An online prediction method for static voltage stability margin based on convolutional neural network

Abstract

The present invention proposes an online prediction method for static voltage stability margin based on convolutional neural network, the method comprising: step 1, performing offline training of convolutional neural network to obtain a static voltage stability margin prediction model; step 2, based on the static voltage stability margin prediction model obtained by training in step 1, real-time prediction of photovoltaic power stability margin. The present invention can predict the limit growth margin of photovoltaic output within a certain period of time in a wide-area transmission network, solves the problem of static voltage stability assessment after large-scale access to photovoltaic power in the transmission network, and provides a scientific basis for scheduling decisions.

Description

Technical Field

[0001] The invention relates to an online prediction method for static voltage stability margin based on a convolutional neural network, and belongs to the technical field of steady-state analysis of power systems. Background Art

[0002] The large-scale access of photovoltaic power to the power system has shown an irreversible trend; the output of photovoltaic power is random and uncertain, and its high proportion of access has aggravated the time-varying and volatile characteristics of the power system, affecting the safe and stable operation of the power system. One of the most significant problems is the voltage stability problem; at the same time, the photovoltaic access method has developed from low-voltage distribution network access to direct access to ultra-high voltage transmission network, which has brought challenges to voltage stability in a wider range; the small-disturbance safety and stability assessment of voltage stability is required to be completed in a shorter time scale, so as to achieve the purpose of real-time prediction of system voltage stability by operation dispatch.

[0003] Static voltage stability mainly studies whether there is a feasible solution to the power flow equation. The most commonly used indicator is the voltage stability margin (VSM), which can be calculated by the continuous power flow method (CPF). However, the topology of large-scale transmission systems is complex and the dimension of the mathematical model is very large. Directly using the continuous power flow method for prediction is inefficient and cannot adapt to changes in network topology caused by maintenance, line transformation, and expansion.

[0004] In recent years, deep learning methods have been used in power systems to a certain extent. They do not require the establishment of complex mathematical models. For the problem of static voltage stability assessment, deep learning can be used to combine the advantages of measurement methods and numerical calculation methods, and the characteristic values ​​of the system can be obtained based on the deterministic linear model as historical training data to ensure the accuracy of the model. The real-time measurement data is then input into the trained model to obtain the results and realize real-time evaluation. Summary of the invention

[0005] The present invention proposes an online prediction method for static voltage stability margin based on convolutional neural network, which aims to solve the problem that static voltage stability cannot be evaluated after large-scale access to photovoltaics in the transmission network.

[0006] The technical solution of the present invention is as follows: 1. A method for online prediction of static voltage stability margin based on convolutional neural network, the method comprising:

[0007] Step 1: Perform offline training of the convolutional neural network to obtain a static voltage stability margin prediction model;

[0008] Step 2: Based on the static voltage stability margin prediction model trained in step 1, the photovoltaic power stability margin is predicted in real time.

[0009] Furthermore, the step 1 performs offline training of a convolutional neural network to obtain a static voltage stability margin prediction model, which specifically includes:

[0010] Step 1.1: Establish the transmission network topology model of the area to be predicted;

[0011] Step 1.2: Define the input data structure for offline training of convolutional neural network;

[0012] Step 1.3: Based on the grid topology information in the transmission network topology model established in step 1.1, randomly generate the basic load data X of each node B , and based on this, generate the target load data X T ;

[0013] The basic load data is randomly generated according to the different proportions of photovoltaic output in the whole system under the initial state, and the target load data is randomly generated according to the different loads, power generation, and photovoltaic growth of each node; the set I constituted by X is used as the input of the training set for offline training of the convolutional neural network, as shown in formula (7):

[0014] I=[X 1 X 2 ...X i ...X h ] (7);

[0015] In formula (7), h is the data volume of the training set used for offline training of the convolutional neural network;

[0016] Step 1.4: Using Matlab and Matpower, perform continuous power flow calculations on the h data in step 1.3 (7) to obtain a parameterized load-type continuous power flow model;

[0017] Step 1.5: Determine the labels of the training set for offline training of the convolutional neural network; define the static voltage stability margin P after photovoltaic access VSM , its expression is shown in formula (10):

[0018]

[0019] where λ max is the load parameter at the voltage critical point, S B is the reference power, ΔP i,PV is the photovoltaic output growth at the i-th node in a certain period of time, that is:

[0020]

[0021] The h P calculated in steps 1.4 and 1.5 areVSM Construct the set Y as the labels of the training set I for offline training of the convolutional neural network:

[0022] Y=[P 1,VSM P 2,VSM ...P h,VSM ] (12);

[0023] Step 1.6: Build a convolutional neural network;

[0024] Step 1.7: Perform normalization preprocessing on the input data of formula (7) and the label data of formula (12), call the Matconvnet tool in Matlab, and use the input data and label data to train the convolutional neural network of step 1.6;

[0025] Step 1.8: Delete the output layer of the convolutional neural network trained in step 1.7, and use the last layer of the fully connected layer as the new output layer to obtain the static voltage stability margin prediction model.

[0026] Furthermore, the power grid topology information in the power transmission network topology model in step 1.1 includes thermal power plant access nodes and photovoltaic access nodes.

[0027] Furthermore, the input data structure in step 1.2 is a three-dimensional matrix composed of two 3×n matrices, that is, matrix X=[X B X T ] 3×n×2 .

[0028] Furthermore, the matrix X=[X B X T ] 3×n×2 Medium X B It is composed of basic load data, and its specific composition is shown in formula (1):

[0029]

[0030] The superscript B of the matrix and the superscript b of the element represent the basic state, n is the number of nodes in the system, and X B The first row of elements is the active power P of the net load of system node i i b , the second row of the matrix is the reactive power of the net load at system node i The third row of the matrix is the active power generated by system node i The expression of each row element is shown in formula (2):

[0031]

[0032] Where i is the system node number, and i = 1, 2, 3..., n; the active power P of the net load in the basic state i b and reactive power It is defined as formula (3):

[0033]

[0034] in, is the active power of the load at node i in the basic state, is the active power of the photovoltaic power of node i in the basic state, is the reactive power of the load at node i in the basic state, is the PV reactive power of node i in the basic state;

[0035] The matrix X=[X B X T ] 3×n×2 Medium X T It is composed of target load data, and its specific composition is shown in equations (4), (5), and (6):

[0036]

[0037]

[0038]

[0039] The superscript T of the matrix and the superscript t of the element represent the target state. is the active power of the load of node i under the target state, is the active power of the photovoltaic power of node i under the target state, is the reactive power of the load at node i in the target state, is the PV reactive power of node i under the target state.

[0040] Furthermore, the load-type continuous power flow model parameterized in step 1.4 is expressed by equation (8):

[0041]

[0042] Where P i b is the active power of the load at node i in the initial state, is the reactive power of the load at node i in the initial state, is the active power generated by node i in the initial state; ΔP i , ΔQ i , ΔPGi is the predetermined load increment and power generation increment, which is obtained by subtracting equation (2) from equation (5), namely:

[0043]

[0044] The λ in formula (8) is a scalar parameter reflecting the load level. After continuous power flow calculation, the λ sequence is obtained: 0<λ≤λ max ; When λ=λ max When , it means that the power of each node increases in a predetermined direction and the power reaches the limit.

[0045] Furthermore, the feature extraction part of the convolutional neural network in step 1.6 includes 5 convolutional layers, five activation layers, and three batch normalization layers; the prediction part of the convolutional neural network includes 4 fully connected layers; the output layer of the convolutional neural network uses the sum of the squares of the difference between the predicted value and the true value as the objective function:

[0046]

[0047] In the above formula, N is the number of batch data; P i It is the output data after the i-th input data passes through the fourth fully connected layer of the network, that is, the predicted static voltage stability margin.

[0048] Furthermore, the purpose of training the convolutional neural network in step 1.7 is to minimize the value of the objective function C by adjusting the network parameters. After several steps of training, C decreases to a stable value and no longer decreases, and the training is stopped.

[0049] Furthermore, the step 2 performs real-time prediction of the photovoltaic power stability margin based on the static voltage stability margin prediction model trained in step 1, specifically including:

[0050] Step 2.1: Get real-time input data X;

[0051] Step 2.2: The real-time input data X is normalized and used as the input of the static voltage stability margin prediction model obtained in step 1.8. The output is the real-time predicted static voltage stability margin.

[0052] Furthermore, the step 2.1 obtains real-time input data X, specifically including: the active power of the load of node i in the basic state Active power of photovoltaic power at node i in basic state Reactive power of the load at node i in the basic state The reactive power of the photovoltaic power plant at node i in the basic state The active power of the load at node i under the target state is collected online in real time by the SCADA system. Reactive power of the load at node i in the target state According to the dispatch load forecast, the active power of the photovoltaic output of node i in the target state is The reactive power of the photovoltaic output of node i under the target state According to the local temperature and illumination data prediction; the real-time input data X = [X B X T ] 3×n×2 .

[0053] Beneficial effects of the present invention:

[0054] 1) The present invention can predict the maximum growth margin of photovoltaic output in a certain period of time in a wide-area transmission network, solve the problem of static voltage stability assessment after large-scale access to photovoltaic power in the transmission network, and provide a scientific basis for scheduling decisions;

[0055] 2) Convolutional neural network is a deep learning neural network that can extract features from a large amount of complex data and has good generalization ability; the present invention uses the constructed load data as the input of the network, uses the calculated static voltage stability margin as the input data label, uses the training samples for offline training to generate a static voltage stability margin prediction model, and then uses the load prediction data obtained online as input, thereby obtaining an online predicted static voltage stability margin;

[0056] 3) The present invention solves the problem of static voltage stability assessment under high-proportion photovoltaic access in the transmission network, and can directly predict the maximum photovoltaic power generation margin under static voltage stability conditions, with the characteristics of small error, fast speed and strong robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Attached Figure 1 It is a schematic flow chart of the method of the present invention.

[0058] Attached Figure 2 It is a schematic diagram of the IEEE-30 node system model.

[0059] Attached Figure 3 It is a schematic diagram of the convolutional neural network structure of the present invention.

[0060] Attached Figure 4 It is a diagram of the data structure of the convolutional neural network training process.

[0061] Attached Figure 5 It is a schematic diagram of the training results of the convolutional neural network.

[0062] Attached Figure 6 It is a schematic diagram of the regression relationship between predicted data and real data. DETAILED DESCRIPTION

[0063] A method for online prediction of static voltage stability margin based on convolutional neural network, the method comprising:

[0064] Step 1: Performing offline training of convolution neural networks (CNN) to obtain a static voltage stability margin prediction model;

[0065] Step 2: Based on the static voltage stability margin prediction model trained in step 1, the photovoltaic power stability margin is predicted in real time.

[0066] The step 1 performs offline training of a convolutional neural network to obtain a static voltage stability margin prediction model, which specifically includes:

[0067] Step 1.1: Establish a transmission network topology model for the area to be predicted; the power grid topology information in the transmission network topology model includes information such as thermal power plant access nodes, photovoltaic access nodes, and other power source access nodes;

[0068] Step 1.2: Define the input data structure for offline training of the convolutional neural network; the input data structure is a three-dimensional matrix composed of two 3×n matrices, that is, matrix X = [X B X T ] 3×n×2 ;

[0069] Among them, X B It is composed of basic load data, and its specific composition is shown in formula (1):

[0070]

[0071] The superscript B of the matrix and the superscript b of the element represent the base state, n is the number of nodes in the system, and X B The first row of elements is the active power P of the net load of system node i i b , the second row of the matrix is the reactive power of the net load at system node i The third row of the matrix is the active power generated by system node i The expression of each row element is shown in formula (2):

[0072]

[0073] Where i is the system node number, and i = 1, 2, 3..., n; the active power P of the net load in the basic state i b and reactive power It is defined as formula (3):

[0074]

[0075] in, is the active power of the load at node i in the basic state, is the active power of the photovoltaic power of node i in the basic state, is the reactive power of the load at node i in the basic state, is the PV reactive power of node i in the basic state;

[0076] X T It is composed of target load data, and its specific composition is shown in equations (4), (5), and (6):

[0077]

[0078]

[0079]

[0080] The superscript T of the matrix and the superscript t of the element represent the target state. is the active power of the load of node i under the target state, is the active power of the photovoltaic power of node i under the target state, is the reactive power of the load at node i in the target state, is the PV reactive power of node i in the target state;

[0081] Step 1.3: Based on the grid topology information in the transmission network topology model established in step 1.1, randomly generate the basic load data X of each node B , and based on this, generate the target load data X T ;

[0082] The basic load data is randomly generated according to the different proportions of photovoltaic output in the whole system under the initial state, and the target load data is randomly generated according to the different loads, power generation, and photovoltaic growth of each node; the set I constituted by X is used as the input of the training set for offline training of the convolutional neural network, as shown in formula (7):

[0083] I=[X 1 X 2 ...X i ...X h ] (7);

[0084] In formula (7), h is the data volume of the training set used for offline training of the convolutional neural network;

[0085] Step 1.4: Using Matlab and Matpower, perform continuous power flow calculations on the h data in step 1.3 (7) to obtain a parameterized load-type continuous power flow model; the parameterized load-type continuous power flow model is expressed by formula (8):

[0086]

[0087] Where P i b is the active power of the load at node i in the initial state, is the reactive power of the load at node i in the initial state, is the active power of node i in the initial state, i.e., the basic load data in equation (2); ΔP i , ΔQ i , ΔPG i is the predetermined load increment and power generation increment, which is obtained by subtracting equation (2) from equation (5), namely:

[0088]

[0089] The λ in formula (8) is a scalar parameter reflecting the load level. After continuous power flow calculation, the λ sequence is obtained: 0<λ≤λ max ; When λ=λ max When , it means that the power of each node increases in a predetermined direction and reaches the limit;

[0090] Step 1.5: Determine the labels of the training set for offline training of the convolutional neural network; define the static voltage stability margin P after photovoltaic access VSM , its expression is shown in formula (10):

[0091]

[0092] where λ max is the load parameter at the voltage critical point, S B is the reference power, ΔP i,PV is the photovoltaic output growth at the i-th node in a certain period of time, that is:

[0093]

[0094] Static voltage stability margin P VSM The growth of the photovoltaic output of the whole system represents the critical point of static voltage collapse, which represents the growth limit of the photovoltaic output of the whole system. VSM Construct the set Y as the labels of the training set I for offline training of the convolutional neural network:

[0095] Y=[P 1,VSM P2,VSM ...P h,VSM ] (12);

[0096] Step 1.6: Establish a convolutional neural network. The feature extraction part of the convolutional neural network includes 5 convolutional layers (Conv), 5 activation layers (using ReLU function), and 3 batch normalization layers (BN, Batch Normalization); the prediction part of the convolutional neural network contains 4 fully connected layers (FC, Full Connection); the output layer of the convolutional neural network uses the sum of the squares of the difference between the predicted value and the true value as the objective function (or loss function, Norm Loss):

[0097]

[0098] In the above formula, N is the number of batch data; P i is the output data of the ith input data after passing through the fourth fully connected layer of the network, that is, the predicted static voltage stability margin;

[0099] Step 1.7: Perform normalization preprocessing on the input data of formula (7) and the label data of formula (12), call the Matconvnet tool in Matlab, and use the input data and label data to train the convolutional neural network of step 1.6; the purpose of training is to minimize the value of the objective function C by adjusting the network parameters. After several steps of training, C decreases to a stable value and no longer decreases, and the training is stopped;

[0100] Step 1.8: Delete the output layer of the convolutional neural network trained in step 1.7, and use the last layer of the fully connected layer as the new output layer to obtain the static voltage stability margin prediction model.

[0101] Step 2: Based on the static voltage stability margin prediction model trained in step 1, real-time prediction of the photovoltaic power stability margin is performed, specifically including:

[0102] Step 2.1: Obtain real-time input data, the active power of the load at node i in the basic state Active power of photovoltaic power at node i in basic state Reactive power of the load at node i in the basic state The reactive power of the photovoltaic power plant at node i in the basic state The active power of the load at node i under the target state is collected online in real time by the SCADA system. Reactive power of the load at node i in the target state According to the dispatch load forecast, the active power of the photovoltaic output of node i in the target state is The reactive power of the photovoltaic output of node i under the target state According to the local temperature and illumination data prediction; the real-time input data X = [X B X T ] 3×n×2 ;

[0103] Step 2.2: The real-time input data X is normalized and used as the input of the static voltage stability margin prediction model obtained in step 1.8. The output is the real-time predicted static voltage stability margin.

[0104] The convolutional neural network designed in the present invention cancels the pooling layer and customizes the error function as the output layer, which can perform regression prediction and directly use the load and photovoltaic power generation data of each node of the ultra-high voltage transmission network as the training data of the convolutional neural network, avoiding complex feature extraction calculations and realizing online prediction of the static voltage stability margin, which has the characteristics of high precision, fast speed and strong robustness.

[0105] The technical solution of the present invention is mainly divided into two parts: the first step is offline training, based on the regional transmission network, taking randomly generated load and photovoltaic output data as input, and taking the static voltage stability margin obtained by continuous power flow calculation as input data label to form a training data set; then the designed convolutional neural network is trained to obtain a convolutional neural network online prediction model; the second step is online prediction, taking the real-time online monitoring data and prediction data of the power system as input of the prediction model to obtain the real-time static voltage stability margin; the present invention improves the traditional convolutional neural network, cancels the pooling layer, and designs a new error function as the objective function of the output layer, which has good generalization performance.

[0106] Example

[0107] The technical solution involved in the present invention is further explained below in conjunction with the accompanying drawings and embodiments.

[0108] Step 1: Build an IEEE-30 node system model in Matlab. This standard system has 6 generators, 30 buses, 41 AC lines, a reference power of 100MVA, and a reference voltage of 230kV. Assume that in addition to the 6 generator access nodes, the remaining 24 nodes are all connected to photovoltaics, such as Figure 2 shown.

[0109] Step 2: Based on the original data of the IEEE-30 node system in step 1, generate training data in batches. The principles of generating data are as follows:

[0110] Basic load data: The basic system photovoltaic penetration rate is set to 10 levels from 0% to 180%, with a step length of 20%; the system photovoltaic penetration rate is defined as: the proportion of the active power of photovoltaic power generation in the whole system to the total system load;

[0111] Target load data: The growth of the whole system load is based on a constant power factor growth, and the growth targets are 80%-260% of the basic load, with a step length of 20%, and 10 settings; the photovoltaic output under the target state is set according to the system photovoltaic penetration rate of 20%-200% in 10 types, with a step length of 20%; 10 groups of photovoltaic output data of 24 photovoltaic access nodes are randomly generated under each photovoltaic penetration rate; 10,000 groups of data can be obtained according to the principle of step 2 above.

[0112] Step 3: Call the Matpower continuous power flow calculation program in Matlab to perform continuous power flow calculation on the 10,000 sets of data in step 2, and calculate 10,000 static voltage stability margin data according to formula (10) as the labels of training data; randomly select 7,000 sets of the above data as training data, and the remaining 3,000 sets as test data.

[0113] Step 4: Build a convolutional neural network, such as Figure 3 As shown, it includes 5 convolution (Conv) layers, 5 activation layers (using ReLU function), 3 batch normalization layers (BN, Batch Normalization), and 4 fully connected layers (FC, FullConnection); the output layer adopts the objective function (Norm Loss) shown in formula (13), Figure 4 It shows the change of data structure when the input data passes through the forward propagation of convolutional neural network.

[0114] Step 5: Call the Matconvnet program in Matlab and use 7000 sets of training data to train the convolutional neural network in step 4; before training, normalize the input data matrix row by row:

[0115]

[0116] x max is the maximum value of each row of the matrix, x min is the minimum value of each row of the matrix, and x′ is the normalized value; the training results are as follows Figure 5 As shown in Figure 1, after 100 steps of training, the objective function converges and the objective function value of the test data (val) reaches 0.015. In fact, the output of the convolutional neural network in step 4 is only the objective function. It can be seen from formula (13) that P VSM The predicted value is generated in the penultimate layer of the network, which is the output of the last fully connected layer. Therefore, the last layer of the network is deleted to obtain the static voltage stability margin prediction model required for prediction.

[0117] Step 6: Use the trained convolutional neural network to predict 3000 sets of test data, and use the average relative error indicator to evaluate the prediction results:

[0118]

[0119] Where y i For the real P VSM , The predicted P VSM , N is the number of samples, the prediction results are shown in the attached Figure 6 As shown; attached Figure 6 The horizontal axis is the real P VSM , the vertical axis is the predicted P VSM , if the predicted value is y and the true value is x, then when all the predicted values ​​are exactly the same as the true value, the regression relationship between the two is y=x; predict the test sample and obtain the regression relationship y=kx+b between the predicted value and the true value. When the regression coefficient k is closer to 1 and the intercept b is closer to 0, the predicted value is closer to the true value and the prediction is more accurate. Figure 4 It can be obtained that for the IEEE-30 node system, the regression coefficient k of the predicted value to the true value is 0.9732, and the intercept b is 0.126, indicating that the static voltage stability margin prediction model proposed in the present invention has good generalization performance; the average relative error ε calculated using formula (15) is 3.53%.

[0120] Step 7: Line maintenance N-1 test, considering that only one line in the power system is under maintenance in the same period of time, if the tie line is not a double-circuit line, the original network topology will be changed; for the 8 lines in the IEEE-30 model shown in Table 1, disconnect one of the lines to simulate the maintenance state; use 3000 test groups to calculate P for the IEEE-30 model after N-1 using the continuous power flow method VSM , and use the static voltage stability margin prediction model obtained in step 5 to predict P VSM ; Regression analysis and error calculation of the calculated values ​​and predicted values ​​can obtain the data shown in Table 1:

[0121] Disconnected Line Prediction error ε Regression coefficient k Intercept b 4-12 9.03％ 1.072 0.008 15-18 4.45％ 0.9686 -0.002 10-20 3.26％ 1.002 0.001 22-24 4.44％ 1.047 -0.01 8-28 6.05％ 0.9157 0.012 5-7 4.75％ 0.9134 0.25 16-17 5.8％ 0.9203 0.12 24-25 3.26％ 0.9859 0.009

[0122] Table 1

[0123] As shown in Table 1, when the system is in the N-1 state, the regression coefficient k is close to 1. The static voltage stability margin prediction model is used to predict P VSMIt still has good generalization; since the 4-12 line is a regional interconnection line, when the 4-12 line is disconnected, the prediction error is significantly improved, reaching 9.03%. For the non-regional interconnection line N-1, when the 8-28 line is disconnected, the prediction error is the largest, which is 6.05%; when the 24-25 and 10-20 lines are disconnected, the error is the smallest, which is 3.26%. It can be seen that in the N-1 state, the prediction method of the present invention still has a high accuracy, is less affected by line maintenance, and has strong robustness.

[0124] This embodiment is an illustration of the present invention. In specific implementations, technicians can replace the IEEE-30 node system with an actual system model, and the convolutional neural network structure of the present invention can also be adjusted according to actual conditions, but it will not deviate from the spirit of the present invention or exceed the scope defined by the attached claims.

Claims

1. An online prediction method for static voltage stability margin based on convolutional neural network, Its characteristics are include: Step 1: Perform offline training of the convolutional neural network to obtain a static voltage stability margin prediction model; Step 2: Based on the static voltage stability margin prediction model trained in step 1, a real-time prediction of the photovoltaic power stability margin is performed; The step 1 performs offline training of a convolutional neural network to obtain a static voltage stability margin prediction model, which specifically includes: Step 1.1: Establish the transmission network topology model of the area to be predicted; Step 1.2: Define the input data structure for offline training of convolutional neural network; Step 1.3: Based on the grid topology information in the transmission network topology model established in step 1.1, randomly generate the basic load data X of each node B , and based on this, generate the target load data X T ; The basic load data is randomly generated according to the different proportions of photovoltaic output in the whole system under the initial state, and the target load data is randomly generated according to the different loads, power generation, and photovoltaic growth of each node; the set I constituted by X is used as the input of the training set for offline training of the convolutional neural network, as shown in formula (7): I=[X 1 X 2 ...X i ...X h ] (7); In formula (7), h is the data volume of the training set used for offline training of the convolutional neural network; Step 1.4: Using Matlab and Matpower, perform continuous power flow calculations on the h data in step 1.3 (7) to obtain a parameterized load-type continuous power flow model; Step 1.5: Determine the labels of the training set for offline training of the convolutional neural network; define the static voltage stability margin P after photovoltaic access VSM , its expression is shown in formula (10): where λ max is the load parameter at the voltage critical point, S B is the reference power, ΔP i,PV is the photovoltaic output growth at the i-th node in a certain period of time, that is: The h P calculated in steps 1.4 and 1.5 are VSM Construct the set Y as the labels of the training set I for offline training of the convolutional neural network: Y=[P 1,VSM P 2,VSM ...P h,VSM ] (12); Step 1.6: Build a convolutional neural network; Step 1.7: Perform normalization preprocessing on the input data of formula (7) and the label data of formula (12), call the Matconvnet tool in Matlab, and use the input data and label data to train the convolutional neural network of step 1.6; Step 1.8: Delete the output layer of the convolutional neural network trained in step 1.7, and use the last layer of the fully connected layer as the new output layer to obtain the static voltage stability margin prediction model; The input data structure in step 1.2 is a three-dimensional matrix consisting of two 3×n matrices, that is, matrix X=[X B X T ] 3×n×2 ; The matrix X=[X B X T ] 3×n×2 Medium X B It is composed of basic load data, and its specific composition is shown in formula (1): The superscript B of the matrix and the superscript b of the element represent the basic state, n is the number of nodes in the system, and X B The first row of elements is the active power P of the net load of system node i i b , the second row of the matrix is the reactive power of the net load at system node i The third row of the matrix is the active power generated by system node i The expression of each row element is shown in formula (2): Where i is the system node number, and i = 1, 2, 3..., n; the active power P of the net load in the basic state i b and reactive power It is defined as formula (3): in, is the active power of the load at node i in the basic state, is the active power of the photovoltaic power of node i in the basic state, is the reactive power of the load at node i in the basic state, is the PV reactive power of node i in the basic state; The matrix X=[X B X T ] 3×n×2 Medium X T It is composed of target load data, and its specific composition is shown in equations (4), (5), and (6): The superscript T of the matrix and the superscript t of the element represent the target state. is the active power of the load of node i under the target state, is the active power of the photovoltaic power of node i under the target state, is the reactive power of the load at node i in the target state, is the PV reactive power of node i under the target state.

2. According to claim 1, a method for online prediction of static voltage stability margin based on convolutional neural network, Its characteristics are The power grid topology information in the power transmission network topology model in step 1.1 includes thermal power plant access nodes and photovoltaic access nodes.

3. According to claim 1, a method for online prediction of static voltage stability margin based on convolutional neural network, Its characteristics are The load-type continuous power flow model parameterized in step 1.4 is expressed by equation (8): Where P i b is the active power of the load at node i in the initial state, is the reactive power of the load at node i in the initial state, is the active power generated by node i in the initial state; ΔP i , ΔQ i , ΔPG i is the predetermined load increment and power generation increment, which is obtained by subtracting equation (2) from equation (5), namely: The λ in formula (8) is a scalar parameter reflecting the load level. After continuous power flow calculation, the λ sequence is obtained: 0<λ≤λ max ; When λ=λ max When , it means that the power of each node increases in a predetermined direction and the power reaches the limit.

4. According to claim 1, a method for online prediction of static voltage stability margin based on convolutional neural network, Its characteristics are The feature extraction part of the convolutional neural network in step 1.6 includes 5 convolutional layers, 5 activation layers, and 3 batch normalization layers; the prediction part of the convolutional neural network includes 4 fully connected layers; the output layer of the convolutional neural network uses the sum of the squares of the difference between the predicted value and the true value as the objective function: In the above formula, N is the number of batch data; P i It is the output data after the i-th input data passes through the fourth fully connected layer of the network, that is, the predicted static voltage stability margin.

5. According to claim 4, a method for online prediction of static voltage stability margin based on convolutional neural network, Its characteristics are The purpose of training the convolutional neural network in step 1.7 is to minimize the value of the objective function C by adjusting the network parameters. After several steps of training, C decreases to a stable value and no longer decreases, and the training is stopped.

6. According to claim 1, a method for online prediction of static voltage stability margin based on convolutional neural network, Its characteristics are The step 2 performs real-time prediction of the photovoltaic power stability margin based on the static voltage stability margin prediction model trained in step 1, specifically comprising: step 2.1: obtaining real-time input data X; Step 2.2: The real-time input data X is normalized and used as the input of the static voltage stability margin prediction model obtained in step 1.

8. The output is the real-time predicted static voltage stability margin.

7. The method for online prediction of static voltage stability margin based on convolutional neural network according to claim 6, Its characteristics are The step 2.1 obtains real-time input data X, specifically including: the active power of the load of node i in the basic state Active power of photovoltaic power at node i in basic state Reactive power of the load at node i in the basic state The reactive power of the photovoltaic power plant at node i in the basic state The active power of the load at node i under the target state is collected online in real time by the SCADA system. Reactive power of the load at node i in the target state According to the dispatch load forecast, the active power of the photovoltaic output of node i in the target state is The reactive power of the photovoltaic output of node i under the target state According to the local temperature and illumination data prediction; the real-time input data X = [X B X T ] 3×n×2 .

Images