Voltage quality composite disturbance identification method considering complex signal time-frequency feature fusion

Abstract

The invention relates to a voltage quality composite disturbance identification method considering complex signal time-frequency feature fusion. The method comprises the following steps: S1, obtaining time-frequency domain distribution information of a voltage quality disturbance signal; s2, dividing disturbance identification input characteristic indexes into time domain statistical characteristics, frequency domain statistical characteristics and energy characteristics; s3, extracting a depth feature from the voltage quality disturbance information, and fusing the depth feature with a disturbance identification input feature index; s4, weighting the fusion feature set; and step S5, adopting an integrated learning device XGBoost based on a Boosting strategy as a disturbance classifier, and inputting the fusion feature set into the XGBoost after empowerment so as to realize identification of the composite disturbance signal. The technical problems of power quality disturbance traceability and power system stability can be solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of voltage quality disturbance identification, and relates to a voltage quality composite disturbance identification method, in particular to a voltage quality composite disturbance identification method considering the fusion of complex signal time-frequency characteristics. Background Art

[0002] Against the backdrop of the energy revolution driven by the "dual carbon" goals, massive heterogeneous distributed energy and nonlinear loads are connected to the power grid, which, while enhancing the flexibility of the power grid, has caused more complex and changeable voltage quality disturbance problems, seriously affecting the stability and economy of the power system operation.

[0003] Voltage quality disturbance identification is the main component of the voltage quality monitoring system in the smart grid paradigm, and is also an important prerequisite for managing and improving voltage quality. However, in the complex process of building a new power system, voltage quality disturbances have changed from traditional single disturbances to mixed disturbances, that is, they are composed of a variety of basic voltage quality disturbances with different disturbance types, different disturbance intensities, and different start and end times.

[0004] Moreover, the composite voltage quality disturbance characteristic quantity is not a simple linear superposition of the basic voltage quality disturbance characteristic quantities, but may be the overlapping and complex intersection of time and frequency domain characteristics, which brings huge challenges to the traditional voltage quality disturbance identification technology.

[0005] Therefore, extracting appropriate voltage quality indicators and accurately identifying voltage quality disturbances through artificial intelligence algorithms has become an important research difficulty in the field of voltage quality in recent years.

[0006] After searching, no public documents of the prior art that are identical or similar to the present invention were found. Summary of the invention

[0007] The purpose of the present invention is to overcome the shortcomings of the prior art and propose a voltage quality composite disturbance identification method considering the fusion of complex signal time-frequency characteristics, which can solve the technical problems of power quality disturbance tracing and power system stability.

[0008] The present invention solves the practical problem by adopting the following technical solutions:

[0009] A voltage quality composite disturbance identification method considering the fusion of complex signal time-frequency characteristics includes the following steps:

[0010] Step S1, using an improved Hilbert-Huang transform (IHHT) method to obtain time-frequency domain distribution information of a voltage quality disturbance signal;

[0011] Step S2: based on the time-frequency domain distribution information of the voltage quality disturbance signal obtained in step S1, design a disturbance identification input feature index, and divide the disturbance identification input feature index into three categories: time-domain statistical features, frequency-domain statistical features, and energy features, so as to construct a more comprehensive feature set from the perspectives of amplitude-frequency-phase-energy, etc.;

[0012] Step S3, using a time-frequency domain joint discrimination enhanced sparse autoencoder, extracting deep features from the voltage quality disturbance information by means of time-frequency domain joint training, and then fusing them with the disturbance identification input feature index manually designed in step S2 to obtain a more comprehensive fusion feature set;

[0013] Step S4, using a dynamic feature weighting method based on double-delayed deep deterministic policy gradient (TD3) and state space, action space and reward function to guide TD3 to learn the optimal dynamic input feature adjustment method, which can adaptively adjust the weight of the input feature according to the change trend of the voltage signal in the offline application stage, and weight the fused feature set of step S3;

[0014] Step S5, using the ensemble learner XGBoost based on the Boosting strategy as the disturbance classifier, inputting the fusion feature set obtained in step S3 into XGBoost after being weighted in step S4, so as to realize the identification of the composite disturbance signal.

[0015] Advantages and beneficial effects of the present invention:

[0016] 1. The present invention proposes a voltage quality composite disturbance identification method that considers the fusion of complex signal time-frequency features, which involves time-frequency domain feature extraction, dynamic feature weight adjustment, and disturbance identification classification model construction, effectively improving the accuracy of voltage quality disturbance identification.

[0017] 2. The present invention proposes an improved Hilbert-Huang transform method to analyze the time-frequency domain characteristics of the disturbance signal, alleviating the problems of modal aliasing and endpoint effect caused by the traditional EMD method.

[0018] 3. The present invention proposes a discriminative enhanced sparse autoencoder, which extracts time-frequency domain features from power quality disturbance information by means of joint training in the time-frequency domain, thereby realizing an effective combination of artificial feature extraction and deep learning adaptive feature extraction. It can not only design indicators based on expert experience but also adaptively learn the complex patterns and features of disturbance signals.

[0019] 4. The present invention proposes a dynamic feature weighting method based on double-delayed deep deterministic policy gradient (TD3), and realizes the classification of disturbance signals through the integrated learner XGBoost, thereby improving the efficiency of disturbance identification while reducing the interference of redundant features and noise features. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a flow chart of a voltage quality composite disturbance identification method considering the fusion of complex signal time-frequency characteristics provided by the present invention;

[0021] Figure 2 It is a partial voltage quality disturbance waveform diagram generated by the mathematical model in the present invention;

[0022] Figure 3 This is a diagram showing the effect of signal decomposition by ICEEMDAN in the present invention;

[0023] Figure 4 It is the distribution diagram of different disturbance signals in the two-dimensional space after the features are extracted by IHHT in the present invention;

[0024] Figure 5 is a marginal spectrum curve diagram of different disturbance signals in the present invention;

[0025] Figure 6 It is a comparison chart of the accuracy of disturbance identification of each model under different signal-to-noise ratios in the present invention. DETAILED DESCRIPTION

[0026] The embodiments of the present invention are further described in detail below with reference to the accompanying drawings:

[0027] A voltage quality composite disturbance identification method considering the fusion of complex signal time-frequency characteristics, such as Figure 1 As shown, the following steps are included:

[0028] Step S1: using an improved Hilbert-Huang transform (IHHT) method to obtain the time-frequency domain distribution information of the voltage quality disturbance signal;

[0029] The improved fully integrated empirical mode decomposition with adaptive noise (ICEEMDAN) is used to replace the empirical mode decomposition (EMD) in the traditional Hilbert-Huang transform. The frequency domain cross-correlation coefficient is used to screen out the best intrinsic mode functions (IMFs) with clear features from a series of IMFs obtained by ICEEMDAN for time-frequency domain feature extraction.

[0030] The specific steps of step S1 include:

[0031] The IHHT in step S1 includes three processes: ICEEMDAN modal decomposition, frequency domain cross-correlation coefficient feature screening, and Hilbert transform:

[0032] Step S101: Inject the modal component of the i-th group of white noise into the original time series S through ICEEMDAN to construct a new time series X (i) :

[0033] X (i)=S+β0E1(ω (i) )

[0034] β0=ε0σ(S) / σ(E1(ω (i) ))

[0035] Where: β0 is the noise coefficient, σ(·) is the standard deviation operator, E i (·) is the i-th mode obtained by EMD decomposition, ω (i) The i-th group of white noise (i=1,2,...,I) obeys N(0,1) distribution, I is the total number of modes decomposed, and ε0 is the amplitude of the white noise.

[0036] The working principle of step S101 is as follows: the traditional HHT consists of two parts: empirical mode decomposition (EMD) and Hilbert transform. The ICEEMDAN introduced in IHHT overcomes the problems of mode aliasing and endpoint effect existing in the traditional EMD.

[0037] Step S102: Calculate the local mean to obtain the residual R1 of the first decomposition:

[0038] R1= <M(X (i) ).

[0039] Where: <·. is the average value calculation, M(·) is the local mean value calculation function of the electric energy signal;

[0040] Step S103: When the modal decomposition number of the sampled signal is 1, the first modal component IMF1 is calculated using the original signal S and the residual signal R1, that is:

[0041] IMF1=S-R1

[0042] Step S104: Add white noise to the residual to construct a new signal to be decomposed R1+β1E2(ω (i) ), and the second set of residuals R2 and their modal components IMF2 can be calculated:

[0043] IMF2=R1-R2=R1- <M(R1+β1E2(ω (i) ))>

[0044] Step S105: Similarly, calculate the residual R of the k-th decomposition k and its modal components IMF k :

[0045] R k = <M(R k-1 +β k-1 E k ω (i) ).

[0046] IMF k =R k-1 -R k

[0047] Where: When k ≥ 1, β k =ε0σ(R k )

[0048] Step S106: Repeat step S105 to obtain all modal components and final residuals;

[0049] Step S107: using the frequency domain correlation coefficient as an evaluation criterion, selecting the best IMF component with clear features from the series of IMF components obtained in step S106 for signal reconstruction and time-frequency domain feature extraction;

[0050] The working principle of the frequency domain correlation coefficient in step S107 is: Assuming G x and G y The signal x is i and i The spectral power, f a is the analysis frequency, then the signal x i and i The cross-correlation coefficient in the frequency domain can be expressed as:

[0051]

[0052] Where: G xi and G yi Respectively represent G x and G y Spectral power at frequency i; and Respectively represent G x and G y The average spectral power of .

[0053] Step S108: for the dominant IMF obtained in step S107, obtain the local amplitude and frequency information existing in the processed signal through Hilbert transform;

[0054] The IMF after Hilbert transformation can be expressed as:

[0055]

[0056] Where: H(·) is the Hilbert transform function; P is the Cauchy principal value; IMF k (t) the value of the kth eigenmode function at time t; τ represents the integral variable;

[0057] Step S109: construct an analytical signal Z based on the Hilbert transform result obtained in step S108 k (t) = IMFk (t)+jH[IMF k (t)], the amplitude function a of each IMF k (t) and the phase function θ k (t) can be calculated as:

[0058]

[0059] Where: Z k (t) represents the analytical signal corresponding to the kth IMF; arctan is the inverse tangent function.

[0060] Step S110: According to the phase function θ obtained in step S109 k (t), calculate the instantaneous frequency, Hilbert spectrum and final marginal spectrum of IHHT, and obtain the time-frequency domain distribution information of the voltage quality disturbance signal for the subsequent characteristic index design:

[0061]

[0062] Where: T is the length of the signal; ω k (t) represents the instantaneous frequency of the kth IMF at time t; H(ω,t) represents the Hilbert spectrum, which accurately reflects the changing trend of the amplitude with duration and frequency; h(ω) represents the marginal spectrum, which reflects the contribution of each frequency to the amplitude.

[0063] Step S2: Based on the time-frequency domain distribution information of the voltage quality disturbance signal obtained in step S1, a disturbance identification input feature index is designed, and the disturbance identification input feature index is divided into three categories: time domain statistical features, frequency domain statistical features and energy features, aiming to construct a more comprehensive feature set from the perspectives of amplitude-frequency-phase-energy, etc.

[0064] The disturbance identification input characteristic index design in step S2 includes time domain statistical characteristics, frequency domain statistical characteristics and energy characteristics;

[0065] The specific steps of step S2 include:

[0066] Step S201: Designing the time domain statistical features as: maximum voltage amplitude, minimum voltage amplitude, average voltage amplitude, voltage standard deviation, voltage amplitude peak factor and disturbance duration.

[0067] The working principle of step S201 is that the original disturbance signal contains important information of the PQD waveform, which enhances the classifier's ability to understand and characterize data.

[0068] Step S202: After Hilbert transform, the instantaneous characteristics of each IMF can be obtained, so the following time domain statistical characteristics are added: the maximum instantaneous amplitude of each IMF, the minimum instantaneous amplitude of each IMF, the average instantaneous amplitude of each IMF, the standard deviation of the instantaneous amplitude of each IMF, and the standard deviation of the instantaneous phase of each IMF;

[0069] Step S203: Design the frequency domain statistical features as: the top three amplitudes in the marginal spectrum and their corresponding frequencies, the minimum marginal spectrum amplitude, the average marginal spectrum amplitude, the standard deviation of the marginal spectrum amplitude, the maximum high-frequency marginal spectrum amplitude, the minimum high-frequency marginal spectrum amplitude, the average high-frequency marginal spectrum amplitude, and the standard deviation of the high-frequency marginal spectrum amplitude.

[0070] The working principle of step S203 is that most traditional studies only focus on the instantaneous amplitude-phase characteristics obtained after the Hilbert transform, and ignore the marginal spectrum characteristics.

[0071] Step S204: Designing energy characteristics as: IMF energy and marginal spectrum total energy.

[0072] IMF energy E imf It reflects the strength of the IMF. The specific calculation method is as follows:

[0073]

[0074] Where: IMF k (i) can be the i-th sampling point in the k-th IMF; N is the total number of sampling points.

[0075] The marginal spectrum is obtained by integrating the Hilbert spectrum on the time axis, reflecting the total energy of the signal at each frequency. Therefore, the total marginal spectrum energy is defined as:

[0076]

[0077] Where: N represents the number of sampling points of the marginal spectrum curve; h(ω,i) represents the i-th sampling point of the marginal spectrum curve h(ω).

[0078] The working principle of step S204 is that the energy feature can quantify the energy distribution of the signal in the entire time-frequency domain, thereby reflecting the intensity or severity of the disturbance.

[0079] Step S3, using a time-frequency domain joint discrimination enhanced sparse autoencoder, extracting deep features from the voltage quality disturbance information by means of time-frequency domain joint training, and then fusing them with the disturbance identification input feature index manually designed in step S2 to obtain a more comprehensive fusion feature set;

[0080] The specific steps of step S3 include:

[0081] Step S301: Based on the encoding function of the autoencoder (AE), the input signal is converted into a compressed feature representation through weighting and biasing:

[0082] H=f e (W e ·X+b e )

[0083] X R =f d (W d ·H+b d )

[0084]

[0085] Where: W e and b e Represent the weight and bias from the input layer to the hidden layer. d and b d X, H and X represent the weights and biases from the hidden layer to the output layer respectively. R Represent the original input, hidden layer output and reconstructed data respectively; f e (·) and f d (·) is the activation function; J AE represents the autoencoder reconstruction error; X i and Represent the i-th original input and reconstructed data respectively.

[0086] Step S302: Based on step S301, a sparse autoencoder (SAE) is constructed. The sparsity constraint affects the behavior of the encoder by adding a sparsity penalty term to the loss function. The KL divergence is used to measure the difference between the actual activation value distribution of the hidden layer and the target sparsity:

[0087]

[0088] Where: ρ represents the sparsity parameter of KL divergence; represents the average activation of hidden layer node j; the purpose of the penalty term is to make active hidden nodes close to 1 and inactive nodes close to 0, forcing the model to learn feature representations with better generalization capabilities.

[0089] Furthermore, the autoencoder loss function after introducing the sparsity constraint can be expressed as:

[0090]

[0091] Where: J SAE (W, b) is the loss function after the sparsity constraint is introduced; β is the weight coefficient of KL divergence.

[0092] The working principle of step S302 is that adding a sparsity constraint to the autoencoder can filter out redundant information in the input data, thereby better focusing on the key features of the data.

[0093] Step S303: Add a discriminator module on the basis of step S302 to construct a discriminant enhanced sparse autoencoder (DESAE) assisted by a category label, thereby improving the quality of feature extraction; during training, the hidden layer features of the SAE are input into the discriminator, and the attention to the disturbance signal is enhanced by the category label information; assuming that a training sample {X, Y} is given, Y represents the disturbance category to which the sample belongs; the discriminator calculates the disturbance category information that it may contain:

[0094]

[0095] Where: d(·) represents the discriminator activation function; W g represents the discriminator weight coefficient; b g Represents the discriminator bias; Represents the perturbation category information output by the discriminator.

[0096] Since the composite disturbance signal may contain multiple disturbance label information, the present invention uses the cross entropy loss function to measure the similarity between the discriminator output result and the actual label of the sample, and implements training by minimizing the following loss function:

[0097]

[0098] Where: K D Represents the number of perturbation categories.

[0099] Therefore, the total loss function of DESAE is defined as:

[0100]

[0101] The working principle of step S303 is that since most of the time periods in the voltage signal containing disturbances are normal signals, it is easy to cause the features extracted by SAE dominated by minimizing reconstruction errors to lack discrimination.

[0102] Step S304: Considering that both the time domain and the frequency domain contain important information reflecting the disturbance type, the DESAE designed based on step S303 simultaneously extracts features from the time domain and the frequency domain to improve the accuracy and robustness of disturbance identification by fusing multimodal features in the time and frequency domains.

[0103] The working principle of step S304 is that the time domain DESAE and the frequency domain DESAE are responsible for reconstructing their input signals respectively, and the shared discriminator simultaneously receives the encoder outputs from the time domain and the frequency domain and calculates the disturbance information that may be contained:

[0104]

[0105] H TF =[H T ,H F ]=[f te (W te ·X T +b te ),f fe (W fe ·X F +b fe )]

[0106] Where: H TF Represents the fused time-frequency domain features; X T , H T 、f te (·),W te , b te are the input time domain signal, hidden layer output, encoder activation function, encoder weight matrix, and encoder bias term of time domain DESAE respectively; X F , H F 、f fe (·),W fe , b fe They are the input frequency domain signal, hidden layer output, encoder activation function, encoder weight matrix, and encoder bias term of frequency domain DESAE respectively.

[0107] Step S305: Under the guidance of the reconstruction loss and the discriminator, the gradient of the time-frequency domain DESAE loss function on the encoder weight coefficient and the bias term is calculated as follows:

[0108]

[0109] Where: and Represent the total loss functions of time domain DESAE and frequency domain DESAE respectively; from the gradient formula, we can see that the encoder extracts time-frequency signal features under the joint correction of reconstruction error, KL divergence, and discriminator error.

[0110] Step S306: The gradient of the loss function with respect to the discriminator weight coefficient and the bias term is calculated as follows:

[0111]

[0112] Finally, the minimization of the loss function of the joint DESAE can be achieved by combining the gradient optimization algorithm with back propagation;

[0113] Step S307: The time-frequency domain DESAE trained in steps S305 and S306 can perform the time-frequency domain feature extraction task, and then be fused with the artificial design index obtained in step S2 to obtain a fused feature set, providing a data basis for the subsequent dynamic feature weighting method.

[0114] Step S4, using a dynamic feature weighting method based on double-delayed deep deterministic policy gradient (TD3) and state space, action space and reward function to guide TD3 to learn the optimal dynamic input feature adjustment method, which can adaptively adjust the weight of the input feature according to the change trend of the voltage signal in the offline application stage, and weight the fused feature set of step S3;

[0115] The specific steps of step S4 include:

[0116] The dynamic feature weighting method based on double-delayed deep deterministic policy gradient (TD3) in step S4 comprises the following steps:

[0117] Step S401: converting the complex vibration signal into a simpler envelope signal through the amplitude envelope to accurately reflect the disturbance of the input signal;

[0118] Therefore, the state space of TD3 is set as:

[0119] S={S E ,S F}

[0120] S F =[F1...F 26 ,H T ,H F ]

[0121] Where: S E represents the amplitude envelope after Hilbert transform; S F Represents the complete set of disturbance features constructed by the present invention.

[0122] The working principle of step S401 is that the variables affecting the control action decision are set as the state quantity of the system. The state space of the intelligent agent of the present invention not only needs to reflect the change trend of the disturbance signal, but also needs to include the input features to be adjusted.

[0123] Step S402: Design a set of input feature weights is the action space of the agent. Indicates the Nth F Hyperparameter changes; N F Represents the number of input features.

[0124] The working principle of step S402 is that in order to obtain the maximum immediate reward, the intelligent body will determine the weight coefficient of each input feature according to the observed system state S.

[0125] Step S403: reward function design. The present invention uses the change in classification accuracy before and after the action as a reward function to motivate the model to implement actions that improve the accuracy of composite disturbance identification.

[0126]

[0127] Where: and They represent the classification loss in the current state and after the action of the agent respectively; BCE(·) is the cross entropy loss calculation function; f C (·) represents the classification function; ⊙ represents the Hadamard product operation; A F Represents the action taken by the agent.

[0128] The working principle of step S403 is that the classifier can be used as a medium to quantify the quality of the agent's actions and provide guidance for the agent's action adjustment through classification loss.

[0129] Step S404: Based on the actions, states and reward functions designed in steps S401 to S403, the TD3 reinforcement learning model is guided to dynamically adjust the input feature weights. First, the pruned dual Q learning technique is introduced to calculate the Q value using two independent Critic networks, and the smaller of the two estimates is selected as the target Q value:

[0130]

[0131] Where: y(t) represents the target Q value; Actions combined with random noise to avoid local optima during the agent’s exploration; Represents the target Critic function; Represents the target Critic network parameters.

[0132] The Q value in step S404 is explained as follows: the Q value represents the expected reward that the agent can obtain after taking action a in state s. It can also be regarded as an evaluation or measurement of the quality of taking a specific action in a specific state.

[0133] Step S405: In step S404, the target strategy smoothing technology is introduced. Clipped random noise is added to the target action to make the Critic function smoother for different actions and reduce the possibility of the strategy using the Critic function error. It can be expressed as:

[0134]

[0135] Where: π λ represents the target policy function; represents the parameters of the target policy network; ε~clip(N(0,σ),-M,M) represents the clipped normal distribution noise, and -M and M represent the upper and lower limits of the clipped noise, respectively.

[0136] Step S406: Based on step S405, a random experience replay method is used to randomly extract batches of data from the experience replay pool as training data. Then, the target Q value is substituted into the Bellman equation to calculate the time difference error and loss function, and the gradient descent algorithm is used to minimize the loss function to implement the training of the Critic network:

[0137]

[0138] Where: Q i represents the Critic function; N SA Represents the number of state-action pairs involved in training; θ i represents the critic network parameters; L(θ i ) represents the Critic network loss function; η C Represents the Critic network learning rate.

[0139] Step S407: During the training process described in step S406, a delayed strategy update technique is adopted for the Actor network, even if the update frequency of the Actor network is lower than that of the Critic network; unlike the Critic network update, the Actor network updates parameters by a gradient ascent method:

[0140]

[0141] Where: θ π is the parameter of the Actor network; J(θ π ) is the objective function of the Actor network; is the objective function relative to θ π The gradient of A is the learning rate of the Actor network.

[0142] Step 408: TD3 network parameter update; in order to avoid error divergence, the target network parameters and Slowly approach θ through a soft update strategy π and θ i :

[0143]

[0144] Where: Represents the soft update factor.

[0145] Step S409: TD3 updated in step S408 can perform adaptive adjustment of feature weights online, and weight the fused feature set in step S3 to provide higher quality input information for the subsequent disturbance identification model.

[0146] Step S5, using the ensemble learner XGBoost based on the Boosting strategy as the disturbance classifier, inputting the fusion feature set obtained in step S3 into XGBoost after being weighted in step S4, so as to realize the identification of the composite disturbance signal;

[0147] The specific steps of step S5 include:

[0148] Step S501: Build an XGBoost model. Perform multiple rounds of training, and in each round a new CART base classifier is trained to correct the errors in the previous round of predictions. These errors are the differences between the actual disturbance type and the current model prediction value.

[0149] Step S502: In each round, a new CART base classifier is fitted based on the error of the previous round, and the objective function is minimized by adjusting the values ​​of the leaf nodes, including the classification error and the regularization term, to prevent the model from overfitting:

[0150]

[0151] Where: Represents the cross entropy loss function, indicating the predicted disturbance type and the true disturbance type y t The error between k represents the classification function of the kth base classifier; Ω is the regularization function, expressed as:

[0152]

[0153] Where: represents the total number of leaf nodes in the kth CART; w k represents the weight coefficient of the kth CART leaf node; μ and λ represent the penalty coefficient and regularization coefficient in XGBoost, respectively.

[0154] Step S503: Each CART base classifier uses a greedy algorithm starting from the root node to recursively select split features and thresholds to maximize information gain. By continuously splitting the data set, multiple leaf nodes are formed. The final XGBoost classification model is the weighted sum of multiple base classifiers.

[0155] Step S504: Use the trained XGBoost model to identify power quality disturbances: Step S4 dynamically adjusts the weight of the input features and inputs them into XGBoost. XGBoost outputs the disturbance identification results based on the features, and finally evaluates the model.

[0156] The present invention is further described below in conjunction with specific examples:

[0157] To identify voltage quality composite disturbances, we first need to find a suitable data set and preprocess the collected data set to provide data support for subsequent case analysis. According to the IEEE STD 1159-2019 standard, voltage quality disturbances are divided into seven single disturbance types: swell, transient pulse, sag, flicker, interruption, oscillation, and harmonics. The general mathematical model can be expressed as:

[0158] Dis=Mul(t)·[sin(ωt)+Add(t)]

[0159] Where: ω is the angular frequency of the voltage signal; Mul(t) represents the multiplicative disturbance; Add(t) represents the additive disturbance.

[0160] In order to verify the effectiveness of the method of the present invention, 13 types of signals are classified into 7 single disturbances and 6 composite disturbances, namely voltage swell (S1), voltage sag (S2), voltage interruption (S3), harmonics (S4), oscillation (S5), transient pulse (S6), voltage flicker (S7), swell and harmonic composite disturbance (S8), sag and harmonic composite disturbance (S9), voltage interruption and harmonic composite disturbance (S10), voltage sag and oscillation composite disturbance (S11), voltage flicker and oscillation composite disturbance (S12), voltage swell, harmonic and oscillation triple composite disturbance (S13). The sample capacity of each disturbance type is 200. Some disturbance signals are as follows: Figure 2 shown.

[0161] In order to verify the ability of the present invention to distinguish the frequency domain features extracted based on IHHT in the classification of power quality disturbances. The effect after decomposition by ICEEMDAN is as follows Figure 3 As shown, it can be seen that each IMF component gradually decreases from high frequency to low frequency, showing the characteristics of different frequency components, and the residual component is a smooth trend signal, indicating that the main high-frequency and medium-frequency components have been effectively extracted. In addition, the energy distribution of each IMF is reasonable, and the correlation between them is low, which ensures the independence and orthogonality of the components, and proves the practicality of the signal decomposition method proposed in the present invention. The IHHT method is used to process each disturbance signal, and then the feature index constructed in step S2 is used for feature extraction, and the features are visualized in two-dimensional space by the principal component analysis method, as shown in FIG. Figure 4As shown. It can be seen that the indicators designed by the present invention basically effectively distinguish the characteristics of different disturbance signals in two-dimensional space. In addition, Figure 5 The marginal spectrum energy of shows an obvious fluctuation trend, which shows that it is reasonable to use it as a feature index in the present invention. However, the distribution of S1, S9 and S13 in the feature space is still relatively close, which also verifies the necessity of further feature extraction by DESAE in the present invention.

[0162] In order to demonstrate the superiority of the IHHT proposed in this invention, the features extracted by HHT and IHHT of 13 disturbance signals are input into the XGBoost model for power quality disturbance identification, and the accuracy comparison results of different signal-to-noise ratios are shown in Table 1. IHHT uses ICEEMDAN to make the decomposed signal quality better than HHT, with better feature quality and noise resistance. Therefore, its classification accuracy is higher than HHT in various noise environments.

[0163] Table 1 Comparison of classification accuracy of features obtained by HHT and IHHT under different noise conditions (%)

[0164]

[0165] In order to prove the effectiveness of the DESAE developed by the present invention, the deep features extracted by the autoencoder (AE), sparse autoencoder (SAE) and DESAE are respectively input into the XGBoost model for disturbance identification, and the accuracy comparison results of different signal-to-noise ratios are shown in Table 2. It can be seen that the DESAE proposed by the present invention shows a classification accuracy that is significantly better than that of the traditional method under various signal-to-noise ratio conditions. Under high signal-to-noise ratios (90dB and 80dB), the accuracy of DESAE reaches 96.74% and 96.58%, respectively, which is significantly higher than the corresponding results of AE and SAE; while under lower signal-to-noise ratios (70dB and 60dB), DESAE still maintains excellent performance, with accuracy rates of 96.46% and 96.14%, respectively. This shows that DESAE can not only effectively extract signal features under ideal conditions, but also maintain robust performance in noisy environments, which fully proves the significant superiority of DESAE developed by the present invention in the extraction and classification of disturbance signal features.

[0166] Table 2 Comparison of classification accuracy of different deep learning methods (%)

[0167]

[0168]

[0169] Furthermore, the classification effect of the dynamic weighted XGBoost proposed in the present invention is demonstrated by examples, which is called TD3-XGBoost in the present invention, and is compared with the back propagation neural network (BPNN), convolutional neural network (CNN) and adaptive boosting (AdaBoost) in boosting ensemble learning. The input of each classification model is the time-frequency domain feature set after the features obtained in step S2 and the features obtained in step S3 are fused. The results of each classifier are shown in Figure 6 As shown. It can be seen that BPNN performs well under noise-free conditions, but as the noise increases, the accuracy decreases significantly, showing sensitivity to noise; LSTM and CNN perform well under high signal-to-noise ratio conditions and have certain noise resistance, but LSTM is slightly less resistant to noise than CNN; AdaBoost has a high accuracy rate under high signal-to-noise ratio, but is more sensitive to noise; TD3-XGBoost proposed in the present invention performs well under all signal-to-noise ratio conditions, and the accuracy rate remains at a high level, especially under low signal-to-noise ratio, it can still maintain a high accuracy rate, showing its excellent noise resistance and robustness, and proving its significant superiority in disturbance signal feature extraction and classification.

[0170] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0171] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0172] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0173] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

Claims

1. A voltage quality composite disturbance identification method considering the fusion of complex signal time-frequency characteristics, characterized by: The following steps are involved: Step S1, using an improved Hilbert-Huang transform (IHHT) method to obtain time-frequency domain distribution information of a voltage quality disturbance signal; Step S2: based on the time-frequency domain distribution information of the voltage quality disturbance signal obtained in step S1, design a disturbance identification input feature index, and divide the disturbance identification input feature index into three categories: time-domain statistical features, frequency-domain statistical features, and energy features, so as to construct a more comprehensive feature set from the perspectives of amplitude-frequency-phase-energy, etc.; Step S3, using a time-frequency domain joint discrimination enhanced sparse autoencoder, extracting deep features from the voltage quality disturbance information by means of time-frequency domain joint training, and then fusing them with the disturbance identification input feature index manually designed in step S2 to obtain a more comprehensive fusion feature set; Step S4, using a dynamic feature weighting method based on double-delayed deep deterministic policy gradient (TD3) and state space, action space and reward function to guide TD3 to learn the optimal dynamic input feature adjustment method, which can adaptively adjust the weight of the input feature according to the change trend of the voltage signal in the offline application stage, and weight the fused feature set of step S3; Step S5, using the ensemble learner XGBoost based on the Boosting strategy as the disturbance classifier, inputting the fusion feature set obtained in step S3 into XGBoost after being weighted in step S4, so as to realize the identification of the composite disturbance signal.

2. The voltage quality composite disturbance identification method considering the fusion of complex signal time-frequency characteristics according to claim 1 is characterized by: The specific steps of step S1 include: Step S101: Inject the modal component of the i-th group of white noise into the original time series S through ICEEMDAN to construct a new time series X (i) : X (i) =S+β0E1(ω (i) ) β0=ε0σ(S) / σ(E1(ω (i) )) Where: β0 is the noise coefficient, σ(·) is the standard deviation operator, E i (·) is the i-th mode obtained by EMD decomposition, ω (i) The i-th group of white noise (i=1,2,…,I) follows N(0,1) distribution, I is the total number of decomposed modes, and ε0 is the amplitude of the white noise; Step S102: Calculate the local mean to obtain the residual R1 of the first decomposition: R1= <M(X (i) ) Where: <·> is the average value calculation, M(·) is the local mean calculation function of the electric energy signal; Step S103: When the modal decomposition number of the sampled signal is 1, the first modal component IMF1 is calculated using the original signal S and the residual signal R1, that is: IMF1=S-R1 Step S104: Add white noise to the residual to construct a new signal to be decomposed R1+β1E2(ω (i) ), and the second set of residuals R2 and their modal components IMF2 can be calculated: IMF2=R1-R2=R1-<M(R1+β1E2(ω (i) )) Step S105: Similarly, calculate the residual R of the k-th decomposition k and its modal components IMF k : R k = <M(R k-1 +b k-1 E k oh (i) ) IMF k =R k-1 -R k Where: When k ≥ 1, β k =ε0σ(R k ) Step S106: Repeat step S105 to obtain all modal components and final residuals; Step S107: using the frequency domain correlation coefficient as an evaluation criterion, selecting the best IMF component with clear features from the series of IMF components obtained in step S106 for signal reconstruction and time-frequency domain feature extraction; Step S108: for the dominant IMF obtained in step S107, obtain the local amplitude and frequency information existing in the processed signal through Hilbert transform; The IMF after Hilbert transformation can be expressed as: Where: H(·) is the Hilbert transform function; P is the Cauchy principal value; IMF k (t) the value of the kth eigenmode function at time t; τ represents the integral variable; Step S109: construct an analytical signal Z based on the Hilbert transform result obtained in step S108 k (t) = IMF k (t)+jH[IMF k (t)], the amplitude function a of each IMF k (t) and the phase function θ k (t) can be calculated as: Where: Z k (t) represents the analytical signal corresponding to the kth IMF; arctan is the inverse tangent function; Step S110: According to the phase function θ obtained in step S109 k (t), calculate the instantaneous frequency, Hilbert spectrum and final marginal spectrum of IHHT, and obtain the time-frequency domain distribution information of the voltage quality disturbance signal for the subsequent characteristic index design: Where: T is the length of the signal; ω k (t) represents the instantaneous frequency of the kth IMF at time t; H(ω,t) represents the Hilbert spectrum, which accurately reflects the changing trend of the amplitude with duration and frequency; h(ω) represents the marginal spectrum, which reflects the contribution of each frequency to the amplitude; K represents the number of intrinsic mode functions.

3. The voltage quality composite disturbance identification method considering the fusion of complex signal time-frequency characteristics according to claim 1 is characterized by: The specific steps of step S2 include: Step S201: Designing the time domain statistical features as: maximum voltage amplitude, minimum voltage amplitude, average voltage amplitude, voltage standard deviation, voltage amplitude peak factor and disturbance duration; Step S202: After Hilbert transform, the instantaneous characteristics of each IMF can be obtained, and the following time domain statistical characteristics are newly added: the maximum instantaneous amplitude of each IMF, the minimum instantaneous amplitude of each IMF, the average instantaneous amplitude of each IMF, the standard deviation of the instantaneous amplitude of each IMF, and the standard deviation of the instantaneous phase of each IMF; Step S203: Design the frequency domain statistical features as: the top three amplitudes in the marginal spectrum and their corresponding frequencies, the minimum marginal spectrum amplitude, the average marginal spectrum amplitude, the standard deviation of the marginal spectrum amplitude, the maximum high-frequency marginal spectrum amplitude, the minimum high-frequency marginal spectrum amplitude, the average high-frequency marginal spectrum amplitude, and the standard deviation of the high-frequency marginal spectrum amplitude; Step S204: Designing energy characteristics as: IMF energy and marginal spectrum total energy; IMF energy E imf It reflects the strength of the IMF. The specific calculation method is as follows: Where: IMF k (i) can be the i-th sampling point in the k-th IMF; N is the total number of sampling points; The marginal spectrum is obtained by integrating the Hilbert spectrum on the time axis, reflecting the total energy of the signal at each frequency. Therefore, the total marginal spectrum energy is defined as: Where: N represents the number of sampling points of the marginal spectrum curve; h(ω,i) represents the i-th sampling point of the marginal spectrum curve h(ω).

4. The voltage quality composite disturbance identification method considering the fusion of complex signal time-frequency characteristics according to claim 1 is characterized by: The specific steps of step S3 include: Step S301: Based on the encoding function of the autoencoder AE, the input signal is converted into a compressed feature representation through weighting and biasing: H=f e (W e ·X+b e ) X R =f d (W d ·H+b d ) Where: W e and b e Respectively represent the weight and bias from the input layer to the hidden layer; W d and b d represent the weights and biases from the hidden layer to the output layer; X, H, and X R Represent the original input, hidden layer output and reconstructed data respectively; f e (·) and f d (·) is the activation function; J AE represents the autoencoder reconstruction error; X i and Represent the i-th original input and reconstructed data respectively; Step S302: Based on step S301, a sparse autoencoder SAE is constructed. The sparsity constraint affects the behavior of the encoder by adding a sparsity penalty term to the loss function. The KL divergence is used to measure the difference between the actual activation value distribution of the hidden layer and the target sparsity: Where: ρ represents the sparsity parameter of KL divergence; represents the average activation of hidden layer node j; the purpose of the penalty term is to make active hidden nodes close to 1 and inactive nodes close to 0, forcing the model to learn feature representations with better generalization capabilities; Furthermore, the loss function of the autoencoder after introducing the sparsity constraint is expressed as: Where: J SAE (W,b) is the loss function after the sparsity constraint is introduced; β is the weight coefficient of KL divergence; Step S303: Add a discriminator module based on step S302 to construct a class label-assisted discriminant enhanced sparse autoencoder DESAE; during training, the hidden layer features of SAE are input into the discriminator, and the class label information is used to assist in enhancing the attention to the disturbance signal; assuming that a training sample {X, Y} is given, Y represents the disturbance category to which the sample belongs; the discriminator calculates the disturbance category information it may contain: Where: d(·) represents the discriminator activation function; W g represents the discriminator weight coefficient; b g Represents the discriminator bias; Represents the perturbation category information output by the discriminator; The cross entropy loss function is used to measure the similarity between the discriminator output and the actual label of the sample, and training is achieved by minimizing the following loss function: Where: K D represents the number of disturbance categories; The total loss function of DESAE is defined as: Step S304: Considering that both the time domain and the frequency domain contain important information reflecting the disturbance type, the DESAE designed in step S303 simultaneously extracts features in the time domain and the frequency domain, so as to improve the accuracy and robustness of disturbance identification by fusing multimodal features in the time and frequency domains; The time domain DESAE and frequency domain DESAE are responsible for reconstructing their input signals respectively. The shared discriminator accepts the encoder outputs from the time domain and frequency domain at the same time and calculates the possible perturbation information: H TF =[H T ,H F ]=[f te (W te ·X T +b te ),f fe (W fe ·X F +b fe )] Where: H TF represents the fused time-frequency domain features; X T , H T 、f te (·),W te , b te are the input time domain signal, hidden layer output, encoder activation function, encoder weight matrix, and encoder bias term of time domain DESAE respectively; X F , H F 、f fe (·),W fe , b fe They are the input frequency domain signal, hidden layer output, encoder activation function, encoder weight matrix, and encoder bias term of frequency domain DESAE respectively; Step S305: Under the guidance of the reconstruction loss and the discriminator, the gradient of the time-frequency domain DESAE loss function on the encoder weight coefficient and the bias term is calculated as follows: Where: and Represent the total loss functions of time domain DESAE and frequency domain DESAE respectively; From the gradient formula, we can know that the encoder extracts the time-frequency signal features under the joint correction of reconstruction error, KL divergence and discriminator error; Step S306: The gradient of the loss function with respect to the discriminator weight coefficient and the bias term is calculated as follows: Finally, the minimization of the loss function of the joint DESAE can be achieved by combining the gradient optimization algorithm with back propagation; Step S307: The time-frequency domain DESAE trained in steps S305 and S306 can perform the time-frequency domain feature extraction task, and then be fused with the artificial design index obtained in step S2 to obtain a fused feature set, providing a data basis for the subsequent dynamic feature weighting method.

5. The voltage quality composite disturbance identification method considering the fusion of complex signal time-frequency characteristics according to claim 1 is characterized by: The specific steps of step S4 include: Step S401: converting the complex vibration signal into a simpler envelope signal through the amplitude envelope to accurately reflect the disturbance of the input signal; Therefore, the state space of TD3 is set as: S={S E ,S F } S F =[F1...F 26 ,H T ,H F ] Where: S E represents the amplitude envelope after Hilbert transform; S F represents the complete set of disturbance features constructed by the present invention; Step S402: Design a set of input feature weights is the action space of the agent; Indicates the Nth F Hyperparameter changes; N F Represents the number of input features; Step S403: reward function design: using the change in classification accuracy before and after the action as a reward function to motivate the model to implement actions that improve the accuracy of compound disturbance identification; Where: and They represent the classification loss in the current state and after the action of the agent respectively; BCE(·) is the cross entropy loss calculation function; f C (·) represents the classification function; ⊙ represents the Hadamard product operation; A F represents the action taken by the agent; Step S404: Based on the actions, states, and reward functions designed in steps S401 to S403, the TD3 reinforcement learning model is guided to dynamically adjust the input feature weights. First, the pruned dual Q learning technique is introduced to calculate the Q value using two independent Critic networks, and the smaller of the two estimates is selected as the target Q value: Where: y(t) represents the target Q value; Actions combined with random noise to avoid local optima during the agent’s exploration; Represents the target Critic function; Represents the target Critic network parameters; The Q value in step S404 is explained as follows: the Q value represents the expected reward that the agent can obtain after taking action a in state s; it can also be regarded as an evaluation or measurement of the quality of taking a specific action in a specific state; Step S405: In step S404, the target strategy smoothing technology is introduced to add clipped random noise to the target action, so that the changes of the Critic function to different actions are smoother, and the possibility of the strategy using the Critic function error is reduced. It can be expressed as: Where: π λ represents the target policy function; represents the parameters of the target policy network; ε~clip(N(0,σ),-M,M) represents the clipped normal distribution noise, -M and M represent the upper and lower limits of the clipped noise respectively; Step S406: Based on step S405, a random experience replay method is used to randomly extract batches of data from the experience replay pool as training data. Then, the target Q value is substituted into the Bellman equation to calculate the temporal difference error and loss function, and the gradient descent algorithm is used to minimize the loss function to implement the training of the Critic network: Where: Q i represents the Critic function; N SA Represents the number of state-action pairs involved in training; θ i represents the critic network parameters; L(θ i ) represents the Critic network loss function; η C Represents the Critic network learning rate; Step S407: During the training process described in step S406, a delayed strategy update technique is adopted for the Actor network, even if the update frequency of the Actor network is lower than that of the Critic network; unlike the Critic network update, the Actor network updates parameters by a gradient ascent method: Where: θ π is the parameter of the Actor network; J(θ π ) is the objective function of the Actor network; is the objective function relative to θ π The gradient of A is the learning rate of the Actor network; Step 408: TD3 network parameter update; in order to avoid error divergence, the target network parameters and Slowly approach θ through a soft update strategy π and θ i : Where: represents the soft update factor; Step S409: TD3 updated in step S408 can perform adaptive adjustment of feature weights online, and weight the fused feature set in step S3 to provide higher quality input information for the subsequent disturbance identification model.

6. The voltage quality composite disturbance identification method considering the fusion of complex signal time-frequency characteristics according to claim 1 is characterized by: The specific steps of step S5 include: Step S501: Build an XGBoost model; perform multiple rounds of training, where a new CART base classifier is trained in each round to correct the errors in the previous round of predictions; these errors are the differences between the actual disturbance type and the current model prediction value; Step S502: In each round, a new CART base classifier is fitted based on the error of the previous round, and the objective function is minimized by adjusting the values ​​of the leaf nodes, including the classification error and the regularization term, to prevent the model from overfitting: Where: Represents the cross entropy loss function, indicating the predicted disturbance type and the true disturbance type y t The error between k represents the classification function of the kth base classifier; Ω is the regularization function, expressed as: Where: represents the total number of leaf nodes in the kth CART; w k represents the weight coefficient of the kth CART leaf node; μ and λ represent the penalty coefficient and regularization coefficient in XGBoost respectively; Step S503: Each CART base classifier uses a greedy algorithm starting from the root node to recursively select split features and thresholds to maximize information gain; by continuously splitting the data set to form multiple leaf nodes, the final XGBoost classification model is the weighted sum of multiple base classifiers; Step S504: Use the trained XGBoost model to identify power quality disturbances: Step S4 dynamically adjusts the weight of the input features and inputs them into XGBoost. XGBoost outputs the disturbance identification results based on the features, and finally evaluates the model.

Images