A data-driven-based energy scheduling fault diagnosis method, device and medium

Abstract

This application discloses a data-driven energy dispatch fault diagnosis method, device, and medium, mainly relating to the field of fault diagnosis technology. It addresses the problems of existing solutions neglecting the multimodal distribution characteristics of energy data, failing to effectively capture the complex coupling relationships between the source-grid-load-storage links, lacking awareness of dispatch strategies, and being sensitive to noise interference. The method includes: obtaining a graph convolutional hidden representation matrix and then calculating a sparse dictionary matrix; optimizing the graph convolutional sparse coding objective function, solving for the sparse coding coefficients, and outputting a feature mining data matrix; calculating a dispatch strategy matching degree matrix based on the feature mining data matrix, using introduced dispatch plan data and real-time operational constraints; fusing the operational status classification vector and the dispatch strategy matching degree matrix to obtain predicted fault categories; and iteratively training a diagnostic model based on the predicted fault categories and labeled fault categories until a well-trained diagnostic model is obtained.

Description

Technical Field

[0001] This application relates to the field of fault diagnosis technology, and in particular to a data-driven method, device and medium for energy dispatch fault diagnosis. Background Technology

[0002] The dispatching, operation, and stable operation of energy systems face unprecedented challenges. On the one hand, the large-scale grid connection of intermittent and fluctuating power sources such as wind and solar power has resulted in highly random and multimodal characteristics in system operation. Traditional fault diagnosis methods based on fixed thresholds and physical models are insufficient to accurately capture their complex dynamic behavior and non-stationary features. On the other hand, to achieve full utilization of new energy sources and economical system operation, advanced dispatching strategies such as automatic generation control and demand-side response are widely used. The system operating point is frequently and proactively adjusted due to dispatching instructions, making it extremely difficult to distinguish between "planned normal adjustments" and "unplanned real faults."

[0003] Existing solutions mainly suffer from the following problems: 1. Conventional normalization methods use global or segmented processing, ignoring the multimodal distribution characteristics of energy data, leading to feature confusion between different operating modes and poor adaptability under non-stationary operating conditions; 2. Conventional feature extraction methods ignore the physical topology of the system, failing to effectively capture the complex coupling relationships between the source-grid-load-storage links, resulting in the loss of interactive information; 3. Existing classifiers lack the ability to perceive scheduling strategies, making it difficult to distinguish between operating condition adjustments caused by normal scheduling instructions and real fault events, resulting in a high false alarm rate; 4. Mainstream diagnostic models are sensitive to noise interference, and their performance degrades significantly when data quality declines, with insufficient robustness affecting the reliability of actual deployment. Summary of the Invention

[0004] This application provides a data-driven method, device, and medium for diagnosing energy dispatching faults to address the problems of existing solutions.

[0005] Firstly, this application provides a data-driven method for diagnosing energy dispatching faults, the method comprising:

[0006] Based on the multi-source synchronous operation data corresponding to the source-grid-load-storage system, a synchronous energy array is constructed; the labeled fault categories corresponding to the synchronous energy array are obtained, and then training samples are constructed.

[0007] The training samples are input into the diagnostic model, and the multi-source synchronous operation data in the synchronous energy array of the training set are normalized; a connection relationship matrix between the multi-source synchronous operation data is constructed; and a normalized adjacency matrix is ​​constructed based on the normalized synchronous energy array and the connection relationship matrix.

[0008] The normalized data matrix is ​​aggregated and transformed using a multi-head graph attention mechanism. Features from a preset historical time window are then concatenated to obtain the graph convolutional hidden representation matrix, which is then used to calculate the sparse dictionary matrix. Based on the sparse dictionary matrix and the sparse coding coefficients, the graph convolutional sparse coding objective function is optimized, the sparse coding coefficients are solved, and the feature mining data matrix is ​​output.

[0009] Based on the feature mining data matrix, the scheduling strategy matching degree matrix is ​​calculated using the introduced scheduling plan data and real-time operation constraints. The running status classification vector and the scheduling strategy matching degree matrix are fused to obtain the predicted fault category. The diagnostic model is iteratively trained according to the predicted fault category and the labeled fault category until a well-trained diagnostic model is obtained.

[0010] In one implementation of this application, the synchronization energy array in the training set is normalized, specifically including:

[0011] The synchronous energy array is mapped to the latent space by a variational autoencoder to obtain the latent variable representation;

[0012] Clustering is performed on the latent variable representations to obtain the clustering results and the number of modes M;

[0013] Based on the clustering results, calculate the center vector and standard deviation vector for each mode;

[0014] The encoder of the variational autoencoder maps the synchronization energy array to a latent variable representation while outputting the mean and variance of a Gaussian distribution. After connecting to the soft classification layer, it outputs an M-dimensional vector, with the components representing the mixing weights of each mode.

[0015] The nonlinear correction term is calculated based on the pre-defined low-dimensional representation of the synchronous energy array in the latent space.

[0016] Calculate the normalized data matrix based on the nonlinear correction term, the synchronous energy array, the standard deviation vector, and the center vector.

[0017] In one implementation of this application, a connection matrix between multi-source synchronously operating data is constructed; based on the normalized synchronous energy array and the connection matrix, a normalized adjacency matrix is ​​constructed, specifically including:

[0018] Based on the physical topology of the source-grid-load-storage system, a connection matrix is ​​constructed among the synchronously operating data from multiple sources.

[0019] Through the formula: Construct a normalized adjacency matrix ;

[0020] in, This represents a normalized adjacency matrix, where the element values ​​are in... Within the range; This represents the self-loop enhanced connectivity matrix; express The degree matrix.

[0021] In one implementation of this application, a multi-head graph attention mechanism is used to aggregate and transform information from a normalized data matrix, and features from a preset historical time window are concatenated to obtain a graph convolutional hidden representation matrix. Then, a sparse dictionary matrix is ​​calculated. Specifically, this includes:

[0022] Through the formula:

[0023] Compute the graph convolution hidden representation matrix;

[0024] in, Representation of the graph convolution hidden representation matrix; This represents the global weight parameter matrix of the graph convolutional layer; Indicates the attention head index; This represents the total number of attention heads in a multi-head attention mechanism; Presentation layer normalization operation; Indicates the first The normalized adjacency matrix is ​​obtained by calculating the attention head; This represents the global bias vector of the graph convolutional layer; The concatenation operation is indicated; This represents the max pooling operation; This represents the feature matrix within a preset historical time window; This indicates a modified linear unit activation function;

[0025] The basic dictionary representation is obtained by linearly transforming the graph convolutional hidden representation matrix with the dictionary learning weight matrix and bias vector and then processing it with the Softmax function. At the same time, the residual terms and gating signals are calculated based on the graph convolutional hidden representation. Then, the basic dictionary representation is combined with the residual terms adjusted by the gating mechanism to obtain the sparse dictionary matrix.

[0026] In one implementation of this application, based on a sparse dictionary matrix and desparse coding coefficients, the graph convolution sparse coding objective function is optimized, the sparse coding coefficients are solved, and a feature mining data matrix is ​​output, specifically including:

[0027] By minimizing the reconstruction error between the product of the normalized data matrix, the sparse dictionary matrix, and the sparse coding coefficient matrix, and simultaneously incorporating sparsity constraints on the sparse coding coefficient matrix and graph smoothing regularization constraints on the sparse dictionary matrix, the optimized graph convolutional sparse coding objective function is obtained through joint learning.

[0028] The sparse coding coefficient matrix is ​​updated in each iteration using an iterative soft thresholding algorithm based on the sparse dictionary matrix, the normalized data matrix, and the graph Laplacian matrix, thus obtaining the sparse coding coefficient matrix.

[0029] The feature mining data matrix is ​​obtained by multiplying the sparse dictionary matrix and the sparse coding coefficient matrix, combining the linear transformation of the graph convolution hidden representation matrix, and then processing it with the modified linear unit activation function.

[0030] In one implementation of this application, based on the feature mining data matrix, a scheduling strategy matching degree matrix is ​​calculated using the introduced scheduling plan data and real-time running constraints, specifically including:

[0031] Based on the feature mining data matrix, a comprehensive index of operating status and its real-time change gradient are introduced, and a classification vector of operating status is constructed by combining multimodal hybrid weights.

[0032] By calculating the bidirectional attention matching degree between the feature mining data matrix and the scheduling plan data and real-time running constraints, a scheduling strategy matching degree matrix is ​​constructed.

[0033] In one implementation of this application, the predicted fault category is obtained by fusing the running state classification vector and the scheduling strategy matching degree matrix; based on the predicted fault category and the labeled fault category, the diagnostic model is iteratively trained until a trained diagnostic model is obtained, specifically including:

[0034] The running status classification vector and the flattened vector of the scheduling policy matching degree matrix are fused together, and the classification is performed by a deep network to obtain the predicted fault category.

[0035] The cross-entropy loss function is used as the total loss during model training.

[0036] The diagnostic model is trained using an iterative method based on gradient descent. When a preset stopping condition is met, the trained diagnostic model is obtained.

[0037] In one implementation of this application, after fusing the running state classification vector and the scheduling strategy matching degree matrix to obtain the predicted fault category; and iteratively training the diagnostic model based on the predicted fault category and the labeled fault category until a trained diagnostic model is obtained, the method further includes:

[0038] Collect multi-source synchronous operation data and corresponding scheduling plans and operation constraints of the source-grid-load-storage system during the period to be diagnosed, input them into the trained diagnostic model, and output fault category labels after multimodal dynamic normalization, feature interaction mining and energy scheduling fault classification processing.

[0039] Secondly, this application provides a data-driven energy dispatch fault diagnosis device, the device comprising:

[0040] processor;

[0041] And a memory storing executable code, which, when executed, causes the processor to perform a data-driven energy dispatch fault diagnosis method as described in any of the preceding claims.

[0042] Thirdly, this application provides a non-volatile computer storage medium storing computer instructions thereon, which, when executed, implement a data-driven energy dispatch fault diagnosis method as described in any of the preceding claims.

[0043] As can be seen from the above technical solutions, this application has the following advantages:

[0044] 1. Multimodal data normalization enhances adaptability:

[0045] By constructing a synchronized energy array and performing normalization, this method effectively addresses the problem of conventional normalization methods neglecting the multimodal distribution characteristics of energy data. Normalization can distinguish features under different operating modes, avoiding feature confusion and thus maintaining the adaptability of the diagnostic model under non-stationary conditions. This approach directly improves the model's responsiveness to complex operating environments and reduces diagnostic errors caused by changes in data distribution.

[0046] 2. Feature extraction for physical topology perception:

[0047] This method explicitly incorporates the physical topology of the source-network-load-storage system by constructing a connectivity matrix and a normalized adjacency matrix. This design enables the feature extraction process to capture the complex coupling relationships between various components, effectively solving the problem of lost interactive information. Through information aggregation and transformation via graph attention mechanisms, the model can gain a more comprehensive understanding of system dynamics and improve the accuracy of fault identification.

[0048] 3. Design of a scheduling strategy-aware classifier:

[0049] By incorporating scheduling plan data and real-time operational constraints, and calculating the scheduling policy matching degree matrix, this method enables the classifier to perceive scheduling policies. By fusing the operational status classification vector with the scheduling policy matching degree matrix, the model can distinguish between operational condition adjustments caused by normal scheduling instructions and real fault events, significantly reducing the false alarm rate. This design directly improves the reliability of diagnostic results and reduces unnecessary operational interventions.

[0050] 4. Sparse coding optimization to improve noise robustness:

[0051] By constructing a graph convolutional sparse coding objective function and solving for the sparse coding coefficients, this method enhances the model's robustness to noise interference. The optimization process of the sparse dictionary matrix and the desparse coding coefficients ensures that the model maintains stable performance even when data quality degrades. This design directly addresses the noise sensitivity of existing models, improves the reliability of practical deployments, and ensures long-term effective operation in complex environments. Attached Figure Description

[0052] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 This is a flowchart of a data-driven energy dispatch fault diagnosis method provided in an embodiment of this application.

[0054] Figure 2 This is a robustness analysis diagram of different methods provided in the embodiments of this application under noisy environments.

[0055] Figure 3 This is a performance comparison chart provided in the embodiments of this application when the signal-to-noise ratio is 5dB.

[0056] Figure 4 This is a comparison chart of the fault diagnosis accuracy of different methods provided in the embodiments of this application.

[0057] Figure 5 This is a schematic diagram of the internal structure of a data-driven energy dispatch fault diagnosis device provided in an embodiment of this application. Detailed Implementation

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] Those skilled in the art should understand that the embodiments described below are merely preferred embodiments of this disclosure and do not imply that this disclosure can only be implemented through these preferred embodiments. These preferred embodiments are merely used to explain the technical principles of this disclosure and are not intended to limit the scope of protection of this disclosure. Based on the preferred embodiments provided by this disclosure, all other embodiments obtained by those skilled in the art without creative effort should still fall within the scope of protection of this disclosure.

[0060] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0061] The technical solutions proposed in the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0062] The embodiment provides a data-driven energy dispatch fault diagnosis method, such as... Figure 1 As shown in the embodiments of this application, the method mainly includes the following steps:

[0063] Step 110: Based on the multi-source synchronous operation data corresponding to the source-grid-load-storage system, construct a synchronous energy array; obtain the labeled fault categories corresponding to the synchronous energy array, and then construct training samples.

[0064] In some embodiments, sensors, smart meters, monitoring and data acquisition systems, and energy management systems deployed on the power supply side (such as wind farms, photovoltaic power plants, and thermal power plants), grid side (such as substations and transmission lines), load side (such as industrial, commercial, and residential loads), and energy storage side (such as battery energy storage systems) of the energy internet system are used for synchronous data acquisition. The acquired data attributes include wind power output power, photovoltaic power output power, thermal power output power, grid frequency, bus voltage, line current, load power, energy storage charging power, energy storage discharging power, energy storage status, transformer temperature, circuit breaker status, protection device action signal, power factor, total harmonic distortion rate, voltage deviation, and frequency deviation. These data are synchronously recorded at a high frequency (such as one sample point per second or per minute) and aligned by timestamps to form an original time series data matrix (synchronous energy array) with a dimension of the total number of samples multiplied by the number of attributes.

[0065] Then, data annotation is performed. The annotation work is based on the rules formulated by domain experts, the action records of protection devices, the system event logs, and manual review. For each data sample at each time point, a fault category label is assigned to it according to the actual operating status of the system and the event records at that time. The annotation categories mainly include: normal, overvoltage, undervoltage, frequency deviation, and line overload.

[0066] Finally, the labeled time series data is divided into training set, validation set and test set according to a certain ratio to complete the construction of training samples.

[0067] Step 120: Input the training samples into the diagnostic model and normalize the multi-source synchronous operation data in the synchronous energy array of the training set; construct the connection relationship matrix between the multi-source synchronous operation data; construct the normalized adjacency matrix based on the normalized synchronous energy array and the connection relationship matrix.

[0068] Data from the synchronous operation of the power source-grid-load-storage system is characterized by multiple sources, non-stationarity, and physical constraint coupling. Different attributes such as power, voltage, and state variables not only have large differences in numerical range, but also exhibit spatiotemporal correlation and dynamic mode switching. In order to effectively cope with the impact of local nonlinearity and multimodal distribution between feature dimensions, data normalization is required. Conventional techniques use global normalization or piecewise normalization, but ignore the multimodal characteristics of the data, resulting in feature confusion between modes and loss of important information.

[0069] This application maps the original data to a latent space using a variational autoencoder and performs clustering to identify data modes. Based on the clustering results, it calculates the center vector and standard deviation vector corresponding to each mode. Simultaneously, it obtains the mixed weights of data belonging to each mode through the encoder network and a soft assignment process. Furthermore, it calculates nonlinear correction terms based on the low-dimensional representation of the data in the latent space. Then, it applies a normalization formula to the original data according to these mode-specific parameters to obtain a normalized data matrix. This process eliminates the differences in numerical distribution among different modes while preserving their important nonlinear structures, enhancing adaptability to non-stationary data. This is expressed as:

[0070] ,

[0071] In the formula, Represents a normalized data matrix with dimension 1. It represents the standardized data after multimodal adaptive processing, eliminates confusion caused by differences in numerical distribution between different modes, and preserves important nonlinear structures within the modes.

[0072] This represents the original data matrix with dimension 1. The data originates from synchronous measurements of the source-grid-load-storage system. Each row represents a multi-attribute observation sample at a given time point, and each column represents a time series of an attribute.

[0073] The total number of samples;

[0074] The number of attributes;

[0075] The modality number represents the total number of modalities dynamically identified through latent space clustering using a variational autoencoder.

[0076] This represents the modal index, with a value range of 100. arrive ;

[0077] Indicates the first The center vectors of each modality are generated by K-means clustering in the latent space of a variational autoencoder, reflecting the typical characteristics of the data within different modalities.

[0078] Indicates the first The standard deviation vector of a modality is used to measure the dispersion of each data attribute within that modality.

[0079] The calculation method is expressed as follows ;

[0080] Indicates the first Number of samples in each modality;

[0081] Indicates the sample index;

[0082] Indicates the first Each original data vector is a raw data matrix. The i-th row vector;

[0083] For small constants, it is preferable to set them. ;

[0084] Represents the original data matrix Belongs to the The mixed weights of each modality satisfy... It is calculated jointly by the encoder network of the variational autoencoder and the soft allocation process;

[0085] In practical implementation,

[0086] Indicates that for the first The nonlinear correction term for each mode is calculated as follows: This is used to enhance the capture of intramodal nonlinearities;

[0087] Represents the original data matrix Low-dimensional representation in the latent space of a variational autoencoder;

[0088] The weight matrix is ​​a trainable parameter used to linearly map latent variables to a space with the same dimensions as the original data.

[0089] The bias vector is a trainable parameter, which is a learnable offset added after the linear transformation.

[0090] In one embodiment, the attributes of the raw data include: wind power output. Photovoltaic output power Thermal power output power Grid frequency Bus voltage Line current Load power Energy storage charging power Energy storage and discharge power Energy storage status Transformer temperature Circuit breaker status , protection device action signal Power factor Total Harmonic Distortion Voltage deviation Frequency deviation These attributes directly or indirectly reflect the multimodal operating status of the source-grid-load-storage system.

[0091] In practical implementation, the number of modes Dynamic identification is achieved through latent space clustering using a variational autoencoder. Specifically, a variational autoencoder is first used to map the data to a low-dimensional latent space. Then, a clustering effectiveness index is applied to the latent space features to help determine the optimal number of clusters. The effectiveness index can be the silhouette coefficient or the Calinski-Harabasz exponent. Finally, the K-means++ algorithm is used for clustering, and the number of cluster centers obtained is the number of modalities dynamically identified. .

[0092] In specific implementation, the first The center vector of each mode It is generated through K-means clustering in the latent space of a variational autoencoder. Specifically, the original data matrix is ​​generated using a trained variational autoencoder. The latent variables of all samples are encoded into a low-dimensional latent space. Then, the K-means++ clustering algorithm is applied to this latent variable space. Each latent variable sample is assigned to the nearest cluster center through iteration, and the centroid of each cluster is recalculated. After clustering, the coordinates of each cluster centroid in the latent space are the center of the mode. Finally, the cluster centers in the latent space are mapped back to the original data space through the decoder of the variational autoencoder to obtain the center vector of each mode.

[0093] In practical implementation, the original data matrix Belongs to the Mixed weights of each modality The data is calculated jointly by the encoder network of the variational autoencoder and the soft allocation process. Specifically, the encoder of the variational autoencoder calculates the original data matrix. While mapping to latent variables, the mean and variance of a Gaussian distribution are output, and then connected to a soft classification layer to output a... A dimensional vector, whose dimensional vector is the first dimensional vector. Each component represents the original data matrix. Belongs to the Mixed weights of each modality .

[0094] In practical implementation, the original data matrix Low-dimensional representation in the latent space of a variational autoencoder The original data matrix is ​​obtained through an encoder network of a variational autoencoder. The input is fed into an encoder network of a variational autoencoder, which typically consists of multiple fully connected or convolutional layers. Through forward propagation, the high-dimensional input data is mapped to a low-dimensional latent space, outputting the mean and variance of the latent variable distribution. Then, a deterministic low-dimensional representation is sampled from this distribution using a reparameterization method, serving as the original data matrix. Low-dimensional representation in the latent space of a variational autoencoder .

[0095] Furthermore, the source-network-load-storage data contains complex physical topological relationships. Conventional feature mining methods, such as principal component analysis or autoencoders, ignore the graph structure and cannot effectively capture the dependencies between nodes, resulting in the loss of interaction information between features and the difficulty in eliminating redundant features.

[0096] This application constructs a physical topology graph of a source-network-load-storage system and applies graph convolutional networks and sparse coding for joint learning to dynamically extract interaction information between features, thereby enhancing feature representation and reducing redundancy. The specific steps are as follows:

[0097] 1) Construct the adjacency matrix of the source-grid-load-storage system:

[0098] Based on the physical topology of the source-grid-load-storage system, a connection matrix between attributes is constructed, specifically defining... This represents the adjacency matrix of the source-grid-load-storage system, with dimension 1. Based on physical topology construction, it is used to characterize the connection relationships between different attributes, where each element indicates whether there is a direct physical connection between the corresponding attributes. That is, the adjacency matrix of the source-grid-load-storage system is represented by the element of the adjacency matrix. Line number The elements of the column are ,when Time indicates the first The first attribute and the first There is a direct physical connection between the attributes, when Time indicates the first The first attribute and the first There is no direct physical connection between the attributes. j is the attribute index, with a value ranging from 1 to... .

[0099] 2) Calculate the normalized adjacency matrix:

[0100] By combining the adjacency matrix representing physical topological connectivity with the identity matrix, a self-loop-enhanced adjacency matrix is ​​obtained. Based on this, the corresponding degree matrix is ​​calculated, and then a symmetric normalization operation is applied to obtain a numerically stable normalized adjacency matrix, expressed as:

[0101] ,

[0102] In the formula, Represents the normalized adjacency matrix with dimension . is a symmetric, positive semi-definite matrix whose element values ​​are in Within the range, it represents the normalized connection strength between attributes after considering the balanced influence of the importance of the node itself and the degree of the neighboring nodes.

[0103] The adjacency matrix of self-loop enhancement is represented as: , dimension Each node is connected to itself, and the diagonal elements are 1, representing an enhanced physical topology relationship that includes self-connections;

[0104] It is an identity matrix with dimension 1. ;

[0105] express The degree matrix is ​​a diagonal matrix, and the diagonal elements in the i-th row and i-th column of the degree matrix are... The calculation method is expressed as ;

[0106] Degree matrix The negative half power is obtained by taking the reciprocal of the square root of the diagonal elements.

[0107] 3) Calculate the graph convolution hidden representation matrix:

[0108] Information aggregation and transformation of the normalized data matrix are performed using a multi-head graph attention mechanism based on the normalized adjacency matrix. This is combined with layer normalization and activation function processing. Simultaneously, max pooling is used to extract summary information from the feature matrix within the historical time window. Finally, the graph convolution output is concatenated with the historical features to obtain a graph convolutional hidden representation matrix that integrates spatiotemporal information, represented as:

[0109] ,

[0110] In the formula, The hidden representation matrix of the graph convolution has dimensions of . The deep feature representation, after being guided by graph structure and enhanced by spatiotemporal information, can more effectively reflect the interaction state and evolution pattern of each attribute in the source-grid-load-storage system.

[0111] The global weight parameter matrix of the graph convolutional layer is a trainable parameter that performs a linear transformation on the node features after aggregation by the adjacency matrix.

[0112] Indicates the attention head index;

[0113] This represents the total number of attention heads in a multi-head attention mechanism;

[0114] Presentation layer normalization operation;

[0115] Indicates the first The normalized adjacency matrix is ​​obtained by calculating the attention head;

[0116] The global bias vector of the graph convolutional layer is a trainable parameter that adds a learnable offset to each hidden feature dimension.

[0117] The concatenation operation is indicated;

[0118] This represents the max pooling operation;

[0119] The feature matrix representing the historical time window is obtained by normalizing the data matrix at the current time point. Take before Composed of time points Matrix;

[0120] The historical time window size is preferably set to 128;

[0121] This represents the dimension of the hidden layer in the graph, with an empirical value of 64.

[0122] This represents the modified linear unit activation function.

[0123] Step 130: Aggregate and transform information in the normalized data matrix through a multi-head graph attention mechanism, splice in the preset historical time window features to obtain the graph convolution hidden representation matrix, and then calculate the sparse dictionary matrix; based on the sparse dictionary matrix and the sparse coding coefficients, optimize the graph convolution sparse coding objective function, solve the sparse coding coefficients, and output the feature mining data matrix.

[0124] The basic dictionary representation is obtained by linearly transforming the graph convolutional hidden representation matrix with the dictionary learning weight matrix and bias vector, and then processing it with the Softmax function. Simultaneously, residual terms and gating signals are calculated based on the graph convolutional hidden representation. Finally, the basic dictionary representation is combined with the gated residual terms to obtain a sparse dictionary matrix with representational capabilities, expressed as:

[0125] ,

[0126] In the formula, Represents a sparse dictionary matrix with dimension . , representing a set of overcomplete bases learned from graph convolutional features that can be sparsely linearly combined to reconstruct the original data;

[0127] This represents the dictionary learning weight matrix, with dimension 1. , are trainable parameters used to linearly map the graph convolutional hidden representation to the dictionary atomic space;

[0128] This represents the dictionary learning bias vector, with dimension . , are trainable parameters used to add learnable offsets;

[0129] This represents the weighting coefficient of the residual connection, which controls the contribution of the residual terms to the dictionary matrix. It is preferably set to 0.5.

[0130] Represents the hyperbolic tangent activation function;

[0131] The weight matrix representing the residual connection mapping is a trainable parameter used to linearly map the graph convolutional hidden representation to an intermediate space that matches the dimension of the target dictionary's atomic space.

[0132] This represents element-wise multiplication;

[0133] This represents the Sigmoid activation function;

[0134] The weight matrix representing the gating mechanism consists of trainable parameters, and the gating signal is calculated based on graph structure information and current features.

[0135] This represents the number of atoms in the dictionary, with an empirical value of 2D.

[0136] This represents the Softmax function.

[0137] The objective function for optimizing graph convolutional sparse coding can be:

[0138] By minimizing the reconstruction error between the product of the normalized data matrix, the sparse dictionary matrix, and the sparse coding coefficient matrix, and simultaneously incorporating sparsity constraints on the sparse coding coefficient matrix and graph smoothing regularization constraints on the sparse dictionary matrix, a sparse representation that can accurately reconstruct the data and conforms to graph structure characteristics is obtained through joint learning. The optimization problem is expressed as:

[0139] ,

[0140] In the formula, This represents the sparse coding coefficient matrix with dimension . By solving this optimization problem, the normalized data matrix is ​​obtained. In sparse dictionary matrix The linear representation coefficients under the given conditions, whose sparsity means that each sample is characterized by only a combination of a few dictionary atoms;

[0141] Denotes the Frobenius norm;

[0142] Represents the L1 norm;

[0143] for transpose;

[0144] The trace operation is used to calculate the sum of the diagonal elements of a matrix.

[0145] The graph Laplacian matrix is ​​a positive semi-definite symmetric matrix used to measure the smoothness of the graph signal's variation on the graph. It is calculated as follows: ;

[0146] This represents the sparsity regularization coefficient, which controls the strength of the sparsity constraint. It is preferably set to 0.01.

[0147] This represents the graph smoothing regularization coefficient, which controls the strength of graph structure preservation; it is preferably set to 0.03.

[0148] Represents the variable and Perform joint minimization.

[0149] Solving for sparse coding coefficients can be done in the following ways:

[0150] By employing an iterative soft thresholding algorithm, the sparse coding coefficient matrix is ​​updated in each iteration based on the sparse dictionary matrix, the normalized data matrix, and the graph Laplacian matrix. This yields a new sparse coding coefficient matrix that simultaneously satisfies sparsity and graph structure preservation, thereby gradually approximating the sparse solution, as shown below:

[0151] ,

[0152] In the formula, Indicates the first The encoding coefficient matrix of the next iteration has dimensions of The sparse representation estimate of the data under the dictionary at the current iteration step is represented by the coding coefficient matrix of the initial iteration. Set as a zero matrix;

[0153] Indicates the first The coding coefficient matrix of the next iteration;

[0154] This represents the soft threshold function, with parameters as follows: The calculation method is expressed as This is used to select input vectors whose absolute values ​​are less than a threshold. The elements are set to zero, and the other elements are shrunk to produce a sparse solution;

[0155] It is the independent variable of the function;

[0156] Represents a symbolic function;

[0157] This represents the function that takes the maximum value.

[0158] The threshold value used for soft thresholding operations;

[0159] This represents the iteration step size parameter, which controls the step size of gradient descent updates. It is preferred to set it to 1.

[0160] This represents the graph guidance coefficient, used to control the influence of the graph Laplacian regularization term in iterative updates, and is preferably set to 0.01.

[0161] The output feature mining data matrix can be specifically:

[0162] The feature mining data matrix is ​​obtained by multiplying the sparse dictionary matrix and the sparse coding coefficient matrix, combined with a linear transformation of the graph convolutional hidden representation matrix, and then processing it with a modified linear unit activation function with leakage, as follows:

[0163] ,

[0164] In the formula, The feature mining data matrix is ​​represented by fusing discriminative features from sparse coding and topological correlation features from graph convolution, and its expressive power is enhanced through nonlinear activation.

[0165] The weight matrix of the feature fusion layer is a trainable parameter used to perform a linear transformation on the graph convolutional hidden representation.

[0166] This represents the modified linear unit activation function with leakage.

[0167] Step 140: Based on the feature mining data matrix, calculate the scheduling strategy matching degree matrix using the introduced scheduling plan data and real-time operation constraints; fuse the operation status classification vector and the scheduling strategy matching degree matrix to obtain the predicted fault category; iteratively train the diagnostic model according to the predicted fault category and labeled fault category until a well-trained diagnostic model is obtained.

[0168] It should be noted that using feature mining data matrices directly for fault classification still faces challenges such as class imbalance, time-varying fault modes, and mismatch with scheduling strategies. Conventional classifiers ignore the coupling between energy scheduling objectives and real-time operating status, making it difficult to distinguish between normal operating condition adjustments caused by scheduling instructions and real faults.

[0169] This application constructs a matching degree matrix between the running state classification vector and the scheduling strategy, and designs a deep classification network with dynamic weight allocation and scheduling awareness to achieve accurate and robust classification of faults. The specific steps are as follows:

[0170] 1) Dynamic construction of runtime state classification vectors:

[0171] Based on the feature mining data matrix, a comprehensive operational status index and its real-time gradient are introduced. Combined with multimodal hybrid weights, an operational status classification vector is constructed that dynamically reflects the system's health and stability, providing richer discrimination criteria for classification. This is represented as:

[0172] ,

[0173] In the formula, This represents the operational status classification vector constructed at time point t, which characterizes the comprehensive operational status by integrating instantaneous features, state change trends, and multimodal information.

[0174] This represents a point-in-time index, with a value range of [value range missing]. arrive ;

[0175] Representing the feature mining data matrix In the The row vector at time point, i.e., the first... Each feature is mined into a data vector;

[0176] This represents a vector concatenation operation;

[0177] This represents the comprehensive index of the operating status at time point t. The discrete gradient at time point t is used to measure the drasticness of instantaneous changes in the operating state, and is calculated as follows: , ,for When, define ;

[0178] This represents the comprehensive operational status index at time t. It is a scalar used to quantify the overall operational health of the system, and its calculation method is expressed as follows: ;

[0179] The comprehensive index weight vector is a trainable parameter used to linearly combine feature mining data vectors into scalars to calculate the comprehensive index of operating status and thus quantify the overall health of the system.

[0180] For Transpose;

[0181] The fluctuation penalty coefficient is preferably set to 0.05. When calculating the comprehensive index of operating status, the characteristic fluctuation is penalized to reflect the system stability.

[0182] The feature mining data matrix represents the data from the first... Time point to the The standard deviation of data within a time window is used to measure the volatility of recent operating characteristics;

[0183] The time window size is preferably set to [value]. It is used to calculate short-term fluctuations in the operating status;

[0184] This indicates that the sample at time t belongs to the th time. Mixed weights for each modality;

[0185] Indicates the relationship with the first The state mapping vector associated with each modality is a trainable parameter used to map modality attribution information to the state space.

[0186] 2) Calculation of the scheduling strategy matching degree matrix:

[0187] To distinguish between scheduling instructions and actual faults, external scheduling plan data and real-time operational constraints are introduced. By calculating the bidirectional attention matching degree between feature mining results and scheduling plans and operational constraints, a scheduling strategy matching degree matrix reflecting the degree of consistency between "actual operation and scheduling intent" is constructed. This matrix serves as the key context for classification decisions and is represented as follows:

[0188] ,

[0189] In the formula, The scheduling strategy matching degree matrix is ​​represented at time t. The value of each element represents the degree of matching between the actual operating characteristics of different devices or bus nodes and their scheduling plan and operating constraints. A low matching degree may indicate a fault.

[0190] This represents the query matrix, which acts as the query in the attention mechanism to match the scheduling plan and runtime constraints, calculating the matching degree. The calculation method is represented as follows: ;

[0191] For splicing operations;

[0192] This represents the runtime state classification vector constructed at time point t. The vector projected through the linear layer projects the running state classification vector to the same dimension as the feature vector;

[0193] The key matrix, used as keys in the attention mechanism, is used to calculate similarity with the query matrix. The calculation method is represented as follows: ;

[0194] The value matrix, used as values ​​in the attention mechanism, is used to aggregate information based on attention weights. Its calculation method is represented as follows: ;

[0195] This represents the scheduling plan vector at time point t, used to specify the scheduled scheduling instructions at the current time point. It includes the planned generation / load value and reserve capacity information, and is a preset external input vector.

[0196] This represents the real-time operating constraint vector at time t, used to specify the real-time operating limits at the current time point, including line capacity limits, voltage safety boundaries, and minimum ramp rate. It is a preset external input vector.

[0197] The query weight matrix is ​​a trainable parameter used to linearly transform the query matrix into the attention space.

[0198] Let be the key weight matrix, which are trainable parameters used to linearly transform the key matrix into the attention space;

[0199] The value weight matrix is ​​a trainable parameter used to linearly transform the value matrix into the attention space.

[0200] The scheduling plan projection weight matrix is ​​a trainable parameter used to project the scheduling plan vector onto a dimension that matches the features.

[0201] The running constraint projection weight matrix is ​​a trainable parameter used to project the running constraint vector onto a dimension that matches the features;

[0202] This represents the dimension of the key vector, used to scale the dot product attention; it is preferably set to [value]. This is to prevent the gradient from vanishing due to excessively large dot product values.

[0203] This represents the exponential linear unit activation function.

[0204] 3) Scheduling-aware deep classification networks:

[0205] The operational status classification vector is fused with the flattened vector of the scheduling policy matching degree matrix, and then classified using a deep network with dynamic feature weight adjustment and fault prototype contrast learning. This network not only learns the classification boundary but also learns the prototype representation of each fault category, resulting in the predicted fault category at each time point, represented as:

[0206] ,

[0207] In the formula, This represents the predicted fault category at time t, characterizing the model's judgment on the system fault type at the current time point, and the output is the fault category label;

[0208] This represents the set of all possible fault categories, including normal, overvoltage, undervoltage, frequency deviation, and line overload.

[0209] This represents the fault category index.

[0210] This indicates that the fault category that maximizes the function value is selected, i.e., the category with the highest probability of being predicted.

[0211] This represents the weight matrix of the first fully connected layer, which are trainable parameters;

[0212] This represents the bias vector of the first fully connected layer, which is a trainable parameter;

[0213] This represents the weight matrix of the second fully connected layer, which are trainable parameters;

[0214] This represents the bias vector of the second fully connected layer, which is a trainable parameter.

[0215] Let represent the fused feature vector at time t, used to fuse information from the running state classification vector and the scheduling strategy matching degree matrix to form the final feature representation for classification. The calculation method is expressed as follows: ;

[0216] This represents the dynamic fusion weight at time t. The weights of the running state vector and the matching degree matrix in the fusion process are adjusted based on the gradient of the changing running state. The calculation method is expressed as follows: ;

[0217] The transpose of the weight vector representing the dynamically fused weights is a trainable parameter;

[0218] This is a matrix flattening operation;

[0219] Indicates a random deactivation operation;

[0220] The weight coefficient of the prototype comparison term is preferably set to 0.1. Adding the prototype comparison term to the classification loss encourages sample features to be close to the prototype of their fault category while moving away from the prototypes of other categories, thereby enhancing intra-class compactness and inter-class separability.

[0221] Indicates Euclidean distance;

[0222] Indicates the first The trainable prototype vector of a fault class is a trainable parameter that is updated along with the network parameters during network training and represents the typical location of the fault class in the feature space.

[0223] The loss function calculation can be specifically as follows:

[0224] During the training of the energy dispatch fault diagnosis model, the calculation of the loss function aims to quantify the difference between the fault categories predicted by the model and the actual categories, and to use this as a guide to optimize the model parameters.

[0225] This application uses the cross-entropy loss function as the total loss for model training, and the specific calculation process is as follows:

[0226] First, for each input training sample, the model passes through the multimodal dynamic normalization module, the feature interaction mining module, and the energy scheduling fault classification module in sequence through forward propagation. Finally, the scheduling-aware deep classification network outputs a probability distribution vector for all possible fault categories.

[0227] Then, the probability distribution vector is compared with the true fault category label of the sample (converted to one-hot encoding form), and the cross-entropy loss is calculated. The cross-entropy loss value measures the degree of mismatch between the probability distribution predicted by the model and the true distribution. The smaller the loss value, the more accurate the model prediction. This loss function serves as the target to be optimized during the model training phase. Through the backpropagation algorithm, the loss gradient is passed to each trainable part of the energy dispatch fault diagnosis model, driving the entire model to perform end-to-end joint learning and parameter updates.

[0228] The iterative training of the energy dispatch fault diagnosis model can be specifically as follows:

[0229] The energy dispatch fault diagnosis model is trained using an iterative method based on gradient descent.

[0230] At the start of training, all trainable parameters in the model are first randomly initialized. Then, the prepared training dataset is divided into multiple batches and fed into the model sequentially for iterative optimization.

[0231] In each iteration, a batch of sample data passes through the three core modules of the model sequentially. First, the data undergoes adaptive transformation by the multimodal dynamic normalization module to obtain a normalized data matrix. This matrix is ​​then input into the feature interaction mining module, which extracts the feature mining data matrix through joint learning of graph convolution and sparse coding. Finally, this feature matrix, along with external scheduling information, is sent to the energy scheduling fault classification module to generate the fault category prediction probability for each sample.

[0232] Then, the cross-entropy loss between the predicted results of all samples in the batch and the true labels is calculated, and the gradient of the loss with respect to each trainable parameter in the model is calculated from top to bottom using the backpropagation algorithm.

[0233] Then, using algorithms such as the adaptive moment estimator optimizer, all parameters are updated based on the calculated gradients to complete one iteration.

[0234] The model training continues iteratively until a preset stopping condition is met. The stopping condition is usually set based on the model's performance on an independent validation set. The main judgment conditions include: the fault classification accuracy on the validation set no longer improves in multiple consecutive training cycles, or the total loss value on the validation set no longer decreases significantly and tends to stabilize. In addition, a maximum number of training iteration cycles is set as a safety guarantee to prevent infinite loops. When any stopping condition is triggered, the iterative training process terminates. The model parameters saved at this time are the final trained energy dispatch fault diagnosis model.

[0235] In addition, once the energy dispatch fault diagnosis model has been trained, it can be put into practical application to realize the automatic diagnosis of faults in the dispatch and operation of the energy system.

[0236] During fault diagnosis, the system first collects or acquires real-time synchronous operation data of the source-grid-load-storage system within the time period to be diagnosed, following the same data specifications and procedures as during the training phase. Simultaneously, the corresponding scheduling plan and operational constraint information are also acquired. Then, this real-time data is input into the pre-trained diagnostic model, and the data flow automatically undergoes the same processing procedure as during training.

[0237] The system undergoes adaptive standardization via a multimodal dynamic normalization module to eliminate multimodal differences. Then, through a feature interaction mining module, deep features containing the physical topological relationships of the system are extracted. Finally, in the energy dispatch fault classification module, real-time dispatch information is combined for comprehensive analysis and decision-making, and the judgment of the system's operating status at the current moment, i.e., the predicted fault category label, is output.

[0238] In one embodiment, such as Figure 2As shown, robustness analysis was conducted under noisy conditions to examine the performance retention of each diagnostic method in the face of data quality degradation, i.e., noise interference, to verify the robustness of the model. The experiment constructed test conditions by simulating different levels of signal-to-noise ratio (SNR). The horizontal axis represents SNR in decibels (dB), with a decrease in value indicating increased noise and deteriorated data quality. The vertical axis represents the diagnostic accuracy percentage. The robustness analysis graphs of different methods under noisy conditions show the curves of performance changing with increasing noise. The performance comparison graph at an SNR of 5 dB specifically compares the performance under the extreme noise condition with the lowest SNR. The curves in the robustness analysis graphs of different methods under noisy conditions clearly show that as the SNR decreases (noise increases), the diagnostic accuracy of all methods shows a downward trend, but... Figure 2 The curve represented by the solid blue line shows the gentlest decline, indicating the slowest performance degradation and the strongest resistance to noise. In the low signal-to-noise ratio (SNR) region, the performance advantage of this technique over other methods is further amplified. Performance comparison chart at a SNR of 5dB. Figure 3 The bar chart provides more direct evidence that this technique can still maintain the highest diagnostic accuracy in a high-noise environment of 5dB, while the performance of other methods deteriorates significantly. This proves the effectiveness of the multimodal dynamic normalization module in this technique. By identifying the intrinsic modes of the data and performing adaptive correction, the model's adaptability to non-stationary and noisy data is enhanced, ensuring the stability of diagnostic performance under real-world complex conditions.

[0239] In one embodiment, such as Figure 4As shown, a comparative experiment on fault diagnosis accuracy was conducted, and the results are presented in a grouped bar chart to evaluate the performance advantages of the energy dispatch fault diagnosis method proposed in this application compared to other conventional fault diagnosis techniques. The experiment compared six fault diagnosis methods: Support Vector Machine (SVM), Random Forest, Gradient Boosting Tree (GPRS), Deep Neural Network (DNN), Long Short-Term Memory (LSTM) Network, and the dispatch-aware deep classification network proposed in this application. SVM, Random Forest, and GPRS are conventional machine learning classification algorithms; Deep Neural Network and LSM are deep learning methods, the former handling static features and the latter specifically handling time-series data. In terms of experimental configuration, the diagnostic accuracy of each method was tested on five fault types. The horizontal axis represents the fault type, including normal, overvoltage, undervoltage, frequency deviation, and line overload; the vertical axis represents the diagnostic accuracy, in percentage. Experimental results show that conventional machine learning methods have relatively low accuracy on complex fault types, while deep learning methods have improved accuracy. However, the method proposed in this application achieves the highest accuracy on all fault types, proving that by constructing a running state classification vector, calculating the scheduling strategy matching degree matrix, and designing a scheduling-aware deep classification network, this application can effectively distinguish between normal operating condition adjustments caused by scheduling instructions and real faults. Combined with the high-quality features provided by the multimodal dynamic normalization and feature interaction mining modules, it achieves accurate and robust diagnosis of energy system faults.

[0240] In addition to the above, which are method embodiments of this application, based on the same application concept, this application also provides a data-driven energy dispatch fault diagnosis device. Figure 5 As shown, the device includes: a processor; and a memory storing executable code thereon, which, when executed, causes the processor to perform a data-driven energy dispatch fault diagnosis method as described in the above embodiments.

[0241] Specifically, the server constructs a synchronous energy array based on the multi-source synchronous operation data corresponding to the source-grid-load-storage system; obtains the labeled fault categories corresponding to the synchronous energy array, and then constructs training samples; inputs the training samples into the diagnostic model, and normalizes the multi-source synchronous operation data in the synchronous energy array of the training set; constructs a connection relationship matrix between the multi-source synchronous operation data; constructs a normalized adjacency matrix based on the normalized synchronous energy array and the connection relationship matrix; performs information aggregation and transformation on the normalized data matrix through a multi-head graph attention mechanism, splices the preset historical time window features, obtains the graph convolutional hidden representation matrix, and then calculates the sparse dictionary matrix; based on the sparse dictionary matrix and the desparse coding coefficients, optimizes the graph convolutional sparse coding objective function, solves the sparse coding coefficients, and outputs the feature mining data matrix; based on the feature mining data matrix, uses the introduced scheduling plan data and real-time operation constraints to calculate the scheduling strategy matching degree matrix; fuses the operation status classification vector and the scheduling strategy matching degree matrix to obtain the predicted fault category; iteratively trains the diagnostic model based on the predicted fault category and the labeled fault category until a well-trained diagnostic model is obtained.

[0242] In addition, this application embodiment also provides a non-volatile computer storage medium storing executable instructions, which, when executed, implement the data-driven energy dispatch fault diagnosis method described above.

[0243] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A data-driven method for energy dispatch fault diagnosis, characterized in that, The method includes: Based on the multi-source synchronous operation data corresponding to the source-grid-load-storage system, a synchronous energy array is constructed; the labeled fault categories corresponding to the synchronous energy array are obtained, and then training samples are constructed. The training samples are input into the diagnostic model, and the multi-source synchronous operation data in the synchronous energy array of the training set are normalized; a connection relationship matrix between the multi-source synchronous operation data is constructed; and a normalized adjacency matrix is ​​constructed based on the normalized synchronous energy array and the connection relationship matrix. The normalized data matrix is ​​aggregated and transformed by a multi-head graph attention mechanism, and the features of a preset historical time window are spliced ​​together to obtain the graph convolutional hidden representation matrix, and then the sparse dictionary matrix is ​​calculated. Based on the sparse dictionary matrix, the objective function of graph convolution sparse coding is optimized, the sparse coding coefficients are solved, and the feature mining data matrix is ​​output. Specifically, the reconstruction error between the normalized data matrix and the product of the sparse dictionary matrix and the sparse coding coefficient matrix is ​​minimized, and the sparsity constraint of the sparse coding coefficient matrix and the graph smoothing regularization constraint of the sparse dictionary matrix are added at the same time. The optimized graph convolution sparse coding objective function is obtained by joint learning. The sparse coding coefficient matrix is ​​updated in each iteration using an iterative soft thresholding algorithm based on the sparse dictionary matrix, the normalized data matrix, and the graph Laplacian matrix, thus obtaining the sparse coding coefficient matrix. The feature mining data matrix is ​​obtained by multiplying the sparse dictionary matrix and the sparse coding coefficient matrix, combining the linear transformation of the graph convolution hidden representation matrix, and then processing it with the modified linear unit activation function. Based on the feature mining data matrix, the scheduling strategy matching degree matrix is ​​calculated using the introduced scheduling plan data and real-time operation constraints; the running status classification vector and the scheduling strategy matching degree matrix are fused to obtain the predicted fault category; the diagnostic model is iteratively trained according to the predicted fault category and the labeled fault category until a well-trained diagnostic model is obtained. The dynamic construction of the operational status classification vector includes: based on the feature mining data matrix, introducing a comprehensive operational status index and its real-time gradient, and combining multimodal hybrid weights to construct an operational status classification vector that can dynamically reflect the system's operational health and stability, providing richer discrimination criteria for classification, expressed as: In the formula, This represents the operational status classification vector constructed at time point t, which characterizes the comprehensive operational status by integrating instantaneous features, state change trends, and multimodal information. This represents a point-in-time index, with a value range of [value range missing]. arrive ; Representing the feature mining data matrix In the The row vector at each time point; This represents a vector concatenation operation; This represents the comprehensive index of the operating status at time point t. The discrete gradient at time point t is used to measure the drasticness of the instantaneous change in the operating state. This indicates that the sample at time t belongs to the th time. Mixed weights for each modality; Indicates the relationship with the first A state mapping vector associated with each modality.

2. The data-driven energy dispatch fault diagnosis method according to claim 1, characterized in that, The synchronization energy array in the training set is normalized, specifically including: The synchronous energy array is mapped to the latent space by a variational autoencoder to obtain the latent variable representation; Clustering is performed on the latent variable representations to obtain the clustering results and the number of modes M; Based on the clustering results, calculate the center vector and standard deviation vector for each mode; The encoder of the variational autoencoder maps the synchronization energy array to a latent variable representation while outputting the mean and variance of a Gaussian distribution. After connecting to the soft classification layer, it outputs an M-dimensional vector, with the components representing the mixing weights of each mode. The nonlinear correction term is calculated based on the pre-defined low-dimensional representation of the synchronous energy array in the latent space. Calculate the normalized data matrix based on the nonlinear correction term, the synchronous energy array, the standard deviation vector, and the center vector.

3. The data-driven energy dispatch fault diagnosis method according to claim 1, characterized in that, Construct a connection matrix among data sources running synchronously from multiple sources; Based on the normalized synchronous energy array and connectivity matrix, a normalized adjacency matrix is ​​constructed, specifically including: Based on the physical topology of the source-grid-load-storage system, a connection matrix is ​​constructed among the synchronously operating data from multiple sources. Through the formula: Construct a normalized adjacency matrix ; in, This represents a normalized adjacency matrix, where the element values ​​are in... Within the range; This represents the self-loop enhanced connectivity matrix; express The degree matrix; Degree matrix The negative half power is obtained by taking the reciprocal of the square root of the diagonal elements.

4. The data-driven energy dispatch fault diagnosis method according to claim 1, characterized in that, The normalized data matrix is ​​aggregated and transformed using a multi-head graph attention mechanism, and features from a preset historical time window are concatenated to obtain the graph convolutional hidden representation matrix. Then, the sparse dictionary matrix is ​​calculated, specifically including: Through the formula: Compute the graph convolution hidden representation matrix; in, Representation of the graph convolution hidden representation matrix; This represents the global weight parameter matrix of the graph convolutional layer; Indicates the attention head index; This represents the total number of attention heads in a multi-head attention mechanism; Presentation layer normalization operation; Indicates the first The normalized adjacency matrix is ​​obtained by calculating the attention head; This represents the global bias vector of the graph convolutional layer; The concatenation operation is indicated; This represents the max pooling operation; This represents the feature matrix within a preset historical time window; This represents the modified linear unit activation function; The basic dictionary representation is obtained by linearly transforming the graph convolutional hidden representation matrix with the dictionary learning weight matrix and bias vector and then processing it with the Softmax function. At the same time, the residual terms and gating signals are calculated based on the graph convolutional hidden representation. Then, the basic dictionary representation is combined with the residual terms adjusted by the gating mechanism to obtain the sparse dictionary matrix.

5. The data-driven energy dispatch fault diagnosis method according to claim 1, characterized in that, Based on the feature mining data matrix, and utilizing the introduced scheduling plan data and real-time operational constraints, a scheduling policy matching degree matrix is ​​calculated, specifically including: Based on the feature mining data matrix, a comprehensive index of operating status and its real-time change gradient are introduced, and a classification vector of operating status is constructed by combining multimodal hybrid weights. By calculating the bidirectional attention matching degree between the feature mining data matrix and the scheduling plan data and real-time running constraints, a scheduling strategy matching degree matrix is ​​constructed.

6. The data-driven energy dispatch fault diagnosis method according to claim 1, characterized in that, By fusing the operational status classification vector with the scheduling strategy matching degree matrix, the predicted fault category is obtained; Based on the predicted fault category and the labeled fault category, the diagnostic model is iteratively trained until a well-trained diagnostic model is obtained. Specifically, this includes: The running status classification vector and the flattened vector of the scheduling policy matching degree matrix are fused together, and the classification is performed by a deep network to obtain the predicted fault category. The cross-entropy loss function is used as the total loss during model training. The diagnostic model is trained using an iterative method based on gradient descent. When a preset stopping condition is met, the trained diagnostic model is obtained.

7. A data-driven energy dispatch fault diagnosis device, characterized in that, The device includes: processor; And a memory having executable code stored thereon, which, when executed, causes the processor to perform a data-driven energy dispatch fault diagnosis method as described in any one of claims 1-6.

8. A non-volatile computer storage medium, characterized in that, It stores computer instructions, which, when executed, implement a data-driven energy dispatch fault diagnosis method as described in any one of claims 1-6.

Images