A Method and Apparatus for Distribution Network Power Flow Calculation and Optimal Scheduling Based on Graph Convolutional Networks

Abstract

This application provides a method and apparatus for power flow calculation and optimal scheduling of distribution networks based on graph convolutional networks, comprising: obtaining node power balance, distributed generation output limits, and voltage constraints from the network correlation matrix and original parameters of the distribution network; adjusting the constraint weights using the correlation matrix to generate a set of nonlinear equations for power flow calculation of the distribution network; identifying overloaded lines and low-voltage nodes from node voltage amplitude, phase angle, and branch power data; generating an optimized operation scheme that meets N-1 safety constraints using load clustering prediction and distributed generation output probability models; and generating a distribution network scheduling strategy; identifying weak links from the distribution network scheduling strategy; predicting the impact of future load growth on voltage and power using global feature vectors; and generating network structure optimization suggestions and equipment upgrade schemes based on distribution network planning standards.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method and apparatus for power flow calculation and optimized scheduling of distribution networks based on graph convolutional networks. Background Technology

[0002] In distribution network operation and management, the integration of multi-perspective data-driven power flow calculation and graph neural network modeling faces complex technical challenges. First, how to effectively integrate heterogeneous data from multiple sources, such as load demand, distributed generation output, and network topology, and standardize them to construct a high-dimensional dataset that comprehensively reflects the distribution network status? Second, in the graph neural network modeling process, how to design appropriate node and edge feature vector encoding methods to accurately represent the attributes and interrelationships of various components in the distribution network? Third, how to optimize the structure of graph convolutional layers, pooling layers, and fully connected layers to achieve effective propagation, aggregation, and compression of network features? Furthermore, when constructing the spatial correlation matrix, how to combine attention mechanisms and multilayer perceptrons to accurately capture the nonlinear relationships between distribution network components? 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 computational efficiency and accuracy? Solving these problems is crucial for improving the accuracy and practicality of distribution network analysis and optimization. Summary of the Invention

[0003] To address the aforementioned issues, this invention proposes a method and apparatus for power flow calculation and optimized scheduling of distribution networks based on graph convolutional networks. This method can optimize scheduling strategies, predict the impact of load growth, provide suggestions for network structure optimization and equipment upgrades, and effectively improve the planning and operation level of distribution networks.

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

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

[0006] S1. Obtain user electricity consumption behavior characteristic data and photovoltaic and wind power distributed power output data through the load demand data acquisition and preprocessing module. Extract the distribution network connection relationship diagram from the network topology information and generate a standardized multi-dimensional dataset.

[0007] S2. Obtain load, power output and voltage attributes from a standardized multi-dimensional dataset, encode and map node feature vectors into high-dimensional node vectors, generate edge feature vectors from line impedance and capacity parameters, propagate and aggregate topological features through graph convolutional layers, and generate global feature vectors by fully connected layers.

[0008] S3. Construct the spatial correlation matrix of the distribution network from the global feature vectors, and process the spatial correlation matrix by the multilayer perceptron to generate the network correlation matrix;

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

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

[0011] S6. Identify overloaded lines and low-voltage nodes from node voltage amplitude, phase angle and branch power data. Use load clustering prediction and distributed power output probability model to generate an optimized operation scheme that meets the N-1 safety constraint and generate a distribution network dispatch strategy.

[0012] Preferably, in S1, user electricity consumption behavior characteristic data is acquired through the load demand data acquisition and preprocessing module; output data of photovoltaic and wind power distributed sources are integrated using a probability distribution model; a distribution network connection diagram is extracted from network topology information; and normalization and standard deviation correction are performed on the acquired multi-source data to generate a standardized multi-dimensional dataset, including:

[0013] The data acquisition module obtains user electricity consumption behavior and load demand data, and the data is then integrated with time series analysis to obtain a preliminary feature set.

[0014] A probability distribution model is used to perform statistical integration on photovoltaic and wind power output data to generate distributed power generation output distribution characteristics.

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

[0016] Normalization is performed on multi-source data to eliminate differences in units and obtain a unified-scale dataset.

[0017] Perform standard deviation correction on a uniform-scale dataset to adjust the range of data fluctuations and generate a corrected dataset.

[0018] Cluster analysis was performed on the corrected dataset using the K-means algorithm to determine the correlation pattern between electricity consumption behavior and power output.

[0019] A standardized multi-dimensional dataset is constructed based on the correlation pattern to determine the characteristics of power distribution network operation.

[0020] Preferably, in S2, load, power output, and voltage attributes are obtained from a standardized multi-dimensional dataset. These are then encoded and mapped into high-dimensional node vectors using node feature vector encoding. Edge feature vectors are generated from line impedance and capacity parameters. Topological features are propagated and aggregated through graph convolutional layers. Node vectors are then compressed through pooling layers. Finally, a fully connected layer generates a global feature vector, including:

[0021] Obtain load attributes and power output and voltage attributes from a standardized multidimensional dataset, and map them into node feature vectors through a preset encoding model to obtain high-dimensional vectors;

[0022] By calculating the edge feature vector using line impedance parameters and capacity parameters, edge features reflecting topological relationships are obtained.

[0023] Graph convolutional layers are used to propagate and aggregate high-dimensional vectors and edge features to obtain convolutional features containing topological structures;

[0024] The convolutional features are reduced in dimensionality using pooling layers to obtain compressed vectors.

[0025] If the dimension of the compressed vector exceeds a preset threshold, the main components are retained through feature filtering to obtain the adjusted compressed vector.

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

[0027] Based on the global feature vector, the system state is determined by a pre-defined classification model, and the results of the operation trend analysis are obtained.

[0028] Preferably, in S3, a distribution network spatial correlation matrix is ​​constructed from global feature vectors. An attention mechanism is used to calculate weights based on inter-node distance and power flow direction, quantifying the correlation between nodes and branches. A multilayer perceptron processes the spatial correlation matrix to generate a network correlation matrix, including:

[0029] A spatial correlation matrix is ​​generated by global features, and preliminary weights are calculated based on the distance between nodes and the power flow direction to obtain the quantized correlation matrix.

[0030] An attention mechanism is used to adjust the initial weights, and weighting coefficients are generated for power flow direction and node distance to obtain an enhanced correlation matrix;

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

[0032] If the dimension of the decomposed feature matrix exceeds the preset threshold, the core components are retained through feature filtering to obtain a simplified feature matrix.

[0033] The simplified feature matrix is ​​transformed using a multilayer perceptron to obtain the initial network correlation matrix.

[0034] Based on the initial network correlation matrix, the correlation strength between nodes and between branches is calculated to obtain the final network correlation matrix;

[0035] By using the final network correlation matrix, the combined effect of spatial correlation and power flow direction is quantified, resulting in a global representation of the system topology.

[0036] Preferably, in S4, node power balance, distributed generation output limits, and voltage constraints are obtained from the network correlation matrix and the original parameters of the distribution network. The constraint weights are adjusted using the correlation matrix to generate a set of nonlinear equations for distribution network power flow calculation, including:

[0037] The initial constraint dataset is obtained by extracting node power, distributed generation output limits, and voltage constraints from the network correlation matrix and distribution network parameters;

[0038] Matrix analysis is used to handle network correlation and determine the weight adjustment coefficients between constraints.

[0039] Update the constraints based on the weight adjustment results and generate the nonlinear equation set required for power flow calculation;

[0040] If the node power exceeds the voltage constraint range, the nonlinear equation system is solved by iterative method to determine the boundary value of the distributed power output limit;

[0041] Obtain the boundary values ​​obtained from the solution, update the distribution network parameters, and obtain the adjusted power flow calculation input dataset;

[0042] The adjusted dataset is verified through power flow calculations to determine the balance state of node power and voltage constraints.

[0043] The Newton-Raphson algorithm is used to solve the nonlinear equations and obtain the stable solution for power flow calculation in the distribution network.

[0044] Preferably, in S5, the initial voltage amplitude and power values ​​are obtained from the global feature vector through a decoding layer. The nonlinear equation system is processed using the Newton-Raphson method combined with a conjugate gradient iterative solver. The target error is defined through a loss function, and the parameters are iteratively updated to generate node voltage amplitude, phase angle, and branch power data, including:

[0045] Initial voltage and power values ​​are obtained by processing global features through a decoding layer.

[0046] The solution set is obtained by processing the nonlinear equations using the Newton-Raphson method combined with a conjugate gradient iterative solver based on the initial voltage and power data.

[0047] The target error value is calculated by combining the solution set with the loss function;

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

[0049] The node voltage and phase angle data are regenerated using a new parameter set;

[0050] Calculate the branch power value based on the node voltage and phase angle data;

[0051] The final data is obtained by comparing the branch power values ​​with the power values ​​to verify consistency.

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

[0053] The K-means algorithm is used to cluster the loads, and the load distribution characteristics are obtained from historical data to obtain the load change trend;

[0054] Based on the load change trend and the distributed power generation output probability model, calculate the output range of each power source and determine the output combination that satisfies the N-1 constraint.

[0055] By using a linear programming algorithm, an optimal scheme that meets the branch power limit is obtained from the power output combination, and the node voltage recovery range is determined.

[0056] Based on the node voltage and branch power adjustment amounts in the optimization scheme, a scheduling strategy is generated to obtain the distribution network operation parameters;

[0057] After obtaining the distribution network operating parameters, if the branch power exceeds the preset threshold, the output of the distributed power source is adjusted, and it is determined whether the overloaded line has been eliminated.

[0058] By adjusting the operating parameters, the voltage offset of each node is calculated to determine whether the low-voltage node has returned to normal.

[0059] Preferably, after S5, calculation results are obtained from node voltage amplitude, phase angle, and branch power data. These results are compared with measured data to calculate the voltage deviation rate and power error percentage. If the error exceeds a preset threshold, the graph convolution kernel size and activation function parameters are adjusted to regenerate the global feature vector. Weak links are identified from the distribution network scheduling strategy. The global feature vector is used to predict the impact of future load growth on voltage and power. Based on distribution network planning standards, network structure optimization suggestions and equipment upgrade plans are generated, including:

[0060] The voltage amplitude and voltage phase angle are obtained by using node voltage and branch power data, and the initial results are calculated.

[0061] By comparing the initial results with the measured data, the voltage deviation rate and power error percentage were obtained.

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

[0063] If the power error percentage exceeds the preset threshold, the activation function parameters are adjusted and the global feature vector is updated.

[0064] The results are recalculated using the updated global feature vector to obtain new voltage deviation rate and power error percentage.

[0065] By comparing the new calculation results with the measured data, it is determined whether the error is within the preset threshold, and the final result is obtained.

[0066] The stable output of the global feature vector is determined based on the final result.

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

[0068] The load demand data acquisition and preprocessing module is used to acquire user electricity consumption behavior characteristic data;

[0069] A data storage module, wherein the data storage module stores a processing program;

[0070] The 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, so as to realize the steps of the distribution network power flow calculation and optimization scheduling method based on graph convolutional network as described above, and generate distribution network scheduling strategy.

[0071] The beneficial effects of using this invention are:

[0072] This invention discloses a method for power flow calculation and optimal scheduling of distribution networks based on graph convolutional networks. This method collects and standardizes multi-source data, extracts distribution network topology features using graph convolutional networks, constructs a spatial correlation matrix, and optimizes power flow calculation constraints. It employs the Newton-Raphson method combined with conjugate gradient iteration to solve a system of nonlinear equations, generating node voltage and branch power data. By comparing with measured data, model parameters are dynamically adjusted to improve calculation accuracy. Based on the calculation results, weak links in the network are identified, and combined with load forecasting and distributed generation output models, an optimal scheduling strategy satisfying N-1 constraints is generated. This invention can also predict the impact of load growth, providing suggestions for network structure optimization and equipment upgrades, effectively improving the planning and operation level of distribution networks. Attached Figure Description

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

[0074] Figure 2 This is a schematic diagram of the power flow calculation and optimization scheduling device for distribution networks based on graph convolutional networks according to the present invention. Detailed Implementation

[0075] To make the objectives, technical solutions, and advantages of this technical solution clearer, the following detailed description, in conjunction with specific embodiments, further illustrates this technical solution. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of this technical solution.

[0076] To further understand the content of this invention, a detailed description of the invention is provided in conjunction with the accompanying drawings and embodiments. The specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0077] like Figure 1 This embodiment of a distribution network power flow calculation and optimal scheduling method based on graph convolutional networks may specifically include:

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

[0079] User electricity consumption behavior and load demand data are acquired through a data acquisition module, and a preliminary feature set is obtained through time series analysis. A probability distribution model is used to statistically integrate photovoltaic and wind power output data to generate distributed power generation output distribution characteristics. A distribution network connection diagram is extracted from the network topology to determine the distribution of power flow paths between nodes. Normalization processing is performed on multi-source data to eliminate dimensional differences, resulting in a unified-scale dataset. Standard deviation correction is applied to the unified-scale dataset to adjust the data fluctuation range, generating a corrected dataset. K-means cluster analysis is performed on the corrected dataset to determine the correlation pattern between electricity consumption behavior and power generation output. Based on the correlation pattern, a standardized multi-dimensional dataset is constructed to determine the characteristic description of distribution network operation.

[0080] Specifically, through the load demand data acquisition and preprocessing module, the system obtains hourly electricity consumption data from smart meters. For example, if a user's electricity consumption on a certain day is [2,8,5,0,7,9,3,1,8,6,4,3] kWh, the system uses the K-means clustering algorithm to perform cluster analysis on this data, identifying the user's typical electricity consumption patterns, such as peak and off-peak hours. In the integration of photovoltaic and wind power distributed generation output data, the system collects the output data of a 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 [0,5,0,5,0,5,0,5,0,5,0,5,0,5] kW. A Gaussian mixture model is used to fit the probability distribution of this data, generating a joint output probability density function to predict the output situation in future periods. From the network topology information, the system extracts the connection graph of the distribution network. For example, a distribution network in a certain area contains 10 nodes and 12 branches, and the connection relationship between the 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 degree and shortest path of each node, providing a foundation for subsequent power flow calculations. Normalization and standard deviation correction are performed on the collected multi-source data. For example, load data, photovoltaic power output data, and wind power output data are normalized to the [0,1] interval, and the Z-score standardization method is used to correct the standard deviation of the data, so that the data in each dimension have the same dimensions and distribution characteristics. Finally, a standardized multi-dimensional dataset is generated for subsequent machine learning model training and predictive analysis.

[0081] S2. Obtain load, power output and voltage attributes from a standardized multi-dimensional dataset, encode and map node feature vectors into high-dimensional node vectors, generate edge feature vectors from line impedance and capacity parameters, propagate and aggregate topological features through graph convolutional layers, compress node vectors through pooling layers, and generate global feature vectors by fully connected layers.

[0082] Load attributes, power output, and voltage attributes from a standardized multi-dimensional dataset are obtained and mapped to node feature vectors using a pre-defined encoding model, resulting in high-dimensional vectors. The node feature vectors are calculated as follows:

[0083]

[0084] Where 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] Edge feature vectors are calculated using line impedance and capacity parameters to obtain edge features reflecting topological relationships. The edge feature calculation is as follows:

[0086]

[0087] Where Zij represents the line impedance, Zbase represents the reference impedance, Sij represents the line capacity, and Smax represents the maximum capacity.

[0088] Graph convolutional layers are used to propagate and aggregate high-dimensional vectors and edge features, resulting in convolutional features that incorporate topological structure. The convolutional features are represented as follows:

[0089]

[0090] Where Hk represents the convolutional feature, σ represents the activation function, Ni represents the set of neighbors of node i, di and dj represent the node degree, and W represents the convolutional kernel parameter matrix.

[0091] Pooling layers are used to reduce the dimensionality of convolutional features, resulting in a compressed vector. If the dimension of the compressed vector exceeds a preset threshold, feature filtering is used to retain the main components, yielding an adjusted compressed vector. A fully connected layer is then used to transform the adjusted compressed vector, resulting in a global feature vector. Based on this global feature vector, a preset classification model is used to determine the system state, providing an operational trend analysis result.

[0092] Specifically, from the standardized multidimensional dataset, load, power output, and voltage attributes are first extracted. For example, load data includes 100MW for node 1 and 150MW for node 2; power output data includes 200MW for node 3 and 250MW for node 4; and voltage data includes 0.5p.u. for node 1 and 0.3p.u. for node 2. Using this data, node features are mapped to high-dimensional node vectors through node feature vector encoding methods, such as Principal Component Analysis (PCA). For example, the feature vector for node 1 is [8,6,4], and the feature vector for node 2 is [7,5,3]. Next, edge feature vectors are generated from line impedance and capacity parameters. For example, line 1 has an impedance of 1Ω and a capacity of 100MVA, while line 2 has an impedance of 2Ω and a capacity of 150MVA. Edge feature vectors are generated through linear transformation, such as [1,2] for line 1 and [15,25] for line 2. Topological features are propagated and aggregated through Graph Convolutional Networks (GCNs). For example, a two-layer GCN can be used, with the first layer having an output dimension of 32 and the second layer having an output dimension of 16. Aggregation functions, such as mean aggregation, are used to obtain the aggregated feature vector for each node; for example, the aggregated feature vector for node 1 is [75, 55, 35]. Dimensionality reduction and compression of the node vectors are then achieved through pooling layers, such as max pooling, which compresses the aggregated feature vector for node 1 to

[75] and for node 2 to

[55] . Finally, a fully connected layer generates a global feature vector. For example, a two-layer fully connected network can be used, with the first layer having an output dimension of 8 and the second layer having an output dimension of 4, resulting in a final global feature vector of [6, 5, 4, 3], which is used for subsequent power system analysis and decision-making.

[0093] S3. Construct a distribution network spatial correlation matrix from global feature vectors, use an attention mechanism to calculate weights based on the distance between nodes and power flow direction, quantify the correlation between nodes and branches, and process the spatial correlation matrix by a multilayer perceptron to generate a network correlation matrix.

[0094] A spatial correlation matrix is ​​generated using global features. Preliminary weights are calculated based on node distances and power flow direction to obtain a quantized correlation matrix. An attention mechanism is used to adjust the preliminary weights, generating weighting coefficients for power flow direction and node distance, resulting in an enhanced correlation matrix. Node-related and branch-related feature components are extracted from the enhanced correlation matrix. Matrix decomposition separates the main structure, yielding a decomposed feature matrix. If the dimension of the decomposed feature matrix exceeds a preset threshold, core components are retained through feature filtering, resulting in a simplified feature matrix. A multilayer perceptron is used to transform the simplified feature matrix, obtaining an initial network correlation matrix. Based on the initial network correlation matrix, the correlation strength between nodes and branches is calculated to obtain the final network correlation matrix. The final network correlation matrix is ​​used to quantify the combined influence of spatial correlation and power flow direction, resulting in a global representation of the system topology.

[0095] Specifically, in distribution network analysis, when constructing a spatial correlation matrix based on global feature vectors, the characteristic relationships between nodes are first extracted from the global feature vectors. 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]. The cosine similarity algorithm is used 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 the initial spatial correlation matrix. Next, an attention mechanism is introduced to dynamically adjust the weights based on 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, so the weight coefficient is 9; the distance between node 1 and node 3 is 15 kilometers, and the power flow direction is negative, so the weight coefficient is 7. Through weighted calculation, the spatial correlation matrix is ​​updated. For example, the correlation value between node 1 and node 2 is updated to 88, and the correlation value between node 1 and node 3 is updated to 66. Subsequently, the updated spatial correlation matrix is ​​input into a multilayer perceptron, which is trained using the ReLU activation function and the Adam optimizer. For example, the output dimension of the first layer perceptron is 16, and that of the second layer is 8, ultimately generating a network correlation matrix. For instance, the network correlation value between node 1 and node 2 is 85, between node 1 and node 3 is 62, and between node 2 and node 3 is 80. This matrix can be used to quantify the correlation between nodes and branches in a distribution network, providing data support for subsequent power system optimization.

[0096] S4. Obtain node power balance, distributed generation output limit and voltage constraint conditions from the network correlation matrix and the original parameters of the distribution network. Adjust the constraint weights using the correlation matrix to generate a set of nonlinear equations for distribution network power flow calculation.

[0097] Initial constraint datasets are obtained by extracting node power, distributed generation output limits, and voltage constraints from the network correlation matrix and distribution network parameters. Matrix analysis is used to handle network correlations and determine the weight adjustment coefficients between constraints. The constraints are updated based on the weight adjustment results, generating the nonlinear equations required for power flow calculation. If node power exceeds the voltage constraint range, the nonlinear equations are solved iteratively to determine the boundary values ​​of the distributed generation output limits. The obtained boundary values ​​are used to update the distribution network parameters, resulting in an adjusted power flow calculation input dataset. Power flow calculations are used to verify the adjusted dataset and determine the balance state between node power and voltage constraints. The Newton-Raphson algorithm is used to finally solve the nonlinear equations, obtaining the stable solution for the distribution network power flow calculation.

[0098] Specifically, firstly, the electrical relationships between nodes are extracted 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. Using the original parameters of the distribution network, such as the load power of node C being 500kW, the maximum output limit of distributed generation D being 300kW, and the allowable voltage range of node E being 95 to 0.5 p.u., node power balance equations are constructed to ensure that total generation equals total load. Next, the weights of constraints are adjusted through the correlation matrix. For example, the voltage constraint weight between node A and node B is adjusted to 2 to reflect their strong correlation. Based on this, a set of nonlinear equations for distribution network power flow calculation is generated and solved iteratively using the Newton-Raphson method. The initial guess value is the node voltage of 0 p.u., and the convergence accuracy is set to 0001. During the iteration process, the power mismatch of each node is calculated, and the node voltage is updated until the convergence condition is met. Through the above steps, accurate calculation of distribution network power flow and dynamic adjustment of constraints are achieved, ensuring stable system operation and efficient management.

[0099] S5. Initial voltage amplitude and power values ​​are obtained from the global feature vector through the decoding layer. The nonlinear equation system is processed by the Newton-Raphson method combined with the conjugate gradient iterative solver. The target error is defined by the loss function and the parameters are iteratively updated to generate node voltage amplitude, phase angle and branch power data.

[0100] Initial voltage and power values ​​are obtained by processing global features through a decoding layer. From this initial data, a Newton-Raphson method combined with a conjugate gradient iterative solver is used to process the nonlinear equations, yielding a solution set. The target error value is calculated using the solution set and a loss function. If the target error value exceeds a preset threshold, the decoding layer parameters are iteratively updated using gradient descent to obtain a new parameter set. Node voltage and phase angle data are regenerated using the new parameter set. Branch power values ​​are calculated based on the node voltage and phase angle data. Consistency is verified by comparing the branch power values ​​with the final power values ​​to obtain the final data.

[0101] Specifically, in power system analysis, the topology and electrical parameters of the power network, such as the node admittance matrix and load data, are first extracted using global feature vectors. Assume the system contains 5 nodes, with the admittance matrix 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]], and the load data P = [8,2,9,5,1] MW and Q = [4,6,5,7,6] MVar. These feature vectors are input into the decoding layer, which employs a fully connected neural network structure with ReLU activation function and 3 layers. The number of neurons in each layer is 10, 20, and 5 respectively. After training, the decoding layer outputs the initial voltage amplitude V0 = [02, 03, 01, 04, 02]pu 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 process the nonlinear equations, with a convergence accuracy of 1e-6 and a maximum number of iterations of 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, updating the voltage magnitude and phase angle. For example, after the first iteration, V1 = [021,031,011,041,021]pu, and θ1 = [01,02,01,03,02]rad. The loss function uses the mean squared error, defined as L = Σ(Vi-Vi0)^2 + Σ(Pi-Pi0)^2, with an initial error of 1. After 5 iterations, the error decreases to 1e-4, satisfying the convergence condition. The final generated node voltage amplitude V = [022,032,012,042,022]pu, phase angle θ = [012,022,012,032,022]rad, and branch power P = [86,26,96,56,16]MW are used for power system stability and economic analysis, providing decision support for grid dispatching.

[0102] S6. Identify overloaded lines and low-voltage nodes from node voltage amplitude, phase angle and branch power data. Use load clustering prediction and distributed power output probability model to generate an optimized operation scheme that meets the N-1 safety constraint and generate a distribution network dispatch strategy.

[0103] By collecting node voltage, phase angle data, and branch power, the load rate of each line and the voltage offset of each node are calculated to identify overloaded lines and low-voltage nodes. The K-means algorithm is used to cluster the load, extracting load distribution characteristics from historical data to obtain load change trends. Based on the load change trends and the distributed generation output probability model, the output range of each power source is calculated, and the output combination satisfying the N-1 constraint is determined. A linear programming algorithm is used to obtain an optimized scheme that satisfies the branch power limit from the output combination, determining the node voltage recovery range. For the node voltage and branch power adjustments in the optimized scheme, a scheduling strategy is generated to obtain the distribution network operating parameters. After obtaining the distribution network operating parameters, if the branch power exceeds a preset threshold, the output of the distributed generation is adjusted to determine whether the overloaded line has been eliminated. Using the adjusted operating parameters, the voltage offset of each node is calculated to determine whether the low-voltage node has returned to normal.

[0104] Specifically, in distribution network dispatching, overloaded lines and low-voltage nodes are first identified by combining node voltage amplitude and phase angle data with branch power information and using power flow calculation algorithms.

[0105] For example, if the current on a certain line reaches 115% of its rated value, it is considered overloaded, while the voltage at a certain node drops to 92 p.u., below the allowable 95 p.u., and is marked as a low-voltage node. Next, the K-means clustering algorithm is used to cluster the load data and predict future load change trends. Assuming the load is divided into three categories representing high, medium, and low load levels, the probability of a high load level in the next hour is predicted to be 30%, a medium load level 50%, and a low load level 20%. Then, based on the distributed generation output probability model, Monte Carlo simulation is used to generate 1000 possible output scenarios, and the system operating state under each scenario is calculated.

[0106] For example, the average output of photovoltaic power is 500kW with a standard deviation of 50kW, while the average output of wind power is 300kW with a standard deviation of 30kW. Based on this, a genetic algorithm is used for optimization, with the objective function being to minimize operating costs while satisfying the N-1 safety constraint. During the optimization process, constraints such as line capacity and node voltage limits are considered, ultimately generating a scheduling strategy.

[0107] For example, the optimization results show that, under the premise of satisfying the N-1 safety constraint, the total system operating cost is 5,000 yuan, of which the distributed power output is 800kW, the main grid power purchase is 1,200kW, and the load demand is 2,000kW. Through the above method, the safe and economical operation of the distribution network is achieved.

[0108] In addition, 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 them with the measured data. If the error exceeds the preset threshold, the graph convolution kernel size and activation function parameters are adjusted to regenerate the global feature vector.

[0109] Voltage amplitude and phase angle are obtained by using node voltage and branch power data to calculate initial results. The initial results are compared with measured data to obtain voltage deviation rate and power error percentage. If the voltage deviation rate exceeds a preset threshold, the graph convolution kernel size is adjusted, and the global feature vector is recalculated. If the power error percentage exceeds a preset threshold, the activation function parameters are adjusted, and the global feature vector is updated. The updated global feature vector is used to recalculate the results, obtaining new voltage deviation rate and power error percentage. The new calculation results are compared with measured data to determine if the error is within the preset threshold, obtaining the final result. The stable output of the global feature vector is determined based on the final result. Specifically, in a power system, firstly, using node voltage amplitude and phase angle data, combined with branch power information, the Newton-Raphson method is used to perform power flow calculations to obtain the voltage amplitude and phase angle of each node.

[0110] For example, the calculated voltage amplitude of a system node is 0.2 pu, and the phase angle is 3 degrees, while the measured data are 0.1 pu and 5 degrees respectively. Calculating the voltage deviation rate reveals an amplitude deviation rate of 99% and a phase angle deviation rate of 64%. Simultaneously, the calculated branch power is 50 MW, while the measured data is 45 MW, resulting in a power error percentage of 0.1%. If the preset thresholds for voltage deviation rate are 1%, phase angle deviation rate are 4%, and power error percentage are 5%, then the phase angle deviation rate is close to the threshold and needs adjustment. In this case, a Graph Convolutional Neural Network (GCN) is used for feature extraction, with an initial kernel size of 3 and an activation function of ReLU. By adjusting the kernel size to 5 and using the LeakyReLU activation function, a new global feature vector is generated.

[0111] For example, after adjustment, the new eigenvectors reduced the deviation rates for predicting node voltage magnitude and phase angle to 8% and 2%, respectively, and the power error percentage to 9%, both meeting the preset threshold requirements. This process is achieved through automated algorithms, ensuring the accuracy and stability of power system analysis.

[0112] This process involves identifying weak links in the distribution network dispatch strategy, using global feature vectors to predict the impact of future load growth on voltage and power, and generating network structure optimization suggestions and equipment upgrade plans based on distribution network planning standards. Weak link features are extracted from dispatch strategy data, and clustering algorithms are used to determine their distribution locations. Global feature vectors are obtained from these locations, and load growth data is used for predictive analysis to obtain future load change trends. Based on these trends, regression algorithms are used to calculate the quantitative values ​​of voltage and power impacts and determine the scope of influence. If the voltage impact exceeds a preset threshold, a network structure adjustment plan is generated based on planning standards to determine the optimization suggestions. If the power impact exceeds a preset threshold, equipment upgrade parameters are matched with planning standards to obtain an equipment upgrade plan. The network structure adjustment plan and equipment upgrade plan are then used, and topology analysis tools are employed to generate an optimized distribution network layout and assess the overall improvement extent. Finally, based on the optimized layout, global feature vector updates and predictive analysis are used to obtain long-term stability indicators for the distribution network.

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

[0114] For example, a depth-first search (DFS) algorithm is used to traverse the distribution network, identifying nodes with voltage below 95 pu and branches with a power factor below 9 as potential weak links. Next, the global eigenvector method is used to extract features from these weak links, and principal component analysis (PCA) is used for dimensionality reduction to extract key feature vectors affecting voltage and power, such as load growth trends and equipment aging levels. Based on this, a time series forecasting model (such as ARIMA) is used to predict future load growth. Assuming an average annual load growth rate of 5%, and combining this with distribution network planning standards, the voltage deviation and power loss of each node over the next five years are calculated.

[0115] For example, forecasts indicate that the voltage at a certain node will drop to 92 pu in three years, exceeding the allowable range. Based on these analyses, recommendations for optimizing the power grid structure are generated, such as adding a new 10kV line near the node and upgrading the existing transformer capacity from 500kVA to 800kVA to meet future load demands. Simultaneously, to address equipment aging, it is recommended to replace equipment with an aging rate exceeding 20% ​​and to adopt an intelligent monitoring system to monitor equipment status in real time, ensuring the stable operation of the distribution network.

[0116] To implement the above method, in a feasible embodiment, the present invention proposes a power flow calculation and optimization scheduling device for distribution networks based on graph convolutional networks, comprising:

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

[0118] Data storage module 30, wherein 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 realize the steps of the distribution network power flow calculation and optimization scheduling method based on graph convolutional network as described above, and generate distribution network scheduling strategy.

[0120] The apparatus provided in this application includes a data processing module 20 and a data storage module 30. The data storage module 30 stores a program or instruction that can run on the data processing module 20. When the program or instruction is executed by the data processing module 20, it implements the various steps of the above-described information processing method embodiments and can achieve the same technical effect. To avoid repetition, it will not be described again here.

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

[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, wherein the application processor mainly handles operations involving the operating system, user interface, and applications, and the modem processor mainly handles wireless communication signals, such as a baseband processor. It is understood that the modem processor may also not be integrated into the data processing module 20.

[0123] The above content is only a preferred embodiment of the present invention. For those skilled in the art, many changes can be made in the specific implementation and application scope based on the ideas of the present invention. As long as these changes do not depart from the concept of the present invention, they all fall within the protection scope of this patent.

Claims

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

2. The distribution network power flow calculation and optimal scheduling method based on graph convolutional networks according to claim 1, characterized in that: In S1, user electricity consumption behavior characteristic data is acquired through the load demand data acquisition and preprocessing module. Probability distribution models are used to integrate photovoltaic and wind power distributed generation output data. A distribution network connection diagram is extracted from network topology information. Normalization and standard deviation correction are performed on the collected multi-source data to generate a standardized multi-dimensional dataset, including: The data acquisition module obtains user electricity consumption behavior and load demand data, and the data is then integrated with time series analysis to obtain a preliminary feature set. A probability distribution model is used to perform statistical integration on photovoltaic and wind power output data to generate distributed power generation output distribution characteristics. Extract the distribution network connection diagram from the network topology to determine the distribution of power flow paths between nodes; Normalization is performed on multi-source data to eliminate differences in units and obtain a unified-scale dataset. Perform standard deviation correction on a uniform-scale dataset to adjust the range of data fluctuations and generate a corrected dataset. Cluster analysis was performed on the corrected dataset using the K-means algorithm to determine the correlation pattern between electricity consumption behavior and power output. A standardized multi-dimensional dataset is constructed based on the correlation pattern to determine the characteristics of power distribution network operation.

3. The distribution network power flow calculation and optimal scheduling method based on graph convolutional networks according to claim 1, characterized in that: In S2, load, power output, and voltage attributes are obtained from a standardized multi-dimensional dataset. These are then encoded and mapped into high-dimensional node vectors using node feature vectors. Edge feature vectors are generated from line impedance and capacity parameters. Topological features are propagated and aggregated through graph convolutional layers. Node vectors are then compressed through pooling layers. Finally, a fully connected layer generates a global feature vector, including: Obtain load attributes and power output and voltage attributes from a standardized multidimensional dataset, and map them into node feature vectors through a preset encoding model to obtain high-dimensional vectors; By calculating the edge feature vector using line impedance parameters and capacity parameters, edge features reflecting topological relationships are obtained. Graph convolutional layers are used to propagate and aggregate high-dimensional vectors and edge features to obtain convolutional features containing topological structures; The convolutional features are reduced in dimensionality using pooling layers to obtain compressed vectors. If the dimension of the compressed vector exceeds a preset threshold, the main components are retained through feature filtering to obtain the adjusted compressed vector. A fully connected layer is used to transform the adjusted compressed vector to obtain the global feature vector; Based on the global feature vector, the system state is determined by a pre-defined classification model, and the results of the operation trend analysis are obtained.

4. The distribution network power flow calculation and optimal scheduling method based on graph convolutional networks according to claim 1, characterized in that: In S3, a distribution network spatial correlation matrix is ​​constructed from global feature vectors. An attention mechanism is used to calculate weights based on node distance and power flow direction, quantifying the correlation between nodes and branches. A multilayer perceptron processes the spatial correlation matrix to generate a network correlation matrix, including: A spatial correlation matrix is ​​generated by global features, and preliminary weights are calculated based on the distance between nodes and the power flow direction to obtain the quantized correlation matrix. An attention mechanism is used to adjust the initial weights, and weighting coefficients are generated for power flow direction and node distance to 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 the decomposed feature matrix; If the dimension of the decomposed feature matrix exceeds the preset threshold, the core components are retained through feature filtering to obtain a simplified feature matrix. The simplified feature matrix is ​​transformed using a multilayer perceptron to obtain the initial network correlation matrix. Based on the initial network correlation matrix, the correlation strength between nodes and between branches is calculated to obtain the final network correlation matrix; By using the final network correlation matrix, the combined effect of spatial correlation and power flow direction is quantified, resulting in a global representation of the system topology.

5. The distribution network power flow calculation and optimal scheduling method based on graph convolutional networks according to claim 1, characterized in that: In S4, node power balance, distributed generation output limits, and voltage constraints are obtained from the network correlation matrix and the original parameters of the distribution network. The constraint weights are adjusted using the correlation matrix to generate a set of nonlinear equations for distribution network power flow calculation, including: The initial constraint dataset is obtained by extracting node power, distributed generation output limits, and voltage constraints from the network correlation matrix and distribution network parameters; Matrix analysis is used to handle network correlation and determine the weight adjustment coefficients between constraints. Update the constraints based on the weight adjustment results and generate the nonlinear equation set required for power flow calculation; If the node power exceeds the voltage constraint range, the nonlinear equation system is solved by iterative method to determine the boundary value of the distributed power output limit; Obtain the boundary values ​​obtained from the solution, update the distribution network parameters, and obtain the adjusted power flow calculation input dataset; The adjusted dataset is verified through power flow calculations to determine the balance state of node power and voltage constraints. The Newton-Raphson algorithm is used to solve the nonlinear equations and obtain the stable solution for power flow calculation in the distribution network.

6. The distribution network power flow calculation and optimal scheduling method based on graph convolutional networks according to claim 1, characterized in that: In S5, initial voltage magnitude and power values ​​are obtained from the global feature vector through a decoding layer. The Newton-Raphson method combined with a conjugate gradient iterative solver is used to process the nonlinear equations. A loss function is used to define the target error and iteratively update the parameters, generating node voltage magnitude, phase angle, and branch power data, including: Initial voltage and power values ​​are obtained by processing global features through a decoding layer. The solution set is obtained by processing the nonlinear equations using the Newton-Raphson method combined with a conjugate gradient iterative solver based on the initial voltage and power data. The target error value is calculated by combining the solution set with the loss function; If the target error value exceeds the preset threshold, the decoding layer parameters are iteratively updated using the gradient descent method to obtain a new parameter set; The node voltage and phase angle data are regenerated using a new parameter set; Calculate the branch power value based on the node voltage and phase angle data; The final data is obtained by comparing the branch power values ​​with the power values ​​to verify consistency.

7. The distribution network power flow calculation and optimal scheduling method based on graph convolutional networks according to claim 1, characterized in that: In S6, by collecting node voltage, phase angle data and branch power, the load rate of each line and the voltage offset of each node are calculated to identify overloaded lines and low-voltage nodes. The K-means algorithm is used to cluster the loads, and the load distribution characteristics are obtained from historical data to obtain the load change trend; Based on the load change trend and the distributed power generation output probability model, calculate the output range of each power source and determine the output combination that satisfies the N-1 constraint. By using a linear programming algorithm, an optimal scheme that meets the branch power limit is obtained from the power output combination, and the node voltage recovery range is determined. Based on the node voltage and branch power adjustment amounts in the optimization scheme, a scheduling strategy is generated to obtain the distribution network operation parameters; After obtaining the distribution network operating parameters, if the branch power exceeds the preset threshold, the output of the distributed power source is adjusted, and it is determined whether the overloaded line has been eliminated. By adjusting the operating parameters, the voltage offset of each node is calculated to determine whether the low-voltage node has returned to normal.

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

9. A distribution network power flow calculation and optimal scheduling device based on graph convolutional networks, characterized in that, include: The load demand data acquisition and preprocessing module is used to acquire user electricity consumption behavior characteristic data; A data storage module, wherein the data storage module stores a processing program; The 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 data storage module, so as to realize the steps of the distribution network power flow calculation and optimization scheduling method based on graph convolutional network as described in any one of claims 1-8, and generate a distribution network scheduling strategy.

Images