Power distribution network state estimation method based on dynamic space-time diagram neural network

Abstract

The invention belongs to the technical field of power system analysis and control, and particularly relates to a power distribution network state estimation method based on a dynamic space-time diagram neural network, a power distribution network is modeled into a graph structure G = (V, E), and a model core is composed of three parts: a dynamic graph convolution module, a time sequence Transform module and a gating fusion module; wherein the dynamic graph convolution module is used for spatial feature extraction; the time sequence Transform module is used for time feature extraction; and the gating fusion module is used for adaptively fusing the spatial features and the time features and finally outputting state estimation values of all nodes. By performing simulation verification on the power distribution network containing high-proportion distributed energy, the method can realize state estimation with higher precision and stronger robustness compared with a traditional WLS and an existing machine learning method under extremely low measurement configuration.

Description

Technical Field

[0001] This invention belongs to the field of power system analysis and control technology, specifically relating to a distribution network state estimation method based on dynamic spatiotemporal graph neural network, which is particularly suitable for active distribution networks with a high proportion of distributed renewable energy access. Background Technology

[0002] State estimation is one of the core functions of the power system energy management system (EMS). It is used to extrapolate the complete and accurate operating state of the system (such as node voltage amplitude and phase angle) based on limited, noisy measurement data (such as voltage, current and power measurements). It is the basis for power grid monitoring, safety assessment and optimization control.

[0003] Traditional distribution network state estimation methods are mainly based on weighted least squares (WLS). While the WLS algorithm is mature, it faces significant challenges in distribution networks: 1) Low measurement redundancy: Incomplete coverage of distribution network measurement devices (PMU, smart meters) leads to underdetermined or ill-conditioned measurement equations; 2) Strong nonlinearity: Distribution networks are radial, resulting in high R / X ratios and stronger model nonlinearity; 3) Sensitivity to poor data: The estimation accuracy of traditional WLS drops sharply when measurement errors or communication failures occur.

[0004] In recent years, machine learning methods have been introduced into state estimation. Some studies have used deep neural networks (DNNs) or long short-term memory networks (LSTMs) to directly map measurement data to system state. While these methods can capture nonlinearities, they severely neglect the inherent topological connections of the distribution network, treating it as a "black box," resulting in poor model interpretability and complete failure when the topology changes.

[0005] Graph neural network (GNN)-based methods model distribution networks as graph structures, effectively utilizing topological information. However, existing GNN methods often employ static adjacency matrices, failing to reflect the bidirectional and drastic power flow fluctuations caused by the integration of intermittent energy sources such as distributed photovoltaic (PV) and wind power. Furthermore, existing methods do not fully leverage the temporal and topological spatial correlations of measurement data, resulting in insufficient estimation accuracy when measurements are missing or noise levels are high. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a power distribution network state estimation method based on a dynamic spatiotemporal graph neural network. This invention aims to solve the problem of poor estimation accuracy of traditional methods in scenarios with insufficient measurement, noise interference, and dynamic changes in topology / operating state.

[0007] This invention is achieved through the following technical solution: a power distribution network state estimation method based on a dynamic spatiotemporal graph neural network, comprising the following steps:

[0008] The distribution network is modeled as a graph structure G=(V, E), where nodes V represent buses and edges E represent lines. The node feature vectors include real-time measurements or pseudo-measurements. The core of the model consists of three parts: a dynamic graph convolution module, a temporal Transformer module, and a gated fusion module.

[0009] The dynamic graph convolution module is used for spatial feature extraction; the temporal Transformer module is used for temporal feature extraction; and the gated fusion module is used to adaptively fuse spatial and temporal features, ultimately outputting the state estimates of all nodes.

[0010] As a further improvement to the technical solution of this invention, a learnable dynamic weight matrix D(t) is introduced into the dynamic graph convolution module to construct a dynamic adjacency matrix A, as follows:

[0011] In the formula: This represents the adjacency matrix adjusted at time t. It dynamically adjusts the weights based on real-time line power flow, reflecting the strength of electrical connections between nodes. This represents the static adjacency matrix, based on the physical topology of the distribution network. Its element values ​​indicate whether there is a connection between nodes, usually 0 or 1; ⊙ represents the Hadamard product, α represents the learnable scaling factor, and D(t) represents the learnable dynamic weight matrix, which is usually obtained by normalizing the estimated power flow amplitude of the line at the current moment, reflecting the real-time changes in the electrical coupling of the line;

[0012] Then, graph convolution is performed using A_dynamic(t):

[0013]

[0014] In the formula: Let D_hat represent the spatial features of the graph convolution output, and let A_dynamic(t) be the degree matrix. Let W_gcn represent the node feature matrix at time t, W_gcn be the trainable weight matrix, and σ be the Sigmoid activation function.

[0015] As a further improvement to the technical solution of this invention, the specific steps for the temporal feature extraction by the temporal Transformer module are as follows:

[0016] Input: For each node in the distribution network, extract its historical features over the past T consecutive time steps to form a time series sample X_i = [x_i(t-T+1), x_i(t-T+2), ..., x_i(t)], where x_i(t) represents the feature vector of the i-th node at time t. The sequences of all nodes together constitute the input tensor X_seq ∈ R^(N×T×F), where N represents the number of nodes, T represents the time step, and F represents the feature dimension.

[0017] Position encoding: Learnable position encoding is added to the input tensor, specifically:

[0018]

[0019] In the formula: Indicates enhanced temporal features. The input tensor PE∈R^(T×d_model) represents a trainable matrix, where R is the real number field, T is the time step, and d_model represents the model dimension of the Transformer.

[0020] Multi-head self-attention mechanism: enhanced temporal features The input is fed into a Transformer encoder consisting of L layers. The core of each layer is a multi-head self-attention mechanism, and its calculation process is as follows:

[0021] First, the input sequence is linearly mapped to a matrix of query Q, key K, and value V:

[0022]

[0023] In the formula: , , ∈R^(d_model×d_k) is the learnable weight matrix, where d_k is the dimension of each head;

[0024] Then, calculate the scaled dot product attention:

[0025]

[0026] In the formula: Q, K, and V are the query, key, and value matrices obtained from the input sequence through linear transformation, respectively. The dimension of the key vector is used to scale the dot product;

[0027] Multi-head attention executes this process in parallel h times, with each head using a different weight matrix, thereby capturing different aspects of the dependencies in the sequence:

[0028]

[0029]

[0030] In the formula: This represents the output of the h-th attention head. , , It is the learnable weight matrix of the h-th attention head. ∈R^(h·d_v×d_model) is the output projection matrix, and Concat represents the concatenation operation;

[0031] Feedforward networks and residual connections:

[0032] Each attention sublayer is followed by a feedforward neural network (FFN), which consists of two linear transformations and a ReLU activation function:

[0033]

[0034] In the formula: and It is a learnable weight matrix. , It is a bias term. This represents the ReLU activation function.

[0035] Each feedforward neural network employs residual connections and layer normalization to stabilize the training process of the deep network, and its output can be represented as:

[0036]

[0037]

[0038] In the formula: Let represent the output of the l-th attention sublayer. This represents the output of the feedforward neural network in the l-th attention sublayer. It is a layer normalization operation. This represents a multi-head attention mechanism.

[0039] Output: After stacking L layers of Transformer encoders, the output H_trans^(L) of the last attention sublayer is taken as the final extracted temporal feature representation H_trans.

[0040] As a further improvement to the technical solution of the present invention, the specific process by which the gated fusion module adaptively fuses spatial and temporal features is as follows:

[0041]

[0042]

[0043] In the formula: G is the fusion gate, and its value is between 0 and 1. It is the Sigmoid activation function, || represents the concatenation operation, W_gate and b_gate are the gating parameters, and the final output H_fused is the state estimate of all nodes.

[0044] As a further improvement to the technical solution of the present invention, the real-time measurement or pseudo-measurement includes voltage amplitude, injected active power, and reactive power.

[0045] As a further improvement to the technical solution of the present invention, the states of all nodes include voltage amplitude and phase angle.

[0046] The power distribution network state estimation method based on dynamic spatiotemporal graph neural network provided by this invention has the following advantages compared with the prior art:

[0047] Construct a dynamic adjacency matrix: dynamically adjust the weights of the adjacency matrix in the graph convolutional network based on real-time or predicted line power flow, so that the model can adapt to changes in the operating state of the distribution network.

[0048] 1. Deep fusion of spatiotemporal features: A parallel-fusion architecture is designed, which uses graph convolutional networks (GCN) to capture spatial topological features, uses temporal Transformers to capture temporal dependent features, and performs adaptive weighted fusion through a gated fusion mechanism to fully explore spatiotemporal correlation information.

[0049] 2. Pseudo-measurement generation and enhancement: Utilizing historical data and topology information, a lightweight generation module generates reliable pseudo-measurements for nodes with missing measurements, effectively improving measurement redundancy and enhancing the input quality of the state estimation model.

[0050] 4. Through simulation verification on a distribution network with a high proportion of distributed energy, this method can achieve higher accuracy and stronger robustness in state estimation than traditional WLS and existing machine learning methods under extremely low measurement configuration. Attached Figure Description

[0051] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0052] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 A schematic diagram illustrating the process of power distribution network modeling. Detailed Implementation

[0054] To better understand the above-mentioned objectives, features, and advantages of the present invention, the solutions of the present invention will be further described below. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other.

[0055] Many specific details are set forth in the following description in order to provide a full understanding of the invention, but the invention may also be practiced in other ways different from those described herein; obviously, the embodiments in the specification are only some embodiments of the invention, and not all embodiments.

[0056] The specific embodiments of the present invention will be described in detail below.

[0057] Example 1

[0058] The power distribution network state estimation method based on Dynamic Spatiotemporal Graph Neural Network (DSTGNN) includes the following steps:

[0059] The distribution network is modeled as a graph structure G=(V, E), where nodes V represent buses and edges E represent lines. The node feature vectors include real-time measurements or pseudo-measurements (voltage amplitude, injected active power, reactive power). The core of the model consists of three parts: a dynamic graph convolution module, a time-series Transformer module, and a gated fusion module.

[0060] The dynamic graph convolution module is used for spatial feature extraction. Specifically, a learnable dynamic weight matrix D(t) is introduced into the dynamic graph convolution module to construct a dynamic adjacency matrix A. D(t) is obtained by normalizing the estimated power flow amplitude of the line at the current moment, reflecting the strength of the line's electrical coupling. The construction of the dynamic adjacency matrix is ​​as follows:

[0061] In the formula: This represents the adjacency matrix adjusted at time t. It dynamically adjusts the weights based on real-time line power flow, reflecting the strength of electrical connections between nodes. This represents the static adjacency matrix, based on the physical topology of the distribution network. Its element values ​​indicate whether there is a connection between nodes, usually 0 or 1; ⊙ represents the Hadamard product, α represents the learnable scaling factor, and D(t) represents the learnable dynamic weight matrix, which is usually obtained by normalizing the estimated power flow amplitude of the line at the current moment, reflecting the real-time changes in the electrical coupling of the line;

[0062] Then, graph convolution is performed using A_dynamic(t):

[0063]

[0064] In the formula: Let D_hat represent the spatial features of the graph convolution output, and let A_dynamic(t) be the degree matrix. Let W_gcn represent the node feature matrix at time t, W_gcn be the trainable weight matrix, and σ be the Sigmoid activation function.

[0065] The Time Series Transformer module is used for time feature extraction. Its core objective is to deeply mine the complex dependencies of each node's measurement data over time, including long-term trends, periodic changes (such as intraday and weekly load variations), and abrupt change patterns (such as transient processes caused by faults or switching operations). The specific steps for time feature extraction using the Time Series Transformer module are as follows:

[0066] Input: For each node in the distribution network, extract its historical features over the past T consecutive time steps to form a time series sample X_i = [x_i(t-T+1), x_i(t-T+2), ..., x_i(t)], where x_i(t) represents the feature vector of the i-th node at time t. The sequences of all nodes together constitute the input tensor X_seq ∈ R^(N×T×F), where N represents the number of nodes, T represents the time step, and F represents the feature dimension.

[0067] Positional Encoding: Learnable positional encoding (PE) is added to the input tensor, specifically as follows:

[0068]

[0069] In the formula: Indicates enhanced temporal features. The input tensor PE ∈ R^(T×d_model) is a trainable matrix, where R is the real number field, T is the time step, and d_model represents the model dimension of the Transformer. In this embodiment, positional encoding enables the model to effectively utilize the temporal order information of the data.

[0070] Multi-head self-attention mechanism: enhanced temporal features The input is fed into a Transformer encoder consisting of L layers. The core of each layer is a multi-head self-attention (MSA) mechanism, and its calculation process is as follows:

[0071] First, the input sequence is linearly mapped to a matrix of query Q, key K, and value V:

[0072]

[0073] In the formula: , , ∈ R^(d_model×d_k) is the learnable weight matrix, where d_k is the dimension of each head;

[0074] Then, calculate the scaled dot product attention:

[0075]

[0076] In the formula: Q, K, and V are the query, key, and value matrices obtained from the input sequence through linear transformation, respectively. The dimension of the key vector is used to scale the dot product;

[0077] Multi-head attention executes this process in parallel h times (i.e., h heads), with each head using a different weight matrix, thereby capturing different aspects of the dependencies in the sequence:

[0078]

[0079]

[0080] In the formula: This represents the output of the h-th attention head. , , It is the learnable weight matrix of the h-th attention head. ∈R^(h·d_v×d_model) is the output projection matrix, and Concat represents the concatenation operation;

[0081] Through the multi-head self-attention mechanism, the model in this embodiment can simultaneously focus on the complex relationship between a point in time and multiple points before and after it, such as identifying that a voltage drop is caused by a previous switching action.

[0082] Feedforward networks and residual connections:

[0083] Each attention sublayer is followed by a feedforward neural network (FFN), which consists of two linear transformations and a ReLU activation function:

[0084]

[0085] In the formula: and It is a learnable weight matrix. , It is a bias term. This represents the ReLU activation function.

[0086] Each feedforward neural network employs residual connections and layer normalization to stabilize the training process of the deep network, and its output can be represented as:

[0087]

[0088]

[0089] In the formula: Let represent the output of the l-th attention sublayer. This represents the output of the feedforward neural network in the l-th attention sublayer. It is a layer normalization operation. This refers to the multi-head attention mechanism.

[0090] Output: After stacking L layers of Transformer encoders, the output H_trans^(L) of the last attention sublayer is taken as the final extracted temporal feature representation H_trans. It integrates all contextual information of each node's own measurement within the entire time window, and has stronger robustness to noise and transient measurement loss.

[0091] The gated fusion module adaptively fuses spatial and temporal features, ultimately outputting state estimates for all nodes. The specific process of this gated fusion module adaptively fusing spatial and temporal features is as follows:

[0092]

[0093]

[0094] In the formula: G is the fusion gate, and its value is between 0 and 1. It is a sigmoid activation function, || represents the concatenation operation, W_gate and b_gate are gating parameters, and the final output H_fused is the state estimate of all nodes (such as voltage magnitude and phase angle).

[0095] Example 2

[0096] (1) Experimental Objective

[0097] This embodiment aims to verify the effectiveness of the model in embodiment 1 in distribution network state estimation, and to compare it with traditional methods and other machine learning methods to demonstrate its superior performance in scenarios with scarce measurement data, noise interference, and high proportion of distributed energy access.

[0098] (2) Experimental platform and data

[0099] Simulation platform: The IEEE 33-node distribution system was used as the test case. This system is a standard radial distribution network, and distributed photovoltaic (PV) systems were connected to multiple nodes to simulate a high proportion of renewable energy scenarios.

[0100] Data generation:

[0101] 2.1 Perform quasi-steady-state power flow calculations to simulate the system operation under different load levels and different photovoltaic outputs (considering daily fluctuations and random fluctuations), generating simulation data for one year with a time resolution of 15 minutes.

[0102] 2.2 Measurement Configuration: Advanced Measurement Systems (AMIs) are installed only at the root nodes (substations) and a few key nodes (10 in total, approximately 30%) of the distribution network. These AMIs provide measurements of voltage amplitude, active power injection, and reactive power injection. Other nodes lack direct measurement capabilities and rely on pseudo-measurements.

[0103] 2.3 Pseudo-measurement: Based on historical load and photovoltaic output curves, load and power generation prediction values ​​are generated for nodes without measurement as pseudo-measurements, with an error set to ±20% (Gaussian distribution) to simulate real-world uncertainties.

[0104] 2.4 Noise Addition: Add 1%~3% Gaussian white noise to all real measurements to simulate sensor error.

[0105] 2.5 Dataset Partitioning: The total amount of data generated is 35,040 time points (365 days * 96 points / day). It is divided into a training set (21,024 samples), a validation set (7,008 samples), and a test set (7,008 samples) in a 6:2:2 ratio.

[0106] (3) Comparison method

[0107] For a fair comparison, the method of this invention is compared with the following three representative methods:

[0108] Traditional Weighted Least Squares (WLS): Serves as a benchmark for traditional methods. It uses the complete network topology and parameters, with input including real-time and pseudo-measurements for all nodes.

[0109] Static Graph Convolutional Networks (GCNs): These networks use a fixed, physically connected adjacency matrix to estimate the state using only the measurement data at the current time step. They represent graph learning methods that utilize only spatial features.

[0110] Long Short-Term Memory (LSTM) network: Each node is trained with an LSTM model, which estimates the state using only its own time-series historical data. It represents a deep learning method that only utilizes time features.

[0111] (4) Evaluation indicators

[0112] The root mean square error of voltage amplitude (RMSE-V) and the mean absolute error of voltage amplitude (MAE-V) are used to quantify the estimation accuracy.

[0113] (5) Implementation details of the model of the present invention

[0114] Dynamic adjacency matrix: The edge weights are dynamically adjusted based on the magnitude of the predicted power flow (calculated from measurements and pseudo-measurements). The scaling factor α is learned during training.

[0115] Dynamic graph convolution module: uses 2 layers Each layer has an output dimension of 32 and uses the ReLU activation function.

[0116] The temporal Transformer module has a time window length of T=16 (i.e., 4 hours of historical data). It employs a 2-layer encoder, a model dimension of d_model=64, and 4 attention heads.

[0117] Training: The Adam optimizer was used with an initial learning rate of 0.001 and a learning rate decay strategy. The batch size was 32. The loss function was the mean square error (MSE) of the voltage amplitude and phase angle estimates.

[0118] (6) Experimental Results and Effects

[0119] Table 1. Overall performance comparison of different methods on the test set.

[0120] method RMSE-V (pu) MAE-V (pu) Traditional WLS 0.0185 0.0142 Static GCN 0.0128 0.0097 LSTM 0.0151 0.0113 This invention 0.0072 0.0054

[0121] As shown in Table 1, the DSTGNN method proposed in this invention significantly outperforms all compared methods in both metrics. Compared to the traditional WLS, RMSE-V is reduced by 61%, demonstrating the advantage of deep learning models in handling nonlinear problems. Compared to static GCN using only spatial features and LSTM using only temporal features, the error is reduced by 44% and 52%, respectively, which fully demonstrates the effectiveness and necessity of fusing spatiotemporal features and introducing a dynamic adjacency matrix.

[0122] (7) Conclusion

[0123] This embodiment has been thoroughly experimentally verified on the IEEE 33-node system. The results show that:

[0124] 1. The DSTGNN state estimation method proposed in this invention has significantly higher accuracy than traditional WLS and single-feature deep learning models.

[0125] 2. The method of the present invention has stronger robustness to measurement noise and data loss, and is more suitable for complex environments in actual power distribution networks where measurements are incomplete and data quality is low.

[0126] 3. The effectiveness of the dynamic adjacency matrix and spatiotemporal feature fusion mechanism is fully supported by experimental data, proving the technological advancement and practicality of this invention.

[0127] This invention provides a high-precision and robust solution to address the state estimation challenges faced by modern distribution networks with a high proportion of renewable energy integration.

[0128] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the present invention. Although detailed descriptions have been provided with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments, and they should all be covered within the protection scope of the claims.

Claims

1. A power distribution network state estimation method based on a dynamic spatio-temporal graph neural network, characterized in that, The method comprises the following steps: The power distribution network is modeled as a graph structure G=(V, E), wherein the nodes V represent busbars, the edges E represent lines, the node feature vectors include real-time measurements or pseudo-measurements, and the model core is composed of three parts: a dynamic graph convolution module, a time series Transformer module, and a gated fusion module; The dynamic graph convolution module is used for spatial feature extraction; the time series Transformer module is used for time feature extraction; and the gated fusion module is used for adaptively fusing the spatial features and the time features, and finally outputs state estimation values of all nodes.

2. The power distribution network state estimation method based on dynamic spatio-temporal graph neural network according to claim 1, characterized in that, In the dynamic graph convolution module, a learnable dynamic weight matrix D(t) is introduced to construct a dynamic adjacency matrix A, as follows: ; wherein: denotes the adjusted adjacency matrix at time t, denotes the static adjacency matrix, and denotes the Hadamard product, denotes the learnable scaling coefficients, and D(t) denotes the learnable dynamic weight matrix; Subsequently using performing graph convolution operations: ； wherein: represents a spatial feature of the graph convolution output, is a degree matrix of is a node feature matrix at time t, is a trainable weight matrix, is a Sigmoid activation function.

3. The power distribution network state estimation method based on dynamic spatio-temporal graph neural network according to claim 2, characterized in that, The specific steps of the time series Transformer module for time feature extraction are as follows: Input: For each node in the power distribution network, extract the historical features of the past T consecutive time steps to form a time series sample X_i= [x_i(t-T+1), x_i(t-T+2),...,x_i(t)], wherein x_i(t) represents the feature vector of the i-th node at the t-th time, and the sequences of all nodes together constitute an input tensor X_seq∈R^(N×T×F), wherein N represents the number of nodes, T represents the time step, and F represents the feature dimension; Position encoding: add a learnable position encoding to the input tensor, and the specific process is as follows: ； wherein: denotes an enhanced timing feature, denotes an input tensor, PE∈R^(T×d_model) is a trainable matrix, R is a real number field, T is a time step, and d_model denotes a model dimension of the Transformer. Multi-head self-attention mechanism: enhanced temporal features The input is fed into a Transformer encoder composed of L layers, the core of each layer being a multi-head self-attention mechanism, whose computation is as follows: First, linearly map the input sequence into query Q, key K, and value V matrices: ； where: , , ∈ R^(d_model x d_k) are learnable weight matrices, d_k is the dimension of each head. Then, calculate the scaled dot-product attention: ； where: Q, K, V are the Query, Key and Value matrices obtained from the input sequence by linear transformations, is the dimension of the Key vector used to scale the dot product; Multi-head attention executes this process h times in parallel, and each head uses different weight matrices to capture different aspects of dependencies in the sequence: ； ； wherein: represents the output of the h-th attention head, , , is a learnable weight matrix of the h-th attention head, is an output projection matrix, and Concat represents a concatenation operation. Feedforward network and residual connection: Each attention sublayer is followed by a feedforward neural network FFN, which is composed of two linear transformations and a ReLU activation function: ； wherein: and are learnable weight matrices, , is a bias term, denotes a ReLU activation function; Each feedforward neural network adopts residual connection and layer normalization to stabilize the training process of the deep network, and the output can be represented as: ； ； wherein: represents an output of the l-th layer attention sub-layer, represents an output of a feed-forward neural network of the l-th layer attention sub-layer, is a layer normalization operation, represents a multi-head attention mechanism; Output: after L-layer Transformer encoder stacking, the output H_trans^(L) of the last attention sublayer is taken as the final extracted time feature representation H_trans.

4. The power distribution network state estimation method based on dynamic spatio-temporal graph neural network according to claim 3, characterized in that, The specific process of the gated fusion module for adaptively fusing the spatial features and the time features is as follows: ； ； where G is the fusion gate with a value between 0 and 1, is a sigmoid activation function, || denotes a concatenation operation, W gate and b gate are gating parameters, and the final output H fused is the state estimation value of all nodes.

5. The power distribution network state estimation method based on dynamic spatio-temporal graph neural network according to any one of claims 1 to 4, characterized in that, The real-time measurements or pseudo-measurements include voltage amplitude, injected active power, and reactive power.

6. The power distribution network state estimation method based on dynamic spatio-temporal graph neural network according to any one of claims 1 to 4, characterized in that, The states of all nodes include voltage amplitude and phase angle.

Images