A safety state monitoring system and method for a smart box-type substation

Abstract

The application discloses a kind of security state monitoring system and method for intelligent box-type substation, it is related to electric power detection technical field, including: the construction equipment failure propagation path set;Excavate weak abnormal feature information in the monitoring signal is covered by background noise;Weak abnormal feature information is input into the state evolution model of corresponding equipment;The propagation process of weak abnormal feature information along equipment failure propagation path set is simulated using random walk algorithm, and the evolution trend of equipment abnormal state is predicted;According to the dynamic deviation degree between the evolution trend of abnormal state and the predetermined security state of equipment, generate equipment risk dynamic distribution curve, determine the risk level of the current security state of box-type substation;The application realizes early abnormal accurate identification and dynamic risk assessment of the security state of box-type substation, effectively improves the predictability and reliability of substation safe operation.

Description

Technical Field

[0001] This invention relates to the field of power detection technology, specifically to a safety status monitoring system and method for intelligent prefabricated substations. Background Technology

[0002] Prefabricated substations, as key infrastructure in power distribution networks, are widely used in urban power distribution, industrial parks, and new energy integration scenarios. Their safe and stable operation is of great significance to ensuring the overall reliability of the power system. In recent years, with the expansion of power grid scale, the diversification of loads, and the improvement of equipment automation, the operating environment of prefabricated substations has become increasingly complex. Because the equipment operates in outdoor or semi-outdoor environments for extended periods, its operating status is highly susceptible to external environmental interference and factors such as equipment aging and wear, leading to a continuous increase in the risk of equipment failure.

[0003] Currently, traditional methods for condition monitoring of equipment in prefabricated substations primarily focus on individual device condition monitoring, typically concentrating only on isolated analysis of device status data and setting basic alarm thresholds. This single-point monitoring approach neglects the interrelationships between the state evolution of different devices, making it difficult to comprehensively reveal potential fault propagation or condition degradation trends among various devices within the substation, and consequently, hindering the effective prediction and management of potential safety risks. Furthermore, because field monitoring data often contains a certain degree of background noise, traditional monitoring methods lack sufficient sensitivity to weak anomalies implicit in the data, potentially leading to the failure to identify early abnormal signals in a timely manner, thereby increasing the likelihood of equipment failures occurring and spreading. Summary of the Invention

[0004] The purpose of this invention is to provide a safety status monitoring system and method for intelligent prefabricated substations to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] In a first aspect, the present invention provides a method for monitoring the safety status of an intelligent prefabricated substation, comprising:

[0007] A fault propagation network model is established based on the topology of the equipment in the prefabricated substation; the potential propagation paths in the fault propagation network model are identified using the dynamic Bayesian probabilistic reasoning method, and a set of equipment fault propagation paths is constructed.

[0008] Based on sparse autoencoder networks, sparse coding reconstruction is performed on multi-source monitoring signals of the device to mine weak abnormal feature information in the monitoring signals that are masked by background noise.

[0009] input the weak abnormal feature information into a state evolution model of a corresponding device; utilize a random walk algorithm to simulate a propagation process of the weak abnormal feature information along a set of device fault propagation paths, and predict an evolution trend of an abnormal state of the device;

[0010] generate a device risk dynamic distribution curve according to a dynamic deviation degree between the evolution trend of the abnormal state and a predetermined safe state of the device, and determine a risk level of a current safe state of the box-type substation.

[0011] In a second aspect, the present application provides a safe state monitoring system for an intelligent box-type substation, which is realized based on the above-mentioned safe state monitoring method for the intelligent box-type substation, and comprises:

[0012] a path construction module, configured to establish a fault propagation network model according to a topological relationship of devices of the box-type substation, and utilize a dynamic Bayesian probability inference method to identify potential propagation paths in the fault propagation network model, and construct a set of device fault propagation paths;

[0013] a feature mining module, configured to perform sparse coding reconstruction on a plurality of monitoring signals of the device based on a sparse self-encoding network, so as to mine weak abnormal feature information in the monitoring signals which is covered by background noise;

[0014] a trend prediction module, configured to input the weak abnormal feature information into a state evolution model of a corresponding device; utilize a random walk algorithm to simulate a propagation process of the weak abnormal feature information along a set of device fault propagation paths, and predict an evolution trend of an abnormal state of the device;

[0015] a risk analysis module, configured to generate a device risk dynamic distribution curve according to a dynamic deviation degree between the evolution trend of the abnormal state and a predetermined safe state of the device, and determine a risk level of a current safe state of the box-type substation.

[0016] In the above technical solution, the present application has the following technical effects and advantages:

[0017] The present application effectively identifies potential fault propagation paths between devices by establishing a fault propagation network model based on a dynamic Bayesian probability inference, realizes accurate prediction and monitoring of fault propagation risks of the box-type substation, and thus overcomes the defect that traditional single-device state monitoring cannot reveal fault cross-device propagation risks, and improves the comprehensiveness and foresight of device operation risk prediction.

[0018] The present application successfully mines the weak abnormal characteristic signal which is difficult to be found by traditional monitoring method and is covered by background noise by using sparse auto-encoding network to sparse coding reconstruction of device multi-source monitoring signal, so as to realize accurate identification and extraction of early weak abnormal state of the device, avoid the problem that potential abnormal state of the device develops into serious failure due to failure to find in time, and improve the sensitivity and reliability of the device state monitoring.

[0019] The present application realizes real-time and dynamic evaluation and early warning of the overall safety state risk of the box-type substation by introducing the extracted weak abnormal characteristic information into the random walk algorithm for state propagation simulation, dynamically predicting the evolution trend of the abnormal state of the device, and further determining the risk level according to the dynamic deviation degree between the predicted trend and the predetermined safety state of the device, thereby effectively improving the fine level and accident prevention ability of the substation safety management. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description only represent some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art according to these drawings.

[0021] Figure 1 A flow chart of a safety state monitoring method for an intelligent box-type substation according to the present application;

[0022] Figure 2 A framework diagram of a safety state monitoring system for an intelligent box-type substation according to the present application. DETAILED DESCRIPTION

[0023] Example implementations will now be described more fully with reference to the accompanying drawings. Example implementations may, however, be implemented in many different forms and should not be construed as limited to the examples set forth herein; rather, these example implementations are provided so that this disclosure will be thorough and complete, and will fully convey the scope of example implementations to those skilled in the art. The accompanying drawings are included to provide a further understanding of example implementations and are incorporated into and constitute a part of this application. The drawings are not intended to be restrictive in any way. Throughout the drawings, like references numerals denote like features, and thus repeated description is omitted for clarity.

[0024] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more exemplary embodiments. Numerous specific details are provided in the following description to give a full understanding of the exemplary embodiments disclosed in this application. However, those skilled in the art will recognize that the technical solutions disclosed in this application can be practiced with one or more specific details omitted, or other methods, components, steps, etc., can be employed. In other instances, well-known structures, methods, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of the disclosure of this application. Example 1

[0025] like Figure 1 As shown in the figure, this embodiment discloses a method for monitoring the safety status of an intelligent prefabricated substation, including:

[0026] S101: Establish a fault propagation network model based on the topology of the prefabricated substation equipment; use dynamic Bayesian probabilistic reasoning to identify potential propagation paths in the fault propagation network model and construct a set of equipment fault propagation paths;

[0027] In one optional implementation, the construction of the set of device fault propagation paths includes:

[0028] Based on the topological connections of device nodes in the fault propagation network model, the initial state transition structure between device nodes is determined, and an initial Bayesian network model is established to obtain the initial state transition probability matrix of the device nodes.

[0029] It should be noted that the fault propagation network model in this embodiment is a directed graph structure established based on the electrical and communication connection topology of the equipment in the actual prefabricated substation. Specifically, each node represents a specific piece of equipment in the substation, and typical equipment includes, but is not limited to, transformers, high-voltage circuit breakers, low-voltage distribution cabinets, protection relays, and smart sensors. The directed edges of the network represent the possible path directions of fault propagation, such as: transformer → high-voltage circuit breaker, high-voltage circuit breaker → low-voltage distribution cabinet, protection relay → high-voltage circuit breaker, etc.

[0030] It should be further noted that this embodiment divides the status of each device node into three typical states: normal, abnormal, and fault. For example, taking the high-voltage circuit breaker node as an example:

[0031] Normal: All monitoring indicators of the equipment are within the rated range;

[0032] Abnormality: The equipment exhibits abnormal indicators, such as reduced insulation resistance, but the fault threshold for equipment shutdown has not been reached.

[0033] Fault: The equipment experiences a significant fault, such as insulation breakdown, causing the equipment to stop operating.

[0034] Based on the above state definition and network topology, an initial Bayesian network model is established, and an initial state transition probability matrix is obtained , which is specifically as follows:

[0035]

[0036] In the formula, each probability element is preliminarily determined by historical operation experience or expert experience, and the specific value is determined according to actual historical data.

[0037] Based on the historical failure data of the equipment, a probability likelihood function of state transition is constructed, and a conditional expectation of the state transition probability of the equipment is derived to obtain an initial estimation of the state transition probability.

[0038] In a specific implementation, the historical failure data of the equipment in the embodiment refers to long-term accumulated historical failure records in the operation process of each equipment in the box-type substation. Typical data includes: specific time of equipment failure, duration, failure mode, failure propagation equipment range, and actual number of state transitions of each equipment.

[0039] Based on the above historical data, a likelihood function of the state transition probability is constructed, and the specific formula is as follows:

[0040]

[0041] In the formula: represents the likelihood function; is a state transition probability parameter to be estimated; represents the state of the equipment at the time of the th historical sample; represents the total number of historical samples; represents the duration length observed by a single historical sample.

[0042] Subsequently, a conditional expectation function of the state transition probability is obtained through Bayesian probability derivation:

[0043]

[0044] In the formula: is an estimation value of the probability parameter obtained in the th iteration, which is used as an initial input value for the next iteration of the EM algorithm.

[0045] The initial estimation value is used as an algorithm input, and the maximum expectation algorithm is used for multiple iteration calculation of the state transition probability matrix, and the convergence of the likelihood function is determined to obtain an optimized state transition probability matrix.

[0046] Specifically, the multiple iteration calculation of the state transition probability matrix by the maximum likelihood expectation algorithm and the convergence of the likelihood function are determined, comprising:

[0047] According to the initial estimate of the state transition probability, the alternating iteration process of the expectation step and the maximization step is performed to update the value of the device state transition probability;

[0048] In implementation, the specific calculation formula of the expectation step (E-step) is as follows:

[0049]

[0050] In the formula: The expected number of state transitions from state to state in the th iteration process is represented by The actual number of state transitions from state to state .

[0051] The specific calculation method of the maximization step (M-step) is as follows:

[0052]

[0053] The above steps are alternately performed to update the state transition probability value.

[0054] After each iteration, the likelihood function value is recalculated and compared with the results of the adjacent two times to determine whether the likelihood function meets the convergence condition;

[0055] In specific implementation, the difference between the likelihood functions obtained by adjacent two iterations is used as the convergence judgment basis, and the specific formula is as follows: Wherein, the threshold is determined by actual experimental data.

[0056] When the likelihood function meets the set convergence condition, the iteration process is stopped, and the optimized state transition probability matrix is obtained;

[0057] In specific implementation, the optimized state transition probability matrix is as follows:

[0058]

[0059] Wherein, each matrix element is the state transition probability obtained after the EM algorithm converges.

[0060] According to the optimized state transition probability matrix, a Monte Carlo simulation process is performed to screen out the fault propagation paths that meet the set probability threshold, and a device fault propagation path set is obtained;

[0061] In practice, the Monte Carlo stochastic simulation method is used to perform a large number of random simulations of equipment state transitions using the optimized state transition probability matrix. For example, after 10,000 random simulations, the probability of occurrence of each fault propagation path is statistically analyzed. ;in, The probability of a fault propagation path occurring. The number of times the path appears. This represents the total number of simulations.

[0062] Further, a probability threshold is set according to actual needs (which can be determined based on experimental data). Paths with a probability exceeding the threshold are considered valid propagation paths, and a set of equipment fault propagation paths is ultimately formed as the basis for subsequent safety status analysis.

[0063] S102: Based on a sparse autoencoder network, the multi-source monitoring signals of the device are sparsely encoded and reconstructed to mine weak abnormal feature information in the monitoring signals that are masked by background noise.

[0064] In one optional implementation, the sparse coding reconstruction of the multi-source monitoring signals of the device includes:

[0065] A multidimensional time-series dataset of equipment operating status is constructed from multi-source monitoring signals, and the initial feature dimensions and feature space are determined.

[0066] It should be noted that the multi-source monitoring signals in this embodiment include, but are not limited to, various types of equipment operation monitoring data such as transformer winding temperature, high-voltage circuit breaker operation count, partial discharge signals of electrical equipment, bus voltage and current signals, and temperature and humidity sensing signals of the internal environment of the transformer substation; the above signals are collected in real time by various sensors to form a multi-source signal set of equipment operation status.

[0067] Specifically, based on the different sampling periods and data types of the monitored signals, each signal is synchronized to a unified timestamp, and a multi-dimensional time series dataset is constructed in units of fixed duration (e.g., 5 minutes). Each time series data sample is specifically defined as follows: ;in, For the first Each dimension corresponds to a one-dimensional monitoring signal vector, representing the time series data of monitoring signals for a single type of device. This refers to the number of signal dimensions, i.e., the number of categories of multi-source monitored signals, such as temperature, partial discharge, voltage, current, etc.

[0068] The initial feature dimension and feature space are the data dimensions and the value space of each dimension, as mentioned above. For example, the transformer winding temperature takes values ​​in the real number range (e.g., The partial discharge signal is the value of the discharge pulse frequency (e.g., ).

[0069] Based on the distribution characteristics of multidimensional time series data, the sparse regularization parameters of the sparse autoencoder network are determined, and the initial parameter set of the network is established accordingly.

[0070] In practical implementation, the sparse autoencoder network (SAE) is used to achieve noise reduction, reconstruction, and weak anomaly feature extraction of multi-source monitoring signals. Sparse regularization parameters (e.g., sparsity parameters) Regularization factor The method for determining the value is as follows: based on the actual distribution characteristics of multidimensional time series data, it is determined through the distribution of a large number of samples.

[0071] Specifically, the loss function (Loss) of a sparse autoencoder network is defined as:

[0072]

[0073] In the formula: Indicates the first The input signal of each sample, For the corresponding reconstructed output; The total number of samples; , The first term represents the network weights and biases; the second term is the weight decay term to prevent overfitting; the third term is the sparsity regularization term, where... divergence Defined as:

[0074]

[0075] in, For the desired sparsity value, Indicates the first The average activation value of each hidden layer neuron; , All of these are regularization parameters determined through experimental data.

[0076] The initial parameter set of the network includes the weight matrices and bias vectors of each layer, and the specific parameter initialization method, such as using the Xavier method.

[0077] A method combining forward propagation and backward iterative optimization is adopted to perform sparse compression coding on multidimensional time series data using the initial parameter set of the network, thereby obtaining preliminary sparse feature codes.

[0078] Specifically, the step of using the network initial parameter set to perform sparse compression encoding on multidimensional time series data includes:

[0079] input the network initial parameter set to a sparse auto-encoding network, calculate preliminary encoding features of the multidimensional time series data through a forward propagation operation;

[0080] In a specific implementation, the sparse auto-encoding network of the embodiment includes an input layer, a hidden encoding layer and an output decoding layer, wherein the number of neurons of the hidden encoding layer is lower than the dimension of the input layer, so as to achieve the purpose of data compression encoding. Given an input multidimensional time series data sample X, the forward propagation calculation content is as follows:

[0081] Hidden encoding layer activation output: ;

[0082] Output decoding layer reconstruction result: ;

[0083] wherein, , are a weight matrix and a bias vector from the input layer to the hidden encoding layer, , are a weight matrix and a bias vector from the hidden encoding layer to the output layer, is a Sigmoid or ReLU activation function.

[0084] The sparse encoding feature vector of the hidden encoding layer output is calculated through the above process .

[0085] Based on the difference between the preliminary encoding features and the original monitoring signal, a sparse regularization loss function is established and the loss gradient value is calculated;

[0086] In a specific implementation, the sparse regularization loss function is defined as follows:

[0087]

[0088] wherein, the first term is a reconstruction error term, the second term is a regularization term for preventing overfitting, and the third term is a sparse constraint term.

[0089] Based on the above loss function, the gradient of the loss function relative to each parameter of the network is calculated, and the gradient calculation method is, for example, a back propagation (BP algorithm), and the specific gradient calculation content is as follows:

[0090] Output layer error term: ;

[0091] Hidden encoding layer error term:

[0092] ;

[0093] Further, the gradient of each parameter is obtained:

[0094]

[0095] wherein: is the derivative of the corresponding activation function.

[0096] updating the network parameters according to the loss gradient value, performing reverse iterative optimization, and obtaining preliminary sparse feature encoding of the multi-dimensional time series data;

[0097] In a specific implementation, the gradient descent algorithm or an improved algorithm thereof (such as the Adam optimizer) is used to update the network parameters:

[0098] Taking the Adam optimization algorithm as an example, the parameter updating rule is: wherein: is the learning rate, respectively represent the first moment and the second moment estimates of the gradient, is a smoothing term to prevent the denominator from being zero.

[0099] Through the above reverse iterative optimization process, the optimized sparse auto-encoding network parameters are obtained through multiple iterations until convergence, and then the preliminary sparse feature encoding of the multi-dimensional time series data is obtained.

[0100] According to the residual distribution between the preliminary sparse feature encoding and the original monitoring signal, the effective residual interval is extracted and feature enhancement reconstruction is performed to mine weak abnormal feature information hidden by background noise.

[0101] In a specific implementation, first, the residual between the original signal and the preliminary sparse feature reconstruction signal is calculated: wherein, represents the multi-dimensional monitoring signal residual signal.

[0102] Further, according to the statistical distribution characteristics of the residual, the residual signal whose residual exceeds a preset threshold interval (the threshold is determined by actual historical data or experiment) is selected as a suspected weak abnormal feature interval, denoted as:

[0103] The effective residual signal interval is subjected to feature enhancement reconstruction, and the specific enhancement method is, for example, wavelet transform or empirical mode decomposition (EMD) processing on the effective residual interval. Taking wavelet transform as an example:

[0104] performing discrete wavelet transform (DWT) on to extract high-frequency detail coefficients;

[0105] performing threshold denoising and enhancement processing (for example, soft threshold denoising method) on the high-frequency coefficients;

[0106] reconstructing the enhanced weak abnormal signal feature through inverse wavelet transform.

[0107] ​Through the above processing, the weak abnormal feature signal covered by the background noise can be further highlighted for subsequent device abnormal state prediction and risk assessment.

[0108] S103: inputting the weak abnormal feature information into a state evolution model of the corresponding device; simulating a propagation process of the weak abnormal feature information along a set of device fault propagation paths by using a random walk algorithm, and predicting an evolution trend of the device abnormal state;

[0109] In an optional embodiment, the prediction of the evolution trend of the device abnormal state comprises:

[0110] Based on the set of device fault propagation paths, an initial propagation weight of a device node in each fault path is calculated, and an initial random walk weight matrix is established.

[0111] It should be noted that the set of fault propagation paths is obtained by step S101, and is specifically represented as a plurality of fault propagation paths, each path being composed of a sequence of device nodes, for example, path examples:

[0112] Path 1: transformer → high-voltage circuit breaker → low-voltage distribution cabinet;

[0113] Path 2: busbar → high-voltage circuit breaker → relay protection device.

[0114] The initial propagation weight calculation method is specifically: the initial propagation probability weight between nodes is determined according to the frequency of actual fault propagation of the device node in the historical data or expert experience. Specifically, for any adjacent node pair on each path , the initial propagation weight is defined as: ; wherein, represents the number of fault propagations from node to node in the historical data or expert experience statistics; the denominator represents the sum of the fault propagation numbers from node to all subsequent nodes.

[0115] The initial propagation weights between device nodes in all paths are calculated according to the above method, and an initial random walk weight matrix is constructed, which has the following matrix form: ; wherein, represents the total number of device nodes in the network, and the matrix element represents the initial probability weight of fault propagation between nodes.

[0116] According to the initial state distribution of the weak abnormal feature information, a path sampling process of random walk is performed to generate a state transition sample set of abnormal feature information.

[0117] In specific implementation, this embodiment first determines the initial state distribution of the abnormal information based on the weak abnormal feature information extracted in step S102. For example, abnormal feature information can be represented as the abnormal signal strength of a specific device, and the initial state probability distribution can be defined as: ;in, Represents device node The probability of an anomaly occurring at the initial anomaly time is determined by the feature enhancement result of step S102.

[0118] Subsequently, with the initial probability distribution Starting from the initial distribution, a random walk method is used to sample paths under the guidance of the random walk weight matrix. For example, each walk extracts the initial abnormal node from the initial distribution, and determines the next node to propagate based on the initial weight matrix. This process is repeated until the set walk length is reached.

[0119] Each random walk generates a path sequence that serves as a state transition sample for anomaly feature information. After performing a large number of random walks (e.g., 100), a set of state transition samples for anomaly feature information is obtained. This is used for subsequent optimization of the random walk weight matrix.

[0120] The initial random walk weight matrix is ​​optimized and adjusted based on the state transition sample set to obtain the optimized random walk weight matrix.

[0121] Specifically, the optimization and adjustment of the initial random walk weight matrix includes:

[0122] Cluster analysis was performed on the state transition sample set to obtain clustering results for different state transition patterns;

[0123] In specific implementation, this embodiment first processes the state transition sample set generated above. Perform cluster analysis, using methods such as K-means clustering or DBSCAN density clustering algorithm.

[0124] For example, taking the K-means clustering method, features are first extracted from each state transition sample sequence, such as using the frequency of occurrence of each node in the sequence or the number of times a specific node sequence appears as a clustering feature:

[0125] Specifically, no. Sample The eigenvector is defined as: In the formula, For nodes In the state transition sequence The frequency or number of times it appears in;

[0126] Subsequently, K-means clustering was performed using the aforementioned feature vectors to obtain a set of different state transition patterns: ,in, For the first Clustering The number of clusters is determined by experiments or statistical data.

[0127] Based on the clustering results, the dominant path for state transition between device nodes is determined, and the corresponding elements in the random walk weight matrix are adjusted according to the dominant path.

[0128] In specific implementation, this embodiment defines the state transition path with the highest percentage of samples within each cluster as the dominant path of that cluster. For example, taking the first... Cluster For example, its advantageous path is defined as:

[0129] Statistical clusters The frequency of occurrence of each state transition path within the system;

[0130] The path that appears most frequently is selected as the dominant path, denoted as: ;

[0131] Subsequently, the elements of the random walk weight matrix are adjusted based on the statistical frequency of each dominant path. The specific adjustment method is as follows:

[0132] If a node pair in a certain advantageous path The frequency of occurrence was significantly higher than that of the initial random weights. If the set value is not met, the weight will be increased accordingly;

[0133] If the value is lower than the initial setting, the weight should be reduced appropriately.

[0134] The specific weight adjustment formula can be defined as follows: In the formula: For node pairs in the clustering results The frequency of occurrence of the dominant path; Average frequency for all node pairs; It is an adjustment factor greater than zero, determined based on actual data.

[0135] The initial random walk weight matrix is ​​optimized and adjusted using the method described above, resulting in the optimized random walk weight matrix. .

[0136] The adjusted random walk weight matrix is ​​subjected to probability normalization to obtain the optimized random walk weight matrix.

[0137] In a specific implementation, after the adjusted random walk weight matrix is obtained, the elements of the matrix are further subjected to probability normalization processing, and the specific formula is defined as: ; wherein, is the weight value of the optimized random walk from the node to the node ; the denominator is the sum of the adjusted weights of all subsequent nodes of the node ; and represents the total number of nodes.

[0138] After the above normalization processing, the final optimized random walk weight matrix is obtained, which satisfies the normalization condition of the probability matrix and serves as a basis for subsequent random walk simulation of abnormal feature propagation.

[0139] The propagation process of abnormal feature information is simulated using the optimized random walk weight matrix, so as to predict the evolution trend of the device abnormal state.

[0140] In an implementation, based on the optimized random walk weight matrix obtained above, the random walk method is used for random simulation again to simulate the dynamic propagation process of abnormal feature information in the device network, and the specific implementation process is as follows:

[0141] First, the propagation starting node of the abnormal feature is determined according to the initial distribution state of the weak abnormal feature .

[0142] Second, according to the optimized random walk weight matrix , the next propagation node of the abnormal feature is determined by random sampling according to the following transition probability in each propagation process: ; wherein, is the transition probability of the abnormal feature information from the current node to the next node .

[0143] After the above method is executed for multiple times (such as 100 times), the frequency of each device node reached by the abnormal feature information is counted, and the propagation trend of the abnormal feature information in a future period of time is predicted according to the frequency, and then the overall evolution trend of the device abnormal state is predicted.

[0144] In another optional implementation, the prediction of the evolution trend of the device abnormal state further includes:

[0145] Based on the random walk algorithm, the stability of the abnormal feature information propagation path is analyzed to obtain the path stability coefficient of each device node.

[0146] In a specific implementation, the stability of the propagation path of the abnormal feature is defined as the frequency of repeated appearance of the same propagation path of the abnormal feature in multiple random simulation processes, representing the stability of the propagation of the abnormal feature along a specific path. The implementation method for calculating the stability coefficient is as follows:

[0147] First, the random walk simulation process is performed, and the number of repeated appearances of the abnormal feature information on each propagation path is counted;

[0148] Second, the path stability coefficient of the node is defined according to the frequency of the appearance of each node on a specific path The specific calculation formula is as follows: In the formula, represents the number of appearances of the node in all path simulation results; represents the total number of all random simulation propagation paths. The higher the path stability coefficient, the stronger the stability of the propagation of the abnormal feature through the node.

[0149] The propagation probability of the abnormal feature information is adjusted according to the path stability coefficient to obtain a corrected abnormal feature propagation probability;

[0150] In a specific implementation, the random walk weight matrix is further adjusted in this step, and the propagation weight corresponding to each node is corrected according to the path stability coefficient obtained above. The specific correction formula is as follows: In the formula, is the corrected abnormal feature propagation probability weight; is the path stability coefficient of the node ; and is a weight correction factor, which is determined through actual experiments.

[0151] After the above correction processing, probability normalization is performed again to ensure that the propagation probability satisfies the basic condition of the probability matrix: In the formula, the denominator is the sum of the corrected weights of the node to all subsequent nodes.

[0152] The random walk simulation process is performed again using the corrected abnormal feature propagation probability to predict the evolution trend of the device abnormal state;

[0153] In a specific implementation, the random walk simulation process is performed again using the corrected random walk weight matrix as the basis, and the specific process is the same as the above implementation method:

[0154] The starting node is selected according to the initial state distribution of the abnormal feature ; and the corrected propagation probability Perform a state transition and simulate a random walk.

[0155] After a sufficient number of simulations, the arrival probability of the abnormal features propagating for each device node is statistically analyzed to obtain the final evolution trend of the abnormal state of the device.

[0156] Compared with the first method, the simulation results after path stability coefficient correction can more accurately reflect the actual anomaly propagation process and are more in line with the physical laws and historical statistical laws of actual equipment anomaly propagation, thereby improving the accuracy and stability of predicting the evolution trend of equipment anomaly status.

[0157] S104: Based on the evolution trend of the abnormal state and the degree of dynamic deviation between the predetermined safety state of the equipment, generate a dynamic risk distribution curve of the equipment to determine the risk level of the current safety state of the prefabricated substation.

[0158] In one optional implementation, determining the risk level of the current safety status of the prefabricated substation includes:

[0159] The initial deviation degree between each equipment node and the predetermined safe state is determined based on the abnormal state evolution trend of the equipment, and the initial deviation index set is obtained.

[0160] In specific implementation, the predetermined safety state in this embodiment is defined as the typical state parameter range when the equipment is running normally, such as the normal reference value range of parameters such as transformer winding temperature, equipment partial discharge frequency, and number of circuit breaker switching operations.

[0161] First, the device abnormal state evolution trend predicted in step S103 is quantified into predicted abnormal state parameter values ​​for each device node, for example:

[0162] The predicted temperature of the transformer windings is 85℃.

[0163] The predicted number of circuit breaker trips is 1500.

[0164] The predicted frequency of partial discharge is 800 Hz.

[0165] Secondly, the difference between the predicted abnormal state parameter values ​​and the predetermined safe state parameter values ​​for each node is calculated and defined as the initial deviation degree. The specific calculation formula is as follows: In the formula: Show node The predicted abnormal state parameter values; This represents the predetermined safety state parameter value (or the upper limit of the normal value) for this node; the initial deviation set is represented as: ,in, This represents the total number of equipment nodes within the transformer substation.

[0166] establish a risk probability distribution model of the device nodes based on the initial deviation index set;

[0167] In a specific implementation, based on the determined initial deviation degree index set , the risk probability of each device node is calculated respectively. Typically, the calculation of the risk probability in this embodiment can adopt an exponential mapping manner, and the specific definition formula is: ; wherein, the risk probability of the device node , the value range is [0, 1]; is the initial deviation degree of the device node ; and is a risk sensitivity coefficient, and the specific value is determined according to historical operation data or experience.

[0168] The above formula means that the greater the device deviation degree, the higher the risk probability; when the deviation degree is 0 (i.e., the device state is completely consistent with the predetermined safe state), the risk probability is 0.

[0169] Through the above manner, the risk probability distribution model of the device nodes is established, and is specifically represented as a risk probability set: .

[0170] In combination with the risk probability distribution of each device node, the device risk dynamic distribution curve of the whole box-type substation is generated;

[0171] In a specific implementation, the obtained risk probability distribution set of the device nodes is used in combination with the topological structure or device logical layout of the box-type substation to generate the whole risk dynamic distribution curve of the box-type substation based on the risk probability of the device nodes.

[0172] Specifically, the horizontal axis of the whole risk dynamic distribution curve of the box-type substation is the device node or node position serial number, and the vertical axis is the risk probability value of the corresponding node; further, the risk probability curve can also be drawn after each device node is sorted according to the actual layout.

[0173] Exemplarily, the risk dynamic distribution curve is represented as:

[0174] Horizontal coordinate: transformer, high-voltage circuit breaker, low-voltage distribution cabinet, relay protection device, etc.

[0175] Vertical coordinate: risk probability value corresponding to each node ;

[0176] After obtaining the risk dynamic distribution curve, the risk probability of each key device node of the current box-type substation can be clearly displayed.

[0177] According to the risk distribution of the equipment risk dynamic distribution curve, the risk level of the current safety state of the box-type substation is determined.

[0178] In a specific implementation, the equipment risk dynamic distribution curve obtained is further analyzed, and the safety risk level of the overall box-type substation is determined through the risk probability distribution characteristics of the equipment nodes. The specific implementation is as follows:

[0179] First, the equipment risk level division standard is defined, for example, the risk probability value of the numerical interval can be used to determine the equipment node risk level, for example:

[0180] When , the node risk level is “safe”;

[0181] When , the node risk level is “general risk”;

[0182] When , the node risk level is “high risk”;

[0183] When , the node risk level is “serious risk”.

[0184] Second, according to the risk level statistics of all equipment nodes in the box-type substation, the comprehensive risk level evaluation rule of the overall box-type substation is determined, and a typical way is as follows:

[0185] If all key equipment nodes are in the “safe” state, the overall risk level of the box-type substation is determined to be “safe”;

[0186] If a small number (for example, no more than 20%) of nodes are in the “general risk”, and the other nodes are safe, the overall risk level is determined to be “general risk”;

[0187] If there is a node risk level of “high risk”, or more than 20% of the nodes reach the “general risk” level, the overall risk level is determined to be “high risk”;

[0188] If any node risk level reaches “serious risk”, the overall box-type substation risk level is determined to be “serious risk”.

[0189] Through the above way, the safety risk level of the current box-type substation can be accurately determined, so that subsequent risk control measures can be taken in time to ensure the safe operation of the box-type substation.

[0190] Embodiment 2

[0191] For example Figure 2As shown, the embodiment does not detail part as shown in embodiment 1, the embodiment discloses a safety state monitoring system for intelligent box-type substation, comprising:

[0192] The path construction module 201 is used for establishing a fault propagation network model according to the topological relationship of the box-type substation equipment, and identifying potential propagation paths in the fault propagation network model by using a dynamic Bayesian probability inference method, thereby constructing a set of equipment fault propagation paths.

[0193] The feature mining module 202 is used for sparse coding reconstruction of the equipment multi-source monitoring signals based on a sparse self-encoding network, so as to mine weak abnormal feature information in the monitoring signals hidden by background noise.

[0194] The trend prediction module 203 is used for inputting the weak abnormal feature information into a state evolution model of the corresponding equipment, simulating the propagation process of the weak abnormal feature information along the set of equipment fault propagation paths by using a random walk algorithm, and predicting the evolution trend of the equipment abnormal state.

[0195] The risk analysis module 204 is used for generating an equipment risk dynamic distribution curve according to the dynamic deviation degree between the evolution trend of the abnormal state and the predetermined safety state of the equipment, and determining the risk level of the current safety state of the box-type substation.

[0196] The above formulas are all dimensionless numerical calculations, the formulas are obtained by software simulation of a large amount of data to obtain a formula of the nearest real situation, and the preset parameters, weights and threshold values in the formula are set by the person skilled in the art according to the actual situation.

[0197] The above only describes certain exemplary embodiments of the present application by way of illustration, and it is needless to say that the described embodiments can be modified in various ways without departing from the spirit and scope of the present application for those skilled in the art. Therefore, the above figures and descriptions are illustrative in nature and should not be understood as limiting the scope of protection of the claims of the present application.

Claims

1. A method for security state monitoring of a smart box-type substation, characterized by, The method comprises the following steps: A fault propagation network model is established according to the topological relationship of the box-type substation equipment; a potential propagation path in the fault propagation network model is identified by using a dynamic Bayesian probability inference method, and a device fault propagation path set is constructed; A sparse auto-encoding network is used to perform sparse coding reconstruction on the device multi-source monitoring signals to mine weak abnormal feature information in the monitoring signals that is covered by background noise; The sparse coding reconstruction on the device multi-source monitoring signals comprises the following steps: A multi-dimensional time series data set of the device operating state is constructed from the multi-source monitoring signals, and an initial feature dimension and a feature space are determined; According to the distribution characteristics of the multi-dimensional time series data, the sparse regularization parameters of the sparse auto-encoding network are determined, and the initial parameter set of the network is established accordingly; A method combining forward propagation and backward iterative optimization is used to perform sparse compression coding processing on the multi-dimensional time series data by using the initial parameter set of the network, and a preliminary sparse feature code is obtained; According to the residual distribution between the preliminary sparse feature code and the original monitoring signal, an effective residual interval is extracted and feature enhancement reconstruction is performed to mine the weak abnormal feature information covered by the background noise; The weak abnormal feature information is input into the state evolution model of the corresponding device; a random walk algorithm is used to simulate the propagation process of the weak abnormal feature information along the device fault propagation path set, and the evolution trend of the device abnormal state is predicted; The prediction of the evolution trend of the device abnormal state comprises the following steps: Based on the device fault propagation path set, the initial propagation weight of the device node in each fault path is calculated, and an initial random walk weight matrix is established; According to the initial state distribution of the weak abnormal feature information, a path sampling process of random walk is performed to generate a state transition sample set of the abnormal feature information; The initial random walk weight matrix is optimized and adjusted according to the state transition sample set to obtain an optimized random walk weight matrix; The propagation process of the abnormal feature information is simulated by using the optimized random walk weight matrix, so as to predict the evolution trend of the device abnormal state; According to the dynamic deviation degree between the evolution trend of the abnormal state and the predetermined safety state of the device, a device risk dynamic distribution curve is generated, and the risk level of the current safety state of the box-type substation is determined.

2. The method for safety state monitoring of a smart cubicle substation according to claim 1, characterized in that, The construction of the device fault propagation path set comprises the following steps: Based on the topological connection relationship of the device nodes in the fault propagation network model, the initial state transition structure between the device nodes is determined, an initial Bayesian network model is established, and an initial state transition probability matrix of the device nodes is obtained; A probability likelihood function of state transition is constructed based on the historical fault data of the device, and the conditional expectation of the state transition probability is derived to obtain an initial estimate of the state transition probability; The initial estimate value is used as the input of the algorithm, and the expectation-maximization algorithm is used to perform multiple iteration calculations on the state transition probability matrix, and the convergence of the likelihood function is determined to obtain an optimized state transition probability matrix; According to the optimized state transition probability matrix, a Monte Carlo simulation process is performed to screen out fault propagation paths that meet a set probability threshold, and a device fault propagation path set is obtained.

3. The method for safety state monitoring of a smart cubicle substation according to claim 2, characterized in that, The multiple iteration calculation of the state transition probability matrix by using the maximum likelihood algorithm, and the convergence of the likelihood function, comprises: According to the initial estimate of the state transition probability, the alternating iteration process of the expectation step and the maximization step is performed, and the numerical value of the device state transition probability is updated; The likelihood function value is recalculated after each iteration, and the adjacent two calculation results are compared to determine whether the likelihood function meets the convergence condition; When the likelihood function meets the set convergence condition, the iteration process is stopped, and the optimized state transition probability matrix is obtained.

4. The method for safety state monitoring of a smart cubicle substation according to claim 3, characterized in that, The sparse compression encoding processing of the multi-dimensional time series data by using the network initial parameter set, comprises: The network initial parameter set is input into the sparse auto-encoding network, and the preliminary encoding features of the multi-dimensional time series data are calculated through forward propagation operation; Based on the difference between the preliminary encoding features and the original monitoring signal, a sparse regularization loss function is established and its loss gradient value is calculated; According to the loss gradient value, the network parameters are updated, the reverse iteration optimization is performed, and the preliminary sparse feature encoding of the multi-dimensional time series data is obtained.

5. The method for safety state monitoring of a smart cubicle substation according to claim 4, characterized in that, The optimization adjustment of the initial random walk weight matrix, comprises: The state transition sample set is subjected to cluster analysis to obtain the clustering results of different state transition modes; The dominant path of state transition between device nodes is determined according to the clustering results, and the corresponding elements in the random walk weight matrix are adjusted according to the dominant path; The adjusted random walk weight matrix is subjected to probability normalization processing to obtain the optimized random walk weight matrix.

6. The method for safety state monitoring of a smart cubicle substation according to claim 5, wherein, The evolution trend of the device abnormal state is predicted, which further comprises: Based on the random walk algorithm, the stability of the abnormal feature information propagation path is analyzed to obtain the path stability coefficient of each device node; The propagation probability weight of the abnormal feature information is adjusted according to the path stability coefficient to obtain the corrected abnormal feature propagation probability; The random walk simulation process is re-executed by using the corrected abnormal feature propagation probability to predict the evolution trend of the device abnormal state.

7. The method for safety state monitoring of a smart cubicle substation according to claim 6, characterized in that, The risk level of the current safety state of the box-type substation is determined, which comprises: According to the evolution trend of the device abnormal state, the initial deviation degree between each device node and the predetermined safety state is determined to obtain an initial deviation index set; The risk probability distribution model of the device node is established by using the initial deviation index set; The device risk dynamic distribution curve of the box-type substation as a whole is generated by combining the risk probability distribution of each device node; According to the risk distribution of the device risk dynamic distribution curve, the risk level of the current safety state of the box-type substation is determined.

8. A safety state monitoring system for a smart box-type substation, realized based on the safety state monitoring method for a smart box-type substation according to any one of claims 1-7, characterized in that, Comprise: The path construction module is used for establishing a fault propagation network model according to the device topology relationship of the box-type substation, and identifying the potential propagation path in the fault propagation network model by using a dynamic Bayesian probability reasoning method to construct a device fault propagation path set; The feature mining module is used for sparse coding reconstruction of the device multi-source monitoring signal based on a sparse auto-encoding network to mine the weak abnormal feature information in the monitoring signal which is covered by background noise; a trend prediction module configured to input the weak abnormal feature information into a state evolution model of the corresponding device, simulate a propagation process of the weak abnormal feature information along a set of device fault propagation paths by using a random walk algorithm, and predict an evolution trend of an abnormal state of the device; a risk analysis module configured to generate a dynamic distribution curve of a risk of the device according to a dynamic deviation degree between the evolution trend of the abnormal state and a predetermined safe state of the device, and determine a risk level of a current safe state of the box-type substation.

Images