Method for enhancing digital twinborn model data of power system in low measurement environment

Abstract

The invention relates to a power system digital twin model data enhancement method in a low measurement environment. The method comprises the following steps: firstly, acquiring initial data of a power system and preprocessing the initial data to obtain a training data set; then, establishing a physical constraint condition; then, a physical constraint generative adversarial network model is constructed based on the physical constraint conditions, the physical constraint generative adversarial network model comprises a generative network and a discrimination network, the generative network is used for repairing data, and the discrimination network is used for discriminating the truth rate of the repaired data; then, training the physical constraint generative adversarial network model based on the training data set, including discriminating network updating and generative network updating; and finally, inputting to-be-repaired power system data into the trained physical constraint generative adversarial network model to obtain enhanced data. The physical consistency and reliability of the data are enhanced, and the precision, stability and generalization ability of the digital twin model are improved.

Description

Technical Field

[0001] The present application relates to the technical field of electromagnetic transient simulation of power systems, and in particular to a method for enhancing data of a digital twin model of a power system in a low-measurement environment. Background Art

[0002] As the scale and complexity of power systems continue to expand, the digitalization and intelligent transformation of power systems has become an important direction for the development of modern power grids. Among them, the digital twin model, as a virtual mapping model that can accurately reflect the operating status of physical objects, has important applications in power system monitoring, operation optimization and fault prevention.

[0003] At present, neural networks, as a tool with powerful fitting ability and nonlinear mapping characteristics, are widely used in the construction of digital twin models and have achieved results to a certain extent. However, this neural network-based method is highly dependent on the quality of measurement data. Specifically, the training of neural networks requires a large amount of high-quality measurement data. In low-measurement environments, the lack of sensor data will significantly affect the learning effect of the neural network and severely limit the accuracy and stability of the model. Therefore, in low-measurement environments, digital twin models based on neural networks cannot effectively reflect the true state and dynamic changes of the power system due to the lack of high-quality data set training.

[0004] In the existing technology, research on neural networks and digital twin models has conducted a lot of exploration on model construction and data fitting capabilities, but most of them ignore the impact of missing measurement data on model performance. In addition, although some methods have achieved certain results by introducing generative adversarial networks to generate supplementary missing data, these methods usually only focus on the statistical characteristics of generated data, while ignoring the physical constraints and mechanism characteristics in the power system, which makes the generated data uncertain and limited in the physical sense. Therefore, although these data-driven methods provide solutions to the problem of missing data in theory, it is still difficult to ensure the consistency of generated data with the actual state of the power system in actual low-measurement environments, which in turn affects the reliability of digital twin models.

[0005] Therefore, in the related technology, there is an urgent need for a method that can solve the problem of missing measurement sample data in a low-measurement environment and ensure the reliability of repair data. Summary of the invention

[0006] Based on this, it is necessary to provide a data enhancement method for the digital twin model of the power system in a low-measurement environment, which can solve the problem of missing measurement sample data in a low-measurement environment and ensure the reliability of repair data, in response to the above-mentioned technical problems.

[0007] In a first aspect, the present application provides a method for enhancing data of a digital twin model of a power system in a low-measurement environment. The method comprises: Obtain the initial data of the power system and preprocess it to obtain a training data set; Establish physical constraints; Based on the physical constraint conditions, a physical constraint generative adversarial network model is constructed, wherein the physical constraint generative adversarial network model includes a generative network and a discriminative network, wherein the generative network is used to repair data, and the discriminative network is used to discriminate the authenticity of the repaired data; Training the physical constraint generative adversarial network model based on the training data set, including discriminative network update and generative network update; The data of the power system to be repaired is input into the trained physical constraint generative adversarial network model to obtain enhanced data.

[0008] Optionally, in one embodiment of the present application, the generative network includes an input layer, a fully connected layer, a batch normalization layer and an output layer, and the discriminative network includes an input layer, a fully connected layer and an output layer, wherein the batch normalization layer is used to standardize the input of each layer to maintain the stability of the input distribution.

[0009] Optionally, in one embodiment of the present application, training the physical constraint generative adversarial network model based on the training data set includes: Selecting a real sample with missing values ​​from the training data set, inputting the real sample and the noise vector into the generation network to obtain a repaired real sample; A mask matrix is ​​determined based on the repaired real sample, the real loss and the repair loss of the discriminant network are calculated based on the mask matrix, the loss of the generator network is calculated based on the repaired real sample and the mask matrix, the real loss, the repair loss and the loss of the discriminant network and the generator network are minimized, and the parameters of the discriminant network and the generator network are updated.

[0010] Optionally, in one embodiment of the present application, the calculation formula of the actual loss is:

[0011] in, Represents the loss of the true value part, The value representing the position is the original true value or the generated network interpolation value. The value representing the position is the original true value, is the probability that the position output by the discriminant network is the true value, That is, the value at that position.

[0012] Optionally, in one embodiment of the present application, the calculation formula of the repair loss is:

[0013] in, Indicates the loss of the repair value part, The value representing the position is the original true value or the generated network interpolation value. The value representing the position is the generated network interpolation value, is the probability of the position value being the interpolated value output by the discriminant network, That is, the value at that position.

[0014] Optionally, in one embodiment of the present application, the loss of the generated network includes statistical loss and physical loss, wherein the calculation formula of the statistical loss is:

[0015] in, is the statistical loss term, The value representing the position is the generated network interpolation value, is the probability that the position value output by the discriminant network is the true value, That is, the value of the position; The calculation formula for physical loss is:

[0016] in, is the physical loss term, It is a physical loss function defined according to the physical constraints of the power system or component. is a random noise vector, To generate a complete sample after network repair.

[0017] Optionally, in one embodiment of the present application, the method further includes: The mean square error, mean absolute error, structural similarity and physical consistency loss are used to evaluate the restoration effect of the physical constraint generative adversarial network model.

[0018] The above-mentioned method for data enhancement of a digital twin model of a power system in a low-measurement environment, first, obtains the initial data of the power system and preprocesses it to obtain a training data set; then, establishes physical constraints; then, constructs a physical constraint generative adversarial network model based on the physical constraints, and the physical constraint generative adversarial network model includes a generative network and a discriminative network, wherein the generative network is used to repair data, and the discriminative network is used to discriminate the authenticity of the repaired data; then, trains the physical constraint generative adversarial network model based on the training data set, including updating the discriminative network and updating the generative network; finally, inputs the power system data to be repaired into the trained physical constraint generative adversarial network model to obtain enhanced data. That is to say, through the introduction of PI-GAN, physical information is embedded in the data generation process of the generative adversarial network, ensuring that the generated data is not only consistent with the real data in a statistical sense, but also can meet the physical characteristics requirements of the power system. In this way, the generated data can not only interpolate the missing data due to sensor failure and other reasons, but also maintain accurate estimation and dynamic simulation capabilities of the system state in a noisy environment. In practical applications, the operating conditions of power systems are complex and changeable. Traditional neural network-based methods often find it difficult to maintain consistent performance under different working conditions. By combining the advantages of physical information and generative adversarial networks, the digital twin model constructed in this application can adapt to different system structures and operating conditions, maintain good predictive capabilities and adaptability, and further enhance its application value in actual power systems. The physical consistency and reliability of data are enhanced, and the accuracy, stability and generalization ability of the digital twin model are improved, thereby providing reliable technical support for power system status monitoring, operation optimization and fault prevention. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 An application environment diagram of a method for enhancing data of a digital twin model of a power system in a low-measurement environment in one embodiment; Figure 2 It is a flowchart of a method for enhancing data of a digital twin model of a power system in a low-measurement environment in one embodiment; Figure 3 A schematic diagram of the structure of PI-GAN in one embodiment; Figure 4 A technical framework diagram of a method for enhancing data of a digital twin model of a power system in a low-measurement environment in one embodiment; Figure 5 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0021] The embodiment of the present application provides a method for enhancing data of a digital twin model of a power system in a low-measurement environment, which can be applied to Figure 1 In the application environment shown. Among them, the terminal communicates with the server through the network. The data storage system can store the data that the server needs to process. The data storage system can be integrated on the server, or it can be placed on the cloud or other network servers. Among them, the terminal can be but not limited to various personal computers, laptops, smart phones, tablets, Internet of Things devices and portable wearable devices. The Internet of Things devices can be smart speakers, smart TVs, smart air conditioners, smart car-mounted devices, etc. Portable wearable devices can be smart watches, smart bracelets, head-mounted devices, etc. The server can be implemented as an independent server or a server cluster consisting of multiple servers.

[0022] In one embodiment, Figure 2 As shown in the figure, a data enhancement method for the digital twin model of a power system in a low-measurement environment is provided. Figure 1 The server in the example is used to illustrate the following steps: S201: Acquire initial data of the power system and perform preprocessing to obtain a training data set.

[0023] In the embodiment of the present application, first, the input and output information of each subsystem or electrical component in the power system, such as voltage, current and other key electrical quantity information, is collected and sorted to ensure data diversity and timeliness. Furthermore, in order to improve the training effect of the subsequent data-driven model, the data of different precisions and dimensions are normalized to the maximum and minimum values ​​to meet the input requirements of the neural network.

[0024] S203: Establish physical constraints.

[0025] In an embodiment of the present application, the steady-state equations and dynamic characteristics of the power system are taken into consideration, historical measurement information and known parameters are combined, a physical model of the target component or subsystem is constructed, and physical constraints are established based on its physical model and input and output electrical quantity information, so as to be combined with subsequent models to ensure that the repair data meets the physical constraints.

[0026] S205: Constructing a physical constraint generative adversarial network model based on the physical constraint conditions, wherein the physical constraint generative adversarial network model includes a generative network and a discriminative network, wherein the generative network is used to repair data, and the discriminative network is used to discriminate the authenticity of the repaired data.

[0027] In the embodiment of the present application, a physical constraint generative adversarial network model is constructed based on the established physical constraint conditions. Different from the general neural network, the generative adversarial network model includes two components: the generative network and the discriminative network. The physical constraint generative adversarial network introduces the guidance of physical constraints on the basis of the traditional generative adversarial network, thereby ensuring that PI-GAN can achieve accurate repair of missing values ​​in accordance with the physical laws of the power system. Its structure is as follows: Figure 3 shown.

[0028] In one embodiment of the present application, the generative network includes an input layer, a fully connected layer, a batch normalization layer and an output layer, and the discriminative network includes an input layer, a fully connected layer and an output layer, wherein the batch normalization layer is used to standardize the input of each layer to maintain the stability of the input distribution.

[0029] In one embodiment of the present application, in the data enhancement process of the digital twin model of the power system in a low-measurement environment, the design of the generation network is particularly critical. The generation network in PI-GAN includes several layers for repairing missing values ​​from some known measurement data to obtain complete electrical measurement information. Specifically, the network parameters and architecture of the generation network in PI-GAN are shown in Table 1 below.

[0030]

[0031] Table 1 The generated network consists of four different types of network layers, namely the input layer, the fully connected layer, the batch normalization layer, and the output layer. The first fully connected layer contains 256 neurons, and the activation function is ReLU (Rectified Linear Unit). The ReLU activation function has the advantages of simple calculation and can alleviate the gradient vanishing problem. Its calculation formula is:

[0032] The second fully connected layer contains 128 neurons, and the activation function is Leaky ReLU. Compared with ReLU, LeakyReLU allows a small negative slope when the input is less than 0, to prevent neurons from "dying" when negative input occurs and affecting the update of gradients. The formula is:

[0033] in, It is a small constant, generally taking a smaller value, such as 0.05.

[0034] After each fully connected layer, a batch normalization layer is also added to standardize the input of each layer to keep the input distribution stable. Batch normalization standardizes the input of each batch to make its mean 0 and standard deviation 1, thereby accelerating the training process and enhancing the stability of the network. Its formula is:

[0035] in, and are the mean and standard deviation of the batch of data, is a small constant used to prevent the denominator from being zero.

[0036] Since the generative network is designed to repair missing samples, its input is the measurement samples with missing values, and its output is the complete repaired samples, so the number of neurons in the output layer is the same as that in the input layer. In addition, the output layer does not use a nonlinear activation function, but a linear activation to ensure the continuity of the generated data, that is,

[0037] The linear activation function ensures that the output value range is not limited to a specific interval, which is conducive to generating real numerical states.

[0038] In terms of the choice of optimizer, in order to accelerate training and improve convergence speed, the Adam optimizer (AdaptiveMoment Estimation) is used, which is one of the widely used optimizers in current deep learning. Adam combines momentum and adaptive learning rate adjustment to assign different learning rates to each parameter in different dimensions. Its update rule is:

[0039]

[0040]

[0041]

[0042] in, is the first-order momentum at the current time step, is the gradient of the current time step, is the second-order momentum at the current time step, is the square of the gradient at the current time step, and are the first-order and second-order momentum decay rates, usually taken as , , is the modified first-order momentum, is the modified second-order momentum, is the updated parameter, is a parameter, is the learning rate, is a decimal value that prevents division by zero, The first-order momentum decay rate of t Second power, is the second-order momentum decay rate of t Power.

[0043] Unlike traditional optimizers, the Adam optimizer exhibits strong robustness during the training process, and can adaptively adjust the update rate of parameters without manually adjusting the learning rate. This enables the generative network to be effectively trained in complex high-dimensional spaces and gradually converge to the global optimal solution, which is suitable for the generative network training process there.

[0044] The discriminant network in PI-GAN is designed to distinguish the true values ​​obtained by measurement in the sample and the repaired values ​​obtained by the generative network. Specifically, the discriminant network determines whether each value in the complete input data comes from the real measurement or is interpolated by the generative network through a binary classification task. The architecture design of the discriminant network is similar to that of the generative network, and also contains multiple fully connected layers to ensure that the potential features of the data can be effectively learned and accurately discriminated. The specific parameters are shown in Table 2 below.

[0045]

[0046] Table 2 First, the input layer receives the repaired data, and the input dimension is consistent with the generation network. The first fully connected layer contains 256 neurons, and the activation function is ReLU. The second fully connected layer contains 128 neurons, and the activation function is also ReLU. By reducing the number of neurons layer by layer, the discriminant network can extract high-order features of the data and then determine the authenticity of each value. The combination of two fully connected layers can effectively learn complex nonlinear relationships and help the discriminant network accurately distinguish generated data from real data in adversarial training.

[0047] The output layer activation function uses the Sigmoid function. The Sigmoid activation function can compress the output value to between 0 and 1, indicating the probability that each value in the input data is a true value or an interpolated value. Its formula is:

[0048] The Sigmoid function is very suitable for binary classification problems. In PI-GAN, the output layer gives the probability that each value in the input data comes from the real measurement. Therefore, choosing the Sigmoid function can naturally express the probability output in this binary classification problem, which helps the discriminant network learn a clearer judgment standard.

[0049] In terms of optimizer, the discriminant network also uses the Adam optimizer for parameter optimization to ensure its smooth update.

[0050] S207: Training the physical constraint generative adversarial network model based on the training data set, including discriminative network update and generative network update.

[0051] In the embodiment of the present application, the adversarial training process of the physical constraint generative adversarial network model PI-GAN can be divided into two main steps: discriminant network update and generative network update. The goal of this process is to enable the generative network G to fool the discriminative network D when generating missing data that conforms to the laws of physics through adversarial training, while the discriminative network improves its own discriminative ability through continuous learning.

[0052] In one embodiment of the present application, training the physical constraint generative adversarial network model based on the training data set includes: S301: Select a real sample with missing values ​​from the training data set, input it and a noise vector into a generation network, and obtain a repaired real sample.

[0053] S303: Determine a mask matrix based on the restored real sample, calculate the real loss and the restoration loss of the discriminant network based on the mask matrix, calculate the loss of the generator network based on the restored real sample and the mask matrix, minimize the real loss, the restoration loss of the discriminant network and the loss of the generator network, and update the parameters of the discriminant network and the generator network.

[0054] In one embodiment of the present application, first, in the discriminant network In the update step, the system randomly selects a real sample with missing values ​​from the training set , the data matrix contains some known real measurement data, and other positions are missing values. Then, a random noise vector z is randomly generated, and the noise vector and the real sample are input into the generation network , and then repair the real sample to get Next, based on Introducing a mask matrix , which is used to identify which of the restored samples are the true values ​​obtained by measurement and which are the restored values ​​obtained by the generative network. Specifically, the mask matrix Elements in The values ​​are:

[0055] The goal of the discriminant network is to target the repaired samples Each value in is classified to determine whether it is the original true value or the generated interpolation value, and the mask matrix , we can calculate the loss of the discriminant network for the true value and the interpolated value separately.

[0056] For the real value part obtained by measurement, in one embodiment of the present application, the calculation formula of the real loss is:

[0057] in, Represents the loss of the true value part, The value representing the position is the original true value or the generated network interpolation value. The value representing the position is the original true value, is the probability that the position output by the discriminant network is the true value, That is, the value at that position.

[0058] For the repair value part obtained by the generation network, in one embodiment of the present application, the calculation formula of the repair loss is:

[0059] in, Indicates the loss of the repair value part, The value representing the position is the original true value or the generated network interpolation value. The value representing the position is the generated network interpolation value, is the probability of the position value being the interpolated value output by the discriminant network, That is, the value at that position.

[0060] During the training process of PI-GAN, this application uniformly uses the Adam optimizer (Adaptive MomentEstimation) to update the parameters of the discriminant network to minimize the total loss. .

[0061] In the generative network In the update step, a real sample with missing values ​​is randomly selected from the training set. , and generate a new random noise vector from the noise distribution , the generative network receives the real sample and the noise vector, and repairs the new sample For the generator network, the goal is to generate realistic interpolated values ​​so that the discriminator network cannot distinguish between these interpolated values ​​and the real values.

[0062] In one embodiment of the present application, the loss of the generated network includes statistical loss and physical loss, wherein the calculation formula of the statistical loss is:

[0063] in, is the statistical loss term, The value representing the position is the generated network interpolation value, is the probability that the position value output by the discriminant network is the true value, That is, the value of the position; The calculation formula for physical loss is:

[0064] in, is the physical loss term, It is a physical loss function defined according to the physical constraints of the power system or component to ensure that the generated interpolation value conforms to the physical laws of the power system. is a random noise vector, To generate a complete sample after network repair.

[0065] Finally, the generated network The total training loss is:

[0066] in, is a parameter that controls the weight of physical loss.

[0067] S209: Input the data of the power system to be repaired into the trained physical constraint generation adversarial network model to obtain enhanced data.

[0068] In the embodiment of the present application, based on the collected measurement samples Perform data integrity check. Specifically, first, filter out meaningless data that is significantly deviated from the sample mean or the effective range of the measurement signal, then mark it as missing values, and finally record the missing value location to facilitate the next step of data repair. For the identified missing value location, the PI-GAN generation network take over With a random noise vector , and output a corresponding repair matrix , that is, generating network interpolation missing values ​​to achieve data repair, so the repaired samples can not only make full use of the original real data, but also restore the real characteristics of the measurement data based on the generation of network interpolation missing data, thus forming complete enhanced data. After completing the repair of the missing sample set of electrical quantities, the repaired sample set is integrated with the known system information to form a more comprehensive, reliable, and high-quality data set suitable for data-driven digital twin models. By combining a complete set of measurement samples, system component parameters, and system operating status, the digital twin model can obtain more complete and reliable environmental and historical information, thereby more accurately reflecting the real state of the power system and providing a solid foundation for subsequent applications.

[0069] In one embodiment of the present application, the method further includes: The mean square error, mean absolute error, structural similarity and physical consistency loss are used to evaluate the restoration effect of the physical constraint generative adversarial network model.

[0070] In one embodiment of the present application, in order to comprehensively evaluate the effect of PI-GAN on missing data repair, a multi-dimensional evaluation index system is established to measure the quality of generated data from different angles to ensure that the repaired data is not only statistically close to the true value distribution, but also conforms to the physical constraints of the power system. In order to comprehensively evaluate the repair effect from multiple dimensions, evaluation indicators including mean square error, mean absolute error, structural similarity and physical consistency loss are introduced.

[0071] Specifically, Mean Squared Error (MSE) is one of the most commonly used error metrics, which is used to measure the numerical deviation between generated data and real data. MSE reflects the overall error level by calculating the square difference between the generated value and the real value. Its formula is as follows:

[0072] in, Is the real data at location The value at Is to generate data at location The value at is the number of all non-missing values ​​in the sample. The smaller the value of MSE, the smaller the gap between the generated data and the real data, and the better the repair effect of the generated network. Due to the existence of the square term, MSE is more sensitive to extreme values ​​and easily amplifies the impact of large errors. Therefore, MSE is used to evaluate the numerical repair accuracy of different repair methods.

[0073] Mean Absolute Error (MAE) is another commonly used error metric, which reflects the magnitude of the error by calculating the absolute difference between the generated value and the true value. Its formula is:

[0074] Unlike MSE, MAE does not square the difference, so it is not sensitive to extreme values. MAE reflects the difference in absolute values ​​between the generated data and the real data. A smaller MAE value means that the error between the generated data and the real data is smaller, that is, the repair accuracy is better.

[0075] Structural Similarity Index (SSIM) is an important indicator for the structural similarity of data. Unlike MSE and MAE, SSIM not only considers the difference in numerical values, but also considers information such as the local structure of the data, which can more comprehensively evaluate the quality of generated data. The calculation formula of SSIM is:

[0076] in, and are the means of the real data and the generated data, respectively. and are the standard deviations of the real data and the generated data, respectively, is the covariance of the two, and It is a constant introduced to stabilize the calculation and is usually a small value.

[0077] The value range of SSIM is 0 to 1. The closer the value is to 1, the higher the structural similarity between the generated data and the real data. This indicator is particularly suitable for evaluating whether the local structure of the data is consistent, and helps to determine the consistency of the data structure between the repaired sample and the real measured sample.

[0078] In the power system, the repaired data must not only be close to the real data in terms of value and structure, but also must meet the physical constraints of the power system. Therefore, physical consistency loss is proposed to evaluate the degree of compliance of the generated data with the physical laws of the power system or components, thereby ensuring the physical rationality of the generated data.

[0079] The calculation of physical consistency loss depends on the specific power system equations. For example, for the power conservation and voltage ratio relationship of the transformer, the following physical loss function can be defined:

[0080] in, is the generated transformer primary voltage, is the generated secondary voltage, and are the currents generated on the primary and secondary sides of the transformer, and is a weight parameter that controls the influence of the two constraints on the total loss, and are the number of coil turns on the primary and secondary sides of the transformer respectively.

[0081] In the above-mentioned data enhancement method for the digital twin model of the power system in a low-measurement environment, first, the initial data of the power system is obtained and preprocessed to obtain a training data set; then, physical constraints are established; then, a physical constraint generative adversarial network model is constructed based on the physical constraints, and the physical constraint generative adversarial network model includes a generative network and a discriminant network, wherein the generative network is used to repair data, and the discriminant network is used to discriminate the authenticity of the repaired data; then, the physical constraint generative adversarial network model is trained based on the training data set, including discriminant network update and generative network update; finally, the power system data to be repaired is input into the trained physical constraint generative adversarial network model to obtain enhanced data. That is to say, through the introduction of PI-GAN, physical information is embedded in the data generation process of the generative adversarial network, ensuring that the generated data is not only consistent with the real data in a statistical sense, but also can meet the physical characteristics requirements of the power system. In this way, the generated data can not only interpolate the missing data due to sensor failure and other reasons, but also maintain accurate estimation and dynamic simulation capabilities of the system state in a noisy environment. In practical applications, the operating conditions of power systems are complex and changeable. Traditional neural network-based methods often find it difficult to maintain consistent performance under different working conditions. By combining the advantages of physical information and generative adversarial networks, the digital twin model constructed in this application can adapt to different system structures and operating conditions, maintain good predictive capabilities and adaptability, and further enhance its application value in actual power systems. The physical consistency and reliability of data are enhanced, and the accuracy, stability and generalization ability of the digital twin model are improved, thereby providing reliable technical support for power system status monitoring, operation optimization and fault prevention.

[0082] In one embodiment of the present application, Figure 4 The figure shows a technical framework diagram of a data enhancement method for a power system digital twin model in a low-measurement environment. It mainly includes data preparation, model training and data enhancement. The data preparation includes data collection and preprocessing, physical model construction and physical constraint arrangement, the model training includes PI-GAN generation network construction, PI-GAN discrimination network construction and PI-GAN adversarial training, and the data enhancement includes missing data identification and repair and enhanced data set construction.

[0083] The following is a specific implementation example to illustrate the specific implementation steps of a method for data enhancement of a digital twin model of a power system in a low-measurement environment of the present application. First, S401, the initial data of the power system is obtained and preprocessed to obtain a training data set. Then, S403, physical constraints are established. Then, S405, a physical constraint generative adversarial network model is constructed based on the physical constraints, and the physical constraint generative adversarial network model includes a generative network and a discriminative network, wherein the generative network is used to repair data, and the discriminative network is used to discriminate the authenticity of the repaired data. Specifically, S407, the generative network includes an input layer, a fully connected layer, a batch normalization layer and an output layer, and the discriminative network includes an input layer, a fully connected layer and an output layer, wherein the batch normalization layer is used to standardize the input of each layer to maintain the stability of the input distribution.

[0084] Afterwards, S409, the physical constraint generative adversarial network model is trained based on the training data set, including updating the discriminant network and the generative network. Specifically, S411-S413, a real sample with missing values ​​is selected from the training data set, and it is input into the generative network with a noise vector to obtain a repaired real sample; a mask matrix is ​​determined based on the repaired real sample, the real loss and repair loss of the discriminant network are calculated based on the mask matrix, the loss of the generative network is calculated based on the repaired real sample and the mask matrix, the real loss, repair loss and loss of the discriminant network and the loss of the generative network are minimized, and the parameters of the discriminant network and the generative network are updated.

[0085] S415, the calculation formula of the actual loss is:

[0086] in, Represents the loss of the true value part, The value representing the position is the original true value or the generated network interpolation value. The value representing the position is the original true value, is the probability that the position output by the discriminant network is the true value, That is, the value at that position.

[0087] S417, the calculation formula of the repair loss is:

[0088] in, Indicates the loss of the repair value part, The value representing the position is the original true value or the generated network interpolation value. The value representing the position is the generated network interpolation value, is the probability of the position value being the interpolated value output by the discriminant network, That is, the value at that position.

[0089] S419, the loss of the generated network includes statistical loss and physical loss, wherein the calculation formula of the statistical loss is:

[0090] in, is the statistical loss term, The value representing the position is the generated network interpolation value, is the probability that the position value output by the discriminant network is the true value, That is, the value of the position; The calculation formula for physical loss is:

[0091] in, is the physical loss term, It is a physical loss function defined according to the physical constraints of the power system or component. is a random noise vector, To generate a complete sample after network repair.

[0092] S421, mean square error, mean absolute error, structural similarity and physical consistency loss are used to evaluate the repair effect of the physical constraint generative adversarial network model. Finally, S423, the power system data to be repaired is input into the trained physical constraint generative adversarial network model to obtain enhanced data.

[0093] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0094] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 5As shown. The computer device includes a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, a mobile cellular network, NFC (near field communication) or other technologies. When the computer program is executed by the processor, a data enhancement method for a digital twin model of a power system in a low measurement environment is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covered on the display screen, or a key, trackball or touchpad set on the housing of the computer device, or an external keyboard, touchpad or mouse.

[0095] Those skilled in the art will understand that Figure 5 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0096] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0097] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but are not limited to this.

[0098] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0099] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A method for data enhancement of a digital twin model of a power system in a low-measurement environment, characterized in that: The method comprises: Obtain the initial data of the power system and preprocess it to obtain a training data set; Establish physical constraints; Based on the physical constraint conditions, a physical constraint generative adversarial network model is constructed, wherein the physical constraint generative adversarial network model includes a generative network and a discriminative network, wherein the generative network is used to repair data, and the discriminative network is used to discriminate the authenticity of the repaired data; Training the physical constraint generative adversarial network model based on the training data set, including discriminative network update and generative network update; The data of the power system to be repaired is input into the trained physical constraint generative adversarial network model to obtain enhanced data.

2. According to a method for enhancing data of a digital twin model of a power system in a low-measurement environment according to claim 1, it is characterized in that: The generative network includes an input layer, a fully connected layer, a batch normalization layer and an output layer, and the discriminative network includes an input layer, a fully connected layer and an output layer, wherein the batch normalization layer is used to standardize the input of each layer to maintain the stability of the input distribution.

3. According to a method for enhancing data of a digital twin model of a power system in a low-measurement environment according to claim 1, it is characterized in that: The training of the physical constraint generative adversarial network model based on the training data set includes: Selecting a real sample with missing values ​​from the training data set, inputting the real sample and the noise vector into the generation network to obtain a repaired real sample; A mask matrix is ​​determined based on the repaired real sample, the real loss and the repair loss of the discriminant network are calculated based on the mask matrix, the loss of the generator network is calculated based on the repaired real sample and the mask matrix, the real loss, the repair loss and the loss of the discriminant network and the generator network are minimized, and the parameters of the discriminant network and the generator network are updated.

4. According to a method for enhancing data of a digital twin model of a power system in a low-measurement environment according to claim 3, it is characterized in that: The calculation formula of the actual loss is: in, Represents the loss of the true value part, The value representing the position is the original true value or the generated network interpolation value. The value representing the position is the original true value, is the probability that the position output by the discriminant network is the true value, That is, the value at that position.

5. According to a method for enhancing data of a digital twin model of a power system in a low-measurement environment according to claim 3, it is characterized in that: The calculation formula of the repair loss is: in, Indicates the loss of the repair value part, The value representing the position is the original true value or the generated network interpolation value. The value representing the position is the generated network interpolation value, is the probability of the position value being the interpolated value output by the discriminant network, That is, the value at that position.

6. According to a method for enhancing data of a digital twin model of a power system in a low-measurement environment according to claim 3, it is characterized in that: The loss of the generated network includes statistical loss and physical loss, wherein the calculation formula of the statistical loss is: in, is the statistical loss term, The value representing the position is the generated network interpolation value, is the probability that the position value output by the discriminant network is the true value, That is, the value of the position; The calculation formula for physical loss is: in, is the physical loss term, It is a physical loss function defined according to the physical constraints of the power system or component. is a random noise vector, To generate a complete sample after network repair.

7. The method for enhancing data of a digital twin model of a power system in a low-measurement environment according to claim 1, characterized in that: The method further comprises: The mean square error, mean absolute error, structural similarity and physical consistency loss are used to evaluate the restoration effect of the physical constraint generative adversarial network model.

Images