Smart grid relay protection reliability evaluation and promotion control method and system

Abstract

The application provides a kind of smart grid relay protection reliability evaluation and promotion control method and system, it is related to smart grid relay protection technical field, including the operation data of relay protection device is collected, after pre-processing and multi-domain feature extraction, input depth hybrid neural network to obtain comprehensive feature vector. Utilize multi-stage fault diagnosis model, respectively identify fault mode, collaborative anomaly and reliability level at device level, interval level and station level, and predict fault propagation path and influence based on causal reasoning network. Finally, construct hierarchical collaborative control strategy, realize the coordinated optimization of protection setting value through multi-objective optimization, distributed game optimization and hierarchical rolling optimization, improve the reliability of relay protection device, thereby reduce the probability of fault propagation, ensure the safe and stable operation of power grid.

Description

Technical Field

[0001] This invention relates to smart grid relay protection technology, and more particularly to a method and system for assessing and improving the reliability of smart grid relay protection. Background Technology

[0002] Relay protection devices are an important component of smart grids, ensuring the safe operation of power systems. With the expansion of power grids and the increasing complexity of equipment, traditional relay protection technologies face challenges such as inaccurate fault mode identification, difficulty in optimizing protection settings, and difficulty in detecting anomalies in inter-device coordination, thus affecting the reliability and efficiency of relay protection devices.

[0003] Currently, fault diagnosis of relay protection devices largely relies on static settings and empirical rules, which cannot adapt to the dynamic changes in the operating state of the power system. Furthermore, traditional methods fail to effectively integrate multiple data sources and complex inter-device relationships, resulting in insufficient accuracy in fault prediction and protection optimization.

[0004] Therefore, there is an urgent need for a new approach that combines big data analytics, artificial intelligence, and optimized control technology to improve the fault diagnosis capabilities and system reliability of relay protection devices, thereby enhancing the safety and stability of smart grids. Summary of the Invention

[0005] This invention provides a method and system for assessing and improving the reliability of smart grid relay protection, which can solve the problems in the prior art.

[0006] A first aspect of the present invention,

[0007] A method for reliability assessment and improvement control of smart grid relay protection is provided, including:

[0008] Operational data of relay protection devices are collected and processed through an adaptive data preprocessing model, dividing the data into steady-state, transient, and fault segments. The data in the steady-state, transient, and fault segments are denoised using an improved empirical mode decomposition method. Multi-domain features are extracted by combining wavelet packet transform and Hilbert-Huang transform to generate an enhanced feature matrix. The enhanced feature matrix is ​​then input into a deep hybrid neural network to extract spatial features. Temporal correlations are captured through recurrent branching. An adaptive fusion layer is used to combine the spatial features and temporal correlations to obtain a comprehensive feature vector.

[0009] inputting the comprehensive feature vector into a pre-constructed multi-stage fault diagnosis model, wherein a fault mode and a performance degradation trend of a single protection device are identified based on a deep belief network at a device level, a coordinated anomaly of adjacent protection devices is analyzed based on an improved conditional random field at an interval level, a reliability level of a protection device group is divided based on a hierarchical clustering method at a station level, the fault mode, the performance degradation trend, the coordinated anomaly and the reliability level are input into a causal reasoning network, a dynamic causal graph is established, a fault propagation path and a propagation probability are predicted based on the dynamic causal graph, a fault influence degree is evaluated in combination with a fuzzy cognitive map, and fault early warning information is output;

[0010] a layered cooperative control strategy is constructed according to the fault early warning information, a protection setting value optimization problem is converted into a multi-objective optimization model at a local control level, a hybrid optimization algorithm is used to solve the multi-objective optimization model to obtain a candidate setting value scheme, a distributed game optimization network is established based on a Nash equilibrium criterion at a global coordination level, and a potential game method is used to coordinate and optimize the candidate setting value scheme, and the coordinated and optimized setting value scheme is input into a layered rolling optimization controller, a time domain optimization problem is established based on model predictive control, closed-loop control is realized through online state estimation and feedback correction, and the reliability of the relay protection device is improved.

[0011] Operation data of the relay protection device are collected, and the data are processed through an adaptive data preprocessing model to divide the data into a steady state segment, a transient state segment and a fault segment, the data of the steady state segment, the transient state segment and the fault segment are denoised through an improved empirical mode decomposition method, and multi-domain features are extracted through wavelet packet transform and Hilbert-Huang transform to generate an enhanced feature matrix including:

[0012] Operation data of the relay protection device are collected, and the operation data include primary equipment current signals, primary equipment voltage signals, protection device action signals, communication link state signals, circuit breaker mechanical characteristic signals and secondary circuit insulation impedance signals, and the operation data are normalized to obtain a standardized data sequence;

[0013] A variational mode decomposition model is established for the standardized data sequence, the variational mode decomposition model establishes a variational optimization objective function including a center frequency term and a bandwidth constraint term, an alternating direction multiplier method is used to iteratively optimize the objective function to obtain an intrinsic mode function group, a cross-correlation coefficient of adjacent intrinsic mode functions is calculated based on energy distribution characteristics of the intrinsic mode function group, an adaptive segmentation threshold is set according to the cross-correlation coefficient, and the standardized data sequence is divided into a steady state segment, a transient state segment and a fault segment;

[0014] The data of the steady-state section, transient section and fault section are respectively subjected to improved empirical mode decomposition denoising processing, and an adaptive soft threshold function based on data distribution is used to process intrinsic mode components, parameters of the adaptive soft threshold function being adjusted according to layers and energy density of the intrinsic mode components, and a local mean envelope is used to replace spline interpolation to eliminate end effect, so as to obtain denoised data;

[0015] The denoised data is subjected to wavelet packet transform, entropy values of wavelet packet decomposition coefficients are calculated, and a node coefficient with the maximum entropy value is selected as a frequency domain feature; the denoised data is subjected to Hilbert-Huang transform, and instantaneous amplitude and instantaneous phase of an analytical signal obtained by the Hilbert-Huang transform are calculated, and a time-frequency feature is obtained according to the instantaneous phase;

[0016] The frequency domain feature vector and the time-frequency feature vector are adaptively weighted and fused, and high-order statistical features based on kernel density estimation, including mean, variance, skewness, kurtosis, quantile features and entropy features, are introduced, and an enhanced feature matrix is formed after dimension reduction by using principal component analysis.

[0017] A variational mode decomposition model is established for the standardized data sequence, the variational mode decomposition model obtains an intrinsic mode function group by constructing a variational optimization objective function containing a center frequency term and a bandwidth constraint term, and iteratively optimizing the objective function by using an alternating direction multiplier method, and correlation coefficients of adjacent intrinsic mode functions are calculated based on energy distribution characteristics of the intrinsic mode function group, and an adaptive segmentation threshold is set according to the correlation coefficients, and the standardized data sequence is divided into a steady-state section, a transient section and a fault section including:

[0018] The standardized data sequence is subjected to discrete sampling and frame processing, each frame of data is converted into an analytical signal by Hilbert transform, and a center frequency sequence is obtained by frequency modulation of the analytical signal, and a reconstruction error matrix and a spectrum constraint matrix are constructed based on the center frequency sequence;

[0019] The reconstruction error matrix and the spectrum constraint matrix are input into a Wiener filter, a frequency response of the Wiener filter is adaptively adjusted according to a bandwidth constraint parameter, and a mode component is output; a soft threshold shrinkage processing is applied to the mode component, and a soft threshold parameter is determined by cross-validation; an updated center frequency sequence is obtained by calculating a power spectrum barycenter based on the processed mode component; the center frequency sequence is optimized by gradient iteration, and an intrinsic mode function group is obtained;

[0020] The instantaneous energy spectrum of each intrinsic mode function in the intrinsic mode function group is calculated, an energy density distribution curve is obtained, and an energy distribution characteristic matrix is obtained by calculating energy aggregation degree and energy dispersion degree based on the energy density distribution curve;

[0021] extracting an energy envelope of adjacent intrinsic mode functions, calculating a normalized cross-correlation function of the energy envelope, calculating a peak value of the normalized cross-correlation function in a sliding time window, and generating a cross-correlation coefficient sequence;

[0022] calculating a mean value and a standard deviation of the cross-correlation coefficient sequence, setting the mean value as a reference threshold value, multiplying an energy concentration and an energy dispersion in the energy distribution feature matrix to obtain a threshold adjustment factor, and setting a product of the reference threshold value and the threshold adjustment factor as an adaptive segmentation threshold value;

[0023] comparing the cross-correlation coefficient sequence with the corresponding adaptive segmentation threshold value in the sliding time window, marking a candidate mutation point when the cross-correlation coefficients of three consecutive windows are all less than the adaptive segmentation threshold value, calculating a local energy density of the candidate mutation point, determining a final modal mutation point position based on a minimum value of the energy density, and dividing a standardized data sequence into a steady-state segment, a transient segment and a fault segment according to the modal mutation point.

[0024] inputting the comprehensive feature vector into a pre-constructed multi-stage fault diagnosis model, wherein a deep belief network is used to identify a fault mode and a performance degradation trend of a single protection device at a device level, an improved conditional random field is used to analyze a cooperative anomaly of adjacent protection devices at an interval level, and a hierarchical clustering method is used to divide a reliability level of a protection device group at a station level, including:

[0025] constructing a multi-stage fault diagnosis model, the multi-stage fault diagnosis model including a device-level diagnosis sub-model, an interval-level diagnosis sub-model and a station-level diagnosis sub-model;

[0026] the device-level diagnosis sub-model adopts a deep belief network, inputs the comprehensive feature vector into the deep belief network, pre-trains layer by layer through a restricted Boltzmann machine, optimizes network weights using a contrastive divergence algorithm and a stochastic gradient descent method, introduces label data for network fine-tuning, and obtains a fault mode probability vector and a performance degradation index of a single protection device;

[0027] the interval-level diagnosis sub-model adopts a conditional random field structure, takes the fault mode probability vector and the performance degradation index as node features, constructs a spatial potential function based on a one-time device topology relationship, constructs a time sequence potential function based on a running state correlation, inputs the fault mode probability vector, the performance degradation index, the spatial potential function and the time sequence potential function into the conditional random field structure, optimizes model parameters through maximum likelihood estimation, and outputs an adjacent device cooperative anomaly probability matrix;

[0028] The station-level diagnosis sub-model adopts a hierarchical clustering structure, fuses the fault mode probability vector, performance degradation index and collaborative anomaly probability matrix to construct a state representation matrix, calculates the inter-device state distance based on the state representation matrix, groups the devices using a hierarchical clustering algorithm with adaptive distance measurement, determines the optimal grouping number through the silhouette coefficient, and obtains the reliability level of the protection device.

[0029] The fault mode, performance degradation trend, collaborative anomaly and reliability level are input into a causal reasoning network to establish a dynamic causal graph, the fault propagation path and propagation probability are predicted based on the dynamic causal graph, the fault impact degree is evaluated in combination with a fuzzy cognitive map, and the fault early warning information includes:

[0030] The fault mode, performance degradation trend, collaborative anomaly and reliability level of the relay protection device are constructed into a node set of the causal reasoning network, the conditional probability table between nodes is calculated based on the conditional independence assumption, the recursive Bayesian estimation method is used to update the node state online, the conditional probability parameters are optimized according to the historical observation data, and a dynamic causal graph reflecting the dynamic correlation between nodes is established;

[0031] Based on the dynamic causal graph, the causal support of the node is calculated by forward propagation, the diagnostic support of the node is calculated by backward propagation, the causal support and diagnostic support are combined to update the node belief value through a normalization factor, the propagation path of the fault in the network is calculated according to the node belief value and the conditional probability, and the fault propagation probability is calculated based on the transmission probability between nodes and the intermediate node state;

[0032] The fault propagation path and fault propagation probability are respectively mapped into the concept node and weight matrix of the fuzzy cognitive map, a fuzzy relationship matrix containing the fault source, fault type, impact range and propagation path is constructed, and the target node state vector is obtained through fuzzy synthesis operation and iterative update of the hyperbolic tangent activation function;

[0033] The direct impact degree is calculated based on the weight matrix and target node state vector of the fuzzy cognitive map, the indirect impact degree is calculated in combination with the fault propagation probability and a preset attenuation factor, and the direct impact degree and indirect impact degree are combined to obtain the cumulative impact degree of the fault by using an exponential weighting method;

[0034] An early warning decision matrix is constructed based on the fault propagation probability and the cumulative impact degree of the fault, the early warning level threshold is determined by using fuzzy analytic hierarchy process, the early warning information is divided into four levels, and the early warning information containing the fault source device, propagation path, risk level and disposal measures is output in a hierarchical and progressive manner according to the early warning level.

[0035] Based on the dynamic causal graph, the causal support of nodes is calculated using forward propagation, and the diagnostic support of nodes is calculated using backward propagation. The causal support and diagnostic support are combined using a normalization factor to update the node belief value. The propagation path of the fault in the network is calculated based on the node belief value and conditional probability. The fault propagation probability is calculated based on the inter-node transmission probability and the intermediate node state, including:

[0036] Obtain the state information of nodes in a dynamic causal graph and construct the conditional probability matrix between nodes;

[0037] In the dynamic causal graph, the set of parent nodes of the current node is identified, the information transmission delay time from each node in the set of parent nodes to the current node is calculated, the temporal weight is determined based on the information transmission delay time, the temporal weight is combined with the physical coupling strength to generate the transmission coefficient, and the causal support of the current node is calculated through forward propagation based on the transmission coefficient and the initial belief value of each node in the set of parent nodes.

[0038] Extract the set of child nodes of the current node from the dynamic causal graph, calculate the state deviation of each node in the set of child nodes, construct a feedback influence function based on the state deviation, multiply the feedback influence function by the distance decay coefficient between nodes to obtain the diagnostic weight, and calculate the diagnostic support of the current node through backpropagation based on the diagnostic weight and the belief value of each node in the set of child nodes.

[0039] The adaptive normalization factor is calculated based on the system's operating status. The causal support and diagnostic support are then weighted and combined using the adaptive normalization factor to update the belief value of the current node.

[0040] Based on the updated current node belief value and the conditional probability matrix, the fault propagation probability between adjacent nodes is calculated, a directed propagation network is constructed based on the fault propagation probability, and the fault propagation path is extracted from the directed propagation network.

[0041] The real-time status values ​​of each node on the critical path of fault propagation are obtained, the real-time status values ​​are converted into a state function, and the cumulative propagation probability on the fault propagation path is calculated based on the state function and the fault propagation probability to obtain the final fault propagation probability.

[0042] Based on the fault warning information, a hierarchical collaborative control strategy is constructed. At the local control layer, the protection setting optimization problem is transformed into a multi-objective optimization model. A hybrid optimization algorithm is used to solve the multi-objective optimization model to obtain candidate setting schemes. At the global coordination layer, a distributed game optimization network is established based on the Nash equilibrium criterion. A potential game method is used to coordinate and optimize the candidate setting schemes, including:

[0043] Receive system fault warning information, and extract the fault type, fault location, fault current, fault duration and fault phase angle from the fault warning information as feature quantities;

[0044] The protection reliability index is calculated based on the fault type and fault current; the protection sensitivity index is calculated based on the fault location and fault current; and the protection selectivity index is calculated based on the fault duration and protection coordination time interval.

[0045] A multi-objective optimization model is constructed at the local control layer based on the protection reliability index, protection sensitivity index, and protection selectivity index.

[0046] The multi-objective optimization model is solved by setting an adaptive crossover operator based on population density to obtain a crossover solution. The crossover solution is mutated by setting an adaptive mutation operator based on fitness value to obtain a mutated solution. A reference point set is constructed based on the mutated solution. The Euclidean distance from the mutated solution to the reference point set is calculated. The solution with the smallest Euclidean distance is selected as a candidate fixed value scheme.

[0047] In the global coordination layer, the protection unit is set as a game participant, and the candidate value scheme is set as the strategy space of the game participant to construct a distributed game optimization network.

[0048] In the distributed game optimization network, the reliability gain of game participants when choosing strategies is calculated based on the protection sensitivity index, the coordination gain between game participants and adjacent protection units is calculated based on the protection coordination time interval, and the interference loss suffered by game participants is calculated based on the electrical distance between adjacent protection units.

[0049] The utility function of the game participants is constructed by taking the reliability gains, coordination gains, and interference losses as the comprehensive gains.

[0050] The electrical connection degree is obtained by calculating the impedance values ​​between protection units in the distributed game optimization network, and the coupling coefficient between game participants is determined based on the electrical connection degree; the cooperation relationship is obtained by calculating the action time difference between protection units, and the interaction utility between game participants is determined based on the cooperation relationship; the potential function is constructed by combining the utility function, coupling coefficient and interaction utility of game participants.

[0051] In the distributed game optimization network, an initial strategy combination for the game participants is generated. The utility function value and potential function value of the game participants under the initial strategy combination are calculated. The initial strategy combination is updated based on the utility function value to obtain an optimized response strategy. The new potential function value under the optimized response strategy is calculated. When the difference between the new potential function value and the potential function value of the previous iteration is less than a preset threshold, the optimized response strategy is output as a coordinated optimization fixed-value scheme that satisfies the Nash equilibrium criterion.

[0052] A second aspect of the present invention provides a smart grid relay protection reliability assessment and improvement control system, comprising:

[0053] The first unit is used to collect the operating data of the relay protection device and process the data through an adaptive data preprocessing model, dividing it into steady-state, transient, and fault segments. The data in the steady-state, transient, and fault segments are respectively processed by an improved empirical mode decomposition method for noise reduction. Multi-domain features are extracted by combining wavelet packet transform and Hilbert-Huang transform to generate an enhanced feature matrix. The enhanced feature matrix is ​​input into a deep hybrid neural network to extract spatial features. Temporal correlation is captured through recurrent branching. An adaptive fusion layer is used to combine the spatial features and temporal correlation to obtain a comprehensive feature vector.

[0054] The second unit is used to input the comprehensive feature vector into a pre-constructed multi-stage fault diagnosis model. At the device level, it identifies the fault modes and performance degradation trends of individual protection devices based on deep belief networks. At the interval level, it analyzes the cooperative anomalies of adjacent protection devices based on improved conditional random fields. At the station level, it divides the reliability levels of protection device groups based on hierarchical clustering methods. The fault modes, performance degradation trends, cooperative anomalies, and reliability levels are input into a causal inference network to establish a dynamic causal graph. Based on the dynamic causal graph, the fault propagation path and propagation probability are predicted. The degree of fault impact is evaluated by combining a fuzzy cognitive graph, and fault warning information is output.

[0055] The third unit is used to construct a hierarchical collaborative control strategy based on the fault warning information. At the local control layer, the protection setting optimization problem is transformed into a multi-objective optimization model, and a hybrid optimization algorithm is used to solve the multi-objective optimization model to obtain candidate setting schemes. At the global coordination layer, a distributed game optimization network is established based on the Nash equilibrium criterion, and the potential game method is used to coordinate and optimize the candidate setting schemes. The coordinated and optimized setting schemes are input into the hierarchical rolling optimization controller, which establishes a time-domain optimization problem based on model predictive control, and achieves closed-loop control through online state estimation and feedback correction to improve the reliability of the relay protection device.

[0056] Third aspect of the present invention

[0057] An electronic device is provided, comprising:

[0058] processor;

[0059] Memory used to store processor-executable instructions;

[0060] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0061] Fourth aspect of the embodiments of the present invention,

[0062] A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0063] In this embodiment, by combining multiple signal processing methods and deep learning models, the fault modes, performance degradation trends, and collaborative anomalies of adjacent protection devices can be more accurately identified. Furthermore, based on a causal inference network, the fault propagation path and impact level are predicted, thereby improving the accuracy of fault early warning. Optimizing protection settings using a multi-objective optimization model and a distributed game-theoretic optimization network enables coordinated optimization of settings at both local and global levels, improving the reliability and adaptability of the protection devices. Employing a hierarchical rolling optimization controller, based on model predictive control, achieves online state estimation and feedback correction. This allows for dynamic adjustment of protection settings according to the system's operating state, realizing closed-loop control and further improving the reliability and stability of the relay protection system. Attached Figure Description

[0064] Figure 1 This is a flowchart illustrating the smart grid relay protection reliability assessment and improvement control method according to an embodiment of the present invention.

[0065] Figure 2 This is a schematic diagram of the structure of the smart grid relay protection reliability assessment and improvement control system according to an embodiment of the present invention. Detailed Implementation

[0066] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0067] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0068] Figure 1 This is a flowchart illustrating the smart grid relay protection reliability assessment and improvement control method according to an embodiment of the present invention, as shown below. Figure 1 As shown, the method includes:

[0069] S101. Collect the operating data of the relay protection device, process the data through an adaptive data preprocessing model, divide it into steady-state, transient, and fault segments, and perform noise reduction processing on the data of the steady-state, transient, and fault segments respectively using an improved empirical mode decomposition method. Combine wavelet packet transform and Hilbert-Huang transform to extract multi-domain features and generate an enhanced feature matrix. Input the enhanced feature matrix into a deep hybrid neural network to extract spatial features, capture temporal correlations through recurrent branches, and use an adaptive fusion layer to combine the spatial features and temporal correlations to obtain a comprehensive feature vector.

[0070] S102. The comprehensive feature vector is input into a pre-constructed multi-stage fault diagnosis model, wherein, at the device level, a deep belief network is used to identify the fault modes and performance degradation trends of individual protection devices; at the interval level, an improved conditional random field is used to analyze the cooperative anomalies of adjacent protection devices; at the station level, a hierarchical clustering method is used to classify the reliability levels of protection device groups; the fault modes, performance degradation trends, cooperative anomalies, and reliability levels are input into a causal inference network to establish a dynamic causal graph; the fault propagation path and propagation probability are predicted based on the dynamic causal graph; and the degree of fault impact is evaluated by combining a fuzzy cognitive graph to output fault warning information.

[0071] S103. Based on the fault warning information, a hierarchical collaborative control strategy is constructed. At the local control layer, the protection setting optimization problem is transformed into a multi-objective optimization model. A hybrid optimization algorithm is used to solve the multi-objective optimization model to obtain candidate setting schemes. At the global coordination layer, a distributed game optimization network is established based on the Nash equilibrium criterion. The potential game method is used to coordinate and optimize the candidate setting schemes. The coordinated and optimized setting schemes are input into the hierarchical rolling optimization controller. The hierarchical rolling optimization controller establishes a time-domain optimization problem based on model predictive control. Closed-loop control is achieved through online state estimation and feedback correction to improve the reliability of the relay protection device.

[0072] In this context, cyclic branching refers to introducing a feedback mechanism into the network structure to capture temporal correlations in the data. This structure can effectively process data with time-series characteristics, such as the operational data of power equipment.

[0073] The optimized setting scheme will be input into the hierarchical rolling optimization controller. The hierarchical rolling optimization controller can make timely adjustments during system operation by dynamically estimating the power system state, thus ensuring the adaptability of the protection settings under different operating conditions.

[0074] In one optional implementation, operational data of the relay protection device is collected and processed using an adaptive data preprocessing model, dividing the data into steady-state, transient, and fault segments. An improved empirical mode decomposition method is applied to denoise the data in the steady-state, transient, and fault segments respectively. Multi-domain features are extracted using wavelet packet transform and Hilbert-Huang transform to generate an enhanced feature matrix, including:

[0075] The operation data of the relay protection device is collected. The operation data includes primary equipment current signal, primary equipment voltage signal, protection device action signal, communication link status signal, circuit breaker mechanical characteristic signal and secondary circuit insulation impedance signal. The operation data is normalized to obtain a standardized data sequence.

[0076] A variational mode decomposition model is established for the standardized data sequence. The variational mode decomposition model constructs a variational optimization objective function containing a center frequency term and a bandwidth constraint term. The objective function is iteratively optimized using the alternating direction multiplier method to obtain an intrinsic mode function set. Based on the energy distribution characteristics of the intrinsic mode function set, the cross-correlation coefficient of adjacent intrinsic mode functions is calculated. An adaptive segmentation threshold is set according to the cross-correlation coefficient to divide the standardized data sequence into a steady-state segment, a transient segment, and a fault segment.

[0077] The data in the steady-state segment, transient segment, and fault segment are respectively subjected to improved empirical mode decomposition denoising processing. An adaptive soft thresholding function based on data distribution is used to process the intrinsic mode components. The parameters of the adaptive soft thresholding function are adjusted according to the number of layers and energy density of the intrinsic mode components. Local mean envelope is used to replace spline interpolation to eliminate endpoint effects, resulting in denoised data.

[0078] The denoised data is subjected to wavelet packet transform, and the entropy of the wavelet packet decomposition coefficients is calculated. The node coefficient with the largest entropy is selected as the frequency domain feature. The denoised data is subjected to Hilbert-Huang transform, and the instantaneous amplitude and instantaneous phase of the resulting analytic signal are calculated. The instantaneous frequency is obtained based on the instantaneous phase to obtain the time-frequency feature.

[0079] The frequency domain feature vector and the time-frequency feature vector are adaptively weighted and fused, and high-order statistical features based on kernel density estimation are introduced, including mean, variance, skewness, kurtosis, quantile features and entropy features. After dimensionality reduction using principal component analysis, an enhanced feature matrix is ​​formed.

[0080] For example, the operation data of the relay protection device is first collected, including primary equipment current signals, primary equipment voltage signals, protection device operation signals, communication link status signals, circuit breaker mechanical characteristic signals, and secondary circuit insulation impedance signals. These signals are acquired by sensors and data acquisition devices and stored in digital form. For instance, current signals can be acquired using current transformers, voltage signals can be acquired using voltage transformers, protection device operation signals can be acquired using digital input modules, communication link status signals can be acquired using network monitoring modules, circuit breaker mechanical characteristic signals can be acquired using limit switches and position sensors, and secondary circuit insulation impedance signals can be acquired using insulation monitoring devices. The collected data includes timestamps for subsequent analysis.

[0081] The collected operational data is normalized. The purpose of normalization is to eliminate the influence of different signal dimensions and orders of magnitude, transforming the data into a uniform range. For example, the min-max normalization method can be used to scale the value of each signal to between 0 and 1.

[0082] A variational mode decomposition model is established for the normalized data sequence. This model constructs a variational optimization objective function containing a center frequency term and a bandwidth constraint term, and iteratively optimizes the objective function using the alternating direction multiplier method to obtain a set of intrinsic mode functions (EMFs). The center frequency term describes the center frequency of each EMF, and the bandwidth constraint term limits the bandwidth of each EMF. The alternating direction multiplier method is an iterative optimization algorithm that decomposes a complex optimization problem into multiple simpler subproblems and solves these subproblems alternately until the global optimum is obtained.

[0083] The cross-correlation coefficients of adjacent intrinsic mode functions (IMFs) are calculated based on the energy distribution characteristics of the IMF set. The cross-correlation coefficients measure the similarity between two IMFs. An adaptive segmentation threshold is set based on the cross-correlation coefficients to divide the standardized data sequence into steady-state, transient, and fault segments. For example, if the cross-correlation coefficients of two adjacent IMFs are greater than a preset threshold, then these two IMFs are considered to belong to the same segment. Assuming the calculated cross-correlation coefficient sequence is [0.9, 0.8, 0.2, 0.9, 0.7], and the preset threshold is 0.5, the data sequence can be divided into three segments: [0.9, 0.8], [0.2], and [0.9, 0.7].

[0084] Improved empirical mode decomposition (EMD) denoising was applied to the data in the steady-state, transient, and fault states. An adaptive soft thresholding function based on data distribution was used to process the intrinsic mode components (IMCs), with parameters adjusted according to the number of IMC layers and energy density. Local mean envelope was used to replace spline interpolation to eliminate endpoint effects, resulting in denoised data. For example, for an IMC, if its energy density was low, it was considered to mainly contain noise, and its amplitude was reduced; conversely, if its energy density was high, it was considered to mainly contain useful signals, and its amplitude was preserved or amplified.

[0085] The denoised data is subjected to wavelet packet transform, and the entropy values ​​of the wavelet packet decomposition coefficients are calculated. The node coefficient with the largest entropy value is selected as the frequency domain feature. Wavelet packet transform is a signal decomposition method that can decompose a signal into sub-signals of different frequency bands. Entropy is used to measure the complexity of a signal. For example, if the entropy value of a node coefficient is large, it is considered that the node contains a lot of information.

[0086] The denoised data is subjected to Hilbert-Huang transform to calculate the instantaneous amplitude and instantaneous phase of the resulting analytic signal. The instantaneous frequency is then calculated based on the instantaneous phase to obtain the time-frequency characteristics. The Hilbert-Huang transform is a time-frequency analysis method that decomposes a signal into a series of intrinsic mode functions (IMFs) and calculates the instantaneous frequency of each IMF.

[0087] The frequency domain feature vector and time-frequency feature vector are adaptively weighted and fused, and high-order statistical features based on kernel density estimation, including mean, variance, skewness, kurtosis, quantile features, and entropy features, are introduced. Principal component analysis (PCA) is then used to reduce the dimensionality, resulting in an enhanced feature matrix. The purpose of adaptive weighted fusion is to assign different weights based on the importance of different features. PCA is a dimensionality reduction method that projects high-dimensional data into a low-dimensional space while retaining key information.

[0088] In this embodiment, by performing multi-domain feature extraction and enhancement on the operational data, the operating status of the power system can be more comprehensively reflected, thereby improving the accuracy of fault identification. Adaptive data preprocessing and denoising effectively suppress the influence of noise and interference, improving the anti-interference capability of the relay protection device. Feature dimensionality reduction and the construction of enhanced feature matrices reduce computational load and improve the operating efficiency of the relay protection device.

[0089] In one optional implementation, a variational mode decomposition model is established for the standardized data sequence. The variational mode decomposition model constructs a variational optimization objective function containing a center frequency term and a bandwidth constraint term. The objective function is iteratively optimized using the alternating direction multiplier method to obtain an intrinsic mode function set. Based on the energy distribution characteristics of the intrinsic mode function set, the cross-correlation coefficients of adjacent intrinsic mode functions are calculated. An adaptive segmentation threshold is set according to the cross-correlation coefficients to divide the standardized data sequence into a steady-state segment, a transient segment, and a fault segment, including:

[0090] The standardized data sequence is discretely sampled and framed. Each frame of data is converted into an analytical signal through Hilbert transform. The analytical signal is frequency modulated to obtain a center frequency sequence. A reconstruction error matrix and a spectral constraint matrix are constructed based on the center frequency sequence.

[0091] The reconstruction error matrix and spectral constraint matrix are input into a Wiener filter, and the frequency response of the Wiener filter is adaptively adjusted according to the bandwidth constraint parameter to output modal components. Soft threshold shrinkage processing is applied to the modal components, and the soft threshold parameter is determined through cross-validation. The power spectrum centroid is calculated based on the processed modal components to obtain an updated center frequency sequence. The center frequency sequence is optimized through gradient iteration to obtain the intrinsic mode function set.

[0092] Calculate the instantaneous energy spectrum of each intrinsic mode function in the intrinsic mode function set to obtain the energy density distribution curve. Based on the energy density distribution curve, calculate the energy concentration and energy dispersion to obtain the energy distribution feature matrix.

[0093] Extract the energy envelopes of adjacent intrinsic mode functions, calculate the normalized cross-correlation function of the energy envelopes, calculate the peak value of the normalized cross-correlation function within a sliding time window, and generate a cross-correlation coefficient sequence;

[0094] Calculate the mean and standard deviation of the cross-correlation coefficient sequence, and set the mean as the baseline threshold; simultaneously, multiply the energy concentration and energy dispersion in the energy distribution feature matrix to obtain the threshold adjustment factor; set the product of the baseline threshold and the threshold adjustment factor as the adaptive segmented threshold;

[0095] Within the sliding time window, the cross-correlation coefficient sequence is compared with the corresponding adaptive segmentation threshold. When the cross-correlation coefficients of three consecutive windows are all less than the adaptive segmentation threshold, they are marked as candidate mutation points. The local energy density of the candidate mutation points is calculated, and the final modal mutation point location is determined based on the minimum energy density. The standardized data sequence is divided into steady-state segment, transient segment, and fault segment according to the modal mutation points.

[0096] For example, firstly, the standardized data sequence is discretized and framed. For instance, for a signal with a sampling frequency of 1 kHz, 100 sampling points can be used as one frame, with a frame shift of 50 sampling points. Then, each frame of data is converted into an analytic signal using a Hilbert transform. The analytic signal contains the original signal and its Hilbert transform, which can be used to extract the instantaneous amplitude and instantaneous frequency of the signal. The analytic signal is then frequency-modulated to obtain a center frequency sequence. The center frequency sequence reflects the change of the signal frequency over time. Based on the center frequency sequence, a reconstruction error matrix and a spectral constraint matrix are constructed. The reconstruction error matrix measures the difference between the reconstructed signal and the original signal, and the spectral constraint matrix is ​​used to limit the bandwidth of the modal components.

[0097] Next, the reconstructed error matrix and spectral constraint matrix are input into the Wiener filter. The frequency response of the Wiener filter is adaptively adjusted according to the bandwidth constraint parameter, outputting modal components. The bandwidth constraint parameter controls the frequency range of the modal components. For example, the bandwidth constraint parameter can be set to 0.1, indicating that the bandwidth of the modal component does not exceed 0.1 times the sampling frequency. Soft thresholding is applied to the modal components, and the soft threshold parameter is determined through cross-validation. Soft thresholding can remove noise and interference from the modal components. For example, 5-fold cross-validation can be used to select the optimal soft threshold parameter. Based on the processed modal components, the power spectrum centroid is calculated to obtain an updated center frequency sequence. The power spectrum centroid reflects the main frequency components of the modal components. The center frequency sequence is optimized through gradient iteration to obtain the intrinsic mode function set. Gradient iteration can gradually adjust the center frequency sequence to approach the optimal solution.

[0098] Then, the instantaneous energy spectrum of each eigenmode function in the eigenmode function set is calculated to obtain the energy density distribution curve. The instantaneous energy spectrum reflects the distribution of signal energy at different frequencies. Based on the energy density distribution curve, the energy concentration and energy dispersion are calculated to obtain the energy distribution characteristic matrix. Energy concentration indicates the degree of concentration of energy near the main frequency components, while energy dispersion indicates the degree of dispersion of energy at different frequencies. For example, the variance of the energy density distribution curve can be calculated as an indicator of energy dispersion.

[0099] Extract the energy envelopes of adjacent intrinsic mode functions (EMFs) and calculate the normalized cross-correlation function of the energy envelopes. The energy envelopes reflect the change of signal energy over time. Calculate the peak value of the normalized cross-correlation function within a sliding time window to generate a cross-correlation coefficient sequence. The sliding time window can be set to 10 frames, or 1 second. The cross-correlation coefficient sequence reflects the correlation between adjacent EMFs.

[0100] Calculate the mean and standard deviation of the cross-correlation coefficient sequence, and set the mean as the baseline threshold. Simultaneously, multiply the energy concentration and energy dispersion in the energy distribution feature matrix to obtain the threshold adjustment factor. The product of the baseline threshold and the threshold adjustment factor is set as the adaptive segmented threshold. The adaptive segmented threshold can be dynamically adjusted according to the energy distribution characteristics of the signal.

[0101] Within a sliding time window, the cross-correlation coefficient sequence is compared with the corresponding adaptive segmentation threshold. A candidate mutation point is marked when the cross-correlation coefficients for three consecutive windows are all less than the adaptive segmentation threshold. The local energy density of the candidate mutation points is calculated, and the final modal mutation point location is determined based on the minimum energy density. The normalized data sequence is then divided into steady-state, transient, and fault segments according to the modal mutation points.

[0102] Suppose we are analyzing a vibration signal containing a fault. Before the fault occurs, the signal energy is mainly concentrated in the low-frequency range, and the cross-correlation coefficient is relatively high. After the fault occurs, the energy distribution of the signal changes, the high-frequency components increase, and the cross-correlation coefficient decreases. By comparing the cross-correlation coefficient with an adaptive segmentation threshold, we can accurately identify the time point of the fault and divide the signal into steady-state, transient, and fault segments.

[0103] In this embodiment, by introducing a center frequency term and a bandwidth constraint term, different modal components in the signal can be extracted more accurately, reducing mode aliasing. Adaptively adjusting the segmentation threshold based on energy distribution characteristics better adapts to the characteristics of different signals, improving the accuracy of state division. Iterative optimization using the alternating direction multiplier method effectively reduces computational complexity and improves the algorithm's operating efficiency.

[0104] In one optional implementation, the comprehensive feature vector is input into a pre-constructed multi-stage fault diagnosis model, wherein, at the device level, fault modes and performance degradation trends of individual protection devices are identified based on deep belief networks; at the interval level, cooperative anomalies of adjacent protection devices are analyzed based on improved conditional random fields; and at the station level, the reliability levels of protection device groups are divided based on hierarchical clustering methods, including:

[0105] A multi-stage fault diagnosis model is constructed, which includes a device-level diagnosis sub-model, a bay-level diagnosis sub-model, and a station-level diagnosis model.

[0106] The device-level diagnostic sub-model adopts a deep belief network. The comprehensive feature vector is input into the deep belief network and pre-trained layer by layer through a restricted Boltzmann machine. The network weights are optimized by contrastive divergence algorithm and stochastic gradient descent method. Label data is introduced to fine-tune the network and obtain the fault mode probability vector and performance degradation index of a single protection device.

[0107] The interval-level diagnostic sub-model adopts a conditional random field structure, uses the fault mode probability vector and performance degradation index as node features, constructs a spatial potential function based on the topological relationship of primary equipment, and constructs a temporal potential function based on the correlation of operating status. The fault mode probability vector, performance degradation index, spatial potential function and temporal potential function are input into the conditional random field structure, and the model parameters are optimized through maximum likelihood estimation to output the cooperative anomaly probability matrix of adjacent devices.

[0108] The station-level diagnostic sub-model adopts a hierarchical clustering structure, which integrates the fault mode probability vector, performance degradation index and collaborative anomaly probability matrix to construct a state representation matrix. Based on the state representation matrix, the state distance between devices is calculated, and the devices are grouped using a hierarchical clustering algorithm with adaptive distance metric. The optimal number of groups is determined by the contour coefficient to obtain the reliability level of the protection device.

[0109] For example, firstly, it is necessary to obtain the comprehensive feature vector of the protection device. This can be achieved in various ways, such as collecting physical quantities of the protection device, such as current, voltage, and temperature, and combining them with historical data such as the device's operating time and number of operations to extract features reflecting the device's status. For instance, the current value, voltage value, and ambient temperature of a certain protection device can be collected, and combined with the device's operating time and number of operations to form a six-dimensional feature vector: [10A, 220V, 25℃, 3600h, 100 times, 0 times]. The last digit, "0 times," indicates that the protection device did not malfunction within the statistical period.

[0110] Next, a multi-stage fault diagnosis model is constructed. This model includes three sub-models: device level, bay level, and station level.

[0111] The device-level diagnostic sub-model employs a deep belief network structure. The acquired comprehensive feature vector is input into a pre-trained deep belief network. This network is optimized through layer-by-layer pre-training and fine-tuning, ultimately outputting a fault mode probability vector and performance degradation index for a single protection device. For example, for the input feature vector [10A, 220V, 25℃, 3600h, 100 times, 0 times], the fault mode probability vector output by the device-level diagnostic sub-model might be [0.1, 0.05, 0.8, 0.05], representing the probabilities of four fault modes: normal, aging, non-operation, and maloperation, respectively. The performance degradation index can be 0.2, indicating a certain degree of performance degradation in the device.

[0112] The interval-level diagnostic sub-model adopts a conditional random field (CRF) structure. The fault mode probability vector and performance degradation index output by the device-level diagnostic sub-model are used as node features. These are combined with the topological relationships and operational state correlations of the primary equipment to construct spatial and temporal potential functions. For example, if two protection devices are geographically close, their spatial potential function values ​​are higher. If the operational states of two protection devices are highly correlated in time, their temporal potential function values ​​are higher. Inputting this information into the CRF model yields the cooperative anomaly probability matrix for adjacent devices. For example, for two adjacent protection devices, their cooperative anomaly probability matrix might be [[0.9, 0.1], [0.1, 0.9]], indicating a high probability that both devices will simultaneously experience anomalies.

[0113] The station-level diagnostic sub-model employs a hierarchical clustering structure. A state characterization matrix is ​​constructed by fusing the fault mode probability vector and performance degradation index output from the device-level diagnostic sub-model with the collaborative anomaly probability matrix output from the interval-level diagnostic sub-model. Based on this matrix, the state distance between devices is calculated, and a hierarchical clustering algorithm with an adaptive distance metric is used to group the devices. For example, the state characterization matrix can reflect the state similarity between different protection devices. The hierarchical clustering algorithm can group protection devices with similar states into the same group. Finally, the optimal number of groups is determined using the silhouette coefficient, yielding the reliability level of the protection devices. For example, protection devices can be divided into three reliability levels: high, medium, and low, corresponding to different levels of reliability.

[0114] In this embodiment, information at the device level, bay level, and station level is comprehensively considered, enabling a more complete analysis of the operating status of the protection device and thus improving the accuracy of fault diagnosis. By analyzing the performance degradation trend of the protection device, early warning of faults can be achieved, avoiding power outages caused by protection device failures. By classifying the protection device into reliability levels, equipment maintenance strategies can be optimized, maintenance efficiency can be improved, and maintenance costs can be reduced.

[0115] In one optional implementation, the failure mode, performance degradation trend, collaborative anomaly, and reliability level are input into a causal inference network to establish a dynamic causal graph. Based on the dynamic causal graph, the failure propagation path and propagation probability are predicted, and the degree of failure impact is assessed by combining a fuzzy cognitive graph. The output failure early warning information includes:

[0116] The fault modes, performance degradation trends, cooperative anomalies, and reliability levels of relay protection devices are used to construct a node set of a causal inference network. Based on the assumption of conditional independence, the conditional probability table between nodes is calculated. The recursive Bayesian estimation method is used to update the node state online. The conditional probability parameters are optimized based on historical observation data to establish a dynamic causal graph that reflects the dynamic relationship between nodes.

[0117] Based on the dynamic causal graph, the causal support of the nodes is calculated by forward propagation, and the diagnostic support of the nodes is calculated by backward propagation. The causal support and diagnostic support are combined and updated by normalization factor. The propagation path of the fault in the network is calculated based on the node belief value and conditional probability. The fault propagation probability is calculated based on the inter-node transmission probability and the intermediate node state.

[0118] The fault propagation path and fault propagation probability are mapped to concept nodes and weight matrices of a fuzzy cognitive graph, respectively. A fuzzy relation matrix containing fault source, fault type, scope of influence, and propagation path is constructed. The target node state vector is obtained by iteratively updating through fuzzy synthesis operation and hyperbolic tangent activation function.

[0119] The direct impact degree is calculated based on the weight matrix and target node state vector of the fuzzy cognitive graph, and the indirect impact degree is calculated by combining the fault propagation probability and the preset attenuation factor. The direct impact degree and the indirect impact degree are combined by an exponential weighting method to obtain the cumulative impact degree of the fault.

[0120] An early warning decision matrix is ​​constructed based on the probability of fault propagation and the cumulative impact of faults. Fuzzy hierarchical analysis is used to determine the threshold of the early warning level. The early warning information is divided into four levels, and early warning information including the fault source equipment, propagation path, risk level and handling measures is output in a hierarchical and progressive manner according to the early warning level.

[0121] For example, firstly, a causal inference network is constructed. Various fault modes of the relay protection device, such as failure to operate, maloperation, and delayed operation, as well as performance degradation trends, such as contact wear and insulation aging, and collaborative anomalies, such as multiple devices malfunctioning simultaneously, and reliability levels, are used as nodes in the causal inference network. Then, based on the conditional independence assumption, the conditional probabilities between nodes are calculated. For example, contact wear increases the probability of delayed operation. A recursive Bayesian estimation method is used to update the node states online, and the conditional probability parameters are optimized based on historical observation data, such as periodic inspection records and operational data. Finally, a dynamic causal graph reflecting the dynamic relationships between nodes is established. For example, a simplified causal graph could contain a causal chain where "contact wear" leads to "increased contact resistance," which in turn leads to "delayed operation" and "heating," ultimately leading to "device failure."

[0122] Next, the fault propagation path and probability are predicted based on the dynamic causal graph. Forward propagation is used to calculate the causal support of nodes, for example, the support of "contact wear" for "slow action". Backward propagation is used to calculate the diagnostic support of nodes, for example, the support of "slow action" for "contact wear". The causal support and diagnostic support are combined using a normalization factor to update the node belief value, representing the confidence of the node in a certain state. The fault propagation path in the network is calculated based on the node belief value and conditional probability. For example, if the belief value of "contact wear" is high, the probability of "slow action" occurring is also considered high, thus forming a propagation path. Then, the fault propagation probability is calculated based on the inter-node transmission probability and the intermediate node state. For example, if the transmission probability from "contact wear" to "slow action" is 0.8, and the transmission probability from "slow action" to "device failure" is 0.6, then the propagation probability of "contact wear" leading to "device failure" is 0.8 * 0.6 = 0.48.

[0123] Then, the fault propagation path and fault propagation probability are mapped to a fuzzy cognitive graph. Fault source, fault type, impact range, and propagation path are used as concept nodes in the fuzzy cognitive graph, and the fault propagation probability is used as the weight matrix. For example, if the propagation probability of "contact wear" leading to "device failure" is 0.48, then the weight of the edge connecting these two concept nodes is set to 0.48. A fuzzy relation matrix containing fault source, fault type, impact range, and propagation path is constructed, and the target node state vector is obtained through fuzzy synthesis operations and iterative updates using the hyperbolic tangent activation function. For example, if "device failure" affects "line tripping," then the state vector of "line tripping" will be updated as the state vector of "device failure" changes.

[0124] Next, the impact of the fault is assessed. The direct impact is calculated based on the weight matrix of the fuzzy cognitive graph and the state vector of the target node. For example, the direct impact of "device failure" on "line tripping". The indirect impact is calculated by combining the fault propagation probability and a preset attenuation factor. For example, the indirect impact of "contact wear" on "line tripping" through "device failure" attenuates as the propagation path length increases. An exponential weighting method is used to combine the direct and indirect impacts to obtain the cumulative fault impact.

[0125] Finally, fault warning information is output. A warning decision matrix is ​​constructed based on the fault propagation probability and the cumulative impact of the fault. Fuzzy hierarchical analysis is used to determine the warning level threshold, dividing the warning information into four levels, such as: minor, moderate, severe, and urgent. Based on the warning level, a hierarchical and progressive approach is used to output warning information including the fault source device, propagation path, risk level, and handling measures. For example, if the propagation probability of "contact wear" leading to "device failure" is high, and the cumulative impact is also high, then an urgent warning information is output, prompting immediate replacement of the device. Assuming the propagation probability from "contact wear" to "device failure" is 0.8, and the direct impact of "device failure" on "line tripping" is 0.9, then the warning information might be: Fault source device: Line 1 protection device; Propagation path: Contact wear -> Device failure -> Line tripping; Risk level: Urgent; Handling measures: Immediately replace Line 1 protection device.

[0126] In this embodiment, the combination of dynamic cause-effect graphs and fuzzy cognitive graphs enables more accurate prediction of fault propagation paths and impact levels, thereby improving the accuracy of early warnings and avoiding false alarms and missed alarms. Through fuzzy hierarchical analysis and an early warning decision matrix, fault risks can be classified, and corresponding measures can be taken according to different levels, improving the effectiveness and practicality of early warnings. Effective early warning of relay protection device faults allows for timely detection and handling of potential faults, preventing fault expansion and propagation, thus improving the reliability and safety of power system operation.

[0127] In one optional implementation, based on the dynamic causal graph, the causal support of nodes is calculated using forward propagation, and the diagnostic support of nodes is calculated using backward propagation. The causal support and diagnostic support are combined using a normalization factor to update the node belief value. The propagation path of the fault in the network is calculated based on the node belief value and conditional probability. The fault propagation probability is calculated based on the inter-node transmission probability and the intermediate node state, including:

[0128] Obtain the state information of nodes in a dynamic causal graph and construct the conditional probability matrix between nodes;

[0129] In the dynamic causal graph, the set of parent nodes of the current node is identified, the information transmission delay time from each node in the set of parent nodes to the current node is calculated, the temporal weight is determined based on the information transmission delay time, the temporal weight is combined with the physical coupling strength to generate the transmission coefficient, and the causal support of the current node is calculated through forward propagation based on the transmission coefficient and the initial belief value of each node in the set of parent nodes.

[0130] Extract the set of child nodes of the current node from the dynamic causal graph, calculate the state deviation of each node in the set of child nodes, construct a feedback influence function based on the state deviation, multiply the feedback influence function by the distance decay coefficient between nodes to obtain the diagnostic weight, and calculate the diagnostic support of the current node through backpropagation based on the diagnostic weight and the belief value of each node in the set of child nodes.

[0131] The adaptive normalization factor is calculated based on the system's operating status. The causal support and diagnostic support are then weighted and combined using the adaptive normalization factor to update the belief value of the current node.

[0132] Based on the updated current node belief value and the conditional probability matrix, the fault propagation probability between adjacent nodes is calculated, a directed propagation network is constructed based on the fault propagation probability, and the fault propagation path is extracted from the directed propagation network.

[0133] The real-time status values ​​of each node on the critical path of fault propagation are obtained, the real-time status values ​​are converted into a state function, and the cumulative propagation probability on the fault propagation path is calculated based on the state function and the fault propagation probability to obtain the final fault propagation probability.

[0134] For example, firstly, the state information of each node in the dynamic causal graph is obtained, such as the node's operating status and performance indicators, and a conditional probability matrix between the nodes is constructed based on this state information. For instance, in a system containing nodes A, B, and C, A is the parent node of B, and B is the parent node of C. If A fails, the probability of B failing is 0.8; if B is normal, the probability of C failing is 0.1; and if B fails, the probability of C failing is 0.9. These probability values ​​constitute the conditional probability matrix.

[0135] Next, identify the set of parent nodes of the current node in the dynamic causal graph. For each current node, find all parent nodes that directly influence it. For example, node B's parent node is A. Then, calculate the information transmission delay time from each node in the parent node set to the current node. For example, the information transmission delay time from A to B is 2 time units. Determine the temporal weight based on the information transmission delay time; the shorter the delay time, the greater the weight. Assuming the weight is inversely proportional to the delay time, the temporal weight from A to B is 1 / 2. Combine the temporal weight with the physical coupling strength to generate the transmission coefficient. The physical coupling strength represents the tightness of the connection between nodes; for example, the connection strength between A and B is 0.9. Assuming the transmission coefficient is equal to the temporal weight multiplied by the physical coupling strength, the transmission coefficient from A to B is (1 / 2)*0.9 = 0.45. Finally, based on the transmission coefficient and the initial belief values ​​of each node in the parent node set (for example, A's initial belief value is 0.2, indicating that the probability of A failing is 0.2), calculate the causal support of the current node through forward propagation. For example, the causal support of B is 0.45 * 0.2 = 0.09.

[0136] Then, the set of child nodes of the current node is extracted from the dynamic cause-effect graph. For example, node B's child node is C. The state deviation of each node in the child node set is calculated. The state deviation represents the difference between the child node's state and its expected state. For example, if C's expected state value is 10 and its actual state value is 8, then the state deviation is 2. A feedback influence function is constructed based on the state deviation; the larger the deviation, the greater the influence. For example, assuming the feedback influence function is equal to the square root of the state deviation, then C's feedback influence function is √2 ≈ 1.41. The diagnostic weight is obtained by multiplying the feedback influence function by the distance attenuation coefficient between nodes. The distance attenuation coefficient represents the weakening effect of distance between nodes; the greater the distance, the smaller the influence. For example, assuming the distance attenuation coefficient between B and C is 0.8, then C's diagnostic weight for B is 1.41 * 0.8 = 1.13. Finally, based on the diagnostic weight and the belief value of each node in the child node set (for example, C's belief value is 0.3), the diagnostic support of the current node is calculated through backpropagation. For example, the diagnostic support for B is 1.13 * 0.3 = 0.34.

[0137] Next, an adaptive normalization factor is calculated based on the system's operating status. For example, the higher the system load, the smaller the normalization factor. The causal support and diagnostic support are then weighted and combined using the adaptive normalization factor to update the belief value of the current node. For example, assuming a normalization factor of 0.5, B's belief value is updated to 0.5*0.09 + 0.5*0.34 = 0.22.

[0138] Based on the updated belief value of the current node and the conditional probability matrix, the probability of fault propagation between adjacent nodes is calculated. For example, if B's ​​belief value is 0.22, then the probability of fault propagation from B to C is 0.22*0.9 + (1-0.22)*0.1 = 0.28. A directed propagation network is constructed based on the fault propagation probability, and fault propagation paths are extracted from the directed propagation network. For example, the propagation path from A to C is A->B->C.

[0139] Finally, the real-time state values ​​of each node on the critical path of fault propagation are obtained and transformed into state functions. For example, if the real-time state value of node B is 8, it is transformed into the state function f(B) = 8. Based on the state function and the fault propagation probability, the cumulative propagation probability on the fault propagation path is calculated to obtain the final fault propagation probability. For example, the fault propagation probability from A to C is the propagation probability from A to B multiplied by the propagation probability from B to C. Assuming the propagation probability from A to B is 0.45 * 0.2 = 0.09, then the fault propagation probability from A to C is 0.09 * 0.28 = 0.025.

[0140] This embodiment comprehensively considers causal relationships, diagnostic information, timing characteristics, and system operating status, enabling more accurate prediction of fault propagation paths and probabilities, thereby improving fault prediction accuracy. By accurately predicting fault propagation paths, preventative measures can be taken in advance to block fault propagation, preventing wider-ranging faults and thus improving system reliability and stability. This method can help identify critical paths and nodes in fault propagation, allowing for targeted optimization of resource allocation, such as strengthening the monitoring and maintenance of critical nodes and improving resource utilization efficiency.

[0141] In one optional implementation, a hierarchical collaborative control strategy is constructed based on the fault warning information. At the local control layer, the protection setting optimization problem is transformed into a multi-objective optimization model, and a hybrid optimization algorithm is used to solve the multi-objective optimization model to obtain candidate setting schemes. At the global coordination layer, a distributed game optimization network is established based on the Nash equilibrium criterion, and a potential game method is used to coordinate and optimize the candidate setting schemes, including:

[0142] Receive system fault warning information, and extract the fault type, fault location, fault current, fault duration and fault phase angle from the fault warning information as feature quantities;

[0143] The protection reliability index is calculated based on the fault type and fault current; the protection sensitivity index is calculated based on the fault location and fault current; and the protection selectivity index is calculated based on the fault duration and protection coordination time interval.

[0144] A multi-objective optimization model is constructed at the local control layer based on the protection reliability index, protection sensitivity index, and protection selectivity index.

[0145] The multi-objective optimization model is solved by setting an adaptive crossover operator based on population density to obtain a crossover solution. The crossover solution is mutated by setting an adaptive mutation operator based on fitness value to obtain a mutated solution. A reference point set is constructed based on the mutated solution. The Euclidean distance from the mutated solution to the reference point set is calculated. The solution with the smallest Euclidean distance is selected as a candidate fixed value scheme.

[0146] In the global coordination layer, the protection unit is set as a game participant, and the candidate value scheme is set as the strategy space of the game participant to construct a distributed game optimization network.

[0147] In the distributed game optimization network, the reliability gain of game participants when choosing strategies is calculated based on the protection sensitivity index, the coordination gain between game participants and adjacent protection units is calculated based on the protection coordination time interval, and the interference loss suffered by game participants is calculated based on the electrical distance between adjacent protection units.

[0148] The utility function of the game participants is constructed by taking the reliability gains, coordination gains, and interference losses as the comprehensive gains.

[0149] The electrical connection degree is obtained by calculating the impedance values ​​between protection units in the distributed game optimization network, and the coupling coefficient between game participants is determined based on the electrical connection degree; the cooperation relationship is obtained by calculating the action time difference between protection units, and the interaction utility between game participants is determined based on the cooperation relationship; the potential function is constructed by combining the utility function, coupling coefficient and interaction utility of game participants.

[0150] In the distributed game optimization network, an initial strategy combination for the game participants is generated. The utility function value and potential function value of the game participants under the initial strategy combination are calculated. The initial strategy combination is updated based on the utility function value to obtain an optimized response strategy. The new potential function value under the optimized response strategy is calculated. When the difference between the new potential function value and the potential function value of the previous iteration is less than a preset threshold, the optimized response strategy is output as a coordinated optimization fixed-value scheme that satisfies the Nash equilibrium criterion.

[0151] For example, firstly, fault warning information is received from the system monitoring device. This information includes the fault type (short circuit, ground fault), fault location (specific line or transformer number), fault current magnitude (current value in amperes), fault duration (time length in milliseconds), and fault phase angle (angle value in degrees). This information will serve as the basis for subsequent analysis and calculations.

[0152] Next, these feature quantities are extracted and processed. Protection reliability indices are calculated based on the fault type and fault current magnitude; for example, by looking up a pre-established table of correspondences between fault types and reliability indices to obtain the corresponding reliability index values. Protection sensitivity indices are calculated based on the fault location and fault current magnitude; for example, by quantifying sensitivity by calculating the ratio of fault current to protection initiation current. Protection selectivity indices are calculated based on the fault duration and protection coordination time interval (the difference between the operating times of two adjacent protection devices); for example, by evaluating selectivity by comparing the fault duration and the protection coordination time interval.

[0153] At the local control layer, a multi-objective optimization model is constructed, aiming to simultaneously optimize protection reliability, sensitivity, and selectivity. These three indicators are used as objective functions, and corresponding constraints are set, such that the operating time of the protection device must be within a certain range.

[0154] To solve this multi-objective optimization model, a hybrid optimization algorithm is employed. First, an adaptive crossover operator is used to generate new candidate solutions. Specifically, the crossover probability is dynamically adjusted based on the population density of the current solution to improve search efficiency. Then, an adaptive mutation operator is used to mutate the crossover solutions, generating more diverse solutions. The mutation probability is dynamically adjusted based on the fitness value of the current solution to balance exploration and exploitation. Finally, a reference point set is constructed based on the mutated solutions, and the Euclidean distance from each mutated solution to the reference point set is calculated. The solution with the smallest distance is selected as the candidate setpoint. The reference points are set as an ideal combination of reliability, sensitivity, and selectivity.

[0155] In the global coordination layer, each protection unit is regarded as a game participant, and the candidate setpoint schemes obtained from the local control layer are used as the strategy space of the game participants to construct a distributed game optimization network.

[0156] In this network, each player calculates their payoff based on their chosen strategy. Reliability payoff is calculated based on protection sensitivity indicators, with higher values ​​indicating higher reliability. Coordination payoff with adjacent protection units is calculated based on protection coordination time intervals, with shorter intervals indicating better coordination. Interference loss is calculated based on the electrical distance (line length) between adjacent protection units, with closer distances indicating greater interference. These three factors are considered together to construct a utility function for each player; for example, adding reliability and coordination payoffs and subtracting interference loss.

[0157] To simulate the mutual influence between protection units, the impedance values ​​between them are calculated to determine the coupling coefficient among the game participants; a smaller impedance value indicates a higher degree of coupling. Simultaneously, the action time difference between protection units is calculated to determine the interaction utility among the game participants; a smaller action time difference indicates a stronger interaction. The utility function, coupling coefficient, and interaction utility are combined to construct a potential function.

[0158] In a distributed game optimization network, initial strategy combinations for the game participants are first randomly generated. Then, the utility function and potential function values ​​for each participant under the current strategy combination are calculated. Next, each participant updates their strategy based on their utility function value, choosing the strategy that maximizes their utility. Finally, a new potential function value is calculated. This process is repeated until the difference between the new potential function value and the potential function value from the previous iteration is less than a preset threshold. The resulting strategy combination is the coordinated optimal constant value scheme that satisfies the Nash equilibrium criterion.

[0159] In this embodiment, a multi-objective optimization model and Nash equilibrium criterion are used to comprehensively consider protection reliability, sensitivity, and selectivity indicators. This effectively avoids maloperation or failure to operate of protection devices, improves the reliability of the protection system, and ensures the safe and stable operation of the power system. Adaptive crossover and mutation operators are employed to dynamically adjust optimization parameters according to the system operating status, improving the algorithm's search efficiency and adaptability, and better handling various complex fault scenarios. Automatic generation of coordinated optimization setting schemes based on fault early warning information reduces manual intervention and debugging work, simplifies the protection setting process, and improves work efficiency.

[0160] Figure 2 This is a schematic diagram of the structure of the smart grid relay protection reliability assessment and improvement control system according to an embodiment of the present invention, as shown below. Figure 2 As shown, the system includes:

[0161] The first unit is used to collect the operating data of the relay protection device and process the data through an adaptive data preprocessing model, dividing it into steady-state, transient, and fault segments. The data in the steady-state, transient, and fault segments are respectively processed by an improved empirical mode decomposition method for noise reduction. Multi-domain features are extracted by combining wavelet packet transform and Hilbert-Huang transform to generate an enhanced feature matrix. The enhanced feature matrix is ​​input into a deep hybrid neural network to extract spatial features. Temporal correlation is captured through recurrent branching. An adaptive fusion layer is used to combine the spatial features and temporal correlation to obtain a comprehensive feature vector.

[0162] The second unit is used to input the comprehensive feature vector into a pre-constructed multi-stage fault diagnosis model. At the device level, it identifies the fault modes and performance degradation trends of individual protection devices based on deep belief networks. At the interval level, it analyzes the cooperative anomalies of adjacent protection devices based on improved conditional random fields. At the station level, it divides the reliability levels of protection device groups based on hierarchical clustering methods. The fault modes, performance degradation trends, cooperative anomalies, and reliability levels are input into a causal inference network to establish a dynamic causal graph. Based on the dynamic causal graph, the fault propagation path and propagation probability are predicted. The degree of fault impact is evaluated by combining a fuzzy cognitive graph, and fault warning information is output.

[0163] The third unit is used to construct a hierarchical collaborative control strategy based on the fault warning information. At the local control layer, the protection setting optimization problem is transformed into a multi-objective optimization model, and a hybrid optimization algorithm is used to solve the multi-objective optimization model to obtain candidate setting schemes. At the global coordination layer, a distributed game optimization network is established based on the Nash equilibrium criterion, and the potential game method is used to coordinate and optimize the candidate setting schemes. The coordinated and optimized setting schemes are input into the hierarchical rolling optimization controller, which establishes a time-domain optimization problem based on model predictive control, and achieves closed-loop control through online state estimation and feedback correction to improve the reliability of the relay protection device.

[0164] A third aspect of the present invention,

[0165] An electronic device is provided, comprising:

[0166] processor;

[0167] Memory used to store processor-executable instructions;

[0168] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0169] Fourth aspect of the embodiments of the present invention,

[0170] A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0171] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0172] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for reliability assessment and improvement control of smart grid relay protection, characterized in that, include: Operational data of relay protection devices are collected and processed through an adaptive data preprocessing model, dividing the data into steady-state, transient, and fault segments. The data in the steady-state, transient, and fault segments are denoised using an improved empirical mode decomposition method. Multi-domain features are extracted by combining wavelet packet transform and Hilbert-Huang transform to generate an enhanced feature matrix. The enhanced feature matrix is ​​then input into a deep hybrid neural network to extract spatial features. Temporal correlations are captured through recurrent branching. An adaptive fusion layer is used to combine the spatial features and temporal correlations to obtain a comprehensive feature vector. The comprehensive feature vector is input into a pre-constructed multi-stage fault diagnosis model. At the device level, a deep belief network is used to identify the fault modes and performance degradation trends of individual protection devices. At the interval level, an improved conditional random field is used to analyze the cooperative anomalies of adjacent protection devices. At the station level, a hierarchical clustering method is used to classify the reliability levels of protection device groups. The fault modes, performance degradation trends, cooperative anomalies, and reliability levels are input into a causal inference network to establish a dynamic causal graph. Based on the dynamic causal graph, the fault propagation path and propagation probability are predicted. The degree of fault impact is evaluated by combining a fuzzy cognitive graph, and fault warning information is output. Based on the fault warning information, a hierarchical collaborative control strategy is constructed. At the local control layer, the protection setting optimization problem is transformed into a multi-objective optimization model, and a hybrid optimization algorithm is used to solve the multi-objective optimization model to obtain candidate setting schemes. At the global coordination layer, a distributed game optimization network is established based on the Nash equilibrium criterion, and the potential game method is used to coordinate and optimize the candidate setting schemes. The coordinated and optimized setting schemes are input into the hierarchical rolling optimization controller, which establishes a time-domain optimization problem based on model predictive control, and achieves closed-loop control through online state estimation and feedback correction, thereby improving the reliability of the relay protection device.

2. The method according to claim 1, characterized in that, Operational data from relay protection devices are collected and processed using an adaptive data preprocessing model, dividing the data into steady-state, transient, and fault segments. An improved empirical mode decomposition method is applied to denoise the data in each segment. Multi-domain features are extracted using wavelet packet transform and Hilbert-Huang transform to generate enhanced feature matrices, including: The operation data of the relay protection device is collected. The operation data includes primary equipment current signal, primary equipment voltage signal, protection device action signal, communication link status signal, circuit breaker mechanical characteristic signal and secondary circuit insulation impedance signal. The operation data is normalized to obtain a standardized data sequence. A variational mode decomposition model is established for the standardized data sequence. The variational mode decomposition model constructs a variational optimization objective function containing a center frequency term and a bandwidth constraint term. The objective function is iteratively optimized using the alternating direction multiplier method to obtain an intrinsic mode function set. Based on the energy distribution characteristics of the intrinsic mode function set, the cross-correlation coefficient of adjacent intrinsic mode functions is calculated. An adaptive segmentation threshold is set according to the cross-correlation coefficient to divide the standardized data sequence into a steady-state segment, a transient segment, and a fault segment. The data in the steady-state segment, transient segment, and fault segment are respectively subjected to improved empirical mode decomposition denoising processing. An adaptive soft thresholding function based on data distribution is used to process the intrinsic mode components. The parameters of the adaptive soft thresholding function are adjusted according to the number of layers and energy density of the intrinsic mode components. Local mean envelope is used to replace spline interpolation to eliminate endpoint effects, resulting in denoised data. The denoised data is subjected to wavelet packet transform, and the entropy of the wavelet packet decomposition coefficients is calculated. The node coefficient with the largest entropy is selected as the frequency domain feature. The denoised data is subjected to Hilbert-Huang transform, and the instantaneous amplitude and instantaneous phase of the resulting analytic signal are calculated. The instantaneous frequency is obtained based on the instantaneous phase to obtain the time-frequency feature. The frequency domain feature vector and the time-frequency feature vector are adaptively weighted and fused, and high-order statistical features based on kernel density estimation are introduced, including mean, variance, skewness, kurtosis, quantile features and entropy features. After dimensionality reduction using principal component analysis, an enhanced feature matrix is ​​formed.

3. The method according to claim 2, characterized in that, A variational mode decomposition model is established for the standardized data sequence. This model constructs a variational optimization objective function containing a center frequency term and a bandwidth constraint term. The objective function is iteratively optimized using the alternating direction multiplier method to obtain an intrinsic mode function set. Based on the energy distribution characteristics of the intrinsic mode function set, the cross-correlation coefficients of adjacent intrinsic mode functions are calculated. An adaptive segmentation threshold is set according to the cross-correlation coefficients to divide the standardized data sequence into steady-state, transient, and fault segments, including: The standardized data sequence is discretely sampled and framed. Each frame of data is converted into an analytical signal through Hilbert transform. The analytical signal is frequency modulated to obtain a center frequency sequence. A reconstruction error matrix and a spectral constraint matrix are constructed based on the center frequency sequence. The reconstruction error matrix and spectral constraint matrix are input into a Wiener filter, and the frequency response of the Wiener filter is adaptively adjusted according to the bandwidth constraint parameter to output modal components. Soft threshold shrinkage processing is applied to the modal components, and the soft threshold parameter is determined through cross-validation. The power spectrum centroid is calculated based on the processed modal components to obtain an updated center frequency sequence. The center frequency sequence is optimized through gradient iteration to obtain the intrinsic mode function set. Calculate the instantaneous energy spectrum of each intrinsic mode function in the intrinsic mode function set to obtain the energy density distribution curve. Based on the energy density distribution curve, calculate the energy concentration and energy dispersion to obtain the energy distribution feature matrix. Extract the energy envelopes of adjacent intrinsic mode functions, calculate the normalized cross-correlation function of the energy envelopes, calculate the peak value of the normalized cross-correlation function within a sliding time window, and generate a cross-correlation coefficient sequence; Calculate the mean and standard deviation of the cross-correlation coefficient sequence, and set the mean as the baseline threshold; simultaneously, multiply the energy concentration and energy dispersion in the energy distribution feature matrix to obtain the threshold adjustment factor; set the product of the baseline threshold and the threshold adjustment factor as the adaptive segmented threshold; Within the sliding time window, the cross-correlation coefficient sequence is compared with the corresponding adaptive segmentation threshold. When the cross-correlation coefficients of three consecutive windows are all less than the adaptive segmentation threshold, they are marked as candidate mutation points. The local energy density of the candidate mutation points is calculated, and the final modal mutation point location is determined based on the minimum energy density. The standardized data sequence is divided into steady-state segment, transient segment, and fault segment according to the modal mutation points.

4. The method according to claim 1, characterized in that, The comprehensive feature vector is input into a pre-constructed multi-stage fault diagnosis model. At the device level, deep belief networks are used to identify the fault modes and performance degradation trends of individual protection devices. At the interval level, improved conditional random fields are used to analyze the cooperative anomalies of adjacent protection devices. At the station level, hierarchical clustering methods are used to classify the reliability levels of protection device groups, including: A multi-stage fault diagnosis model is constructed, which includes a device-level diagnosis sub-model, a bay-level diagnosis sub-model, and a station-level diagnosis model. The device-level diagnostic sub-model adopts a deep belief network. The comprehensive feature vector is input into the deep belief network and pre-trained layer by layer through a restricted Boltzmann machine. The network weights are optimized by contrastive divergence algorithm and stochastic gradient descent method. Label data is introduced to fine-tune the network and obtain the fault mode probability vector and performance degradation index of a single protection device. The interval-level diagnostic sub-model adopts a conditional random field structure, uses the fault mode probability vector and performance degradation index as node features, constructs a spatial potential function based on the topological relationship of primary equipment, and constructs a temporal potential function based on the correlation of operating status. The fault mode probability vector, performance degradation index, spatial potential function and temporal potential function are input into the conditional random field structure, and the model parameters are optimized through maximum likelihood estimation to output the cooperative anomaly probability matrix of adjacent devices. The station-level diagnostic sub-model adopts a hierarchical clustering structure, which integrates the fault mode probability vector, performance degradation index and collaborative anomaly probability matrix to construct a state representation matrix. Based on the state representation matrix, the state distance between devices is calculated, and the devices are grouped using a hierarchical clustering algorithm with adaptive distance metric. The optimal number of groups is determined by the contour coefficient to obtain the reliability level of the protection device.

5. The method according to claim 1, characterized in that, The fault modes, performance degradation trends, collaborative anomalies, and reliability levels are input into a causal inference network to establish a dynamic causal graph. Based on this dynamic causal graph, the fault propagation path and propagation probability are predicted. The impact of the fault is then assessed using a fuzzy cognitive graph, and fault warning information is output, including: The fault modes, performance degradation trends, cooperative anomalies, and reliability levels of relay protection devices are used to construct a node set of a causal inference network. Based on the assumption of conditional independence, the conditional probability table between nodes is calculated. The recursive Bayesian estimation method is used to update the node state online. The conditional probability parameters are optimized based on historical observation data to establish a dynamic causal graph that reflects the dynamic relationship between nodes. Based on the dynamic causal graph, the causal support of the nodes is calculated by forward propagation, and the diagnostic support of the nodes is calculated by backward propagation. The causal support and diagnostic support are combined and updated by normalization factor. The propagation path of the fault in the network is calculated based on the node belief value and conditional probability. The fault propagation probability is calculated based on the inter-node transmission probability and the intermediate node state. The fault propagation path and fault propagation probability are mapped to concept nodes and weight matrices of a fuzzy cognitive graph, respectively. A fuzzy relation matrix containing fault source, fault type, scope of influence, and propagation path is constructed. The target node state vector is obtained by iteratively updating through fuzzy synthesis operation and hyperbolic tangent activation function. The direct impact degree is calculated based on the weight matrix and target node state vector of the fuzzy cognitive graph, and the indirect impact degree is calculated by combining the fault propagation probability and the preset attenuation factor. The direct impact degree and the indirect impact degree are combined by an exponential weighting method to obtain the cumulative impact degree of the fault. An early warning decision matrix is ​​constructed based on the probability of fault propagation and the cumulative impact of faults. Fuzzy hierarchical analysis is used to determine the threshold of the early warning level. The early warning information is divided into four levels, and early warning information including the fault source equipment, propagation path, risk level and handling measures is output in a hierarchical and progressive manner according to the early warning level.

6. The method according to claim 5, characterized in that, Based on the dynamic causal graph, the causal support of nodes is calculated using forward propagation, and the diagnostic support of nodes is calculated using backward propagation. The causal support and diagnostic support are combined using a normalization factor to update the node belief value. The propagation path of the fault in the network is calculated based on the node belief value and conditional probability. The fault propagation probability is calculated based on the inter-node transmission probability and the intermediate node state, including: Obtain the state information of nodes in a dynamic causal graph and construct the conditional probability matrix between nodes; In the dynamic causal graph, the set of parent nodes of the current node is identified, the information transmission delay time from each node in the set of parent nodes to the current node is calculated, the temporal weight is determined based on the information transmission delay time, the temporal weight is combined with the physical coupling strength to generate the transmission coefficient, and the causal support of the current node is calculated through forward propagation based on the transmission coefficient and the initial belief value of each node in the set of parent nodes. Extract the set of child nodes of the current node from the dynamic causal graph, calculate the state deviation of each node in the set of child nodes, construct a feedback influence function based on the state deviation, multiply the feedback influence function by the distance decay coefficient between nodes to obtain the diagnostic weight, and calculate the diagnostic support of the current node through backpropagation based on the diagnostic weight and the belief value of each node in the set of child nodes. The adaptive normalization factor is calculated based on the system's operating status. The causal support and diagnostic support are then weighted and combined using the adaptive normalization factor to update the belief value of the current node. Based on the updated current node belief value and the conditional probability matrix, the fault propagation probability between adjacent nodes is calculated, a directed propagation network is constructed based on the fault propagation probability, and the fault propagation path is extracted from the directed propagation network. The real-time status values ​​of each node on the critical path of fault propagation are obtained, the real-time status values ​​are converted into a state function, and the cumulative propagation probability on the fault propagation path is calculated based on the state function and the fault propagation probability to obtain the final fault propagation probability.

7. The method according to claim 1, characterized in that, Based on the fault warning information, a hierarchical collaborative control strategy is constructed. In the local control layer, the protection setting optimization problem is transformed into a multi-objective optimization model. A hybrid optimization algorithm is used to solve the multi-objective optimization model to obtain candidate setting schemes. A distributed game optimization network is established at the global coordination layer based on the Nash equilibrium criterion. The potential game method is used to coordinate and optimize the candidate fixed-value schemes, including: Receive system fault warning information, and extract the fault type, fault location, fault current, fault duration and fault phase angle from the fault warning information as feature quantities; The protection reliability index is calculated based on the fault type and fault current; the protection sensitivity index is calculated based on the fault location and fault current; and the protection selectivity index is calculated based on the fault duration and protection coordination time interval. A multi-objective optimization model is constructed at the local control layer based on the protection reliability index, protection sensitivity index, and protection selectivity index. The multi-objective optimization model is solved by setting an adaptive crossover operator based on population density to obtain a crossover solution. The crossover solution is mutated by setting an adaptive mutation operator based on fitness value to obtain a mutated solution. A reference point set is constructed based on the mutated solution. The Euclidean distance from the mutated solution to the reference point set is calculated. The solution with the smallest Euclidean distance is selected as a candidate fixed value scheme. In the global coordination layer, the protection unit is set as a game participant, and the candidate value scheme is set as the strategy space of the game participant to construct a distributed game optimization network. In the distributed game optimization network, the reliability gain of game participants when choosing strategies is calculated based on the protection sensitivity index, the coordination gain between game participants and adjacent protection units is calculated based on the protection coordination time interval, and the interference loss suffered by game participants is calculated based on the electrical distance between adjacent protection units. The utility function of the game participants is constructed by taking the reliability gains, coordination gains, and interference losses as the comprehensive gains. The electrical connection degree is obtained by calculating the impedance values ​​between protection units in the distributed game optimization network, and the coupling coefficient between game participants is determined based on the electrical connection degree; the cooperation relationship is obtained by calculating the action time difference between protection units, and the interaction utility between game participants is determined based on the cooperation relationship; the potential function is constructed by combining the utility function, coupling coefficient and interaction utility of game participants. In the distributed game optimization network, an initial strategy combination for the game participants is generated. The utility function value and potential function value of the game participants under the initial strategy combination are calculated. The initial strategy combination is updated based on the utility function value to obtain an optimized response strategy. The new potential function value under the optimized response strategy is calculated. When the difference between the new potential function value and the potential function value of the previous iteration is less than a preset threshold, the optimized response strategy is output as a coordinated optimization fixed-value scheme that satisfies the Nash equilibrium criterion.

8. A smart grid relay protection reliability assessment and improvement control system, used to implement the method of any one of claims 1-7, characterized in that, include: The first unit is used to collect the operating data of the relay protection device and process the data through an adaptive data preprocessing model, dividing it into steady-state, transient, and fault segments. The data in the steady-state, transient, and fault segments are respectively processed by an improved empirical mode decomposition method for noise reduction. Multi-domain features are extracted by combining wavelet packet transform and Hilbert-Huang transform to generate an enhanced feature matrix. The enhanced feature matrix is ​​input into a deep hybrid neural network to extract spatial features. Temporal correlation is captured through recurrent branching. An adaptive fusion layer is used to combine the spatial features and temporal correlation to obtain a comprehensive feature vector. The second unit is used to input the comprehensive feature vector into a pre-constructed multi-stage fault diagnosis model. At the device level, it identifies the fault modes and performance degradation trends of individual protection devices based on deep belief networks. At the interval level, it analyzes the cooperative anomalies of adjacent protection devices based on improved conditional random fields. At the station level, it divides the reliability levels of protection device groups based on hierarchical clustering methods. The fault modes, performance degradation trends, cooperative anomalies, and reliability levels are input into a causal inference network to establish a dynamic causal graph. Based on the dynamic causal graph, the fault propagation path and propagation probability are predicted. The degree of fault impact is evaluated by combining a fuzzy cognitive graph, and fault warning information is output. The third unit is used to construct a hierarchical collaborative control strategy based on the fault warning information. At the local control layer, the protection setting optimization problem is transformed into a multi-objective optimization model, and a hybrid optimization algorithm is used to solve the multi-objective optimization model to obtain candidate setting schemes. At the global coordination layer, a distributed game optimization network is established based on the Nash equilibrium criterion, and the potential game method is used to coordinate and optimize the candidate setting schemes. The coordinated and optimized setting schemes are input into the hierarchical rolling optimization controller, which establishes a time-domain optimization problem based on model predictive control, and achieves closed-loop control through online state estimation and feedback correction to improve the reliability of the relay protection device.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to 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 the processor, they implement the method described in any one of claims 1 to 7.

Images