Power quality disturbance identification method based on artificial intelligence

Abstract

The present invention relates to the field of artificial intelligence technology, and in particular to an artificial intelligence-based power quality disturbance identification method. The artificial intelligence-based power quality disturbance identification method comprises the following main steps: step S1, data acquisition and labeling; step S2, disturbance data expansion, using a generative adversarial network algorithm based on quantum spectral decomposition for sample generation; step S3, feature extraction model training, using a training process of a neural network algorithm based on dynamic ecological optimization; step S4, feature dimensionality reduction model training, using an autoencoder neural network based on feature refinement as a dimensionality reduction model; step S5, classifier model training, using an extreme learning machine classification algorithm based on a manifold mapping strategy; and step S6, power quality disturbance identification.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an artificial intelligence-based power quality disturbance identification method. Background Art

[0002] With the increasing complexity of power systems and the increasing prominence of power quality issues, the monitoring and identification of power quality disturbances have become increasingly important. Power quality disturbances may cause a series of power equipment failures and energy efficiency losses, affecting the stability of industrial production and residents' lives. With the development of artificial intelligence technology, especially the application of deep learning and generative adversarial networks, it has become possible to use data-driven methods to automatically identify power quality disturbances.

[0003] The invention patent with publication number CN 116881808 A proposes a method for identifying power quality disturbance signals based on machine learning, including S1, applying DTCWT-LPP to the original disturbance signal, extracting corresponding feature quantities and performing dimensionality reduction processing; S2, after the original disturbance signal is processed by DTCWT-LPP, the corresponding feature quantities such as standard deviation, mean, maximum value, minimum value, etc. are obtained, which are input as input samples into local density clustering for classification and identification, and a corresponding classification model for power quality disturbance identification is established; S3, when new data is input, the power quality is classified according to the power quality classification model after DTCWT-LPP transformation, etc.

[0004] Existing power quality disturbance identification methods usually rely on traditional data generation methods. The models perform poorly when dealing with complex disturbances. Traditional neural network methods are often used, which are prone to gradient vanishing, gradient exploding, or falling into local optimal solutions during training. Existing feature dimensionality reduction methods usually lack a dynamic adjustment mechanism for feature importance. In addition, when traditional extreme learning machines process complex power quality disturbance data, their hidden layer weights and biases are fixed and lack adaptability to the statistical properties of the features.

[0005] In order to solve the above problems, a power quality disturbance identification method based on artificial intelligence is proposed. Summary of the Invention

[0006] Technical problems to be solved

[0007] 1) Existing technologies typically rely on traditional data generation methods, which cannot effectively expand the sample diversity of complex power quality disturbance data. This results in poor model performance and poor generalization ability when processing complex disturbances.

[0008] 2) Traditional neural network methods are prone to gradient vanishing, gradient exploding, or falling into local optimal solutions during training, resulting in unstable feature extraction and affecting their ability to process high-dimensional power quality disturbance data.

[0009] 3) Existing feature dimensionality reduction methods generally lack a dynamic adjustment mechanism for feature importance, resulting in a large amount of redundant information. The reduced feature space may not effectively retain the key information of the original data, affecting the efficiency and accuracy of the model.

[0010] 4) When processing complex power quality disturbance data, traditional extreme learning machines have fixed hidden layer weights and biases, lacking adaptability to the statistical properties of the features, resulting in poor classification performance for complex disturbance data. Furthermore, existing classifiers are not robust enough to abnormal data, making it difficult to cope with noise and outliers in disturbance data in real-world scenarios.

[0011] Technical solution: An artificial intelligence-based power quality disturbance identification method with the following main steps:

[0012] Step S1: data collection and annotation;

[0013] Step S2: perturbation data expansion, using a generative adversarial network algorithm based on quantum spectral decomposition to generate samples;

[0014] Step S3: feature extraction model training, using a training process of a neural network algorithm based on dynamic ecological optimization;

[0015] Step S4: feature dimensionality reduction model training, using an autoencoder neural network based on feature refinement as the dimensionality reduction model;

[0016] Step S5: classifier model training, using an extreme learning machine classification algorithm based on a manifold mapping strategy;

[0017] Step S6: Identify power quality disturbances.

[0018] Furthermore, the data collected in step S1 is mainly time-series voltage signals; the data comes from multiple power system monitoring stations equipped with high-precision voltage sensors; the collection method is real-time data streaming, the signal is read directly from the sensor through the embedded system, and then transmitted to the central data processing center through a secure network protocol, and the collected disturbance data is stored in a high-performance server in CSV format;

[0019] The labeling of each disturbance category is completed by power system experts.

[0020] Furthermore, the perturbation data expansion in step S2 and the training process of the generative adversarial network algorithm based on quantum spectral decomposition are as follows:

[0021] Step S201: Initialize the quantum circuit at the beginning of training. The circuit consists of two parts: the generator and the discriminator. Each part consists of multiple qubits. The qubits are set to the ground state during initialization. The Hadamard gate is used to initialize each qubit to a superposition state, which is expressed as:

[0022] Where, represents the initial quantum state, represents the Hadamard gate, represents the number of qubits, is the tensor product symbol, Indicates that the Hadamard gate acts simultaneously on qubits, means that all qubits are in ground state;

[0023] Step S202: convert the perturbation data into the state on the quantum bit. The perturbation data is mapped into the quantum state using the quantum encoding function as follows: Where, Indicates the The quantum state after quantum encoding, is the quantum encoding function, For the perturbation data samples;

[0024] Among them, the implementation of the quantum coding function is expressed as: Where, Indicates that to qubits perform tensor product operations, represents a revolving door around the y-axis, For the The rotation angle of the revolving door of the perturbed data sample is obtained by mapping the normalization function, which linearly normalizes each dimension of the perturbed data sample to Interval, and mapped to the rotation angle, the calculation method is expressed as: Where, represents all perturbation data sample sets, represents the minimum function, represents a maximum function; in step S203, the generator transforms the quantum state through a quantum gate sequence to generate a new perturbation data sample, wherein the method of adjusting the quantum state by using parameterized quantum gates is expressed as:

[0025] Where, Indicates the The quantum state after quantum encoding, represents the quantum gate sequence of the generator, represents the parameter set of the parameterized quantum gate, represents the quantum state output by the generator;

[0026] In step S204, the task of the discriminator is to distinguish the perturbation data generated by the generator from the perturbation data in the real perturbation data set. The discriminator also uses a quantum gate sequence to judge the state of the perturbation data, which is expressed as: Where, represents the quantum gate sequence of the discriminator, represents the discriminator parameters, Denotes the discriminator evaluation The probability that a perturbed data sample is the real perturbation data, represents the quantum state output by the generator, represents the conjugate state of the quantum state output by the generator;

[0027] Among them, for the training of the generator, the quantum gate sequence is a complex sequence composed of multiple quantum gates, and the calculation method is expressed as: Where, It is a combination of a CNOT gate and a parameterized revolving gate, each gate is based on the parameters Adjustment, parameters are learned through training data, The generator's The rotation angle of the revolving door, The number of rotation angle parameters for the generator's revolving door;

[0028] Moreover, for the training of the discriminator, the calculation method of the quantum gate sequence is expressed as: Where, It is a combination of the CNOT gate and the parameterized revolving gate of the discriminator, each gate is based on the parameters Adjustment, parameters are learned through training data, The discriminator The rotation angle of the revolving door, is the number of rotation angle parameters of the revolving door of the discriminator;

[0029] Step S205: Using quantum state entanglement quantization and optimization strategies, the degree of entanglement between generated quantum states is measured, and the parameters of the quantum gate are adjusted accordingly to optimize the complexity and authenticity of the generated perturbation data, thereby enhancing the effectiveness of the generative adversarial network in simulating complex perturbation data. The entanglement metric is defined as The entanglement metric is used to evaluate the degree of entanglement between the current quantum states. The goal of the entanglement metric is to maximize the entanglement between quantum states, thereby increasing the diversity and complexity of the generated perturbation data. The calculation method of the entanglement degree is expressed as:

[0030] Where, is the entanglement degree calculation function of the entanglement meter, is the reduced density matrix of the data generated by the generator, Represents trace operation;

[0031] Step S206: The generator and the discriminator are iteratively trained under the adversarial framework. The generator attempts to deceive the discriminator and generate samples that are closer and closer to the real perturbation data, while the discriminator strives to improve its ability to distinguish true and false perturbation data. Through this adversarial process, the generator learns how to improve the quality and diversity of its generated samples. The training objectives of the generator and the discriminator are constrained by the adversarial loss function, which is expressed as:

[0032] Where, is the loss function of adversarial training, Express expectations, represents the real perturbation data input to the discriminator, Indicates that it obeys a specific distribution. represents the real perturbation data distribution, represents random noise, represents the noise distribution, represents the perturbation data generated by the generator, It represents the probability that the discriminator evaluates the input true perturbation data sample to be the true perturbation data, represents the probability that the perturbation data sample generated by the discriminator is the real perturbation data, is a hyperparameter used to control the influence strength of the entanglement regularization term. Step S207: After each iteration, the quality of the generated perturbation data is evaluated by quantum state measurement, and the parameters of the quantum circuit are adjusted according to the measurement results to optimize the authenticity and diversity of the generated samples;

[0033] After each round of training, the measurement operation The quality of the generated perturbation data is evaluated and the generator is optimized accordingly, which is expressed as: Where, represents the learning rate of the generator, represents the gradient of the loss function of adversarial training with respect to the generator parameters, It is the parameter update operation;

[0034] The dynamic parameter adjustment mechanism allows the discriminator to automatically adjust its parameters according to the statistical characteristics of the generated perturbation data during training, thereby optimizing the discriminator's ability to evaluate the authenticity of the generated perturbation data. The dynamic parameter adjustment mechanism dynamically adjusts the discriminator's parameters according to the statistical distance between the discrimination results of the previous batch of perturbation data and the true perturbation data. The Wasserstein distance is used as a measure of the statistical distance, and the calculation method is expressed as: Where, represents the Wasserstein distance, measure and The distance between two distributions is a measure of the difference between the true distribution and the generated distribution. represents the joint distribution, is the infimum or minimum symbol, which refers to finding a possible joint distribution Minimize the expected distance, In the joint distribution Down, Right expectations, are values ​​from the true distribution, are values ​​from the generating distribution, is the L2 norm, and are the perturbation data distributions obtained from the real perturbation data and the generated perturbation data, respectively. is the joint distribution of all possible

[0035] Among them, based on the Wasserstein distance, the update method of the discriminator is expressed as: Where, is the learning rate of the discriminator, is the preset ideal Wasserstein distance, which is used to indicate the ideal distance between two distributions. represents the gradient of the loss function of adversarial training with respect to the discriminator parameters, is a sign function that indicates the direction of parameter adjustment;

[0036] Step S208: Repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed.

[0037] Furthermore, the feature extraction model training in step S3 is based on the training process of the dynamic ecological optimization neural network algorithm as follows:

[0038] Step S301: In the initialization phase, multiple populations are constructed. Each population represents a set of network parameters. The initial weights and biases of each population are generated by a random normal distribution, expressed as: Where, is the standard deviation of the initialization distribution of the neural network weight parameters, is the standard deviation of the initialization distribution of the bias parameters of the neural network, Indicates that it obeys a specific distribution. is the initial weight of the neural network, is the initial bias of the neural network, represents a normal distribution, represents the weight of the neural network, Represents the bias of the neural network;

[0039] Step S302: Perform fitness evaluation on each population. The fitness calculation is based on the prediction error of the network and the complexity of the network parameters. The calculation method of the fitness function is defined as follows: Where, is the fitness function of the neural network, is the loss function of the neural network, specifically the mean square error loss, It is the regularization term of the neural network, specifically the L2 regularization term, is the regularization coefficient of the neural network;

[0040] Step S303: Apply selection pressure to the population based on the fitness results. The population with low fitness will face a higher risk of elimination. The roulette wheel method is used to select the population with higher fitness for reproduction. The calculation method of the selection probability is expressed as: Where, is the probability of selecting a combination of neural network weight and bias parameters, For the The neural network weight parameters corresponding to the population, For the The neural network bias parameters corresponding to the population, is the population size, is the inhibition coefficient;

[0041] In order to make the competition between different populations more dynamic, it is assumed that the competition between populations depends not only on the current fitness, but also on the mutual inhibition effect between populations. The inhibition coefficient is used to represent the inhibition effect between populations, and the calculation method is expressed as: Where, is the L2 norm, is the adjustment factor of the inhibition coefficient, is the distance measure between population parameters, is the neural network weight parameter corresponding to the first population, is the neural network weight parameter corresponding to the second population;

[0042] Among them, the distance measurement between population parameters adopts Euclidean distance, and the calculation method is expressed as: Where, is the dimension of the weight vector, is the first population corresponding to the neural network weight parameter parameters, is the weight parameter of the neural network corresponding to the second population. parameters;

[0043] Step S304: The resources in the environment are redistributed according to the fitness of the population. The population with more resources has a higher reproduction rate and survival rate. The resource allocation is proportional to the fitness of the population. The new resource amount is defined as: Where, is the total resource amount, is the new resource volume; It is a resource quantity regulating factor used to control the competitive ability of a population in an ecological niche;

[0044] Step S305: Perform crossover and mutation of the population. The population with high fitness will generate a new population through crossover and mutation, simulating the genetic process of organisms. The way in which the population generates new individuals through crossover and mutation is expressed as: Where, is the neural network weight parameter after the crossover operation, is the bias parameter of the neural network after the crossover operation, is the cross ratio, is the neural network bias parameter corresponding to the first population, is the neural network bias parameter corresponding to the second population;

[0045] Among them, the mutation operation randomly perturbs the parameters according to a certain probability, which can be expressed as: Where, It is The mutation intensity of the iteration, is the neural network weight parameter after mutation operation, is the bias parameter of the neural network after the mutation operation;

[0046] Among them, the dynamic mutation rate can increase the complexity of mutation and dynamically adjust with the number of iterations. As the number of iterations increases, the mutation rate gradually decreases, ensuring that the algorithm conducts extensive searches in the early stage and fine-tunes in the later stage. The adjustment method is expressed as: Where, is the initial mutation rate, is the mutation rate attenuation factor, is the current iteration number;

[0047] Step S306: Repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed.

[0048] Furthermore, in step S4, the feature dimensionality reduction model training adopts an autoencoder neural network based on feature refinement as the dimensionality reduction model:

[0049] The autoencoding neural network based on feature refinement consists of three parts: encoder, decoder and feature adjustment module, where:

[0050] The encoder is responsible for mapping high-dimensional input data into a low-dimensional feature space.

[0051] The decoder is used to reconstruct the reduced-dimensional features back to the original space.

[0052] The feature adjustment module dynamically adjusts the feature space after dimensionality reduction through recursive feature adaptive optimization;

[0053] Specifically, the training process of the autoencoder neural network algorithm based on feature refinement is as follows:

[0054] Step S401: Assume that the data input to the autoencoder neural network is , the encoder uses a multi-layer nonlinear mapping structure to map high-dimensional data to the initial low-dimensional feature space, which can be expressed as: Where, Represents the initial low-dimensional features; is the weight matrix of the encoder; is the bias vector of the encoder, is the multi-layer Sigmoid activation function of the encoder;

[0055] Step S402: After the low-dimensional features are generated, the feature adjustment module automatically generates feature weights based on the feature importance in the current feature space. When the module is initialized, it assigns the same initial weight to all features so that they can be gradually adjusted according to the feature contribution in subsequent steps. The calculation method of the initial weight matrix is ​​expressed as: Where, is the initial weight matrix; is the feature weight vector, Each element in Initialized to the same value, indicating that all features have the same importance in the initial stage, is a function for extracting the diagonal elements of a matrix;

[0056] Among them, the adjusted features can be expressed as: Where, is the feature representation after feature weight adjustment;

[0057] Step S403: The feature adjustment module recursively optimizes the initially generated low-dimensional features. In each iteration, the module adjusts the weights of each feature based on the performance of the previous round of features, gradually enhancing those features with important influences and gradually weakening redundant or noise features. In the round iteration, the weight update rule is as follows: Where, Indicates the The feature weights of the round iteration, Indicates the The feature weights of the round iteration, is the learning rate of the autoencoder neural network, is the loss function of the autoencoder neural network, is the label data, Represents the gradient of the loss function with respect to the feature weight, where the loss function of the autoencoder neural network is the reconstruction error loss function;

[0058] Step S404: To ensure the effectiveness of the dimensionality reduction process, the decoder remaps the low-dimensional features back to the high-dimensional space to ensure that no important information is lost during the dimensionality reduction process. The reconstruction process of the decoder is expressed as: Where, is the reconstructed high-dimensional data, is the weight matrix of the decoder, is the decoder bias vector, is the multi-layer Sigmoid activation function of the decoder;

[0059] Step S405: Repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed.

[0060] Furthermore, in step S5, the training process of the classifier model training and the extreme learning machine classification algorithm based on the manifold mapping strategy is as follows:

[0061] Step S501: Initialize the extreme learning machine and set the weight matrix from the input layer to the hidden layer of the extreme learning machine to be , the bias vector from the input layer to the hidden layer of the extreme learning machine is , the initialization method is expressed as: Where, is the random initialization function, is the randomization range, is the dimension of the perturbation data after dimensionality reduction input to the extreme learning machine, is the number of hidden layer nodes of the extreme learning machine;

[0062] Step S502: The configuration of hidden layer nodes adopts a feature symbiosis strategy to dynamically adjust weights and biases to adapt to the statistical characteristics of the input features, so as to enhance the adaptability of the model to complex perturbation data structures. The adjustment of weights depends on the variability of the input features, and the adjustment method is expressed as: Where, and are the hidden layer weights and biases of the adjusted extreme learning machine, is the adjustment coefficient used to adjust the sensitivity of the weight to the statistical attributes of the feature, is a coefficient used to adjust the sensitivity of the bias to the statistical properties of the feature, Represent the variance and mean of the features of the perturbation data after dimensionality reduction input to the extreme learning machine, represents the perturbation data after dimensionality reduction input to the extreme learning machine, and They represent the skewness and kurtosis of the features of the perturbed data after dimensionality reduction that are input to the extreme learning machine, is a small constant that prevents the denominator from being zero;

[0063] Step S503: The disturbance data after feature dimensionality reduction passes through a hidden layer based on manifold mapping. The hidden layer based on manifold mapping can map the input disturbance data to a high-dimensional space in a nonlinear manner to extract more information that is helpful for classification. The output of the hidden layer after processing is expressed as: Where, is the output of the hidden layer of the extreme learning machine, is the activation function of the hidden layer of the extreme learning machine;

[0064] Among them, let the input of the activation function of the hidden layer of the extreme learning machine be , the activation function of the hidden layer of the extreme learning machine is an activation function based on a mixture of polynomials and exponentials, which can increase the model's ability to capture nonlinear features. The calculation method is expressed as: Where, is the Sigmoid activation function, is the hyperbolic tangent function, is the first nonlinear parameter, is the second nonlinear parameter, is the third nonlinear parameter;

[0065] Step S504: In the output layer of the extreme learning machine, the classifier is trained using a fast learning algorithm. The output layer weight is determined by minimizing the output error. The calculation method is expressed as: Where, is the output layer weight of the extreme learning machine, is the transpose of the output of the hidden layer of the extreme learning machine, is the target output, specifically the label vector of the perturbed data sample, is the identity matrix, is the robustness enhancement function of the extreme learning machine;

[0066] Among them, the calculation method of the robustness enhancement function of the extreme learning machine is expressed as: Where, yes The L1 norm of yes The square of the L2 norm, is a parameter that controls the strength of L1 regularization, yes The square of the L2 norm of ;

[0067] Step S505: adjust the weights and biases of the hidden layer nodes according to the training results, and test the performance of the model through the cross-validation method to ensure that the classifier can show good generalization ability on the unseen perturbation data after dimensionality reduction.

[0068] Furthermore, in step S6, the power quality disturbance identification includes 7 types of single disturbances and 13 types of compound disturbances:

[0069] The seven single disturbances include: voltage sag, voltage swell, voltage interruption, transient pulse, transient oscillation, voltage flicker and harmonics;

[0070] Among them, the 13 types of composite disturbances include: voltage sag + harmonics, voltage swell + harmonics, voltage interruption + harmonics, voltage flicker + harmonics, voltage sag + transient oscillation, voltage swell + transient oscillation, voltage sag + voltage flicker, voltage swell + voltage flicker, voltage flicker + transient pulses, voltage swell + harmonics + transient oscillation, voltage sag + harmonics + transient oscillation, voltage flicker + harmonics + transient pulses and voltage sag + harmonics + transient oscillation + transient pulses.

[0071] Beneficial effects: 1) By generating adversarial networks based on quantum spectral decomposition, more diverse and realistically distributed perturbation data are generated, the energy quality perturbation data is expanded, the superposition and entanglement characteristics of quantum states are used to simulate the distribution of complex perturbation data, and the complexity and authenticity of the generated samples are optimized using quantum state entanglement measurement, which significantly improves the training effect of the model and enhances the model's ability to recognize complex perturbation data, solves the problems of insufficient perturbation data samples and poor generalization ability, and avoids the limitations of traditional data augmentation methods in processing complex perturbation signals.

[0072] 2) By extracting features through a neural network based on dynamic ecological optimization, simulating the competition mechanism of biological populations in the ecosystem, and optimizing the weights and biases of the neural network, we effectively solve the problems of gradient vanishing and gradient exploding that are easily encountered in neural network training. This effectively improves the stability of the neural network in feature extraction, avoids the risk of falling into local optimal solutions, and significantly improves the network's feature capture ability on high-dimensional perturbation data.

[0073] 3) An autoencoding neural network based on feature refinement is used for feature dimensionality reduction. Through a recursive feature adaptive optimization module, feature weights are dynamically adjusted to enhance the expression of important features and weaken noise features, thereby effectively reducing the data dimension and achieving effective compression of the feature space. This not only maintains the key information in the disturbed data, but also reduces the computational burden, improves the efficiency of the model, and ensures that the features after dimensionality reduction retain key information.

[0074] 4) An extreme learning machine classification algorithm based on a manifold mapping strategy dynamically adjusts the weights and biases of hidden layer nodes and, combined with feature statistical properties, improves the classifier's ability to capture the nonlinear features of perturbed data. This enhances the classifier's ability to discern the perturbed features of complex perturbed data, improving classification accuracy. Furthermore, the use of L1 regularization improves the model's adaptability to anomalous data, making the classifier more robust in real-world scenarios. This enhanced robustness to anomalous data addresses the poor performance of traditional extreme learning machines when dealing with complex perturbed data. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0076] Figure 1 It is a training flow chart of the autoencoding neural network algorithm based on feature refinement of the present invention;

[0077] Figure 2 It is a comparison diagram of data distribution of different generation methods in the present invention;

[0078] Figure 3 This is a comparison chart of the training process of different optimization methods in the present invention;

[0079] Figure 4 It is a feature space visualization comparison diagram in the present invention;

[0080] Figure 5 This is a performance comparison chart of different classifiers in the present invention on various types of power disturbances. DETAILED DESCRIPTION

[0081] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments.

[0082] In the description of the present invention, it should be understood that the terms "upper", "lower", "front", "back", "left", "right", "top", "bottom", "inside", "outside", etc., indicating directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore should not be understood as limiting the present invention.

[0083] Reference Figure 1-3 , the present invention provides an embodiment:

[0084] The main steps of the power quality disturbance identification method based on artificial intelligence are as follows:

[0085] Step S1: data collection and annotation;

[0086] Step S2: perturbation data expansion, using a generative adversarial network algorithm based on quantum spectral decomposition to generate samples;

[0087] Step S3: feature extraction model training, using a training process of a neural network algorithm based on dynamic ecological optimization;

[0088] Step S4: feature dimensionality reduction model training, using an autoencoder neural network based on feature refinement as the dimensionality reduction model;

[0089] Step S5: classifier model training, using an extreme learning machine classification algorithm based on a manifold mapping strategy;

[0090] Step S6: Identify power quality disturbances.

[0091] In step S1, the training data collected for the power quality disturbance identification model is mainly the time-series voltage signal of the power quality disturbance; the collected data comes from multiple power system monitoring stations, which are equipped with high-precision voltage sensors that can monitor voltage changes in real time and capture various power quality disturbance events;

[0092] The disturbance data is collected in real-time data streams. The embedded system reads the signal directly from the sensor and transmits it to the central data processing center via a secure network protocol. The collected disturbance data is stored in a high-performance server in CSV format.

[0093] The labeling of each disturbance category is completed by power system experts.

[0094] Among them, the perturbation data expansion in step S2 and the training process of the generative adversarial network algorithm based on quantum spectral decomposition are as follows:

[0095] In the task of the present invention, the collection, acquisition, labeling and preprocessing of training data are time-consuming and labor-intensive, and insufficient training samples can easily lead to poor generalization ability of the model and affect the accuracy of the model. The present invention uses a generative adversarial network algorithm based on quantum spectral decomposition to generate samples, thereby achieving perturbation data expansion;

[0096] Based on the traditional generative adversarial network algorithm, the measurement of quantum states and noise suppression strategies are used to improve the quality and diversity of generated perturbation data, and the superposition and entanglement characteristics of quantum states are used to effectively simulate the distribution of complex perturbation data.

[0097] Step S201: Initialize the quantum circuit at the beginning of training. The circuit consists of two parts: the generator and the discriminator. Each part consists of multiple qubits. The qubits are set to the ground state during initialization. The Hadamard gate is used to initialize each qubit to a superposition state, which is expressed as: Where, represents the initial quantum state, represents the Hadamard gate, represents the number of qubits, is the tensor product symbol, Indicates that the Hadamard gate acts simultaneously on qubits, means that all qubits are in ground state;

[0098] Step S202: convert the perturbation data into the state on the quantum bit. The perturbation data is mapped into the quantum state using the quantum encoding function as follows: Where, Indicates the The quantum state after quantum encoding, is the quantum encoding function, For the perturbation data samples;

[0099] Among them, the implementation of the quantum coding function is expressed as: Where, Indicates that to qubits perform tensor product operations, represents a revolving door around the y-axis, For the The rotation angle of the revolving door for each perturbation data sample;

[0100] The rotation angle of the revolving door is obtained by mapping the normalization function, which linearly normalizes each dimension of the perturbation data sample to Interval, and mapped to the rotation angle, the calculation method is expressed as: Where, represents all perturbation data sample sets, represents the minimum function, represents the maximum value function;

[0101] Step S203: The generator transforms the quantum state through a quantum gate sequence to generate a new perturbation data sample. The method of adjusting the quantum state using the parameterized quantum gate is expressed as: Where, Indicates the The quantum state after quantum encoding, represents the quantum gate sequence of the generator, represents the parameter set of the parameterized quantum gate, represents the quantum state output by the generator;

[0102] In step S204, the task of the discriminator is to distinguish the perturbation data generated by the generator from the perturbation data in the real perturbation data set. The discriminator also uses a quantum gate sequence to judge the state of the perturbation data, which is expressed as: Where, represents the quantum gate sequence of the discriminator, represents the discriminator parameters, Denotes the discriminator evaluation The probability that a perturbed data sample is the real perturbation data, represents the quantum state output by the generator, represents the conjugate state of the quantum state output by the generator;

[0103] Among them, for the training of the generator, the quantum gate sequence is a complex sequence composed of multiple quantum gates, and the calculation method is expressed as: Where, It is a combination of a CNOT gate and a parameterized revolving gate, each gate is based on the parameters Adjustment, parameters are learned through training data, The generator's The rotation angle of the revolving door, The number of rotation angle parameters for the generator's revolving door;

[0104] Moreover, for the training of the discriminator, the calculation method of the quantum gate sequence is expressed as: Where, It is a combination of the CNOT gate and the parameterized revolving gate of the discriminator, each gate is based on the parameters Adjustment, parameters are learned through training data, The discriminator The rotation angle of the revolving door, is the number of rotation angle parameters of the revolving door of the discriminator;

[0105] Step S205: Using quantum state entanglement quantization and optimization strategies, the degree of entanglement between generated quantum states is measured, and the parameters of the quantum gate are adjusted accordingly to optimize the complexity and authenticity of the generated perturbation data, thereby enhancing the effectiveness of the generative adversarial network in simulating complex perturbation data. The entanglement metric is defined as The entanglement metric is used to evaluate the degree of entanglement between the current quantum states. The goal of the entanglement metric is to maximize the entanglement between quantum states, thereby increasing the diversity and complexity of the generated perturbation data. The calculation method of the entanglement degree is expressed as: Where, is the entanglement degree calculation function of the entanglement meter, is the reduced density matrix of the data generated by the generator, Represents trace operation;

[0106] Step S206: The generator and the discriminator are iteratively trained under the adversarial framework. The generator attempts to deceive the discriminator and generate samples that are closer and closer to the real perturbation data, while the discriminator strives to improve its ability to distinguish true and false perturbation data. Through this adversarial process, the generator learns how to improve the quality and diversity of its generated samples. The training objectives of the generator and the discriminator are constrained by the adversarial loss function, which is expressed as: Where, is the loss function of adversarial training, Express expectations, represents the real perturbation data input to the discriminator, Indicates that it obeys a specific distribution. represents the real perturbation data distribution, represents random noise, represents the noise distribution, represents the perturbation data generated by the generator, It represents the probability that the discriminator evaluates the input true perturbation data sample to be the true perturbation data, represents the probability that the perturbation data sample generated by the discriminator is the real perturbation data, is a hyperparameter used to control the influence of the entanglement regularization term, Set to 0.1.

[0107] Step S207: After each iteration, the quality of the generated perturbation data is evaluated by quantum state measurement, and the parameters of the quantum circuit are adjusted according to the measurement results to optimize the authenticity and diversity of the generated samples;

[0108] After each round of training, the measurement operation The quality of the generated perturbation data is evaluated and the generator is optimized accordingly, which is expressed as: Where, represents the learning rate of the generator, represents the gradient of the loss function of adversarial training with respect to the generator parameters, It is the parameter update operation; Set to 0.01.

[0109] The dynamic parameter adjustment mechanism allows the discriminator to automatically adjust its parameters according to the statistical characteristics of the generated perturbation data during training, thereby optimizing the discriminator's ability to evaluate the authenticity of the generated perturbation data. The dynamic parameter adjustment mechanism dynamically adjusts the discriminator's parameters according to the statistical distance between the discrimination results of the previous batch of perturbation data and the true perturbation data. The Wasserstein distance is used as a measure of the statistical distance, and the calculation method is expressed as: Where, represents the Wasserstein distance, measure and The distance between two distributions is a measure of the difference between the true distribution and the generated distribution. represents the joint distribution, is the infimum or minimum symbol, which refers to finding a possible joint distribution Minimize the expected distance, In the joint distribution Down, Right expectations, are values ​​from the true distribution, are values ​​from the generating distribution, is the L2 norm, and are the perturbation data distributions obtained from the real perturbation data and the generated perturbation data, respectively. is the joint distribution of all possible

[0110] Among them, based on the Wasserstein distance, the update method of the discriminator is expressed as: Where, is the learning rate of the discriminator, is the preset ideal Wasserstein distance, which is used to indicate the ideal distance between two distributions. represents the gradient of the loss function of adversarial training with respect to the discriminator parameters, is a sign function that indicates the direction of parameter adjustment; Set to 0.01;

[0111] Step S208: Repeat the above steps until a preset stop iteration condition is met, indicating that the model training is completed. The preset stop iteration condition is reaching a preset maximum number of iterations, which is set to 1000 times.

[0112] The feature extraction model training in step S3 is based on the training process of the neural network algorithm of dynamic ecological optimization as follows:

[0113] The expanded data is input into a feature extraction model for training. In existing technologies, some solutions use neural networks for feature extraction. In certain neural network structures, problems such as gradient vanishing, gradient exploding, or falling into local optimal solutions may occur, affecting training stability and model performance. The present invention uses a 6-layer fully connected neural network for feature extraction. This neural network is based on dynamic ecological optimization and is inspired by the concepts of ecosystem dynamic balance and species competition. By simulating the competition and adaptation process of multiple biological populations under limited resources, the weights and bias parameters of the neural network are optimized.

[0114] When neural networks are used to adjust parameters in dynamic ecological optimization algorithms, each population represents a set of possible network parameters (i.e., the weights and biases of the neural network). The diversity of the population directly affects the network's ability to explore the solution space. By defining the fitness function, not only error minimization is considered, but also parameter sparsity and regularization factors. This allows the algorithm to automatically balance the complexity and generalization ability of the model during the optimization process, thereby prompting the parameters to evolve towards a better solution.

[0115] Step S301: In the initialization phase, multiple populations are constructed. Each population represents a set of network parameters. The initial weights and biases of each population are generated by a random normal distribution, expressed as: Where, is the standard deviation of the initialization distribution of the neural network weight parameters, is the standard deviation of the initialization distribution of the bias parameters of the neural network, Indicates that it obeys a specific distribution. is the initial weight of the neural network, is the initial bias of the neural network, represents a normal distribution, represents the weight of the neural network, Represents the bias of the neural network; Set to 0.1, Set to 0.2;

[0116] Step S302: Perform fitness evaluation on each population. The fitness calculation is based on the prediction error of the network and the complexity of the network parameters. The calculation method of the fitness function is defined as follows: Where, is the fitness function of the neural network, is the loss function of the neural network, specifically the mean square error loss, It is the regularization term of the neural network, specifically the L2 regularization term, is the regularization coefficient of the neural network; Set to 0.01;

[0117] Step S303: Apply selection pressure to the population based on the fitness results. The population with low fitness will face a higher risk of elimination. The roulette wheel method is used to select the population with higher fitness for reproduction. The calculation method of the selection probability is expressed as: Where, is the probability of selecting a combination of neural network weight and bias parameters, For the The neural network weight parameters corresponding to the population, For the The neural network bias parameters corresponding to the population, is the population size, is the inhibition coefficient;

[0118] In order to make the competition between different populations more dynamic, it is assumed that the competition between populations depends not only on the current fitness, but also on the mutual inhibition effect between populations. The inhibition coefficient is used to represent the inhibition effect between populations, and the calculation method is expressed as: Where, is the L2 norm, is the adjustment factor of the inhibition coefficient, is the distance measure between population parameters, is the neural network weight parameter corresponding to the first population, is the neural network weight parameter corresponding to the second population; Set to 0.1;

[0119] Among them, the distance measurement between population parameters adopts Euclidean distance, and the calculation method is expressed as: Where, is the dimension of the weight vector, is the first population corresponding to the neural network weight parameter parameters, is the weight parameter of the neural network corresponding to the second population. parameters;

[0120] Step S304: The resources in the environment are redistributed according to the fitness of the population. The population with more resources has a higher reproduction rate and survival rate. The resource allocation is proportional to the fitness of the population. The new resource amount is defined as: Where, is the total resource amount, Set to twice the population size, is the new resource volume; It is a resource quantity regulating factor used to control the competitive ability of a population in an ecological niche; Set to 0.4;

[0121] Step S305: Perform crossover and mutation of the population. The population with high fitness will generate a new population through crossover and mutation, simulating the genetic process of organisms. The way in which the population generates new individuals through crossover and mutation is expressed as: Where, is the neural network weight parameter after the crossover operation, is the bias parameter of the neural network after the crossover operation, is the cross ratio, is the neural network bias parameter corresponding to the first population, is the neural network bias parameter corresponding to the second population;

[0122] Among them, the mutation operation randomly perturbs the parameters according to a certain probability, which can be expressed as: Where, It is The mutation intensity of the iteration, is the neural network weight parameter after mutation operation, is the bias parameter of the neural network after the mutation operation, Set to 0.2;

[0123] Among them, the dynamic mutation rate can increase the complexity of mutation and dynamically adjust with the number of iterations. As the number of iterations increases, the mutation rate gradually decreases, ensuring that the algorithm conducts extensive searches in the early stage and fine-tunes in the later stage. The adjustment method is expressed as: Where, is the initial mutation rate, is the mutation rate attenuation factor, is the current iteration number; Set to 0.01, Set to 0.001;

[0124] Step S306: Repeat the above steps until a preset stop iteration condition is met, indicating that the model training is completed. The preset stop iteration condition is reaching a preset maximum number of iterations, which is set to 1000 times.

[0125] In step S4, the feature dimensionality reduction model training adopts an autoencoder neural network based on feature refinement as the dimensionality reduction model:

[0126] The data after feature extraction is input into the feature dimensionality reduction model for training the feature dimensionality reduction model. The present invention adopts the autoencoder neural network based on feature refinement as the dimensionality reduction model.

[0127] The autoencoding neural network based on feature refinement consists of three parts: encoder, decoder, and feature adjustment module, where:

[0128] The encoder is responsible for mapping high-dimensional input data into a low-dimensional feature space.

[0129] The decoder is used to reconstruct the reduced-dimensional features back to the original space to ensure the reversibility of the dimensionality reduction process.

[0130] The feature adjustment module dynamically adjusts the feature space after dimensionality reduction through recursive feature adaptive optimization, so that important features can be strengthened and secondary features are gradually weakened, so that the feature representation after dimensionality reduction has better simplicity while retaining data information.

[0131] Specifically, the training process of the autoencoder neural network algorithm based on feature refinement is as follows:

[0132] Step S401: Assume that the data input to the autoencoder neural network is , the encoder uses a multi-layer nonlinear mapping structure to map high-dimensional data to the initial low-dimensional feature space, which can be expressed as: Where, Represents the initial low-dimensional features; is the weight matrix of the encoder; is the bias vector of the encoder, is the multi-layer Sigmoid activation function of the encoder;

[0133] Step S402: After the low-dimensional features are generated, the feature adjustment module automatically generates feature weights based on the feature importance in the current feature space. When the module is initialized, it assigns the same initial weight to all features so that they can be gradually adjusted according to the feature contribution in subsequent steps. The calculation method of the initial weight matrix is ​​expressed as: Where, is the initial weight matrix; is the feature weight vector, Each element in Initialized to the same value, indicating that all features have the same importance in the initial stage, is a function for extracting the diagonal elements of a matrix;

[0134] Among them, the adjusted features can be expressed as: Where, is the feature representation after feature weight adjustment;

[0135] Step S403: The feature adjustment module recursively optimizes the initially generated low-dimensional features. In each iteration, the module adjusts the weights of each feature based on the performance of the previous round of features, gradually enhancing those features with important influences and gradually weakening redundant or noise features. In the round iteration, the weight update rule is as follows: Where, Indicates the The feature weights of the round iteration, Indicates the The feature weights of the round iteration, is the learning rate of the autoencoder neural network, is the loss function of the autoencoder neural network, is the label data, Represents the gradient of the loss function with respect to the feature weight, where the loss function of the autoencoder neural network is the reconstruction error loss function; Set to 0.01;

[0136] Step S404: To ensure the effectiveness of the dimensionality reduction process, the decoder remaps the low-dimensional features back to the high-dimensional space to ensure that no important information is lost during the dimensionality reduction process. The reconstruction process of the decoder is expressed as: Where, is the reconstructed high-dimensional data, is the weight matrix of the decoder, is the decoder bias vector, is the multi-layer Sigmoid activation function of the decoder;

[0137] Step S405: Repeat the above steps until the preset stop iteration condition is met, which means the model training is completed. The preset stop iteration condition is to reach the preset maximum number of iterations, which is set to 1000 times.

[0138] Among them, the training process of the classifier model training in step S5 and the extreme learning machine classification algorithm based on the manifold mapping strategy is as follows:

[0139] The reduced-dimensional data is input into a classifier to train the classifier model. The present invention adopts an extreme learning machine classification algorithm based on a manifold mapping strategy to classify perturbed data. On the basis of the traditional extreme learning machine, the hidden layer nodes of the extreme learning machine algorithm are optimized through the manifold mapping strategy to enhance the learning ability and generalization performance of the model.

[0140] Step S501: Initialize the extreme learning machine and set the weight matrix from the input layer to the hidden layer of the extreme learning machine to be , the bias vector from the input layer to the hidden layer of the extreme learning machine is , the initialization method is expressed as: Where, is the random initialization function, is the randomization range, is the dimension of the perturbation data after dimensionality reduction input to the extreme learning machine, is the number of hidden layer nodes of the extreme learning machine;

[0141] Step S502: The configuration of hidden layer nodes adopts a feature symbiosis strategy to dynamically adjust weights and biases to adapt to the statistical characteristics of the input features, so as to enhance the adaptability of the model to complex perturbation data structures. The adjustment of weights depends on the variability of the input features, and the adjustment method is expressed as: Where, and are the hidden layer weights and biases of the adjusted extreme learning machine, is the adjustment coefficient used to adjust the sensitivity of the weight to the statistical attributes of the feature, is a coefficient used to adjust the sensitivity of the bias to the statistical properties of the feature, Represent the variance and mean of the features of the perturbation data after dimensionality reduction input to the extreme learning machine, represents the perturbation data after dimensionality reduction input to the extreme learning machine, and They represent the skewness and kurtosis of the features of the perturbed data after dimensionality reduction that are input to the extreme learning machine, is a small constant that prevents the denominator from being zero; Set to 0.01.

[0142] Step S503: The disturbance data after feature dimensionality reduction passes through a hidden layer based on manifold mapping. The hidden layer based on manifold mapping can map the input disturbance data to a high-dimensional space in a nonlinear manner to extract more information that is helpful for classification. The output of the hidden layer after processing is expressed as: Where, is the output of the hidden layer of the extreme learning machine, is the activation function of the hidden layer of the extreme learning machine;

[0143] Among them, let the input of the activation function of the hidden layer of the extreme learning machine be , the activation function of the hidden layer of the extreme learning machine is an activation function based on a mixture of polynomials and exponentials, which can increase the model's ability to capture nonlinear features. The calculation method is expressed as: Where, is the Sigmoid activation function, is the hyperbolic tangent function, is the first nonlinear parameter, is the second nonlinear parameter, is the third nonlinear parameter; Set to 0.1, Set to 0.4, Set to 0.4.

[0144] Step S504: In the output layer of the extreme learning machine, the classifier is trained using a fast learning algorithm. The output layer weight is determined by minimizing the output error. The calculation method is expressed as: Where, is the output layer weight of the extreme learning machine, is the transpose of the output of the hidden layer of the extreme learning machine, is the target output, specifically the label vector of the perturbed data sample, is the identity matrix, is the robustness enhancement function of the extreme learning machine;

[0145] In order to improve the robustness of the model to outliers in the training data, a penalty term based on the L1 norm is used, making the model less sensitive to outliers in the perturbation data during training. The calculation method of the robustness enhancement function of the extreme learning machine is expressed as: Where, yes The L1 norm of yes The square of the L2 norm, is a parameter that controls the strength of L1 regularization, yes The square of the L2 norm of ; Set to 0.2;

[0146] Step S505: adjust the weights and biases of the hidden layer nodes according to the training results, and test the performance of the model through the cross-validation method to ensure that the classifier can show good generalization ability on the unseen perturbation data after dimensionality reduction.

[0147] In step S6, power quality disturbance identification is performed, and the types of disturbance identification include 7 single disturbances and 13 compound disturbances:

[0148] The seven single disturbances include: voltage sag, voltage swell, voltage interruption, transient pulse, transient oscillation, voltage flicker and harmonics;

[0149] Among them, the 13 types of composite disturbances include: voltage sag + harmonics, voltage swell + harmonics, voltage interruption + harmonics, voltage flicker + harmonics, voltage sag + transient oscillation, voltage swell + transient oscillation, voltage sag + voltage flicker, voltage swell + voltage flicker, voltage flicker + transient pulses, voltage swell + harmonics + transient oscillation, voltage sag + harmonics + transient oscillation, voltage flicker + harmonics + transient pulses and voltage sag + harmonics + transient oscillation + transient pulses.

[0150] The trained models are used to process new samples to realize power quality disturbance identification. The collected original data are input into the trained feature extraction and feature dimension reduction models for feature processing. Furthermore, the processed features are input into the classifier model for classification to obtain the disturbance identification results.

[0151] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. The power quality disturbance identification method based on artificial intelligence is characterized by: Here are the steps: Step S1: data collection and annotation; Step S2: perturbation data expansion, using a generative adversarial network algorithm based on quantum spectral decomposition to generate samples; Step S3: feature extraction model training, using a training process of a neural network algorithm based on dynamic ecological optimization; Step S301: In the initialization phase, multiple populations are constructed. Each population represents a set of network parameters. The initial weights and biases of each population are generated by a random normal distribution. Step S302: Perform fitness evaluation on each population. The fitness calculation is based on the prediction error of the network and the complexity of the network parameters. The calculation method of the fitness function is defined as follows: Fitness(W p ,b p )=exp(-L(W p ,b p )-λ r R(w P ,b p )) Where, Fitness() is the fitness function of the neural network, L() is the loss function of the neural network, specifically the mean square error loss, R() is the regularization term of the neural network, specifically the L2 regularization term, and λ r is the regularization coefficient of the neural network; Step S303: Apply selection pressure to the population based on the fitness results. The population with low fitness will face a higher risk of elimination. The roulette wheel method is used to select the population with higher fitness for reproduction. The calculation method of the selection probability is expressed as: Where, P s (W p ,b p ) is the selection probability of the combination of neural network weights and bias parameters, W i is the neural network weight parameter corresponding to the i-th population, b i is the bias parameter of the neural network corresponding to the i-th population, N p is the population size, η pq is the inhibition coefficient; Step S304: The resources in the environment are redistributed according to the fitness of the population. The population with more resources has a higher reproduction rate and survival rate. The resource allocation is proportional to the fitness of the population. Step S305: Perform crossover and mutation of the population. The population with high fitness will generate a new population through crossover and mutation, simulating the genetic process of organisms. The way in which the population generates new individuals through crossover and mutation is expressed as: Where, is the neural network weight parameter after the crossover operation, is the bias parameter of the neural network after the cross operation, α prer is the cross ratio, W p1 is the neural network bias parameter corresponding to the first population, W p2 is the neural network bias parameter corresponding to the second population; Step S306: Repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed. Step S4: feature dimensionality reduction model training, using an autoencoder neural network based on feature refinement as the dimensionality reduction model; The autoencoding neural network based on feature refinement consists of three parts: encoder, decoder and feature adjustment module, where: The encoder is responsible for mapping high-dimensional input data into a low-dimensional feature space. The decoder is used to reconstruct the reduced-dimensional features back to the original space. The feature adjustment module dynamically adjusts the feature space after dimensionality reduction through recursive feature adaptive optimization. Step S5: classifier model training, using the extreme learning machine classification algorithm based on the manifold mapping strategy, the process is as follows; Step S501: Initialize the extreme learning machine. Step S502: The configuration of hidden layer nodes adopts a feature symbiosis strategy, dynamically adjusting weights and biases to adapt to the statistical characteristics of input features, so as to enhance the adaptability of the model to complex perturbation data structures. Step S503: The disturbance data after feature dimensionality reduction passes through a hidden layer based on manifold mapping. The hidden layer based on manifold mapping can map the input disturbance data to a high-dimensional space in a nonlinear manner to extract more information that is helpful for classification. Step S504: In the output layer of the extreme learning machine, the classifier is trained using a fast learning algorithm, and the output layer weight is determined by minimizing the output error. Step S505: adjust the weights and biases of the hidden layer nodes according to the training results, and test the performance of the model through the cross-validation method to ensure that the classifier can show good generalization ability on the unseen perturbation data after dimensionality reduction. Step S6: Identify power quality disturbances.

2. The method for identifying power quality disturbances according to claim 1, wherein: The data collected in step S1 is a time-series voltage signal; the data comes from multiple power system monitoring stations equipped with high-precision voltage sensors; the collection method is real-time data streaming, the signal is read directly from the sensor through the embedded system, and then transmitted to the central data processing center through a secure network protocol. The collected disturbance data is stored in a high-performance server in CSV format; The labeling of each disturbance category is completed by power system experts.

3. The method for identifying power quality disturbances according to claim 2, wherein: The perturbation data expansion in step S2 and the training process of the generative adversarial network algorithm based on quantum spectral decomposition are as follows: Step S201: Initialize the quantum circuit at the beginning of training. The circuit consists of two parts: the generator and the discriminator. Each part consists of multiple qubits. The qubits are set to the ground state during initialization. Hadamard gates are used to initialize each qubit to a superposition state. Step S202: convert the perturbation data into the state on the quantum bit. The perturbation data is mapped into the quantum state using the quantum encoding function as follows: |φ c,i >=ENC c (x c,i ) Where, |φ c,i > represents the quantum state after the i-th quantum encoding, ENC c () is the quantum encoding function, x c,i is the i-th perturbation data sample; Step S203: The generator transforms the quantum state through a quantum gate sequence to generate a new perturbation data sample. In step S204, the task of the discriminator is to distinguish the perturbation data generated by the generator from the perturbation data in the real perturbation data set. The discriminator also uses a quantum gate sequence to judge the state of the perturbation data, which is expressed as: Where, represents the quantum gate sequence of the discriminator, Ce c represents the discriminator parameters, P c (real|x c,i ) represents the probability that the discriminator evaluates that the i-th perturbation data sample is the true perturbation data, represents the quantum state output by the generator, represents the conjugate state of the quantum state output by the generator; Among them, for the training of the generator, the quantum gate sequence is a complex sequence composed of multiple quantum gates, and the calculation method is expressed as: Where, It is a combination of a CNOT gate and a parameterized revolving gate, each gate is based on the parameter De c.k Adjustment, parameters are learned through training data, De c.k is the rotation angle of the kth revolving door of the generator, K c The number of rotation angle parameters for the generator's revolving door; Moreover, for the training of the discriminator, the calculation method of the quantum gate sequence is expressed as: Where, It is a combination of CNOT gate and parameterized rotation gate of the discriminator, each gate is based on the parameter Ce c.m Adjustment, parameters are learned through training data, Ce c.m is the rotation angle of the kth revolving door of the discriminator, M c is the number of rotation angle parameters of the revolving door of the discriminator; Step S205: Using quantum state entanglement quantization and optimization strategies, the degree of entanglement between generated quantum states is measured, and the parameters of the quantum gate are adjusted accordingly to optimize the complexity and authenticity of the generated perturbation data, thereby enhancing the effectiveness of the generative adversarial network in simulating complex perturbation data. The entanglement metric is defined as E c The entanglement metric is used to evaluate the degree of entanglement between the current quantum states. The goal of the entanglement metric is to maximize the entanglement between the quantum states, thereby increasing the diversity and complexity of the generated perturbation data. Step S206: The generator and the discriminator are iteratively trained under the adversarial framework. The generator attempts to deceive the discriminator and generate samples that are closer and closer to the real perturbation data, while the discriminator strives to improve its ability to distinguish true and false perturbation data. Through this adversarial process, the generator learns how to improve the quality and diversity of its generated samples. The training objectives of the generator and the discriminator are constrained by the adversarial loss function. Step S207: After each iteration, the quality of the generated perturbation data is evaluated by quantum state measurement, and the parameters of the quantum circuit are adjusted according to the measurement results to optimize the authenticity and diversity of the generated samples; After each round of training, the operation M is measured c Evaluate the quality of the generated perturbation data and optimize the generator accordingly, The dynamic parameter adjustment mechanism allows the discriminator to automatically adjust its parameters according to the statistical characteristics of the generated perturbation data during training, thereby optimizing the discriminator's ability to evaluate the authenticity of the generated perturbation data. The dynamic parameter adjustment mechanism dynamically adjusts the discriminator's parameters according to the statistical distance between the discrimination results of the previous batch of perturbation data and the true perturbation data, using Wasserstein distance as a measure of statistical distance. Step S208: Repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed.

4. The method for identifying power quality disturbances according to claim 3, wherein: In step S6, the power quality disturbance identification includes 7 types of single disturbances and 13 types of compound disturbances: The seven single disturbances include: voltage sag, voltage swell, voltage interruption, transient pulse, transient oscillation, voltage flicker and harmonics; Among them, the 13 types of composite disturbances include: voltage sag + harmonics, voltage swell + harmonics, voltage interruption + harmonics, voltage flicker + harmonics, voltage sag + transient oscillation, voltage swell + transient oscillation, voltage sag + voltage flicker, voltage swell + voltage flicker, voltage flicker + transient pulses, voltage swell + harmonics + transient oscillation, voltage sag + harmonics + transient oscillation, voltage flicker + harmonics + transient pulses and voltage sag + harmonics + transient oscillation + transient pulses.

5. The method for identifying power quality disturbances according to claim 3, wherein: In the perturbation data expansion in step S2, in the training process of the generative adversarial network algorithm based on quantum spectral decomposition: In step S202, the implementation of the quantum coding function is expressed as: Where, Indicates the order from 1st to nth c qubits perform tensor product operations, RY c (θ c,j (x c,i )) represents a revolving door around the y-axis, θ c,j (x c,i ) is the rotation angle of the revolving door for the i-th disturbance data sample; The rotation angle of the revolving door is obtained by mapping the normalization function. Each dimension of the perturbation data sample is linearly normalized to the interval [0, 1] and mapped to the rotation angle. The calculation method is expressed as: Where x c Represents all perturbation data sample sets, min() represents the minimum function, and max() represents the maximum function.

6. The method for identifying power quality disturbances according to claim 3, wherein: In the perturbation data expansion in step S2, in the training process of the generative adversarial network algorithm based on quantum spectral decomposition: In step S207, the Wasserstein distance is used as a measure of the statistical distance, and the calculation method is expressed as follows: Where W() represents the Wasserstein distance, W(ρ real ,ρ gen ) Measure ρ real and ρ gen The distance between two distributions is a measure of the difference between the true distribution and the generated distribution, γ dar Represents the joint distribution, inf is the lower bound or minimum symbol, which refers to finding the possible joint distribution γ dar Minimize the expected distance, Indicates that in the joint distribution γ dar Next, (x tcs ,y tcs )Expectations for x tcs is the value from the true distribution, y tcs is the value from the generating distribution, |||| is the L2 norm, ρ real and ρ gen are the perturbation data distributions obtained from the real perturbation data and the generated perturbation data, Γ(ρ real ,ρ gen ) is the joint distribution of all possible Among them, based on the Wasserstein distance, the update method of the discriminator is expressed as: Where η φ is the learning rate of the discriminator, W0 is the preset ideal Wasserstein distance, which is used to indicate the ideal distance between the two distributions. It represents the gradient of the loss function of adversarial training with respect to the discriminator parameters, and sgn is a sign function used to indicate the direction of parameter adjustment.

7. The method for identifying power quality disturbances according to claim 1, wherein: In the feature extraction model training in step S3, the training process of the neural network algorithm based on dynamic ecological optimization is as follows: In step S303, selection pressure is applied to the population according to the fitness result, and the selection probability is calculated. In order to make the competition between different populations more dynamic, it is assumed that the competition between populations depends not only on the current fitness, but also on the mutual inhibition effect between populations. The inhibition coefficient is used to represent the inhibition effect between populations, and the calculation method is expressed as: Where |||| is the L2 norm, λ pq is the adjustment factor of the inhibition coefficient, d(W p1 ,W p2 ) is the distance metric between population parameters, W p1 is the neural network weight parameter corresponding to the first population, W p2 is the neural network weight parameter corresponding to the second population; Among them, the distance measurement between population parameters adopts Euclidean distance, and the calculation method is expressed as: Where mps is the dimension of the weight vector, W p1,i is the i-th parameter of the neural network weight parameter corresponding to the first population, W p2,i is the i-th parameter of the neural network weight parameter corresponding to the second population.

Images