A prediction method for a generator rotor angle prediction model after a power system fault

Abstract

The invention relates to the technical field of electric power systems, and in particular to a prediction method for a generator rotor angle prediction model after a fault in an electric power system, the method comprising: obtaining a large number of training samples through simulation, extracting characteristic quantities of a single generator as original data; using the generator rotor angle prediction model to take L moments of rotor angle characteristics as input, and obtaining an output formula of the generator rotor angle prediction model; minimizing the distance between the predicted value and the true value of the rotor angle of each layer, and obtaining rotor angle data at a specific moment; assigning different weights to errors at different time points, and using the Mahalanobis distance to measure the distance between the predicted value and the true value; and discretizing the prediction model in the time dimension. The invention can accurately predict the generator rotor angle after a fault, alleviates the dependence on input data, and can also ensure the accuracy of the prediction result, thereby providing strong technical support for the stable operation of the electric power system.

Description

Technical Field

[0001] The invention belongs to the technical field of electric power systems, and in particular relates to a prediction method of a generator rotor angle prediction model after a fault in an electric power system. Background Art

[0002] For transient stability assessment (TSA), most studies generally use simplified binary state variables to describe system stability, that is, the system state is simplified to 0 or 1, representing instability and stability respectively. This division ignores the complexity and continuity of the system in the dynamic process after the fault. The rotor angle can directly reflect the stability of the power system. Therefore, the regression task of TSA is mainly to predict the change trajectory of the rotor angle after the fault. By predicting the future rotor angle, the subsequent control time can be effectively shortened. Thereby improving the efficiency of system response and control strategy.

[0003] In the power system, the phasor measurement unit (PMU) as the basic information collection unit of the wide area measurement system (WAMS) has laid a solid data foundation for the application of artificial intelligence technology in the field of power systems. Usually, the solution process of the machine learning power system method based on data mining must first be trained with a large amount of power system data to build a relevant problem-solving model.

[0004] However, machine learning models are often viewed as highly closed “black boxes” whose internal working mechanisms and decision-making processes are often opaque. And the reliability of the data used for training will greatly affect the effectiveness of the prediction model. The measurement accuracy of the PMU may be affected by time synchronization errors or phase errors. In addition, the data may also be interfered by different noise distributions, which can significantly affect the prediction performance. Summary of the invention

[0005] In order to solve the problem that the existing machine learning models have unstable model performance in the field of power systems and cannot well predict the generator rotor angle after a fault, the present invention proposes a prediction method for a generator rotor angle prediction model after a power system fault, which can accurately predict the generator rotor angle after a fault, alleviates the dependence on input data, and can also ensure the accuracy of the prediction results, providing strong technical support for the stable operation of the power system.

[0006] The technical solution of the present invention is as follows:

[0007] A prediction method for a generator rotor angle prediction model after a power system fault, the prediction method comprising the following steps:

[0008] Step 1: Obtain a large number of training samples through simulation and extract the characteristic quantities of a single generator as raw data;

[0009] Step 2: construct a Lipschitz continuous deep neural network rotor angle prediction model through a multi-layer perceptron. The Lipschitz continuous deep neural network rotor angle prediction model takes the rotor angle features at L moments as input and outputs the rotor angle at the next moment, that is, the predicted value of the rotor angle;

[0010] Step 3: Minimize the distance between the predicted value and the true value of each layer and fit the rotor angle data at a specific moment;

[0011] Step 4: Assign different weights to the errors at different time points, and use the Mahalanobis distance to measure the distance between the predicted value and the true value;

[0012] Step 5: Discretize the Lipschitz continuous deep neural network rotor angle prediction model in the time dimension, and integrate the rotor motion equation into the Lipschitz continuous deep neural network rotor angle prediction model.

[0013] Furthermore, the characteristic quantity of the single generator described in step 1 includes the rotor angle and angular velocity information of the single generator set.

[0014] Furthermore, the output formula of the Lipschitz continuous deep neural network rotor angle prediction model described in step 2 is as follows:

[0015] (1);

[0016] in, is the predicted value of the rotor angle of the i-th layer, x is the data of the rotor angle, and are the output values ​​of the models at the i-th layer and i-1-th layer of the prediction model, respectively. and Represent the weight and bias of the i-th layer respectively, is the Leakyrelu nonlinear activation function;

[0017] (2);

[0018] in, , , and , , are the weights and biases of the 1st, 2nd, and Lth layers of the model respectively.

[0019] Furthermore, step 3 is specifically as follows:

[0020] The output of the i-th layer corresponds to the rotor angle at the i-th moment, by minimizing and the actual rotor angle value at the i-th moment The distance is realized as shown below:

[0021] (3);

[0022] in, is the real rotor angle value at the i-th moment Distance The value of x when the minimum value is reached, are the trainable parameters of the model, is the output value of the i-th layer of the prediction model, is the predicted value of the rotor angle at the i-th layer;

[0023] and the actual rotor angle value at the i-th moment Distance The expression is as follows:

[0024] (4);

[0025] Among them, N is the number of samples.

[0026] Furthermore, step 4 is specifically as follows:

[0027] Different weights are assigned to the losses of different layers, and the weight values ​​of different layers are obtained by the following formula (5):

[0028] (5);

[0029] in, is the weight of the i-th layer, It is expressed as the true angular velocity deviation at the i-th moment, L is the number of model layers, and i is the current number of layers;

[0030] Normalize the obtained weight value, the formula is as follows:

[0031] (6);

[0032] in, is the weight of the jth layer, is the weight of the i-th layer after normalization;

[0033] For the predicted rotor angle and the true value of the rotor angle The covariance matrix C of is as follows:

[0034] (7);

[0035] Therefore, the covariance between the true and predicted values ​​is It is shown in the following formula (8):

[0036] (8);

[0037] in, represents the predicted value of the rotor angle at the i-th layer, represents the true value of the rotor angle at the jth moment, represents the expected value, , It represents the expected value of the product of the difference between the predicted value of the rotor angle and its expected value, and the difference between the actual value of the rotor angle and its expected value;

[0038] The distance between the predicted rotor angle and the true rotor angle The Mahalanobis distance is used as follows:

[0039] (9);

[0040] Therefore, the data loss term As shown in the following formula (10):

[0041] (10);

[0042] Furthermore, step 5 is specifically as follows:

[0043] Each prediction process is discretized into L prediction cycles, with a sampling period of Δt. The rotor angle data with a length of L is rolled input to predict the rotor angle at time L+1. The rotor angle output at the previous moment and the rotor angle output at the next moment in the prediction model satisfy the discretized form of the generator rotor motion equation, as shown in the following equation (11):

[0044] (11);

[0045] in, and are the weights and biases of the i+1th and ith layers of the model respectively; is the true angular velocity of the generator at time i, are the prediction model output values ​​at the i-th moment and the i-1-th moment respectively.

[0046] Furthermore, when training the model, the overall loss is divided into two aspects: The overall loss function is shown in formula (12):

[0047] (12);

[0048] in, is the data loss term, is a physical constraint;

[0049] As shown in the following formula (13):

[0050] (13);

[0051] in, is the generator angular velocity deviation at the i-th moment, The complex trapezoidal formula is used to obtain the value, as shown in the following formula (14):

[0052] (14);

[0053] in, is the output of the i+1th layer, are the generator angular velocity deviations at the 1st, j+1th, and Lth moments respectively.

[0054] Furthermore, after step 5, a rolling prediction method is used to input the generator rotor angle data of L cycles into the prediction model, and the prediction model outputs the generator rotor angle of the next cycle.

[0055] Compared with the prior art, the present invention has the following beneficial effects:

[0056] (1) The present invention embeds physical laws into the model architecture, which can effectively prevent the model from learning the wrong data model and making wrong predictions, improve the accuracy and stability of the prediction results, and effectively reduce the dependence of the machine learning model on data.

[0057] (2) The present invention ensures that each layer in the model accurately fits the rotor angle at a specific moment by minimizing the distance between the output of each layer and the rotor angle at the actual moment, and each layer of the model outputs the rotor angle at the current fitting moment. The layer-by-layer accurate fitting method helps to improve the overall fitting accuracy of the model to the rotor angle change trend. At the same time, the present invention integrates the generator rotor motion equation into the machine learning model by embedding the generator rotor motion equation into the hierarchical structure of the model. The angular velocity deviation is integrated by the complex trapezoidal formula to obtain the deviation of the rotor angle. This deviation is then used to fine-tune the internal parameters of the model to ensure the physical relationship between the output at the previous moment and the output at the next moment, so as to ensure the accuracy and reliability of the model and reduce the impact of adverse factors in the data.

[0058] (3) The present invention proposes time-adaptive weights, which target the error between the predicted value and the true value of each layer. The weight value increases with time, ensuring that the final prediction result pays more attention to the data at the nearest moment, thereby enhancing the accuracy of the prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1It is the predicted diagram of the rotor trajectory of the G1 unit under steady state;

[0060] Figure 2 It is the prediction of the rotor trajectory of the G1 unit under unstable state;

[0061] Figure 3 It is the wiring diagram of the IEEE39 node system;

[0062] FIG. 4 is a schematic diagram of a prediction model of the present invention. DETAILED DESCRIPTION

[0063] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0064] See also Figure 1-2 A prediction method for a generator rotor angle prediction model after a power system fault, the prediction method comprising the following steps:

[0065] Step 1: Obtain a large number of training samples through simulation and extract the characteristic quantities of a single generator as raw data;

[0066] Step 2: construct a Lipschitz continuous deep neural network rotor angle prediction model through a multi-layer perceptron. The Lipschitz continuous deep neural network rotor angle prediction model takes the rotor angle features at L moments as input and outputs the rotor angle at the next moment, that is, the predicted value of the rotor angle;

[0067] Step 3: Minimize the distance between the predicted value and the true value of each layer and fit the rotor angle data at a specific moment;

[0068] Step 4: Assign different weights to the errors at different time points, and use the Mahalanobis distance to measure the distance between the predicted value and the true value;

[0069] Step 5: Discretize the Lipschitz continuous deep neural network rotor angle prediction model in the time dimension, and integrate the rotor motion equation into the Lipschitz continuous deep neural network rotor angle prediction model.

[0070] The specific model configuration of the present invention is as follows:

[0071] The number of layers L is set to 4; the number of layers of PINN and MLP without physical constraints is also 4, and the number of neurons in each layer is 200; the LSTM structure is 2 LSTM layers and 1 fully connected layer, and the GRU structure is similar to LSTM, which is 2 GRU layers plus 1 fully connected layer.

[0072] The multi-layer perceptron is composed of multiple fully connected layers, with a specific structure of input layer, hidden layer and output layer, and an activation function is selected after each layer to introduce nonlinear characteristics. The layers of the present invention are interconnected, and each layer has an output, indicating that this layer predicts the rotor angle at the corresponding moment.

[0073] During implementation, the simulation software uses PSD-BPA, such as Figure 3 The test system shown uses the IEEE-39 node system, covering the impact of speed regulators, voltage regulators, and new energy factors such as wind power generation and photovoltaic power generation. The system's operating status has five load levels: 90%, 95%, 100%, 105% and 110%. The fault locations occurred at 10%, 50% and 90% of each line. Finally, a total of 10,000 samples were generated in the IEEE-39 node system. G1 unit samples were extracted from the 10,000 fault samples at a ratio of 4:1 to form the training set and test set.

[0074] In one embodiment of the present invention, the characteristic quantity of the single generator mentioned in step 1 includes the rotor angle and angular velocity information of the single generator set.

[0075] In one embodiment of the present invention, the output formula of the Lipschitz continuous deep neural network rotor angle prediction model described in step 2 is as follows:

[0076] (1);

[0077] in, is the predicted value of the rotor angle of the i-th layer, x is the data of the rotor angle, are the output values ​​of the models at the i-th layer and i-1-th layer of the prediction model, respectively. Represent the weight and bias of the i-th layer respectively, is the Leakyrelu nonlinear activation function;

[0078] (2);

[0079] in, are the weights and biases of the 1st, 2nd, and Lth layers of the model respectively.

[0080] In order to enable the network model to accurately fit the changing trend of the rotor angle data, in another embodiment of the present invention, step 3 is specifically as follows: the output of each layer is regarded as the rotor angle state of the rotor at a specific time point, and the output of the i-th layer corresponds to the rotor angle at the i-th moment, and the i-th layer output corresponds to the i-th moment rotor angle. and the actual rotor angle value at the i-th moment The distance is realized as shown below:

[0081] (3);

[0082] in, is the real rotor angle value at the i-th moment Distance The value of x when the minimum value is reached, are the trainable parameters of the model, is the output value of the i-th layer of the prediction model, is the predicted value of the rotor angle at the i-th layer;

[0083] and the actual rotor angle value at the i-th moment Distance The expression is as follows:

[0084] (4);

[0085] Among them, N is the number of samples.

[0086] In one embodiment of the present invention, step 4 is specifically as follows: different weights are assigned to the losses of different layers, and the weight values ​​of different layers are obtained by the following formula (5):

[0087] (5);

[0088] in, is the weight of the i-th layer, It is expressed as the true angular velocity deviation at the i-th moment, L is the number of model layers, and i is the current number of layers;

[0089] Normalize the obtained weight value, the formula is as follows:

[0090] (6);

[0091] in, is the weight of the jth layer, is the weight of the i-th layer after normalization;

[0092] For the predicted rotor angle and the true value of the rotor angle The covariance matrix C of is as follows:

[0093] (7);

[0094] Therefore, the covariance between the true and predicted values ​​is It is shown in the following formula (8):

[0095] (8);

[0096] in, represents the predicted value of the rotor angle at the i-th layer, represents the true value of the rotor angle at the jth moment, represents the expected value, , It represents the expected value of the product of the difference between the predicted value of the rotor angle and its expected value, and the difference between the actual value of the rotor angle and its expected value;

[0097] The distance between the predicted rotor angle and the true rotor angle The Mahalanobis distance is used as follows:

[0098] (9);

[0099] Therefore, the data loss term As shown in the following formula (10):

[0100] (10);

[0101] In addition, in order to make the model fit the actual generator rotor motion state, the rotor motion equation is introduced to guide model training. Each layer of the model receives the output of the previous layer as its input. Therefore, the predicted value at the previous moment and the predicted value at the next moment should be constrained by the generator rotor motion equation. In one embodiment of the present invention, step 5 is specifically:

[0102] Each prediction process is discretized into L prediction cycles, with a sampling period of Δt. The rotor angle data with a length of L is rolled input to predict the rotor angle at time L+1. The rotor angle output at the previous moment and the rotor angle output at the next moment in the prediction model satisfy the discretized form of the generator rotor motion equation, as shown in the following equation (11):

[0103] (11);

[0104] in, are the weights and biases of the i+1th and ith layers of the model respectively; is the true angular velocity of the generator at time i, are the prediction model output values ​​at the i-th moment and the i-1-th moment respectively.

[0105] In another embodiment of the present invention, when the model is trained, the overall loss is divided into two aspects: The overall loss function is shown in formula (12):

[0106] (12);

[0107] in, is the data loss term, is a physical constraint;

[0108] As shown in the following formula (13):

[0109] (13);

[0110] in, is the generator angular velocity deviation at the i-th moment, The complex trapezoidal formula is used to obtain the value, as shown in the following formula (14):

[0111] (14);

[0112] in, is the output of the i+1th layer, are the generator angular velocity deviations at the 1st, j+1th, and Lth moments respectively.

[0113] The core function of the loss function is to quantify the difference between the model's predicted value and the true value. By calculating this difference, the loss function provides a clear optimization goal for the model, which is to minimize the loss value. The loss function is a necessary part of deep learning, and the model is trained by minimizing the loss function. During the training process, this loss value is passed to each layer of the network through back propagation to guide the update direction and amplitude of the network parameters. Compared with ordinary prediction models, when there are bad factors in the input data, there are physical constraints to ensure that the upper and lower layer outputs conform to the rotor equation and are not easily affected by interference factors.

[0114] Therefore, when there are adverse factors in the input data, the network can make effective internal adjustments to ensure the accuracy and physical consistency of the prediction.

[0115] In one embodiment of the present invention, after step 5, a rolling prediction method is used to input generator rotor angle data of L cycles into the prediction model, and the prediction model outputs the generator rotor angle of the next cycle.

[0116] The following is an example of the present invention in the implementation process: Calculate column 1:

[0117] The rotor angle of the G1 unit 36-43 cycles after the fault is input in a rolling manner, and the prediction target is 40-44 cycles. Taking the prediction of the 40th cycle of the G1 unit 36-39 cycles as an example, the specific prediction process is as follows:

[0118] (1) The 36-39 cycle data of the G1 unit is input into the model, so that each layer of the model fits the changing trend of the given cycle. At this time, the prediction error of each layer is They are 0.00435, 0.00372, 0.00481, and 0.00262 respectively.

[0119] (2) In order to more accurately reflect the error between the predicted value and the true value, different weights are assigned to the errors at different time points, which can be calculated by equations (5) and (6): 0.1524, 0.1905, 0.2286, and 0.4285, respectively. Giving larger weights to recent moments can effectively increase the attention paid to recent errors.

[0120] (3) The layers are constrained by the rotor equation, which will produce a physical constraint loss, which can be calculated by equations (13) and (14), which are 0.01028, 0.01526, and 0.00971, respectively. The physical loss is conducive to the internal adjustment of the model, so that the physical consistency of the front and back outputs is enhanced.

[0121] (4) Finally, the predicted value of the 40th cycle is 23.4726 degrees. The rotor trajectories of stable and unstable states are randomly selected for prediction, and the basic model without physical constraints and PINN are compared. The prediction results are shown in Tables 1, 2 and 3.

[0122] Table 1 Mean square error (MSE) and absolute percentage error (MAPE) for different forecast periods

[0123]

[0124] Table 2 Actual value and predicted value of rotor angle under stable condition

[0125]

[0126] Table 3 Actual value and predicted value of rotor angle under unstable condition

[0127]

[0128] As shown in Tables 1, 2 and 3, Table 1 shows the MSE and MAPE of each prediction period of different models, and Tables 2 and 3 are the comparisons of the actual value and predicted value of the rotor angle under stable and unstable conditions respectively. Compared with other existing models, the invention can more accurately predict the rotor angle trajectory after a fault.

[0129] Calculation example 2:

[0130] Since in practical applications, data may be subject to varying degrees of noise interference during the process of collection and transmission, the prediction model is required to have strong noise resistance. In order to verify that the prediction model with embedded physical laws proposed in the present invention has strong noise resistance and can effectively alleviate data dependence, the present invention and the above-mentioned comparison model were subjected to noise experiments. G1 unit samples were extracted from 10,000 fault samples at a ratio of 4:1 to form a training set and a test set. Gaussian white noise with a signal-to-noise ratio of 40, 30, and 20 dB was added to the test set data to simulate the error caused by noise during the measurement or transmission of actual data.

[0131] Take the prediction of the 40th cycle of the G1 unit from the 36th to the 39th cycle and the addition of 40dB noise as an example. The specific process is as follows:

[0132] (1) Input the 36-39 cycle data of the G1 unit after adding 40dB noise into the model, so that each layer of the model fits the changing trend of the given cycle. At this time, the prediction error of each layer is They are 0.0612, 0.0818, 0.0627, and 0.0512 respectively.

[0133] (2) In order to more accurately reflect the error between the predicted value and the true value, different weights are assigned to the errors at different time points, which can be calculated by equations (5) and (6): 0.1524, 0.1905, 0.2286, and 0.4285, respectively. Giving larger weights to recent moments can effectively increase the attention paid to recent errors.

[0134] (3) The layers are constrained by the rotor equation, which will produce a physical constraint loss, which can be calculated by equations (13) and (14), which are 0.01028, 0.01526, and 0.00971, respectively. The physical loss is conducive to the internal adjustment of the model, so that the physical consistency of the front and back outputs is enhanced.

[0135] (4) Finally, the predicted value of the 40th cycle is 24.0212 degrees. The prediction results of MLP and PINN without physical constraints and LSTM and GRU with different noises are shown in Table 4.

[0136] Table 4 Prediction performance of each model under different noise levels

[0137]

[0138] As shown in Table 4, Table 4 shows the impact of noise with signal-to-noise ratios of 40dB, 30dB, and 20dB added to the test data on the prediction performance of each model. Compared with other existing models, the invention has better noise resistance and can effectively alleviate the dependence on data.

[0139] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A prediction method for a generator rotor angle prediction model after a power system fault, characterized in that: The prediction method comprises the following steps: Step 1: Obtain a large number of training samples through simulation and extract the characteristic quantity of a single generator as the original data; Step 2: construct a Lipschitz continuous deep neural network rotor angle prediction model through a multi-layer perceptron. The Lipschitz continuous deep neural network rotor angle prediction model takes the rotor angle features at L moments as input and outputs the rotor angle at the next moment, that is, the predicted value of the rotor angle; Step 3: Minimize the distance between the predicted value and the true value of each layer and fit the rotor angle data at a specific moment; Step 4: Assign different weights to the errors at different time points, and use the Mahalanobis distance to measure the distance between the predicted value and the true value; Step 5: discretize the Lipschitz continuous deep neural network rotor angle prediction model in the time dimension, and integrate the rotor motion equation into the Lipschitz continuous deep neural network rotor angle prediction model; Step 3 is as follows: The output of the i-th layer corresponds to the rotor angle at the i-th moment, by minimizing v (i) (x) and the actual rotor angle value δ at the i-th moment i The distance is realized as shown below: Among them, argmin is the actual rotor angle value at the i-th moment The distance S(v (i) (x),δ i ) is the value of x when it reaches the minimum value, θ is the trainable parameter of the model, and v (i) (x) is the output value of the i-th layer of the prediction model, is the predicted value of the rotor angle at the i-th layer; v (i) (x) and the actual rotor angle value at the i-th moment The distance S(v (i) (x),δ i ) is as follows: Where N is the number of samples; Step 4 is as follows: Different weights are assigned to the losses of different layers, and the weight values ​​of different layers are obtained by the following formula (5): in, is the weight of the i-th layer, Δω i It is expressed as the true angular velocity deviation at the i-th moment, L is the number of model layers, and i is the current number of layers; Normalize the obtained weight value, the formula is as follows: in, is the weight of the jth layer, w i is the weight of the i-th layer after normalization; For the predicted rotor angle And the true value of the rotor angle δ = {δ1, δ2, δ3, δ4, ...δ n The covariance matrix C of} is as follows: Therefore, the covariance C between the true value and the predicted value is ij As shown in the following formula (8): in, represents the predicted value of the rotor angle at the i-th layer, δ j represents the true value of the rotor angle at the jth moment, E[·] represents the expected value, i,j=1,2,3…n, It represents the expected value of the product of the difference between the predicted value of the rotor angle and its expected value, and the difference between the actual value of the rotor angle and its expected value; The distance between the predicted rotor angle and the true rotor angle The Mahalanobis distance is used as follows: Therefore, the data loss term L mse As shown in the following formula (10): Step 5 is as follows: Each prediction process is discretized into L prediction cycles, with a sampling period of Δt. The rotor angle data of length L is input in a rolling manner to predict the rotor angle at time L+1. The rotor angle output at the previous moment and the rotor angle output at the next moment in the prediction model satisfy the discretized form of the generator rotor motion equation, as shown in the following equation (11): Among them, W (i+1 )、W (i) and b (i+1) 、b (i) are the weights and biases of the i+1th and ith layers of the model respectively; ω i is the true angular velocity of the generator at time i, v (i) (x) and v (i-1) (x) are the output values ​​of the prediction model at the i-th moment and the i-1-th moment respectively.

2. The prediction method of the generator rotor angle prediction model after a power system fault according to claim 1 is characterized in that: The characteristic quantities of the single generator described in step 1 include the rotor angle and angular velocity information of the single generator set.

3. The prediction method of the generator rotor angle prediction model after a power system fault according to claim 1 is characterized in that: The output formula of the Lipschitz continuous deep neural network rotor angle prediction model described in step 2 is as follows: in, is the predicted value of the rotor angle of the i-th layer, x is the data of the rotor angle, v (i) (x) and v (i-1) (x) are the output values ​​of the i-th layer and i-1-th layer of the prediction model, respectively, W (i) and b (i) Represent the weight and bias of the i-th layer respectively, σ is the Leakyrelu nonlinear activation function; v(x)=σ(W (1) (σ(W (2) (…σ(W (L) (x)+b (L) ))+b (2) ))+b (1) ) (2) Among them, W (1) , W (2) , W (L) and b (1) 、b (2) 、b (L) are the weights and biases of the 1st, 2nd, and Lth layers of the model respectively.

4. The prediction method of the generator rotor angle prediction model after a power system fault according to claim 1 is characterized in that: When training the model, the overall loss is divided into two aspects. Loss is the overall loss function as shown in formula (12): Loss=L mse +L physic (12) Among them, L mse is the data loss term, L physic is a physical constraint; L physic As shown in the following formula (13): Among them, Δω i is the generator angular velocity deviation at the i-th moment, Δω i The complex trapezoidal formula is used to obtain the value, as shown in the following formula (14): Among them, v( i+1) (x) is the output of the i+1th layer, Δω1, Δω j+1 and Δω L are the generator angular velocity deviations at the 1st, j+1th, and Lth moments respectively.

5. The prediction method of the generator rotor angle prediction model after a power system fault according to claim 1 is characterized in that: After step 5, a rolling prediction method is used to input the generator rotor angle data of L cycles into the prediction model, and the generator rotor angle of the next cycle is output through the prediction model.

Images