State estimation method based on adaptive space-time diagram neural network

Abstract

The invention relates to a power distribution network state estimation method based on an adaptive space-time diagram neural network, and the method mainly comprises the following steps: S1, collecting historical and real-time measurement data of a power distribution network, and carrying out the preprocessing of the data, so as to guarantee the integrity of the data, provide high-quality input data for a model, and improve the estimation precision and stability of the model; s2, discrete wavelet transform is carried out on historical measurement data, multi-scale decomposition is achieved, and low-frequency and high-frequency components are extracted; a double-branch time sequence fusion module is constructed, global trend and local fluctuation features are respectively captured through a dynamic attention mechanism and a time convolution network, and the features are efficiently fused by means of adaptive weights. S3, in the real-time data processing process, branch measurement features are extracted through a multi-layer perceptron (MLP) and mapped to nodes of the whole network, dynamic integration of historical data and real-time data is achieved, and therefore the real-time performance and accuracy of state estimation of the power distribution network are improved.

Description

Technical Field

[0001] The present invention proposes a state estimation method based on an adaptive spatiotemporal graph neural network, which belongs to the technical field of smart grids. Background Art

[0002] As distribution networks gradually evolve toward intelligence, automation, and digitalization, state estimation has become a key technology for optimizing the operation and management of smart grids. Traditional distribution network state estimation methods often rely on physical models and empirical rules. However, these methods often struggle to cope with the highly complex and dynamically changing power flow characteristics of distribution networks in complex dynamic environments, resulting in insufficient estimation accuracy and robustness. Therefore, improving the accuracy, robustness, and real-time performance of distribution network state estimation has become a major challenge in power system research.

[0003] In recent years, deep learning and graph neural networks (GNNs) have gradually become important technical tools for distribution network state estimation due to their excellent performance in spatiotemporal data modeling. GNNs can effectively handle the complex topological relationships between nodes in a distribution network and capture both global and local spatiotemporal dynamic characteristics. However, in practical applications, accurately integrating multiple data sources of the distribution network (such as historical and real-time data) while improving the model's ability to express spatiotemporal characteristics remains an urgent problem. Among existing methods, although graph convolutional networks (GCNs) and graph attention networks (GATs) have made some progress in modeling graph-structured data, these methods often ignore the temporal characteristics of distribution networks and their multi-scale dynamic changes. Therefore, how to effectively combine time series data with spatial topological information to design adaptable spatiotemporal graph neural networks remains an important research direction. Summary of the Invention

[0004] Purpose of the Invention: This invention aims to provide a distribution network state estimation method based on an adaptive spatiotemporal graph neural network. By leveraging technologies such as multi-scale analysis, dynamic attention mechanisms, and graph convolutional networks, this method improves the accuracy and robustness of distribution network state estimation in complex dynamic environments. This method can effectively process historical and real-time data, combine static and dynamic topological structures, and dynamically integrate multi-level features to improve the accuracy and stability of distribution network state estimation in real time.

[0005] Technical solution: The present invention discloses a distribution network state estimation method based on an adaptive spatiotemporal graph neural network, comprising the following steps:

[0006] S1: Collect historical and real-time measurement data of the distribution network and preprocess the data to ensure data integrity, thereby providing high-quality input data for the model and improving its estimation accuracy and stability.

[0007] S2: Perform a discrete wavelet transform on historical measurement data to achieve multi-scale decomposition and extract low-frequency and high-frequency components. A dual-branch time series fusion module is constructed, capturing global trends and local fluctuations through a dynamic attention mechanism and a temporal convolutional network, respectively. Adaptive weights are then used to efficiently fuse these features. Furthermore, a hybrid graph learning module is designed, combining static physical topology with adaptive dynamic topology, optimizing the adjacency matrix and integrating graph attention with graph convolutional features to enhance the expressive power of spatiotemporal features.

[0008] S3: During real-time data processing, a multi-layer perceptron (MLP) is used to extract branch measurement features and map them to nodes across the entire network, enabling dynamic integration of historical and real-time data, thereby improving the real-time performance and accuracy of distribution network state estimation.

[0009] As an improvement of the present invention, step S1 includes the following steps:

[0010] S11: In a distribution network, historical and real-time data are collected from N nodes, covering T time steps. Each node contains M-dimensional features (such as voltage amplitude, voltage phase angle, active power, and reactive power). The observation vector for node i at time t is expressed as: The historical data tensor of the entire network is: X = [X (1) ,X (2) ,…,X (T) ]∈R T×N×M ,in is the state matrix of the entire network at time t.

[0011] S12: In the distribution network sensor data, due to equipment failure, communication interruption and other reasons, data may be missing, affecting subsequent analysis. Therefore, the time neighborhood linear interpolation method is used to fill the missing values. For node i in the feature m dimension, the missing data at time t The interpolation calculation is as follows:

[0012]

[0013] Among them, t (t-1) is the previous valid timestamp of the missing time point t, t (t+1) is the next valid timestamp after the missing time point t, and It is the effective measurement value of the adjacent moments.

[0014] S13: There may be outliers in the distribution network data (such as extreme data caused by sensor failure), which need to be detected and corrected in a timely manner. For each node i’s feature m, calculate the statistic based on the time dimension:

[0015]

[0016] Among them, μ i,m is the mean of feature m of node i in the time series, reflecting the long-term steady-state level, σ i,m is the standard deviation, which represents the range of data fluctuation. T-1: The sample standard deviation is calculated using unbiased estimation (Bessel correction).

[0017] Determine outliers based on the 3σ criterion:

[0018]

[0019] in, is the value of feature m of node i at time t. Usually set γ = 3, then the judgment For outliers. Replace with the local window mean:

[0020]

[0021] Where K is the window radius (the default K=3, covering the previous and next 3 time steps, a total of 7 points).

[0022] S14: Since data from different measurement points have different physical dimensions, which may affect model training, the data needs to be normalized. The goal of normalization is to convert the data to a fixed range (such as [0,1]), thereby improving the stability and convergence speed of the model. For feature m of node i, the normalization calculation is as follows:

[0023]

[0024] Among them, x i,m,min is the local minimum value of feature m of node i in its time series, X i,m,max is the local maximum of feature m of node i in its time series.

[0025] As an improvement of the present invention, step S2 includes the following steps:

[0026] S21: First, use DWT to analyze the input node historical load data X=[x1,x2,...,x T ]∈R M×T×N Decomposition is performed, where M is the number of node variables; T is the length of the historical load sequence; and N is the number of distribution network nodes. DWT realizes time-frequency analysis of time series signals through multi-level decomposition. In the k-th layer decomposition process, the time series signal is decomposed into frequency representations of different scales using low-pass filters L and high-pass filters H, so that the model can simultaneously process the short-term and long-term characteristics of time series data. Specifically, the output of the k-th layer decomposition is the low-frequency approximation coefficient CL(k) and high frequency detail coefficient CH (k) ,Right now:

[0027] CL (k) ,CH (k) =DWT(CL (k-1) ) ↓

[0028] Among them, CL (0) = X is the original signal. Each layer of decomposition is achieved through convolution and downsampling:

[0029]

[0030] Among them, s = 1, 2, ..., M / 2 represents the index after downsampling; M is the CL after k-1 layer decomposition (k-1) Under the dual-scale effect of high-pass filter L and low-pass filter H, the number of samples of the output signal of DWT is the original signal CL (k-1) According to the Nyquist sampling theorem, half of the samples can be removed by downsampling, retaining the main information of the signal. In the decomposition result, the approximate coefficient CL (k) Describes the long-term trend of the time series and reflects its basic characteristics. (k) It describes the short-term changes of the time series and captures the subtle fluctuations of the signal. As the decomposition level increases, the frequency resolution of the original signal gradually improves.

[0031] After k-layer decomposition, DWT outputs a set of k+1 coefficients:

[0032] C=[CH (1) ,CH (2) ,…,CH (k) ,CL (k) ]

[0033] In order to effectively extract the long-term and short-term dynamic information from the historical load, a band-selective masking operator is designed based on the multi-scale coefficient set obtained by k-layer wavelet decomposition to separate the long-term and short-term dynamic features. Specifically, the masking function M is constructed. L =[0,0,...,0,I(k)], through the operator M L (C)=C⊙M L Achieve progressive suppression of high-frequency coefficients and retain low-frequency components CL (k) . Design complementary mask M H =[I(1),I(2),...,I(k),0], through M H (C)=C⊙M H Extract high-frequency features CH (k) .

[0034] On this basis, the inverse discrete wavelet transform (IDWT) is used to reconstruct the processed low-frequency and high-frequency coefficients to extract the heterogeneous pattern of the signal. This process reconstructs the signal X by upsampling and convolution operations on the low-frequency and high-frequency coefficients. L ,X H ∈R M×T×N , which correspond to the time domain characteristics of the node historical load at different scales. The specific reconstruction process can be expressed as:

[0035] X L =IDWT(M L (C)

[0036] X H =IDWT(M H (C)

[0037] Through the multi-scale feature extraction and reconstruction strategy based on wavelet decomposition, the effective separation of long-term and short-term dynamic characteristics of historical loads and the fine representation of time domain information are achieved.

[0038] S22: Based on the results of the multi-scale time feature decomposition module, this application designs a dual-branch time series fusion module. Specifically, for the long-term trend component X output by the multi-scale time feature decomposition module, L ∈R M×T×N and the short-term volatility component X H ∈R M×T×N , respectively input into the dynamic attention temporal mechanism and the temporal convolutional network (TCN) to achieve independent modeling of long-term and short-term features. To further enhance the collaborative modeling capability of multi-scale features, an adaptive feature fusion mechanism is designed to dynamically adjust the weight distribution of features at different time scales.

[0039] Dynamic attention time mechanism: In the state estimation of distribution networks, the time dependency of historical state information is often affected by the mutual influence of different features and nodes. Therefore, how to effectively model these complex global time dependencies from the historical state sequence of the node has become a core challenge. Therefore, this application introduces a dynamic attention time mechanism, which aims to accurately model the global time dependency of the load sequence through a dual-stream feature extraction mechanism and adaptive time weight allocation, and improve the temporal expression capability of the model. The dynamic attention time mechanism mainly includes correlation modeling of the variable domain and node domain and global time correlation modeling.

[0040] In the variable domain, the interaction between different variables on the same node is considered. To this end, a two-dimensional channel convolution is introduced to calculate the weighted combination between variables:

[0041]

[0042] Where w∈R M×1×1 is the unbiased convolution kernel weight. Since the size of the convolution kernel is (1,1), the weighting operation is only performed on the feature dimension, so the output time step T and the number of nodes N remain unchanged, and the final output is Represents the characteristics of each variable at different time steps.

[0043] In the node domain, we consider the dependency relationship between different nodes of the same variable. To do this, we first transform the input value dimension, and then use two-dimensional channel convolution to perform weighted calculations on the relationship between different nodes:

[0044]

[0045] Among them, permute is a dimension conversion function, which converts the 0th dimension and the 2nd dimension to obtain It is the node domain output, representing the characteristics of each node at different time steps.

[0046] In getting the variable domain characteristics and node domain characteristics Finally, an adaptive temporal attention mechanism is adopted to enable the model to focus on key time steps by learning the dynamic weight distribution between time steps, thereby improving the ability of time-dependent modeling.

[0047] First, in order to quantify the mutual influence between time steps, a trainable parameter w1∈R is introduced N×M Learn the mapping relationship between variable domain and node domain features, and calculate the attention weights between time steps:

[0048]

[0049] Where A∈R T×T is the temporal attention weight matrix, which describes the degree of correlation between different time steps.

[0050] Secondly, in order to avoid the possible numerical scale unevenness problem of the original attention matrix A, direct use may lead to numerical instability during gradient update and affect the convergence of the model, so a trainable transformation matrix w2∈R is introduced T×1 Adjust the weight relationship between different time steps to enhance the influence of the current time step. Then, use Softmax normalization to ensure that the time attention distribution is more stable and reasonable:

[0051]

[0052] Finally, the information on the time axis is weighted and aggregated through the Einstein summation convention, so that the model can more effectively integrate global time-dependent information and improve the time series modeling capability. The final output result is X ATT ∈R M×1×N The specific calculation formula is:

[0053]

[0054] Temporal Convolutional Network: TCN is a deep learning model specifically for time series data, which can efficiently extract sequence features through one-dimensional convolution. Compared with CNN, TCN introduces dilated convolution and causal convolution, which not only effectively expands the receptive field, but also strictly guarantees the causal structure, so that the output of the model at each time step depends only on the current and past inputs, avoiding future information leakage, thereby meeting the causal consistency requirements in time series modeling. In addition, the local convolution mechanism of TCN is highly sensitive to short-term fluctuations in time series, which makes it show good modeling capabilities in capturing rapidly changing dynamic features. Therefore, in this application, TCN is used to extract the short-term fluctuation component X. H ∈R M×T×N Perform feature extraction. First, reorganize X into H The variable dimension M and node dimension N in are merged into the channel dimension, and we get To adapt to the subsequent one-dimensional convolution operation. Then, TCN is used to extract features from the input time series to obtain the output X TCN ∈R (M ×N)×1 .

[0055] Adaptive weight feature fusion mechanism: dynamic attention time mechanism in the dual-branch temporal fusion module and the trend component X of TCN output ATT and the fluctuation component X TCN Traditional fixed-weight fusion methods (such as direct addition or concatenation) are unable to flexibly adjust the importance of different components according to different scenarios, affecting the model's ability to model complex time series patterns. This paper introduces an adaptive weight mechanism to achieve dynamic fusion of trend and fluctuation components. To this end, a learnable parameter α is used to achieve dynamic weight allocation:

[0056] α sigmoid =Sigmoid(α)

[0057] X fusion =α sigmoid ·X ATT +(1-α sigmoid )·X TCN

[0058] Among them, X fusion ∈R M×Nis the output result of the fusion.

[0059] S23: In the task of state estimation of distribution networks, traditional static topology modeling methods rely on physical electrical connections and characterize fixed electrical correlations between nodes through admittance matrices. However, this static modeling approach cannot effectively capture the dynamic coupling characteristics between nodes that evolve over time, especially in complex scenarios such as load fluctuations and distributed power supply access, where the spatiotemporal correlation of the system becomes more complex. Therefore, this application proposes a topology-enhanced adaptive dynamic graph learning method that improves the modeling capability of the model for state variables by dynamically adjusting the topology structure. On the basis of maintaining the physical topology structure, a dynamic correlation weight adjustment mechanism is considered to more accurately characterize the dynamic coupling relationship between nodes.

[0060] First, the adjacency matrix A of the distribution network topology adj First, after normalization, the normalized matrix is ​​obtained The calculation formula is:

[0061]

[0062] Where D is the degree matrix. Then, we construct the trainable node embedding matrix E1, E2∈R N×d , and through the linear transformation layer parameter w e,1 、w e,2 ∈R N×d And the nonlinear activation function generates the dynamic association matrix A graph .

[0063] M1=tanh(αE1w e,1 )

[0064] M2=tanh(αE2w e,2 )

[0065]

[0066] Where N is the number of nodes, d is the dimension of the node embedding matrix, and α is a hyperparameter that controls the node saturation rate to avoid gradient vanishing. Finally, a weighted fusion strategy is adopted to combine the static topology information and the dynamic association matrix to generate the final adjacency matrix:

[0067]

[0068] Here, λ is a learnable balancing factor that adjusts the fusion ratio between static and dynamic topologies. By introducing a learnable mechanism, the model can adaptively balance physical priors and data-driven information in different scenarios, thereby improving the accuracy and robustness of distribution network state estimation.

[0069] After obtaining the dynamic graph structure Aadaptive After that, the correlation between the nodes of the distribution network is first captured by the GAT layer and the GCN layer, and the output H is obtained. gat and H gcn Subsequently, in order to combine the advantages of the GAT layer and the GCN layer, this section adopts a gated fusion method, which adjusts the weighted fusion ratio of the GAT layer output and the GCN layer output through a learnable gating coefficient γ, as shown below:

[0070] H fused =γ·H gat +(1-γ)·H gcn

[0071] Among them, H fused is the output of the adaptive dynamic graph learning module. Through the gating mechanism, the model can adaptively adjust the fusion ratio of the multi-head graph attention layer and the graph convolution layer in different situations, thereby better utilizing topological information and optimizing the information transmission and fusion process.

[0072] As an improvement of the present invention, step S3 includes the following steps:

[0073] S31: In the distribution network operation scenario set by this application, only some branches are equipped with real-time measurement devices. How to effectively extract the real-time measurement data features of the branches and integrate them with the historical load data features of the nodes becomes a key issue. To this end, this application adopts a real-time measurement feature extraction method based on a multilayer perceptron (MLP), and maps the extracted features to the nodes of the entire network and integrates them with the historical load features of the nodes, ultimately achieving distribution network state estimation.

[0074] The real-time measurement data in the distribution network can be expressed as a two-dimensional matrix D∈R K×M , where K represents the number of real-time measurement branches and M is the feature dimension of each branch measurement data. To extract effective features from the measurement data, MLP is used to perform nonlinear mapping on it, as shown in the formula:

[0075] D f =σ2(σ1(D·w d,1 +b1)·w d,2 +b2)

[0076] Among them, D f ∈R K×C is the output feature of MLP; C is the feature dimension extracted by MLP; w d,1 ∈R M×2C 、w d,2 ∈R 2C×C are the weight matrices of the input layer and the hidden layer respectively; b1, b2∈R Cis the bias term; σ1 and σ2 are ReLU activation functions.

[0077] Since the real-time measurement data only covers part of the branches, and the historical load characteristics cover N nodes in the entire network, in order to achieve multi-source data fusion, we first introduce a linear mapping to convert the branch real-time measurement data characteristics D obtained from MLP into f Mapping to all network nodes is as follows:

[0078]

[0079] Among them, w d,3 ∈R K×N is the mapping matrix, D N ∈R C×N It is the real-time measurement feature mapped on all nodes in the network.

[0080] Secondly, in order to effectively combine the node historical load characteristics H fused Real-time measurement characteristics of branch D N In this section, an adaptive attention weighting mechanism is designed to dynamically adjust the weights during the fusion process by calculating the relative importance between historical load and real-time measurement. The adaptive attention weight α is shown in the following formula:

[0081]

[0082] Among them, σ4 is the Sigmoid activation function; w d,4 , b4 are the weight matrix and bias term of adaptive attention respectively.

[0083] Again, the adaptive weight α is used to adjust the historical node feature H fused With real-time measurement feature D N Perform weighted fusion:

[0084] F=α·H fused +(1-α)·D N

[0085] Where F∈R C×N is the feature representation after multi-source data fusion. Through this weighting mechanism, the fusion of historical data and real-time measurement data can more accurately reflect the current status of the distribution network.

[0086] Finally, the fused features are further processed by linear mapping, as shown in the following formula:

[0087] Y=F T w d,5

[0088] Among them, w d,5 ∈R C×1 is a linear mapping matrix; Y∈RN×1 Through the above steps, the features extracted from the node historical load data and the branch real-time measurement data are successfully fused to obtain the final distribution network state estimation result.

[0089] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the method for estimating the state of a distribution network based on an adaptive spatiotemporal graph neural network is implemented.

[0090] A computer-readable storage medium stores computer instructions, which, when executed by a processor, implement the distribution network state estimation method based on an adaptive spatiotemporal graph neural network.

[0091] Beneficial effects of the present invention: The distribution network state estimation method based on the adaptive spatiotemporal graph neural network of the present invention has the following significant beneficial effects: 1. By efficiently extracting branch measurement features and integrating them into the nodes of the entire network through a multi-layer perceptron (MLP), the present invention realizes the seamless fusion of historical data and real-time data. This not only improves the real-time performance and accuracy of distribution network state estimation, but also lays the foundation for more accurate load forecasting and fault detection. 2. Through multi-scale feature extraction and dynamic time feature modeling, the present invention can more accurately capture the spatiotemporal characteristics of the distribution network, especially the response to complex load changes and dynamic events. This significantly improves the accuracy of distribution network state estimation and can more efficiently cope with various complex operating conditions. 3. The distribution network state estimation method based on STGNN proposed in the present invention is significantly superior to comparative methods such as GCN, GAT, LSTM and GRU in the IEEE33 node system. By integrating the distribution network topology and node load time series characteristics, the limitations of the single feature model are overcome, and higher-precision state estimation is achieved. In both measurement loss and noise interference test scenarios, the present invention demonstrates excellent performance advantages. Especially in complex scenarios with high-intensity noise environments and multiple missing measurements, the estimation results can still be kept stable and reliable, providing strong support for the actual operation monitoring of the distribution network. BRIEF DESCRIPTION OF THE DRAWINGS

[0092] Figure 1 is a flow chart of the method of the present invention;

[0093] Figure 2 It is the topology diagram of IEEE33-node distribution network system. DETAILED DESCRIPTION

[0094] The specific embodiments of the present invention will be described below in conjunction with the accompanying drawings so that those skilled in the art can better understand the present invention.

[0095] Example: Figure 1As shown, a distribution network state estimation method based on an adaptive spatiotemporal graph neural network includes the following steps:

[0096] S1: Collect historical and real-time measurement data of the distribution network and preprocess the data to ensure data integrity, thereby providing high-quality input data for the model and improving its estimation accuracy and stability.

[0097] S2: Perform a discrete wavelet transform on historical measurement data to achieve multi-scale decomposition and extract low-frequency and high-frequency components. A dual-branch time series fusion module is constructed, capturing global trends and local fluctuations through a dynamic attention mechanism and a temporal convolutional network, respectively. Adaptive weights are then used to efficiently fuse these features. Furthermore, a hybrid graph learning module is designed, combining static physical topology with adaptive dynamic topology, optimizing the adjacency matrix and integrating graph attention with graph convolutional features to enhance the expressive power of spatiotemporal features.

[0098] S3: During real-time data processing, a multi-layer perceptron (MLP) is used to extract branch measurement features and map them to nodes across the entire network, enabling dynamic integration of historical and real-time data, thereby improving the real-time performance and accuracy of distribution network state estimation.

[0099] Wherein, the step S1 includes the following steps:

[0100] S11: In a distribution network, historical and real-time data are collected from N nodes, covering T time steps. Each node contains M-dimensional features (such as voltage amplitude, voltage phase angle, active power, and reactive power). The observation vector for node i at time t is expressed as: The historical data tensor of the entire network is: X = [X (1) ,X (2) ,…,X (T) ]∈R T×N×M ,in is the state matrix of the entire network at time t.

[0101] S12: In the distribution network sensor data, due to equipment failure, communication interruption and other reasons, data may be missing, affecting subsequent analysis. Therefore, the time neighborhood linear interpolation method is used to fill the missing values. For node i in the feature m dimension, the missing data at time t The interpolation calculation is as follows:

[0102]

[0103] Among them, t (t-1) is the previous valid timestamp of the missing time point t, t (t+1) is the next valid timestamp after the missing time point t, and It is the effective measurement value of the adjacent moments.

[0104] S13: There may be outliers in the distribution network data (such as extreme data caused by sensor failure), which need to be detected and corrected in a timely manner. For each node i’s feature m, calculate the statistic based on the time dimension:

[0105]

[0106] Among them, μ i,m is the mean of feature m of node i in the time series, reflecting the long-term steady-state level, σ i,m is the standard deviation, which represents the range of data fluctuation. T-1: The sample standard deviation is calculated using unbiased estimation (Bessel correction).

[0107] Determine outliers based on the 3σ criterion:

[0108]

[0109] in, is the value of feature m of node i at time t. Usually set γ = 3, then the judgment For outliers. Replace with the local window mean:

[0110]

[0111] Where K is the window radius (the default K=3, covering the previous and next 3 time steps, a total of 7 points).

[0112] S14: Since data from different measurement points have different physical dimensions, which may affect model training, the data needs to be normalized. The goal of normalization is to convert the data to a fixed range (such as [0,1]), thereby improving the stability and convergence speed of the model. For feature m of node i, the normalization calculation is as follows:

[0113]

[0114] Among them, x i,m,min is the local minimum value of feature m of node i in its time series, X i,m,max is the local maximum of feature m of node i in its time series.

[0115] Wherein, the step S2 includes the following steps:

[0116] S21: First, use DWT to analyze the input node historical load data X=[x1,x2,...,x T ]∈R M×T×NDecomposition is performed, where M is the number of node variables; T is the length of the historical load sequence; and N is the number of distribution network nodes. DWT realizes time-frequency analysis of time series signals through multi-level decomposition. In the k-th layer decomposition process, the time series signal is decomposed into frequency representations of different scales using low-pass filters L and high-pass filters H, so that the model can simultaneously process the short-term and long-term characteristics of time series data. Specifically, the output of the k-th layer decomposition is the low-frequency approximation coefficient CL (k) and high frequency detail coefficient CH (k) ,Right now:

[0117] CL (k) ,CH (k) =DWT(CL (k-1) ) ↓

[0118] Among them, CL (0) = X is the original signal. Each layer of decomposition is achieved through convolution and downsampling:

[0119]

[0120] Among them, s = 1, 2, ..., M / 2 represents the index after downsampling; M is the CL after k-1 layer decomposition (k-1) Under the dual-scale effect of high-pass filter L and low-pass filter H, the number of samples of the output signal of DWT is the original signal CL (k-1) According to the Nyquist sampling theorem, half of the samples can be removed by downsampling, retaining the main information of the signal. In the decomposition result, the approximate coefficient CL (k) Describes the long-term trend of the time series and reflects its basic characteristics. (k) It describes the short-term changes of the time series and captures the subtle fluctuations of the signal. As the decomposition level increases, the frequency resolution of the original signal gradually improves.

[0121] After k-layer decomposition, DWT outputs a set of k+1 coefficients:

[0122] C=[CH (1) ,CH (2) ,…,CH (k) ,CL (k) ]

[0123] In order to effectively extract the long-term and short-term dynamic information from the historical load, a band-selective masking operator is designed based on the multi-scale coefficient set obtained by k-layer wavelet decomposition to separate the long-term and short-term dynamic features. Specifically, the masking function M is constructed. L =[0,0,...,0,I(k)], through the operator M L (C)=C⊙M LAchieve progressive suppression of high-frequency coefficients and retain low-frequency components CL (k) . Design complementary mask M H =[I(1),I(2),...,I(k),0], through M H (C)=C⊙M H Extract high-frequency features CH (k) .

[0124] On this basis, the inverse discrete wavelet transform (IDWT) is used to reconstruct the processed low-frequency and high-frequency coefficients to extract the heterogeneous pattern of the signal. This process reconstructs the signal X by upsampling and convolution operations on the low-frequency and high-frequency coefficients. L ,X H ∈R M×T×N , which correspond to the time domain characteristics of the node historical load at different scales. The specific reconstruction process can be expressed as:

[0125] X L =IDWT(M L (C)

[0126] X H =IDWT(M H (C)

[0127] Through the multi-scale feature extraction and reconstruction strategy based on wavelet decomposition, the effective separation of long-term and short-term dynamic characteristics of historical loads and the fine representation of time domain information are achieved.

[0128] S22: Based on the results of the multi-scale time feature decomposition module, this application designs a dual-branch time series fusion module. Specifically, for the long-term trend component X output by the multi-scale time feature decomposition module, L ∈R M×T×N and the short-term volatility component X H ∈R M×T×N , respectively input into the dynamic attention temporal mechanism and the temporal convolutional network (TCN) to achieve independent modeling of long-term and short-term features. To further enhance the collaborative modeling capability of multi-scale features, an adaptive feature fusion mechanism is designed to dynamically adjust the weight distribution of features at different time scales.

[0129] Dynamic attention time mechanism: In the state estimation of distribution networks, the time dependency of historical state information is often affected by the mutual influence of different features and nodes. Therefore, how to effectively model these complex global time dependencies from the historical state sequence of the node has become a core challenge. Therefore, this application introduces a dynamic attention time mechanism, which aims to accurately model the global time dependency of the load sequence through a dual-stream feature extraction mechanism and adaptive time weight allocation, and improve the temporal expression capability of the model. The dynamic attention time mechanism mainly includes correlation modeling of the variable domain and node domain and global time correlation modeling.

[0130] In the variable domain, the interaction between different variables on the same node is considered. To this end, a two-dimensional channel convolution is introduced to calculate the weighted combination between variables:

[0131]

[0132] Where w∈R M×1×1 is the unbiased convolution kernel weight. Since the size of the convolution kernel is (1,1), the weighting operation is only performed on the feature dimension, so the output time step T and the number of nodes N remain unchanged, and the final output is Represents the characteristics of each variable at different time steps.

[0133] In the node domain, we consider the dependency relationship between different nodes of the same variable. To do this, we first transform the input value dimension, and then use two-dimensional channel convolution to perform weighted calculations on the relationship between different nodes:

[0134]

[0135] Among them, permute is a dimension conversion function, which converts the 0th dimension and the 2nd dimension to obtain It is the node domain output, representing the characteristics of each node at different time steps.

[0136] In getting the variable domain characteristics and node domain characteristics Finally, an adaptive temporal attention mechanism is adopted to enable the model to focus on key time steps by learning the dynamic weight distribution between time steps, thereby improving the ability of time-dependent modeling.

[0137] First, in order to quantify the mutual influence between time steps, a trainable parameter w1∈R is introduced N×M Learn the mapping relationship between variable domain and node domain features, and calculate the attention weights between time steps:

[0138]

[0139] Where A∈RT×T is the temporal attention weight matrix, which describes the degree of correlation between different time steps.

[0140] Secondly, in order to avoid the possible numerical scale unevenness problem of the original attention matrix A, direct use may lead to numerical instability during gradient update and affect the convergence of the model, so a trainable transformation matrix w2∈R is introduced T×1 Adjust the weight relationship between different time steps to enhance the influence of the current time step. Then, use Softmax normalization to ensure that the time attention distribution is more stable and reasonable:

[0141]

[0142] Finally, the information on the time axis is weighted and aggregated through the Einstein summation convention, so that the model can more effectively integrate global time-dependent information and improve the time series modeling capability. The final output result is X ATT ∈R M×1×N The specific calculation formula is:

[0143]

[0144] Temporal Convolutional Network: TCN is a deep learning model specifically for time series data, which can efficiently extract sequence features through one-dimensional convolution. Compared with CNN, TCN introduces dilated convolution and causal convolution, which not only effectively expands the receptive field, but also strictly guarantees the causal structure, so that the output of the model at each time step depends only on the current and past inputs, avoiding future information leakage, thereby meeting the causal consistency requirements in time series modeling. In addition, the local convolution mechanism of TCN is highly sensitive to short-term fluctuations in time series, which makes it show good modeling capabilities in capturing rapidly changing dynamic features. Therefore, in this application, TCN is used to extract the short-term fluctuation component X. H ∈R M×T×N Perform feature extraction. First, reorganize X into H The variable dimension M and node dimension N in are merged into the channel dimension, and we get To adapt to the subsequent one-dimensional convolution operation. Then, TCN is used to extract features from the input time series to obtain the output X TCN ∈R (M ×N)×1 .

[0145] Adaptive weight feature fusion mechanism: dynamic attention time mechanism in the dual-branch temporal fusion module and the trend component X of TCN output ATT and the fluctuation component X TCNTraditional fixed-weight fusion methods (such as direct addition or concatenation) are unable to flexibly adjust the importance of different components according to different scenarios, affecting the model's ability to model complex time series patterns. This paper introduces an adaptive weight mechanism to achieve dynamic fusion of trend and fluctuation components. To this end, a learnable parameter α is used to achieve dynamic weight allocation:

[0146] α sigmoid =Sigmoid(α)

[0147] X fusion =α sigmoid ·X ATT +(1-α sigmoid )·X TCN

[0148] Among them, X fusion ∈R M×N is the output result of the fusion.

[0149] S23: In the task of state estimation of distribution networks, traditional static topology modeling methods rely on physical electrical connections and characterize fixed electrical correlations between nodes through admittance matrices. However, this static modeling approach cannot effectively capture the dynamic coupling characteristics between nodes that evolve over time, especially in complex scenarios such as load fluctuations and distributed power supply access, where the spatiotemporal correlation of the system becomes more complex. Therefore, this application proposes a topology-enhanced adaptive dynamic graph learning method that improves the modeling capability of the model for state variables by dynamically adjusting the topology structure. On the basis of maintaining the physical topology structure, a dynamic correlation weight adjustment mechanism is considered to more accurately characterize the dynamic coupling relationship between nodes.

[0150] First, the adjacency matrix A of the distribution network topology adj First, after normalization, the normalized matrix is ​​obtained The calculation formula is:

[0151]

[0152] Where D is the degree matrix. Then, we construct the trainable node embedding matrix E1, E2∈R N×d , and through the linear transformation layer parameter w e,1 、w e,2 ∈R N×d And the nonlinear activation function generates the dynamic association matrix A graph .

[0153] M1=tanh(αE1w e,1 )

[0154] M2=tanh(αE2w e,2 )

[0155]

[0156] Where N is the number of nodes, d is the dimension of the node embedding matrix, and α is a hyperparameter that controls the node saturation rate to avoid gradient vanishing. Finally, a weighted fusion strategy is adopted to combine the static topology information and the dynamic association matrix to generate the final adjacency matrix:

[0157]

[0158] Here, λ is a learnable balancing factor that adjusts the fusion ratio between static and dynamic topologies. By introducing a learnable mechanism, the model can adaptively balance physical priors and data-driven information in different scenarios, thereby improving the accuracy and robustness of distribution network state estimation.

[0159] After obtaining the dynamic graph structure A adaptive After that, the correlation between the nodes of the distribution network is first captured by the GAT layer and the GCN layer, and the output H is obtained. gat and H gcn Subsequently, in order to combine the advantages of the GAT layer and the GCN layer, this section adopts a gated fusion method, which adjusts the weighted fusion ratio of the GAT layer output and the GCN layer output through a learnable gating coefficient γ, as shown below:

[0160] H fused =γ·H gat +(1-γ)·H gcn

[0161] Among them, H fused is the output of the adaptive dynamic graph learning module. Through the gating mechanism, the model can adaptively adjust the fusion ratio of the multi-head graph attention layer and the graph convolution layer in different situations, thereby better utilizing topological information and optimizing the information transmission and fusion process.

[0162] Wherein, the step S3 includes the following steps:

[0163] S31: In the distribution network operation scenario set by this application, only some branches are equipped with real-time measurement devices. How to effectively extract the real-time measurement data features of the branches and integrate them with the historical load data features of the nodes becomes a key issue. To this end, this application adopts a real-time measurement feature extraction method based on a multilayer perceptron (MLP), and maps the extracted features to the nodes of the entire network and integrates them with the historical load features of the nodes, ultimately achieving distribution network state estimation.

[0164] The real-time measurement data in the distribution network can be expressed as a two-dimensional matrix D∈R K×M, where K represents the number of real-time measurement branches and M is the feature dimension of each branch measurement data. To extract effective features from the measurement data, MLP is used to perform nonlinear mapping on it, as shown in the formula:

[0165] D f =σ2(σ1(D·w d,1 +b1)·w d,2 +b2)

[0166] Among them, D f ∈R K×C is the output feature of MLP; C is the feature dimension extracted by MLP; w d,1 ∈R M×2C 、w d,2 ∈R 2C×C are the weight matrices of the input layer and the hidden layer respectively; b1, b2∈R C is the bias term; σ1 and σ2 are ReLU activation functions.

[0167] Since the real-time measurement data only covers part of the branches, and the historical load characteristics cover N nodes in the entire network, in order to achieve multi-source data fusion, we first introduce a linear mapping to convert the branch real-time measurement data characteristics D obtained from MLP into f Mapping to all network nodes is as follows:

[0168]

[0169] Among them, w d,3 ∈R K×N is the mapping matrix, D N ∈R C×N It is the real-time measurement feature mapped on all nodes in the network.

[0170] Secondly, in order to effectively combine the node historical load characteristics H fused Real-time measurement characteristics of branch D N In this section, an adaptive attention weighting mechanism is designed to dynamically adjust the weights during the fusion process by calculating the relative importance between historical load and real-time measurement. The adaptive attention weight α is shown in the following formula:

[0171]

[0172] Among them, σ4 is the Sigmoid activation function; w d,4 , b4 are the weight matrix and bias term of adaptive attention respectively.

[0173] Again, the adaptive weight α is used to adjust the historical node feature H fused With real-time measurement feature D N Perform weighted fusion:

[0174] F=α·H fused +(1-α)·D N

[0175] Where F∈R C×N is the feature representation after multi-source data fusion. Through this weighting mechanism, the fusion of historical data and real-time measurement data can more accurately reflect the current status of the distribution network.

[0176] Finally, the fused features are further processed by linear mapping, as shown in the following formula:

[0177] Y=F T w d,5

[0178] Among them, w d,5 ∈R C×1 is a linear mapping matrix; Y∈R N×1 Through the above steps, the features extracted from the node historical load data and the branch real-time measurement data are successfully fused to obtain the final distribution network state estimation result.

[0179] Case Analysis

[0180] The test system of the present invention is a 33-node distribution network, and the system structure is as follows: Figure 2 shown.

[0181] The main algorithms for comparison include graph convolutional networks (GCN), graph attention networks (GAT), long short-term memory networks (LSTM), and gated recurrent units (GRU). Mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean square error (RMSE) are used:

[0182]

[0183] Among them, y i is the true value of the i-th sample, is the estimated value of the i-th sample, N is the number of time steps.

[0184]

[0185] Among them, y i is the true value of the i-th sample, is the estimated value of the i-th sample, N is the number of time steps.

[0186]

[0187] Among them, y i is the true value of the i-th sample, is the estimated value of the i-th sample, N is the number of time steps.

[0188] The error comparison results in Table 1 demonstrate that our proposed method comprehensively outperforms traditional methods in the IEEE 33-node distribution network state estimation task. For voltage amplitude estimation, our proposed method achieves significantly higher MAE (0.2627×10⁻³), MAPE (0.0272%), and RMSE (0.3455×10⁻³), respectively, compared to the next-best GAT model. For voltage phase angle estimation, our proposed method achieves MAE (0.4860×10⁻³) and MAPE (0.6412%) reductions of 34.3% and 51.8%, respectively, compared to the best comparison method, the GRU.

[0189] Table 1 Comparison of state estimation errors of different methods for IEE33 nodes

[0190]

[0191] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A distribution network state estimation method based on an adaptive spatiotemporal graph neural network is characterized by: The following steps are involved: S1: Collect historical and real-time measurement data of the distribution network and pre-process the data to ensure data integrity, thereby providing high-quality input data for the model and improving its estimation accuracy and stability. S2: Perform discrete wavelet transform on historical measurement data to achieve multi-scale decomposition, extract low-frequency and high-frequency components, and build a dual-branch time series fusion module. Through the dynamic attention mechanism and temporal convolutional network, the global trend and local fluctuation characteristics are captured respectively, and the features are efficiently fused with the help of adaptive weights. At the same time, a hybrid graph learning module is designed to combine static physical topology with adaptive dynamic topology, optimize the adjacency matrix, and integrate the graph attention mechanism and graph convolution features to enhance the expression ability of spatiotemporal features. S3: During real-time data processing, a multi-layer perceptron (MLP) is used to extract branch measurement features and map them to nodes across the entire network, enabling dynamic integration of historical and real-time data, thereby improving the real-time performance and accuracy of distribution network state estimation.

2. The distribution network state estimation method based on adaptive spatiotemporal graph neural network according to claim 1 is characterized in that: Step S1 includes the following steps: S11: In the distribution network, historical data and real-time data are collected from N nodes, covering T time steps. Each node contains M-dimensional features. The observation vector of node i at time t is expressed as: The historical data tensor of the entire network is: X = [X (1) ,X (2) ,…,X (T) ]∈R T×N×M ,in is the state matrix of the entire network at time t, S12: In the distribution network sensor data, the time neighborhood linear interpolation method is used to fill the missing values. For node i in the feature m dimension, the missing data at time t The interpolation calculation is as follows: Among them, t (t-1) is the previous valid timestamp of the missing time point t, t (t+1) is the next valid timestamp after the missing time point t, and is the effective measurement value of the adjacent moments. S13: There may be outliers in the distribution network data, which need to be detected and corrected in a timely manner. For each node i’s feature m, the statistics are calculated based on the time dimension: Among them, μ i,m is the mean of feature m of node i in the time series, reflecting the long-term steady-state level, σ i,m is the standard deviation, which represents the range of data fluctuation. T-1: the sample standard deviation is calculated using unbiased estimation (Bessel correction). Determine outliers based on the 3σ criterion: in, is the value of feature m of node i at time t, if Usually set γ = 3, then the judgment For outliers, Replace with the local window mean: Among them, K is the window radius, the default K = 3, covering the three time steps before and after, a total of 7 points, S14: Standardize the data. The goal of standardization is to convert the data to a fixed range, thereby improving the stability and convergence speed of the model. For the feature m of node i, the standardization calculation is as follows: Among them, x i,m,min is the local minimum value of feature m of node i in its time series, X i,m,max is the local maximum of feature m of node i in its time series.

3. The distribution network state estimation method based on adaptive spatiotemporal graph neural network according to claim 1 is characterized in that: Step S2 includes the following steps, S21: First, use DWT to analyze the input node historical load data X=[x1,x2,...,x T ]∈R M×T×N Decomposition is performed, where M is the number of node variables; T is the length of the historical load sequence; N is the number of distribution network nodes. DWT realizes time-frequency analysis of time series signals through multi-level decomposition. In the k-th layer decomposition process, the time series signal is decomposed into frequency representations of different scales using low-pass filters L and high-pass filters H, so that the model can simultaneously process the short-term and long-term characteristics of time series data. Specifically, the output of the k-th layer decomposition is the low-frequency approximation coefficient CL (k) and high frequency detail coefficient CH (k) ,Right now: CL (k) ,CH (k) =DWT(CL (k-1) ) ↓ Initial conditions, CL (0) =X is the original signal, and each layer of decomposition is achieved through convolution and downsampling: Among them, s = 1, 2, ..., M / 2 represents the index after downsampling; M is the CL after k-1 layer decomposition (k-1) The length of DWT, under the dual-scale effect of low-pass filter L and high-pass filter H, the number of samples of the output signal is the original signal CL (k-1) half, After k-layer decomposition, DWT outputs a set of k+1 coefficients: C=[CH (1) ,CH (2) ,…,CH (k) ,CL (k) ] In order to effectively extract the long-term and short-term dynamic information from the historical load, a frequency band selective mask function is designed to separate the long-term and short-term dynamic features based on the multi-scale coefficient set obtained by k-layer wavelet decomposition. Specifically, a high-frequency mask M is constructed. L =[0,0,...,0,1], M L Represents the mask matching the C dimension, through M L (C)=C⊙M L Achieve progressive suppression of high-frequency coefficients and retain low-frequency components CL (k) , design complementary low-frequency mask M H =[1,1,...,1,0], M H Represents the mask matching the C dimension, through M H (C)=C⊙M H Extract high-frequency features CH (k) , where ⊙ is the Hadamard product, On this basis, the inverse discrete wavelet transform (IDWT) is used to reconstruct the processed low-frequency and high-frequency coefficients to extract the heterogeneous mode of the signal. This process performs upsampling and convolution operations on the low-frequency and high-frequency coefficients, and finally obtains the reconstructed low-frequency time domain signal and high-frequency time domain signal X. L ,X H ∈R M×T×N , which correspond to the time domain characteristics of the node historical load at different scales. The specific reconstruction process is expressed as: X L =IDWT(M L (C)) X H =IDWT(M H (C)) Among them, IDWT represents inverse wavelet transform, M L (C) and M H (C) is the low-frequency coefficient after masking and the high-frequency coefficient after processing; Through the multi-scale feature extraction and reconstruction strategy based on wavelet decomposition, the effective separation of long-term and short-term dynamic characteristics of historical loads and the fine representation of time domain information are achieved; S22: Based on the results of the multi-scale time feature decomposition module, a dual-branch time series fusion module is designed. Specifically, for the long-term trend component X output by the multi-scale time feature decomposition module, L ∈R M×T×N and the short-term volatility component X H ∈R M×T×N , where M is the number of node variables, T is the length of the historical load sequence, and N is the number of distribution network nodes. These are input into the dynamic attention time mechanism and the temporal convolutional network (TCN) respectively to achieve independent modeling of long-term and short-term features. To further enhance the collaborative modeling capability of multi-scale features, an adaptive feature fusion mechanism is designed to dynamically adjust the weight distribution of features at different time scales. Dynamic attention time mechanism: The dynamic attention time mechanism mainly includes the correlation modeling of variable domain, node domain and global time correlation modeling. In the variable domain, a two-dimensional channel convolution is introduced to calculate the weighted combination between variables: Where w∈R M×1×1 is the unbiased convolution kernel weight. Since the size of the convolution kernel is (1,1), the weighted operation is only performed on the feature dimension. Therefore, the output historical load sequence length T and the number of distribution network nodes N remain unchanged, and the final output is Represents the characteristics of each variable at different time steps, In the node domain, the input value is first transformed in dimension, and then a two-dimensional channel convolution is used to perform weighted calculations on the relationships between different nodes: Among them, X L ∈R M×T×N is a low-frequency time domain signal, and permute is a dimension conversion function that converts the 0th dimension and the 2nd dimension to obtain w n is the learnable weight coefficient, is the low-frequency time domain signal of the nth node, is the node domain output, which represents the characteristics of each node at different time steps. In getting the variable domain characteristics and node domain characteristics Finally, an adaptive temporal attention mechanism is used to enable the model to focus on key time steps by learning the dynamic weight distribution between time steps, thereby improving the ability of time-dependent modeling. First, in order to quantify the mutual influence between time steps, a trainable parameter w1∈R is introduced N×M Learn the mapping relationship between variable domain and node domain features, and calculate the attention weights between time steps: Where A∈R T×T is the time attention weight matrix, describing the degree of correlation between different time steps, Secondly, we introduce a trainable transformation matrix w2∈R T×1 Adjust the weight relationship between different time steps to enhance the influence of the current time step. Then, use Softmax normalization to obtain a new time attention weight matrix Ensure that the distribution of time attention is more stable and reasonable: Finally, the information on the time axis is weighted and aggregated through the Einstein summation convention, so that the model can more effectively integrate global time-dependent information and improve the time series modeling capability, and the final output result is X ATT ∈R M×1×N The specific calculation formula is: Temporal Convolutional Network: Use TCN to classify the short-term fluctuation component X H ∈R M×T×N To extract features, first, reorganize X into H The number of node variables M in the distribution network is combined with the number of distribution network nodes N to obtain To adapt to the subsequent one-dimensional convolution operation, the input time series is subjected to feature extraction using TCN to obtain the output X TCN ∈R(M×N)×1, Adaptive weight feature fusion mechanism: Adaptive weight mechanism to achieve trend component X ATT and the fluctuation component X TCN Dynamic fusion, for this purpose, the learnable parameter α is used to achieve dynamic weight distribution: a sigmoid =Sigmoid(a) X fusion =a sigmoid ·X ATT +(1-a sigmoid )·X TCN Among them, X fusion ∈R M×N is the fusion output result, S23: In the distribution network state estimation task, the details are as follows: First, the adjacency matrix A of the distribution network topology adj First, after normalization, the normalized matrix is ​​obtained The calculation formula is: Where D is the degree matrix, then construct the trainable node embedding matrix E1, E2∈R N×d , and through the linear transformation layer parameter w e,1 、w e,2 ∈R N×d And the nonlinear activation function generates the dynamic association matrix A graph , M1=tanh(αE1w e,1 ) M2=tanh(αE2w e,2 ) Where N is the number of nodes in the distribution network, d is the dimension of the node embedding matrix, and α is a hyperparameter that controls the node saturation rate to avoid gradient vanishing. Finally, a weighted fusion strategy is adopted to combine the static topology information and the dynamic association matrix to generate the final adjacency matrix: Among them, λ is a learnable balance coefficient, which is used to adjust the fusion ratio between static topology and dynamic topology. By introducing a learnable mechanism, the model can adaptively balance physical priors and data-driven information in different scenarios, thereby improving the accuracy and robustness of distribution network state estimation. After obtaining the dynamic graph structure A adaptive After that, the correlation between the nodes of the distribution network is first captured by the GAT layer and the GCN layer, and the output H is obtained. gat and H gcn , Subsequently, in order to combine the advantages of the GAT layer and the GCN layer, a gated fusion method is adopted. A learnable gating coefficient γ is used to adjust the weighted fusion ratio of the GAT layer output and the GCN layer output, as shown below: H fused =γ·H gat +(1-γ)·H gcn Among them, H fused For the output of the adaptive dynamic graph learning module, through the gating mechanism, the model can adaptively adjust the fusion ratio of the multi-head graph attention layer and the graph convolution layer in different situations, so as to better utilize the topological information and optimize the information transmission and fusion process.

4. The distribution network state estimation method based on adaptive spatiotemporal graph neural network according to claim 1 is characterized in that: Step S3 includes the following steps: S31: A real-time measurement feature extraction method based on a multilayer perceptron (MLP) is used, and the extracted features are mapped to the nodes of the entire network and integrated with the node historical load features to finally achieve distribution network state estimation. The real-time measurement data in the distribution network is represented as a two-dimensional matrix D∈R K×P , where K represents the number of real-time measurement branches, P is the number of variables in the measurement data of each branch, and in order to extract effective features from the measurement data, MLP is used to perform nonlinear mapping on it, as shown in the formula: D f =σ2(σ1(D·w d,1 +b1)·w d,2 +b2) Among them, D f ∈R K×C is the output feature of MLP; C is the feature dimension extracted by MLP; w d,1 ∈R P×2C 、w d,2 ∈R 2C×C are the weight matrices of the input layer and the hidden layer respectively; b1, b2∈R C is the bias term; σ1 and σ2 are ReLU activation functions, First, a linear mapping is introduced to convert the branch real-time measurement data feature D obtained from MLP f Mapping to all network nodes is as follows: Among them, w d,3 ∈R K×N is the mapping matrix, D f Transpose, D N ∈R C×N To map the real-time measurement characteristics on all network nodes, Secondly, in order to effectively combine the node historical load characteristics H fused Real-time measurement characteristics of branch D N , an adaptive attention weighting mechanism is designed to achieve dynamic adjustment of weights in the fusion process by calculating the relative importance between historical load and real-time measurement. The adaptive attention weight β is shown in the following formula: Among them, σ4 is the Sigmoid activation function; w d,4 , b4 are the weight matrix and bias term of adaptive attention respectively, Again, the adaptive weight β is used to adjust the historical node feature H fused With real-time measurement feature D N Perform weighted fusion: F=β·H fused +(1-β)·D N Where F∈R C×N It is the feature representation after multi-source data fusion. Through this weighting mechanism, the fusion of historical data and real-time measurement data can more accurately reflect the current state of the distribution network. Finally, the fused features are further processed by linear mapping, as shown in the following formula: Y=F T ·w d,5 Among them, w d,5 ∈R C×1 is a linear mapping matrix; Y∈R N×1 As the output, through the above steps, the features extracted from the node historical load data and the branch real-time measurement data are successfully fused to obtain the final distribution network state estimation result.

5. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, it implements the distribution network state estimation method based on the adaptive spatiotemporal graph neural network as described in any one of claims 1 to 4 above.

6. A computer-readable storage medium having computer instructions stored thereon, characterized in that: When the computer instructions are executed by a processor, the distribution network state estimation method based on the adaptive spatiotemporal graph neural network as described in any one of claims 1 to 4 is implemented.

Images