nodes, representing the bus and load nodes in the distribution network; and E is the set of edges, representing the power lines in the distribution network.

[0043] The node feature matrix is ​​formed by stacking the feature vectors of all nodes into a matrix X, with a size of |V|×d, where |V| is the number of nodes and d is the dimension of the node features.

[0044] The edge feature matrix is ​​formed by stacking the feature vectors of each edge into a matrix E, which has a size of |E|×g, where |E| is the number of edges and g is the feature dimension of each edge.

[0045] (2.4) Data normalization

[0046]

[0047] In the formula, u i ' represents the i-th normalized historical measurement, u i Let be the i-th original historical measurement value, min(·) be the minimum value in the set of data, and max(·) be the maximum value in the set of data.

[0048] Furthermore, in step 3, graph data is used for training, and historical measurement data of node voltage amplitude and branch power are read. The measurement data is collected by SCADA measurement device. The read historical measurements are input into message passing neural network model for training. The target output of the model is the voltage amplitude of the node and the branch phase angle difference.

[0049] The specific definition of a message-passing neural network is as follows:

[0050] For the i-th node, the following expression is given:

[0051]

[0052] In the formula, Let N(i) represent the information received by node i at time t+1; N(i) represent all neighboring nodes of node i. It is the state of node i at time t; e ij These are the edge features between nodes i and j; It is the node message passing function at time t; this formula indicates that the information received by node i comes from... and the state of adjacent node j at time t and the edge feature e connecting them ij Once information is collected, the node status is updated.

[0053]

[0054] in, For the node update function, this function uses and As input, generate the node state at time t+1. The node message passing function is as follows:

[0055]

[0056] in, It is a learnable parameter matrix; Let be the activation function of the fully connected layer at time t. Let be the edge characteristics of edge ij at time t;

[0057] Based on the node modeling described above, the edges in the graph are modeled using the following formula:

[0058]

[0059] in, This represents the information received by edge ij at time t+1. and These are the edge message passing function and state update function at time t. The edge message passing function represents the state of edge ij at time t+1.

[0060]

[0061] in, It is a learnable parameter matrix. Let be the activation function of the fully connected layer at time t;

[0062] The final output of the model is:

[0063]

[0064] In the formula, y represents the final output vector of the model, T represents the last time step of the message passing phase, R(·) is a fully connected layer whose parameters are determined through network training, and the state update function adopts a gated recurrent unit (GRU):

[0065]

[0066] Furthermore, in step 4, the target output of the model is the voltage magnitude of the node and the branch phase angle difference. The specific conversion formulas for the branch phase angle difference and the node phase angle are as follows:

[0067] θ=(B T B) -1 B T Δθ

[0068] In the formula, θ is the node voltage phase angle estimation vector, Δθ is the branch phase angle difference estimation vector, and B is the node-branch correlation matrix.

[0069] Beneficial effects: Compared with the prior art, the technical solution of the present invention has the following beneficial technical effects:

[0070] This invention introduces an online neural network (ONN) to quickly identify the real-time topology of a distribution network, making it particularly suitable for distribution network scenarios with frequent changes. Compared to traditional state estimation methods that rely on a fixed topology, this invention does not depend on a static topology and can dynamically adjust when the topology changes, improving the system's adaptability and flexibility.

[0071] Improving state estimation accuracy: State estimation is performed using a message-passing neural network (MPNN), which models the node and line information of the distribution network as a graph structure, making full use of the physical topology characteristics of the system. MPNN can effectively integrate node and edge information and perform multiple rounds of iterative updates, significantly improving the accuracy of distribution network state estimation, especially performing well in the case of sparse measurement data.

[0072] Addressing the problem of insufficient measurement data: Distribution networks typically have a limited number of measurement points, leading to significant observability challenges in state estimation. This invention combines pseudo-measurement data with actual measurement data, employing a deep learning model for topology identification and state estimation. This reduces reliance on highly redundant measurement data and effectively solves the problem of insufficient measurement data.

[0073] Improved computational efficiency: The MPNN in this invention combines the advantages of graph neural networks, significantly improving the computational efficiency of state estimation in large-scale power distribution networks by parallel computing of information transmission between nodes and edges. Compared with traditional weighted least squares (WLS) and mixed integer linear programming (MILP) methods, this invention can significantly reduce computational complexity and shorten computation time, making it suitable for application in practical large-scale power distribution systems.

[0074] Adaptive adjustment: This invention introduces the HBP (Hedge Backpropagation) mechanism through neural networks, which can adaptively adjust the layer depth and weight parameters of the neural network, and quickly adjust the model for different topologies to ensure the accuracy and real-time performance of state estimation.

[0075] To address the aforementioned problems, this invention proposes a joint topology identification and state estimation method based on a message passing neural network (MPNN) and online neural networks (ONN). This method can quickly identify the real-time topology of a distribution network and construct a graph dataset using measurement data, then perform efficient state estimation via MPNN. This invention significantly improves the accuracy and efficiency of distribution network state estimation even under conditions of frequent topology changes and insufficient measurement data. Attached Figure Description

[0076] Figure 1 This is a flowchart of the method for realizing distribution network topology identification and state estimation based on message passing neural network and online deep learning as described in the invention;

[0077] Figure 2 This is a diagram of a 33-node three-phase balance test system;

[0078] Figure 3 This is a diagram of a 123-node three-phase imbalance test system;

[0079] Figure 4 This is a graph of the loss function of the topology recognition model on the training set.

[0080] Figure 5 This is a distribution diagram of the weight coefficients of each layer of the topology recognition model after training and adjustment.

[0081] Figure 6The figure shows a comparison of the state estimation results of three state estimation methods on the continuous time section topology 3 of the 123-node system. (a) Estimation error of node voltage amplitude, (b) Estimation error of node voltage phase angle. Detailed Implementation

[0082] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0083] like Figure 1 As shown, this invention proposes a method for distribution network topology identification and state estimation based on message-passing neural networks and online deep learning, including the following steps:

[0084] Step 1: Establish a distribution network topology identification model. Use smart meters or SCADA systems in the distribution network to collect voltage amplitude and power flow data in real time. Input the real-time collected measurement data into an online neural network model. After training with historical data, the model identifies the current topology based on the input voltage and power data. Introduce a hedging backpropagation mechanism into the neural network. By weighted summarizing the output results of each layer, the network depth is dynamically adjusted to ensure convergence when facing different topology changes and to identify the accurate topology.

[0085] Step 2: Construct a graph dataset, representing nodes in the distribution network as nodes in the graph and lines in the distribution network as edges in the graph; construct a graph dataset using node voltage amplitude, power injection features, and line power as edge features and the topology information obtained from the topology identification model in Step 1, and perform data normalization processing. This dataset will be used as training data for the message passing neural network.

[0086] Step 3: Establish a message-passing neural network power distribution network state estimation model. Use the constructed graph dataset to train the message-passing neural network offline. During the training process, compare the state estimates of the system node voltage amplitude and branch phase angle difference output by the message-passing graph neural network model with the true values, calculate the loss function, update the weights and parameters of the message-passing graph neural network model through the backpropagation algorithm, and obtain a message-passing graph neural network model adapted to different topologies through multiple iterations.

[0087] Step 4: Use the message-passing neural network distribution network state estimation model for estimation, including: constructing graph data of the current section based on the normalized current section measurement information and distribution network topology information, and inputting the graph data of the current section into the message-passing graph neural network model trained offline in Step 3 to obtain the state estimates of the system node voltage amplitude and branch phase angle difference of the current section.

[0088] Furthermore, the output of the distribution network topology identification model established in step 1 is the on / off status of the lines, as detailed below:

[0089] (1.1) Switching circuit status

[0090]

[0091] In the formula, a i y represents the state of the i-th switch circuit, where 0 indicates the circuit is open and 1 indicates the circuit is closed; y represents the output of the online neural network, with λ1 and λ2 set to constant values; when the output of the neural network exceeds λ2, the circuit switch is considered closed; otherwise, the circuit switch is considered open; in this scenario, λ1 = λ2 = 0.5;

[0092] (1.2) Measurement Equation

[0093] The measurement equations for historical measurement information can be expressed in the following form:

[0094] zh(x)=e

[0095] In the formula, z represents the measurement vector, x represents the state variable vector, h(·) represents the AC or DC power flow equation, and e represents the measurement residual vector.

[0096] For AC systems, the measurements include node voltage magnitude measurement, node active power injection measurement, node reactive power injection measurement, branch active power measurement, and branch reactive power measurement. The power flow equations are as follows:

[0097]

[0098] θ ii' =θ i -θ i' {i,i'}∈N AC

[0099] In the formula, P AC,ii’ and Q AC,ii’ V represents the active and reactive power of the branch connected to the i-th AC node and the i'-th AC node, respectively. i and V i’ θ represents the voltage magnitudes of the i-th AC node and the i'-th AC node, respectively. ii’ θ represents the phase angle difference between the branches connected to the i-th and i'-th AC nodes. i and θ i’ P represents the phase angles of the i-th and i'-th communication nodes, respectively. AC,i and Q AC,i Let represent the active power injected and the reactive power injected into the i-th AC node, respectively. and These represent the measurement errors of active power and reactive power in the branches connected to the i-th and i'-th AC nodes, respectively. and N represents the measurement error of the active power injection and the measurement error of the reactive power injection at the i-th AC node, respectively. AC G represents the set of nodes in the communication system. ii’ and B ii’ These represent the real and imaginary parts of the admittance matrix of the branch connected to the i-th and i'-th AC nodes, respectively;

[0100] (1.3) The specific details of the hedging back propagation mechanism are as follows:

[0101]

[0102] h 0 =z=[V,P node Q node ,P line Q line ]

[0103] In the formula, L represents the total number of layers in the neural network, and w l and b l Let Θ be the weight matrix and bias vector of hidden layer l. l and α l Here are the weight matrix and weight coefficients for the l-th layer. h represents the output of the l-th layer. l h represents the activation value of the l-th layer. 0 This represents the initial activation values ​​of the model, including the voltage amplitude measurement V and the node active power injection measurement P. node Reactive power injection measurement Q node Active power and flow rate measurement of the line P line Reactive power flow measurement Q line F(z) represents the output of the neural network.

[0104] Furthermore, in step 2, the graph dataset is constructed by abstracting the electrical components of the power distribution network system into a graph structure, and the data is normalized. The specific process is as follows:

[0105] (2.1) Node definition and feature extraction

[0106] In a power distribution network, components such as power buses, load nodes, and distributed generation sources are defined as nodes in the diagram. The voltage amplitude and power state information of each node are used as feature vectors and input into the neural network. The specific extraction steps are as follows:

[0107] Each physical node of a busbar and load node in the distribution network corresponds to a node in the diagram;

[0108] Based on data obtained from SCADA systems or smart meters, the voltage amplitude, active power, and reactive power of each node are extracted, and the node features are represented as a vector:

[0109] x i =[V i ,P i Q i ]

[0110] V i ∈V

[0111] P i ∈P node

[0112] Q i ∈Q node

[0113] In the formula, x i V represents the characteristics of node i. i P represents the voltage amplitude measurement at node i. i Q i These represent the active and reactive power measurements at node i, respectively.

[0114] (2.2) Definition and Feature Extraction of Edges

[0115] e.g. ij =[P ij Q ij ]

[0116] P ij ∈P line

[0117] Q ij ∈Q line

[0118] In the formula, eg ij P represents the feature of edge ij. ij Q represents the active power measurement of side ij. ij This indicates the measurement of reactive power on side ij;

[0119] (2.3) Constructing a graph dataset

[0120] Graph data construction: Based on the network structure obtained from topology identification, a graph G(V,E) is generated, where V is the set of nodes, representing the bus and load nodes in the distribution network; and E is the set of edges, representing the power lines in the distribution network.

[0121] The node feature matrix is ​​formed by stacking the feature vectors of all nodes into a matrix X, with a size of |V|×d, where |V| is the number of nodes and d is the dimension of the node features.

[0122] The edge feature matrix is ​​formed by stacking the feature vectors of each edge into a matrix E, which has a size of |E|×g, where |E| is the number of edges and g is the feature dimension of each edge.

[0123] (2.4) Data normalization

[0124]

[0125] In the formula, u i ' represents the i-th normalized historical measurement, u i Let be the i-th original historical measurement value, min(·) be the minimum value in the set of data, and max(·) be the maximum value in the set of data.

[0126] Furthermore, in step 3, graph data is used for training, and historical measurement data of node voltage amplitude and branch power are read. The measurement data is collected by SCADA measurement device. The read historical measurements are input into message passing neural network model for training. The target output of the model is the voltage amplitude of the node and the branch phase angle difference.

[0127] The specific definition of a message-passing neural network is as follows:

[0128] For the i-th node, the following expression is given:

[0129]

[0130] In the formula, Let N(i) represent the information received by node i at time t+1; N(i) represent all neighboring nodes of node i. It is the state of node i at time t; e ij These are the edge features between nodes i and j; It is the node message passing function at time t; this formula indicates that the information received by node i comes from... and the state of adjacent node j at time t and the edge feature e connecting them ij Once information is collected, the node status is updated.

[0131]

[0132] in, For the node update function, this function uses and As input, generate the node state at time t+1. The node message passing function is as follows:

[0133]

[0134] in, It is a learnable parameter matrix; Let be the activation function of the fully connected layer at time t. Let be the edge characteristics of edge ij at time t;

[0135] Based on the node modeling described above, the edges in the graph are modeled using the following formula:

[0136]

[0137] in, This represents the information received by edge ij at time t+1. and These are the edge message passing function and state update function at time t. The edge message passing function represents the state of edge ij at time t+1.

[0138]

[0139] in, It is a learnable parameter matrix. Let be the activation function of the fully connected layer at time t;

[0140] The final output of the model is:

[0141]

[0142] In the formula, y represents the final output vector of the model, T represents the last time step of the message passing phase, R(·) is a fully connected layer whose parameters are determined through network training, and the state update function adopts a gated recurrent unit (GRU):

[0143]

[0144] Furthermore, in step 4, the target output of the model is the voltage magnitude of the node and the branch phase angle difference. The specific conversion formulas for the branch phase angle difference and the node phase angle are as follows:

[0145] θ=(B T B) -1 B T Δθ

[0146] In the formula, θ is the node voltage phase angle estimation vector, Δθ is the branch phase angle difference estimation vector, and B is the node-branch correlation matrix.

[0147] Case Analysis

[0148] 1) Example Explanation

[0149] The testing system of this invention is a 33-node three-phase balanced distribution network and a 123-node three-phase unbalanced distribution network. The system structure is as follows: Figure 2 and Figure 3 As shown. In the 33-node system, nodes 18, 22, 25, and 33 are connected to distributed photovoltaic (PV) systems, while in the 123-node system, nodes 152, 197, 49, and 250 are connected to distributed PV systems. The topology and measurement configurations of the two test systems are as follows:

[0150] Table 1 Topology configurations of different test systems

[0151]

[0152] Table 2 Measurement configurations for different test systems

[0153]

[0154] 2) Distribution network topology identification model

[0155] First, a distribution network topology identification model is trained based on historical data. Figure 4 The loss function curves of two test systems during the topology model training process are presented. Topologies 1-4 were used for training with known historical topologies, while the remaining topologies represent new topologies, demonstrating the model's recognition ability after dynamic adjustments during training. It can be seen that the loss of both models gradually decreases as training progresses. The enhanced 33 bus system converges and stabilizes after 30 iterations, while the 123 bus system converges after 20 iterations. Tests show that after training with topo1-4, ONN can not only accurately recognize the trained topologies but also demonstrates a certain ability to recognize new topologies: the recognition accuracy for topo6 and topo7 reaches 100%, and the recognition accuracy for topo8 reaches 99%.

[0156] When the topology changes, the proposed trained model can dynamically adjust the weight parameters of each layer, enabling the model to quickly adapt to the new topology data. The model is configured with 5 layers, with the weights of each layer initialized to 0.2. It can be seen that as training progresses, the model adjusts the weights of each layer, eventually stabilizing at... Figure 5 The values ​​shown are relevant. By adjusting the model's parameters, the model can quickly adapt to new topologies and achieve accurate topology recognition even on untrained topologies.

[0157] 3) Comparison of the algorithm of this invention with other methods

[0158] There are two main comparison algorithms: WLS and deep neural network (DNN). Both comparison algorithms and the method proposed in this invention are used at each time segment. The accuracy is estimated using the mean absolute error (AAE).

[0159]

[0160] In the formula: n is the number of state variables; x real The truth value is calculated to represent the power flow. This represents the predicted output value of the model.

[0161] As shown in Table 3, in the 33-node system, the proposed MPNN method exhibits higher estimation accuracy compared to the WLS and DNN algorithms. Across different topologies, the MPNN consistently demonstrates low error values, indicating its effectiveness and accuracy in this system. Compared to the traditional WLS method, the proposed MPNN method shows slightly lower estimation accuracy for voltage amplitude in the IEEE 123 bus system, but still meets the required accuracy (within 1.8e-3). However, for voltage phase angle estimation, the MPNN method shows slightly higher accuracy than the WLS method. Figure 6 The estimation results for 50 consecutive time intervals of topo 3 in the IEEE 123 bus system are given.

[0162] Table 3 Comparison of state estimation errors of different methods under different test systems

[0163]

[0164] 4) Computational efficiency

[0165] Distribution network nodes are numerous and widely distributed, making computational efficiency and time one of the main metrics for ensuring the practical application of algorithms in engineering. As shown in Table 4, in a 33-node system, the estimation time of the WLS algorithm is 0.0824 seconds, while the estimation time of the Proposed algorithm is only 0.0249 seconds, representing an efficiency improvement of 81.92%. For a 123-node system, the computation time of WLS significantly increases to 0.6873 seconds, while the proposed method only requires 0.0728 seconds, an improvement of 91.59%. These results demonstrate that as the system scale increases, the proposed method has a significant advantage in estimation efficiency, especially in large-scale systems where the time growth of the proposed algorithm is much shorter than that of the WLS algorithm.

[0166] Table 4 Computation time for different state estimation methods

[0167]

Claims

1. A power distribution network topology identification and state estimation method based on message passing neural network and online deep learning, characterized in that, The method comprises the following steps: Step 1, establishing a power distribution network topology identification model, and collecting voltage amplitude and power flow data of the power distribution network in real time by using smart meters or SCADA systems in the power distribution network; The measured data collected in real time are input into the online neural network model, the model is trained based on historical data, and the current topology structure is identified according to the input voltage and power data; the hedging backpropagation mechanism is introduced in the neural network, the output results of each layer are weighted and summarized, the network depth is dynamically adjusted, the convergence is ensured when facing different topology changes, and the accurate topology structure is identified; Step 2, constructing a graph data set, representing the nodes in the power distribution network as nodes in the graph, and representing the lines in the power distribution network as edges in the graph; using the node voltage amplitude, power injection characteristics and line power as edge features and the topology information obtained in step 1, the graph data set is constructed, and data normalization processing is performed, which will be used as training data for the message passing neural network; Step 3, establishing a message passing neural network power distribution network state estimation model, using the constructed graph data set to train the message passing neural network offline, comparing the state estimation values of the system node voltage amplitude and branch phase angle difference output by the message passing graph neural network model with the true values during the training process, calculating the loss function, updating the weights and parameters of the message passing graph neural network model through the backpropagation algorithm, and obtaining the message passing graph neural network model suitable for different topologies through multiple iterations; Step 4, using the message passing neural network power distribution network state estimation model for estimation, including: constructing the graph data of the current section according to the normalized current section measurement information and power distribution network topology information, and inputting the graph data of the current section into the message passing graph neural network model trained offline in step 3 to obtain the state estimation values of the system node voltage amplitude and branch phase angle difference of the current section; The output of the topology identification model established in step 1 is the open state of the line, which is as follows: (1.1) Switch line state wherein a i represents the state of the ith switch line, 0 indicating that the line is open, and 1 indicating that the line is closed; y represents the output of the online neural network, and λ1 and λ2 are set as constant values; when the output of the neural network exceeds λ2, it is considered that the line switch is closed; otherwise, it is considered that the line switch is opened; in this scenario, λ1=λ2=0.5; (1.2) Measurement equation The measurement equation of the historical measurement information is as follows: z-h(x)=e In the formula, z represents the measurement vector, x represents the state variable vector, h(·) represents the power flow equation of alternating current or direct current, and e represents the measurement residual error vector; For an alternating current system, the measurement includes node voltage amplitude measurement, node active power injection measurement, node reactive power injection measurement, branch active power measurement and branch reactive power measurement, and the power flow equation is as follows: θ ii' = θ i - θ i' {i,i'}∈N AC In the formula, P AC,ii’ and Q AC,ii ' represents the active and reactive power of the branch connected to the i-th AC node and the i'-th AC node, respectively, V i and V i ' represents the voltage magnitude of the i-th AC node and the i'-th AC node, respectively, θ ii ' represents the phase angle difference between the branches connecting the i-th and i'-th AC nodes, θ i and θ i ' represents the phase angle of the i-th communication node and the i'-th communication node, respectively, P AC i and Q AC,i Let represent the active power injected and the reactive power injected into the i-th AC node, respectively. and These represent the measurement errors of active power and reactive power in the branches connected to the i-th and i'-th AC nodes, respectively. and N represents the measurement error of the active power injection and the measurement error of the reactive power injection at the i-th AC node, respectively. AC G represents the set of nodes in the communication system. ii 'and B ii ' represents the real and imaginary parts of the admittance matrix of the branch connected to the i-th AC node and the i'-th AC node, respectively; (1.3) The hedging backpropagation mechanism is as follows h 0 = z = [V, P node , Q node , P line , Q line ] where L represents the total number of layers of the neural network, w l and b l are the weight matrix and bias vector of the hidden layer l, Θ l and a l are the weight matrix and weight coefficient of the lth layer, represents the output of the lth layer, h l represents the activation value of the lth layer, h 0 represents the initial activation value of the model, including the voltage amplitude measurement V, the node active power injection measurement P node and the reactive power injection measurement Q node , the line active power flow measurement P line and the reactive power flow measurement Q line , and F(z) represents the output of the neural network.

2. The power distribution network topology identification and state estimation method based on message passing neural network and online deep learning according to claim 1, characterized in that, In step 2, the graph data set is constructed by abstracting the electrical elements of the power distribution system into a graph structure, and the data is normalized, and the specific process is as follows: (2.1) Definition and feature extraction of nodes In the distribution network, power bus, load node and distributed power and other elements are defined as nodes in the graph, the voltage amplitude, power state information of each node as the feature vector of the node input to the neural network, the specific extraction steps are as follows: Each bus, load node in the distribution network is corresponding to a node in the graph; According to the data obtained from SCADA system or smart meter, the voltage amplitude, active power and reactive power of each node are extracted, and the node feature is represented as a vector: x i = [V i , P i , Q i ] V i ∈V P i ∈P node Q i ∈Q node where x i represents the feature of node i, V i represents the voltage magnitude measurement of node i, P i , Q i represent the active and reactive power measurements of node i, respectively; (2.2) Definition of edge and feature extraction eg ij = [P ij , Q ij ] P ij ∈P line Q ij ∈Q line where eg ij denotes the characteristic of edge ij, P ij denotes the active power measurement of edge ij, Q ij denotes the reactive power measurement of edge ij; (2.3) Construction of graph data set Graph data construction, based on the network structure obtained by topology identification, generate graph G(V, E), where V is the node set, representing the bus and load node in the distribution network; E is the edge set, representing the power line in the distribution network; Node feature matrix, stack all node feature vectors into a matrix X, whose size is |V|xd, where |V| is the number of nodes, and d is the dimension of node feature; Edge feature matrix, stack each edge feature vector into a matrix E, whose size is |E|xg, where |E| is the number of edges, and g is the feature dimension of each edge; (2.4) Data normalization where u i is the ith normalized historical measurement value, u i is the ith raw historical measurement value, min( ) is the minimum value in a set of data, and max( ) is the maximum value in a set of data.

3. The power distribution network topology identification and state estimation method based on message passing neural network and online deep learning according to claim 2, characterized in that, In step 3, the graph data is used for training, the historical measurement data of node voltage amplitude and branch power are read, and the measurement data is collected by SCADA measurement device; the read historical measurement is input into the message passing neural network model for training, and the target output of the model is the voltage amplitude and branch phase angle difference of the node; The message passing neural network is defined as follows: For the ith node, the following expression is given: where, denotes the information received by node i at time t+1 ; N(i) denotes all the neighboring nodes of node i; is the state of node i at time t; e ij is the edge feature between node i and node j; is the node message passing function at time t; this equation indicates that the information received by node i comes from and the state of neighboring node j at time t and the edge feature e ij connecting them; once the information is collected, the node state is updated; where, is the node update function that takes and as input and generates the node state at time t+1 The node message passing function is as follows: wherein, is a learnable parameter matrix; is an activation function of the node fully connected layer at time t, is an edge feature of edge ij at time t; On the basis of the above node modeling, the edge in the graph is modeled, and the specific formula is as follows: where, denotes the information received by edge ij at time t + 1, and is the edge message passing function and state update function at time t, denotes the state of edge ij at time t + 1, the edge message passing function is as follows: wherein, is a learnable parameter matrix, is an activation function of the edge fully connected layer at time t. The final output of the model is: In the formula, y represents the final output vector of the model, T represents the last time step of the message passing stage, R(·) is the full connection layer, and its parameters are determined by network training, and the state update function adopts the gated recurrent unit GRU:

4. The power distribution network topology identification and state estimation method based on message passing neural network and online deep learning according to claim 3, characterized in that, In step 4, the target output of the model is the voltage amplitude and branch phase angle difference of the node, and the specific conversion formula of the branch phase angle difference and the node phase angle is as follows: θ = (B T B)- 1 B T Δθ In the formula, θ is the node voltage phase angle estimation vector, Δθ is the branch phase angle difference estimation vector, and B is the node-branch association matrix.

Images