Power distribution network state estimation method and device based on improved graph neural network

Abstract

The invention discloses a power distribution network state estimation method and device based on an improved graph neural network, and the method comprises the steps: carrying out the feature extraction of input power distribution network measurement data through constructing a multi-scale convolutional neural network module, and obtaining multi-level features; a multi-head attention mechanism is introduced to calculate correlation weights among the multi-level features, and a weight matrix is obtained; performing weighted adjustment on the multi-level features according to the weight matrix to obtain fusion features; iteratively updating the state of the power grid node through the gating mechanism graph neural network; constructing a power distribution network state estimation model based on MHA-MSCNN-GGNN based on a multi-scale convolutional neural network module, a multi-head attention mechanism and a gating mechanism graph neural network; and analyzing and evaluating the input target measurement data of the power distribution network through the power distribution network state estimation model, and outputting the state of the power distribution network. According to the method, the space-time dependency relationship of the nodes and the edges in the power distribution network can be effectively captured while the expression capability of the complex network state is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of distribution network state estimation, and in particular to a distribution network state estimation method and device based on an improved graph neural network. Background Art

[0002] As power grids become increasingly intelligent, the safe and stable operation of distribution networks, a key link directly serving users, is crucial and significantly impacts the reliability and power quality of the entire power system. The large-scale integration of distributed energy resources, such as wind and photovoltaic power generation, into the power grid has led to increasingly complex distribution network operations. State estimation, a supporting technology for real-time monitoring, fault diagnosis, and optimized dispatching of distribution networks, ensures safe and stable operation, optimizes distribution network planning and construction, and enhances the management of distributed energy access. It also supports intelligent control of distribution networks.

[0003] Currently, three typical approaches are used in the field of distribution network state estimation: analytical methods based on physical models, data-driven intelligent algorithms, and hybrid approaches that combine the two. The weighted least squares (WLS) method is the most mature method for static state estimation in distribution networks. Under ideal conditions, it offers high estimation accuracy and computational efficiency. However, with the widespread integration of renewable energy sources (such as photovoltaic and wind power), the instability and volatility of their output have increased the uncertainty of distribution network state estimation. Furthermore, due to errors in distribution network measurement equipment, packet loss during communication, and external environmental factors (such as climate change and temperature fluctuations), the data input to the state estimation model often contains noise or even corrupted data. Traditional model-based state estimation algorithms are susceptible to abnormal data, resulting in large estimation errors and inaccurate reflection of the actual operating state of the power system. While data-driven intelligent algorithms address computational efficiency and error correction issues in large-scale network state estimation, most data-driven models fail to consider the topology of the distribution network, resulting in insufficient interpretability and difficulty in establishing a relationship between system state and network structure.

[0004] Therefore, how to invent a distribution network state estimation method based on an improved graph neural network that can effectively capture the spatiotemporal dependencies between nodes and edges in the distribution network while improving the ability to express complex network states has become an urgent problem to be solved. Summary of the Invention

[0005] To this end, the present invention provides a distribution network state estimation method and device based on an improved graph neural network. By constructing a distribution network state estimation model based on MHA-MSCNN-GGNN, the topological structure of the distribution network and historical measurement data are combined, and the features of the measurement data are preliminarily extracted at multiple levels through a multi-scale convolutional neural network. The MHA is used to enhance the model's attention to the characteristics of different nodes, thereby improving the model's ability to express complex network states. At the same time, the GGNN is used to effectively capture the spatiotemporal dependencies between nodes and edges in the distribution network.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a distribution network state estimation method based on an improved graph neural network, comprising:

[0007] Constructing a multi-scale convolutional neural network module; extracting features from input distribution network measurement data through the multi-scale convolutional neural network module to obtain multi-level features of power grid nodes;

[0008] Based on the multi-level features, a multi-head attention mechanism is introduced; the correlation weights between the multi-level features are calculated by the attention heads of the multi-head attention mechanism to obtain a weight matrix; the multi-level features are weighted and adjusted according to the weight matrix, and feature splicing processing is performed to obtain a fusion feature;

[0009] Inputting the fused features into a gated graphical neural network; dynamically controlling the propagation and fusion of grid node information over time steps through the update gate and reset gate mechanisms in the gated graphical neural network, and iteratively updating the states of the grid nodes;

[0010] Based on the multi-scale convolutional neural network module, the multi-head attention mechanism and the gating mechanism graphical neural network, a distribution network state estimation model based on MHA-MSCNN-GGNN is constructed; the input distribution network target measurement data is analyzed and evaluated by the distribution network state estimation model based on MHA-MSCNN-GGNN, and the distribution network state is output.

[0011] As a preferred solution for a distribution network state estimation method based on an improved graph neural network, the multi-scale convolutional neural network module is provided with a plurality of convolution branches; the convolution branches are configured with convolution kernels of set sizes; small convolution kernels are used to capture local detail features and instantaneous fluctuation patterns of measurement data; medium convolution kernels are used to extract medium-range measurement correlation features; and large convolution kernels are used to capture global trend features and long-term dependencies.

[0012] The expression of the multi-level feature is:

[0013] y=α1y1+α2y2+α3y3

[0014] Where y is a multi-level feature; y1, y2, and y3 are features extracted by different convolution branches respectively; α n is the number of weights assigned to different convolution kernel sizes, n = 1, 2, 3.

[0015] As a preferred solution of the distribution network state estimation method based on the improved graph neural network, the expression of the fusion feature is:

[0016] MultiHead(Q,K,V)=concat(head1,...,head h )W O

[0017] head i =Attention(QW i Q ,KW i K ,VW i V )

[0018] Where Q is the query; K is the key; V is the value; QW i Q ,KW i K ,VW i V are the linear transformation matrices of query, key, and value respectively; h is the number of attention heads; W O The linear transformation matrix of the output is (hd v ,d model ),d v is the dimension of the value, d model The dimensions of the model.

[0019] As a preferred solution for the distribution network state estimation method based on the improved graph neural network, the propagation model expression of the gating mechanism graph neural network is:

[0020]

[0021] Where, is the node state vector, representing the hidden state of node v at time t; T is the transpose symbol, which is used to adjust the dimension to meet the requirements of matrix multiplication; b is a learnable bias term that can enhance the nonlinear expression ability of the model; is the aggregated message vector, which represents the messages aggregated by node v from its neighbors; To update the threshold and control the forgotten information; To reset the threshold, control the generation of new information; A v is the weight matrix, is the transformation matrix of neighbor messages; W z ,Wt Both are learnable weight matrices, representing the transformation matrix of aggregated messages in the gating mechanism; U z ,U t Both are learnable weight matrices, representing the transformation matrix of the historical state in the gating mechanism; σ As the activation function, the sigmoid function is usually used in gated graph neural networks.

[0022] As a preferred solution of the distribution network state estimation method based on the improved graph neural network, the root mean square error, mean absolute percentage error and determination coefficient are used to evaluate the distribution network state estimation method based on the improved graph neural network.

[0023] The estimation results of the distribution network state estimation model based on MHA-MSCNN-GGNN are evaluated;

[0024] The calculation formula of the root mean square error is:

[0025]

[0026] Where RMSE is the root mean square error; is the estimate of the i-th state, y i is the corresponding true value, n is the total number of state quantities;

[0027] The calculation formula for the mean absolute percentage error is:

[0028]

[0029] Where, MAPE is the root mean square error;

[0030] The calculation formula of the determination coefficient is:

[0031]

[0032] Where R 2 is the root mean square error; is the mean of the observations.

[0033] The present invention also provides a distribution network state estimation device based on an improved graph neural network, based on the above distribution network state estimation method based on an improved graph neural network, comprising:

[0034] A multi-scale convolutional neural network module construction unit is used to construct a multi-scale convolutional neural network module; the multi-scale convolutional neural network module is used to extract features of the input distribution network measurement data to obtain multi-level features of the power grid nodes;

[0035] A multi-head attention mechanism introduction and processing unit is used to introduce a multi-head attention mechanism based on the multi-level features; calculate the correlation weights between the multi-level features through the attention heads of the multi-head attention mechanism to obtain a weight matrix; perform weighted adjustment on the multi-level features according to the weight matrix, and perform feature splicing processing to obtain fused features;

[0036] a gated graphical neural network processing unit, configured to input the fusion features into the gated graphical neural network; dynamically control the propagation and fusion of grid node information over time steps through the update gate and reset gate mechanisms in the gated graphical neural network, and iteratively update the states of the grid nodes;

[0037] The MHA-MSCNN-GGNN-based distribution network state estimation model construction and processing unit is used to construct an MHA-MSCNN-GGNN-based distribution network state estimation model based on the multi-scale convolutional neural network module, the multi-head attention mechanism and the gated mechanism graphical neural network; the MHA-MSCNN-GGNN-based distribution network state estimation model is used to analyze and evaluate the input distribution network target measurement data and output the distribution network state.

[0038] As a preferred solution of a distribution network state estimation device based on an improved graph neural network, in the multi-scale convolutional neural network module construction unit, the multi-scale convolutional neural network module is provided with a plurality of convolution branches; the convolution branches are configured with convolution kernels of set sizes; small convolution kernels are used to capture local detail features and instantaneous fluctuation patterns of measurement data; medium convolution kernels are used to extract medium-range measurement correlation features; and large convolution kernels are used to capture global trend features and long-term dependencies;

[0039] The expression of the multi-level feature is:

[0040] y=α1y1+α2y2+α3y3

[0041] Where y is a multi-level feature; y1, y2, and y3 are features extracted by different convolution branches respectively; α n is the number of weights assigned to different convolution kernel sizes, n = 1, 2, 3.

[0042] As a preferred solution of the distribution network state estimation device based on the improved graph neural network, the multi-head attention mechanism is introduced into the processing unit, and the expression of the fusion feature is:

[0043] MultiHead(Q,K,V)=concat(head1,...,head h )W O

[0044] head i =Attention(QW i Q ,KW i K ,VW i V )

[0045] Where Q is the query; K is the key; V is the value; QW i Q ,KW i K ,VW i V are the linear transformation matrices of query, key, and value respectively; h is the number of attention heads; W O The linear transformation matrix of the output is (hd v ,d model ),d v is the dimension of the value, d model The dimensions of the model.

[0046] As a preferred solution of a distribution network state estimation device based on an improved graph neural network, in the gate control mechanism graph neural network processing unit, the propagation model expression of the gate control mechanism graph neural network is:

[0047]

[0048] Where, is the node state vector, representing the hidden state of node v at time t; T is the transpose symbol, which is used to adjust the dimension to meet the requirements of matrix multiplication; b is a learnable bias term that can enhance the nonlinear expression ability of the model; is the aggregated message vector, which represents the messages aggregated by node v from its neighbors; To update the threshold and control the forgotten information; To reset the threshold, control the generation of new information; A v is the weight matrix, is the transformation matrix of neighbor messages; W z ,W t Both are learnable weight matrices, representing the transformation matrix of aggregated messages in the gating mechanism; U z ,U t Both are learnable weight matrices, representing the transformation matrix of the historical state in the gating mechanism; σ As the activation function, the sigmoid function is usually used in gated graph neural networks.

[0049] As a preferred solution of the distribution network state estimation device based on the improved graph neural network, in the distribution network state estimation model construction and processing unit based on MHA-MSCNN-GGNN, the estimation results of the distribution network state estimation model based on MHA-MSCNN-GGNN are evaluated by the root mean square error, mean absolute percentage error and determination coefficient;

[0050] The calculation formula of the root mean square error is:

[0051]

[0052] Where RMSE is the root mean square error; is the estimate of the i-th state, y i is the corresponding true value, n is the total number of state quantities;

[0053] The calculation formula for the mean absolute percentage error is:

[0054]

[0055] Where, MAPE is the root mean square error;

[0056] The calculation formula of the determination coefficient is:

[0057]

[0058] Where R 2 is the root mean square error; is the mean of the observations.

[0059] The present invention has the following advantages: the present invention constructs a multi-scale convolutional neural network module; extracts features of input distribution network measurement data through the multi-scale convolutional neural network module to obtain multi-level features of power grid nodes; introduces a multi-head attention mechanism based on the multi-level features; calculates the correlation weights between the multi-level features through the attention heads of the multi-head attention mechanism to obtain a weight matrix; weights are adjusted on the multi-level features according to the weight matrix, and feature splicing processing is performed to obtain fusion features; the fusion features are input into a gated mechanism graphical neural network; the propagation and fusion of power grid node information in time steps are dynamically controlled through the update gate and reset gate mechanism in the gated mechanism graphical neural network, and the state of the power grid nodes is iteratively updated; based on the multi-scale convolutional neural network module, the multi-head attention mechanism and the gated mechanism graphical neural network, a distribution network state estimation model based on MHA-MSCNN-GGNN is constructed; the input distribution network target measurement data is analyzed and evaluated by the distribution network state estimation model, and the distribution network state is output. The GNN-based state estimation model of the present invention reduces the problem of traditional methods' over-reliance on physical models, while avoiding the limitation of pure data-driven methods' lack of physical interpretability. The addition of the gating mechanism enables the model to have stronger dynamic adaptability, and can capture time-varying processes such as load or generator output fluctuations. MSCNN enriches the model's ability to extract node information through different convolution kernel sizes, providing more comprehensive node features for subsequent neural networks. The addition of MHA improves the model's ability to estimate complex network states, making the model more robust and able to provide reliable estimation results even in the face of missing or inaccurate data. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the embodiments or the description of the prior art. Obviously, the drawings described below are merely exemplary, and those skilled in the art can, without inventive effort, derive other implementation drawings based on the provided drawings.

[0061] The structures, proportions, sizes, etc. illustrated in this specification are intended solely to complement the contents disclosed herein and to facilitate understanding and reading by persons skilled in the art. They are not intended to limit the conditions under which the present invention may be implemented and therefore have no substantive technical significance. Any structural modifications, changes in proportions, or adjustments in sizes, without affecting the efficacy and objectives of the present invention, shall remain within the scope of the technical contents disclosed herein.

[0062] Figure 1This is a flow chart of a method for estimating the state of a distribution network based on an improved graph neural network provided in Example 1 of the present invention;

[0063] Figure 2 Schematic diagram of a multi-scale convolution structure in a distribution network state estimation method based on an improved graph neural network provided in Example 1 of the present invention;

[0064] Figure 3 Schematic diagram of a multi-head attention mechanism in a distribution network state estimation method based on an improved graph neural network provided in Example 1 of the present invention;

[0065] Figure 4 This is a flow diagram of a distribution network state estimation model based on MHA-MSCNN-GGNN in a distribution network state estimation method based on an improved graph neural network provided in Example 1 of the present invention;

[0066] Figure 5 A schematic diagram of an IEEE 33 node system in a possible embodiment provided in Example 1 of the present invention;

[0067] Figure 6 This is a schematic diagram comparing the state estimation results of GCN at a certain time section in a possible embodiment provided in Example 1 of the present invention;

[0068] Figure 7 This is a schematic diagram comparing the state estimation results of MSCNN-GGNN at a certain time section in a possible embodiment provided in Example 1 of the present invention;

[0069] Figure 8 This is a schematic diagram comparing the state estimation results of MHA-MSCNN-GGNN at a certain time section in a possible embodiment provided in Example 1 of the present invention;

[0070] Figure 9 This is a schematic diagram of the architecture of a distribution network state estimation device based on an improved graph neural network provided in Example 2 of the present invention. DETAILED DESCRIPTION

[0071] The following describes the implementation of the present invention using specific embodiments. Those skilled in the art will readily understand the other advantages and benefits of the present invention from the disclosure herein. Obviously, the embodiments described are only a portion of the present invention, not all of it. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are intended to fall within the scope of protection of the present invention.

[0072] Example 1

[0073] See also Figure 1Embodiment 1 of the present invention provides a distribution network state estimation method based on an improved graph neural network, comprising the following steps:

[0074] S1. Construct a multi-scale convolutional neural network module; extract features from input distribution network measurement data through the multi-scale convolutional neural network module to obtain multi-level features of power grid nodes;

[0075] S2. Based on the multi-level features, a multi-head attention mechanism is introduced; the correlation weights between the multi-level features are calculated by the attention heads of the multi-head attention mechanism to obtain a weight matrix; the multi-level features are weighted and adjusted according to the weight matrix, and feature splicing is performed to obtain a fusion feature;

[0076] S3. Inputting the fused features into a gated graphical neural network; dynamically controlling the propagation and fusion of grid node information over time steps through the update gate and reset gate mechanisms in the gated graphical neural network, and iteratively updating the states of the grid nodes;

[0077] S4. Based on the multi-scale convolutional neural network module, the multi-head attention mechanism and the gating mechanism graph neural network, a distribution network state estimation model based on MHA-MSCNN-GGNN is constructed; the input distribution network target measurement data is analyzed and evaluated by the distribution network state estimation model based on MHA-MSCNN-GGNN, and the distribution network state is output.

[0078] In this embodiment, in step S1, a multi-scale convolutional neural network module is constructed; the multi-scale convolutional neural network module is used to extract features of the input distribution network measurement data to obtain multi-level features of the power grid nodes;

[0079] Specifically, a multi-scale convolutional neural network module (MSCNN) is constructed, and the MSCNN is used to extract the features of the input information. By designing a parallel multi-branch convolutional structure, a multi-level feature representation of the distribution network measurement data is achieved. For each convolution branch, convolution kernels of different sizes are configured. Small convolution kernels focus on capturing local detail features and instantaneous fluctuation patterns of the measurement data; medium convolution kernels can extract medium-range measurement correlation features; and large convolution kernels are responsible for capturing global trend features and long-term dependencies. Each convolution branch uses the same deep architecture but different receptive fields to ensure that the network can understand the dynamic characteristics of the measurement data at different scales at the same time. At the same time, a Bottleneck structure with residual connections is used within each branch, and 1×1 convolution is used to achieve dimensionality increase and reduction of feature channels, thereby improving computational efficiency while ensuring feature expression capabilities. This multi-granularity feature extraction mechanism provides a richer feature representation for subsequent graph neural network processing.

[0080] In this embodiment, Figure 2 As shown in Figure 1, MSCNN deeply mines the features of input information to obtain its hidden feature information. Through a parallel multi-branch architecture, cross-scale feature extraction of input information is achieved, and operational features at different granularities can be captured simultaneously. Since the data to be processed is feature information rather than image information, a one-dimensional CNN is used for feature extraction. The detailed parameters of each branch of the multi-scale convolutional neural network module are shown in Table 1:

[0081]

[0082] Table 1 Multi-scale convolutional neural network module branch parameters

[0083] In this example, three independent encoders (Encoder1, Encoder2, and Encoder3) process the input features independently. Each branch uses a different convolution kernel size, which can be adjusted based on the number of nodes to accommodate distribution networks of varying sizes. Each branch contains multiple Bottleneck layers, which aim to reduce computational effort by compressing and expanding channel dimensions, enhancing nonlinearity while preserving feature representation.

[0084] Assume the input is (where B is the batch size, C is the number of channels, and L is the sequence length), the outputs of the three branches are:

[0085]

[0086] Where, Conv k×1 is a 1D convolution with kernel size k, Represents three cascaded Bottleneck layers.

[0087] Finally, multi-scale features are obtained by additive fusion:

[0088] y=α1y1+α2y2+α3y3

[0089] Where y is a multi-level feature; y1, y2, and y3 are features extracted by different convolution branches respectively; α n is the number of weights assigned to different convolution kernel sizes, n = 1, 2, 3.

[0090] In this embodiment, in step S2, a multi-head attention mechanism is introduced based on the multi-level features; the correlation weights between the multi-level features are calculated by the attention heads of the multi-head attention mechanism to obtain a weight matrix; the multi-level features are weighted and adjusted according to the weight matrix, and feature splicing processing is performed to obtain a fusion feature;

[0091] Specifically, the introduction of the attention mechanism module can enhance the model's ability to extract features from key nodes in the distribution network, suppress invalid feature information, and thus improve the efficiency of the model. The present invention adopts a multi-head attention mechanism to improve the efficiency of CNN, because the node features of the distribution network may include a variety of electrical measurement data such as active power and reactive power. The multi-head attention mechanism enables the model to focus on key information on different electrical feature dimensions at the same time by deploying multiple independent attention heads in parallel. This multi-perspective feature learning method enables the model to have a more comprehensive understanding of the operating status of the distribution network. During the feature processing process, each attention head calculates the correlation weights between node features, and after softmax normalization, strengthens the representation of important features (such as voltage fluctuations at key nodes) while weakening the influence of noise and redundant information (such as measurement errors). This dynamic feature selection mechanism makes the model robust to the problems of partial measurement missing or uneven data quality that may occur in the distribution network.

[0092] In this embodiment, Figure 3 As shown in Figure 2, the introduction of the attention mechanism module can dynamically weight the feature information of key measurement nodes in the distribution network, thereby effectively suppressing noise interference in state estimation. The calculation formula of the single-head attention mechanism is:

[0093]

[0094] Where Q is the query; K is the key; V is the value; d k is the dimension of the key, which determines the scale of the attention calculation, the scaling factor, and the representation capacity of each attention head; T is the transpose symbol, which is used to adjust the dimension to meet the requirements of matrix multiplication.

[0095] MHA computes multiple attention heads in parallel. For each head, it calculates a separate attention output, then concatenates the outputs of all heads together and obtains the final result after a linear transformation:

[0096] MultiHead(Q,K,V)=concat(head1,...,head h )W O

[0097] head i =Attention(QW i Q ,KW i K ,VW i V )

[0098] In the formula, QW i Q ,KW iK ,VW i V are the linear transformation matrices of query, key, and value respectively; h is the number of attention heads; W O The linear transformation matrix of the output is (hd v ,d model ),d v is the dimension of the value, d model The dimensions of the model.

[0099] Because the inputs for distribution network state estimation are typically node active and reactive power, the model's multi-head attention mechanism uses two parallel heads for attention calculation. This multi-head attention mechanism not only enhances the model's ability to extract data features but also improves its ability to model long-range dependencies and global information, making it particularly suitable for state estimation in complex, nonlinear systems such as distribution networks.

[0100] In this embodiment, in step S3, the fused features are input into a gated graphical neural network; through the update gate and reset gate mechanism in the gated graphical neural network, the propagation and fusion of the grid node information in the time step are dynamically controlled, and the state of the grid node is iteratively updated;

[0101] GGNN is a graph neural network based on a gating mechanism, designed to process node feature propagation and state updates in graph-structured data. Its core concept is to introduce the concept of the gated recurrent unit (GRU) into graph neural networks. Through the update gate and reset gate mechanisms, it dynamically controls the propagation and fusion of node information at different time steps. Local features extracted by MS-CNN and global correlation features enhanced by MHA are input into the GGNN for topological information propagation. The update gate controls the proportion of historical state information retained, and the reset gate determines the degree of fusion of neighborhood features. This gating mechanism can capture the spatiotemporal characteristics of node states in distribution networks.

[0102] Specifically, the fused features are fed into a gated graph neural network. Information is propagated between nodes in the graph via a message passing mechanism, and gating units are used to update node features. Data such as power and voltage in the distribution network vary over time, exhibiting temporal regularities. GGNN can capture these time series features and, by analyzing graph-structured data at different moments, improve the accuracy and reliability of state estimation in dynamic environments.

[0103] The propagation model expression of the gating mechanism graph neural network is:

[0104]

[0105] Where, is the node state vector, representing the hidden state of node v at time t; T is the transpose symbol, which is used to adjust the dimension to meet the requirements of matrix multiplication; b is a learnable bias term that can enhance the nonlinear expression ability of the model; is the aggregated message vector, which represents the messages aggregated by node v from its neighbors; To update the threshold and control the forgotten information; To reset the threshold, control the generation of new information; A v is the weight matrix, is the transformation matrix of neighbor messages; W z ,W t Both are learnable weight matrices, representing the transformation matrix of aggregated messages in the gating mechanism; U z ,U t Both are learnable weight matrices, representing the transformation matrix of the historical state in the gating mechanism; σ As the activation function, the sigmoid function is usually used in gated graph neural networks.

[0106] In this embodiment, in step S4, a distribution network state estimation model based on MHA-MSCNN-GGNN is constructed based on the multi-scale convolutional neural network module, the multi-head attention mechanism and the gating mechanism graph neural network; the input distribution network target measurement data is analyzed and evaluated by the distribution network state estimation model based on MHA-MSCNN-GGNN, and the distribution network state is output.

[0107] The distribution network state estimation model based on the MHA-MSCNN-GGNN architecture comprises a multi-head attention mechanism (MHA), a multi-scale convolutional neural network module (MSCNN), and a gating-based graph neural network (GGNN). The MSCNN module extracts multi-level features from raw measurement data, achieving a multi-dimensional representation of the distribution network's operating status through a parallel convolutional architecture. This hierarchical feature extraction mechanism provides information-rich and dimensionally regular feature representations for subsequent attention mechanism and graph neural network processing. MHA, as a data enhancement step, simultaneously focuses on different types of distribution network data features, uncovering hidden characteristic patterns in power data. This enables the model to accurately capture the complex mapping relationship between power and voltage even when faced with large amounts of complex power data. The GGNN effectively addresses the problem of previous data-driven models that ignore the actual topology of the distribution network and lack physical interpretability. Furthermore, the GGNN uses gated neural units to deeply mine the temporal correlations of the measurement data, capturing trends and periodicity in the temporal dimension, complementing the spatial feature information previously obtained through the MSCNN and MHA.

[0108] Specifically, such as Figure 4As shown in the figure, a distribution network state estimation model based on the multi-head attention mechanism (MHA), a multi-scale convolutional neural network module (MSCNN), and a gating mechanism graph neural network (GGNN) is constructed. The MHA-MSCNN-GGNN-based distribution network state estimation model analyzes and evaluates the input distribution network target measurement data, and outputs the distribution network state after processing through the pooling layer, the fully connected layer, and the ReLU activation function layer.

[0109] In this embodiment, the root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R 2 ,R-Squared) is used as an indicator to evaluate and compare the effects of different models.

[0110] The root mean square error (RMSE) is used to measure the absolute deviation between the estimated value and the true value, and its calculation formula is:

[0111]

[0112] Where RMSE is the root mean square error; is the estimate of the i-th state, y i is the corresponding true value, n is the total number of state quantities;

[0113] The mean absolute percentage error (MAPE) is used to reflect the relative size of the estimation error, which can reflect the ratio of the error of the predicted value to the observed value. Its calculation formula is:

[0114]

[0115] Where, MAPE is the root mean square error;

[0116] Coefficient of determination R 2 It can measure the similarity of the model's fit to the data, and its calculation formula is:

[0117]

[0118] Where R 2 is the root mean square error; y is the mean of the observations. R 2 The value range of is [0,1]. The closer it is to 1, the better the model fitting effect is.

[0119] Through the above three evaluation indicators, RMSE provides absolute error measurement, MAPE focuses on relative error evaluation, and R 2The reliability of the model is verified from a statistical perspective, and the combination of the three can more comprehensively estimate and compare the model performance.

[0120] In one possible embodiment, a state estimation algorithm based on a standard IEEE 33-node system is provided as follows:

[0121] The topology of the IEEE 33-node system is as follows Figure 5 As shown. A simulation dataset was constructed using New York City electricity load data. 92 days of historical load data from October 1, 2024, to December 31, 2024, were selected as the actual node load, with a data interval of 5 minutes (288 sampling points per day). Power flow calculations were performed using the Python package Pandapower, obtaining a total of 25,920 sets of time section data. 80% of the dataset data was used as the training set, and 20% as the validation set. A continuous partitioning method was used in the time series (the first 80% period was used for training, and the last 20% was used for validation) to better simulate the time series prediction scenario during actual deployment. All input features were Min-Max normalized to the range [0, 1], and the output target values retained their original dimensions for intuitive evaluation of the prediction error.

[0122] The GCN and MSCNN-GGNN models were selected for comparison with the MHA-MSCNN-GGNN model. The GCN model adopts the classic graph convolution structure and realizes the topological propagation of node features through the adjacency matrix; the MSCNN-GGNN model introduces a multi-scale convolutional encoder and a gated graph neural network based on the GCN. The 3 / 5 / 7 convolution kernels set in parallel capture node features in different ranges respectively, and the GGNN layer realizes dynamic feature updates through reset gates and update gates.

[0123] All three models use the same hyperparameters and optimizer configurations, as shown in Table 2:

[0124] parameter Numerical parameter Numerical Initial learning rate 0.001 Batch size 256 Attenuation coefficient #1 0.9 Optimization solver AdamW Attenuation Coefficient #2 0.999 Maximum number of iterations 200

[0125] Table 2 Main hyperparameter settings of the model

[0126] Compared with the standard Adam optimizer, the AdamW optimizer can more effectively prevent overfitting by decoupling weight decay and gradient updates. The two decay coefficients control the exponential decay rates of the first-order and second-order moment estimates, respectively. The classic combination of 0.9 / 0.999 is suitable for most non-stationary distribution network measurement data.

[0127] IEEE33 node state estimation results and analysis:

[0128] The accuracy of the three models in estimating voltage amplitude and phase angle is shown in Table 3:

[0129]

[0130] Table 3 IEEE33 node evaluation indicators

[0131] Table 3 shows that for voltage amplitude, both the MSCNN-GGNN and MHA-MSCNN-GGNN achieve high accuracy. The proposed model's RMSE decreases by approximately 4.12% compared to the model without MHA, and its MAPE decreases by approximately 5.11%. While GCN still maintains good fitting capabilities, its error level is significantly higher than the previous two. This performance difference is even more pronounced in the voltage phase angle estimation task. While more complex models often imply higher model complexity, the runtime differences among the three models are similar in terms of computational efficiency, indicating that the increased model complexity does not impose a significant computational burden.

[0132] In order to more intuitively compare the state estimation accuracy of different models, Figure 6 、 Figure 7 、 Figure 8 The comparison diagrams of the state estimation results of the three models at the same time section are shown respectively; Figure 6 As shown in , the GCN model shows obvious estimation instability in some node areas; Figure 7 As shown in Figure 2, the node change trend of MSCNN-GGNN is consistent with the true value, but there is still a certain amount of error; Figure 8 As shown in Figure 3, MHA-MSCNN-GGNN has the best fitting effect, and the predicted values are almost consistent with the true values except for a few nodes.

[0133] IEEE33 node state estimation results and analysis in the presence of noise:

[0134] To improve the model's noise immunity, we generated Gaussian noise with a mean of zero and standard deviations of 10% and 20% of the original standard deviation based on the standard deviation of the measured data, simulating the measurement errors and communication noise present in actual distribution networks. The model accuracy after adding noise is shown in Tables 4 and 5:

[0135]

[0136] Table 4 IEEE33 node evaluation indicators under 10% Gaussian noise

[0137]

[0138] Table 5 IEEE33 node evaluation indicators under 20% Gaussian noise

[0139] By analyzing the model's performance under Gaussian noise of varying intensities, we observed a clear model-dependent effect of noise level on state estimation accuracy. As noise intensity increased, the accuracy of each model decreased. The MHA-MSCNN-GGNN achieved minimal errors in both voltage amplitude and phase angle, maintaining good fitting performance even in high noise conditions. Under 20% Gaussian noise, the MAPE error for voltage amplitude was reduced by 36% compared to the model without MHA, and the voltage phase angle error was reduced by approximately 2.8%.

[0140] IEEE118 node state estimation results and analysis:

[0141] To further verify the generalization performance of the present invention under different power grid scales, this embodiment expands the experimental scenario to a standard IEEE 118-node system for comparative experiments, maintaining the same evaluation indicators as the 33-node system. The experimental results are shown in Table 6:

[0142]

[0143] Table 6 IEEE118 node evaluation indicators

[0144] The analysis results show that the advantages of the present invention can be maintained when the scale of the power grid is expanded. Its voltage amplitude RMSE is reduced by 21.6% compared with MSCNN-GGNN, and the phase angle estimation accuracy is improved by 46.9%. The traditional GCN model shows obvious degradation in the phase angle estimation task. 2 The R value is only 0.508, indicating that the model fitting ability is poor. However, the improved model still maintains an R value of more than 0.998 in the 118-node system. 2 The value of 100% for the 33-node system verifies its good scalability. As the node scale increases, the performance gap between models becomes larger than that of the 33-node system, indicating that complex topology will affect model performance to a certain extent, and the challenges posed by basic graph neural networks are more significant.

[0145] Through case analysis with two control models, GCN and MSCNN-GGNN, experimental results on IEEE33 and IEEE118 nodes demonstrate that the proposed method achieves higher accuracy in voltage amplitude and phase angle. Comparative experiments with Gaussian noise of varying intensities further demonstrate that the proposed method possesses greater robustness and anti-interference capabilities, demonstrating both technical advantages and application value.

[0146] In summary, the present invention has the following advantages: the present invention constructs a multi-scale convolutional neural network module; the multi-scale convolutional neural network module extracts features of the input distribution network measurement data to obtain multi-level features of the power grid nodes; based on the multi-level features, a multi-head attention mechanism is introduced; the attention heads of the multi-head attention mechanism calculate the correlation weights between the multi-level features to obtain a weight matrix; the multi-level features are weighted and adjusted according to the weight matrix, and feature splicing processing is performed to obtain fusion features; the fusion features are input into a gated mechanism graph neural network; the update gate and reset gate mechanism in the gated mechanism graph neural network dynamically control the propagation and fusion of power grid node information in time steps, and iteratively update the state of the power grid nodes; based on the multi-scale convolutional neural network module, the multi-head attention mechanism and the gated mechanism graph neural network, a distribution network state estimation model based on MHA-MSCNN-GGNN is constructed; the input distribution network target measurement data is analyzed and evaluated by the distribution network state estimation model, and the distribution network state is output. The GNN-based state estimation model of the present invention reduces the problem of traditional methods' over-reliance on physical models, while avoiding the limitation of pure data-driven methods' lack of physical interpretability. The addition of the gating mechanism enables the model to have stronger dynamic adaptability, and can capture time-varying processes such as load or generator output fluctuations. MSCNN enriches the model's ability to extract node information through different convolution kernel sizes, providing more comprehensive node features for subsequent neural networks. The addition of MHA improves the model's ability to estimate complex network states, making the model more robust and able to provide reliable estimation results even in the face of missing or inaccurate data.

[0147] It should be noted that the method of the embodiments of the present disclosure can be performed by a single device, such as a computer or server. The method of the embodiments of the present disclosure can also be applied in a distributed scenario, where multiple devices cooperate to perform the method. In such a distributed scenario, one of the multiple devices may only perform one or more steps of the method of the embodiments of the present disclosure, and the multiple devices will interact with each other to complete the method.

[0148] It should be noted that the above description is limited to some embodiments of the present disclosure. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in an order different from that described in the above embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require the specific order or sequential order shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0149] Example 2

[0150] See also Figure 9 Embodiment 2 of the present invention further provides a distribution network state estimation device based on an improved graph neural network, comprising:

[0151] A multi-scale convolutional neural network module construction unit 001 is used to construct a multi-scale convolutional neural network module; the multi-scale convolutional neural network module is used to extract features of the input distribution network measurement data to obtain multi-level features of the power grid nodes;

[0152] The multi-head attention mechanism introduction and processing unit 002 is used to introduce a multi-head attention mechanism based on the multi-level features; calculate the correlation weights between the multi-level features through the attention heads of the multi-head attention mechanism to obtain a weight matrix; perform weighted adjustment on the multi-level features according to the weight matrix, and perform feature splicing processing to obtain fused features;

[0153] The gated mechanism graph neural network processing unit 003 is used to input the fusion feature into the gated mechanism graph neural network; through the update gate and reset gate mechanism in the gated mechanism graph neural network, the propagation and fusion of the power grid node information in the time step is dynamically controlled, and the state of the power grid node is iteratively updated;

[0154] The MHA-MSCNN-GGNN-based distribution network state estimation model construction and processing unit 004 is used to construct a distribution network state estimation model based on MHA-MSCNN-GGNN based on the multi-scale convolutional neural network module, the multi-head attention mechanism and the gate control mechanism graphical neural network; the input distribution network target measurement data is analyzed and evaluated by the MHA-MSCNN-GGNN-based distribution network state estimation model, and the distribution network state is output.

[0155] In this embodiment, in the multi-scale convolutional neural network module construction unit 001, the multi-scale convolutional neural network module is provided with a plurality of convolution branches; the convolution branches are configured with convolution kernels of set sizes; small convolution kernels are used to capture local detail features and instantaneous fluctuation patterns of measurement data; medium convolution kernels are used to extract medium-range measurement correlation features; and large convolution kernels are used to capture global trend features and long-term dependencies.

[0156] The expression of the multi-level feature is:

[0157] y=α1y1+α2y2+α3y3

[0158] Where y is a multi-level feature; y1, y2, and y3 are features extracted by different convolution branches respectively; α n is the number of weights assigned to different convolution kernel sizes, n = 1, 2, 3.

[0159] In this embodiment, the multi-head attention mechanism is introduced into the processing unit 002, and the expression of the fusion feature is:

[0160] MultiHead(Q,K,V)=concat(head1,...,head h )W O

[0161] head i =Attention(QW i Q ,KW i K ,VW i V )

[0162] Where Q is the query; K is the key; V is the value; QW i Q ,KW i K ,VW i V are the linear transformation matrices of query, key, and value respectively; h is the number of attention heads; W O The linear transformation matrix of the output is (hd v ,d model ),d v is the dimension of the value, d model The dimensions of the model.

[0163] In this embodiment, in the gating mechanism graph neural network processing unit 003, the propagation model expression of the gating mechanism graph neural network is:

[0164]

[0165] Where, is the node state vector, which represents the hidden state of node v at time t. T is the transposed symbol, which is used to adjust the dimension to meet the requirements of matrix multiplication. b is a learnable bias term that can enhance the nonlinear expression ability of the model. is the aggregated message vector, which represents the messages aggregated by node v from its neighbors; To update the threshold and control the forgotten information; To reset the threshold, control the generation of new information; A v is the weight matrix, is the transformation matrix of neighbor messages; W z ,W t Both are learnable weight matrices, representing the transformation matrix of aggregated messages in the gating mechanism; U z ,U tBoth are learnable weight matrices, representing the transformation matrix of the historical state in the gating mechanism; σ As the activation function, the sigmoid function is usually used in gated graph neural networks.

[0166] In this embodiment, in the MHA-MSCNN-GGNN-based distribution network state estimation model construction and processing unit 004, the estimation result of the MHA-MSCNN-GGNN-based distribution network state estimation model is evaluated by the root mean square error, the mean absolute percentage error and the determination coefficient;

[0167] The calculation formula of the root mean square error is:

[0168]

[0169] Where RMSE is the root mean square error; is the estimate of the i-th state, y i is the corresponding true value, n is the total number of state quantities;

[0170] The calculation formula for the mean absolute percentage error is:

[0171]

[0172] Where, MAPE is the root mean square error;

[0173] The calculation formula of the determination coefficient is:

[0174]

[0175] Where R 2 is the root mean square error; y is the mean of the observations.

[0176] It should be noted that the information interaction, execution process, etc. between the modules of the above-mentioned system are based on the same concept as the method embodiment in Example 1 of the present application, and the technical effects they bring are the same as those of the method embodiment of the present application. For specific contents, please refer to the description in the method embodiment shown above in the present application, and no further details will be given here.

[0177] Example 3

[0178] Embodiment 3 of the present invention provides a non-transitory computer-readable storage medium, in which a program code for a distribution network state estimation method based on an improved graph neural network is stored. The program code includes instructions for executing embodiment 1 or any possible implementation thereof.

[0179] Computer-readable storage media can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that includes one or more available media. The available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

[0180] Example 4

[0181] Embodiment 4 of the present invention provides an electronic device, including: a memory and a processor;

[0182] The processor and the memory communicate with each other through a bus; the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute a distribution network state estimation method based on an improved graph neural network according to Example 1 or any possible implementation thereof.

[0183] Specifically, the processor can be implemented by hardware or by software. When implemented by hardware, the processor can be a logic circuit, an integrated circuit, etc.; when implemented by software, the processor can be a general-purpose processor, which is implemented by reading software code stored in a memory. The memory can be integrated into the processor or located outside the processor and exist independently.

[0184] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present invention is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable systems. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from a website, computer, server or data center to another website, computer, server or data center via a wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode.

[0185] Obviously, those skilled in the art will appreciate that the various modules or steps of the present invention described above can be implemented using a general-purpose computing system. They can be centralized on a single computing system or distributed across a network of multiple computing systems. Alternatively, they can be implemented using program code executable by a computing system, and thus, they can be stored in a storage system and executed by the computing system. In some cases, the steps shown or described herein can be performed in a different order than that shown, or they can be fabricated into separate integrated circuit modules, or multiple modules or steps can be fabricated into a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0186] Although the present invention has been described in detail above using general descriptions and specific embodiments, it will be apparent to those skilled in the art that modifications and improvements may be made thereto. Therefore, such modifications and improvements, without departing from the spirit of the present invention, are intended to be within the scope of protection claimed herein.

Claims

1. A distribution network state estimation method based on an improved graph neural network, characterized in that: include: Construct a multi-scale convolutional neural network module; The multi-scale convolutional neural network module is used to extract features from the input distribution network measurement data to obtain multi-level features of the power grid nodes; Based on the multi-level features, a multi-head attention mechanism is introduced; the correlation weights between the multi-level features are calculated by the attention heads of the multi-head attention mechanism to obtain a weight matrix; the multi-level features are weighted and adjusted according to the weight matrix, and feature splicing processing is performed to obtain a fusion feature; Inputting the fused features into a gated graphical neural network; dynamically controlling the propagation and fusion of grid node information over time steps through the update gate and reset gate mechanisms in the gated graphical neural network, and iteratively updating the states of the grid nodes; Based on the multi-scale convolutional neural network module, the multi-head attention mechanism and the gating mechanism graphical neural network, a distribution network state estimation model based on MHA-MSCNN-GGNN is constructed; the input distribution network target measurement data is analyzed and evaluated by the distribution network state estimation model based on MHA-MSCNN-GGNN, and the distribution network state is output.

2. A distribution network state estimation method based on an improved graph neural network according to claim 1, characterized in that: The multi-scale convolutional neural network module is provided with a plurality of convolution branches; the convolution branches are configured with convolution kernels of set sizes; small convolution kernels are used to capture local detail features and instantaneous fluctuation patterns of measurement data; medium convolution kernels are used to extract measurement-related features in a medium range; Large convolution kernels are used to capture global trend features and long-term dependencies; The expression of the multi-level feature is: y=α1y1+α2y2+α3y3 Where y is a multi-level feature; y1, y2, and y3 are features extracted by different convolution branches respectively; α n is the number of weights assigned to different convolution kernel sizes, n = 1, 2, 3.

3. A distribution network state estimation method based on an improved graph neural network according to claim 2, characterized in that: The expression of the fusion feature is: MultiHead(Q,K,V)=concat(head1,...,head h )W O head i =Attention(QW i Q ,KW i K ,VW i V ) Where Q is the query; K is the key; V is the value; QW i Q ,KW i K ,VW i V are the linear transformation matrices of query, key, and value respectively; h is the number of attention heads; W O The linear transformation matrix of the output is (hd v ,d model ),d v is the dimension of the value, d model The dimensions of the model.

4. A distribution network state estimation method based on an improved graph neural network according to claim 3, characterized in that: The propagation model expression of the gating mechanism graph neural network is: Where, is the node state vector, which represents the hidden state of node v at time t; T is the transpose symbol, which is used to adjust the dimension to meet the requirements of matrix multiplication; b is a learnable bias term that can enhance the nonlinear expression ability of the model; is the aggregated message vector, which represents the messages aggregated by node v from its neighbors; To update the threshold and control the forgotten information; To reset the threshold, control the generation of new information; A v is the weight matrix, is the transformation matrix of neighbor messages; W z ,W t Both are learnable weight matrices, representing the transformation matrix of aggregated messages in the gating mechanism; U z ,U t Both are learnable weight matrices, representing the transformation matrix of the historical state in the gating mechanism; σ As the activation function, the sigmoid function is usually used in gated graph neural networks.

5. A distribution network state estimation method based on an improved graph neural network according to claim 4, characterized in that: The estimation results of the distribution network state estimation model based on MHA-MSCNN-GGNN are evaluated by root mean square error, mean absolute percentage error and determination coefficient; The calculation formula of the root mean square error is: Where RMSE is the root mean square error; is the estimate of the i-th state, y i is the corresponding true value, n is the total number of state quantities; The calculation formula for the mean absolute percentage error is: Where, MAPE is the root mean square error; The calculation formula of the determination coefficient is: Where R 2 is the root mean square error; is the mean of the observations.

6. A distribution network state estimation device based on an improved graph neural network, adopting a distribution network state estimation method based on an improved graph neural network according to any one of claims 1 to 5, characterized in that: include: A multi-scale convolutional neural network module building unit, used to build a multi-scale convolutional neural network module; The multi-scale convolutional neural network module is used to extract features from the input distribution network measurement data to obtain multi-level features of the power grid nodes; A multi-head attention mechanism introduction and processing unit is used to introduce a multi-head attention mechanism based on the multi-level features; calculate the correlation weights between the multi-level features through the attention heads of the multi-head attention mechanism to obtain a weight matrix; perform weighted adjustment on the multi-level features according to the weight matrix, and perform feature splicing processing to obtain fused features; a gated graphical neural network processing unit, configured to input the fusion features into the gated graphical neural network; dynamically control the propagation and fusion of grid node information over time steps through the update gate and reset gate mechanisms in the gated graphical neural network, and iteratively update the states of the grid nodes; The MHA-MSCNN-GGNN-based distribution network state estimation model construction and processing unit is used to construct a distribution network state estimation model based on MHA-MSCNN-GGNN based on the multi-scale convolutional neural network module, the multi-head attention mechanism and the gated mechanism graphical neural network; the input distribution network target measurement data is analyzed and evaluated by the MHA-MSCNN-GGNN-based distribution network state estimation model, and the distribution network state is output.

7. The distribution network state estimation device based on an improved graph neural network according to claim 6, characterized in that: In the multi-scale convolutional neural network module construction unit, the multi-scale convolutional neural network module is provided with a plurality of convolution branches; the convolution branches are configured with convolution kernels of set sizes; the small convolution kernels are used to capture local detail features and instantaneous fluctuation patterns of the measurement data; the medium convolution kernels are used to extract measurement-related features in a medium range; Large convolution kernels are used to capture global trend features and long-term dependencies; The expression of the multi-level feature is: y=α1y1+α2y2+α3y3 Where y is a multi-level feature; y1, y2, and y3 are features extracted by different convolution branches respectively; α n is the number of weights assigned to different convolution kernel sizes, n = 1, 2, 3.

8. The distribution network state estimation device based on an improved graph neural network according to claim 7, characterized in that: The multi-head attention mechanism is introduced into the processing unit, and the expression of the fusion feature is: MultiHead(Q,K,V)=concat(head1,...,head h )W O head i =Attention(QW i Q ,KW i K ,VW i V ) Where Q is the query; K is the key; V is the value; QW i Q ,KW i K ,VW i V are the linear transformation matrices of query, key, and value respectively; h is the number of attention heads; W O The linear transformation matrix of the output is (hd v ,d model ),d v is the dimension of the value, d model The dimensions of the model.

9. The distribution network state estimation device based on an improved graph neural network according to claim 8, characterized in that: In the gate control mechanism graph neural network processing unit, the propagation model expression of the gate control mechanism graph neural network is: Where, is the node state vector, which represents the hidden state of node v at time t; T is the transpose symbol, which is used to adjust the dimension to meet the requirements of matrix multiplication; b is a learnable bias term that can enhance the nonlinear expression ability of the model; is the aggregated message vector, which represents the messages aggregated by node v from its neighbors; To update the threshold and control the forgotten information; To reset the threshold, control the generation of new information; A v is the weight matrix, is the transformation matrix of neighbor messages; W z ,W t Both are learnable weight matrices, representing the transformation matrix of aggregated messages in the gating mechanism; U z ,U t Both are learnable weight matrices, representing the transformation matrix of the historical state in the gating mechanism; σ As the activation function, the sigmoid function is usually used in gated graph neural networks.

10. The distribution network state estimation device based on improved graph neural network according to claim 9, characterized in that: In the MHA-MSCNN-GGNN-based distribution network state estimation model construction and processing unit, the estimation result of the MHA-MSCNN-GGNN-based distribution network state estimation model is evaluated by the root mean square error, mean absolute percentage error and determination coefficient; The calculation formula of the root mean square error is: Where RMSE is the root mean square error; is the estimate of the i-th state, y i is the corresponding true value, n is the total number of state quantities; The calculation formula for the mean absolute percentage error is: Where, MAPE is the root mean square error; The calculation formula of the determination coefficient is: Where R 2 is the root mean square error; is the mean of the observations.

Images