Fault detection method and system for virtual power plant microgrid

Abstract

The application discloses a fault detection method and system for a virtual power plant microgrid, and comprises the following steps: step 1, acquiring three-phase voltage and current historical data of a power grid bus as input signals; step 2, decomposing the input signals into multiple modes by using a variational mode decomposition method; step 3, calculating the energy of each mode and the correlation coefficient between each mode and the input signals, establishing a phase asymmetry index and a disturbance frequency index, and using the weighted sum of the energy of each mode and the correlation coefficient, the weighted sum of the phase asymmetry index and the correlation coefficient, and the weighted sum of the disturbance frequency index and the correlation coefficient to constitute a training feature vector; step 4, training a fault detection model by using the training feature vector; obtaining a feature vector according to real-time data of three-phase voltage and three-phase current of the power grid bus according to steps 1 to 3; and outputting a fault probability by the fault detection model according to the feature vector, and there is a fault when the fault probability is greater than a set threshold. The application has the advantages of high sensitivity to faults, high robustness and the like.

Description

Technical Field

[0001] The present invention belongs to the technical field of virtual power plant fault detection, and in particular relates to a method and system for virtual power plant microgrid fault detection. Background Art

[0002] A virtual power plant (VPP) is a system that aggregates and coordinates the management of distributed power sources, energy storage, electric vehicles, and controllable loads through communication and control technologies. Its core goal is to enhance the aggregation capabilities of distributed resources and their participation in the grid, thereby increasing system flexibility and stability. Within the VPP architecture, microgrids, as the primary controllable unit, undertake the critical tasks of local energy scheduling and autonomous operation. However, due to the complex types of power sources connected to microgrids and their fluctuating operating states, faults can occur unexpectedly and unpredictably. Failure to detect them promptly can impact the operational safety and dispatch reliability of the entire VPP. Therefore, developing an efficient, real-time, and knowledge-free fault detection method for microgrids is crucial for the safe operation of VPPs.

[0003] Existing microgrid fault detection technologies can be primarily categorized into model-based and data-driven approaches. Model-based approaches rely on the microgrid's topology, electrical parameters, and control strategies to construct equivalent mathematical models or physical mechanism models for fault identification. However, in practice, microgrids are complex in structure and their operating states are highly variable, making modeling challenging. Furthermore, these models are sensitive to parameter perturbations and structural changes, resulting in unstable detection accuracy. Furthermore, these approaches typically require extensive manual parameter tuning, resulting in poor adaptability and difficulty meeting the rapid deployment requirements of virtual power plants with diverse resource access. Data-driven approaches collect electrical data and utilize machine learning and signal analysis to achieve state identification and anomaly detection. While this improves adaptability to a certain extent, existing approaches still suffer from the following challenges: They often rely on supervised learning algorithms, requiring the predefined fault types and the collection of large numbers of labeled samples, making them difficult to adapt to unknown or rare faults in practice. Furthermore, feature extraction often relies on traditional transforms such as Fourier transforms and wavelets, which have limited ability to extract non-stationary electrical signals.

[0004] In summary, the existing microgrid fault detection methods still have problems such as insufficient model generalization ability, weak feature extraction ability, and strong dependence on annotation in the virtual power plant scenario. A new online fault detection method with high robustness, low dependence, and strong adaptability is urgently needed to solve these problems. Summary of the Invention

[0005] In order to address the deficiencies in the prior art, the purpose of the present invention is to provide a method and system for virtual power plant microgrid fault detection, which can quickly and accurately detect grid faults without building a grid model, thereby ensuring the safe operation of the grid.

[0006] The present invention adopts the following technical solutions.

[0007] The present invention proposes a fault detection method for a virtual power plant microgrid, comprising:

[0008] Step 1: Obtain historical data of three-phase voltage and three-phase current of the power grid bus as input signals;

[0009] Step 2: Decompose the input signal into multiple modes using the variational mode decomposition method;

[0010] Step 3: Calculate the energy of each mode and the correlation coefficient between each mode and the input signal, establish a phase asymmetry index and a disturbance frequency index, and form a training feature vector with the weighted sum of the energy and correlation coefficient of each mode, the weighted sum of the phase asymmetry index and the correlation coefficient, and the weighted sum of the disturbance frequency index and the correlation coefficient;

[0011] Step 4: Use the training feature vector to train the established fault detection model; according to the real-time data of the three-phase voltage and three-phase current of the power grid bus, obtain the feature vector according to steps 1 to 3; based on the feature vector, the fault detection model outputs the fault probability. If the fault probability is greater than the set threshold, a fault exists; otherwise, no fault exists.

[0012] In step 1, historical data of the three-phase voltage and three-phase current of the power grid bus are obtained; historical instantaneous values ​​of the three-phase voltage and three-phase current in the sliding window are extracted to form a three-phase voltage sequence and a three-phase current sequence respectively; symmetrical component transformation is performed on the three-phase voltage sequence and the three-phase current sequence to obtain positive sequence components, negative sequence components and zero sequence components as multiple input signals.

[0013] In step 2, the variational mode decomposition method is used to decompose each input signal in turn, and the augmented Lagrangian function is introduced and the alternating direction multiplier method is used to obtain the modal , the frequency bands of each mode do not overlap and the bandwidth is minimized.

[0014] In step 3, the modal Perform Hilbert transform to get the mode The analytical signal form of as follows:

[0015]

[0016] Where, 、 Mode The instantaneous amplitude and instantaneous phase of is the imaginary number symbol, For modal The Hilbert transform of

[0017] Extract the instantaneous phase of each mode from the analytical signal form.

[0018] The absolute value of the deviation between the mean of the instantaneous phase of each mode in the positive region and the mean in the negative region is used as the phase asymmetry index, as follows:

[0019]

[0020] Where, For modal The phase asymmetry index, 、 Mode The time sample set in the positive and negative areas, For collection The number of elements contained in For collection The number of elements contained in .

[0021] The instantaneous phase reversal rate of each mode is used as the disturbance frequency index, as follows:

[0022]

[0023] Where, For modal The disturbance frequency index, is the time window length, that is, the length of the sequence, is the modal signal The instantaneous frequency, It is a logical judgment function. If the conditions in the brackets are met, it is 1, otherwise it is 0.

[0024] The correlation coefficient between each mode and each input signal , as shown below:

[0025]

[0026] Where, 、 Mode and input signal The variance of Indicates the first Positive voltage sequence component , the first Negative sequence voltage component , the first voltage zero-sequence component , the first positive sequence current component , the first Negative sequence current component and the first zero-sequence current component , To calculate the covariance.

[0027] In step 4, a fault detection model is established based on a deep feedforward neural network and trained using the training feature vector. The training objective is to minimize the composite loss function and use the Adam optimizer for backpropagation optimization to train the fault detection model.

[0028] According to the real-time data of the three-phase voltage and three-phase current of the power grid bus, the characteristic vector is obtained according to steps 1 to 3;

[0029] According to the feature vector, the output layer of the trained fault detection model uses the Sigmoid function to output the system failure probability. If the failure probability is greater than the set threshold, it is determined that a fault exists; if the failure probability is less than or equal to the set threshold, it is determined that no fault exists.

[0030] Composite loss function as follows:

[0031]

[0032] Where, is the sample size, is the weight of the contrast loss, The weight of KL divergence, For the The uncertainty perception weight of the samples, For the Adaptive focal loss for samples, For the The contrast loss of samples, For the KL divergence of samples;

[0033]

[0034] Where, The output of the Sigmoid function used in the output layer;

[0035]

[0036] Where, For the The adaptively adjusted class weights of samples, To control the weighted degree coefficient of difficult samples, For the The true label corresponding to each sample is 0 or 1;

[0037]

[0038] Where, For contrasting boundaries;

[0039]

[0040] Where, is the temperature parameter, is a small constant that is numerically stable.

[0041] The present invention also proposes a fault detection system for a virtual power plant microgrid, comprising:

[0042] An acquisition module is used to obtain historical data of the three-phase voltage and three-phase current of the power grid bus as input signals;

[0043] A modal decomposition module is used to decompose the input signal into multiple modes using a variational modal decomposition method;

[0044] The eigenvector calculation module is used to calculate the energy of each mode and the correlation coefficient between each mode and the input signal, establish the phase asymmetry index and the disturbance frequency index, and form the training eigenvector with the weighted sum of the energy and correlation coefficient of each mode, the weighted sum of the phase asymmetry index and the correlation coefficient, and the weighted sum of the disturbance frequency index and the correlation coefficient;

[0045] A model building module, used for training the established fault detection model using the training feature vector;

[0046] The fault detection module is used to call the eigenvector calculation module to obtain the eigenvector based on the real-time data of the three-phase voltage and three-phase current of the power grid bus. Based on the eigenvector, the fault detection model outputs the fault probability. If the fault probability is greater than the set threshold, a fault exists; otherwise, no fault exists.

[0047] The beneficial effects of the present invention, compared to existing technologies, include at least one aspect: the proposed method uses the three-phase voltage and current signals at the microgrid bus terminal as input, combines signal decomposition technology with supervised learning methods, and constructs an online detection mechanism for abnormal operating conditions. First, variational modal decomposition is used to perform nonlinear and non-stationary time series decomposition on the bus signal to extract multi-scale modal components. Then, a multidimensional feature vector is constructed based on two new features: modal energy distribution characteristics and phase asymmetry measurement and phase flip rate. Finally, a deep neural network model with an improved loss function is trained to accurately identify abnormal samples during operation.

[0048] The detection method only relies on the single-point sampling signal at the bus end, reducing the complexity of multi-point synchronous acquisition and communication overhead. It is suitable for distributed deployment and unified monitoring of multiple microgrid units in virtual power plants.

[0049] The method developed in the present invention has data-driven characteristics and does not involve complex grid topology and mechanism. It can be conveniently used for fault diagnosis of microgrids with various topologies without the need to construct different grid models in different situations like the model-based method. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a flow chart of the microgrid fault detection method according to the present invention;

[0051] Figure 2 This is a loss curve diagram of the training process of the microgrid fault detection method according to the present invention;

[0052] Figure 3 This is a confusion matrix diagram for test data detection used in the microgrid fault detection method of the present invention. DETAILED DESCRIPTION

[0053] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. The embodiments described in this application are only part of the embodiments of the present invention, not all of them. Based on the spirit of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0054] The present invention proposes a fault detection method for a virtual power plant microgrid, such as Figure 1 As shown, including:

[0055] Step 1: obtaining positive sequence components, negative sequence components, and zero sequence components of historical data of three-phase voltage and three-phase current of a power grid bus as multiple input signals.

[0056] It is worth noting that the present invention obtains historical data of the three-phase voltage and three-phase current of the power grid bus and performs symmetrical component transformation to obtain multiple input signals, which is a non-restrictive and preferred choice.

[0057] Specifically, step 1 includes:

[0058] Step 1.1, obtain the historical data of three-phase voltage and three-phase current of the power grid bus;

[0059] Obtain the historical voltage and current data at the busbar where the distribution line in the power grid is connected, and perform symmetrical component transformation. The historical voltage and current obtained are three-phase AC signals, which are signed instantaneous value sequences obtained through sampling. In order to capture the fault transition process and high-order harmonic components, the sampling frequency On the order of 5-10kHz.

[0060] Step 1.2, extract the historical instantaneous values ​​of the three-phase voltage and three-phase current within the sliding window to form a three-phase voltage sequence and a three-phase current sequence respectively;

[0061] Use sliding windows to extract data to facilitate real-time online operation of the algorithm and obtain the three-phase voltage sequence 、 、 , three-phase current sequence 、 、 ,in, , , is the sliding window length.

[0062] Step 1.3: Perform symmetrical component transformation on the three-phase voltage sequence and the three-phase current sequence to highlight the unbalanced characteristics and enhance the separability of the fault information; as shown below:

[0063]

[0064]

[0065]

[0066] Where, The first A voltage positive sequence component, The first A negative sequence voltage component, The first A voltage zero-sequence component, , ;

[0067] Similarly, we can transform the first positive sequence current component , the first Negative sequence current component , the first zero-sequence current component .

[0068] by 、 、 、 、 and as multiple input signals in subsequent steps.

[0069] Step 2: Decompose each input signal into multiple modes using the variational mode decomposition method; the frequency bands of each mode do not overlap and the bandwidth is minimized.

[0070] The variational mode decomposition method is used to decompose each input signal in turn, and the augmented Lagrangian function is introduced and the alternating direction multiplier method is used to iteratively update the solution, and finally the modal , , use the variational mode decomposition algorithm to decompose the components obtained in step 1 to obtain several modes containing information in different frequency bands, so as to reconstruct the original signal and minimize the bandwidth of each mode. Then remove several high-frequency modes to reduce noise interference, and finally obtain multiple modes.

[0071] It is worth noting that the present invention adopts the variational mode decomposition method to decompose each input signal in turn, which is a non-restrictive and preferred choice.

[0072] Step 3: Calculate the energy, phase asymmetry index, and instantaneous phase flip rate of each mode, as well as the correlation coefficient between each mode and each input signal; and construct a training feature vector with the weighted sum of the energy and correlation coefficient of each mode, the weighted sum of the phase asymmetry index and the correlation coefficient of each mode, and the weighted sum of the instantaneous phase flip rate and the correlation coefficient of each mode.

[0073] Specifically, step 3 includes:

[0074] Step 3.1, calculate the energy of each mode in the sliding window ,as follows:

[0075]

[0076] Step 3.2, for the modal Perform Hilbert transform to extract the instantaneous phase of each mode;

[0077] Modal Perform Hilbert transform to get the mode The analytical signal form of as follows:

[0078]

[0079] Where, 、 Mode The instantaneous amplitude and instantaneous phase of is the imaginary number symbol, For modal The Hilbert transform of is defined as:

[0080]

[0081] Where, is the Cauchy principal value integral;

[0082] Step 3.3: The absolute value of the deviation between the mean of the instantaneous phase of each mode in the positive region and the mean in the negative region is used as the phase asymmetry index, as follows:

[0083]

[0084] Where, For modal The phase asymmetry index, 、 Mode The time sample set in the positive and negative areas, For collection The number of elements contained in For collection The number of elements contained in .

[0085] In step 3.4, the instantaneous phase reversal rate (MPR) of each mode is used as the disturbance frequency indicator, as follows:

[0086]

[0087] Where, For modal The disturbance frequency index, is the time window length, that is, the length of the sequence, is the modal signal The instantaneous frequency, It is a logical judgment function, which is 1 if the conditions in the brackets are met, otherwise it is 0;

[0088] In the embodiment, a larger MPR indicates that the phase evolution process of the modal signal is more unstable, and there are frequent disturbances or non-stationary faults. Under normal operating conditions, Most of the changes are smooth, with very little sign change and MPR close to zero; abnormal conditions such as short circuit, grounding, and impact disturbance will cause With violent fluctuations, the number of sign reversals increases, so the MPR increases significantly.

[0089] Building on traditional modal energy features, this invention introduces two new features: phase asymmetry and phase reversal rate, based on phase dynamics. These features effectively improve the accuracy and robustness of fault detection. The phase asymmetry metric compares the instantaneous phase mean of the positive and negative half-cycles of the modal signal, sensitively capturing imbalances and structural offsets in grid operation and significantly discriminating against minor asymmetric faults (such as single-phase grounding). The phase reversal rate measures the stability of the phase evolution direction by counting the number of instantaneous frequency sign reversals, demonstrating high sensitivity to non-stationary faults such as impact short circuits and high-frequency disturbances.

[0090] Step 3.5: To enhance the distinguishability between features and faults, calculate the correlation coefficient between each mode and each input signal. , as significance weights, as follows:

[0091]

[0092] Where, 、 Mode and input signal The variance of express 、 、 、 、 and , To calculate the covariance;

[0093] Step 3.6: The training feature vector is constructed by taking the weighted sum of the energy and correlation coefficient of each mode, the weighted sum of the phase asymmetry index and the correlation coefficient, and the weighted sum of the disturbance frequency index and the correlation coefficient. ;

[0094] In the embodiment, the training feature vector .

[0095] Step 3.7, training feature vector Perform normalization processing;

[0096] The training feature vector is normalized using zero mean and unit variance Normalization is performed to improve the model training effect; for each feature dimension , calculate the mean of the training feature vector and standard deviation , and then mapped to the normalized training feature vector , as shown below:

[0097]

[0098] Step 4: Establish a fault detection model based on a deep feedforward neural network and train the fault detection model using the training feature vector. Repeat steps 1 to 3 to obtain the feature vector based on the real-time data of the three-phase voltage and three-phase current of the power grid bus. Based on the feature vector, use the trained fault detection model to output the system fault probability value. If the fault probability value is greater than the set threshold, it is determined that a fault exists. If the fault probability value is less than or equal to the set threshold, it is determined that no fault exists.

[0099] Specifically, step 4 includes:

[0100] Step 4.1: Design and initialize a deep feedforward neural network to establish a fault detection model. Multiple linear layers and nonlinear activation functions (ReLU) are stacked alternately, including Dropout to suppress overfitting. The output layer uses the Sigmoid function to normalize the system fault probability, that is, output the system fault probability value. ;

[0101] Step 4.2: In order to improve the model’s ability to identify abnormal samples, reduce the performance degradation caused by class imbalance, and enhance the discrimination of boundary samples, a composite loss function that integrates multiple mechanisms is designed. , as shown below:

[0102]

[0103] Where, is the sample size, is the weight of the contrast loss, The weight of KL divergence, For the The uncertainty perception weight of the samples, For the Adaptive focal loss for samples, For the The contrast loss of samples, For the KL divergence of samples;

[0104] Among them, the uncertainty-aware weight of each sample is used to weight the entire loss, satisfying the following relationship:

[0105]

[0106] Where, is the output of the Sigmoid function; in the embodiment, hour Reaching the maximum value, that is, the most uncertain, When it approaches 0 or 1 Approaching 0;

[0107] No. The adaptive focal loss of samples satisfies the following relationship:

[0108]

[0109] Where, For the The adaptively adjusted class weights of samples, To control the weighted degree coefficient of difficult samples, For the The true label corresponding to each sample is 0 or 1;

[0110] No. The contrast loss of each sample satisfies the following relationship:

[0111]

[0112] Where, It is a contrast boundary used to encourage negative sample predictions to stay away from positive samples;

[0113] No. The KL divergence (KL Regularization) of samples satisfies the following relationship:

[0114]

[0115] Where, is the temperature parameter, is a small constant with stable numerical value, and the preferred value in the embodiment is 0.01.

[0116] An adaptive focal contrast loss function is introduced into the deep neural network model, comprehensively considering sample difficulty, class imbalance, adaptive focus, and modal distribution distance constraints. The focal loss suppresses the influence of easily classified samples and strengthens attention to boundary samples; the contrast loss effectively improves the distinction between normal and abnormal feature spatial distributions; and the KL divergence regularization and uncertainty weighting mechanism further enhance the model's stability and generalization capabilities under complex working conditions.

[0117] Step 4.3, using the composite loss function The minimum is the training target, and the Adam optimizer is combined with backpropagation optimization to train the fault detection model.

[0118] Step 4.4, according to the real-time data of the three-phase voltage and three-phase current of the power grid bus, obtain the characteristic vector according to steps 1 to 3;

[0119] Step 4.5: Input the feature vector into the trained fault detection model, and the output layer uses the Sigmoid function to output the system failure probability. , probability and set threshold If the failure probability Greater than the set threshold Then it is determined that there is a fault; if the fault probability Less than or equal to the set threshold It is determined that there is no fault. In this embodiment, the threshold is set to 0.618;

[0120] A deep feedforward neural network model with adaptive focal contrast loss is established to train the microgrid time series characteristics to achieve accurate identification of abnormal states of the power system, such as Figure 2 , which is the loss curve of 100 rounds of DNN model training after collecting training data according to steps 1 to 3 and normalizing it and inputting the improved loss function.

[0121] like Figure 3 As shown in the figure, based on the feature extraction and improved loss function training strategy proposed in this invention, the model's prediction results on the collected test data are pre-confusion matrix. The test set contains a total of 720 samples, including 491 normal samples and 229 abnormal samples. The model accurately identified 489 of the 491 normal samples and successfully detected 227 abnormal samples, with an extremely low false positive rate and zero missed detection rate, demonstrating extremely high detection accuracy and robustness.

[0122] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions, which should all be included in the scope of the claims of the present invention.

Claims

1. A fault detection method for a virtual power plant microgrid, characterized in that: include: Step 1: Obtain historical data of three-phase voltage and three-phase current of the power grid bus as input signals; Step 2: Decompose the input signal into multiple modes using the variational mode decomposition method; Step 3: Calculate the energy of each mode and the correlation coefficient between each mode and the input signal, establish a phase asymmetry index and a disturbance frequency index, and form a training feature vector with the weighted sum of the energy and correlation coefficient of each mode, the weighted sum of the phase asymmetry index and the correlation coefficient, and the weighted sum of the disturbance frequency index and the correlation coefficient; The absolute value of the deviation between the mean of the instantaneous phase of each mode in the positive region and the mean in the negative region is used as the phase asymmetry index, as follows: Where, For modal The phase asymmetry index, 、 Mode The time sample set in the positive and negative areas, For collection The number of elements contained in For collection The number of elements contained in For modal The instantaneous phase of The instantaneous phase reversal rate of each mode is used as the disturbance frequency index, as follows: Where, For modal The disturbance frequency index, is the time window length, the length of the sequence, is the modal signal The instantaneous frequency, It is a logical judgment function, which is 1 if the conditions in the brackets are met, otherwise it is 0; Step 4: train the established fault detection model using the training feature vector; obtain the feature vector according to steps 1 to 3 based on the real-time data of the three-phase voltage and three-phase current of the power grid bus; According to the feature vector, the fault detection model outputs the fault probability. If the fault probability is greater than the set threshold, a fault exists; otherwise, no fault exists.

2. The fault detection method for a virtual power plant microgrid according to claim 1, characterized in that: In step 1, historical data of the three-phase voltage and three-phase current of the power grid bus are obtained; historical instantaneous values ​​of the three-phase voltage and three-phase current in the sliding window are extracted to form a three-phase voltage sequence and a three-phase current series respectively; symmetrical component transformation is performed on the three-phase voltage sequence and the three-phase current series to obtain positive sequence components, negative sequence components and zero sequence components as multiple input signals.

3. The fault detection method for a virtual power plant microgrid according to claim 2, characterized in that: In step 2, the variational mode decomposition method is used to decompose each input signal in turn, and the augmented Lagrangian function is introduced and the alternating direction multiplier method is used to obtain the modal , the frequency bands of each mode do not overlap and the bandwidth is minimized.

4. The fault detection method for a virtual power plant microgrid according to claim 3, characterized in that: In step 3, the modal Perform Hilbert transform to get the mode The analytical signal form of as follows: Where, 、 Mode The instantaneous amplitude and instantaneous phase of is the imaginary number symbol, For modal The Hilbert transform of Extract the instantaneous phase of each mode from the analytical signal form.

5. The fault detection method for a virtual power plant microgrid according to claim 3, characterized in that: The correlation coefficient between each mode and each input signal , as shown below: Where, 、 Mode and input signal The variance of Indicates the first Positive voltage sequence component , the first Negative sequence voltage component , the first voltage zero-sequence component , the first positive sequence current component , the first Negative sequence current component and the first zero-sequence current component , To calculate the covariance.

6. The fault detection method for a virtual power plant microgrid according to claim 1, characterized in that: In step 4, a fault detection model is established based on a deep feedforward neural network and trained using the training feature vector. The training objective is to minimize the composite loss function and use the Adam optimizer for backpropagation optimization to train the fault detection model. According to the real-time data of the three-phase voltage and three-phase current of the power grid bus, the characteristic vector is obtained according to steps 1 to 3; According to the feature vector, the output layer of the trained fault detection model uses the Sigmoid function to output the system failure probability. If the failure probability is greater than the set threshold, it is determined that a fault exists; if the failure probability is less than or equal to the set threshold, it is determined that no fault exists.

7. The fault detection method for a virtual power plant microgrid according to claim 6, characterized in that: Composite loss function as follows: Where, is the sample size, is the weight of the contrast loss, The weight of KL divergence, For the The uncertainty perception weight of the samples, For the Adaptive focal loss for samples, For the The contrast loss of samples, For the KL divergence of samples; Where, The output of the Sigmoid function used in the output layer; Where, For the The adaptively adjusted class weights of samples, To control the weighted degree coefficient of difficult samples, For the The true label corresponding to each sample is 0 or 1; Where, For contrasting boundaries; Where, is the temperature parameter, is a small constant that is numerically stable.

8. A fault detection system for a virtual power plant microgrid, used to implement the fault detection method for a virtual power plant microgrid according to any one of claims 1 to 7, characterized in that: include: An acquisition module is used to obtain historical data of the three-phase voltage and three-phase current of the power grid bus as input signals; A modal decomposition module is used to decompose the input signal into multiple modes using a variational modal decomposition method; The eigenvector calculation module is used to calculate the energy of each mode and the correlation coefficient between each mode and the input signal, establish the phase asymmetry index and the disturbance frequency index, and form the training eigenvector with the weighted sum of the energy and correlation coefficient of each mode, the weighted sum of the phase asymmetry index and the correlation coefficient, and the weighted sum of the disturbance frequency index and the correlation coefficient; The absolute value of the deviation between the mean of the instantaneous phase of each mode in the positive region and the mean in the negative region is used as the phase asymmetry index, as follows: Where, For modal The phase asymmetry index, 、 Mode The time sample set in the positive and negative areas, For collection The number of elements contained in For collection The number of elements contained in For modal The instantaneous phase of The instantaneous phase reversal rate of each mode is used as the disturbance frequency index, as follows: Where, For modal The disturbance frequency index, is the time window length, the length of the sequence, is the modal signal The instantaneous frequency, It is a logical judgment function, which is 1 if the conditions in the brackets are met, otherwise it is 0; A model building module, used for training the established fault detection model using the training feature vector; The fault detection module is used to call the eigenvector calculation module to obtain the eigenvector based on the real-time data of the three-phase voltage and three-phase current of the power grid bus. Based on the eigenvector, the fault detection model outputs the fault probability. If the fault probability is greater than the set threshold, a fault exists; otherwise, no fault exists.

Images