Virtual power plant platform source network load storage equipment real-time monitoring and optimizing method and system

Abstract

The invention provides a virtual power plant platform source network load storage equipment real-time monitoring and optimization method and system, and relates to the technical field of virtual power plant intelligent optimization scheduling, and the method comprises the steps: collecting equipment operation data, carrying out the noise reduction and feature compression to obtain feature data, and recognizing the equipment fault type based on the feature data; carrying out decoupling reconstruction on the feature data in time and space dimensions to obtain an equipment incidence matrix, constructing an equipment topological graph based on the equipment incidence matrix and a credible fault type, carrying out bidirectional tracking on the equipment topological graph according to a fault diffusion probability, and generating a fault propagation path; grading and sorting the equipment on the fault propagation path, and outputting a fault influence prediction result and an equipment priority list; and carrying out cooperative scheduling optimization on the equipment in the virtual power plant by adopting a hierarchical reinforcement learning algorithm to realize cooperative scheduling optimization of the equipment in the virtual power plant.

Description

Technical Field

[0001] The present invention relates to the intelligent optimization scheduling technology of virtual power plants, and particularly to a real-time monitoring and optimization method and system for source-network-load-storage equipment of a virtual power plant platform. Background Art

[0002] Existing power dispatch systems mainly rely on traditional monitoring methods to collect and analyze the operating states of source-network-load-storage equipment. However, in a complex operating environment, the equipment data is greatly interfered by noise, resulting in low data quality and affecting the accuracy of fault identification. At the same time, traditional methods usually perform fault diagnosis based on fixed rules or empirical judgments, and it is difficult to make full use of historical data for accurate matching, resulting in large errors in the identification of equipment fault types and being unable to effectively evaluate the risk degree and credibility of faults.

[0003] In addition, current equipment monitoring methods mostly focus on the independent analysis of individual equipment, ignoring the dynamic coupling relationship between equipment. As a result, when a equipment fault occurs, it is difficult to accurately predict its propagation path and influence range, unable to effectively evaluate the diffusion trend of the fault, and affecting the efficiency of fault early warning and emergency dispatch. At the same time, existing dispatch optimization methods are mostly based on static dispatch strategies and fail to dynamically adjust equipment combinations and control strategies in combination with real-time fault analysis results, making it difficult to meet the optimal dispatch requirements in the case of sudden faults.

[0004] Therefore, there is an urgent need for a method that can improve the equipment data processing ability, accurately identify faults and dynamically evaluate their propagation impacts, and at the same time combine optimization dispatch strategies to improve the system stability, so as to improve the operation reliability and dispatch efficiency of source-network-load-storage equipment. Summary of the Invention

[0005] An embodiment of the present invention provides a real-time monitoring and optimization method and system for source-network-load-storage equipment of a virtual power plant platform, which can solve the problems in the prior art.

[0006] In the first aspect of the embodiment of the present invention, A real-time monitoring and optimization method for source-network-load-storage equipment of a virtual power plant platform is provided, including: Collecting equipment operation data and performing noise reduction and feature compression to obtain feature data, identifying the equipment fault type based on the feature data, matching the equipment fault type with historical fault data to generate a fault risk score, and verifying the credibility of the equipment fault type based on the fault risk score to obtain a credible fault type; Decouple and reconstruct the feature data in the time and space dimensions to obtain the device association matrix. Based on the device association matrix and the credible fault types, construct a device topology graph, dynamically partition the device topology graph and extract local features, and calculate the fault diffusion probability through feature transfer between regions; perform bidirectional tracking on the device topology graph according to the fault diffusion probability to generate a fault propagation path; rank the devices on the fault propagation path hierarchically and output the fault impact prediction result and the device priority list; Adopt a hierarchical reinforcement learning algorithm to optimize the cooperative scheduling of devices in a virtual power plant. Among them, the upper-layer agent generates a device combination plan according to the fault impact prediction result and the device priority list, and the lower-layer agent generates device control instructions based on the device combination plan and the fault risk score to achieve the cooperative scheduling optimization of the virtual power plant devices.

[0007] In an alternative embodiment, Collect device operation data and perform noise reduction and feature compression to obtain feature data. Identify the device fault types based on the feature data, match the device fault types with the historical fault data to generate a fault risk score, and verify the credibility of the device fault types based on the fault risk score. The credible fault types obtained include: Perform wavelet transform on the device operation data to obtain multi-scale coefficients, calculate the ratio of adjacent scale coefficients to construct a feature matrix; perform singular value decomposition on the feature matrix to obtain a signal subspace and a noise subspace, calculate the orthogonal vector product of the signal subspace and the noise subspace to obtain an adaptive threshold, project the device operation data onto the signal subspace and reconstruct it based on the adaptive threshold to obtain noise-reduced data; Construct an autoencoder model, where the encoding part maps the noise-reduced data to intermediate features, and obtains latent variables by passing the intermediate features through a non-linear activation function; construct a discriminator network, input the latent variables and a preset Gaussian distribution sample into the discriminator network to obtain a discrimination probability, and calculate the cross-entropy based on the discrimination probability and the true label to obtain an adversarial loss; perform gradient descent optimization on the latent variables according to the adversarial loss to obtain fault features; Perform sliding convolution on the fault features through a one-dimensional convolution kernel to obtain a time series feature sequence, and use a gated recurrent unit to calculate the state vector for each time step of the time series feature sequence; calculate the conditional entropy of the state vector at different time steps to obtain time series weights, perform weighted summation of the time series weights and the time series feature sequence to obtain global features, and input the global features into a softmax classifier to obtain the initial fault type; Stratify historical fault samples according to operating condition parameters based on conditional entropy, calculate the fault type transition frequency within each layer of samples to obtain a state transition matrix, splice the global features and the current operating condition parameters and input them into the state transition matrix to obtain a prediction probability, calculate the mutual information between the initial fault type and the prediction probability to obtain a credibility score, and confirm it as a credible fault type when the credibility score exceeds a preset threshold.

[0008] In an alternative embodiment, Stratifying historical fault samples according to operating condition parameters based on conditional entropy, calculating the fault type transition frequency within each layer of samples to obtain a state transition matrix, splicing the global features and the current operating condition parameters and inputting them into the state transition matrix to obtain a prediction probability, and calculating the mutual information between the initial fault type and the prediction probability to obtain a credibility score includes: Divide the numerical interval of the operating condition parameters into multiple sub-intervals, count the occurrence frequency of the fault type within each sub-interval, and calculate the conditional entropy of the operating condition parameters with respect to the fault type; calculate the mean and standard deviation of the conditional entropy, and obtain an adaptive threshold based on the mean minus the standard deviation; construct a stratified parameter set for the operating condition parameters with conditional entropy less than the adaptive threshold; Calculate the local density and minimum distance of the historical fault samples in the space corresponding to the stratified parameter set, determine the density peak point as the clustering center based on the local density, and assign the sample points to the clustering center based on the minimum distance to obtain the operating condition layer; Within each operating condition layer, calculate the total number of samples of the fault type and the number of transition samples between fault types; when the total number of samples is greater than the preset sample number threshold, calculate the ratio of the number of transition samples to the total number of samples to obtain the fault type transition frequency, otherwise set the fault type transition frequency to zero; construct a state transition matrix based on the fault type transition frequency; Obtain the global features and the current operating condition parameters of the sample to be verified, and determine the target operating condition layer to which it belongs based on the current operating condition parameters; combine the global features and the current operating condition parameters to form a feature vector, and input it into the state transition matrix corresponding to the target operating condition layer to obtain a prediction probability distribution; Obtain the initial fault type probability distribution of the sample to be verified, divide the probability of each fault type in the initial fault type probability distribution by the probability of the corresponding fault type in the prediction probability distribution to obtain a probability ratio; obtain the credibility score based on the sum of the products of the logarithm of the probability ratio and the initial fault type probability.

[0009] In an alternative embodiment, Decouple and reconstruct the feature data in the time and space dimensions to obtain a device association matrix, construct a device topology graph based on the device association matrix and the credible fault types, perform dynamic partitioning on the device topology graph and extract local features, calculate the fault diffusion probability through feature transfer between regions; perform bidirectional tracking on the device topology graph according to the fault diffusion probability to generate a fault propagation path; rank the devices on the fault propagation path hierarchically and output the fault impact prediction result and the device priority list, including: Decompose the feature data in the time dimension, decompose the feature data into hourly subsequences, daily subsequences and weekly subsequences, extract the periodic patterns and state trends in the hourly subsequences, daily subsequences and weekly subsequences, and construct a time dimension feature vector; Extract the physical connection distance, electrical loop connection and device functional relationship in the feature data, calculate the correlation coefficient between devices, and construct a space dimension feature vector; Perform a tensor fusion operation on the time dimension feature vector and the space dimension feature vector to obtain a device association matrix; Construct a device topology graph including device nodes and connection edges based on the device association matrix and the credible fault types, perform dynamic partitioning on the device topology graph using a spectral clustering algorithm with adaptive similarity to obtain multiple device partitions, calculate the device density distribution within each device partition to obtain the eigenvalue within the region, and calculate the connection edge weight between the device partitions to obtain the eigenvalue between regions; construct a transfer function including a time delay factor and an attenuation coefficient, input the eigenvalue within the region and the eigenvalue between regions into the transfer function, and calculate the fault diffusion probability between device partitions; Mark the position of the faulty device as the starting node in the device topology graph, perform forward propagation calculation based on the fault diffusion probability to obtain an impact propagation path; mark the device nodes on the impact propagation path with a fault diffusion probability exceeding a preset probability threshold as the positions of affected devices, calculate the transfer importance of each node in the device topology graph, and obtain the position of the source device through backward tracking; Combine and optimize the impact propagation path and the position of the source device to obtain a fault propagation path, calculate the risk level of each device on the fault propagation path for hierarchical ranking, and output the fault impact prediction result and the device priority list.

[0010] In an alternative embodiment, Construct a transfer function including a time delay factor and an attenuation coefficient, input the eigenvalue within the region and the eigenvalue between regions into the transfer function, and calculate the fault diffusion probability between device partitions, including: Obtain the physical propagation delay and response delay between devices, classify them according to the transfer type to obtain a delay classification result, perform segmented mapping on the delay classification result to obtain a delay mapping value, and construct a time delay factor based on the delay mapping value; Collect the physical distance, electrical impedance, and functional correlation degree between the acquisition devices, calculate the spatial attenuation value according to the physical distance, calculate the signal attenuation value according to the electrical impedance, calculate the coupling attenuation value according to the functional correlation degree, perform hierarchical combination on the spatial attenuation value, signal attenuation value, and coupling attenuation value to obtain an attenuation combination value, and construct an attenuation coefficient based on the attenuation combination value; construct a non-linear transfer function according to the time delay factor and the attenuation coefficient; Calculate the spatial distribution density value and the distance distribution value between devices in the area, calculate the area density feature according to the spatial distribution density value and the distance distribution value, and perform feature fusion on the area density feature and the device operation parameters to obtain the in-area feature value; calculate the physical connection quantity value and the energy transfer power value of the devices at the area boundary, calculate the connection transmission feature according to the physical connection quantity value and the energy transfer power value, and perform feature fusion on the connection transmission feature and the device operation state to obtain the inter-area feature value; establish device connection constraints and energy transfer constraints according to the in-area feature value and the inter-area feature value; Classify the historical fault samples according to the fault type and propagation path to obtain training samples, input the training samples into the non-linear transfer function, and perform iterative training using the device connection constraints and energy transfer constraints to obtain the correction values of the time delay factor and the attenuation coefficient; input the in-area feature value and the inter-area feature value into the non-linear transfer function, and perform dynamic correction on the propagation calculation result according to the correction values, and output the fault diffusion probability between device partitions.

[0011] In an alternative embodiment, Adopt a hierarchical reinforcement learning algorithm to perform collaborative scheduling optimization on the devices in the virtual power plant. Among them, the upper-layer agent generates a device combination plan according to the fault impact prediction result and the device priority list, and the lower-layer agent generates a device control instruction based on the device combination plan and the fault risk score. The realization of the collaborative scheduling optimization of the virtual power plant devices includes: Obtain the device fault impact prediction value, device operation state value, and device priority score, construct the upper-layer state space, and construct the upper-layer action space with the device combination plan; construct the upper-layer reward function based on the weighted combination of the system reliability index, operation efficiency index, and scheduling cost; Construct the lower-layer state space with the device combination plan, device fault risk score, and real-time operation parameters, and construct the lower-layer action space with the device control instruction; construct the lower-layer reward function based on the dynamic weight combination of the system stability index, response time index, and the fault risk score; Calculate the collaborative impact degree of devices based on the predicted results of fault impacts and the priority list, and use the collaborative impact degree of devices as a weight coefficient to construct weighted state features; obtain historical scheduling data to calculate the scheduling correlation strength between devices, adaptively adjust the weighted state features based on the scheduling correlation strength, input the adjusted features into the upper-layer state space, calculate the upper-layer state-action value according to the upper-layer reward function, and generate a device combination plan; Perform deep feature fusion on the device combination plan and the fault risk score to obtain device correlation features, input the device correlation features into the lower-layer state space, calculate the lower-layer state-action value based on the lower-layer reward function, and generate device control instructions; Execute the device control instructions to obtain the control results, calculate the reward value according to the control results, determine the sampling weights of the experience replay pool based on the historical reward distributions of different scheduling scenarios, and sample training samples from the historical data; online update the collaborative impact degree and the scheduling correlation strength based on the training samples and the reward value; Input the updated collaborative impact degree and the scheduling correlation strength into the upper-layer agent and the lower-layer agent respectively, generate an optimized device combination plan and optimized device control instructions, and realize the collaborative scheduling optimization of virtual power plant devices.

[0012] In an alternative embodiment, Execute the device control instructions to obtain the control results, calculate the reward value according to the control results, determine the sampling weights of the experience replay pool based on the historical reward distributions of different scheduling scenarios, and sample training samples from the historical data; online update the collaborative impact degree and the scheduling correlation strength based on the training samples and the reward value includes: Execute the device control instructions to obtain the control results composed of power balance degree, response time, energy loss, and system stability; calculate the system reliability score based on the power balance degree and the system stability, calculate the scheduling efficiency score based on the response time and the energy loss, and use the weighted combination of the system reliability score and the scheduling efficiency score as the reward value; Classify and store the state transition samples according to the scheduling scenarios of different device combination plans to construct an experience replay pool, where the state transition samples include device combination plans, device control instructions, reward values, and post-control states; determine the sampling weights according to the historical reward distributions of the scheduling scenarios, and select training samples from the experience replay pool based on the sampling weights; Extract the response timing features and power coupling features between devices from the device combination plans in the training samples to construct a device collaborative impact matrix; construct a collaborative impact degree loss function according to the temporal correlation between the device collaborative impact matrix and the reward value, and optimize the collaborative impact degree loss function to online update the collaborative impact degree; Extract the scheduling dependency features and control coupling features between devices based on the device control instructions in the training samples, and construct a device scheduling association matrix; construct a scheduling association strength loss function according to the dynamic correlation between the device scheduling association matrix and the reward value, and optimize the scheduling association strength loss function to update the scheduling association strength online.

[0013] In the second aspect of the embodiments of the present invention, Provide a real-time monitoring and optimization system for source-network-load-storage devices of a virtual power plant platform, including: The first unit is used to collect device operation data, perform noise reduction and feature compression to obtain feature data, identify device fault types based on the feature data, match the device fault types with historical fault data to generate a fault risk score, and verify the credibility of the device fault types based on the fault risk score to obtain credible fault types; The second unit is used to decouple and reconstruct the feature data in the time and space dimensions to obtain a device association matrix, construct a device topology map based on the device association matrix and the credible fault types, dynamically partition the device topology map and extract local features, calculate the fault diffusion probability through feature transfer between regions; perform two-way tracking on the device topology map according to the fault diffusion probability to generate a fault propagation path; rank the devices on the fault propagation path hierarchically, and output a fault impact prediction result and a device priority list; The third unit is used to perform collaborative scheduling optimization on the devices in the virtual power plant by using a hierarchical reinforcement learning algorithm. Among them, the upper-layer agent generates a device combination plan according to the fault impact prediction result and the device priority list, and the lower-layer agent generates device control instructions based on the device combination plan and the fault risk score to achieve the collaborative scheduling optimization of the virtual power plant devices.

[0014] In the third aspect of the embodiments of the present invention, Provide an electronic device, including: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0015] In the fourth aspect of the embodiments of the present invention, Provide a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0016] In this embodiment, by denoising and feature compression of the device operation data, the data quality is improved, noise interference is reduced, and the accuracy of fault identification is ensured. At the same time, through the historical fault data matching and risk scoring mechanism, the types of device faults can be accurately identified, the credibility of fault diagnosis is improved, misjudgment and missed judgment are avoided, and thus the accuracy and reliability of device operation and maintenance are improved. This solution constructs an association matrix based on device feature data, which can dynamically analyze the coupling relationship between devices, and through topology graph analysis and fault diffusion probability calculation, predict the fault propagation path. This can not only early warn the devices that may be affected, but also rank them according to the degree of influence between devices, facilitating accurate positioning of the key fault sources, improving the pertinence and efficiency of fault handling, and reducing the impact of faults on the entire system. In terms of scheduling optimization, a hierarchical reinforcement learning algorithm is adopted to achieve the coordinated optimization scheduling of the source-network-load-storage devices. The upper-layer intelligent agent formulates a reasonable device combination plan according to the fault impact prediction results and the device priority list, and the lower-layer intelligent agent generates specific control instructions based on the device combination plan and the risk score, so as to realize the dynamic optimization control of the devices, improve the intelligent level of scheduling, and ensure the safety, economy and stability of the power system under complex working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a schematic flow chart of the real-time monitoring and optimization method for the source-network-load-storage devices of the virtual power plant platform in the embodiment of the present invention; Figure 2 is an effect simulation diagram of the real-time monitoring and optimization method for the source-network-load-storage devices of the virtual power plant platform in the embodiment of the present invention; Figure 3 is an analysis diagram of the fault propagation time delay and attenuation characteristics in the embodiment of the present invention; Figure 4 is a schematic structural diagram of the real-time monitoring and optimization system for the source-network-load-storage devices of the virtual power plant platform in the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0019] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0020] Figure 1 This is a schematic flowchart of the real-time monitoring and optimization method for the virtual power plant platform's source-network-load-storage equipment in the embodiments of the present invention. As Figure 1 shown, the method includes: Collect equipment operation data, perform noise reduction and feature compression to obtain feature data, identify the equipment failure type based on the feature data, match the equipment failure type with historical failure data for similarity, generate a failure risk score, and verify the credibility of the equipment failure type based on the failure risk score to obtain a credible failure type; Decouple and reconstruct the feature data in the time and space dimensions to obtain an equipment association matrix, construct an equipment topology map based on the equipment association matrix and the credible failure type, perform dynamic partitioning on the equipment topology map and extract local features, calculate the fault diffusion probability through feature transfer between regions; perform two-way tracking on the equipment topology map according to the fault diffusion probability to generate a fault propagation path; rank the equipment on the fault propagation path hierarchically and output the fault impact prediction result and the equipment priority list; Adopt a hierarchical reinforcement learning algorithm to optimize the coordinated scheduling of equipment in the virtual power plant. Among them, the upper-layer agent generates an equipment combination plan according to the fault impact prediction result and the equipment priority list, and the lower-layer agent generates equipment control instructions based on the equipment combination plan and the failure risk score to achieve the coordinated scheduling optimization of the virtual power plant equipment.

[0021] Among them, the equipment operation data refers to the time-series data generated by various equipment in the virtual power plant during operation, including but not limited to parameters such as voltage, current, power, temperature, frequency, and equipment operation status. These data may contain noise and redundant information, so noise reduction and feature compression processing are required before analysis to remove irrelevant information and reduce the computational complexity.

[0022] The feature data refers to the key information extracted from the equipment operation data that can characterize the equipment operation status. After noise reduction and feature compression, the original data is transformed into a data set with a smaller dimension but still retaining the essential features of the equipment operation, which is convenient for subsequent analysis and modeling.

[0023] The equipment failure type is the classification result of the abnormal states or operation failures that the equipment may encounter, usually including short-circuit faults, overload faults, over-temperature faults, electrical faults, etc. The identification of the failure type is based on the feature data and combined with historical data for classification to determine the current failure category of the equipment.

[0024] The historical failure data refers to the failure information recorded by the equipment during past operation, including failure type, occurrence time, impact range, repair measures, etc. This data is used for similarity matching with the currently detected failure type to evaluate the potential risk of the current failure.

[0025] The fault risk score is a metric calculated based on the similarity between the current device fault types and historical fault data. This score is used to quantify the severity of the device fault and the potential impact the fault may have on the system. A higher score indicates a greater risk of the fault.

[0026] A credible fault type refers to a fault category verified through the fault risk score. Since there may be misjudgments during the fault identification process, it is necessary to verify the credibility of the initially identified fault types based on historical data to ensure the reliability of the final result.

[0027] The device topology map is a network structure diagram constructed based on the device association matrix and credible fault types, where nodes represent devices and edges represent the association relationships between devices. This topology map is used to analyze the impact propagation paths between devices, especially when a fault occurs, to help identify other devices that may be affected.

[0028] Dynamic partitioning refers to dividing the device network of the entire virtual power plant into several local areas using an adaptive algorithm based on the structure of the device topology map and the device operating status. This method can reduce the computational complexity and ensure that the affected areas can be focused on during fault analysis.

[0029] Local features refer to the key feature information extracted within a specific partition of the device topology map. These features are used to describe the operating status of devices within the area, the fault distribution, and the potential scope of fault impact. Inter-region feature transfer refers to the process of information exchange between different dynamic partitions of the device topology map. Since there are mutual influences between devices, the analysis of a single region may not comprehensively reflect the state of the entire system. Therefore, it is necessary to transfer fault-related features between regions to calculate the possibility of fault diffusion.

[0030] Device hierarchical ranking is a process of evaluating the priority of devices based on the fault propagation path. Devices more severely affected by the fault will be assigned a higher priority so that they can be given priority consideration for adjustment or protection during scheduling optimization.

[0031] The fault impact prediction result is the predicted analysis result of the possible scope and degree of fault impact, which combines the fault propagation path, device priority, and historical fault impact patterns to help the scheduling system make preventive decisions.

[0032] The device priority list is a list of device priorities generated based on the device hierarchical ranking result, which contains the scheduling priority information of each device. During the scheduling optimization process, high-priority devices are usually given priority to be included in the scheduling plan to reduce system risks.

[0033] In an alternative embodiment, Collect the operation data of the device, perform noise reduction and feature compression on the data to obtain feature data, identify the device fault type based on the feature data, match the device fault type with the historical fault data for similarity, generate a fault risk score, and verify the credibility of the device fault type based on the fault risk score. The obtained credible fault types include: Perform wavelet transform on the device operation data to obtain multi-scale coefficients, calculate the ratio of adjacent scale coefficients to construct a feature matrix; perform singular value decomposition on the feature matrix to obtain a signal subspace and a noise subspace, calculate the orthogonal vector product of the signal subspace and the noise subspace to obtain an adaptive threshold, project the device operation data onto the signal subspace and reconstruct it based on the adaptive threshold to obtain noise-reduced data; Construct an autoencoder model, where the encoding part maps the noise-reduced data to intermediate features, and obtains latent variables by passing the intermediate features through a non-linear activation function; construct a discriminator network, input the latent variables and a preset Gaussian distribution sample into the discriminator network to obtain a discrimination probability, calculate the cross-entropy based on the discrimination probability and the true label to obtain an adversarial loss; optimize the latent variables by gradient descent according to the adversarial loss to obtain fault features; Perform sliding convolution on the fault features through a one-dimensional convolution kernel to obtain a time series feature sequence, and use a gated recurrent unit to calculate the state vector for each time step of the time series feature sequence; calculate the conditional entropy of the state vector at different time steps to obtain time series weights, perform weighted summation of the time series weights and the time series feature sequence to obtain global features, and input the global features into a softmax classifier to obtain an initial fault type; Stratify the historical fault samples based on the operating condition parameters according to the conditional entropy, calculate the fault type transition frequency within each layer of samples to obtain a state transition matrix, input the global features concatenated with the current operating condition parameters into the state transition matrix to obtain a prediction probability, calculate the mutual information between the initial fault type and the prediction probability to obtain a credibility score, and confirm it as a credible fault type when the credibility score exceeds a preset threshold.

[0034] Exemplarily, first, perform wavelet transform on the device operation data to decompose the original signal into coefficients of different scales to extract information in each frequency band. Calculate the ratio of adjacent scale coefficients and construct a feature matrix based on the ratio to maintain the multi-scale characteristics of the data. On this basis, perform singular value decomposition on the feature matrix to divide the data into a signal subspace and a noise subspace. By calculating the orthogonal vector product of the signal subspace and the noise subspace, an adaptive threshold is obtained, and the original device operation data is projected onto the signal subspace using this threshold to remove noise and reconstruct the noise-reduced data.

[0035] Construct an autoencoder model to further extract features from the denoised data. The encoding part maps the denoised data into intermediate features and obtains latent variables through a non-linear activation function. To enhance the generalization ability of the model, a discriminator network is introduced. This network takes the latent variables and a preset Gaussian distribution sample as inputs and outputs a discrimination probability. Calculate the cross-entropy loss based on the discrimination probability and the true label, and optimize the latent variables through adversarial training to obtain more discriminative fault features.

[0036] Perform temporal modeling on the fault features, using a one-dimensional convolutional kernel for sliding convolution operations to capture the local temporal features of the data. Use a gated recurrent unit to calculate the temporal features step by step over time to obtain state vectors, and determine the importance of the temporal features by calculating the conditional entropy at different time steps. Weightedly sum the state vectors at each time step according to their importance to obtain global features, and input the global features into a softmax classifier to identify the initial fault type of the device.

[0037] Perform hierarchical processing on the historical fault data based on the conditional entropy, and divide the sample set according to the operating condition parameters. For each hierarchical set, count the transition frequencies between different fault types to construct a state transition matrix. Concatenate the global features with the current operating condition parameters and input them into the state transition matrix to calculate the prediction probabilities of each fault type. Calculate the credibility score using the mutual information between the initial fault type and the prediction probabilities. When the credibility score exceeds the set threshold, confirm that the fault type is a credible fault type.

[0038] In this embodiment, the multi-scale feature analysis method is adopted, which improves the accuracy of signal processing and avoids the defect that key information may be lost in traditional methods. By constructing an autoencoder and a discriminator network, the extraction of fault features is made more accurate, overcoming the problem of insufficient learning ability for high-dimensional data in existing methods. Combining the temporal analysis method effectively solves the problem of time information loss caused by single-frame data analysis, enabling the accurate capture of the fault evolution process. In addition, the introduction of statistical analysis of historical fault data makes fault identification under different operating conditions more reasonable, improving the stability and reliability of prediction. Overall, this solution not only improves the accuracy of equipment fault detection but also reduces the false alarm rate, making equipment management more intelligent.

[0039] In an alternative embodiment, Performing hierarchical processing on historical fault samples according to operating condition parameters based on conditional entropy, calculating the fault type transition frequency within each layer of samples to obtain a state transition matrix, inputting the global features concatenated with the current operating condition parameters into the state transition matrix to obtain prediction probabilities, and calculating the mutual information between the initial fault type and the prediction probabilities to obtain a credibility score includes: Divide the numerical range of the operating condition parameters into multiple sub - ranges, count the occurrence frequency of the fault types in each sub - range, and calculate the conditional entropy of the operating condition parameters with respect to the fault types; calculate the mean and standard deviation of the conditional entropy, and obtain the adaptive threshold based on the mean minus the standard deviation; construct a hierarchical parameter set for the operating condition parameters whose conditional entropy is less than the adaptive threshold. Calculate the local density and minimum distance of the historical fault samples in the space corresponding to the hierarchical parameter set, determine the density peak points as the clustering centers based on the local density, and assign the sample points to the clustering centers based on the minimum distance to obtain the operating condition layers. Within each operating condition layer, calculate the total number of samples of the fault types and the number of transfer samples between the fault types; when the total number of samples is greater than the preset sample number threshold, calculate the ratio of the number of transfer samples to the total number of samples to obtain the fault type transfer frequency, otherwise set the fault type transfer frequency to zero; construct a state transition matrix based on the fault type transfer frequency. Obtain the global features and current operating condition parameters of the sample to be verified, and determine the target operating condition layer to which it belongs based on the current operating condition parameters; combine the global features and the current operating condition parameters to form a feature vector, and input it into the state transition matrix corresponding to the target operating condition layer to obtain the predicted probability distribution. Obtain the initial fault type probability distribution of the sample to be verified, divide the probability of each fault type in the initial fault type probability distribution by the probability of the corresponding fault type in the predicted probability distribution to obtain the probability ratio; obtain the credibility score based on the sum of the products of the logarithm of the probability ratio and the initial fault type probability.

[0040] Exemplarily, first, it is necessary to conduct an in - depth analysis of the historical fault data to determine the relationship between the operating condition parameters and the fault types. Operating condition parameters refer to the descriptive data of the system state under different operating environments, such as temperature, pressure, rotational speed, etc. At each moment, the system has a set of specific operating condition parameter values, and there is a certain correlation between these values and the fault types.

[0041] To effectively analyze these operating condition parameters, first, it is necessary to divide the numerical range of the operating condition parameters. Divide the numerical range of the operating condition parameters into multiple sub - ranges, and the operating condition parameter values within each sub - range correspond to different fault types. Specifically, first, through the statistics of the historical fault data, determine the minimum and maximum values that each operating condition parameter can take in the historical samples, and based on these values, perform interval division in an equal - interval or other way. Within each interval, count the occurrence frequency of the corresponding fault types. For example, if a certain operating condition parameter value is in the first interval, count the frequency of the fault types occurring in this interval.

[0042] Calculate the conditional entropy corresponding to each operating condition parameter value. The conditional entropy is used to measure the distribution uncertainty of fault types under given operating condition parameters. The calculation of the conditional entropy is based on the probability distribution of fault types within each interval, indicating how uncertain the fault types are within a given operating condition parameter interval. If the conditional entropy of a certain operating condition parameter is low, it means that this parameter has a strong predictive ability for fault types, that is, the fault types are relatively clear; if the conditional entropy is high, it means that this operating condition parameter has a weak predictive ability for fault types. To further screen out important operating condition parameters, first calculate the mean and standard deviation of the conditional entropy, and these two values can reflect the distribution of the entire data set. Then, determine the adaptive threshold based on the method of subtracting the standard deviation from the mean. The adaptive threshold is dynamically adjusted by subtracting one standard deviation from the mean, and it can adjust the tolerance for the conditional entropy according to different operating condition data. If the conditional entropy of a certain operating condition parameter is lower than this threshold, it can be considered that this operating condition parameter has good discrimination ability for the prediction of fault types and is worthy of further attention. If the conditional entropy is higher than the adaptive threshold, the prediction effect of this operating condition parameter is poor and may need to be excluded.

[0043] Construct a hierarchical parameter set based on the screened important operating condition parameters. The hierarchical parameter set is a space containing multiple dimensions, each dimension represents an operating condition parameter, and the data of each dimension is affected by the conditional entropy. By screening out the operating condition parameters with lower conditional entropy, a space containing important operating condition parameters is obtained.

[0044] Conduct further analysis on historical fault samples. In the hierarchical parameter set, first calculate the local density and minimum distance of each sample. The local density reflects the density of samples around a certain position. The local density can be achieved by calculating the number of samples within a certain range around each sample point. The minimum distance refers to the closest distance between a sample point and its surrounding sample points. By calculating the local density and minimum distance, the density peak points in the sample space can be found, and these peak points represent the clustering centers. On this basis, based on the minimum distance method, other samples are assigned to the closest clustering center, thus dividing the samples into different operating condition layers.

[0045] In each working condition layer, it is necessary to calculate the total number of samples of the fault types and the number of transfer samples between the fault types. The number of transfer samples refers to the number of times that samples transfer from one fault type to another in a certain working condition layer. The transfer frequency of the fault types is represented by the ratio of the number of transfer samples to the total number of samples. If the total number of samples in a certain working condition layer exceeds the preset threshold, the transfer frequency is calculated according to the proportion of the transfer samples. If the total number of samples is small, the transfer frequency is set to zero. This transfer frequency reflects the conversion law between different fault types under specific working conditions. Based on these transfer frequencies, a state transition matrix can be constructed, and the state transition matrix can show the transition probability of the fault types under different working conditions.

[0046] When the system needs to perform fault prediction on new samples to be verified, it is first necessary to extract the global features of the samples to be verified and the current working condition parameters. The global features may include the historical fault records of the samples, system configurations, etc., while the working condition parameters describe the current operating state of the system. Based on the current working condition parameters, the target working condition layer to which the sample belongs can be determined. Then, the global features and the current working condition parameters are combined into a feature vector, and the feature vector contains all the key information of the sample. Next, the feature vector is input into the state transition matrix corresponding to the target working condition layer, and the system will output the probability distribution of the sample belonging to different fault types.

[0047] The system needs to compare the initial fault type probability distribution of the samples to be verified with the predicted probability distribution of the target working condition layer. By taking the logarithm of the ratio of the initial probability to the predicted probability and multiplying it by the initial fault type probability, a credibility score is obtained. This credibility score represents the reliability of the current prediction. If the credibility score is high, it indicates that the fault type prediction result of the system for this sample is relatively reliable; if the credibility score is low, it is necessary to further check the data or adjust the model.

[0048] The existing technologies usually rely on fixed thresholds or traditional machine learning methods in fault prediction, and do not conduct sufficient hierarchical analysis on the relationship between working condition parameters and fault types, resulting in low accuracy and credibility of the prediction results. In addition, some methods lack in-depth mining of historical fault data and fail to effectively utilize the state transition law to model the fault evolution trend, so it is difficult to dynamically adapt to the changes in fault patterns under different working condition conditions.

[0049] This application conducts hierarchical analysis on historical fault samples based on conditional entropy, screens out operating condition parameters that have a greater impact on fault types, and constructs a hierarchical parameter set, thereby achieving more refined fault prediction. Further, based on the local density and minimum distance method, the density peak points are determined, and the samples are divided into different operating condition layers to make full use of the spatial distribution characteristics of the data and improve the rationality of operating condition modeling. On this basis, a state transition matrix is constructed by calculating the fault type transition frequency, enabling the prediction model to learn the fault evolution laws under different operating conditions. For the samples to be verified, this application combines the global features with the operating condition parameters, calculates the prediction probability in combination with the state transition matrix, and then calculates the credibility score through mutual information to quantify the reliability of the prediction results.

[0050] Compared with the prior art, the improvement of this application lies in introducing conditional entropy to screen operating condition parameters, conducting operating condition layering based on density peaks, and making predictions in combination with the state transition matrix, overcoming the problems of insufficient adaptability of traditional methods to complex operating conditions and difficult measurement of prediction credibility. Through these optimization measures, this application can improve the accuracy of fault prediction under different operating condition conditions and provide a credibility assessment of the prediction results, thereby enhancing the intelligent diagnosis ability and reliability of the system.

[0051] Figure 2 It is a simulation diagram of the effect of the real-time monitoring and optimization method for the source-network-load-storage equipment of the virtual power plant platform in the embodiment of the present invention. As Figure 2 shown, this figure shows the conditional entropy distribution of different operating condition parameters and the evaluation results of feature importance. The horizontal axis in the figure represents six key operating condition parameters: temperature, pressure, rotational speed, vibration, current, and voltage; the vertical axis represents the conditional entropy value, ranging from 0 to 0.8. The comparison results between this technical solution (circular markers) and the traditional entropy value method (square markers) clearly show the advantages of this solution.

[0052] In terms of the rotational speed parameter, this technical solution obtains the lowest conditional entropy value of 0.64, indicating that this parameter has the strongest prediction ability for fault types. In contrast, the conditional entropy corresponding to the traditional entropy value method is 0.44, and the prediction ability is significantly weaker. The pressure parameter ranks second, with the conditional entropy of this technical solution being 0.56 and that of the traditional method being 0.38. The conditional entropy values of the temperature parameter are 0.44 and 0.30 respectively, those of the vibration parameter are 0.50 and 0.34 respectively, those of the current parameter are 0.38 and 0.24 respectively, and those of the voltage parameter are 0.30 and 0.18 respectively.

[0053] Overall, the proposed technical solution has obtained higher conditional entropy values for all operating condition parameters, with an average increase of approximately 47%. This indicates that it makes more efficient use of the operating condition parameters and can better capture the correlation between the parameters and the fault types. Particularly for the two key parameters of rotational speed and pressure, the performance improvement is most significant, reaching 45.5% and 47.4% respectively, fully demonstrating the superiority of this solution in feature importance assessment.

[0054] In an optional embodiment, the feature data is decoupled and reconstructed in the time and space dimensions to obtain an equipment correlation matrix. Based on the equipment correlation matrix and the credible fault types, an equipment topology graph is constructed. The equipment topology graph is dynamically partitioned and local features are extracted. The fault diffusion probability is calculated through feature transfer between regions. Bidirectional tracking is performed on the equipment topology graph according to the fault diffusion probability to generate a fault propagation path. The equipment on the fault propagation path is ranked hierarchically, and the fault impact prediction result and the equipment priority list are output, including: The feature data is decomposed in the time dimension, and the feature data is decomposed into hourly subsequences, daily subsequences, and weekly subsequences. The periodic patterns and state trends in the hourly subsequences, daily subsequences, and weekly subsequences are extracted to construct a time dimension feature vector; The physical connection distance, electrical circuit connection, and equipment functional relationship in the feature data are extracted, the correlation coefficient between equipment is calculated, and a space dimension feature vector is constructed; The time dimension feature vector and the space dimension feature vector are subjected to tensor fusion operation to obtain an equipment correlation matrix; Based on the equipment correlation matrix and the credible fault types, an equipment topology graph including equipment nodes and connection edges is constructed. The equipment topology graph is dynamically partitioned into multiple equipment partitions by using a spectral clustering algorithm with adaptive similarity. The equipment density distribution within each equipment partition is calculated to obtain the eigenvalue within the region, and the connection edge weight between the equipment partitions is calculated to obtain the eigenvalue between regions. A transfer function including a time delay factor and an attenuation coefficient is constructed, and the eigenvalue within the region and the eigenvalue between regions are input into the transfer function to calculate the fault diffusion probability between equipment partitions; In the equipment topology graph, the position of the faulty equipment is marked as the starting node, and the influence propagation path is calculated through forward propagation based on the fault diffusion probability; on the influence propagation path, the equipment nodes with a fault diffusion probability exceeding a preset probability threshold are marked as the positions of affected equipment, the transfer importance of each node in the equipment topology graph is calculated, and the position of the source equipment is obtained through backward tracking; The influence propagation path and the position of the source equipment are combined and optimized to obtain a fault propagation path. The risk levels of the equipment on the fault propagation path are calculated and ranked hierarchically, and the fault impact prediction result and the equipment priority list are output.

[0055] Exemplarily, first, the feature data is decomposed in the time dimension. The feature data usually includes the operating status, performance indicators, fault records, etc. of various devices, and these data are presented in the form of a time series. Through the time dimension decomposition, the data is decomposed into hourly, daily, and weekly subsequences. Each subsequence represents the data characteristics at different time granularities. During the decomposition process, the hourly subsequence can capture the dynamic changes of the device in a short period of time, the daily subsequence reveals the stability and fluctuations of the device in a longer period, and the weekly subsequence helps to identify the trend changes of the device in a longer time range. By analyzing these subsequences, the periodic patterns and state trends are extracted. For example, the device may exhibit periodic faults at certain time points, or there are obvious trend changes during certain periods, and this information can help to identify the key characteristics of the device operation.

[0056] Next, the physical connection distance, electrical circuit connection, and device functional relationship in the feature data are extracted from the spatial dimension perspective. The physical connection distance refers to the distance between devices in physical space. For example, whether the devices are located in the same computer room or adjacent positions. The electrical circuit connection refers to the connection situation between devices through circuits, reflecting the mutual dependence of devices in aspects such as power supply or data transmission. The device functional relationship refers to whether there is a dependence or cooperation relationship between devices in terms of function. For example, the failure of one device may cause the shutdown of another device. By extracting these features, the correlation coefficient between devices can be calculated, indicating the degree of association between them. The higher the correlation coefficient, the stronger the association between devices, and the greater the probability of fault spread between them. According to the calculated correlation coefficient, a spatial dimension feature vector is constructed, and each feature vector reflects the mutual relationship of devices in space.

[0057] The time dimension feature vector and the spatial dimension feature vector are combined through tensor fusion operations to obtain a device association matrix. Tensor fusion operation is a way to combine multi-dimensional data, which synthesizes the information of the time dimension and the spatial dimension, thereby constructing a complete device association matrix. This matrix can accurately describe the mutual relationship of devices under different time and space conditions, reflecting the potential risk of fault propagation.

[0058] Based on the device association matrix and the trusted fault types, further construct a device topology graph that includes device nodes and connection edges. The device nodes represent each device, while the connection edges represent the relationships between devices. By analyzing the connectivity between devices, use the spectral clustering algorithm with adaptive similarity to dynamically partition the device topology graph. The spectral clustering algorithm can automatically divide devices into different partitions according to the similarity between devices, and the devices in each partition have strong correlations with each other. After obtaining the device partitions, calculate the device density distribution within each partition, which represents the distribution of devices within each device partition, and obtain the regional eigenvalue of this partition. At the same time, calculate the connection edge weights between device partitions, and these weights represent the strength of the relationships between partitions, that is, the possibility of fault propagation between different regions. Based on these intra-region and inter-region eigenvalues, a transfer function can be constructed. The transfer function can synthesize the delay factor and the attenuation coefficient to simulate the propagation of device faults between device partitions.

[0059] By inputting the intra-region and inter-region eigenvalues into the transfer function, calculate the fault diffusion probability between device partitions. The fault diffusion probability reflects the possibility that when a fault occurs in a certain device partition, this fault spreads to other regions. This process helps to identify the key regions of fault propagation and predict the devices that may be affected.

[0060] In the device topology graph, mark the location of the faulty device as the starting node. The starting node represents the device where the fault occurs, and from this device, the fault will spread to the surrounding devices. Based on the fault diffusion probability between devices, perform forward propagation calculation to obtain the influence propagation path. On the propagation path, the fault diffusion probability of each device node will be calculated to determine which devices will be affected by the fault. When the fault diffusion probability of a device on the propagation path exceeds the preset probability threshold, mark its location as the affected device location. This threshold is set according to the device importance and historical fault data to screen out those devices that may cause serious faults.

[0061] By calculating the transfer importance of each node in the device topology graph, trace back in the reverse direction to obtain the location of the source device. The transfer importance refers to the degree of the role of the device in the process of fault propagation, and it reflects the influence of the device on fault diffusion. By tracing back in the reverse direction, the source device of the fault can be found, and the key devices that cause fault propagation can be identified.

[0062] Finally, the influencing propagation path and the location of the source device are combined and optimized to obtain a complete fault propagation path. The fault propagation path shows the propagation process of the fault from the source device to other devices. By grading and sorting the risk levels of each device on this path, a priority list of devices can be obtained, which helps the operation and maintenance personnel to focus on high-risk devices first, so as to take countermeasures in advance and reduce the impact of the fault on the overall system.

[0063] In this embodiment, by fusing tensors to combine temporal and spatial features, the constructed device association matrix can comprehensively reflect the mutual relationship between devices, and combined with the fault type, predict its possible fault diffusion path. The construction of the device topology map and the dynamic partitioning algorithm further optimize the partition management of the devices, making the identification of the fault propagation path more accurate and avoiding the limitations of the single-factor influence in the traditional method. Based on the spectral clustering algorithm of adaptive similarity, it can dynamically adjust the device partitions, calculate the density distribution and connection weights of each partition, and accurately identify the key areas of fault diffusion. During the fault propagation process, the transmission importance of the devices can guide the tracing of the fault source, so that the fault response is more timely and potential losses are reduced. This solution not only improves the accuracy of device fault prediction, but also optimizes the device maintenance strategy, helps the operation and maintenance personnel to accurately identify high-risk devices before the fault occurs, realizes priority processing, thus improving the reliability and stability of the system and reducing the impact of the fault on the overall system.

[0064] In an alternative embodiment, Construct a transfer function including a time delay factor and an attenuation coefficient, input the eigenvalue within the region and the eigenvalue between regions into the transfer function, and the calculation of the fault diffusion probability between device partitions includes: Obtain the physical propagation delay and response delay between devices, classify the delay classification results according to the transfer type, perform segmented mapping on the delay classification results to obtain delay mapping values, and construct a time delay factor based on the delay mapping values; Collect the physical distance, electrical impedance and functional correlation degree between devices, calculate the spatial attenuation value according to the physical distance, calculate the signal attenuation value according to the electrical impedance, calculate the coupling attenuation value according to the functional correlation degree, perform hierarchical combination on the spatial attenuation value, signal attenuation value and coupling attenuation value to obtain an attenuation combination value, and construct an attenuation coefficient based on the attenuation combination value; construct a non-linear transfer function according to the time delay factor and the attenuation coefficient; Calculate the spatial distribution density value of devices within the calculation region and the distance distribution value between devices, calculate the regional density feature based on the spatial distribution density value and the distance distribution value, and perform feature fusion on the regional density feature and the device operation parameters to obtain the feature value within the region; calculate the physical connection quantity value and the energy transfer power value of the devices at the regional boundary, calculate the connection transmission feature based on the physical connection quantity value and the energy transfer power value, and perform feature fusion on the connection transmission feature and the device operation state to obtain the feature value between regions; establish device connection constraints and energy transfer constraints based on the feature value within the region and the feature value between regions; Classify historical fault samples according to fault types and propagation paths to obtain training samples, input the training samples into a non-linear transfer function, and perform iterative training using device connection constraints and energy transfer constraints to obtain the correction values of the time delay factor and the attenuation coefficient; input the feature value within the region and the feature value between regions into the non-linear transfer function, dynamically correct the propagation calculation result according to the correction values, and output the fault diffusion probability between device partitions.

[0065] Exemplarily, first, obtain the physical propagation delay and the response delay between devices. The physical propagation delay refers to the time for a signal to propagate in a transmission medium (such as a cable, optical fiber, etc.), which is usually related to the propagation distance and the physical properties of the transmission medium. The response delay is the time required for a device to respond after receiving a signal, including the processing delay of the device and the signal transmission time. These two delay values can be calculated by real-time monitoring of the communication delay between devices and the response speed of the devices. When classifying these delay data, it is first necessary to classify them according to different types of propagation paths and transmission media, such as distinguishing between wired and wireless signal propagation, or dividing by network topology. Perform segmented mapping on these classification results, map the delay values into different delay intervals, and thus obtain the delay mapping values. The delay mapping values represent the degree of signal propagation delay between devices, and then a time delay factor can be constructed through these mapping values, which can reflect the influence of the signal propagation delay between devices.

[0066] Then, collect the physical distance, electrical impedance, and functional correlation degree between devices. The physical distance refers to the straight-line distance between devices, and this value can be obtained through a sensor network or a positioning system between devices. The electrical impedance refers to the impedance value of the electrical lines between devices, which is usually determined by factors such as the material and length of the cable, and the signal frequency, and can be measured by a network analysis instrument. The functional correlation degree measures the degree of mutual dependence of devices in terms of function, and is usually evaluated according to the cooperation situation of devices during operation. For example, if two devices jointly control a key load in a power system, then the functional correlation degree between them is relatively high. When calculating the attenuation value based on these parameters, first calculate the spatial attenuation value through the physical distance. Usually, the spatial attenuation value is inversely proportional to the distance between devices; the electrical impedance affects the attenuation degree of the signal, so it is necessary to calculate the signal attenuation value according to the resistance value; the functional correlation degree reflects the coupling effect when devices work together, and calculate the coupling attenuation value. Then, weight and combine these three attenuation values to form an attenuation combined value, and further construct an attenuation coefficient to reflect the comprehensive attenuation effect during signal propagation between devices.

[0067] Next, construct a non-linear transfer function based on the time delay factor and the attenuation coefficient. The transfer function can describe the propagation law of signals or fault information between devices. The time delay factor and the attenuation coefficient affect the propagation speed and intensity of signals. Therefore, in the transfer function, these two are used as inputs for non-linear modeling to reflect the propagation characteristics between actual devices. Machine learning algorithms can be used for model training, enabling the transfer function to adaptively adjust according to historical data and optimize the propagation results.

[0068] When calculating the spatial distribution density value of devices in the area and the distance distribution value between devices, it is necessary to consider the distribution of devices in the area. The spatial distribution density value represents the number density of devices in a specific area, and is usually calculated through the layout data of devices in space; the distance distribution value between devices represents the distance distribution situation between devices in the area, which needs to be calculated based on the positioning information of devices. Calculate the area density feature based on these two values, and this feature can reflect the fault propagation potential of devices in the area. If the device density in the area is relatively high, it may lead to the rapid propagation of fault information between devices. Integrate the area density feature with the operating parameters of devices (such as load, temperature, operating status, etc.) to obtain the feature value in the area, and this value can comprehensively reflect the fault propagation risk of devices in the area.

[0069] When calculating the physical connection quantity value and energy transfer power value of the boundary devices of the calculation region, it is first necessary to know the physical connection number between the device and other external devices. The physical connection number can be counted through the interface data or communication protocol of the device. The energy transfer power value represents the energy transmission efficiency between devices and is usually related to the operating state of the device and the energy transmission efficiency. These connection and transmission characteristics are combined with the operating state of the device (such as workload, fault type, etc.) through feature fusion to obtain the inter-region feature value, which can reflect the energy transmission ability of the device between regions and its importance in fault propagation.

[0070] Based on the feature values within the region and between regions, connection constraints and energy transfer constraints of the devices can be established. The connection constraints reflect the strength of the physical connections between devices, and the energy transfer constraints represent the resistance of energy transmission between devices. These constraint conditions are used to limit the signal propagation paths between devices, thereby more accurately simulating fault propagation.

[0071] By classifying historical fault samples, the fault types and propagation paths can be labeled to generate training samples. These training samples are input into the non-linear transfer function, and through iterative training, the time delay factor and attenuation coefficient are adjusted so that the transfer function can more accurately reflect the actual fault propagation process. The connection constraints and energy transfer constraints of the devices are used to optimize the propagation results during the training process, making the propagation process more in line with the actual situation.

[0072] Finally, based on the corrected time delay factor and attenuation coefficient, the intra-region feature value and inter-region feature value are input into the non-linear transfer function for dynamic correction, and the fault diffusion probability between device partitions is output. These probability values can help the operation and maintenance personnel accurately predict the expansion range of faults and provide a basis for maintenance decisions.

[0073] Existing technologies usually rely on simple linear models or static rules in fault diffusion prediction and are difficult to accurately characterize the fault propagation characteristics in complex device environments. For example, some methods only evaluate the signal attenuation between devices based on a single physical parameter, such as distance or electrical impedance, while ignoring the influence of functional correlation on fault propagation. In addition, traditional methods lack a fine-grained modeling of the time delay factor between devices, resulting in an inaccurate description of the time characteristics of the signal propagation process, thus affecting the prediction accuracy.

[0074] This application improves the calculation method of the fault diffusion probability by constructing a non - linear transfer function that includes a time - delay factor and an attenuation coefficient. By introducing the time - delay factor, through the classification and segmented mapping of physical propagation delay and response delay, the time characteristics of signal transmission between devices are accurately characterized. Considering physical distance, electrical impedance, and functional correlation comprehensively, an attenuation coefficient is constructed to comprehensively evaluate the attenuation of signal propagation between devices. By calculating the regional density characteristics and connection transmission characteristics, the integration of intra - regional and inter - regional characteristics is realized, enabling the fault propagation prediction to not only consider the operating state of devices but also combine their spatial distribution characteristics, improving the comprehensiveness and accuracy of the prediction. In addition, through iterative optimization of the time - delay factor and attenuation coefficient with training samples, the model can be adaptively adjusted to improve the robustness and applicability of the prediction.

[0075] The improvement starting point of this application is to enhance the accuracy and adaptability of fault propagation prediction and overcome the problem of insufficient adaptability of traditional methods to complex device environments. After the improvement, the model can more accurately reflect the dynamic process of fault propagation, provide a more accurate prediction basis for device operation and maintenance, thereby optimizing the maintenance strategy, improving the reliability of the system, reducing the maintenance cost, and ensuring the long - term stable operation of the device.

[0076] Figure 3 It is an analysis diagram of the fault propagation time - delay and attenuation characteristics in the embodiments of the present invention. As Figure 3 shown, this diagram shows the change characteristics of the propagation intensity of fault signals at different physical distances. The horizontal axis represents the physical distance between devices (0 - 50 meters), and the vertical axis represents the propagation intensity of the fault signal (0.0 - 0.8). Three different propagation models are compared in the figure: the technical solution of this application (circular markers), the traditional exponential attenuation model (square markers), and the linear attenuation model (triangle markers).

[0077] It can be seen from the data that within the short - distance range (0 - 10 meters), all three models show a relatively high propagation intensity, with an initial value of 0.75 for all. However, as the distance increases, the attenuation characteristics of different models show significant differences. The technical solution of this application maintains a propagation intensity of 0.68 at 10 meters, while the exponential model and the linear model drop to 0.55 and 0.50 respectively. In the medium - distance range (20 - 30 meters), the attenuation of the technical solution of this application is more gentle, still maintaining a propagation intensity of 0.52 at 20 meters, while the other two models drop to 0.35 and 0.30 respectively.

[0078] Especially in the long-distance range (40 - 50 meters), this technical solution exhibits better propagation characteristics. At 50 meters, it can still maintain a propagation intensity of 0.12, while the exponential model and the linear model drop to 0.08 and 0.05 respectively. This indicates that by introducing the dynamic adjustment of the time delay factor and the attenuation coefficient, this technical solution significantly improves the propagation characteristics of long-distance fault signals, increases the effective distance of fault detection, and makes the fault warning of long-distance devices more reliable. Overall, the non-linear transfer function of this technical solution shows better propagation characteristics in various distance ranges, which is of great significance for the fault warning of large equipment groups.

[0079] In an alternative embodiment, a hierarchical reinforcement learning algorithm is used to optimize the collaborative scheduling of devices in a virtual power plant. Among them, the upper-layer agent generates a device combination plan based on the fault impact prediction result and the device priority list, and the lower-layer agent generates device control instructions based on the device combination plan and the fault risk score. The realization of the collaborative scheduling optimization of virtual power plant devices includes: Obtain the device fault impact prediction value, the device operation status value, and the device priority score, construct the upper-layer state space, and construct the upper-layer action space with the device combination plan; construct the upper-layer reward function based on the weighted combination of the system reliability index, the operation efficiency index, and the scheduling cost; Construct the lower-layer state space with the device combination plan, the device fault risk score, and the real-time operation parameters, and construct the lower-layer action space with the device control instructions; construct the lower-layer reward function based on the dynamic weight combination of the system stability index, the response time index, and the fault risk score; Calculate the device collaborative influence degree based on the fault impact prediction result and the priority list, construct the weighted state feature with the device collaborative influence degree as the weight coefficient; obtain the historical scheduling data to calculate the scheduling association strength between devices, adaptively adjust the weighted state feature based on the scheduling association strength, input the adjusted feature into the upper-layer state space, calculate the upper-layer state action value according to the upper-layer reward function, and generate the device combination plan; Perform deep feature fusion on the device combination plan and the fault risk score to obtain the device association feature, input the device association feature into the lower-layer state space, and calculate the lower-layer state action value based on the lower-layer reward function to generate the device control instruction; Execute the device control instruction to obtain the control result, calculate the reward value according to the control result, determine the sampling weight of the experience replay pool based on the historical reward distribution of different scheduling scenarios, and sample the training samples from the historical data; perform online update on the collaborative influence degree and the scheduling association strength based on the training samples and the reward value; The updated collaborative influence degree and scheduling correlation strength are respectively input into the upper-layer intelligent agent and the lower-layer intelligent agent to generate an optimized device combination plan and an optimized device control instruction, so as to realize the collaborative scheduling optimization of virtual power plant devices.

[0080] Exemplarily, first, the fault impact prediction value, real-time operation status information, and device priority score of each device are obtained through the monitoring system and the data acquisition module. The fault impact prediction value can be obtained through a deep learning prediction model based on the device's historical fault records and maintenance logs, and its value describes the impact degree of the device on the overall system when a fault may occur in the future; the operation status information includes multiple indicators such as the current load, temperature, and vibration of the device; the device priority score is determined based on the key degree and historical reliability of the device in the virtual power plant. For example, if the fault prediction of a certain key device shows an impact of about five percent, the status indicators are all in good condition, and the priority score is rated as the highest, the system will assign a higher weight to this device in subsequent scheduling.

[0081] Next, the upper-layer state space is constructed using the above-collected data, and each device is presented in this space as a description vector containing fault prediction, operation status, and priority score. Historical scheduling data and expert experience are used to construct the action space of the device combination plan, and its content is various combinations of switching and operation modes between different devices. For example, in a certain scheduling scenario, the device combination plan may be that device A and device C are running simultaneously while device B is in a standby state. At the same time, by analyzing the device fault prediction results and the priority list, the collaborative influence degree between devices is calculated, that is, the degree of mutual influence between devices when a fault occurs is described, and this description is presented in the form of text and data cases. For example, the collaborative influence between device A and device B is rated as medium, indicating that there is a certain dependence between the two in scheduling.

[0082] Furthermore, based on the overall reliability, operation efficiency, and scheduling cost of the system, a certain rule (such as an empirical rule or a weight strategy formulated by experts) is used to construct the upper-layer reward function. When the device combination plan can reduce the scheduling cost while ensuring the reliability and efficient operation of the system, this plan will receive a higher reward. In the construction process of the reward function, it does not rely on mathematical formulas, but through a comprehensive description of the contributions of each indicator. For example, if the device combination plan can increase the system operation efficiency by nearly ten percent and significantly reduce the scheduling cost at the same time, its reward value will increase accordingly.

[0083] In terms of lower-layer scheduling, the device combination scheme generated by the upper layer, the failure risk scores of each device, and the real-time operating parameters are comprehensively combined to form the lower-layer state space. The lower-layer action space is defined as a set of device control instructions, which cover operations such as startup, shutdown, load adjustment, and fault self-check. The lower-layer reward function depends on the dynamic performance of system stability, response time, and device failure risk, and is determined through text descriptions and empirical data. For example, when the device control instructions can reduce the response time to a predetermined level and improve system stability at the same time, positive incentives are given.

[0084] In the state feature construction stage, the collaborative influence degree of the device is obtained through fault impact prediction and priority list calculation. This influence degree is combined with the original state features as a weight coefficient to generate weighted state features. For example, if the device A has a relatively high failure risk and a high priority, a larger weight is added to its state description. At the same time, the scheduling association strength between devices is extracted from historical scheduling data to describe the frequency and dependence of collaborative scheduling between devices. If device A and device B have often run collaboratively in the past, the information of these two devices is appropriately enhanced in the weighted state features. The weighted state features after adaptive adjustment are input into the upper-layer state space for the upper-layer agent to evaluate the state-action values of various device combination schemes, and finally an optimal or approximately optimal device combination scheme is generated. This scheme can be described as a combined configuration in which device A, device B, and device C are in the running, standby, or standby states respectively.

[0085] Subsequently, the device combination scheme output by the upper layer and the real-time failure risk score undergo deep feature fusion to generate device association features. This process comprehensively considers the operating state, failure risk, and collaborative influence of each device. The generated device association features reflect the relevance between devices in a detailed description manner, such as information that the risk score of device A is moderately high and device B has a high scheduling association with device A. These device association features are passed as input to the lower-layer state space. Based on the pre-set reward function, the lower-layer agent generates highly targeted and responsive device control instructions through the evaluation of state-action values. These instructions may include operations such as load adjustment of the device, switching of operating modes, and early warning detection.

[0086] After the device control instruction is executed, the system collects the control results in real time, such as the device response time, load change situation, and scheduling effect, as feedback data. The feedback data is used to calculate the actual reward value, and the sampling weights of each training sample in the experience replay pool are determined according to the historical reward distribution under different scheduling scenarios. For example, in a certain scheduling, if the device scheduling success rate reaches more than 90% and the response time is significantly shortened, the data sample in this scheduling scenario is given a higher sampling weight for preferential use in subsequent training. The training samples extracted from the experience replay pool, combined with the actual reward value, are used to online update the parameters of the collaborative influence degree and scheduling correlation strength between devices. The update process adopts a descriptive feedback mechanism, that is, the previously set weights are fine-tuned according to the newly sampled data to ensure that the description of the collaborative effect between devices and the scheduling relevance can timely reflect the current system operation state.

[0087] The updated collaborative influence degree and scheduling correlation strength are respectively fed back to the upper-layer intelligent agent and the lower-layer intelligent agent, prompting the upper layer to further optimize the device combination plan, and the lower layer to generate more accurate device control instructions. The entire scheduling process forms a closed-loop feedback system, which adjusts in real time and continuously improves the overall performance of the device scheduling plan, ensuring that the virtual power plant system can cope with various emergencies and maintain a stable and efficient operation state.

[0088] In this embodiment, the intelligent collaborative scheduling and adaptive optimization of virtual power plant devices are realized through hierarchical reinforcement learning, and a closed-loop scheduling system is constructed by making full use of fault prediction, device status, and priority information. The system can generate and adjust the device combination plan and control instructions in real time, improve the overall operation efficiency and system reliability, and at the same time effectively reduce the scheduling cost and fault risk. The online update of the collaborative influence degree and scheduling correlation strength between devices makes the scheduling decision more accurate, and has the ability to quickly respond to sudden faults, optimize load distribution, and dynamically warn. The overall effect is reflected in the continuous improvement of intelligent scheduling management, stable and efficient operation, and self-learning ability, thereby enhancing the security and economic benefits of the virtual power plant.

[0089] In an alternative embodiment, Execute the device control instruction to obtain the control result, calculate the reward value according to the control result, determine the sampling weight of the experience replay pool based on the historical reward distribution of different scheduling scenarios, and sample training samples from the historical data; the online update of the collaborative influence degree and scheduling correlation strength based on the training samples and the reward value includes: Execute the device control instruction to obtain the control result composed of power balance degree, response time, energy loss, and system stability; calculate the system reliability score based on the power balance degree and system stability, calculate the scheduling efficiency score based on the response time and energy loss, and use the weighted combination of the system reliability score and the scheduling efficiency score as the reward value; Classify and store state transition samples according to the scheduling scenarios of different device combination schemes to construct an experience replay pool. The state transition samples include device combination schemes, device control instructions, reward values, and post-control states. Determine the sampling weights based on the historical reward distribution of the scheduling scenarios, and select training samples from the experience replay pool based on the sampling weights. Extract the response timing features and power coupling features between devices based on the device combination schemes in the training samples, and construct a device collaboration influence matrix. Construct a collaboration influence degree loss function based on the temporal correlation between the device collaboration influence matrix and the reward value, and optimize the collaboration influence degree loss function to update the collaboration influence degree online. Extract the scheduling dependency features and control coupling features between devices based on the device control instructions in the training samples, and construct a device scheduling association matrix. Construct a scheduling association strength loss function based on the dynamic correlation between the device scheduling association matrix and the reward value, and optimize the scheduling association strength loss function to update the scheduling association strength online.

[0090] Exemplarily, evaluate the effects of the device control instructions executed in the virtual power plant to obtain four key indicators: The power balance degree reflects the matching degree between the output power of the source-network-load-storage devices and the load demand; the response time characterizes the time interval from when the device receives the control instruction to reaching the target state; the energy loss measures the loss level during the device regulation process; the system stability depicts the fluctuations of key parameters such as voltage and frequency. Calculate the system reliability score based on the power balance degree and the system stability. When the power balance degree is greater than 95% and the fluctuation range of the key parameters is less than 3% of the rated value, the system reliability score is high; calculate the scheduling efficiency score based on the response time and the energy loss. When the average response time is less than 10 seconds and the total energy loss rate is lower than 5%, the scheduling efficiency score is high. Combine the two types of scores with weights to obtain the comprehensive reward value.

[0091] Establish an experience replay pool storage mechanism for different scenarios. For different scheduling scenarios such as peak-valley regulation, accident recovery, and economic dispatch, construct independent experience replay pools respectively. Each state transition sample contains four elements: The device combination scheme records the device combination participating in the scheduling; the device control instruction contains the target state and control parameters of each device; the reward value reflects the quality of the control effect; the post-control state describes the final operating state of the system. Statistically analyze the reward distribution characteristics of each scenario based on historical scheduling data, calculate the scenario sampling weights, and the higher the reward value, the greater the sampling probability of the scenario. Select training samples from the experience replay pools of different scenarios according to the sampling weights.

[0092] Analyze the response characteristics of each device in the equipment combination plan, and extract the timing characteristics such as the start-stop sequence of the equipment and the power adjustment rate; analyze the power mutual feedback relationship between the devices, and extract the power coupling characteristics such as the equipment capacity matching degree and transmission loss. Based on the extracted characteristics, construct a device collaborative influence matrix, and the matrix elements represent the degree of collaborative influence between the devices. By analyzing the correlation between the collaborative influence matrix and the historical reward sequence, construct a collaborative influence degree loss function, which reflects the deviation between the current collaborative influence degree and the influence degree corresponding to the optimal reward. Through optimization algorithms such as gradient descent, the collaborative influence degree is updated in real time.

[0093] Analyze the execution process of the device control instructions, and extract the scheduling dependence characteristics such as the instruction response sequence and control authority; analyze the control association between the devices, and extract the control coupling characteristics such as the control parameter coupling degree and state constraint. Based on the extracted characteristics, construct a device scheduling association matrix, and the matrix elements represent the strength of the scheduling association between the devices. By analyzing the correlation between the scheduling association matrix and the dynamic reward change, construct a scheduling association strength loss function, which depicts the difference between the current association strength and the strength corresponding to the optimal scheduling effect. Through the online learning method, the scheduling association strength is updated dynamically.

[0094] Existing technologies usually adopt rule-driven or simple static models in equipment scheduling optimization, and it is difficult to accurately adapt to the dynamic changes of complex scheduling scenarios. For example, some methods only adjust the device control strategy based on fixed rules and fail to make full use of historical data for learning, resulting in difficult optimization of scheduling efficiency and system reliability. In addition, existing methods lack accurate modeling of the collaborative influence degree and scheduling association strength of devices, ignoring the dynamic interaction relationship between devices, which affects the scheduling optimization effect.

[0095] This application improves the accuracy and adaptability of device scheduling decisions by constructing an optimization method based on experience replay and online update. An experience replay pool is constructed for different scheduling scenarios, and the sampling weights are dynamically adjusted based on the historical reward distribution, making the training samples more representative, thereby enhancing the generalization ability of the model. An online update mechanism for the collaborative influence degree and scheduling association strength is proposed. By extracting the response timing characteristics, power coupling characteristics, scheduling dependence characteristics, and control coupling characteristics between devices, a collaborative influence matrix and a scheduling association matrix are constructed respectively, and a loss function is constructed based on the correlation of the reward values to achieve adaptive optimization of key parameters. Through methods such as gradient optimization, the model can dynamically adjust the collaborative influence degree and scheduling association strength between devices according to the actual scheduling effect, improving the intelligent level of the scheduling strategy. Considering indicators such as power balance degree, system stability, response time, and energy loss, the reward calculation method is optimized, so that the scheduling strategy can take into account both system reliability and scheduling efficiency.

[0096] The improvement starting point of this application lies in enhancing the accuracy and intelligent level of equipment scheduling optimization, and overcoming the problem of insufficient adaptability of traditional methods to dynamic scheduling scenarios. After the improvement, the model can make full use of historical data to dynamically optimize scheduling decisions, making the equipment scheduling strategy more flexible and adaptable, thereby enhancing the stability of the system, reducing energy consumption, and improving the overall scheduling efficiency to achieve better resource allocation.

[0097] The core innovation of this invention lies in the organic combination of equipment fault diagnosis, propagation prediction, and scheduling optimization. Through in-depth mining of equipment operation data, a complete fault warning and processing framework is established. The method not only considers the operation status of a single device but also fully considers the mutual influence between devices, making fault warning and processing more accurate and effective. At the same time, a hierarchical reinforcement learning algorithm is used for equipment scheduling, enabling the system to have stronger environmental adaptability and optimization effects.

[0098] By improving the accuracy of fault warning, the operation and maintenance cost is significantly reduced; by optimizing fault propagation prediction, the efficiency of fault handling is improved; by improving equipment scheduling, the overall operation efficiency of the system is enhanced. These innovations not only solve the key problems in the existing technology but also provide new technical ideas for the intelligent operation of virtual power plants.

[0099] Figure 4 It is a schematic structural diagram of the real-time monitoring and optimization system for the source-network-load-storage equipment of the virtual power plant platform in the embodiment of the present invention. As Figure 4 shown, the system includes: The first unit is used to collect equipment operation data, perform noise reduction and feature compression to obtain feature data, identify the equipment fault type based on the feature data, match the equipment fault type with historical fault data to generate a fault risk score, and verify the credibility of the equipment fault type based on the fault risk score to obtain a credible fault type; The second unit is used to decouple and reconstruct the feature data in the time and space dimensions to obtain an equipment association matrix, construct an equipment topology map based on the equipment association matrix and the credible fault type, dynamically partition the equipment topology map and extract local features, calculate the fault diffusion probability through feature transfer between regions; perform two-way tracking on the equipment topology map according to the fault diffusion probability to generate a fault propagation path; rank the equipment on the fault propagation path hierarchically and output the fault impact prediction result and the equipment priority list; The third unit is used to adopt a hierarchical reinforcement learning algorithm to perform collaborative scheduling optimization on the equipment in the virtual power plant. Among them, the upper-layer intelligent agent generates an equipment combination plan according to the fault impact prediction result and the equipment priority list, and the lower-layer intelligent agent generates an equipment control instruction based on the equipment combination plan and the fault risk score to achieve the collaborative scheduling optimization of the virtual power plant equipment.

[0100] In a third aspect of the embodiments of the present invention, a kind of electronic device is provided, including: a processor; a memory for storing instructions executable by the processor; wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0101] In a fourth aspect of the embodiments of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0102] The present invention can be a method, a device, a system and / or a computer program product. The computer program product can include a computer-readable storage medium, on which computer-readable program instructions for executing various aspects of the present invention are uploaded.

[0103] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for real-time monitoring and optimization of source, grid, load and storage equipment on a virtual power plant platform, characterized in that: include: Collect equipment operation data and perform noise reduction and feature compression to obtain feature data. Identify equipment fault types based on feature data, match equipment fault types with historical fault data for similarity, generate fault risk scores, and verify the credibility of equipment fault types based on the fault risk scores to obtain credible fault types. The feature data is decoupled and reconstructed in time and space dimensions to obtain the device association matrix. Based on the device association matrix and the trusted fault type, the device topology map is constructed. The device topology map is dynamically partitioned and local features are extracted. The fault diffusion probability is calculated by transferring features between regions. Perform bidirectional tracing on the device topology map based on the fault diffusion probability to generate the fault propagation path; Classify and sort the devices on the fault propagation path, and output the fault impact prediction results and device priority list; A hierarchical reinforcement learning algorithm is used to coordinate and optimize the equipment scheduling in the virtual power plant. The upper-level agent generates an equipment combination plan based on the fault impact prediction results and the equipment priority list, and the lower-level agent generates equipment control instructions based on the equipment combination plan and fault risk score, thereby realizing the coordinated scheduling optimization of the virtual power plant equipment.

2. The method according to claim 1, characterized in that Collect equipment operation data and perform noise reduction and feature compression to obtain feature data. Identify equipment fault types based on feature data, match equipment fault types with historical fault data for similarity, generate fault risk scores, and verify the credibility of equipment fault types based on the fault risk scores. The credible fault types obtained include: Performing wavelet transform on the equipment operation data to obtain multi-scale coefficients, calculating the ratio of adjacent scale coefficients to construct a feature matrix; performing singular value decomposition on the feature matrix to obtain a signal subspace and a noise subspace, calculating the orthogonal vector product of the signal subspace and the noise subspace to obtain an adaptive threshold, projecting the equipment operation data to the signal subspace and reconstructing based on the adaptive threshold to obtain denoised data; Construct an autoencoder model, in which the encoding part maps the denoised data into intermediate features, and obtains latent variables through a nonlinear activation function. Construct a discriminator network, input the latent variables and preset Gaussian distribution samples into the discriminator network to obtain the discrimination probability, and calculate the cross entropy based on the discrimination probability and the true label to obtain the adversarial loss. Perform gradient descent optimization on the latent variables according to the adversarial loss to obtain the fault features. The fault feature is subjected to sliding convolution through a one-dimensional convolution kernel to obtain a time series feature sequence, and a gated recurrent unit is used to calculate the time series feature sequence time by time step to obtain a state vector; the conditional entropy of the state vector at different time steps is calculated to obtain a time series weight, and the time series weight is weighted and summed with the time series feature sequence to obtain a global feature, and the global feature is input into a softmax classifier to obtain an initial fault type; Based on conditional entropy, historical fault samples are stratified according to operating parameters, and the fault type transfer frequency is calculated in each layer of samples to obtain a state transfer matrix. The global features are concatenated with the current operating parameters and input into the state transfer matrix to obtain the predicted probability. The mutual information between the initial fault type and the predicted probability is calculated to obtain a credibility score. When the credibility score exceeds a preset threshold, it is confirmed as a credible fault type.

3. The method according to claim 2, characterized in that Based on conditional entropy, historical fault samples are stratified according to operating condition parameters, the fault type transition frequency is calculated in each layer of samples to obtain a state transfer matrix, the global features are spliced ​​with the current operating condition parameters and input into the state transfer matrix to obtain the predicted probability, and the mutual information between the initial fault type and the predicted probability is calculated to obtain the credibility score, including: Divide the numerical interval of the operating condition parameter into multiple sub-intervals, count the occurrence frequency of the fault type in each sub-interval, and calculate the conditional entropy of the operating condition parameter to the fault type; calculate the mean and standard deviation of the conditional entropy, and obtain an adaptive threshold based on the mean minus the standard deviation; construct a hierarchical parameter set for the operating condition parameters whose conditional entropy is less than the adaptive threshold; Calculate the local density and minimum distance of the historical fault samples in the space corresponding to the hierarchical parameter set, determine the density peak point as the cluster center based on the local density, and assign the sample points to the cluster center based on the minimum distance to obtain the working condition layer; In each operating condition layer, the total number of samples of the fault type and the number of transfer samples between the fault types are calculated; when the total number of samples is greater than the preset sample number threshold, the ratio of the transfer sample number to the total number of samples is calculated to obtain the fault type transfer frequency, otherwise the fault type transfer frequency is set to zero; a state transfer matrix is ​​constructed based on the fault type transfer frequency; Obtain the global features and current operating parameters of the sample to be verified, and determine the target operating layer to which it belongs based on the current operating parameters; combine the global features with the current operating parameters to form a feature vector, and input the state transfer matrix corresponding to the target operating layer to obtain the predicted probability distribution; Obtain the initial fault type probability distribution of the sample to be verified, divide the probability of each fault type in the initial fault type probability distribution by the probability of the corresponding fault type in the predicted probability distribution to obtain a probability ratio; and obtain a credibility score based on the sum of the logarithm of the probability ratio and the product of the initial fault type probability.

4. The method according to claim 1, characterized in that: The feature data is decoupled and reconstructed in time and space dimensions to obtain the device association matrix. Based on the device association matrix and the trusted fault type, the device topology map is constructed. The device topology map is dynamically partitioned and local features are extracted. The fault diffusion probability is calculated by transferring features between regions. Perform bidirectional tracing on the device topology map based on the fault diffusion probability to generate the fault propagation path; The devices on the fault propagation path are ranked and sorted, and the fault impact prediction results and device priority list are output, including: Decomposing the feature data in a time dimension, decomposing the feature data into hourly subsequences, daily subsequences, and weekly subsequences, extracting periodic patterns and state trends in the hourly subsequences, daily subsequences, and weekly subsequences, and constructing a time dimension feature vector; Extracting physical connection distance, electrical circuit connection and equipment function relationship in the feature data, calculating correlation coefficients between equipment, and constructing a spatial dimension feature vector; Performing a tensor fusion operation on the time dimension feature vector and the space dimension feature vector to obtain a device association matrix; Based on the device association matrix and the trusted fault type, a device topology graph including device nodes and connection edges is constructed, a spectral clustering algorithm with adaptive similarity is used to dynamically partition the device topology graph to obtain multiple device partitions, the device density distribution within each device partition is calculated to obtain the regional eigenvalue, and the connection edge weights between the device partitions are calculated to obtain the inter-regional eigenvalue; a transfer function including a delay factor and an attenuation coefficient is constructed, the intra-regional eigenvalue and the inter-regional eigenvalue are input into the transfer function, and the fault diffusion probability between the device partitions is calculated; The location of the faulty device is marked as the starting node in the device topology map, and a forward propagation calculation is performed based on the fault diffusion probability to obtain an impact propagation path; the device nodes whose fault diffusion probability exceeds a preset probability threshold are marked on the impact propagation path as the affected device locations, and the transfer importance of each node in the device topology map is calculated, and the source device location is obtained by reverse tracing; The impact propagation path and the source device location are combined and optimized to obtain the fault propagation path, the risk level of each device on the fault propagation path is calculated and ranked, and the fault impact prediction result and the device priority list are output.

5. The method according to claim 4, characterized in that Construct a transfer function including a delay factor and an attenuation coefficient, input the characteristic value within the region and the characteristic value between regions into the transfer function, and calculate the fault diffusion probability between device partitions, including: Obtaining physical propagation delay and response delay between devices, and obtaining delay classification results according to transmission type classification, performing segment mapping on the delay classification results to obtain delay mapping values, and constructing a delay factor based on the delay mapping values; The physical distance, electrical impedance and functional correlation between the acquisition devices are calculated, the spatial attenuation value is calculated according to the physical distance, the signal attenuation value is calculated according to the electrical impedance, the coupling attenuation value is calculated according to the functional correlation, the spatial attenuation value, the signal attenuation value and the coupling attenuation value are hierarchically combined to obtain an attenuation combination value, and an attenuation coefficient is constructed based on the attenuation combination value; a nonlinear transfer function is constructed according to the delay factor and the attenuation coefficient; Calculate the spatial distribution density value of the equipment in the area and the distance distribution value between the equipment, calculate the regional density characteristics according to the spatial distribution density value and the distance distribution value, and fuse the regional density characteristics with the equipment operation parameters to obtain the characteristic value within the area; calculate the physical connection quantity value and the energy transfer power value of the equipment at the boundary of the area, calculate the connection transmission characteristics according to the physical connection quantity value and the energy transfer power value, and fuse the connection transmission characteristics with the equipment operation status to obtain the inter-area characteristic value; establish the equipment connection constraint and the energy transfer constraint according to the characteristic value within the area and the characteristic value between the areas; Historical fault samples are classified according to fault type and propagation path to obtain training samples, the training samples are input into a nonlinear transfer function, and iterative training is performed using device connection constraints and energy transfer constraints to obtain correction values ​​of delay factors and attenuation coefficients; the intra-regional eigenvalues ​​and inter-regional eigenvalues ​​are input into the nonlinear transfer function, the propagation calculation results are dynamically corrected according to the correction values, and the fault diffusion probability between device partitions is output.

6. The method according to claim 1, characterized in that A hierarchical reinforcement learning algorithm is used to coordinate and optimize the equipment in the virtual power plant. The upper-level agent generates an equipment combination plan based on the fault impact prediction results and the equipment priority list, and the lower-level agent generates equipment control instructions based on the equipment combination plan and fault risk score. The coordinated scheduling optimization of virtual power plant equipment includes: Obtain the predicted value of equipment failure impact, equipment operation status value and equipment priority score, construct the upper state space, and construct the upper action space with the equipment combination plan; construct the upper reward function based on the weighted combination of system reliability index, operation efficiency index and scheduling cost; The equipment combination scheme, equipment failure risk score and real-time operation parameters are used to construct a lower state space, and the equipment control instructions are used to construct a lower action space; a lower reward function is constructed based on a dynamic weight combination of a system stability index, a response time index and the failure risk score; The equipment collaborative impact is calculated based on the fault impact prediction results and the priority list, and the equipment collaborative impact is used as the weight coefficient to construct the weighted state feature; historical scheduling data is obtained to calculate the scheduling association strength between devices, and the weighted state feature is adaptively adjusted based on the scheduling association strength. The adjusted feature is input into the upper state space, and the upper state action value is calculated according to the upper reward function to generate the equipment combination plan; The equipment combination scheme and the fault risk score are deeply fused to obtain equipment-related features, and the equipment-related features are input into the lower-level state space. The lower-level state action value is calculated based on the lower-level reward function to generate equipment control instructions. Execute device control instructions to obtain control results, calculate reward values ​​based on control results, determine sampling weights of the experience replay pool based on historical reward distributions of different scheduling scenarios, and obtain training samples from historical data; perform online updates on collaborative influence and scheduling association strength based on training samples and reward values; The updated collaborative influence and scheduling association strength are input into the upper-level intelligent agent and the lower-level intelligent agent respectively to generate optimized equipment combination plans and optimized equipment control instructions, thereby realizing the collaborative scheduling optimization of virtual power plant equipment.

7. The method according to claim 6, characterized in that Execute device control instructions to obtain control results, calculate reward values ​​based on the control results, determine the sampling weights of the experience replay pool based on the historical reward distribution of different scheduling scenarios, and sample training samples from historical data; Online updates of the collaborative impact and scheduling association strength based on training samples and reward values ​​include: Execute the device control instruction to obtain the control result consisting of power balance, response time, energy loss and system stability; calculate the system reliability score based on the power balance and system stability, calculate the dispatch efficiency score based on the response time and energy loss, and use the weighted combination of the system reliability score and the dispatch efficiency score as the reward value; According to the scheduling scenarios of different equipment combination schemes, state transition samples are classified and stored to build an experience replay pool, wherein the state transition samples include equipment combination schemes, equipment control instructions, reward values, and post-control states; sampling weights are determined according to the historical reward distribution of the scheduling scenarios, and training samples are selected from the experience replay pool based on the sampling weights; Extract response timing characteristics and power coupling characteristics between devices based on the device combination scheme in the training sample, and construct a device synergy influence matrix; construct a synergy influence loss function according to the temporal correlation between the device synergy influence matrix and the reward value, and optimize the synergy influence loss function to update the synergy influence online; Based on the device control instructions in the training samples, the scheduling dependency characteristics and control coupling characteristics between devices are extracted to construct a device scheduling association matrix; according to the dynamic correlation between the device scheduling association matrix and the reward value, a scheduling association strength loss function is constructed, and the scheduling association strength is updated online by optimizing the scheduling association strength loss function.

8. A real-time monitoring and optimization system for source, grid, load and storage equipment of a virtual power plant platform, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to collect equipment operation data and perform noise reduction and feature compression to obtain feature data, identify equipment fault types based on the feature data, perform similarity matching between the equipment fault types and historical fault data, generate fault risk scores, and perform credibility verification on the equipment fault types based on the fault risk scores to obtain credible fault types; The second unit is used to decouple and reconstruct the feature data in time and space dimensions to obtain the device association matrix, build the device topology map based on the device association matrix and the trusted fault type, dynamically partition the device topology map and extract local features, and calculate the fault diffusion probability through inter-region feature transfer; Perform bidirectional tracing on the device topology map based on the fault diffusion probability to generate the fault propagation path; Classify and sort the devices on the fault propagation path, and output the fault impact prediction results and device priority list; The third unit is used to use a hierarchical reinforcement learning algorithm to coordinate and optimize the equipment in the virtual power plant. The upper-level intelligent agent generates an equipment combination plan based on the fault impact prediction results and the equipment priority list, and the lower-level intelligent agent generates equipment control instructions based on the equipment combination plan and fault risk score to achieve coordinated scheduling optimization of virtual power plant equipment.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images