Graph convolutional network-based power flow calculation and optimal scheduling method and device for power distribution network

Abstract

The invention provides a power distribution network load flow calculation and optimal scheduling method and device based on a graph convolutional network, and the method comprises the steps: obtaining node power balance, distributed power supply output limitation and voltage constraint conditions from a network correlation matrix and original parameters of a power distribution network, adjusting the weight of the constraint conditions through the correlation matrix, and obtaining a load flow calculation result; generating a power distribution network load flow calculation nonlinear equation set; identifying an overload line and a low-voltage node from node voltage amplitude, phase angle and branch power data, generating an optimized operation scheme meeting N-1 security constraints by using a load clustering prediction and distributed power supply output probability model, and generating a power distribution network scheduling strategy; weak links are identified from a power distribution network scheduling strategy, the influence of future load increase on voltage and power is predicted by using global feature vectors, and a grid structure optimization suggestion and an equipment upgrading scheme are generated according to a power distribution network planning standard.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular, to a method and device for power flow calculation and optimal scheduling of a distribution network based on a graph convolutional network. Background Art

[0002] In the operation and management of a distribution network, the integration of multi-perspective data-driven power flow calculation and graph neural network modeling faces complex technical challenges. First, how to effectively integrate multi-source heterogeneous data such as load demand, distributed power generation output, and network topology, and perform standardized processing to construct a high-dimensional data set that comprehensively reflects the state of the distribution network? Second, in the process of graph neural network modeling, how to design a suitable method for encoding node and edge feature vectors to accurately express the attributes and mutual relationships of each component of the distribution network? Third, how to optimize the structures of the graph convolutional layer, pooling layer, and fully connected layer to achieve effective propagation, aggregation, and compression of network features? In addition, when constructing the spatial correlation matrix, how to combine the attention mechanism and multi-layer perceptron to accurately capture the non-linear relationships between the components of the distribution network? Finally, how to combine the output of the graph neural network model with traditional power flow calculation methods to design a hybrid algorithm to improve the calculation efficiency and accuracy? The solution of these problems is crucial for improving the accuracy and practicality of distribution network analysis and optimization. Summary of the Invention

[0003] To solve the above problems, the present invention proposes a method and device for power flow calculation and optimal scheduling of a distribution network based on a graph convolutional network, which can optimize the scheduling strategy, predict the impact of load growth, provide suggestions for grid structure optimization and equipment upgrade, and effectively improve the planning and operation level of the distribution network.

[0004] To achieve the above object, the technical solution adopted by the present invention is:

[0005] A method for power flow calculation and optimal scheduling of a distribution network based on a graph convolutional network, the method comprising:

[0006] S1. Obtain user power consumption behavior feature data through a load demand data acquisition and preprocessing module, obtain photovoltaic and wind power distributed power generation output data, extract a distribution network connection relationship graph from network topology structure information, and generate a standardized multi-dimensional data set;

[0007] S2. Obtain load, power generation output, and voltage attributes from the standardized multi-dimensional data set, map them to high-dimensional node vectors using node feature vector encoding, generate edge feature vectors from line impedance and capacity parameters, propagate and aggregate topological structure features through a graph convolutional layer, and generate a global feature vector by a fully connected layer;

[0008] S3. Construct a distribution network spatial correlation matrix from the global feature vector, process the spatial correlation matrix by a multi-layer perceptron, and generate a network correlation matrix;

[0009] S4. Obtain the node power balance, distributed power generation output limit, and voltage constraint conditions from the network correlation matrix and the original parameters of the distribution network, and generate a non-linear equation set for distribution network power flow calculation;

[0010] S5. Obtain the initial voltage amplitude and power value from the global eigenvector through the decoding layer conversion, and generate node voltage amplitude, phase angle, and branch power data;

[0011] S6. Identify overloaded lines and low-voltage nodes from the node voltage amplitude, phase angle, and branch power data, and generate an optimal operation plan that meets the N-1 security constraint by using load clustering prediction and distributed power generation output probability model, and generate a distribution network dispatching strategy.

[0012] Preferably, in S1, obtain user electricity consumption behavior characteristic data through the load demand data acquisition and preprocessing module, integrate the output data of photovoltaic and wind power distributed power generations by using the probability distribution model, extract the distribution network connection relationship diagram from the network topology structure information, perform normalization and standard deviation correction on the collected multi-source data, and generate a standardized multi-dimensional data set, including:

[0013] Obtain user electricity consumption behavior and load demand data through the data acquisition module, and fuse time series analysis to obtain a preliminary feature set;

[0014] Perform statistical integration on the photovoltaic output and wind power output data by using the probability distribution model to generate the distribution characteristics of distributed power generation output;

[0015] Extract the distribution network connection relationship diagram from the network topology structure to determine the distribution of power flow paths between nodes;

[0016] Perform normalization processing on the multi-source data to eliminate the dimension difference and obtain a unified scale data set;

[0017] Perform standard deviation correction on the unified scale data set to adjust the data fluctuation range and generate a corrected data set;

[0018] Perform clustering analysis on the corrected data set by using the K-means algorithm to judge the association mode between electricity consumption behavior and power generation output;

[0019] Construct a standardized multi-dimensional data set according to the association mode to determine the description of the distribution network operation characteristics.

[0020] Preferably, in S2, obtain the load, power generation output, and voltage attributes from the standardized multi-dimensional data set, map them to high-dimensional node vectors by using the node feature vector encoding, generate edge feature vectors from the line impedance and capacity parameters, propagate and aggregate the topology structure features through the graph convolution layer, reduce the dimension of the node vector by the pooling layer, and generate a global eigenvector by the fully connected layer, including:

[0021] Obtain the load attributes, power output, and voltage attributes in the standardized multi-dimensional dataset, map them to node feature vectors through a preset coding model, and obtain high-dimensional vectors;

[0022] Calculate the edge feature vectors through the line impedance parameters and capacity parameters to obtain the edge features reflecting the topological relationship;

[0023] Use the graph convolutional layer to propagate and aggregate the high-dimensional vectors and edge features to obtain the convolutional features containing the topological structure;

[0024] Reduce the dimension of the convolutional features through the pooling layer to obtain the compressed vectors;

[0025] If the dimension of the compressed vector exceeds the preset threshold, retain the main components through feature screening to obtain the adjusted compressed vector;

[0026] Use the fully connected layer to transform the adjusted compressed vector to obtain the global feature vector;

[0027] For the global feature vector, judge the system state through a preset classification model to obtain the operation trend analysis result.

[0028] Preferably, in S3, construct a spatial correlation matrix of the distribution network from the global feature vector, use the attention mechanism to calculate the weights according to the distance between nodes and the power flow direction, quantify the correlation degree between nodes and branches, and process the spatial correlation matrix by a multi-layer perceptron to generate a network correlation matrix, including:

[0029] Generate a spatial correlation matrix through the global features, calculate the preliminary weights according to the distance between nodes and the power flow direction to obtain the quantified correlation matrix;

[0030] Use the attention mechanism to adjust the preliminary weights, generate weighted coefficients for the power flow direction and node distance to obtain the enhanced correlation matrix;

[0031] Extract the node-related and branch-related feature components from the enhanced correlation matrix, separate the main structure through matrix decomposition to obtain the decomposed feature matrix;

[0032] If the dimension of the decomposed feature matrix exceeds the preset threshold, retain the core components through feature screening to obtain the refined feature matrix;

[0033] For the refined feature matrix, use a multi-layer perceptron for transformation processing to obtain the initial network correlation matrix;

[0034] According to the initial network correlation matrix, calculate the association strength between nodes and between branches to obtain the final network correlation matrix;

[0035] Quantify the comprehensive impact of spatial correlation and power flow direction through the final network correlation matrix to obtain a global representation of the system topology.

[0036] Preferably, in S4, obtain the node power balance, distributed power generation output limit, and voltage constraint conditions from the network correlation matrix and the original distribution network parameters, and use the correlation matrix to adjust the constraint condition weights to generate a non-linear equation set for distribution network power flow calculation, including:

[0037] Extract the node power, distributed power generation output limit, and voltage constraint conditions through the network correlation matrix and distribution network parameters to obtain an initial constraint data set;

[0038] Use the matrix analysis method to process the network correlation and determine the weight adjustment coefficient between constraint conditions;

[0039] Update the constraint conditions according to the weight adjustment result to generate a non-linear equation set required for power flow calculation;

[0040] If the node power exceeds the voltage constraint range, solve the non-linear equation set by the iterative method to judge the boundary value of the distributed power generation output limit;

[0041] Obtain the solved boundary value, update the distribution network parameters, and obtain an adjusted input data set for power flow calculation;

[0042] Verify the adjusted data set through power flow calculation to determine the balance state of node power and voltage constraints;

[0043] Finally, solve the non-linear equation set using the Newton-Raphson algorithm to obtain a stable solution for distribution network power flow calculation.

[0044] Preferably, in S5, obtain the initial voltage amplitude and power value by converting the global eigenvector through the decoding layer, use the Newton-Raphson method combined with the conjugate gradient iterative solver to process the non-linear equation set, define the target error through the loss function and iteratively update the parameters to generate node voltage amplitude, phase angle, and branch power data, including:

[0045] Obtain the initial voltage and power value data by processing the global features through the decoding layer;

[0046] Use the Newton-Raphson method combined with the conjugate gradient iterative solver to process the non-linear equation set from the initial voltage and power value data to obtain a solution set;

[0047] Calculate the target error value through the solution set combined with the loss function;

[0048] If the target error value exceeds the preset threshold, iteratively update the decoding layer parameters by the gradient descent method to obtain a new parameter set;

[0049] Regenerate the node voltage and phase angle data through a new parameter set;

[0050] Calculate the branch power value according to the node voltage and phase angle data;

[0051] Verify the consistency through the comparison between the branch power value and the power value to obtain the final data.

[0052] Preferably, in S6, by collecting the node voltage, phase angle data and branch power, calculate the load rate of each line and the node voltage offset, and determine the overloaded lines and low-voltage nodes;

[0053] Use the K-means algorithm to cluster the load, obtain the load distribution characteristics from the historical data, and get the load change trend;

[0054] According to the load change trend and the distributed power output probability model, calculate the output range of each power source, and judge the output combination that meets the N-1 constraint;

[0055] Through the linear programming algorithm, obtain the optimization scheme that meets the branch power limit from the output combination, and determine the node voltage recovery range;

[0056] Generate a scheduling strategy for the node voltage and branch power adjustment amount in the optimization scheme to obtain the distribution network operation parameters;

[0057] After obtaining the distribution network operation parameters, if the branch power exceeds the preset threshold, adjust the distributed power output and judge whether the overloaded lines are eliminated;

[0058] Calculate the node voltage offset of each node through the adjusted operation parameters to determine whether the low-voltage nodes return to normal.

[0059] Preferably, after S5, obtain the calculation results from the node voltage amplitude, phase angle and branch power data, compare with the measured data to calculate the voltage deviation rate and the power error percentage. If the error exceeds the preset threshold, adjust the graph convolution kernel size and the activation function parameters, regenerate the global feature vector, identify the weak links from the distribution network scheduling strategy, use the global feature vector to predict the impact of future load growth on voltage and power, and generate the grid structure optimization suggestions and equipment upgrade plans according to the distribution network planning standards, including:

[0060] Obtain the voltage amplitude and voltage phase angle through the node voltage and branch power data, and calculate the initial result;

[0061] Compare the initial result with the measured data to obtain the voltage deviation rate and the power error percentage;

[0062] If the voltage deviation rate exceeds the preset threshold, adjust the graph convolution kernel size and recalculate the global feature vector;

[0063] If the power error percentage exceeds the preset threshold, adjust the activation function parameters and update the global feature vector;

[0064] Recalculate the result using the updated global feature vector to obtain the new voltage deviation rate and power error percentage;

[0065] Compare the newly calculated result with the measured data to determine whether the error is within the preset threshold to obtain the final result;

[0066] Determine the stable output of the global feature vector based on the final result.

[0067] In another technical solution, a device for power flow calculation and optimal scheduling of a distribution network based on a graph convolutional network includes:

[0068] A load demand data acquisition and preprocessing module for obtaining user electricity consumption behavior characteristic data;

[0069] A data storage module that stores a processing program;

[0070] A data processing module is electrically connected to the data storage module and the load demand data acquisition and preprocessing module. The data processing module is used to receive user electricity consumption behavior characteristic data, load and execute the processing program stored in the memory to implement each step of the power flow calculation and optimal scheduling method of the distribution network based on a graph convolutional network as described in any one of the above, and generate a distribution network scheduling strategy.

[0071] The beneficial effects of using the present invention are:

[0072] The present invention discloses a power flow calculation and optimal scheduling method for a distribution network based on a graph convolutional network. The method collects multi-source data and performs standardized processing, uses a graph convolutional network to extract the topological features of the distribution network, constructs a spatial correlation matrix, and optimizes the power flow calculation constraint conditions. The Newton-Raphson method is combined with conjugate gradient iteration to solve the non-linear equations to generate node voltage and branch power data. By comparing with the measured data, the model parameters are dynamically adjusted to improve the calculation accuracy. According to the calculation results, the weak links of the network are identified, and combined with the load prediction and distributed power generation output model, an optimal scheduling strategy that meets the N-1 constraint is generated. The present invention can also predict the impact of load growth and provide suggestions for the optimization of the grid structure and equipment upgrade, effectively improving the planning and operation level of the distribution network. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 It is a flowchart of the power flow calculation and optimal scheduling method for a distribution network based on a graph convolutional network of the present invention.

[0074] Figure 2 It is a schematic diagram of the device for power flow calculation and optimal scheduling of a distribution network based on a graph convolutional network of the present invention. Detailed implementation manners

[0075] To make the objectives, technical solutions and advantages of the present technical solution clearer and more understandable, the present technical solution will be further described in detail below in conjunction with specific implementation manners. It should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present technical solution.

[0076] To further understand the content of the present invention, the present invention will be described in detail in conjunction with the accompanying drawings and embodiments. The present application will be further described in detail below in conjunction with the accompanying drawings and embodiments. It can be understood that the specific embodiments described herein are only used to explain the relevant invention and are not intended to limit the invention. Additionally, it should be noted that for the sake of description, only the parts related to the invention are shown in the drawings.

[0077] As Figure 1 , a method for power flow calculation and optimal scheduling of a distribution network based on a graph convolutional network in this embodiment may specifically include:

[0078] S1. Obtain user electricity consumption behavior characteristic data through a load demand data acquisition and preprocessing module, integrate the output data of photovoltaic and wind power distributed power sources by using a probability distribution model, extract a connection relationship graph of the distribution network from the network topology structure information, and perform normalization and standard deviation correction on the collected multi-source data to generate a standardized multi-dimensional data set.

[0079] Obtain user electricity consumption behavior and load demand data through a data acquisition module, and fuse time series analysis to obtain a preliminary feature set. Use a probability distribution model to perform statistical integration on the photovoltaic output and wind power output data to generate the output distribution characteristics of the distributed power sources. Extract a connection relationship graph of the distribution network from the network topology structure to determine the distribution of the power flow paths between nodes. Perform normalization processing on the multi-source data to eliminate the dimension difference and obtain a unified scale data set. Perform standard deviation correction on the unified scale data set to adjust the data fluctuation range and generate a corrected data set. Perform clustering analysis on the corrected data set through the K-means algorithm to judge the association mode between the electricity consumption behavior and the power source output. Construct a standardized multi-dimensional data set according to the association mode to determine the description of the distribution network operation characteristics.

[0080] Specifically, through the load demand data acquisition and preprocessing module, the system obtains the hourly electricity consumption data of users from smart meters. For example, the electricity consumption of a certain user on a certain day is [2, 8, 5, 0, 7, 9, 3, 1, 8, 6, 4, 3] kWh. The K-means clustering algorithm is used to perform clustering analysis on these data to identify the typical electricity consumption patterns of users, such as peak hours and off-peak hours. In the integration of the output data of photovoltaic and wind power distributed power sources, the system collects the output data of a certain photovoltaic power station as [0, 0, 5, 2, 0, 5, 8, 0, 5, 0, 5, 0] kW, and the output data of a wind farm as [0, 5, 0, 5, 0, 5, 0, 5, 0, 5, 0, 5] kW. The Gaussian mixture model is used to fit the probability distribution of these data to generate a joint output probability density function for predicting the output situation in future periods. From the network topology structure information, the system extracts the connection graph of the distribution network. For example, a certain regional distribution network contains 10 nodes and 12 branches, and the connection relationship between nodes is {(1, 2), (2, 3), (3, 4), (4, 5), (5, 6), (6, 7), (7, 8), (8, 9), (9, 10), (10, 1), (2, 5), (6, 9)}. Graph theory algorithms are used to calculate the degrees and shortest paths of each node, providing a basis for subsequent power flow calculations. Normalization and standard deviation correction are performed on the collected multi-source data. For example, the load data, photovoltaic output data, and wind power output data are respectively normalized to the [0, 1] interval, and the Z-score standardization method is used to perform standard deviation correction on the data, so that the data in each dimension has the same dimension and distribution characteristics, and finally a standardized multi-dimensional data set is generated for subsequent machine learning model training and prediction analysis.

[0081] S2. Obtain load, power source output, and voltage attributes from the standardized multi-dimensional data set, map them to high-dimensional node vectors using node feature vector encoding, generate edge feature vectors from line impedance and capacity parameters, propagate and aggregate topological structure features through graph convolutional layers, reduce the dimension of node vectors through pooling layers, and generate global feature vectors by fully connected layers.

[0082] Obtain the load attributes, power source output, and voltage attributes in the standardized multi-dimensional data set, and map them to node feature vectors through a preset encoding model to obtain high-dimensional vectors. The calculation method of the node feature vector is as follows:

[0083]

[0084] Among them, Fn represents the node feature vector, M represents the data dimension, Pi represents the power attribute, Vi represents the voltage attribute, Li represents the load attribute, and We represents the encoding weight matrix.

[0085] Calculate the edge eigenvector through the line impedance parameter and the capacity parameter to obtain the edge feature reflecting the topological relationship. The edge feature calculation is as follows:

[0086]

[0087] Among them, Zij represents the line impedance, Zbase represents the base impedance, Sij represents the line capacity, and Smax represents the maximum capacity.

[0088] Use the graph convolutional layer to propagate and aggregate the high-dimensional vector and the edge feature to obtain the convolutional feature containing the topological structure. The convolutional feature representation is as follows:

[0089]

[0090] Among them, Hk represents the convolutional feature, σ represents the activation function, Ni represents the neighbor set of node i, di and dj represent the node degrees, and W represents the convolutional kernel parameter matrix.

[0091] Reduce the dimension of the convolutional feature through the pooling layer to obtain the compressed vector. If the dimension of the compressed vector exceeds the preset threshold, retain the main components through feature screening to obtain the adjusted compressed vector. Use the fully connected layer to transform the adjusted compressed vector to obtain the global feature vector. For the global feature vector, judge the system state through the preset classification model to obtain the operation trend analysis result.

[0092] Specifically, from the standardized multi-dimensional dataset, the load, power output, and voltage attributes are first extracted. For example, the load data includes the load of node 1 being 100 MW and the load of node 2 being 150 MW. The power output data includes the output of node 3 being 200 MW and the output of node 4 being 250 MW. The voltage data includes the voltage of node 1 being 0.5 p.u. and the voltage of node 2 being 0.3 p.u. Using these data, through the node feature vector encoding method, such as principal component analysis (PCA), the node features are mapped into high-dimensional node vectors. For example, the feature vector of node 1 is [8, 6, 4], and the feature vector of node 2 is [7, 5, 3]. Then, the edge feature vectors are generated from the line impedance and capacity parameters. For example, the impedance of line 1 is 1 Ω and the capacity is 100 MVA, and the impedance of line 2 is 2 Ω and the capacity is 150 MVA. The edge feature vectors are generated through linear transformation. For example, the edge feature vector of line 1 is [1, 2], and the edge feature vector of line 2 is [15, 25]. The topological structure features are propagated and aggregated through the graph convolutional layer (GCN). For example, two layers of GCN are used, the output dimension of the first layer is 32, and the output dimension of the second layer is 16. Through the aggregation function such as mean aggregation, the aggregated feature vector of each node is obtained. For example, the aggregated feature vector of node 1 is [75, 55, 35]. The node vectors are reduced in dimension and compressed by the pooling layer. For example, using max pooling, the aggregated feature vector of node 1 is compressed to

[75] , and the aggregated feature vector of node 2 is compressed to

[55] . Finally, the global feature vector is generated by the fully connected layer. For example, two layers of fully connected networks are used, the output dimension of the first layer is 8, and the output dimension of the second layer is 4. The finally generated global feature vector is [6, 5, 4, 3], which is used for subsequent power system analysis and decision-making.

[0093] S3. Construct the spatial correlation matrix of the distribution network from the global feature vector, use the attention mechanism to calculate the weights according to the distance between nodes and the power flow direction, quantify the correlation degree between nodes and branches, and process the spatial correlation matrix by the multi-layer perceptron to generate the network correlation matrix.

[0094] Generate a spatial correlation matrix through global features, calculate preliminary weights based on the distance between nodes and power flow direction, and obtain a quantized correlation matrix. Use the attention mechanism to adjust the preliminary weights, generate weighted coefficients for power flow direction and node distance, and obtain an enhanced correlation matrix. Extract node-related and branch-related feature components from the enhanced correlation matrix, separate the main structure through matrix decomposition, and obtain a decomposed feature matrix. If the dimension of the decomposed feature matrix exceeds the preset threshold, retain the core components through feature screening to obtain a refined feature matrix. For the refined feature matrix, use a multi-layer perceptron for transformation processing to obtain an initial network correlation matrix. According to the initial network correlation matrix, calculate the association strength between nodes and between branches to obtain the final network correlation matrix. Through the final network correlation matrix, quantify the comprehensive influence of spatial correlation and power flow direction to obtain a global representation of the system topology.

[0095] Specifically, in the analysis of the distribution network, when constructing a spatial correlation matrix based on the global feature vector, first extract the feature relationships between nodes from the global feature vector. For example, the global feature vector of node 1 is [6, 5, 4, 3], that of node 2 is [5, 4, 3, 2], and that of node 3 is [4, 3, 2, 1]. Use the cosine similarity algorithm to calculate the feature similarity between nodes. For example, the similarity between node 1 and node 2 is 98, the similarity between node 1 and node 3 is 95, and the similarity between node 2 and node 3 is 97, forming an initial spatial correlation matrix. Then, introduce the attention mechanism to dynamically adjust the weights according to the physical distance and power flow direction between nodes. For example, the distance between node 1 and node 2 is 10 kilometers and the power flow direction is positive, with a weight coefficient of 9; the distance between node 1 and node 3 is 15 kilometers and the power flow direction is negative, with a weight coefficient of 7. Through weighted calculation, update the spatial correlation matrix. For example, the correlation value between node 1 and node 2 is updated to 88, and that between node 1 and node 3 is updated to 66. Subsequently, input the updated spatial correlation matrix into a multi-layer perceptron, and use the ReLU activation function and Adam optimizer for training. For example, the output dimension of the first layer of the perceptron is 16, the second layer is 8, and finally generate a network correlation matrix. For example, the network correlation value between node 1 and node 2 is 85, that between node 1 and node 3 is 62, and that between node 2 and node 3 is 80. This matrix can be used to quantify the correlation degree between nodes and branches in the distribution network, providing data support for subsequent power system optimization.

[0096] S4. Obtain the node power balance, distributed power generation output limit, and voltage constraint conditions from the network correlation matrix and the original parameters of the distribution network, and use the correlation matrix to adjust the weights of the constraint conditions to generate a non-linear equation system for distribution network power flow calculation.

[0097] Extract the node power, distributed power generation output limits, and voltage constraint conditions from the network correlation matrix and distribution network parameters to obtain the initial constraint dataset. Use matrix analysis methods to process the network correlation and determine the weight adjustment coefficients between the constraint conditions. Update the constraint conditions according to the weight adjustment results to generate the non-linear equations required for power flow calculation. If the node power exceeds the voltage constraint range, solve the non-linear equations by the iterative method to determine the boundary values of the distributed power generation output limits. Obtain the solved boundary values, update the distribution network parameters, and obtain the adjusted input dataset for power flow calculation. Verify the adjusted dataset through power flow calculation to determine the balance state of the node power and voltage constraints. Use the Newton-Raphson algorithm to finally solve the non-linear equations to obtain the stable solution of the distribution network power flow calculation.

[0098] Specifically, first, extract the electrical relationships between nodes from the network correlation matrix. For example, the correlation coefficient between node A and node B is 85, indicating a strong correlation in power flow between the two. Use the original distribution network parameters, such as the load power of node C being 500 kW, the maximum output limit of distributed power source D being 300 kW, and the allowable voltage range of node E being 95 to 05 p.u., to construct a node power balance equation to ensure that the total power generation equals the total load. Then, adjust the weights of the constraint conditions through the correlation matrix. For example, adjust the voltage constraint weight between node A and node B to 2 to reflect their strong correlation. On this basis, generate the non-linear equations for distribution network power flow calculation and use the Newton-Raphson method for iterative solution with the initial guess value of the node voltage being 0 p.u. and the convergence accuracy being set to 0001. During the iteration process, calculate the power mismatch of each node and update the node voltage until the convergence condition is met. Through the above steps, the accurate calculation of the distribution network power flow and the dynamic adjustment of the constraint conditions are realized, ensuring the stable operation and efficient management of the system.

[0099] S5. Obtain the initial voltage amplitude and power value by converting the global eigenvector through the decoding layer. Use the Newton-Raphson method combined with the conjugate gradient iterative solver to process the non-linear equations, define the target error through the loss function, and iteratively update the parameters to generate the node voltage amplitude, phase angle, and branch power data.

[0100] Obtain the initial voltage and power value data by processing the global features through the decoding layer. Use the Newton-Raphson method combined with the conjugate gradient iterative solver to process the non-linear equations from the initial voltage and power value data to obtain the solution set. Calculate the target error value through the solution set combined with the loss function. If the target error value exceeds the preset threshold, iteratively update the parameters of the decoding layer by the gradient descent method to obtain a new parameter set. Regenerate the node voltage and phase angle data through the new parameter set. Calculate the branch power value according to the node voltage and phase angle data. Verify the consistency through the comparison between the branch power value and the power value to obtain the final data.

[0101] Specifically, in power system analysis, the topological structure and electrical parameters of the power network are first extracted through global eigenvectors, such as the nodal admittance matrix and load data. Assume the system contains 5 nodes, the admittance matrix is Y = [[5, -2, 0, 0, 0], [-2, 6, -1, 0, 0], [0, -1, 7, -3, 0], [0, 0, -3, 8, -2], [0, 0, 0, -2, 4]], the load data is P = [8, 2, 9, 5, 1] MW, and Q = [4, 6, 5, 7, 6] MVar. These eigenvectors are input into the decoding layer. The decoding layer adopts a fully connected neural network structure, the activation function is ReLU, the number of network layers is 3, and the number of neurons in each layer is 10, 20, and 5 respectively. After training, the decoding layer outputs the initial voltage magnitude V0 = [02, 03, 01, 04, 02] p.u. and the initial power value P0 = [85, 25, 95, 55, 15] MW. Next, the Newton-Raphson method combined with a conjugate gradient iterative solver is used to solve the non-linear equations, with the convergence accuracy set to 1e-6 and the maximum number of iterations set to 100. In each iteration, the Jacobian matrix J and the residual vector F are calculated. For example, in the first iteration, J = [[5, -3, 0, 0, 0], [-3, 8, -2, 0, 0], [0, -2, 0, -4, 0], [0, 0, -4, 5, -3], [0, 0, 0, -3, 7]], and F = [05, 03, 02, 04, 03]. The conjugate gradient method is used to solve for Δx = J^(-1)F, and the voltage magnitude and phase angle are updated. For example, after the first iteration, V1 = [021, 031, 011, 041, 021] p.u., and θ1 = [01, 02, 01, 03, 02] rad. The loss function uses the mean square error, defined as L = Σ(Vi - Vi0)^2 + Σ(Pi - Pi0)^2, the initial error is 1, and after 5 iterations, the error drops to 1e-4, meeting the convergence condition. Finally, the nodal voltage magnitude V = [022, 032, 012, 042, 022] p.u., the phase angle θ = [012, 022, 012, 032, 022] rad, and the branch power P = [86, 26, 96, 56, 16] MW are generated. These data are used for power system stability and economic analysis, providing decision support for power grid scheduling.

[0102] S6. Identify overloaded lines and low-voltage nodes from the nodal voltage magnitude, phase angle, and branch power data, generate an optimized operation plan that meets the N-1 security constraint using load clustering prediction and the distributed power output probability model, and generate a distribution network scheduling strategy.

[0103] By collecting node voltage, phase angle data and branch power, calculate the load rate of each line and the node voltage offset, and determine the overloaded lines and low-voltage nodes. Use the K-means algorithm to cluster the load, obtain the load distribution characteristics from historical data, and get the load change trend. According to the load change trend and the distributed power output probability model, calculate the output range of each power source, and judge the output combination that meets the N-1 constraint. Through the linear programming algorithm, obtain the optimization scheme that meets the branch power limit from the output combination, and determine the node voltage recovery range. For the node voltage and branch power adjustment amounts in the optimization scheme, generate a scheduling strategy to obtain the operating parameters of the distribution network. After obtaining the operating parameters of the distribution network, if the branch power exceeds the preset threshold, adjust the distributed power output, and judge whether the overloaded lines are eliminated. Through the adjusted operating parameters, calculate the node voltage offset of each node, and determine whether the low-voltage nodes return to normal.

[0104] Specifically, in the distribution network scheduling, first, through the node voltage amplitude and phase angle data, combined with the branch power information, use the power flow calculation algorithm to identify the overloaded lines and low-voltage nodes.

[0105] For example, if the current of a certain line reaches 115% of the rated value, it is determined to be overloaded, and if the voltage of a certain node drops to 92 p.u., which is lower than the allowed 95 p.u., it is marked as a low-voltage node. Then, use the K-means clustering algorithm to perform clustering analysis on the load data to predict the future load change trend. Suppose the load is divided into three categories, representing high, medium, and low load levels respectively, and the probability of the high load level in the next hour is predicted to be 30%, the medium load level is 50%, and the low load level is 20%. Then, based on the distributed power output probability model, use the Monte Carlo simulation to generate 1000 possible output scenarios, and calculate the system operating state under each scenario.

[0106] For example, the average value of the photovoltaic output is 500 kW, the standard deviation is 50 kW, the average value of the wind power output is 300 kW, and the standard deviation is 30 kW. On this basis, use the genetic algorithm for optimization, with the objective function of minimizing the operating cost while meeting the N-1 security constraint. During the optimization process, consider the constraints such as line capacity and node voltage limit, and finally generate a scheduling strategy.

[0107] For example, the optimization result shows that under the premise of meeting the N-1 security constraint, the total system operating cost is 5000 yuan, where the distributed power output is 800 kW, the main network power purchase is 1200 kW, and the load demand is 2000 kW. Through the above method, the safe and economic operation of the distribution network is realized.

[0108] In addition, calculation results are obtained from the node voltage amplitude, phase angle, and branch power data. The voltage deviation rate and power error percentage are calculated by comparing with the measured data. If the error exceeds the preset threshold, the size of the graph convolution kernel and the parameters of the activation function are adjusted, and the global feature vector is regenerated.

[0109] The voltage amplitude and voltage phase angle are obtained from the node voltage and branch power data, and the initial result is calculated. By comparing the initial result with the measured data, the voltage deviation rate and power error percentage are obtained. If the voltage deviation rate exceeds the preset threshold, the size of the graph convolution kernel is adjusted, and the global feature vector is recalculated. If the power error percentage exceeds the preset threshold, the parameters of the activation function are adjusted, and the global feature vector is updated. The result is recalculated using the updated global feature vector to obtain the new voltage deviation rate and power error percentage. By comparing the newly calculated result with the measured data, it is determined whether the error is within the preset threshold to obtain the final result. The stable output of the global feature vector is determined according to the final result. Specifically, in the power system, first, based on the node voltage amplitude and phase angle data, combined with the branch power information, the power flow calculation is carried out using the Newton-Raphson method to obtain the voltage amplitude and phase angle of each node.

[0110] For example, in a certain system, the calculated value of the node voltage amplitude is 0.2 pu, and the phase angle is 3 degrees, while the measured data are 0.1 pu and 5 degrees respectively. By calculating the voltage deviation rate, it is found that the amplitude deviation rate is 99%, and the phase angle deviation rate is 64%. At the same time, the calculated value of the branch power is 50 MW, the measured data is 45 MW, and the power error percentage is 1%. If the preset voltage deviation rate threshold is 1%, the phase angle deviation rate threshold is 4%, and the power error percentage threshold is 5%, then the phase angle deviation rate is close to the threshold and needs to be adjusted. At this time, a graph convolutional neural network (GCN) is used for feature extraction. The initial convolution kernel size is 3, and the activation function is ReLU. By adjusting the convolution kernel size to 5 and using the LeakyReLU activation function, the global feature vector is regenerated.

[0111] For example, after adjustment, the deviation rates of the new feature vector in predicting the node voltage amplitude and phase angle are reduced to 8% and 2% respectively, and the power error percentage is reduced to 9%, all meeting the requirements of the preset threshold. This process is realized through an automated algorithm, ensuring the accuracy and stability of power system analysis.

[0112] Identify weak links from the distribution network dispatching strategy, use the global feature vector to predict the impact of future load growth on voltage and power, and generate suggestions for grid structure optimization and equipment upgrade plans according to the distribution network planning standards. Extract the characteristics of weak links from the dispatching strategy data, and use the clustering algorithm to determine the distribution locations of weak links. Obtain the global feature vector from the distribution locations of weak links, conduct predictive analysis on the load growth data, and obtain the future load change trend. According to the future load change trend, calculate the quantitative values of voltage impact and power impact through the regression algorithm, and judge the impact range. If the voltage impact exceeds the preset threshold, generate a grid structure adjustment plan according to the planning standards and determine the content of the optimization suggestions. If the power impact exceeds the preset threshold, match the equipment upgrade parameters through the planning standards to obtain the equipment upgrade plan. Obtain the grid structure adjustment plan and the equipment upgrade plan, use the topology analysis tool to generate the optimized layout of the distribution network system, and judge the overall improvement amplitude. For the optimized layout, update the predictive analysis in combination with the global feature vector to obtain the long-term stability index of the distribution network system.

[0113] Specifically, in the process of identifying weak links from the distribution network dispatching strategy, first collect the real-time operation data of the distribution network, including node voltage, branch current, power factor, etc., and use the network topology analysis method based on graph theory to construct the physical model of the distribution network.

[0114] For example, use the depth-first search (DFS) algorithm to traverse the distribution network, identify the nodes with voltage lower than 95 pu and the branches with power factor lower than 9 as potential weak links. Then, use the global feature vector method to extract the characteristics of these weak links, reduce the dimension through the principal component analysis (PCA) algorithm, and extract the key feature vectors affecting voltage and power, such as load growth trend, equipment aging degree, etc. On this basis, use the time series prediction model (such as ARIMA) to predict the future load growth. Assuming that the average annual load growth rate is 5%, combined with the distribution network planning standards, calculate the voltage deviation and power loss of each node in the next 5 years.

[0115] For example, the prediction results show that the voltage of a certain node will drop to 92 pu in 3 years, exceeding the allowable range. According to these analysis results, generate suggestions for grid structure optimization, such as adding a new 10 kV line near this node and upgrading the original transformer capacity from 500 kVA to 800 kVA to meet the future load demand. At the same time, for the problem of equipment aging, it is recommended to replace the equipment with an aging rate exceeding 20%, and use an intelligent monitoring system to monitor the equipment status in real time to ensure the stable operation of the distribution network.

[0116] To implement the above method, in a feasible embodiment, the present invention proposes a distribution network power flow calculation and optimal dispatching device based on a graph convolutional network, including:

[0117] The load demand data acquisition and preprocessing module 10 is used to obtain user electricity consumption behavior characteristic data;

[0118] The data storage module 30 stores a processing program;

[0119] The data processing module 20 is electrically connected to the data storage module 30 and the load demand data acquisition and preprocessing module 10. The data processing module 20 is used to receive user electricity consumption behavior characteristic data, load and execute the processing program stored in the memory, so as to implement the steps of the above-mentioned power flow calculation and optimal scheduling method for the distribution network based on the graph convolutional network, and generate a distribution network scheduling strategy.

[0120] The device provided by the embodiment of the present application includes a data processing module 20 and a data storage module 30. A program or instruction that can run on the data processing module 20 is stored on the data storage module 30. When the program or instruction is executed by the data processing module 20, the steps of the above-mentioned information processing method embodiment are implemented, and the same technical effects can be achieved. To avoid repetition, it will not be elaborated here.

[0121] The data storage module 30 can be used to store software programs and various data. The data storage module 30 may mainly include a first storage area for storing programs or instructions and a second storage area for storing data. Among them, the first storage area can store an operating system, application programs or instructions required for at least one function (such as a sound playback function, an image playback function, etc.). In addition, the data storage module 30 can include a volatile memory or a non-volatile memory, or the data storage module 30 can include both a volatile and a non-volatile memory. Among them, the non-volatile memory can be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory can be a random access memory (RAM), a static random access memory (SRAM), a dynamic random access memory (DRAM), a synchronous dynamic random access memory (SDRAM), a double data rate synchronous dynamic random access memory (DDR SDRAM), an enhanced synchronous dynamic random access memory (ESDRAM), a synch link dynamic random access memory (SLDRAM), and a direct rambus random access memory (DRRAM). The data storage module 30 in the embodiments of the present application includes but is not limited to these and any other suitable types of memories.

[0122] The data processing module 20 may include one or more processing units; optionally, the data processing module 20 integrates an application processor and a modem processor. Among them, the application processor mainly processes operations related to the operating system, user interface, and application programs, etc., and the modem processor mainly processes wireless communication signals, such as a baseband processor. It can be understood that the above-mentioned modem processor may not be integrated into the data processing module 20 either.

[0123] The above content is only a preferred embodiment of the present invention. For those of ordinary skill in the art, many changes can be made in the specific implementation manners and application scopes according to the idea of the technical content of the present application. As long as these changes do not depart from the concept of the present invention, they all belong to the protection scope of this patent.

Claims

1. A power flow calculation and optimal scheduling method for a distribution network based on a graph convolutional network, characterized in that, The method includes: S1. Obtain user electricity consumption behavior characteristic data through a load demand data acquisition and preprocessing module, obtain photovoltaic and wind power distributed power generation output data, extract the connection relationship diagram of the distribution network from the network topology structure information, and generate a standardized multi-dimensional data set; S2. Obtain load, power generation output, and voltage attributes from the standardized multi-dimensional data set, use node feature vector encoding to map to high-dimensional node vectors, generate edge feature vectors from line impedance and capacity parameters, propagate and aggregate topology structure features through a graph convolutional layer, and generate a global feature vector by a fully connected layer; S3. Construct a spatial correlation matrix of the distribution network from the global feature vector, and process the spatial correlation matrix by a multi-layer perceptron to generate a network correlation matrix; S4. Obtain node power balance, distributed power generation output limit, and voltage constraint conditions from the network correlation matrix and the original parameters of the distribution network, and generate a non-linear equation set for distribution network power flow calculation; S5. Obtain the initial voltage amplitude and power value through conversion by a decoding layer from the global feature vector, and generate node voltage amplitude, phase angle, and branch power data; S6. Identify overloaded lines and low-voltage nodes from the node voltage amplitude, phase angle, and branch power data, use load clustering prediction and distributed power generation output probability models to generate an optimized operation plan that meets the N-1 security constraint, and generate a distribution network scheduling strategy.

2. The power flow calculation and optimal scheduling method of the distribution network based on the graph convolutional network according to claim 1, wherein: In S1, obtain user electricity consumption behavior characteristic data through a load demand data acquisition and preprocessing module, integrate photovoltaic and wind power distributed power generation output data using a probability distribution model, extract the connection relationship diagram of the distribution network from the network topology structure information, perform normalization and standard deviation correction on the collected multi-source data, and generate a standardized multi-dimensional data set, including: Obtain user electricity consumption behavior and load demand data through a data acquisition module, and fuse time series analysis to obtain a preliminary feature set; Adopt a probability distribution model to perform statistical integration on photovoltaic output and wind power output data, and generate distributed power generation output distribution characteristics; Extract the connection relationship diagram of the distribution network from the network topology structure, and determine the distribution of power flow paths between nodes; Perform normalization processing on the multi-source data to eliminate the dimension difference and obtain a unified scale data set; Perform standard deviation correction on the unified scale data set to adjust the data fluctuation range and generate a corrected data set; Perform clustering analysis on the corrected data set through the K-means algorithm to judge the association mode between electricity consumption behavior and power generation output; Construct a standardized multi-dimensional data set according to the association mode, and determine the description of the distribution network operation characteristics.

3. The power flow calculation and optimal scheduling method for a distribution network based on a graph convolutional network according to claim 1, characterized in that: In S2, obtain load, power generation output, and voltage attributes from the standardized multi-dimensional data set, use node feature vector encoding to map to high-dimensional node vectors, generate edge feature vectors from line impedance and capacity parameters, propagate and aggregate topology structure features through a graph convolutional layer, reduce the dimension of the node vector through a pooling layer, and generate a global feature vector by a fully connected layer, including: Obtain the load attribute and power generation output and voltage attributes in the standardized multi-dimensional data set, and map them to node feature vectors through a preset encoding model to obtain high-dimensional vectors; Calculate edge feature vectors through line impedance parameters and capacity parameters to obtain edge features reflecting topological relationships; Use a graph convolutional layer to propagate and aggregate high-dimensional vectors and edge features to obtain convolutional features containing topological structures; Reduce the dimension of the convolutional features through a pooling layer to obtain a compressed vector; If the dimension of the compressed vector exceeds a preset threshold, retain the main components through feature screening to obtain an adjusted compressed vector; Use a fully connected layer to transform the adjusted compressed vector to obtain a global feature vector; For the global feature vector, judge the system state through a preset classification model to obtain the operation trend analysis result.

4. The power flow calculation and optimal scheduling method for a distribution network based on a graph convolutional network according to claim 1, characterized in that: In S3, construct a spatial correlation matrix of the distribution network from the global feature vector, use the attention mechanism to calculate weights according to the distance between nodes and the power flow direction, quantify the correlation degree between nodes and branches, and process the spatial correlation matrix by a multi-layer perceptron to generate a network correlation matrix, including: Generate a spatial correlation matrix through the global feature, calculate preliminary weights based on the distance between nodes and the power flow direction to obtain a quantified correlation matrix; Use the attention mechanism to adjust the preliminary weights, generate weighted coefficients for the power flow direction and node distance to obtain an enhanced correlation matrix; Extract the node-related and branch-related feature components from the enhanced correlation matrix, and separate the main structure through matrix decomposition to obtain a decomposed feature matrix; If the dimension of the decomposed feature matrix exceeds the preset threshold, retain the core components through feature screening to obtain a refined feature matrix; For the refined feature matrix, use a multi-layer perceptron for transformation processing to obtain an initial network correlation matrix; According to the initial network correlation matrix, calculate the association strength between nodes and between branches to obtain the final network correlation matrix; Through the final network correlation matrix, quantify the comprehensive influence of spatial correlation and power flow direction to obtain a global representation of the system topology.

5. The power flow calculation and optimal scheduling method of a distribution network based on a graph convolutional network according to claim 1, wherein: In S4, obtain the node power balance, distributed power generation output limit, and voltage constraint conditions from the network correlation matrix and the original parameters of the distribution network, and use the correlation matrix to adjust the constraint condition weights to generate a non-linear equation set for distribution network power flow calculation, including: Extract the node power, distributed power generation output limit, and voltage constraint conditions through the network correlation matrix and the distribution network parameters to obtain an initial constraint data set; Use matrix analysis methods to process network correlation and determine the weight adjustment coefficient between constraint conditions; Update the constraint conditions according to the weight adjustment result to generate a non-linear equation set required for power flow calculation; If the node power exceeds the voltage constraint range, solve the non-linear equation set by the iterative method to judge the boundary value of the distributed power generation output limit; Obtain the solved boundary value, update the distribution network parameters to obtain an adjusted input data set for power flow calculation; Verify the adjusted data set through power flow calculation to determine the balance state of node power and voltage constraints; Use the Newton-Raphson algorithm to finally solve the non-linear equation set to obtain a stable solution for distribution network power flow calculation.

6. The power flow calculation and optimal scheduling method for a distribution network based on a graph convolutional network according to claim 1, characterized in that: In S5, the initial voltage amplitude and power value are obtained by conversion through the decoding layer from the global feature vector. The Newton-Raphson method combined with the conjugate gradient iterative solver is used to process the non-linear equations. The objective error is defined through the loss function and the parameters are iteratively updated to generate the node voltage amplitude, phase angle, and branch power data, including: The initial voltage and power value data are obtained by processing the global features through the decoding layer; The non-linear equations are processed using the Newton-Raphson method combined with the conjugate gradient iterative solver from the initial voltage and power value data to obtain the solution set; The objective error value is calculated through the solution set combined with the loss function; If the objective error value exceeds the preset threshold, the parameters of the decoding layer are iteratively updated by the gradient descent method to obtain a new parameter set; The node voltage and phase angle data are regenerated through the new parameter set; The branch power value is calculated based on the node voltage and phase angle data; The final data is obtained by verifying the consistency through the comparison of the branch power value and the power value.

7. The power flow calculation and optimal scheduling method of a distribution network based on a graph convolutional network according to claim 1, characterized in that: In S6, by collecting the node voltage, phase angle data, and branch power, the load rate of each line and the node voltage offset are calculated to determine the overloaded lines and low-voltage nodes; The K-means algorithm is used to cluster the load, and the load distribution characteristics are obtained from the historical data to get the load change trend; According to the load change trend and the distributed power generation output probability model, the output range of each power source is calculated to judge the output combination that meets the N-1 constraint; Through the linear programming algorithm, an optimized solution that meets the branch power limit is obtained from the output combination to determine the node voltage recovery range; For the node voltage and branch power adjustment amounts in the optimized solution, a scheduling strategy is generated to obtain the operation parameters of the distribution network; After obtaining the operation parameters of the distribution network, if the branch power exceeds the preset threshold, the output of the distributed power generation is adjusted to judge whether the overloaded line is eliminated; Through the adjusted operation parameters, the node voltage offset of each node is calculated to determine whether the low-voltage nodes return to normal.

8. The method for power flow calculation and optimal scheduling of a distribution network based on a graph convolutional network according to claim 1, characterized in that: After S5, the calculation results are obtained from the node voltage amplitude, phase angle, and branch power data, and the voltage deviation rate and power error percentage are calculated by comparing with the measured data. If the error exceeds the preset threshold, the size of the graph convolution kernel and the parameters of the activation function are adjusted, and the global feature vector is regenerated. The weak links are identified from the distribution network scheduling strategy, and the impact of future load growth on voltage and power is predicted using the global feature vector. According to the distribution network planning standards, optimization suggestions for the grid structure and equipment upgrade plans are generated, including: The voltage amplitude and voltage phase angle are obtained from the node voltage and branch power data to calculate the initial result; The voltage deviation rate and power error percentage are obtained by comparing the initial result with the measured data; If the voltage deviation rate exceeds the preset threshold, the size of the graph convolution kernel is adjusted and the global feature vector is recalculated; If the power error percentage exceeds the preset threshold, the parameters of the activation function are adjusted to update the global feature vector; The result is recalculated using the updated global feature vector to obtain the new voltage deviation rate and power error percentage; By comparing the newly calculated result with the measured data, it is judged whether the error is within the preset threshold to obtain the final result; The stable output of the global feature vector is determined according to the final result.

9. A power flow calculation and optimal scheduling device for a distribution network based on a graph convolutional network, characterized in that, Including: A load demand data acquisition and preprocessing module for obtaining user power consumption behavior characteristic data; A data storage module that stores a processing program; A data processing module is electrically connected to the data storage module and the load demand data acquisition and preprocessing module. The data processing module is configured to receive user power consumption behavior characteristic data, load and execute the processing program stored in the memory, so as to implement the steps of the power distribution network power flow calculation and optimal scheduling method based on the graph convolutional network as described in any one of the above 1-8, and generate a power distribution network scheduling strategy.

Images