Load impact type identification method and system for distributed power grid

Abstract

The invention discloses a distributed power grid-oriented load impact type identification method and system. The method comprises the steps of obtaining an initial training set of a target power system; performing sample expansion on the gradient update quantity initial training set by adopting a preset generative adversarial network introduced with a local curvature sensing mechanism to obtain an expanded training set; wherein the gradient update quantity expansion training set comprises a plurality of power data samples; expanding the training set through the gradient update quantity, training the initial model, and obtaining a target identification model; wherein the gradient update quantity target identification model is used for identifying the to-be-detected power data and outputting a corresponding load impact type. According to the embodiment of the invention, by introducing the preset generative adversarial network of the local curvature sensing mechanism, the generated power data sample of which the geometric structure is close to the real power data is obtained, the expansion of the training sample can be realized, and the identification accuracy and generalization ability of the target identification model on the load impact type are improved.

Description

Technical Field

[0001] The present invention relates to the fields of power grid fault identification and artificial intelligence technology, and in particular to a load impact type identification method and system for distributed power grids. Background Art

[0002] With the widespread integration of distributed energy resources and the increasing diversification of electricity loads, the operating environment of the power grid has become more complex and dynamic. In this context, load shocks of various sizes can occur within the grid. These shocks can be caused by a variety of factors, such as the startup and shutdown of large-scale equipment, load fluctuations caused by changing weather conditions, or failures of internal grid equipment. These load shocks not only pose challenges to the stability, reliability, and security of the grid, but can also trigger a series of problems, such as voltage fluctuations, frequency deviations, and even localized power outages. Therefore, effectively identifying the types of load shocks in distributed power grids has become a pressing issue. Summary of the Invention

[0003] The purpose of the embodiments of the present invention is to provide a load shock type identification method and system for distributed power grids. By introducing a preset generative adversarial network with a local curvature perception mechanism, generated power data samples with a geometric structure close to real power data are obtained, which can achieve the expansion of training samples and thereby improve the recognition accuracy and generalization ability of the target identification model for load shock types.

[0004] A first embodiment of the present invention provides a method for identifying load impact types in a distributed power grid, comprising: Obtain an initial training set of the target power system; Using a preset generative adversarial network that introduces a local curvature perception mechanism, the initial training set is expanded to obtain an expanded training set; wherein the expanded training set includes a plurality of power data samples; The initial model is trained by using the expanded training set to obtain a target recognition model; wherein the target recognition model is used to identify the power data to be detected and output the corresponding load impact type.

[0005] Optionally, the generator loss in the preset generative adversarial network includes: adversarial loss, multi-scale local curvature loss and high-order curvature alignment loss.

[0006] Optionally, the multi-scale local curvature loss is calculated using the following formula: ; in, is the multi-scale local curvature loss function; For the generator; The set of real power data samples input to the preset generative adversarial network for the current batch; is the second derivative of the th generated power data sample at the th scale; is the second derivative of the th real power data sample at the th scale; is the third derivative of the th generated power data sample; is the third derivative of the th real power data sample; is the th generated power data sample; is the th real power data sample; is the first features of the th generated power data sample; is the first features of the th real power data sample; is the curvature loss balance coefficient; is the total number of scales; is the number of samples input to the preset generative adversarial network for the current batch; is a positive integer; is a positive integer and ; is the L2 norm.

[0007] Optionally, the generator loss is calculated by the following formula: ; where is the loss function of the generator; is the set of real power data samples input to the preset generative adversarial network for the current batch; is the set of latent space vectors corresponding to the set of real power data samples; is the generator function; is the discriminator function; is the latent variable interaction term; represents the expectation when the generator input is ; is the multi-scale local curvature loss function; is the high-order curvature term of the th generated power data sample; is the high-order curvature term of the th real power data sample; is the weight coefficient of the first generator loss; is the weight coefficient of the second generator loss; is the number of samples in the current batch input to the preset generative adversarial network.

[0008] Optionally, the method further includes: During the training of the initial model, calculating the gradient update amount of the current weight matrix according to the augmented training set; wherein, the gradient update amount is at least composed of the following three components: an interference cancellation term, a perturbation term, and a non-linear adaptive adjustment term; the interference cancellation term is used to correct the gradient update direction according to the historical gradient vector of the power data sample; the perturbation term is used to adjust the gradient update direction according to the deviation degree between the power data sample and the sample mean vector; the non-linear adaptive adjustment term is used to adjust the gradient update direction according to the non-linear relationship in the power data sample.

[0009] Optionally, the interference cancellation term is calculated by the following formula: ; wherein, is the interference cancellation term; is the learning rate adjustment coefficient of the initial model; is the model output corresponding to the i-th power data sample; represents the gradient update direction of the i-th power data sample; [[ID=2,6]] is the historical gradient vector of the i-th power data sample; is the number of samples in the current batch input to the initial model for training.

[0010] Optionally, the perturbation term is calculated by the following formula: ; wherein, is the perturbation term; is the i-th power data sample; is the sample mean vector; is the model output corresponding to the i-th power data sample; represents the gradient update direction of the i-th power data sample; is the number of samples in the current batch input to the initial model for training.

[0011] Optionally, the non-linear adaptive adjustment term is obtained through the following steps: Performing high-order feature expansion on the power data sample to obtain a corresponding non-linear feature vector; wherein, the non-linear feature vector includes: the original features of the power data sample, the square terms of the original features, and the combined cross terms between the original features. Calculate the response intensity of the current weight matrix to the non - linear feature vector to obtain a non - linear adjustment factor; wherein, the non - linear adjustment factor is used to characterize the adaptability of the initial model to the non - linear relationship; Adjust the gradient update direction based on the non - linear adjustment factor and the influence coefficient factor to obtain the non - linear adaptive adjustment term.

[0012] Optionally, the method further includes: Update the current weight matrix based on the gradient update amount to obtain a transitional weight matrix; Perform smooth correction on the transitional weight matrix based on the difference between the weight matrices of adjacent iteration periods and the viscosity factor to obtain an updated weight matrix.

[0013] An embodiment of the second aspect of the present invention provides a load impact type identification system for a distributed power grid, including: An initial training set acquisition module, configured to acquire an initial training set of a target power system; An augmented training set acquisition module, configured to use a preset generative adversarial network introducing a local curvature perception mechanism to perform sample augmentation on the initial training set to obtain an augmented training set; wherein, the augmented training set includes a number of power data samples; A target identification model construction module, configured to train an initial model through the augmented training set to obtain a target identification model; wherein, the target identification model is used to identify the power data to be detected and output the corresponding load impact type.

[0014] Compared with the prior art, an embodiment of the present invention provides a load impact type identification method and system for a distributed power grid. The method includes: acquiring an initial training set of a target power system; using a preset generative adversarial network introducing a local curvature perception mechanism to perform sample augmentation on the initial training set to obtain an augmented training set; wherein, the augmented training set includes a number of power data samples; training an initial model through the augmented training set to obtain a target identification model; wherein, the target identification model is used to identify the power data to be detected and output the corresponding load impact type. By introducing a preset generative adversarial network with a local curvature perception mechanism, the embodiment of the present invention can obtain generated power data samples with a geometric structure approximating real power data, realize the augmentation of training samples, and thus improve the recognition accuracy and generalization ability of the target identification model for load impact types. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a flowchart of an embodiment of the load impact type identification method for a distributed power grid provided by the present invention; Figure 2 It is a schematic flow chart of an embodiment for training a preset generative adversarial network provided by the present invention; Figure 3 It is a schematic structural diagram of an embodiment of a load impact type identification system for a distributed power grid provided by the present invention. Detailed implementation manners

[0016] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the technical field of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0017] See Figure 1 , which is a schematic flow chart of an embodiment of a load impact type identification method for a distributed power grid provided by the present invention.

[0018] An embodiment of the first aspect of the present invention provides a load impact type identification method for a distributed power grid, including steps S1 to S3, specifically as follows: Step S1: Obtain an initial training set of a target power system; Step S2: Use a preset generative adversarial network introducing a local curvature perception mechanism to perform sample expansion on the initial training set to obtain an expanded training set; wherein, the expanded training set includes a number of power data samples; Step S3: Train an initial model through the expanded training set to obtain a target identification model; wherein, the target identification model is used to identify the power data to be detected and output the corresponding load impact type.

[0019] It should be noted that in traditional methods, the collection of training data is often restricted by factors such as time, space, and equipment, which not only results in insufficient sample quantity but also may lead to uneven data distribution, thus having an adverse impact on the generalization ability and classification accuracy of the model.

[0020] The embodiment of the present invention introduces a local curvature perception mechanism into a generative adversarial network (GAN), and by calculating the curvature difference between the generated power data and the real power data, ensures that the geometric structure of the generated power data approximates the distribution characteristics of the real power data (such as voltage fluctuation trend, current mutation pattern, meteorological fluctuation trend), and avoids generating "pseudo power data".

[0021] In addition, in the embodiments of the present invention, the preset generative adversarial network is used to expand the samples, which can significantly increase the scale of the training data, enabling the model to learn the characteristics of various load shocks more robustly and comprehensively during training, thereby accurately identifying four types: normal fluctuations, minor shocks, medium shocks, and severe shocks, and reducing the risk of misclassification caused by insufficient samples.

[0022] In step S1, in the embodiments of the present invention, the multi-source monitoring data of the target power system is collected and preprocessed to construct an initial training set. In other words, by collecting data such as current, voltage, equipment status, and meteorological parameters in the target power system, and through operations such as data cleaning, feature extraction, discretization processing, and expert annotation, an initial training set is constructed.

[0023] The data collected in the embodiments of the present invention has a wide range of sources, covering various real-time monitoring devices in the power system, including but not limited to current sensors, voltage sensors, load monitoring devices, equipment status monitoring systems, and external meteorological data sources. The following are the main methods and sources of data collection: (1) Real-time monitoring: Through sensors in power substations and distribution networks, numerical values such as current, voltage, and load fluctuations are collected in real time and transmitted to the data collection system. Each collection device collects data at a fixed time interval, and the data period can be flexibly adjusted according to requirements, usually from 1 second to 10 seconds.

[0024] (2) Equipment status information: With the help of the real-time status monitoring system of the equipment, the switch status, operation mode, and fault information of various equipment (such as transformers, circuit breakers, etc.) are obtained.

[0025] (3) Meteorological data: By accessing meteorological monitoring stations, external meteorological data such as local temperature, humidity, wind speed, and precipitation are obtained; since these information are related to the load fluctuations of the power grid to a certain extent, they need to be collected synchronously.

[0026] All the above collected data are uniformly stored in the data warehouse of the cloud platform and marked with time stamps to ensure that the original data has time series attributes. The data is stored in a structured JSON format to improve the efficiency of subsequent processing and parsing.

[0027] The collected raw data needs to go through multiple steps of preprocessing to ensure its quality and applicability. The main processing flow is as follows: (1) Data cleaning and outlier detection: There may be noise, missing values, or outliers in the raw data, which will have an adverse impact on subsequent analysis and modeling. First, abnormal data is detected through statistical analysis methods (such as the Z-score method), and then the raw data is cleaned by using interpolation methods or directly deleting outliers.

[0028] (2)Feature Transformation and Data Standardization: Perform feature transformation and standardization on the time series data such as the collected current, voltage, load, and meteorological parameters. Specifically, standardize the values of current, voltage, etc. to ensure that different features are calculated on the same scale. In addition, use the sliding window technique to extract statistical features such as the maximum value, minimum value, mean value, and variance within different time windows.

[0029] (3)Discretization and Feature Construction: For time series data, transform it into discrete features through discretization methods. For example, through discretization methods, divide the current values into different intervals (such as 0–10A, 10–20A, 20–30A, etc.), and generate corresponding discrete feature values according to the interval to which the current belongs.

[0030] After completing the data preprocessing, label the collected training data. Specifically, based on features such as the amplitude of current and voltage fluctuations and the rate of load change, experts manually label the type of load impact corresponding to the data at each moment. The types of load impacts include the following four: normal fluctuation, minor impact, medium impact, and severe impact.

[0031] It can be understood that the collection, labeling, and preprocessing of training data are time-consuming and laborious, and the lack of training samples (i.e., power data samples) is likely to lead to poor generalization ability of the model and affect the accuracy of the model.

[0032] Aiming at the deficiencies in the scale, distribution, and diversity of power data samples, the embodiment of the present invention uses a preset generative adversarial network based on local curvature perception for sample generation, and then realizes the expansion of power data samples to obtain an expanded training set; in the expanded training set, the generated power data samples and the real power data samples can be collectively referred to as power data samples. Specifically, by adopting a local curvature perception mechanism and multi-scale curvature modeling, optimize the geometric structure of the generated power data samples, and improve the authenticity, diversity, and adaptability to complex power data structures of the generated power data samples.

[0033] The embodiment of the present invention not only effectively expands the scale of power data samples, alleviates the problem of insufficient samples, but also improves the diversity and distribution balance of samples in the expanded training set, providing a solid data basis for the subsequent training and optimization of the model.

[0034] In an optional embodiment, the generator loss in the preset generative adversarial network includes: adversarial loss, multi-scale local curvature loss, and high-order curvature alignment loss.

[0035] Further, the multi-scale local curvature loss is calculated by the following formula: ; Where is a multi-scale local curvature loss function; is the generator; is the set of real power data samples input to the preset generative adversarial network in the current batch; is the second derivative of the th generated power data sample at the is the second derivative of the th real power data sample at the is the third derivative of the is the third derivative of the is the th generated power data sample; is the th real power data sample; is the first features of the th generated power data sample; is the first features of the is the curvature loss balance coefficient; is the total number of scales; is the number of samples input to the preset generative adversarial network in the current batch; is a positive integer; ; is the L2 norm.

[0036] Furthermore, the generator loss is calculated by the following formula: ; where, is the loss function of the generator; is the set of real power data samples input to the preset generative adversarial network in the current batch; is the set of latent space vectors corresponding to the set of real power data samples; is the generator function; is the discriminator function; is the latent variable interaction term; represents the expectation when the generator input is ; is the multi-scale local curvature loss function; is the high-order curvature term of the th generated power data sample; is the The high-order curvature terms of a real power data sample; is the loss weight coefficient of the first generator; is the loss weight coefficient of the second generator; is the number of samples input to the preset generative adversarial network in the current batch.

[0037] It should be noted that, as Figure 2 shown, it is a schematic flowchart of an embodiment for training the preset generative adversarial network provided by the present invention. In Figure 2 , first, the generator and the discriminator are initialized, and then it enters the iterative training stage. In each iteration, the generator generates data, and the discriminator determines the authenticity of the data; then, the geometric structure difference of the generated data is evaluated through the multi-scale local curvature loss, and the geometric structure of the generated data is optimized accordingly. Subsequently, the losses of the generator and the discriminator are calculated respectively; among them, the loss of the generator combines the discrimination result and the geometric structure optimization objective, while the loss of the discriminator is based on its classification ability for real and fake data. Using these losses, the parameters of the generator and the discriminator are updated through backpropagation. The training process continues to loop until the maximum number of iterations is reached, and finally, the training of the preset generative adversarial network (i.e., the power data augmentation model) is completed. The above training process not only covers the basic training steps of the GAN but also introduces multi-scale geometric structure optimization to improve the authenticity of the generated power data samples.

[0038] Specifically, in the training process of the preset generative adversarial network introducing the local curvature perception mechanism, the main steps are as follows: Step 1: According to the framework of the generative adversarial network, the generator adopts a multi-layer neural network structure, with the input being a randomly sampled latent space vector and the feature vector of the original power data (i.e., the real power data sample), and the output being the generated power data (i.e., the generated power data sample); the discriminator is constructed based on a multi-layer convolutional network and is used to output the authenticity probability after inputting the power data.

[0039] Let the generator be , the input of the generator is (i.e., the set of latent space vectors corresponding to the real power data sample set), the real power data is (i.e., the set of real power data samples input to the preset generative adversarial network in the current batch), the generated power data output by the generator is (i.e., the set of generated power data samples output by the generator in the current batch), the discriminator is , the input of the discriminator is or ; the output probability of the discriminator is , and , which represents the probability that the discriminator determines the power data to be real; the weight parameters of the generator and the discriminator are and , and are initialized by random assignment.

[0040] Step 2: Optimize the geometric structure of the generated power data through a local curvature perception mechanism; among them, the local curvature perception mechanism means that during the training process of the generator, calculate the curvature difference between the generated power data and the real power data at different scales, and capture complex geometric features by combining high-order derivative information to constrain the output of the generator to approximate the geometric distribution of the real power data in the local area. At this stage, optimize the local curvature perception mechanism by calculating the multi-scale local curvature loss function, which is expressed as: ; where is the multi-scale local curvature loss function; is the generator; is the set of real power data samples input to the preset generative adversarial network in the current batch; is the generator outputs the second-order derivative of the th generated power data sample at the th scale, which is used to characterize the local curvature of the th generated power data sample; is the second-order derivative of the th real power data sample at the th scale, which is used to characterize the local curvature of the ] th real power data sample; is the third-order derivative of the th generated power data sample output by the generator, which is used to characterize the high-order geometric change of the th generated power data sample; is the third-order derivative of the th real power data sample, which is used to characterize the high-order geometric change of the th real power data sample; is the th generated power data sample; is the th real power data sample; is the first th features of the th generated power data sample; is the first th features of the th real power data sample; is the curvature loss balance coefficient; ​is the total number of scales; is the number of samples input to the preset generative adversarial network in the current batch; is a positive integer; is a positive integer and ; is the L2 norm. Preferably, is set to 0.2.

[0041] It should be noted that the total number of scales refers to the local geometric structure modeling range of power data samples at different time windows and feature combination levels. For example, =1 means that within a short time window (such as 1 second), basic features are used for curvature calculation to capture micro mutations; as increases, the time window length and feature combination complexity are successively expanded to model the trend changes and high-order structure features at the meso and macro levels, so as to achieve accurate simulation and structure preservation of load disturbance behaviors at different levels.

[0042] Step 3: Enhance the generator's ability to capture complex feature relationships through latent variable interaction modeling, perform non-linear mapping on the latent space vector and the input features, enhance the deep interaction relationship between the two, and improve the diversity of the generated power data. The calculation method of the latent variable interaction term is expressed as: ; In the formula, is the latent variable interaction term, is the ReLU activation function, indicating the use of the ReLU activation function for non-linear feature extraction; is the Sigmoid activation function; is the weight parameter of the generator; is the bias parameter of the generator; ⊙ is the element-wise multiplication; is the input of the generator and the real power data vector concatenation operation; is the weight coefficient of the polynomial term; are the first k values of the generator input; are the first k values of the real power data. Preferably, is set to 0.2.

[0043] Furthermore, based on the latent variable interaction term, the final output of the generator is expressed as: ; In the formula, ⊙ is the element-wise multiplication; is the generator function.

[0044] Step 4: Based on the traditional adversarial loss, construct the loss function of the generator using the multi-scale local curvature loss function to balance the optimization objectives of power data distribution and local structure. The calculation method is expressed as: ; where is the loss function of the generator, which constrains the training process of the generator; is the set of real power data samples input to the preset generative adversarial network in the current batch; is the set of latent space vectors corresponding to the set of real power data samples; is the generator function; is the discriminator function; is the latent variable interaction term; denotes expectation; denotes the expectation when the input of the generator is ; is the multi-scale local curvature loss function; is the higher-order curvature term of the th generated power data sample output by the generator, expressed as ; is the third derivative of the th generated power data sample, is the second derivative of the th generated power data sample. When calculating the second derivative and third derivative of the generated power data sample, the derivatives of the values of each feature dimension are calculated separately and then accumulated; is the higher-order curvature term of the th real power data sample, and the calculation method is similar to that of the higher-order curvature term of the generated power data sample, i.e., ; is the first generator loss weight coefficient; is the second generator loss weight coefficient; is the number of samples input to the preset generative adversarial network in the current batch.

[0045] Step 5: The training process of the discriminator is constrained by the loss function of the discriminator to enhance the sensitivity to the subtle differences in the generated power data. The calculation method is expressed as: ; In the formula, is the loss function of the discriminator; denotes expectation, denotes the expectation when the input is ; denotes the expectation when the input is .

[0046] Step 6: Update the parameters of the generator and discriminator using an alternating iteration strategy. The generator minimizes the total loss through gradient descent, and the discriminator maximizes the adversarial loss through gradient ascent. The calculation method is expressed as follows: Update of generator parameters: ; Update of discriminator parameters: ; In the formula, is the learning rate of the preset generative adversarial network; is the gradient of the loss function of the generator with respect to ; is the gradient of the loss function of the discriminator with respect to ; and [[ID=2�]] are the weight parameters of the generator and discriminator respectively; represents the parameter update operation. Preferably, is set to 0.001.

[0047] Step 7: Repeat the above steps until the preset stop iteration condition is met, which indicates that the training of the preset generative adversarial network (i.e., the power data augmentation model) is completed. Usually, the stop iteration condition can be set to reach the preset maximum number of iterations. In the embodiments of the present invention, the preferred maximum number of iterations can be set to 1000 times to ensure that the model converges sufficiently and has good generalization ability.

[0048] Step 8: After the training of the power data augmentation model is completed, use the trained model to augment the number of samples. For example, if the original number of collected samples is 800, and 200 new samples are generated through the power data augmentation model, then the augmented power data sample set (i.e., the augmented training set) will contain 1000 samples.

[0049] In an alternative embodiment, the method further includes: During the training of the initial model, calculate the gradient update amount of the current weight matrix according to the augmented training set; wherein, the gradient update amount is at least composed of the following three components: interference cancellation term, perturbation term, and non-linear adaptive adjustment term; the interference cancellation term is used to correct the gradient update direction according to the historical gradient vector of the power data sample; the perturbation term is used to adjust the gradient update direction according to the deviation degree between the power data sample and the sample mean vector; the non-linear adaptive adjustment term is used to adjust the gradient update direction according to the non-linear relationship in the power data sample.

[0050] Further, the interference cancellation term is calculated by the following formula: ; Wherein, is the interference cancellation term; is the learning rate adjustment coefficient of the initial model; is the model output corresponding to the i-th power data sample; represents the gradient update direction of the i-th power data sample; is the historical gradient vector of the i-th power data sample; is the number of samples input to the initial model for training in the current batch.

[0051] Further, the perturbation term is calculated by the following formula: ; wherein, is the perturbation term; is the i-th power data sample; is the sample mean vector; is the model output corresponding to the i-th power data sample; represents the gradient update direction of the i-th power data sample, is the number of samples input to the initial model for training in the current batch.

[0052] Further, the non-linear adaptive adjustment term is obtained through the following steps: Perform high-order feature expansion on the power data sample to obtain a corresponding non-linear feature vector; wherein, the non-linear feature vector includes: the original features of the power data sample, the square terms of the original features, and the combined cross terms between the original features; Calculate the response intensity of the current weight matrix to the non-linear feature vector to obtain a non-linear adjustment factor; wherein, the non-linear adjustment factor is used to characterize the adaptability of the initial model to the non-linear relationship; Based on the non-linear adjustment factor and the influence coefficient factor, adjust the current gradient update direction to obtain the non-linear adaptive adjustment term.

[0053] Further, the method further includes: Update the current weight matrix based on the gradient update amount to obtain a transition weight matrix; Based on the difference between the weight matrices of adjacent iteration cycles and the viscosity factor, perform smooth correction on the transition weight matrix to obtain an updated weight matrix.

[0054] It should be noted that after obtaining the augmented training set, the embodiments of the present invention will perform model training on the initial model based on the augmented training set to obtain the final target identification model. The selection of the initial model has high flexibility, including but not limited to neural network models, support vector machines, and extreme learning machines. Preferably, the initial model is set as an extreme learning machine. In addition, the embodiments of the present invention combine the augmented training set and the gradient update optimization algorithm, so that the target identification model can significantly improve the accuracy and robustness in the high-dimensional power data classification task.

[0055] Next, taking the initial model set as an extreme learning machine (ELM) as an example, the training process of obtaining the target identification model will be described in detail.

[0056] It should be noted that in order to solve a series of problems in high-dimensional power data classification, such as noise interference, redundant features, and inconsistent feature scales, the embodiments of the present invention optimize the extreme learning machine algorithm to improve the accuracy of the classifier (i.e., the target identification model) and accelerate the training process. Specifically, by introducing an interference cancellation term, a perturbation term, and a non-linear adaptive adjustment term, the gradient update direction is dynamically adjusted to reduce the influence of noise and redundant features.

[0057] In the training process of the extreme learning machine in the embodiments of the present invention, an interference cancellation mechanism based on the gradient direction is introduced; among them, the interference cancellation mechanism based on the gradient direction first calculates the gradient value of each training sample and adjusts the weights of the hidden layer according to the gradient direction.

[0058] In the forward propagation process of the extreme learning machine, the activation method is expressed as: ; In the formula, is the Sigmoid activation function; is the weight matrix of the hidden layer of the extreme learning machine, is the bias term of the extreme learning machine; is the output of the extreme learning machine (i.e., the output set of the extreme learning machine in the current batch); is the training sample input to the extreme learning machine (i.e., the power data sample set input to the extreme learning machine in the current batch). In other words, in the current batch training, each power data sample passes through the same and , and obtains the corresponding extreme learning machine output .

[0059] The embodiments of the present invention correct the gradient update direction through the interference cancellation mechanism, specifically by calculating the interference cancellation term. The calculation method is based on the weighted sum of the absolute value of the current gradient and the historical gradient vector to reduce the gradient interference between samples, so as to stabilize the training direction. Specifically as follows: ; In the formula, is the interference cancellation term; is the learning rate adjustment coefficient of the extreme learning machine; represents the loss function of the extreme learning machine with respect to gradient; is the parameter of the extreme learning machine, including the weight matrix and the bias term; is the loss function of the extreme learning machine; is the historical gradient vector used for gradient direction adjustment, representing the accumulation of the historical gradient influence of the samples.

[0060] It should be noted that through the chain rule, we can obtain , and the gradient direction is the batch gradient update direction, reflecting the fastest descent direction of the loss function L under the current parameter θ in the current training batch. represents the set of historical gradient information of all samples in the current training batch, which is an aggregated representation used to adjust the current gradient direction.

[0061] To further illustrate the gradient of the loss function with respect to , and the acquisition method of the historical gradient vector after gradient adjustment, taking the sample as the granularity, the calculation formula is expanded as: ; In the formula, is the historical gradient vector of the i-th power data sample, is the output of the model (extreme learning machine) corresponding to the i-th power data sample, is the number of samples input to the extreme learning machine in the current batch.

[0062] It should be noted that the historical gradient vector is the cumulative average of the gradient vectors corresponding to the i-th sample in historical training, that is, the gradient values corresponding to the i-th sample in previous training are successively accumulated and then divided by the number of training rounds it participates in, used to characterize the historical optimization experience of this sample. represents the gradient update direction of the i-th power data sample.

[0063] During the training process, in order to avoid overfitting and improve training efficiency, the robustness of the extreme learning machine training process is enhanced through the perturbation factor to ensure the best training effect under different types of power data. The calculation formula is expressed as: ; In the formula, is the learning rate of the extreme learning machine, is the current weight matrix of the gradient update amount (i.e., the weight update amount), is the weight matrix of the hidden layer of the extreme learning machine at the t-th iteration (i.e., the current weight matrix), is the transition weight matrix of the hidden layer of the extreme learning machine at the (t + 1)-th iteration. Preferably, is set to 0.01.

[0064] In the embodiment of the present invention, the gradient update amount is at least composed of the following three components: interference cancellation term, perturbation term, and non-linear adaptive adjustment term; in this way, the influences of gradient correction, perturbation, and non-linear relationship can be integrated, the update direction can be optimized, and the generalization ability can be balanced. The calculation formula is expressed as: ; In the formula, is the first adjustment parameter, is the second adjustment parameter, is the perturbation term (i.e., the perturbation factor), is the interference cancellation term, is the non-linear adaptive adjustment term. Preferably, is set to 0.2, is set to 0.3.

[0065] In the embodiment of the present invention, the perturbation term (perturbation factor) is calculated based on the dot product sum of the difference between the input sample deviating from the mean value and the gradient, so as to enhance the robustness of the model to the change of the input distribution and prevent overfitting. The calculation formula is expressed as: ; In the formula, is the i-th power data sample input to the extreme learning machine, is the sample mean vector input to the extreme learning machine, is the number of samples input to the extreme learning machine in the current batch.

[0066] In the process of gradient update in the embodiment of the present invention, a non-linear adaptive adjustment term is also introduced. First, construct the non-linear feature vector of the power sample, and combine the original input vector, the square terms of its components, and the cross products between the features into a new high-order feature vector to construct the non-linear relationship between the input features, so as to improve the perception ability of the model to the complex patterns of power data.

[0067] After completing the construction of non-linear features, combined with the calculation mechanism of the non-linear adjustment factor, that is, by performing a weighted product of the above non-linear feature vector and the weight matrix of the extreme learning machine in the current iteration round, and through normalization processing by the Sigmoid function, an adjustment factor between 0 and 1 is obtained, which is used to quantify the response degree of the model to the non-linear structure in the current sample, so as to provide an adaptive adjustment basis for parameter update and ensure that the model has higher sensitivity and adaptability to non-linear features.

[0068] After obtaining the non-linear adjustment factor, further construct a non-linear adaptive adjustment term. By weighted combining this factor with the current gradient information, an additional gradient term that can dynamically adjust the update direction and amplitude is generated. Combining the compensation mechanism for the non-linear structure of the data during the weight update process enables the training process to not only follow the guidance of minimizing errors but also effectively capture the high-order structure features in the input space, enhancing the generalization ability of the model and improving the stability and accuracy of classifying complex power data.

[0069] In specific implementation, define the non-linear adaptive adjustment term as a correction term related to the non-linear relationship of power data. This term is calculated based on the high-order features (such as quadratic terms, cross terms, etc.) of the input power data. Assume that each sample of power data contains several features, and its non-linear feature vector is the high-order transformation form of the power data input to the extreme learning machine, expressed as: ; In the formula, is the matrix composed of the non-linear feature vectors of each sample in the power data sample set in the current batch, represents the matrix composed of the square terms of the power data features corresponding to each sample, represents the matrix of cross terms (such as , etc., , , are the three power data features of the power data sample) corresponding to each sample.

[0070] Taking the sample as the granularity, the calculation formula is expanded and expressed as: ; where, is the non-linear feature vector of the i-th power data sample, represents the square term of the power data feature in the i-th power data sample, represents the cross term between the power data features in the i-th power data sample.

[0071] Furthermore, a non - linear adjustment factor is defined to quantify the non - linear relationship in the power data samples and is calculated based on the internal structure of the non - linear feature vectors. Specifically, by calculating the response intensity of the current weight matrix to the non - linear feature vectors, the non - linear adjustment factor is obtained, and the corresponding calculation formula is as follows: ; In the formula, is the non - linear adjustment factor, which characterizes the adaptability of the model to the non - linear relationship in the power data samples, is the number of samples input to the extreme learning machine in the current batch.

[0072] Taking samples as the granularity, the calculation formula is expanded as: .

[0073] Furthermore, the non - linear adaptive adjustment term is obtained by adjusting the current gradient update direction based on the non - linear adjustment factor and the influence coefficient factor, and the corresponding calculation formula is as follows: ; Taking samples as the granularity, the calculation formula is expanded as: ; In the formula, is the weight hyper - parameter of the non - linear adaptive adjustment term, which is used to control the influence of the non - linear adjustment term on the total weight update. represents the current gradient calculation, indicating how the model adjusts the weights according to the error; ; is the number of samples input to the extreme learning machine in the current batch. Preferably, is set to 0.2.

[0074] It should be noted that in the weight update process of this embodiment of the present invention, through the perturbation factor and the non - linear adaptive adjustment term, the robustness of the model is improved, over - fitting is prevented, and the update direction is optimized.

[0075] In addition, in order to avoid the negative impact of too large a parameter update amplitude on the model stability, this embodiment of the present invention introduces a viscous control strategy. When updating the parameters each time, the update step size is adjusted through the historical weight changes, so as to avoid oscillation or too fast convergence.

[0076] Specifically, the current weight matrix is updated based on the gradient update amount to obtain a transition weight matrix; then, based on the difference between the weight matrices of adjacent iteration cycles and the viscous factor, the transition weight matrix is smoothed and corrected to obtain an updated weight matrix, and the corresponding calculation formula is as follows: ; In the formula, is the viscosity factor; is the weight matrix of the hidden layer of the extreme learning machine at the (t + 1)-th iteration (i.e., the updated weight matrix), which is used as the training parameter for the next iteration; is the weight matrix of the hidden layer of the extreme learning machine at the (t - 1)-th iteration; is the transition weight matrix of the hidden layer of the extreme learning machine at the (t + 1)-th iteration.

[0077] It should be noted that the viscosity factor is calculated based on the average amplitude of historical weight changes to dynamically control the update step size, thereby suppressing oscillations and ensuring stable convergence. Its calculation formula is expressed as: ; In the formula, T is the total number of iterations of the extreme learning machine. Preferably, T is set to 400.

[0078] After multiple rounds of training, the classifier finally outputs the classification result of each sample. This result is calculated through the output layer of the extreme learning machine and the final classification result is obtained through the activation function. The calculation formula is expressed as: ; In the formula, is the Softmax activation function, is the classification output of the extreme learning machine.

[0079] After the extreme learning machine model (initial model) is trained, a target identification model is obtained. This target identification model is used to identify the power data to be detected and output the corresponding load impact type; that is, the collected power data (power data to be detected) is input into the trained target identification model for classification, and then the classification result is obtained. In the embodiments of the present invention, the classification categories include four types: normal fluctuation, slight impact, medium impact, and severe impact.

[0080] See Figure 3 , which is a schematic structural diagram of an embodiment of a load impact type identification system for a distributed power grid provided by the present invention.

[0081] The second aspect of the embodiments of the present invention provides a load impact type identification system for a distributed power grid, including: An initial training set acquisition module 11, configured to acquire an initial training set of a target power system; An augmented training set acquisition module 12, configured to use a preset generative adversarial network introducing a local curvature perception mechanism to perform sample augmentation on the initial training set to obtain an augmented training set; wherein, the augmented training set includes a plurality of power data samples; The target identification model construction module 13 is configured to train an initial model through the augmented training set to obtain a target identification model; wherein, the target identification model is used to identify the power data to be detected and output the corresponding load impact type.

[0082] It should be noted that the load impact type identification system for a distributed power grid provided in the second aspect embodiment of the present invention can implement all the processes of the load impact type identification method for a distributed power grid described in any embodiment of the first aspect. The functions and achieved technical effects of each module and unit in the system are respectively the same as those of the load impact type identification method for a distributed power grid described in any embodiment of the first aspect, and will not be elaborated here.

[0083] The above are only the preferred embodiments of the present invention. It should be pointed out that for those of ordinary skill in the art, without departing from the technical principles of the present invention, several improvements and modifications can be made, and these improvements and modifications should also be regarded as the protection scope of the present invention.

Claims

1. A method for identifying load impact types for a distributed power grid, characterized in that, Including: Obtain an initial training set of the target power system; Use a preset generative adversarial network introducing a local curvature perception mechanism to perform sample augmentation on the initial training set to obtain an augmented training set; wherein, the augmented training set includes a number of power data samples; Train an initial model through the augmented training set to obtain a target identification model; wherein, the target identification model is used to identify the power data to be detected and output the corresponding load impact type.

2. The method for identifying load impact types for a distributed power grid according to claim 1, characterized in that The generator loss in the preset generative adversarial network includes: adversarial loss, multi-scale local curvature loss, and high-order curvature alignment loss.

3. The load impact type identification method for a distributed power grid according to claim 2, wherein, The multi-scale local curvature loss is calculated by the following formula: ； Among them, is the multi-scale local curvature loss function; is the generator; is the set of real power data samples input to the preset generative adversarial network in the current batch; is the th second derivative of the generated power data sample at the th scale; is the th second derivative of the real power data sample at the th scale; is the third derivative of the th generated power data sample; is the third derivative of the th real power data sample; is the th generated power data sample; is the th real power data sample; is the first features of the th generated power data sample; is the first features of the th real power data sample; is the curvature loss balance coefficient; is the total number of scales; is the number of samples input to the preset generative adversarial network in the current batch; is a positive integer; is a positive integer and ; is the L2 norm.

4. The load impact type identification method for a distributed power grid according to claim 2, characterized in that, The generator loss is calculated by the following formula: ； Among them, is the loss function of the generator; is the set of real power data samples input to the preset generative adversarial network in the current batch; is the set of latent space vectors corresponding to the set of real power data samples; is the generator function; is the discriminator function; is the latent variable interaction term; indicates that the input of the generator is the corresponding expectation when; is the multi-scale local curvature loss function; is the higher-order curvature term of the th generated power data sample; is the higher-order curvature term of the th real power data sample; is the first generator loss weight coefficient; is the number of samples input to the preset generative adversarial network in the current batch.

5. The method for identifying load impact types for a distributed power grid according to claim 1, wherein The method further includes: During the training process of the initial model, calculate the gradient update amount of the current weight matrix according to the augmented training set; wherein, the gradient update amount is composed of at least the following three components: interference cancellation term, perturbation term, and non-linear adaptive adjustment term; the interference cancellation term is used to correct the gradient update direction according to the historical gradient vector of the power data sample; the perturbation term is used to adjust the gradient update direction according to the deviation degree of the power data sample from the sample mean vector; the non-linear adaptive adjustment term is used to adjust the gradient update direction according to the non-linear relationship in the power data sample.

6. The method for identifying load impact types for a distributed power grid according to claim 5, characterized in that The interference cancellation term is calculated by the following formula: ； Among them, is the interference cancellation term; is the learning rate adjustment coefficient of the initial model; is the model output corresponding to the i-th power data sample; represents the gradient update direction of the i-th power data sample; is the historical gradient vector of the i-th power data sample; is the number of samples input to the initial model for training in the current batch.

7. The load impact type identification method for a distributed power grid according to claim 5, characterized in that, The perturbation term is calculated by the following formula: ； Among them, is the disturbance term; is the i-th power data sample; is the sample mean vector; is the model output corresponding to the i-th power data sample; represents the gradient update direction of the i-th power data sample; is the number of samples input to the initial model for training in the current batch.

8. The method for identifying load impact types for a distributed power grid according to claim 5, wherein The non-linear adaptive adjustment term is obtained through the following steps: Perform high-order feature expansion on the power data sample to obtain a corresponding non-linear feature vector; wherein, the non-linear feature vector includes: the original feature of the power data sample, the square term of the original feature, and the combined cross term between the original features; Calculate the response intensity of the current weight matrix to the non-linear feature vector to obtain a non-linear adjustment factor; wherein, the non-linear adjustment factor is used to characterize the adaptability of the initial model to the non-linear relationship; Based on the non-linear adjustment factor and the influence coefficient factor, adjust the gradient update direction to obtain the non-linear adaptive adjustment term.

9. The load impact type identification method for a distributed power grid according to claim 5, wherein The method further includes: Update the current weight matrix based on the gradient update amount to obtain a transition weight matrix; Based on the weight matrix difference between adjacent iteration cycles and the viscosity factor, perform smooth correction on the transition weight matrix to obtain an updated weight matrix.

10. A load impact type identification system for a distributed power grid, characterized in that, Including: Initial training set acquisition module, used to obtain an initial training set of the target power system; Augmented training set acquisition module, used to use a preset generative adversarial network introducing a local curvature perception mechanism to perform sample augmentation on the initial training set to obtain an augmented training set; wherein, the augmented training set includes a number of power data samples; Target identification model construction module, used to train an initial model through the augmented training set to obtain a target identification model; wherein, the target identification model is used to identify the power data to be detected and output the corresponding load impact type.

Images