Load flow calculation node spatial feature extraction method based on large electric power model

Abstract

The invention discloses a power flow calculation node spatial feature extraction method based on a large electric power model, and the method generates a multi-scene operation model based on historical data, meteorological conditions and load prediction, carries out the dimension reduction processing of the multi-scene operation model, retains key features, and guarantees the simplification and accuracy of the model. And secondly, converting the data after dimension reduction into a graph model, dividing the power system into a plurality of independent sub-regions by using a graph partitioning algorithm, and simplifying the Jacobian matrix and remarkably reducing the calculation complexity by eliminating interconnected nodes of which the power injection is zero. Meanwhile, the node parallel computing technology is adopted, load flow calculation of all the sub-regions is achieved, cross-region communication is reduced, and efficient power system load flow analysis is achieved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of node space feature extraction, and particularly relates to a power flow calculation node space feature extraction method based on a power large model. BACKGROUND

[0002] With the increasing complexity of modern power systems, especially under the background of large-scale renewable energy access, load fluctuation and weather condition change, traditional power system power flow calculation methods are facing great challenges. These traditional methods usually rely on a single scenario and static model, and cannot effectively handle the complex dynamic characteristics and uncertain factors in the power system. At the same time, with the expansion of the power grid scale, the number of nodes and the size of the network of the power system increase sharply, leading to the complexity of power flow calculation rising exponentially, which brings a huge computing burden to real-time calculation and system optimization.

[0003] In order to improve the calculation efficiency, in recent years, power flow analysis methods based on graph model and parallel computing have emerged. However, the existing methods still face the problems of too large Jacobian matrix dimension and too heavy cross-regional node communication burden when dealing with large-scale power systems. In addition, the traditional power flow calculation method fails to fully consider the influence of load, weather and other dynamic factors on the power system, resulting in limited adaptive scheduling capability of the system.

[0004] Therefore, the present application is proposed. SUMMARY

[0005] The technical problem solved by the present application is to overcome the shortcomings of the prior art and provide a power flow calculation node space feature extraction method based on a power large model, which solves the problems proposed in the background.

[0006] To solve the above technical problems, the basic idea of the technical solution of the present application is as follows:

[0007] A power flow calculation node space feature extraction method based on a power large model, comprising the following steps:

[0008] S1: Based on historical data, weather conditions and load prediction, a plurality of power system operation scenarios are generated, then the scenario data generated by multi-scenario modeling is processed by dimension reduction to retain the key features of the power system, and finally the dimension-reduced power system data is converted into a graph model, and the entire system is divided into a plurality of independent sub-regions using a graph partitioning algorithm;

[0009] S2: In the constructed graph model, a Jacobian matrix is constructed according to the power balance equation of each sub-region, the matrix size is reduced by identifying and eliminating interconnected nodes with zero power injection, and the voltage and power distribution of each node are adjusted in real time;

[0010] S3: In the optimized subsystem, apply node parallel computing technology to perform power flow calculation for each sub-region simultaneously using the simplified Jacobian matrix, and only perform data exchange across regional nodes;

[0011] S4: Based on the results of parallel power flow calculation, use the spatio-temporal graph convolution network to extract the spatio-temporal features between nodes in the power system;

[0012] S5: According to the results of spatio-temporal feature extraction, apply the deep deterministic policy gradient algorithm to dynamically adjust the operation strategy of the power system.

[0013] Optionally, based on historical data, weather conditions and load prediction, the step of generating multiple power system operation scenarios is:

[0014] Step S101: Collect historical operation data and historical weather data of the power system, then analyze future weather conditions using a weather model, and finally combine the weather conditions with the historical data to generate different weather scenarios;

[0015] Step S102: Use time series analysis algorithm to predict future load demand and generate multiple load prediction scenarios, including load demand under different time periods, different seasons and extreme weather conditions;

[0016] Step S103: Combine historical data, future weather conditions and load prediction to form multiple power system operation scenarios. After obtaining the operation scenarios, different operation scenarios can be generated according to different weather conditions and load demand.

[0017] Optionally, the step of dimensionality reduction processing of the scenario data generated by multi-scenario modeling to retain the key features of the power system is:

[0018] Step S111: Extract the operating state variables of the power system from the scenario data generated by multi-scenario modeling, including but not limited to point voltage, power generation, load demand, and line power;

[0019] Step S112: Apply principal component analysis to the extracted high-dimensional feature data by linear transformation to project high-dimensional data into low-dimensional space while retaining maximum variance, expressed as: Y = W T (X - μ), where: X represents the original high-dimensional data matrix, n is the number of samples, and p is the number of dimensions, μ represents the mean vector of the data, used to centralize the data, W is the transformation matrix, containing the principal component vectors projected into the low-dimensional space, k < p is the dimensionality after dimensionality reduction, Y is the low-dimensional data representation after dimensionality reduction, is the error term, used to capture the noise or information that the model fails to explain, W satisfies the orthogonal condition, i.e. W T W = I, representing that each principal component is independent of each other;

[0020] Step S113: After dimensionality reduction, the most important principal components are selected according to the principal component contribution degree, and the secondary features with low contribution degree are discarded, and finally the features related to system voltage fluctuation, power generation fluctuation and load change are retained, and the expression is: wherein W * is the final selected principal component matrix, containing the principal component vectors that meet the conditions, represents the contribution degree of the i-th principal component, λ i represents the contribution degree of the i-th principal component, is the eigenvalue corresponding to the i-th principal component, representing the data variance explained by the principal component, is the sum of eigenvalues of all principal components, representing the total variance of the data, p is the number of original data features, k is the first k principal components selected, and the cumulative contribution degree is a set threshold, R i is the correlation between the i-th principal component and the target feature, β is the threshold of the correlation, w i is the feature vector of the i-th principal component, x target is the feature vector of the power system, |w i | and |x target | are the norms of the corresponding vectors, respectively.

[0021] Optionally, the step of converting the dimensionality-reduced power system data into a graph model and using a graph partitioning algorithm to divide the entire system into multiple independent sub-regions is:

[0022] Step S121: The dimensionality-reduced power system data is converted into a graph model, in which each node in the power system is a node in the graph, and the transmission lines between the nodes are edges in the graph, and the nodes and edges represent the actual physical devices in the power grid and the power transmission relationship between them;

[0023] Step S122: According to the dimensionality-reduced power system data, the nodes in the graph model are assigned weights according to their power load, power generation, and node voltage, and then the edges are assigned weights based on the resistance, reactance and power transmission capacity of the transmission lines, and the expression is: wherein W is a set of weights in the graph model, s i is the weight of node i, which is assigned according to the power load, power generation and node voltage, w ij is the weight of edge ij, is a set of nodes in the power system, and ε is a set of edges in the power system;

[0024] Step S123: the entire power system graph is divided into multiple relatively independent sub-regions by using a spectral graph decomposition algorithm, and the power system is divided into closely connected sub-regions by reducing the edge connection between different regions.

[0025] Optionally, in the constructed graph model, a Jacobian matrix is constructed according to the power balance equation of each sub-region, the matrix size is reduced by identifying and eliminating the interconnection nodes with zero power injection, and the step of updating the simplified Jacobian matrix to reduce the calculation complexity is:

[0026] Step S201: after the graph partition is completed, a power balance equation is established for each node in each sub-region, and for each node, an active power balance equation and a reactive power balance equation are established to describe the influence of the voltage phase angle and the voltage amplitude on the node power, respectively;

[0027] Step S202: according to the power balance equation, an initial Jacobian matrix is constructed, wherein the Jacobian matrix is composed of four sub-matrices representing the partial derivative relationship between active power and reactive power and voltage phase angle and voltage amplitude, and the expression is: Wherein, J is the Jacobian matrix. P is the active power, Q is the reactive power, θ is the voltage phase angle of the node, and V is the voltage amplitude of the node, represents the partial derivative of the active power with respect to the voltage phase angle, represents the partial derivative of the active power with respect to the voltage amplitude, represents the partial derivative of the reactive power with respect to the voltage phase angle, represents the partial derivative of the reactive power with respect to the voltage amplitude;

[0028] Step S203: after the interconnection nodes are identified, the interconnection nodes and their related power balance equations are removed from the Jacobian matrix, and by merging the equations of the adjacent nodes of the interconnection nodes, the overall integrity of the system is maintained while the number of equations is reduced;

[0029] Step S204: after the interconnection nodes are removed, the Jacobian matrix is updated and simplified. At this time, the new Jacobian matrix only contains the equations related to the important power generation nodes, load nodes and key transmission nodes, and the redundant equations related to the interconnection nodes are removed.

[0030] Optionally, in the optimized subsystem, a node parallel computing technology is applied, the simplified Jacobian matrix is used to simultaneously perform power flow calculation of each sub-region, and only necessary data exchange is performed between cross-region nodes.

[0031] Step S301: In the subsystems after graph partitioning and optimization, each sub-region is assigned to different computing nodes as an independent computing task, and then after the computing task is assigned, the parallel computing environment is initialized and the local power flow calculation is processed using independent computing resources;

[0032] Step S302: Each subsystem uses the simplified Jacobian matrix to perform power flow calculation and iteratively solves the power flow calculation by Newton-Raphson method, and then during the parallel calculation, for the boundary nodes between the partitions, data exchange is only performed when the boundary conditions change or the cross-region power flow has an impact;

[0033] Step S303: At the end of each computing iteration, check whether the power flow calculation results of each subsystem converge. The calculation results of the cross-region nodes are also checked synchronously. If the calculation results of all subsystems meet the convergence condition within the tolerance range, the parallel calculation ends; otherwise, the next round of iteration is continued;

[0034] Step S304: After the power flow calculation of all subsystems converges, the calculation results of each region are summarized to generate a power flow calculation report for the entire system.

[0035] Optionally, based on the results of the parallel power flow calculation, the step of extracting the space-time features between nodes in the power system using a space-time graph convolution network is:

[0036] Step S401: According to the power flow calculation results, each node in the power system is taken as a node in the graph, and the power transmission line between the nodes is taken as an edge in the graph to obtain a space-time graph structure. At the same time, a time dimension is constructed, and each time step of the power flow calculation is regarded as a time series feature;

[0037] Step S402: The constructed space-time graph is input into the space-time graph convolution network. The space-time graph convolution network models the spatial structure and time dynamics simultaneously to extract the spatial dependency between nodes and the dynamic features on the time series. At the same time, the spatial convolution layer of the space-time graph convolution network extracts the power transmission relationship between nodes through graph convolution, and the time convolution layer extracts the time series change features of each node through time series analysis;

[0038] Step S403: In the space-time graph convolution network, the spatial features of each node in the power system are first extracted through the graph convolution layer, and in the time convolution layer of the space-time graph convolution network, the dynamic changes of each node at multiple time steps are extracted through time series convolution;

[0039] Step S404: In the last layer of the space-time graph convolution network, the space-time features of each node are output, including the power transmission features in the spatial dimension and the dynamic fluctuation features in the time dimension.

[0040] Optionally, a spatio-temporal graph convolutional network is also needed to be constructed between steps S3, the spatio-temporal graph convolutional network has a graph convolutional layer and a temporal convolutional layer, which are respectively used to capture the spatial dependency between nodes and the dynamic changes in the time dimension, wherein the steps of constructing the spatio-temporal graph convolutional network are:

[0041] Step A1: On the basis of the graph structure, the time dimension is added, and the time series features are constructed for each node. Then, the spatio-temporal graph data is constructed by combining the spatial graph structure and the time series features, and the spatio-temporal graph data is divided into a training set and a test set according to a ratio of 70% and 30%;

[0042] Step A2: The spatial dependency between nodes in the power system is captured by constructing a graph convolutional layer, the time series changes of the node state are extracted by combining a temporal convolutional layer, and the integrated dynamic behavior of the nodes in the power system is captured by integrating the spatio-temporal features through a fusion layer. Then, the weights and bias terms of the spatio-temporal graph convolutional network are initialized, and the learning rate, optimizer and other hyperparameters are set;

[0043] Step A3: The training set is input into the designed spatio-temporal graph convolutional network for training. At the same time, the spatio-temporal graph convolutional network calculates the spatio-temporal features of each node through forward propagation in each iteration, and updates the network weights through back propagation. Moreover, the training process uses the power flow calculation data as the label;

[0044] Step A4: After the network training is completed, the performance of the spatio-temporal graph convolutional network is evaluated using the validation set data, the accuracy of the network in capturing the spatio-temporal features of the power system is checked, and the parameters are adjusted according to the validation results.

[0045] Optionally, according to the results of spatio-temporal feature extraction, the steps of dynamically adjusting the operation strategy of the power system by applying the deep deterministic policy gradient algorithm are as follows:

[0046] Step S501: The spatio-temporal features of the power system extracted by the spatio-temporal graph convolutional network are input as input into the deep reinforcement learning framework. In the reinforcement learning framework, the deep deterministic policy gradient algorithm is used to establish a policy network and a Q value network, wherein the policy network is used to input the current spatio-temporal features and output the corresponding optimal generation dispatch and load adjustment, and the Q value network is used to evaluate the expected return after executing a specific strategy in a given state, which is used to guide the training of the policy network;

[0047] Step S502: At each time step, the policy network outputs the corresponding operation strategy based on the current input system state, and after executing the strategy each time, the reward value is calculated according to the execution result, wherein the reward value is evaluated based on multiple objective functions of the power system, and the reward value reflects the effect of the current strategy, which is used to optimize the policy network. The expected return of each strategy is calculated through the Q value network, and this feedback is used to optimize the strategy selection in the next step.

[0048] Step S503: After the end of each time step, based on the feedback and rewards of the policy execution, the policy network and the Q-value network are updated using the policy gradient method and error minimization of the Q-value network;

[0049] Step S504: After the policy update, the power system will adjust the operation according to the new policy, and finally the power system enters the next time step, and the state change of the system will be fed back to the reinforcement learning model as the spatiotemporal features again, starting a new round of decision-making process.

[0050] Optionally, the steps of establishing the policy network and the Q-value network using the deep deterministic policy gradient algorithm in step S501 are as follows:

[0051] Step S5011: A multi-layer neural network is established as the policy network to output a continuous action vector according to the current state of the power system;

[0052] Step S5012: A multi-layer neural network is constructed as the Q-value network to output the Q-value of the action in the current state, according to the current state of the power system and the action output by the policy network. Meanwhile, the Q-value network inputs the combination of the state and the action through forward propagation, and outputs the corresponding Q-value through the nonlinear mapping of the multi-layer neural network.

[0053] After adopting the above technical solutions, the present application has the following advantages compared with the prior art. Of course, any product implementing the present application does not necessarily need to achieve all the advantages described below:

[0054] By generating a multi-scenario operation model based on historical data, weather conditions and load prediction, and performing dimensionality reduction processing, the key features are retained, ensuring the simplification and accuracy of the model. Secondly, the dimensionality-reduced data is converted into a graph model, and the power system is divided into multiple independent sub-regions using a graph partitioning algorithm. By eliminating the interconnection nodes with zero power injection, the Jacobian matrix is simplified, significantly reducing the computational complexity. At the same time, node parallel computing technology is adopted to realize the power flow calculation of each sub-region, and to reduce cross-region communication, achieving efficient power system power flow analysis. Finally, based on the parallel power flow calculation results, the spatiotemporal features of the power system are extracted using a spatiotemporal graph convolution network, and the system operation strategy is dynamically adjusted using a deep deterministic policy gradient algorithm, ensuring the adaptive optimization of the power system under complex environments and improving the stability and flexibility of the system.

[0055] The specific embodiments of the present application will be described in further detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0056] The drawings in the following description are only some embodiments, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings. In the drawings:

[0057] Figure 1 Flow chart of node space feature extraction method for power flow calculation based on large power model;

[0058] It should be noted that these drawings and written descriptions are not intended to limit the scope of the concept of the present application in any way, but to illustrate the concept of the present application to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0059] The present application will now be further described in detail in conjunction with the drawings.

[0060] Please refer to Figure 1 In this embodiment, a node space feature extraction method for power flow calculation based on large power model is provided, which includes the following steps:

[0061] S1: Based on historical data, weather conditions and load prediction, a plurality of power system operation scenarios are generated, then the scene data generated by multi-scene modeling is processed by dimension reduction to retain the key features of the power system, and finally the power system data after dimension reduction is converted into a graph model, and the entire system is divided into a plurality of independent sub-regions using a graph partitioning algorithm;

[0062] S2: In the constructed graph model, a Jacobian matrix is constructed according to the power balance equation of each sub-region, the matrix size is reduced by identifying and eliminating interconnected nodes with zero power injection, and the voltage and power distribution of each node is adjusted in real time;

[0063] S3: In the optimized subsystem, node parallel computing technology is applied, and the simplified Jacobian matrix is used to simultaneously perform power flow calculation of each sub-region, and only data exchange is performed between cross-region nodes;

[0064] S4: Based on the results of parallel power flow calculation, the spatiotemporal features between nodes in the power system are extracted using a spatiotemporal graph convolution network;

[0065] S5: According to the results of spatiotemporal feature extraction, a deep deterministic policy gradient algorithm is applied to dynamically adjust the operation strategy of the power system.

[0066] In this embodiment, the step of generating a plurality of power system operation scenarios based on historical data, weather conditions and load prediction is:

[0067] Step S101: Collect historical operation data and historical meteorological data of the power system, then analyze future meteorological conditions using a meteorological model, and finally combine the meteorological conditions with the historical data to generate different meteorological scenarios;

[0068] Step S102: Use a time series analysis algorithm to predict future load demand and generate multiple load prediction scenarios, including load demand under different time periods, seasons, and extreme weather conditions.

[0069] Step S103: Combine historical data, future meteorological conditions, and load prediction to form multiple power system operation scenarios. After obtaining the operation scenarios, different operation scenarios can be generated according to different meteorological conditions and load demand.

[0070] It is necessary to note that by combining historical data, meteorological conditions, and load prediction to generate multiple power system operation scenarios, a more comprehensive and flexible power system modeling method is provided. First, historical operation data and meteorological data of the power system are collected, and future meteorological conditions are predicted using a meteorological model to generate different meteorological scenarios, enabling the model to respond to different meteorological changes. Second, a time series analysis algorithm is used to predict future load demand, generating load scenarios under different time periods, seasons, and extreme weather conditions, ensuring the adaptability of the power system under various load conditions. Finally, historical data, meteorological conditions, and load prediction are combined to generate multiple power system operation scenarios, effectively simulating and evaluating the performance of the power system under varying meteorological and load conditions. This method improves the accuracy and adaptability of the power system model, making power dispatching and optimization more forward-looking and robust.

[0071] The step of reducing the dimensionality of the scenario data generated by the multi-scenario modeling in this embodiment is:

[0072] Step S111: Extract the operating state variables of the power system from the scenario data generated by the multi-scenario modeling, including but not limited to point voltage, power generation, load demand, and line power.

[0073] Step S112: Apply principal component analysis to the extracted high-dimensional feature data by linear transformation to project high-dimensional data into low-dimensional space while retaining maximum variance, expressed as: Y = W T (X-μ), where: X represents the original high-dimensional data matrix, n is the number of samples, and p is the number of dimensions, μ represents the mean vector of the data, used to centralize the data, is the transformation matrix, containing the principal component vectors projected into the low-dimensional space, k < p is the dimensionality after dimensionality reduction, is the low-dimensional data representation after dimensionality reduction, is the error term, used to capture noise or information that the model fails to explain, W satisfies the orthogonal condition, i.e. W T W = I, representing that each principal component is independent of each other;

[0074] Step S113: After dimensionality reduction, the most important principal components are selected according to the principal component contribution, and the secondary features with lower contribution are discarded, finally the features related to system voltage fluctuation, power generation fluctuation and load change are retained, and the expression is: where W * is the final selected principal component matrix, containing the principal component vectors that meet the conditions, represents the contribution of the i-th principal component, λ i represents the contribution of the i-th principal component, is the eigenvalue corresponding to the i-th principal component, representing the data variance explained by the principal component, is the sum of eigenvalues of all principal components, representing the total variance of the data, p is the number of original data features, k is the first k selected principal components, and the cumulative contribution degree is the set threshold, R i is the correlation between the i-th principal component and the target feature (such as voltage fluctuation, power generation fluctuation, load change), β is the threshold of correlation, w i is the feature vector of the i-th principal component, x target is the specific feature vector of the power system, |w i | and |x target | are the norms of the corresponding vectors. Through principal component analysis (PCA) dimensionality reduction processing of high-dimensional data generated by multi-scenario modeling of the power system, the processing efficiency of complex data of the power system and the accuracy of the model are greatly improved. First, the key variables in the operation state of the power system are extracted, such as voltage, power generation, load demand and line power, which comprehensively covers the core features of the power system. Then, through principal component analysis, the high-dimensional feature data is projected into a low-dimensional space, retaining the principal components with the largest variance in the data, thus simplifying the data structure, reducing redundant information, and ensuring that important information is not lost. After dimensionality reduction, the system selects the key features related to voltage fluctuation, power generation fluctuation and load change according to the principal component contribution, and discards the secondary features, so that the model can focus on the core factors affecting the operation of the power system. This method not only improves the calculation efficiency and reduces the computational complexity caused by data dimensionality, but also enhances the model's interpretability, making the operation analysis of the power system more targeted and accurate, thus improving the optimization and decision-making ability of the power system.

[0075] The step of the embodiment is to convert the reduced dimension power system data into a graph model, and divide the entire system into multiple independent sub-regions using a graph partitioning algorithm:

[0076] Step S121: Convert the reduced dimension power system data into a graph model, in which each node in the power system is a node in the graph, and the transmission line between each node is an edge in the graph, and the nodes and edges represent the actual physical devices in the power grid and the power transmission relationship between them;

[0077] Step S122: According to the reduced dimension power system data, the weights of the nodes in the graph model are assigned according to the power load, power generation, and node voltage, and then the weights of the edges are assigned based on the resistance, reactance, and power transmission capacity of the transmission line, and the expression is: where W is the weight set in the graph model, s i is the weight of node i, which is assigned according to the power load, power generation, and node voltage, w ij is the weight of edge ij, is the node set in the power system, and ε is the edge set in the power system;

[0078] Step S123: Use a spectral graph decomposition algorithm to divide the entire power system graph into multiple relatively independent sub-regions, and divide the power system into tightly connected sub-regions by reducing the edge connections between different regions. By converting the reduced dimension power system data into a graph model, structured modeling and efficient analysis of complex power systems are achieved. First, the graph model represents each node (such as a power station, a transformer station) of the power system as a node in the graph, and the transmission line as an edge in the graph, so that the physical devices and power transmission relationships of the power system can be displayed in the form of an intuitive graph structure. Secondly, in step S122, the weights of the nodes are assigned according to the power load, power generation, and node voltage, and the weights of the edges are assigned based on the resistance, reactance, and power transmission capacity of the transmission line, ensuring the accuracy and physical consistency of the graph model, and enhancing the accuracy of the graph model in power system analysis. Finally, the spectral graph decomposition algorithm is used to divide the power system into multiple relatively independent sub-regions, reducing the edge connections between different regions and improving the manageability and parallel computing efficiency of the system.

[0079] In the graph model constructed in this embodiment, the step of constructing a Jacobian matrix according to the power balance equation of each sub-region, reducing the matrix size by identifying and eliminating interconnected nodes with zero power injection, and updating the simplified Jacobian matrix to reduce the computational complexity is:

[0080] Step S201: After the graph partition is completed, power balance equations are established for nodes in each sub-region. For each node, active power balance equation and reactive power balance equation are established to describe the influence of voltage phase angle and voltage amplitude on node power, respectively;

[0081] Step S202: According to the power balance equation, an initial Jacobian matrix is constructed. The Jacobian matrix is composed of four sub-matrices, which represent the partial derivative relationship between active power and reactive power and voltage phase angle and voltage amplitude. The expression is: Wherein, J is the Jacobian matrix. P is the active power, Q is the reactive power, θ is the voltage phase angle of the node, V is the voltage amplitude of the node, represents the partial derivative of active power with respect to voltage phase angle, represents the partial derivative of active power with respect to voltage amplitude, represents the partial derivative of reactive power with respect to voltage phase angle, represents the partial derivative of reactive power with respect to voltage amplitude, the expression of the partial derivative of active power with respect to voltage phase angle is: The expression of the partial derivative of active power with respect to voltage amplitude is: The expression of the partial derivative of reactive power with respect to voltage phase angle is: The expression of the partial derivative of reactive power with respect to voltage amplitude is: Wherein, G ij and B ij are the conductance and susceptance between node i and node j, θ ij = θ i -θ j is the voltage phase angle difference between node i and j;

[0082] Step S203: After identifying the interconnected nodes, remove the interconnected nodes and their related power balance equations from the Jacobian matrix, and keep the overall integrity of the system by merging the equations of the adjacent nodes of the interconnected nodes, while reducing the number of equations;

[0083] Step S204: After the interconnected nodes are removed, update and simplify the Jacobian matrix. At this time, the new Jacobian matrix only contains equations related to important power generation nodes, load nodes and key transmission nodes, and removes redundant equations related to interconnected nodes.

[0084] The efficiency and accuracy of power flow calculation are improved by optimizing and simplifying the Jacobian matrix of the power system. First, after graph partitioning, the active and reactive power balance equations are established for each sub-region node, accurately describing the influence of voltage phase angle and voltage amplitude on node power, providing a reliable mathematical basis for subsequent calculations. Then, by constructing the Jacobian matrix, the partial derivative relationship between active and reactive power and voltage phase angle and amplitude is systematically processed, forming four sub-matrices, ensuring the accuracy of power flow calculation of the power system. The mathematical expression of the Jacobian matrix ensures that the power balance between nodes in the system can be achieved through accurate numerical calculation. Subsequently, by identifying and removing the interconnection nodes and their related equations, the size of the Jacobian matrix is further reduced, and the computational complexity is reduced. Finally, by updating the Jacobian matrix, only the equations of important nodes are retained, and redundant information is removed, making the power flow calculation more concise and efficient. This method not only improves the speed of power flow calculation and reduces the consumption of computing resources, but also improves the reliability of power system analysis, especially suitable for real-time optimization and management of large-scale complex power networks.

[0085] In the optimized subsystem of the present embodiment, the node parallel computing technology is applied to simultaneously perform power flow calculation of each sub-region using the simplified Jacobian matrix, and the step of only performing necessary data exchange for cross-region nodes is

[0086] Step S301: In the subsystem after graph partitioning and optimization, each sub-region is assigned to different computing nodes as an independent computing task. Then, after assigning the computing task, the parallel computing environment is initialized. Subsequently, each subsystem uses independent computing resources to process local power flow calculation. Necessary input data, including node voltage, power balance equation, and simplified Jacobian matrix of the region, are assigned to each computing unit during the initialization process.

[0087] Step S302: Each subsystem uses the simplified Jacobian matrix to perform power flow calculation and iteratively solves the power flow calculation by the Newton-Raphson method. Then, during parallel computing, only data exchange is performed for boundary nodes between partitions when boundary conditions change or cross-region power flow has an impact.

[0088] Step S303: At the end of each computing iteration, check whether the power flow calculation results of each subsystem converge. The calculation results of cross-region nodes are also checked synchronously. If the calculation results of all subsystems meet the convergence conditions within the tolerance range, the parallel computing ends; otherwise, the next iteration is continued.

[0089] Step S304: After the power flow calculation of all subsystems converges, the calculation results of each region are summarized to generate a power flow calculation report for the entire system.

[0090] The efficiency and accuracy of power system power flow calculation are optimized through parallel computing technology. First, the power system is divided into multiple sub-regions after graph partitioning, and each sub-region is assigned to a different computing node as an independent computing task, making full use of distributed computing resources and improving computing speed. Each subsystem obtains the necessary input data during initialization, including node voltage, power balance equation and simplified Jacobian matrix, ensuring the accuracy of the computing task. Second, each subsystem uses a simplified Jacobian matrix for power flow calculation and iteratively solves it using the Newton-Raphson method, effectively reducing computational complexity. At the same time, data exchange only occurs when the boundary conditions change or the cross-regional power flow is affected, avoiding unnecessary communication overhead and further improving the efficiency of parallel computing. Finally, by checking the convergence of the power flow calculation results of each subsystem and aggregating the calculation results globally, a power flow calculation report for the entire system is generated, ensuring the accuracy and systematicness of the calculation. This method greatly improves the efficiency of power flow calculation for large-scale power systems and the convergence speed of the calculation results, enabling the power system to respond quickly and perform real-time scheduling and optimization.

[0091] In this embodiment, based on the results of parallel power flow calculation, the steps of extracting the space-time features between nodes in the power system using a space-time graph convolution network are as follows:

[0092] Step S401: According to the power flow calculation results, each node in the power system is taken as a node in the graph, and the power transmission line between the nodes is taken as an edge in the graph, to obtain a space-time graph structure. At the same time, a time dimension is constructed, and each time step of the power flow calculation is regarded as a time series feature.

[0093] Step S402: The constructed space-time graph is input into the space-time graph convolution network. The space-time graph convolution network models the spatial structure and temporal dynamics simultaneously, extracting the spatial dependency between nodes and the dynamic features on the time series. At the same time, the spatial convolution layer of the space-time graph convolution network extracts the power transmission relationship between nodes through graph convolution, and the time convolution layer extracts the time series change features of each node through time series analysis.

[0094] Step S403: In the space-time graph convolution network, the spatial features of each node in the power system are first extracted through the graph convolution layer. In the time convolution layer of the space-time graph convolution network, the dynamic changes of each node at multiple time steps are extracted through time series convolution.

[0095] Step S404: In the last layer of the space-time graph convolution network, the space-time features of each node are output, including the power transmission features in the spatial dimension and the dynamic fluctuation features in the time dimension.

[0096] By introducing a spatio-temporal graph convolutional network (STGCN) to model and extract spatial and temporal features in the power system, the accuracy and dynamic response capability of power system load flow analysis are greatly improved. First, step S401 constructs the nodes and transmission lines in the power system into a spatio-temporal graph structure, combining the time dimension, each time step of the load flow calculation is regarded as a time series feature, so that the spatio-temporal dynamic relationship of the power system is fully modeled. Second, steps S402 to S403 extract the power transmission relationship between nodes in the spatial dimension through the spatio-temporal graph convolutional network (STGCN), identify the dependency of each node, and extract the time series change characteristics of the node in the time dimension through the time convolution, effectively capturing the dynamic fluctuations of the power system at different time steps. Finally, in the last layer of the STGCN, the output spatio-temporal features comprehensively consider the information in the spatial and temporal dimensions, these features not only include the structural information in the power transmission network, but also reflect the dynamic patterns of the power system over time. This method enhances the prediction and scheduling capability of the power system in complex scenarios, improves the comprehensive understanding of the state of each node in the power grid, and effectively deals with dynamic factors such as power load fluctuations and power generation changes, improving the stability, flexibility and real-time decision-making capability of the power grid.

[0097] In the embodiment, a spatio-temporal graph convolutional network is also needed to be constructed between steps S3, the spatio-temporal graph convolutional network has a graph convolutional layer and a time convolutional layer, which are used to capture the spatial dependency between nodes and the dynamic changes in the time dimension respectively, wherein the steps of constructing the spatio-temporal graph convolutional network are:

[0098] Step A1: On the basis of the graph structure, add a time dimension to construct a time series feature for each node, then combine the spatial graph structure and the time series feature to construct a spatio-temporal graph data, and divide the spatio-temporal graph data into a training set and a test set according to a ratio of 70% and 30%;

[0099] Step A2: Capture the spatial dependency between nodes in the power system through the construction of a graph convolutional layer, extract the time series change of the node state in combination with a time convolutional layer, and integrate the spatio-temporal features through a fusion layer to capture the comprehensive dynamic behavior of the nodes in the power system, then initialize the weights and bias terms of the spatio-temporal graph convolutional network, set the learning rate, optimizer and other hyperparameters;

[0100] Step A3: input the training set into the designed spatio-temporal graph convolutional network for training, at the same time, the spatio-temporal graph convolutional network calculates the spatio-temporal features of each node through forward propagation in each iteration, and updates the network weights through back propagation, and the training process uses the load flow calculation data as the label;

[0101] Step A4: After the network training is completed, the validation set data is used to evaluate the performance of the spatio-temporal graph convolution network, check the accuracy of the network in capturing the spatio-temporal features of the power system, and adjust the parameters according to the validation results. By constructing and training the spatio-temporal graph convolution network (STGCN), the extraction accuracy of the spatio-temporal features of the power system and the dynamic prediction ability are improved. First, step A1 adds a time dimension on the basis of the graph structure, combines the time series features of the nodes, constructs a spatio-temporal graph data, and reasonably divides it into a training set and a test set, ensuring that the model can be trained and verified within a wide range of data. Second, step A2 captures the spatial dependence between nodes in the power system through the graph convolution layer, and extracts the dynamic changes of the time series of the nodes through the time convolution layer, so that the model can pay attention to the spatial structure and time trend of the power system at the same time. Through the fusion layer, the integrated dynamic behavior of the nodes is captured to ensure a comprehensive understanding of the complex state of the power system. In step A3, the STGCN model updates the weights through forward propagation and backward propagation in each training iteration, improves the learning effect of the model on spatio-temporal features, and uses power flow calculation data as labels to ensure the accuracy of the model training. Finally, step A4 evaluates the network performance through the validation set, checks the accuracy of the network in capturing the spatio-temporal features of the power system, and adjusts the parameters according to the validation results to further optimize the model performance.

[0102] According to the results of spatio-temporal feature extraction, the step of dynamically adjusting the operation strategy of the power system in this embodiment is:

[0103] Step S501: The spatio-temporal features of the power system extracted by the spatio-temporal graph convolution network are input into the deep reinforcement learning framework. In the reinforcement learning framework, the deep deterministic policy gradient algorithm is used to establish a policy network and a Q-value network, respectively. The policy network is used to input the current spatio-temporal features and output the corresponding optimal generation scheduling and load adjustment. The Q-value network is used to evaluate the expected return after executing a specific policy in a given state, and is used to guide the training of the policy network.

[0104] Step S502: At each time step, the policy network outputs the corresponding operation strategy based on the current input system state, and calculates the reward value after executing the strategy each time. The reward value is evaluated based on multiple objective functions of the power system, and reflects the effect of the current strategy, which is used to optimize the policy network. The expected return of each strategy is calculated through the Q-value network, and this feedback is used to optimize the selection of the next strategy.

[0105] Step S503: After each time step, the policy network and the Q-value network are updated using the policy gradient method and the error minimization of the Q-value network based on the feedback and rewards of the policy execution.

[0106] Step S504: After the policy is updated, the power system will adjust the operation according to the new policy, and finally the power system enters the next time step. The state change of the system will be fed back to the reinforcement learning model again as the spatiotemporal features, and a new round of decision-making process will begin. By combining the spatiotemporal graph convolution network (STGCN) and the deep reinforcement learning framework, the intelligent scheduling and optimization ability of the power system is significantly improved. The spatiotemporal features of the power system extracted by the STGCN are used as input, which are introduced into the deep reinforcement learning framework. The deep deterministic policy gradient (DDPG) algorithm is used to establish the policy network and the Q-value network respectively, so that the model can output the optimal generation scheduling and load adjustment strategy based on the current system state. At the same time, the Q-value network evaluates the long-term return of the strategy, ensuring the reliability and foresight of the decision. Further, the reward value is calculated through the execution result of the strategy, and the reward value is evaluated according to the multi-objective function of the power system, ensuring that the selection of each strategy can optimize the overall performance of the system. The policy network and the Q-value network are continuously optimized based on the feedback and rewards of the policy execution, and through the policy gradient method and error minimization, the model can better adjust the operation strategy at each time step. Finally, step S504 updates the policy to adjust the operation of the power system in real time, ensuring that the system state can be fed back in the next time step and continue to be optimized.

[0107] In this embodiment, step S501 uses the deep deterministic policy gradient algorithm to establish the policy network and the Q-value network, which are:

[0108] Step S5011: According to the current state of the power system, a multi-layer neural network is established as the policy network to output a continuous action vector.

[0109] Step S5012: According to the current state of the power system and the action output by the policy network, a multi-layer neural network is constructed as the Q-value network to output the Q-value of the action in the current state. At the same time, the Q-value network inputs the combination of the state and the action through forward propagation, and outputs the corresponding Q-value through the nonlinear mapping of the multi-layer neural network.

[0110] The application innovatively improves the decision-making accuracy and overall optimization capability of power system dispatching through the strategy network and Q-value network design of the multi-layer neural network. First, the strategy network can generate continuous dispatching actions according to the complex dynamic state of the current power system, making the dispatching strategy more flexible and accurate, and adapting to diversified power demand and system changes. Compared with traditional fixed rules or discrete action decision-making, continuous action output can provide more detailed adjustments, making the grid dispatching more smooth and accurate, thereby reducing the impact of sudden changes on grid operation. Second, the Q-value network evaluates the actions output by the strategy network to deeply measure the long-term impact of the current decision on future system state, considering not only short-term benefits but also long-term stability and overall performance. In this way, the system no longer focuses on the current demand when making dispatching decisions, but also considers the optimization and continuous stability of the system in the future. Finally, this method combines the strategy network and Q-value network to provide an optimization framework that balances short-term and long-term goals, greatly improving the adaptability and overall dispatching capability of the power system, ensuring efficient operation and continuous optimization of the system.

[0111] Experimental Example: A certain regional power grid is facing the problem of power load fluctuation after large-scale renewable energy access, leading to unstable operation of the power grid. In order to improve the adaptability and efficiency of the power grid, the power flow calculation node space feature extraction method of the present application is applied to optimize the operation of the power grid.

[0112] Experimental Group: First, collect the historical operation data of the power system, including power generation, load demand, voltage level, and corresponding meteorological data (such as wind speed, solar radiation intensity). In addition, use time series analysis models to predict load demand under different time periods, different seasons, and extreme weather conditions, forming multiple load prediction scenarios. Combine historical data, weather conditions, and load prediction to generate multiple power system operation scenarios. Apply principal component analysis (PCA) to the generated high-dimensional scenario data for dimensionality reduction, retaining key features of the power system such as voltage fluctuation, power generation fluctuation, and load change.

[0113] Convert the dimensionally reduced power system data into a graph model, representing the nodes and transmission lines of the power grid as nodes and edges in the graph, respectively, and assigning weights to the nodes based on their voltage, load, and other information, and assigning weights to the edges based on the resistance and reactance of the transmission lines.

[0114] Graph partitioning and Jacobian matrix simplification: use spectral graph decomposition algorithms to divide the entire power system into multiple sub-regions. Nodes within each region have strong connections, and regions are connected through fewer edges. According to the power balance equation of each sub-region, construct the Jacobian matrix and simplify the calculation scale by eliminating the equations of interconnected nodes.

[0115] Each sub-region is assigned to an independent computing node, and the power flow in each region is iteratively solved using the Newton-Raphson method with node parallel computing technology. During parallel computing, only necessary data exchange is performed when there is power flow or load change across the region nodes, greatly improving the computing efficiency.

[0116] Results convergence and aggregation: when the power flow calculation results of each region meet the convergence conditions, the cross-region calculation results are aggregated to generate the power flow calculation report of the entire power system.

[0117] According to the power flow calculation results, each node and its power transmission line in the power grid is constructed into a space-time graph model. The spatial dependence relationship and dynamic characteristics in the time dimension between nodes in the power system are extracted using the space-time graph convolution network.

[0118] The space-time features extracted by the space-time graph convolution network are input into the deep reinforcement learning framework, and the deep deterministic policy gradient (DDPG) algorithm is applied to optimize the operation strategy of the power grid. According to the space-time features of the system, the generation dispatching and load adjustment strategy are dynamically adjusted.

[0119] In large-scale power grids, the main source of computational complexity is the number of nodes and power transmission lines. Traditional methods treat the entire power grid as a whole for calculation, resulting in a large Jacobian matrix that needs to be processed. By dividing the power grid into multiple independent sub-regions, the graph partitioning technique can reduce the calculation scale of each region, greatly reducing the dimension of the Jacobian matrix, thereby simplifying the calculation process.

[0120] The process of traditional power flow calculation and dispatching method includes the following steps:

[0121] Collect the load data, power generation data and network topology information of the power system. Construct a global Jacobian matrix for the entire power system (all nodes and lines) for power flow calculation using the Newton-Raphson method. In the entire power grid, the power flow is calculated by the Newton-Raphson method, and the voltage, power and other parameters of all nodes are iteratively updated. This calculation requires multiple iterations until the system converges.

[0122] Then, based on historical load data, predict future power demand. Next, based on the predicted load demand, develop a fixed generation plan and allocate the output power of power plants, without considering real-time load fluctuations and weather changes. Then, execute the pre-set generation plan, without adjusting the power generation according to real-time power demand. The implementation results are shown in Table 1:

[0123] Table 1 Comparison of control group and experimental group

[0124]

[0125]

[0126] In the control group, the traditional global power flow calculation method is used to complete the power flow calculation of the entire power system, which takes 120 minutes. This is because this method requires unified calculation of all nodes and transmission lines in the power grid. The construction of the global Jacobian matrix is a very complex and computationally intensive process. Since the size of the Jacobian matrix is proportional to the number of nodes in the power system, and the larger the system, the more complex the matrix, the calculation process requires repeated iteration, resulting in a longer calculation time. In the experimental group, by introducing the graph partitioning technique and node parallel computing, the calculation time is significantly reduced to 72 minutes, a reduction of 40%. This is because the partitioned power grid is divided into multiple independent sub-regions, each region can be calculated separately, and parallel computing allows multiple regions to be calculated simultaneously, greatly shortening the overall calculation time.

[0127] The traditional global power flow calculation method requires processing all nodes of the entire power system in an iterative process, which not only increases complexity but also leads to high system computing burden. Through the method of the experimental group, the global Jacobian matrix is divided into multiple local Jacobian matrices, which are applied in sub-regions to handle local power grid power flow problems. This localized processing greatly reduces the computational complexity of each calculation process and reduces the number of iterations for global calculation. In multiple regions, local Jacobian matrix calculations are performed simultaneously in parallel, speeding up the convergence speed of the overall power flow calculation. In addition, on the basis of parallel computing, data exchange between cross-regional nodes is only performed when necessary, effectively reducing the communication burden of the system, making the use of computing resources more efficient.

[0128] In terms of power dispatch, the traditional method is usually based on static dispatch strategy, and the allocation of generation and load is determined according to the preset load demand and generation capacity. This method cannot cope with dynamic changes such as load fluctuations and renewable energy generation fluctuations. The traditional static dispatch method is difficult to adjust generation and load in real time, resulting in insufficient flexibility and adaptability of the power system under different operating conditions, which in turn leads to power loss problems, with a power loss rate of up to 15%. In the experimental group, a dynamic dispatch strategy combining spatiotemporal graph convolutional networks and deep reinforcement learning is adopted. By using real-time load data and generation fluctuations of the power system, the experimental group can effectively predict and adjust the generation and load distribution in the system, making the system have stronger adaptability and real-time optimization capability.

[0129] Through this dynamic scheduling, the experimental group reduces power loss to 12.75%, achieving about 15% energy saving effect. The introduction of dynamic scheduling strategy enables the system to more accurately respond to changes in power demand in complex environments, reducing unnecessary power transmission losses and excessive power generation. This not only optimizes the operation efficiency of the power system, but also greatly reduces energy consumption. In practical applications, this scheduling optimization reduces the cost of power generation from 100,000 yuan to 85,000 yuan, saving about 15% of energy costs, which has a significant impact on the economic benefits of the power system.

[0130] In addition, the parallel computing method of the experimental group not only reduces the calculation time, but also significantly improves the convergence speed of the system. The traditional method usually needs to go through 15 iterations to reach the convergence state of the system, while the experimental group uses the parallel computing method after partitioning the graph, only 9 iterations are needed to reach convergence. This is because in the process of partitioning and parallel processing, the calculation of sub-regions is independent and simultaneous, effectively avoiding the cumulative error in global calculation, thereby accelerating the convergence of power flow calculation.

[0131] Finally, the traditional power flow calculation method often needs to exchange data frequently between multiple regions to ensure the balance and power flow distribution of the global system. However, this frequent data exchange not only consumes a lot of system communication resources, but also increases the calculation complexity. In the parallel computing scheme of the experimental group, the system only exchanges data across regions under key nodes and necessary conditions, effectively reducing unnecessary communication demand, reducing network load, and further improving the calculation efficiency and overall operation performance of the power grid.

[0132] The present application is not limited to the above-mentioned embodiments, and anyone should know that any structural changes made under the inspiration of the present application fall within the protection scope of the present application. The technical, shape, and structure parts not described in detail in the present application are well-known technologies.

Claims

1. A method for extracting spatial features of nodes in power flow calculations based on a large power model, characterized in that, Includes the following steps: S1: Based on historical data, meteorological conditions and load forecasts, multiple power system operation scenarios are generated. Then, the scenario data generated by multi-scenario modeling is dimensionality reduced to retain the key features of the power system. Finally, the dimensionality-reduced power system data is transformed into a graph model, and the entire system is divided into multiple independent sub-regions using a graph partitioning algorithm. S2: In the constructed graph model, a Jacobian matrix is ​​constructed based on the power balance equation of each sub-region. The matrix size is reduced by identifying and eliminating interconnect nodes with zero power injection, and the voltage and power distribution of each node are adjusted in real time. S3: In the optimized subsystem, node parallel computing technology is applied, and the simplified Jacobian matrix is ​​used to perform power flow calculations in each sub-region simultaneously. Cross-region nodes only exchange data. S4: Based on the results of parallel power flow calculation, spatiotemporal features between nodes in the power system are extracted using a spatiotemporal graph convolutional network. S5: Based on the results of spatiotemporal feature extraction, apply the deep deterministic strategy gradient algorithm to dynamically adjust the operation strategy of the power system.

2. The method for extracting node spatial features for power flow calculation based on a large power model according to claim 1, characterized in that, The steps for generating multiple power system operation scenarios based on historical data, meteorological conditions, and load forecasts are as follows: Step S101: Collect historical operating data and historical meteorological data of the power system, then use meteorological models to analyze future meteorological conditions, and finally combine the meteorological conditions with historical data to generate different meteorological scenarios. Step S102: Use time series analysis algorithms to predict future load demand and generate multiple load forecast scenarios, including load demand under different time periods, different seasons, and extreme weather conditions; Step S103: Combine historical data, future weather conditions and load forecasts to form multiple power system operation scenarios. After obtaining the operation scenarios, different operation scenarios can be generated according to different weather conditions and load demands.

3. The method for extracting node spatial features for power flow calculation based on a large power model according to claim 1, characterized in that, The steps for dimensionality reduction of scene data generated from multi-scenario modeling, while retaining the key features of the power system, are as follows: Step S111: Extract the operating state variables of the power system from the scenario data generated by multi-scenario modeling. The operating state variables include, but are not limited to, point voltage, power generation, load demand, and line power. Step S112: Apply principal component analysis to the extracted high-dimensional feature data. Through linear transformation, project the high-dimensional data into a low-dimensional space while preserving the maximum variance. The expression is: Y = W T (X-μ), where: This represents the original high-dimensional data matrix, where n is the number of samples and p is the number of dimensions. A mean vector representing the data, used to centralize the data. It is the transformation matrix, which contains the principal component vectors projected onto the lower-dimensional space, where k < p is the dimension after dimensionality reduction. It is a low-dimensional representation of the data after dimensionality reduction. This is the error term, used to capture noise that the model fails to explain. W satisfies the orthogonality condition, i.e., W T W = I, which means that the principal components are independent of each other; Step S113: After dimensionality reduction, the most important principal components are selected based on their contribution, and secondary features with lower contributions are discarded. Finally, features related to system voltage fluctuations, power generation fluctuations, and load changes are retained, and their expressions are as follows: Among them, W * It is the final principal component matrix to be selected and retained, containing principal component vectors that meet the criteria. λ represents the contribution of the i-th principal component. i This represents the contribution of the i-th principal component. It is the eigenvalue corresponding to the i-th principal component, representing the variance of the data explained by that principal component. It is the sum of the eigenvalues ​​of all principal components, representing the total variance of the data. p is the number of features in the original data, and k is the selected top k principal components that satisfy the cumulative contribution. It is the set threshold, R i β is the correlation between the i-th principal component and the target feature, β is the correlation threshold, and w i x is the eigenvector of the i-th principal component. target It is a specific eigenvector of the power system, |w i | and |x target | are the norms of the corresponding vectors.

4. The method for extracting node spatial features for power flow calculation based on a large power model according to claim 1, characterized in that, The steps to transform the dimensionality-reduced power system data into a graph model and then divide the entire system into multiple independent sub-regions using a graph partitioning algorithm are as follows: Step S121: Convert the dimensionality-reduced power system data into a graph model. In this graph model, each node in the power system is a node in the graph, and the transmission lines between each node are edges in the graph. The nodes and edges represent the actual physical equipment in the power grid and the power transmission relationship between them. Step S122: Based on the dimensionality-reduced power system data, assign weights to the nodes in the graph model according to their power load, generation capacity, and node voltage. Then, assign weights to the edges based on the resistance, reactance, and power transmission capacity of the transmission lines. The expression is as follows: Where W is the set of weights in the graphical model, s i The weight of node i is allocated based on power load, generation capacity, and node voltage. ij It is the weight of edge ij. ε is the set of nodes in the power system, and ε is the set of edges in the power system. Step S123: Use the spectral decomposition algorithm to divide the entire power system graph into multiple relatively independent sub-regions, and reduce the edge connections between different regions to divide the power system into closely connected sub-regions.

5. The method for extracting node spatial features for power flow calculation based on a large power model according to claim 1, characterized in that, In the constructed graph model, the Jacobian matrix is ​​constructed based on the power balance equation of each sub-region. The matrix size is reduced by identifying and eliminating interconnect nodes with zero power injection. The simplified Jacobian matrix is ​​then updated to reduce computational complexity. Step S201: After the graph is partitioned, establish power balance equations for nodes in each sub-region. For each node, establish active power balance equations and reactive power balance equations to describe the effects of voltage phase angle and voltage amplitude on node power, respectively. Step S202: Based on the power balance equation, construct the initial Jacobian matrix, which consists of four submatrices representing the partial derivatives of active power and reactive power with respect to voltage phase angle and voltage amplitude, respectively. Their expressions are as follows: Where J is the Jacobian matrix, P is the active power, Q is the reactive power, θ is the voltage phase angle at the node, and V is the voltage magnitude at the node. This represents the partial derivative of active power with respect to the voltage phase angle. This represents the partial derivative of active power with respect to voltage amplitude. This represents the partial derivative of reactive power with respect to the voltage phase angle. This represents the partial derivative of reactive power with respect to voltage amplitude; Step S203: After identifying the interconnect nodes, remove the interconnect nodes and their associated power balance equations from the Jacobian matrix. By merging the equations of the adjacent nodes of the interconnect nodes, the overall integrity of the system is maintained, while the number of equations is reduced. Step S204: After the interconnection nodes are removed, update and simplify the Jacobian matrix. At this time, the new Jacobian matrix only contains equations related to important generation nodes, load nodes and key transmission nodes, and removes redundant equations related to interconnection nodes.

6. The method for extracting node spatial features for power flow calculation based on a large power model according to claim 1, characterized in that, In the optimized subsystem, node-parallel computing technology is applied, and power flow calculations for each sub-region are performed simultaneously using a simplified Jacobian matrix. The steps for cross-region nodes to exchange data are as follows: Step S301: In the subsystem after graph partitioning and optimization, each sub-region is assigned as an independent computing task to different computing nodes. Then, after assigning computing tasks, the parallel computing environment is initialized and independent computing resources are used to process local power flow calculations. Step S302: Each subsystem uses the simplified Jacobian matrix to perform power flow calculation and iteratively solves the power flow calculation using the Newton-Raphson method. Then, during the parallel computing process, data exchange is performed only at the boundary nodes between partitions when the boundary conditions change or when cross-regional power flow has an impact. Step S303: At the end of each calculation iteration, check whether the power flow calculation results of each subsystem have converged, and simultaneously check the calculation results of cross-region nodes. If the calculation results of all subsystems meet the convergence condition within the tolerance range, the parallel calculation ends; otherwise, continue to the next round of iteration. Step S304: After the power flow calculations of all subsystems converge, the calculation results between each sub-region will be synchronized through cross-regional nodes to form the global power flow distribution of the entire power system and summarize the calculation results of each region to generate a power flow calculation report for the entire system.

7. The method for extracting node spatial features for power flow calculation based on a large power model according to claim 1, characterized in that, Based on the results of parallel power flow calculation, the steps for extracting the spatiotemporal features between nodes in a power system using a spatiotemporal graph convolutional network are as follows: Step S401: Based on the power flow calculation results, each node in the power system is taken as a node in the graph, and the power transmission lines between nodes are taken as edges in the graph to obtain the spatiotemporal graph structure. At the same time, the time dimension is constructed, and each time step of the power flow calculation is regarded as a time series feature. Step S402: Input the constructed spatiotemporal graph into the spatiotemporal graph convolutional network. The spatiotemporal graph convolutional network extracts the spatial dependencies between nodes and the dynamic features in the time series by simultaneously modeling the spatial structure and temporal dynamics. At the same time, the spatial convolutional layer of the spatiotemporal graph convolutional network extracts the power transmission relationship between nodes through graph convolution, and the temporal convolutional layer extracts the temporal change features of each node through time series analysis. Step S403: In the spatiotemporal graph convolutional network, the spatial features of each node in the power system are first extracted through the graph convolutional layer. In the temporal convolutional layer of the spatiotemporal graph convolutional network, the dynamic changes of each node at multiple time steps are extracted through time series convolution. Step S404: In the last layer of the spatiotemporal graph convolutional network, output the spatiotemporal features of each node. These features include power transmission features in the spatial dimension and dynamic fluctuation features in the temporal dimension.

8. The method for extracting node spatial features for power flow calculation based on a large power model according to claim 1, characterized in that, Before step S3, a spatiotemporal graph convolutional network needs to be constructed. This network has graph convolutional layers and temporal convolutional layers, used to capture spatial dependencies and dynamic changes over time, respectively. The steps for constructing the spatiotemporal graph convolutional network are as follows: Step A1: Based on the graph structure, add the time dimension to construct time series features for each node. Then, combine the spatial graph structure and time series features to construct spatiotemporal graph data, and divide the spatiotemporal graph data into training set and test set according to the ratio of 70% and 30% respectively. Step A2: By constructing a graph convolutional layer to capture the spatial dependencies between nodes in the power system, combining it with a temporal convolutional layer to extract the temporal series changes of node states, and integrating spatiotemporal features through a fusion layer to capture the comprehensive dynamic behavior of nodes in the power system, then initializing the weights and biases of the spatiotemporal graph convolutional network, and setting the learning rate, optimizer, and hyperparameters. Step A3: Input the training set into the designed spatiotemporal graph convolutional network for training. At the same time, in each iteration, the spatiotemporal graph convolutional network calculates the spatiotemporal features of each node through forward propagation and updates the network weights through backward propagation. Moreover, the training process uses power flow calculation data as labels. Step A4: After the network training is completed, use validation set data to evaluate the performance of the spatiotemporal graph convolutional network, check the accuracy of the network in capturing the spatiotemporal features of the power system, and adjust the parameters according to the validation results.

9. The method for extracting node spatial features for power flow calculation based on a large power model according to claim 1, characterized in that, Based on the results of spatiotemporal feature extraction, the steps for dynamically adjusting the power system's operating strategy using the deep deterministic strategy gradient algorithm are as follows: Step S501: The spatiotemporal features of the power system extracted by the spatiotemporal graph convolutional network are used as input and imported into the deep reinforcement learning framework. In the reinforcement learning framework, the deep deterministic policy gradient algorithm is used to establish the policy network and the Q-value network respectively. The policy network is used to input the current spatiotemporal features and output the corresponding optimal generation scheduling and load adjustment. The Q-value network is used to evaluate the expected return after executing a specific policy in a given state and to guide the training of the policy network. Step S502: At each time step, the policy network outputs the corresponding operating policy based on the current input system state and calculates the reward value based on the execution result after each policy execution. The reward value is evaluated based on multiple objective functions of the power system. The reward value reflects the effect of the current policy and is used to optimize the policy network. The expected return of each policy is calculated through the Q-value network, and this feedback is used to optimize the next policy selection. Step S503: After each time step, based on the feedback and reward of policy execution, update the policy network and Q-value network using the policy gradient method and error minimization of the Q-value network; Step S504: After the policy is updated, the power system will adjust its various operations according to the new policy. Finally, the power system enters the next time step, and the system's state changes will be fed back into the reinforcement learning model as spatiotemporal features to start a new round of decision-making.

10. The method for extracting node spatial features for power flow calculation based on a large power model according to claim 9, characterized in that, Step S501 involves using the deep deterministic policy gradient algorithm to establish the policy network and the Q-value network, respectively: Step S5011: Establish a multi-layer neural network as the policy network to output continuous action vectors based on the current state of the power system; Step S5012: Based on the current state of the power system and the action output by the policy network, a multi-layer neural network is constructed as the Q-value network, which outputs the Q-value of the action in the current state. At the same time, the Q-value network takes the combination of state and action as input through forward propagation, passes it through the nonlinear mapping of the multi-layer neural network, and outputs the corresponding Q-value.

Images