Model-driven graph convolutional neural network power flow calculation method

Abstract

The present invention discloses a model-driven graph convolutional neural network power flow calculation method, the steps of which include: 1) establishing a linear power flow equation for a power system; 2) decentralizing the linear power flow equation for the power system to obtain a node characteristic state update equation; 3) establishing a model-driven graph convolution equation according to the node characteristic state update equation; 4) establishing a model-driven graph convolutional neural network according to the model-driven graph convolutional equation; 5) acquiring basic data of the power system and inputting it into the model-driven graph convolutional neural network to obtain the power system power flow. The present invention can adapt to topology and new energy uncertainties.

Description

Technical Field

[0001] The present invention relates to the field of power systems and automation thereof, and in particular to a model-driven graph convolutional neural network power flow calculation method. Background Art

[0002] The recent trend of developing renewable energy and the expected changes in the coming years will bring great challenges to the operation of power systems. In order to consider the uncertainty of renewable energy and intermittent loads, probabilistic analysis methods have been studied by many scholars, such as probabilistic power flow. Probabilistic analysis methods can obtain a reasonable statistical assessment of system risks by solving a large number of samples randomly generated under a given distribution. However, the traditional numerical solution method has a heavy cumulative computational burden and cannot obtain calculation results in a limited time, which limits the practical application of probabilistic analysis.

[0003] With the development of artificial intelligence technology, neural networks have been widely used in power flow calculations due to their advantages of high accuracy and fast speed. Power flow calculation methods based on deep neural networks, convolutional neural networks, extreme learning machines and graph convolutional neural networks have been proposed in large numbers. Neural networks learn the mapping distribution of inputs and outputs in a large number of power flow samples to achieve learning of power flow mapping. However, due to its strong dependence on training data, if there is a difference between the actual scene and the data distribution of the training set, the trained neural network cannot accurately predict the power flow results. At the same time, when the topology of the power system changes, the power flow distribution of the system will change dramatically. In order to ensure the accuracy of the neural network prediction, it will be necessary to collect a large number of samples under the new topology and retrain the neural network. However, in actual engineering, the system topology often changes due to factors such as maintenance or failure. The existing data-driven power flow calculation method cannot adapt to the actual situation of topology changes and uncertain new energy.

[0004] In summary, there is an urgent need to study a data-driven power flow calculation method that can adapt to topological and new energy uncertainties. Summary of the invention

[0005] The object of the present invention is to provide a model-driven graph convolutional neural network power flow calculation method, comprising the following steps:

[0006] 1) Establish the linear power flow equation of the power system.

[0007] The linear power flow equation of the power system is as follows:

[0008]

[0009] Where V i 、V j is the voltage amplitude of node i and node j; G ij , B ijare the i-th row and j-th column elements in the conductance matrix and susceptance matrix respectively; P i , Q i Inject active power and reactive power into the node respectively; G i，shunt , B i，shunt are the conductance and susceptance of node i to ground respectively; N(i) represents the set of nodes connected to node i; θ ij is the voltage phase angle difference between node i and node j.

[0010] The state variables in the linear power flow equation of the power system include voltage amplitude and phase angle.

[0011] 2) The linear power flow equation of the power system is decentralized to obtain the node characteristic state update equation.

[0012] The steps of decentralizing the linear power flow equation of the power system include:

[0013] 2.1) Taking the parameters in the linear power flow equation of the power system as As the node independent variable, establish the state update equation, that is,

[0014]

[0015] Where N(i) represents the set of nodes connected to node i; θ i ,θ j is the voltage phase angle between node i and node j;

[0016] 2.2) Introducing the intermediate parameter α i , intermediate parameter β i , intermediate parameter δ i , intermediate parameter γ i , intermediate parameter λ i , intermediate parameter ζ i , and simplify formula (2) to obtain:

[0017]

[0018] Among them, the intermediate parameter α i , intermediate parameter β i , intermediate parameter δ i , intermediate parameter γ i , intermediate parameter λ i , intermediate parameter ζ i They are as follows:

[0019]

[0020] In the formula, G ii , B ii are the i-th row and j-th column elements in the conductance matrix and susceptance matrix respectively;

[0021] 2.3) Establish the node characteristic state update equation, namely:

[0022]

[0023] 3) Establish a model-driven graph convolution equation based on the node feature state update equation.

[0024] The model-driven graph convolution equation is as follows:

[0025]

[0026] In the formula, are the graph convolution kernels for voltage amplitude and phase angle, respectively. are the voltage amplitude vector and phase angle vector respectively; σ() is the graph convolution function;

[0027] Among them, the parameters parameter They are as follows:

[0028]

[0029] In the formula, are the voltage amplitude vector and phase angle vector respectively. i , Q i Inject active power and reactive power into the nodes respectively.

[0030] 4) Establishing a model-driven graph convolutional neural network based on the model-driven graph convolutional equation.

[0031] The model-driven graph convolutional neural network includes a feature extraction module and a prediction module.

[0032] The input of the feature extraction module includes the node initial state voltage amplitude V0, phase angle θ0, node injected active power P, node injected reactive power Q, conductance G and susceptance B.

[0033] The feature extraction module includes K layers of model-driven graph convolution layers and pooling layers. The input of the first layer of model-driven graph convolution layer includes the node initial state voltage amplitude V0, phase angle θ0, node injected active power P, node injected reactive power Q, conductance G and susceptance B. The output of the Kth layer of model-driven graph convolution layer includes the voltage amplitude V K ∈R N×n and phase angle θ K ∈R N×n . N×n is the dimension.

[0034] Among them, the output of the l-th model-driven graph convolutional layer is as follows:

[0035]

[0036] Where V l ,θ l The voltage amplitude and phase angle output by the convolutional layer of the driving graph of the l-th layer model.

[0037] Among them, the parameter α l-1 , parameter β l-1 , parameter δ, parameter ζ, parameter λ, and parameter γ are as follows:

[0038]

[0039] In the formula, G ndiag , B ndiag are the nodal conductance and susceptance matrices with the diagonal elements removed;

[0040] The input of the pooling layer is the output of the K-th model-driven graph convolutional layer, and the output is as follows:

[0041] H=V K ||pool(V K )||θ K ||pool(θ K ) (10)

[0042] In the formula, pool(·) represents mean pooling in the feature channel. || represents concatenation of features. It represents the power system flow characteristics.

[0043] The prediction module includes a fully connected layer. The input of the prediction module includes the power system flow characteristics H, and the output is as follows:

[0044]

[0045] In the formula, are the weights and biases related to the node voltage amplitude in the fully connected neural network; W θ , B θ are the weights and biases related to the node voltage phase angle in the fully connected neural network; V out ,θ out are the predicted node voltage amplitude and phase angle respectively; PF out , QF out are the predicted branch active and reactive power respectively. A is the node branch connection matrix; W PF , B PF are the weight and bias related to the branch active power respectively; W QF , B QF are the weight and bias related to branch reactive power respectively;

[0046] The model-driven graph convolutional neural network is trained by historical trend samples or simulated samples, and the samples include training samples of different new energy distributions under different topologies.

[0047] 5) Obtain basic data of the power system and input it into the model-driven graph convolutional neural network to obtain the power system flow.

[0048] The basic data of the power system include the node initial state voltage amplitude V0, phase angle θ0, node injected active power P, node injected reactive power Q, conductance G and susceptance B.

[0049] The technical effect of the present invention is unquestionable. The present invention can adapt to the uncertainty of topology and new energy, and has good adaptability to unknown load distribution. At the same time, the power flow calculated by the present invention has higher accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 The graph convolution is embedded in the graph convolutional neural network driven by the model. The left side is the calculation process of neighborhood aggregation based on the power flow equation, and the right side is the forward propagation process of the graph convolutional neural network;

[0051] Figure 2 The overall architecture of the model-driven graph convolutional neural network;

[0052] Figure 3 is the loss value of neural network training;

[0053] Figure 4 The test accuracy of the model after the load distribution changes; Figure 4 (a) is the test accuracy of the model after the load distribution mean is changed; Figure 4 (b) is the test accuracy of the model after the load distribution standard deviation is changed. DETAILED DESCRIPTION

[0054] The present invention is further described below in conjunction with the embodiments, but it should not be understood that the above subject matter of the present invention is limited to the following embodiments. Without departing from the above technical ideas of the present invention, various substitutions and changes are made according to the common technical knowledge and customary means in the art, which should all be included in the protection scope of the present invention.

[0055] Embodiment 1:

[0056] See also Figure 1 to Figure 2 ,The model-driven graph convolutional neural network power flow calculation method includes the following steps:

[0057] 1) Establish the linear power flow equation of the power system.

[0058] The linear power flow equation of the power system is as follows:

[0059]

[0060] Where V i 、V j is the voltage amplitude of node i and node j; G ij , B ij are the i-th row and j-th column elements in the conductance matrix and susceptance matrix respectively; P i , Q i Inject active power and reactive power into the node respectively; G i，shunt , B i，shunt are the conductance and susceptance of node i to ground respectively; N(i) represents the set of nodes connected to node i; θ ij is the voltage phase angle difference between node i and node j.

[0061] The state variables in the linear power flow equation of the power system include voltage amplitude and phase angle.

[0062] 2) The linear power flow equation of the power system is decentralized to obtain the node characteristic state update equation.

[0063] The steps of decentralizing the linear power flow equation of the power system include:

[0064] 2.1) Taking the parameters in the linear power flow equation of the power system as As the node independent variable, establish the state update equation, that is,

[0065]

[0066] Where N(i) represents the set of nodes connected to node i; θ i ,θ j is the voltage phase angle between node i and node j;

[0067] 2.2) Introducing the intermediate parameter α i , intermediate parameter β i , intermediate parameter δ i , intermediate parameter γ i , intermediate parameter λ i , intermediate parameter ζ i , and simplify formula (2) to obtain:

[0068]

[0069] Among them, the intermediate parameter α i , intermediate parameter β i , intermediate parameter δ i , intermediate parameter γ i , intermediate parameter λ i , intermediate parameter ζ i They are as follows:

[0070]

[0071] In the formula, G ii , B ii are the i-th row and j-th column elements in the conductance matrix and susceptance matrix respectively;

[0072] 2.3) Establish the node characteristic state update equation, namely:

[0073]

[0074] 3) Establish a model-driven graph convolution equation based on the node feature state update equation.

[0075] The model-driven graph convolution equation is as follows:

[0076]

[0077] In the formula, are the graph convolution kernels for voltage amplitude and phase angle, respectively. are the voltage amplitude vector and phase angle vector respectively; σ() is the graph convolution function;

[0078] Among them, the parameters parameter They are as follows:

[0079]

[0080] In the formula, are the voltage amplitude vector and phase angle vector respectively. i , Q i Inject active power and reactive power into the nodes respectively.

[0081] 4) Establishing a model-driven graph convolutional neural network based on the model-driven graph convolutional equation.

[0082] The model-driven graph convolutional neural network includes a feature extraction module and a prediction module.

[0083] The input of the feature extraction module includes the node initial state voltage amplitude V0, phase angle θ0, node injected active power P, node injected reactive power Q, conductance G and susceptance B.

[0084] The feature extraction module includes K layers of model-driven graph convolution layers and pooling layers. The input of the first layer of model-driven graph convolution layer includes the node initial state voltage amplitude V0, phase angle θ0, node injected active power P, node injected reactive power Q, conductance G and susceptance B. The output of the Kth layer of model-driven graph convolution layer includes the voltage amplitude V K ∈R N×n and phase angle θ K∈R N×n . N×n is the dimension.

[0085] Among them, the output of the l-th model-driven graph convolutional layer is as follows:

[0086]

[0087] Where V l ,θ l The voltage amplitude and phase angle output by the convolutional layer of the driving graph of the l-th layer model.

[0088] Among them, the parameter α l-1 , parameter β l-1 , parameter δ, parameter ζ, parameter λ, and parameter γ are as follows:

[0089]

[0090] In the formula, G ndiag , B ndiag are the nodal conductance and susceptance matrices with the diagonal elements removed;

[0091] The input of the pooling layer is the output of the K-th model-driven graph convolutional layer, and the output is as follows:

[0092] H=V K ||pool(V K )||θ K ||pool(θ K ) (10)

[0093] In the formula, pool(.) represents mean pooling in the feature channel. || represents concatenation of features. It represents the power system flow characteristics.

[0094] The prediction module includes a fully connected layer. The input of the prediction module includes the power system flow characteristics H, and the output is as follows:

[0095]

[0096] In the formula, are the weights and biases related to the node voltage amplitude in the fully connected neural network; W θ , B θ are the weights and biases related to the node voltage phase angle in the fully connected neural network; V out ,θ out are the predicted node voltage amplitude and phase angle respectively; PF out , QF out are the predicted branch active and reactive power respectively. A is the node branch connection matrix; W PF , B PFare the weight and bias related to the branch active power respectively; W QF , B QF are the weight and bias related to branch reactive power respectively;

[0097] The model-driven graph convolutional neural network is trained by historical trend samples or simulated samples, and the samples include training samples of different new energy distributions under different topologies.

[0098] 5) Obtain basic data of the power system and input it into the model-driven graph convolutional neural network to obtain the power system flow.

[0099] The basic data of the power system include the node initial state voltage amplitude V0, phase angle θ0, node injected active power P, node injected reactive power Q, conductance G and susceptance B.

[0100] Embodiment 2:

[0101] The model-driven graph convolutional neural network power flow calculation method includes the following steps:

[0102] First, according to the domain aggregation method in the graph convolutional neural network, the linear power flow equation of the power system is decentralized to form a state equation of the central node for updating the state of adjacent nodes. Then the coupling equations of the same nodes in all state equations are solved to realize the decoupling of the state equations. After obtaining the decoupled state update method, the equation is applied to the domain aggregation in the graph convolutional neural network to form a model-driven graph convolutional layer. Finally, multiple model graph convolutional layers are connected in sequence and feature global pooling layers and fully connected layers are added to form a model-driven graph convolutional neural network. The model-driven graph convolutional neural network is trained by extracting training samples of different new energy distributions under different topologies through the Monte Carlo simulation method, and finally the power flow calculation model under the system is obtained. Finally, the IEEE14-node system power flow is simulated and calculated to verify the effectiveness and accuracy of the present invention.

[0103] Embodiment 3:

[0104] The model-driven graph convolutional neural network power flow calculation method includes the following steps:

[0105] 1) Graph convolution based on linear power flow model

[0106] The present invention transforms the linear power flow equation according to the process of domain aggregation, derives a model-driven domain aggregation method, and applies the derived domain aggregation method to graph convolution to form a model-driven graph convolution layer. Finally, the system topology is embedded in the neural network. The present invention regards the state variables (voltage amplitude and phase angle) as node features in the graph and derives them using the linear power flow equation. The linear power flow model is expressed as:

[0107]

[0108] Where V i ,θ i is the node voltage amplitude and phase angle; G ij , B ij are the i-th row and j-th column elements in the conductance matrix and susceptance matrix respectively. i , Q i They are the active and reactive power injected by the node respectively.

[0109] In (1), Treat it as a node independent variable, move all items on the right side containing the state variables of node i to the left, and move the remaining items to the right to form a state update equation.

[0110]

[0111] In the formula, N(t) represents the set of nodes connected to node i;

[0112] On the right side of (2), it mainly includes the node input features (node ​​injected active and reactive power) and the features obtained from adjacent nodes, and the left side of the equal sign is the coupling relationship of node features. For simplicity, the right side is set to α i and β i , the coefficients on the left are δ i , γ i , i , i . Formula (2) can be simplified to:

[0113]

[0114] in:

[0115]

[0116] So keep the right side α i and β i , and solve the coupled linear equation on the left. The node feature state update equation is obtained, that is, the model-driven domain aggregation method, such as:

[0117]

[0118] The derived domain aggregation method is introduced into the graph convolution expression, and finally the model-driven graph convolution method is obtained, which is expressed as follows:

[0119]

[0120] in:

[0121]

[0122] In (6), are graph convolution kernels for voltage amplitude and phase angle, respectively. The proposed graph convolution method not only embeds the power system topology into the neural network, but also retains the coupling characteristics of the power flow when aggregating node features. By embedding the graph convolution neural network with the proposed model-driven graph convolution, the embedding of the power system topology and physical relationship is achieved, such as Figure 1 shown.

[0123] 2) Model-driven graph convolutional neural network

[0124] According to the proposed model-driven graph convolution method, the present invention constructs a model-driven graph convolutional neural network. The neural network consists of two parts, namely, a feature extraction module and a prediction module. Figure 2 shown.

[0125] Feature extraction module

[0126] The feature extraction module is mainly composed of multiple layers of model-driven graph convolution layers and pooling layers. First, the node voltage amplitude is set to 1 and the phase angle is set to 0 in the initial state of the node. Then the input features of the node include V0, θ0, P, Q, G, B. Then the model-driven graph convolution layer of the lth layer can be expressed as a matrix:

[0127]

[0128] in:

[0129]

[0130] The first part of the feature extraction module in the model-driven graph convolutional neural network is formed by connecting multiple layers of model-driven graph convolution. When a model-driven graph convolutional neural network with K layers of graph convolution is built in an N-node system, and the last convolutional layer contains n feature channels, the feature obtained by convolution is V K , Then the obtained features are mean pooled in each feature channel, and the pooled feature vector is added to the feature vector of each node to finally form the complete feature vector of the node, which is mathematically expressed as:

[0131] H=V K ||pool(V K )||θ K ||pool(θ K ) (10)

[0132] Where pool(.) represents mean pooling in the feature channel; || represents concatenation of features; Represents the complete feature obtained by concatenating the global feature vector obtained after feature pooling with the input local feature vector.

[0133] Prediction Module

[0134] The prediction module is mainly built by the fully connected layer. Since the flow calculation includes the node voltage amplitude and phase angle, branch active and reactive power. However, the feature extraction module does not extract the branch feature vector, and the branch active and reactive power prediction cannot be performed. Therefore, the present invention converts the node feature vector into the branch feature vector, which is mainly converted through the node branch connection matrix A. Then the output of the prediction module can be expressed as:

[0135]

[0136] in are the trainable weights and biases in a fully connected neural network; V out ,θ out are the predicted node voltage amplitude and phase angle respectively; PF out , QF out They are the predicted branch active and reactive power respectively.

[0137] Finally, the feature extraction module and the prediction module are combined to form a model-driven graph convolutional neural network. By using historical flow samples or simulated samples, the proposed model-driven neural network can accurately calculate the flow when the topology or new energy distribution changes.

[0138] Embodiment 4:

[0139] Experiments on the model-driven graph convolutional neural network power flow calculation method, including:

[0140] The basic data of the system in this embodiment refers to the IEEE 14-node system. Wind farms and photovoltaic power plants are added to the system to take into account the uncertainty of renewable energy. Assume that the wind speed obeys the Weibull distribution, with a scale parameter of 2.016 and a shape parameter of 5.089. The maximum output of a single wind turbine is 70MW. The solar irradiance obeys the Beta distribution with α=2.06 and β=2.5. The maximum output of a single photovoltaic unit is 45MW. Add a wind turbine to each of the 4 and 11 nodes, and add a photovoltaic unit to each of the 3 and 8 nodes. Assume that the load obeys a normal distribution with a standard deviation of 0.1 and a mean of the system default load. In order to consider the situation under different topologies, this embodiment collects samples of branch N-1 and N-2 faults for training and testing. In the simulation, 1 or 2 transmission lines in the power system are randomly selected to consider accidental events. Through random sampling, 10,000 training samples and 1,000 test samples are collected.

[0141] In this embodiment, the proposed method is compared with the existing deep neural network (DNN), graph neural network (GNN) and graph convolutional neural network (GCN). The above three neural networks and the proposed method are built under the Tensorflow deep learning framework, and the adam optimizer is used to train the neural network. Finally, the curve of the loss of the neural network training with the number of training times is obtained, as shown in Figure 2. Figure 3 . Then the trained neural network model is used to predict the samples of the test set, and the predicted probability accuracy is shown in Table 1. The probability accuracy is the probability that the neural network prediction error is less than the threshold. In this embodiment, the thresholds for node voltage amplitude and phase angle, branch active power and reactive power are 0.001pu, 0.01rad, 5MW and 5MWar respectively.

[0142] Table 1

[0143]

[0144] exist Figure 3 It can be observed that the graph convolutional neural network and the model-driven graph convolutional neural network can converge after 1000 trainings. However, the model-driven graph convolutional neural network proposed in the present invention has stronger convergence and can converge to a smaller loss in the training set. It can be observed from Table 1 that only the prediction accuracy of the proposed method can reach more than 95%. Therefore, in summary, the proposed method has higher accuracy in calculating the trend.

[0145] Embodiment 5:

[0146] Experiments on the model-driven graph convolutional neural network power flow calculation method, including:

[0147] This example uses the IEEE 30 node system, randomly selects nodes to add wind turbines and photovoltaic units (the settings of each unit are the same as in Example 1), and makes the new energy penetration rate reach 35%. Assume that the load follows a normal distribution, its standard deviation is a random number between [0, 0.1], and its mean is [Pd default -0.5, Pd default +0.5] interval. For topology changes, when extracting samples, a branch is randomly selected to fail, and finally 10,000 samples are randomly extracted for training. For sample extraction of the test set, when extracting a single sample, a branch is also randomly selected to fail, and the standard deviation and average value of the load distribution are modified, and finally a single sample is obtained by power flow calculation. For different standard deviations and average values, 1,000 samples are collected for testing, and finally the test results of all cases are obtained, such as Figure 4 shown.

[0148] from Figure 4It can be observed that when the load fluctuations included in the training set, the prediction errors of different neural networks are relatively low, but when the load fluctuation distribution of the test data is not in the training set, the prediction errors of the deep neural network and the graph convolutional neural network rise sharply, but the error of the model-driven graph convolutional neural network power flow calculation method proposed in the present invention in test five rises very slowly. This proves that the model-driven graph convolutional neural network power flow calculation method proposed in the present invention has good adaptability to unknown load distributions.

[0149] Embodiment 6:

[0150] Experiments on the model-driven graph convolutional neural network power flow calculation method, including:

[0151] Based on Example 5, this embodiment fixes the load distribution so that its standard deviation is 0.1 and the average value is the system default load. Then 1000 samples are extracted as the test set, and in the process of extracting a single sample, two branches are randomly selected to have faults to simulate a topology different from the training set. Finally, the model trained in Example 2 is used to test the extracted samples, and the test results are shown in Table 2.

[0152] Table 2

[0153] method <![CDATA[P v ]]> <![CDATA[P θ ]]> <![CDATA[P PL ]]> <![CDATA[P QL ]]> DNN 43.60％ 50.28％ 78.60％ 88.18％ GCN 48.98％ 51.21％ 91.65％ 97.32％ Model-driven GCN 92.95％ 97.65％ 99.24％ 99.77％

[0154] From the data in Table 2, it can be seen that for samples that do not contain topology in the training set, the deep neural network and the graph convolutional neural network are completely unable to adapt, and the prediction accuracy is less than 50%. However, the method of the present invention can still maintain a prediction accuracy of more than 90%. It can be seen that the method proposed in the present invention has good topology adaptability.

Claims

1. Model-driven graph convolutional neural network power flow calculation method, characterized in that: The following steps are involved: 1) Establish the linear power flow equation of the power system; 2) Decentralize the linear power flow equation of the power system to obtain the node characteristic state update equation; 3) Establish a model-driven graph convolution equation based on the node feature state update equation; 4) establishing a model-driven graph convolutional neural network according to the model-driven graph convolutional equation; 5) Obtain basic data of the power system and input it into the model-driven graph convolutional neural network to obtain the power system flow; The linear power flow equation of the power system is as follows: Where V i 、V j is the voltage amplitude of node i and node j; G ij , B ij are the i-th row and j-th column elements in the conductance matrix and susceptance matrix respectively; P i , Q i Inject active power and reactive power into the node respectively; G i，shunt , B i，shunt are the conductance and susceptance of node i to ground respectively; N(i) represents the set of nodes connected to node i; θ ij is the voltage phase angle difference between node i and node j; The steps of decentralizing the linear power flow equation of the power system include: 2.1) Taking the parameters in the linear power flow equation of the power system as As the node independent variable, establish the state update equation, that is, Where N(i) represents the set of nodes connected to node i; θ i ,θ j is the voltage phase angle between node i and node j; 2.2) Introducing the intermediate parameter α i , intermediate parameter β i , intermediate parameter δ i , intermediate parameter γ i , intermediate parameter λ i , intermediate parameter ζ i , and simplify formula (2) to obtain: Among them, the intermediate parameter α i , intermediate parameter β i , intermediate parameter δ i , intermediate parameter γ i , intermediate parameter λ i , intermediate parameter ζ i They are as follows: In the formula, G ii , B ii are the i-th row and i-th column elements in the conductance matrix and susceptance matrix respectively; 2.3) Establish the node characteristic state update equation, namely: The model-driven graph convolution equation is as follows: In the formula, are graph convolution kernels for voltage amplitude and phase angle, respectively; are the voltage amplitude vector and phase angle vector respectively; σ() is the graph convolution function; Among them, the parameters parameter They are as follows: In the formula, are the voltage amplitude vector and phase angle vector respectively; P i , Q i Inject active power and reactive power into the nodes respectively.

2. The model-driven graph convolutional neural network power flow calculation method according to claim 1, characterized in that: The state variables in the linear power flow equation of the power system include voltage amplitude and phase angle.

3. The model-driven graph convolutional neural network power flow calculation method according to claim 1, characterized in that: The model-driven graph convolutional neural network includes a feature extraction module and a prediction module; The input of the feature extraction module includes the node initial state voltage amplitude V0, phase angle θ0, node injected active power P, node injected reactive power Q, conductance G and susceptance B; The feature extraction module includes K layers of model-driven graph convolution layers and pooling layers; the input of the first layer of model-driven graph convolution layers includes the node initial state voltage amplitude V0, phase angle θ0, node injected active power P, node injected reactive power Q, conductance G and susceptance B; the output of the Kth layer of model-driven graph convolution layers includes the voltage amplitude V K ∈R N×n and phase angle θ K ∈R N×n ; N×n is the dimension; Among them, the output of the l-th model-driven graph convolutional layer is as follows: Where V l ,θ l The voltage amplitude and phase angle output by the convolutional layer of the driving graph of the l-th layer model; Among them, the parameter α l-1 , parameter β l-1 , parameter δ, parameter ζ, parameter λ, and parameter γ are as follows: In the formula, G ndiag , B ndiag are the nodal conductance and susceptance matrices with the diagonal elements removed; The input of the pooling layer is the output of the K-th model-driven graph convolutional layer, and the output is as follows: H=V K |||pool(V K )||θ K ||pool(θ K ) (10) In the formula, pool(·) represents mean pooling in the feature channel; || represents concatenation of features; Indicates the power system flow characteristics; The prediction module includes a fully connected layer; the input of the prediction module includes the power system flow characteristics H, and the output is as follows: In the formula, are the weights and biases related to the node voltage amplitude in the fully connected neural network; W θ , B θ are the weights and biases related to the node voltage phase angle in the fully connected neural network; V out ,θ out are the predicted node voltage amplitude and phase angle respectively; PF out , QF out are the predicted branch active and reactive power respectively; A is the node branch connection matrix; W PF , B PF are the weight and bias related to the branch active power respectively; W QF , B QF are the weight and bias related to branch reactive power respectively.

4. The model-driven graph convolutional neural network power flow calculation method according to claim 1, characterized in that: The model-driven graph convolutional neural network is trained by historical trend samples or simulated samples; the samples include training samples of different new energy distributions under different topologies.

5. The model-driven graph convolutional neural network power flow calculation method according to claim 1, characterized in that: The basic data of the power system include the node initial state voltage amplitude V0, phase angle θ0, node injected active power P, node injected reactive power Q, conductance G and susceptance B.

Images