A substation misoperation mode mining method for large-scale historical data

Abstract

The application discloses a substation misoperation mode mining method for large-scale historical data, relates to the technical field of substation misoperation analysis and mode mining, and comprises the following steps: a time sequence dynamic data graph of a substation device is established, equipment state evaluation is performed in combination with equipment time sequence data and health state modeling; cross-modal feature alignment and dynamic fusion are performed, and deep representation of multi-source data is extracted; mode recognition and risk assessment are performed by using causal reasoning and explainability analysis. According to the method, the time sequence dynamic data graph and the health state of the equipment are constructed, the health condition of the equipment can be monitored in real time, cross-modal feature alignment and dynamic fusion are performed, deep representation of multi-source data is extracted, different types of data can be comprehensively utilized, the comprehensive understanding of the equipment state is improved, the causal relationship between various factors can be accurately identified by combining causal reasoning and explainability analysis, and in-depth mode recognition and risk assessment can be performed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of substation misoperation analysis and pattern mining, and particularly relates to a substation misoperation pattern mining method for large-scale historical data. BACKGROUND

[0002] At present, the misoperation analysis of substations still mainly depends on the experience judgment and manual analysis of operation and maintenance personnel, and such a traditional analysis method excessively depends on personal experience, lacks systematicness and replicability, and thus is unable to meet the increasing demand for large-scale data analysis, especially in a complex substation operation environment, and it is difficult to accurately grasp the complex correlation between various influencing factors by relying on manual experience, and the overall assessment of system risks is also limited.

[0003] In terms of data analysis, the prior art often excessively focuses on single-dimensional data detection, such as only focusing on the abnormality of equipment operation parameters, and such one-sided analysis ignores the mutual influence of environmental factors, human operation and other multiple dimensions in the operation process of substations, resulting in insufficient in-depth analysis of the correlation between equipment groups, and thus it is difficult to timely discover and prevent complex faults caused by the interaction of multiple devices, in addition, the existing method also rarely considers the aging phenomenon of devices over time, lacks comprehensive health assessment of the life cycle of devices, and affects the scientificity of preventive maintenance strategies and equipment updating and reconstruction decisions.

[0004] More importantly, the current analysis method often regards each misoperation as an independent event, ignoring the internal relationship that may exist between different misoperations, and such isolated analysis perspective makes it difficult for operation and maintenance personnel to discover the correlation patterns between misoperations and unable to effectively predict and prevent possible chain faults. In the coordinated operation of large substation groups, such a problem is particularly evident, which may lead to the evolution of local problems into systemic risks, therefore, how to accurately identify and predict potential misoperation risks from massive historical operation data has become a technical challenge that needs to be solved, and it is necessary to establish an effective data processing mechanism and deeply understand the operation rules behind the data. SUMMARY

[0005] In view of the above problems, the present application is proposed.

[0006] Therefore, the technical problem solved by the present application is that the existing substation misoperation analysis method excessively depends on the experience judgment of operation and maintenance personnel, lacks systematic analysis of multi-dimensional data, is insufficient in the assessment of the life cycle and change trend of devices, and how to comprehensively consider the correlation between devices to effectively identify and predict potential misoperation risks.

[0007] To solve the above technical problems, the application provides the following technical scheme: a substation misoperation mode mining method for large-scale historical data, comprising establishing a time sequence dynamic data graph of substation equipment, combining time sequence data and health state modeling of the equipment to perform equipment state evaluation; performing cross-modal feature alignment and dynamic fusion to extract deep representation of multi-source data; and using causal reasoning and explainability analysis to perform mode recognition and risk assessment.

[0008] As a preferred scheme of the substation misoperation mode mining method for large-scale historical data, the establishment of the time sequence dynamic data graph of the substation equipment comprises collecting substation equipment operation data, establishing a substation equipment operation data matrix, and being expressed as:

[0009] X∈R N×D

[0010] wherein X is the substation equipment operation data matrix, R is a real number set, N is the total number of equipment in the substation, and D is the feature dimension of the collected operation data of each equipment; based on the substation equipment operation data matrix, a substation equipment correlation matrix is established, and the Pearson correlation coefficient between the equipment is calculated, and is expressed as:

[0011] P[i,j]=cov(X[i,:],X[j,:]) / (stdX[i,:]*X[j,:])

[0012] wherein P is the substation equipment correlation matrix, P[i,j] is the Pearson correlation coefficient between equipment i and equipment j, cov(X[i,:],x[j,:]) is the covariance between the feature vectors of equipment i and equipment j, and stdX[i,:]*X[j,:] is the standard deviation of the feature vectors of equipment i and equipment j; an adaptive threshold is used to determine the associated equipment, and is expressed as:

[0013] θ=mean(P)+α*std(P)

[0014] wherein θ is the adaptive threshold, mean(P) is the mean of the substation equipment correlation matrix P, α is an adjustable parameter, and std(P) is the standard deviation of the substation equipment correlation matrix P; when the correlation coefficient P[i,j] is greater than the threshold θ, equipment i and equipment j are marked as associated equipment, and an adjacency matrix A is generated, a spatial heterogeneous graph network structure of the substation equipment is established, including an equipment layer, a function layer and a system layer, and is expressed as:

[0015] A=[A1[i,j],A2[i,j],A3[i,j]]

[0016] Wherein, A1 is the device layer adjacency matrix, A2 is the function layer adjacency matrix, A3 is the system layer adjacency matrix, A1[i,j] is the physical connection association of device i and device j in the device layer, A2[i,j] is the association of device i and device j in the function layer, and A3[i,j] is the interaction association of device i and device j in the system layer; A1, A2 and A3 are spliced in the feature dimension, nonlinearly transformed through the full connection layer FC, and the meta-path weight matrix is obtained through the softmax function, the meta-path weight matrix is fused with the substation device association matrix, and the fused substation device association matrix is established, which is represented as:

[0017] P' = Q1A1 + Q2A2 + Q3A3

[0018] Wherein, P' is the fused substation device association matrix, Q1 is the device layer weight, Q2 is the function layer weight, and Q3 is the system layer weight; the cross-layer attention mechanism is introduced, the features of nodes in different layers are extracted and fused through the multi-layer perception, the node attention coefficients β of different layers are calculated according to the softmax function ij , the updated node is calculated by weighted summation, and is represented as:

[0019]

[0020] Wherein, is the feature of the updated node i, h j is the feature vector of node j, β ij is the attention coefficient from node i to node j; the time window sequence is constructed, the window size is set to 12 hours, the periodic change of the device state is captured, for any two time steps t and s, the attention score is calculated, and is represented as:

[0021] score(t,s) = v T *tanh(W q *X t +W k *X s )

[0022] Wherein, score is the attention score, W q , W k are learnable parameter matrices, the dimensions of W q , W k are d x D, d is 128, v is a d-dimensional attention vector, tanh is the hyperbolic tangent activation function, T is the transpose operation of the attention vector v, X t is the device state feature vector at time step t, and X s is the device state feature vector at time step s; the attention score is normalized to obtain the attention weight γ through the softmax functionts , is expressed as:

[0023] γ ts = softmax(score(t,s))

[0024] The encoder-decoder structure based on graph convolution adopts an encoder of a multi-layer graph convolution network for normalizing an adjacent matrix, and a decoder reconstructs original operation data features through linear transformation and an activation function, and is expressed as:

[0025] X' = LeakyReLU(H^(L)W d )

[0026] wherein X' is a reconstructed substation equipment operation data matrix output by the decoder, LeakyReLU is a nonlinear activation function, H^(L) is the output of the last layer of the encoder, W d is a weight matrix of the decoder; the parameters of the encoder are optimized, and a reconstruction loss is calculated, and is expressed as:

[0027]

[0028] wherein L rec is the reconstruction loss, and F is the Frobenius norm.

[0029] As a preferred scheme of the substation misoperation mode mining method for large-scale historical data, wherein: the device state evaluation combined with the time series data and health state modeling comprises collecting multi-dimensional time series operation feature data of the substation equipment, including equipment operation parameters, environmental data, equipment aging indicators and historical operation records, establishing a device time series operation feature sequence, and is expressed as:

[0030] S ∈ R^(T' × D')

[0031] wherein S is the device time series operation feature sequence, T' is the time span of the historical data, and D' is the feature dimension of the multi-dimensional time series operation feature data collected for each device; the time series operation features of the device are processed through a Transformer encoder, a sine-cosine position encoding method is adopted to capture the position information in the device time series operation feature sequence, and is expressed as:

[0032] PE(pos,2u) = sin(pos / 10000^(2u / o))

[0033] PE(pos,2u+1) = cos(pos / 10000^(2u / o))

[0034] wherein PE(pos, 2u) is an encoding value of an even dimension at position pos, PE(pos, 2u+1) is an encoding value of an odd dimension at position pos, pos is a position index in the time-series running feature sequence, pos takes a value of [0, T'-1], o is a hidden layer dimension set to 256, u is a dimension index, u takes a value of [0, d / 2-1]; through the multi-head self-attention mechanism, long-range dependencies in the time-series running feature sequence are captured, and an attention head is calculated, denoted as:

[0035] head x =Attention(QW x Q ,KWK x K ,VW x V

[0036] MultiHead=Concat(head1,...,head n )W O

[0037] wherein head x is an output of the xth attention head, Q, K, and V are a query matrix, a key matrix, and a value matrix respectively, W x Q , W x K , and W x V are parameter matrices of the xth attention head, n is the number of attention heads set to 8, and W O is an output projection matrix; an index correlation matrix C is constructed, denoted as:

[0038] C∈R^(M×M)

[0039] C[i,j]=MultiHeadAttention(x i ,x j )

[0040] wherein C is the index correlation matrix, M is the total number of health indicators, x i and x j are feature vectors of the ith and jth health indicators respectively, and MultiHeadAttention is the multi-head attention mechanism; a graph attention network is introduced for deep feature extraction, denoted as:

[0041] δ ij =softmax(LeakyReLU(v T [Wh i ||Wh j ]))​

[0042]

[0043] where h i , h j are the feature representations of node i and node j, W is the feature transformation weight matrix; v' is the attention vector, simoid is the activation function, δ ij is the attention weight of node i to node j.

[0044] As a preferred scheme of the substation misoperation mode mining method for large-scale historical data, the cross-modal feature alignment includes establishing a contrast learning model, selecting sample pairs using a contrast learning framework, and selecting sample pairs based on device features. When selecting a positive sample pair, different modal data of the same substation device are paired. When selecting a negative sample pair, the same modal data of different devices are selected as negative samples. Standardized temperature scaling is used, and the contrast learning model is optimized according to a cosine similarity function to calculate a contrast loss function, which is represented as:

[0045] L con =-log(exp(sim(v i ,v j ) / τ) / sum(exp(sim(v i ,v k ) / τ)))

[0046] where L con is the contrast loss function, sim is the cosine similarity function, v i and v j are the i-th and j-th feature vectors in the positive sample pair, sum is the normalization function, τ is the temperature parameter set to 0.07, and v k is the negative sample feature vector. The variational lower bound is optimized, and a variational lower bound of mutual information is constructed, which is represented as:

[0047] I(A;B)≥E[f ε (a,b)]-log(E[exp(f ε (a,b'))])

[0048] where I(A;B) is the mutual information of modal A and modal B, A and B are data of two different modalities, f ε is a neural network mapping function determined by the parameter ε, a is a sample in modal A, b is a sample in modal B, b' is a sample randomly sampled from the edge distribution of modal B, E[f ε (a,b)] is the expected score of the positive sample pair, and log(E[exp(f ε (a,b'))]) is the negative sample pair.where I (a, b) is the influence of the negative sample pair (a, b); set the maximization of the sum of mutual information between all modal pairs as the optimization objective, denoted as:

[0049] I' = max ε sum(I(A i ;B i ))

[0050] where I' is the optimization objective, (A i ;B i ) is the i-th pair of modal data.

[0051] As a preferred scheme of the substation misoperation mode mining method for large-scale historical data, wherein: the dynamic fusion includes integrating feature information from different modalities through dynamic weight distribution and cross-modal attention; the dynamic weight distribution includes generating a modality importance score using a multilayer perceptron by evaluating the quality, confidence and reliability of the modal data, denoted as:

[0052] s i = MLP (Concat [quality i , confidence i , reliability i ])

[0053] where s i is the modality importance score, quality i is the data quality score of the i-th modality, with a value range of [0, 1], confidence i is the confidence score of the i-th modality, with a value range of [0, 1], reliability i is the reliability score of the i-th modality, with a value range of [0, 1], and MLP is a multilayer perceptron with two hidden layers, with hidden layer dimensions of 64 and 32, respectively; the normalized weight of the modality importance score is calculated by a softmax function, denoted as:

[0054] w i = softmax (s i )

[0055] where w i is the normalized weight of the i-th modality; based on the normalized weight, the weighted fused feature is calculated, denoted as:

[0056] z = sum (w i *h' i )

[0057] where z is the weighted fused feature representation, and h' iis the feature representation vector of the i-th modality; the cross-modal attention includes cross-modal information interaction, and attention coefficients between modalities are calculated, denoted as:

[0058]

[0059] wherein, is the attention coefficient from the modality i to the modality j, W1 and W2 are trainable transformation matrices, h′ j is the feature representation vector of the j-th modality; based on the attention coefficients between modalities, the feature representation of cross-modal interaction is calculated, denoted as:

[0060]

[0061] wherein, z i is the updated representation of the i-th modality incorporating information of other modalities; the updated representations of each modality are adaptively fused, denoted as:

[0062] h′ fused =FC(Concat[z i ,...,z m ])

[0063] wherein, h′ fused is the final fused feature representation, FC is a full connection layer, and [z i ,...,z m ] is the updated feature representation of m modalities.

[0064] As a preferred scheme of the substation misoperation mode mining method for large-scale historical data, wherein: the utilization of causal reasoning includes constructing a causal graph by using a structural causal model, and quantifying the causal effect of influencing factors based on counterfactual reasoning; the causal graph is constructed, and the key variables and causal relationships are established, denoted as:

[0065] G=(V',E')

[0066] wherein, G is a causal graph, V' is a set of key variable nodes in a substation system, and E' is a set of directed edges between nodes; based on the Bayesian information criterion, the causal graph is evaluated, and a score function is calculated, denoted as:

[0067] Score(G)=BIG(G)=logL(G)-g / 2*loge

[0068] wherein, Score(G) is a score function, BIC is a Bayesian information criterion, L(G) is a data likelihood function under the causal graph structure G, g is the total number of model parameters, and e is the number of training samples; by maximizing the score function, an optimal causal structure is generated, denoted as:

[0069] G * = argmax G Score(G)

[0070] Where G * is the optimal causal structure; based on counterfactual reasoning, the average treatment effect is calculated, denoted as:

[0071] ATE = E[Y|do(A' = 1)] - E[Y|do(A' = 0)]

[0072] Where ATE is the average treatment effect, Y is the misoperation risk indicator, A' is the intervention variable to be evaluated, do(A' = a') is the intervention operation of artificially setting variable A' to value a', E[Y|do(A' = a')] is the expectation of misoperation risk indicator Y after artificially setting variable A' to value a'; the backdoor criterion is used for causal inference to accurately estimate the causal effect from the observed data, denoted as:

[0073] P'(Y|do(A' = a')) = sum(P(Y|A' = a', C' = c')P(C' = c'))

[0074] Where P'(Y|do(A' = a')) is the conditional probability distribution of misoperation risk indicator Y after artificially setting variable A' to value a', C' is the set of adjustment variables that satisfy the backdoor criterion; based on causal inference, the robustness test mechanism includes evaluating the potential impact of unobserved confounding factors on causal effect estimation through sensitivity analysis, using a double robust estimation method to model the treatment assignment mechanism and the outcome generation mechanism, and introducing confidence interval estimation to characterize the uncertainty degree of causal effect estimation.

[0075] As a preferred scheme of the substation misoperation mode mining method for large-scale historical data, wherein: the interpretability analysis includes constructing a three-layer attention visualization framework for key components in the spatial heterogeneous graph network structure of the substation equipment, at the node level, the importance score of the hidden representation of each node is scored through a multilayer perceptron, denoted as:

[0076] μ l = softmax(MLP(p l ))

[0077] Where μ l is the attention score of node l, p l is the hidden state vector of node l; at the edge level, the attention mechanism is jointly represented based on the source node and the target node, denoted as:

[0078] μ l′ = softmax(MLP([ps ; p t ])

[0079] where μ l′ is the importance score of an edge, p s , p t are the latent representations of the source and target nodes connected by the edge respectively; at the path level, the overall importance is evaluated by aggregating the representations of all nodes in the path, denoted as:

[0080] μ path = softmax(MLP(sum(pl for l in path)))

[0081] where μ path is the importance score of a path, path is a complete path in the graph, sum is the summation of the latent representations of all nodes in the path; the complex data patterns are transformed into actionable operational guidance through decision rules extraction, approximated by CART decision tree, the training objective function of the decision tree is denoted as:

[0082] loss = MSE + λ * tree complexity

[0083] where loss is the training objective function of the decision tree, MSE is the mean squared error of the prediction, tree complexity is the complexity of the decision tree, λ is the balancing factor; based on the trained decision tree, a rule extraction method is designed to determine the importance threshold, denoted as:

[0084] rules = extract rules (tree, threshold = 0.1)

[0085] where rules is the output decision rule, extract rulesTo extract rules, tree is a trained decision tree, and threshold = 0.1 is an importance threshold set to 1; starting from the root node, traverse the decision tree, and extract decision rules of nodes with split gain greater than the importance threshold threshold, including device load rules, temperature anomaly rules, operation sequence rules, environmental condition rules, and device state combination rules; the device load rule includes that if the transformer load rate exceeds 85% and the duration exceeds 30 minutes, and the ambient temperature is higher than 35℃, there is a device overload risk, and load transfer needs to be performed immediately; if the breaker switch frequency exceeds 2 times the normal value within 1 hour, and the operation interval is less than 15 minutes, there is a misoperation risk, and manual confirmation is needed; the temperature anomaly rule includes that if the transformer oil temperature rising rate exceeds 2℃ / h, and the ambient temperature is stable in the past 3 hours, there may be an internal fault of the transformer, and a warning needs to be started; when the bus connection point temperature is higher than the surrounding temperature by 30℃ or more, and the load current does not reach 80% of the rated value, there is a poor contact risk; the operation sequence rule includes that if multiple switch devices of the same loop are continuously operated within 10 minutes, and the standard confirmation process is not performed, the operation has a logical error risk; when the maintenance operation is not performed according to the five-prevention steps, and the weather condition on site is poor, the operation needs to be suspended and re-evaluated; the environmental condition rule includes that when the relative humidity exceeds 85% and the insulation monitoring value is lower than 90% of the standard threshold, there is an insulation damage risk, and monitoring needs to be strengthened; if lightning weather occurs and the lightning protection equipment detection result is abnormal, an emergency lightning protection plan needs to be started immediately; the device state combination rule includes that when the state indicators of multiple associated devices simultaneously occur abnormally, and the time interval of abnormal occurrence is less than 5 minutes, there may be a systematic fault risk; if the monitoring parameters of the main transformer and the related circuit breaker simultaneously fluctuate in a short time, and the fluctuation amplitude exceeds 50% of the normal value, a linkage protection mechanism needs to be started.

[0086] Another object of the present application is to provide a substation misoperation mode mining system for large-scale historical data, which can solve the problem of lack of systematic analysis of multi-dimensional data in the current substation misoperation analysis technology by establishing a time sequence dynamic data graph of substation equipment, combining device time sequence data and health state modeling for device state evaluation.

[0087] As a preferred scheme of the substation misoperation mode mining system for large-scale historical data, wherein: a dynamic graph network construction and feature learning module, a multi-modal deep representation learning module, a causal reasoning and explainability analysis module are included; the dynamic graph network construction and feature learning module is used to establish a time series dynamic data graph of the substation equipment, and the equipment state is evaluated in combination with the time series data and health state modeling of the equipment; the multi-modal deep representation learning module is used to perform cross-modal feature alignment and dynamic fusion, and extract deep representations of multi-source data; the causal reasoning and explainability analysis module is used to perform mode recognition and risk assessment by using causal reasoning and explainability analysis.

[0088] A computer device comprises a memory and a processor, the memory stores a computer program, and the processor executes the computer program to implement the steps of the substation misoperation mode mining method for large-scale historical data.

[0089] A computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of the substation misoperation mode mining method for large-scale historical data.

[0090] The beneficial effects of the present application: the substation misoperation mode mining method for large-scale historical data provided by the present application can realize accurate evaluation of the equipment running state by constructing a time series dynamic data graph of the equipment, in combination with the time series data and health state of the equipment, effectively capture the changes of the equipment and the environment, thereby real-time monitoring the health status of the equipment, reducing the risk of misoperation caused by equipment failure, through cross-modal feature alignment and dynamic fusion, extracting deep representations of multi-source data, which can comprehensively utilize different types of data to improve the overall understanding of the equipment state, through effective integration of multi-dimensional information, the recognition ability of the model to the complex substation environment running condition is enhanced, which helps to more accurately predict potential misoperation, optimize operation and maintenance decision and resource allocation, combined with causal reasoning and explainability analysis, the causal relationship between various factors can be accurately identified, thereby deep mode recognition and risk assessment can be performed, which can reveal the key factors affecting system stability, and then realize effective risk prevention, and the decision-making level of operation and maintenance personnel is improved by producing transparent decision basis, the present application achieves better effects in efficiency, adaptability and reliability. BRIEF DESCRIPTION OF DRAWINGS

[0091] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0092] Figure 1 A flow chart of a substation misoperation mode mining method for large-scale historical data is provided for the first embodiment of the present application.

[0093] Figure 2 A module schematic diagram of a substation misoperation mode mining system for large-scale historical data is provided for the third embodiment of the present application. DETAILED DESCRIPTION

[0094] In order to make the above objectives, characteristics and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should fall within the scope of protection of the present application.

[0095] Embodiment 1, with reference to Figure 1 For an embodiment of the present application, a substation misoperation mode mining method for large-scale historical data is provided, comprising:

[0096] S1: Establishing a time sequence dynamic data graph of substation equipment, combining device time sequence data and health state modeling to perform device state evaluation.

[0097] Further, establishing a time sequence dynamic data graph of substation equipment includes collecting substation equipment operation data, establishing a substation equipment operation data matrix, and being expressed as:

[0098] X∈R N×D

[0099] Wherein, X is a substation equipment operation data matrix, R is a real number set, N is the total number of devices in the substation, D is the feature dimension of the collected operation data of each device, including voltage, current, temperature and other key operation parameters; based on the substation equipment operation data matrix, a substation equipment correlation matrix is established, and the Pearson correlation coefficient between devices is calculated, and is expressed as:

[0100] P[i,j]=cov(X[i,:],X[j,:]) / (stdX[i,:]*X[j,:])

[0101] Wherein, P is a substation equipment correlation matrix, P[i,j] is the Pearson correlation coefficient between device i and device j, cov(X[i,:],X[j,:]) is the covariance between the feature vectors of device i and device j, and stdX[i,:]*X[j,:] is the standard deviation of the feature vectors of device i and device j; through an adaptive threshold, the associated devices are determined, and are expressed as:

[0102] θ = mean(P) + a * std(P)

[0103] wherein θ is an adaptive threshold, mean(P) is the mean value of the substation equipment association matrix P, a is an adjustable parameter, the value range of a is recommended to be between [0.8, 1.2], which is used to control the sparsity degree of the association edge, std(P) is the standard deviation of the substation equipment association matrix P; when the correlation coefficient P[i,j] is greater than the threshold θ, the device i and the device j are marked as associated devices, and an adjacency matrix A is generated to establish a substation equipment spatial heterogeneous graph network structure, including a device layer, a function layer and a system layer, which is represented as:

[0104] A = [A1[i,j], A2[i,j], A3[i,j]]

[0105] wherein A1 is a device layer adjacency matrix, A2 is a function layer adjacency matrix, and A3 is a system layer adjacency matrix, A1[i,j] is the physical connection association of device i and device j in the device layer, A2[i,j] is the association of device i and device j in the function layer, and A3[i,j] is the interaction association of device i and device j in the system layer; in the substation equipment spatial heterogeneous graph network structure, each layer describes the association characteristics in different dimensions respectively, the device layer is mainly constructed based on the physical connection relationship, the actual electrical connection mode between devices is analyzed to form the bottom layer network topology, in this layer, the main devices such as transformers, circuit breakers and disconnectors are nodes in the graph, and the physical connection between them constitutes the relationship of the edge; the function layer focuses on expressing the business function association between devices, reflecting the cooperative working relationship of devices in the operation process of the substation, for example, the devices belonging to the same protection loop have close functional association, even if they are not adjacent in the physical space; the system layer describes the system-level interaction relationship between device groups from a higher level, including the coupling effect between multiple subsystems such as electrical systems and secondary systems.

[0106] A1, A2 and A3 are spliced in the feature dimension, nonlinear transformation is performed through a fully connected layer FC, and an element path weight matrix is obtained through a softmax function, the element path weight matrix is fused with the substation equipment association matrix to establish a fused substation equipment association matrix, which is represented as:

[0107] P' = Q1A1 + Q2A2 + Q3A3

[0108] wherein P' is the fused substation equipment association matrix, Q1 is the device layer weight, Q2 is the function layer weight, and Q3 is the system layer weight; this adaptive weighting method can dynamically adjust the importance of information in different levels according to specific scenarios.

[0109] The cross-layer attention mechanism is introduced, the nodes of different layers are extracted and fused by the multi-layer perception, and the node attention coefficient β of different layers is calculated according to the softmax function ij The updated node is calculated by weighted summation, which is represented as:

[0110]

[0111] wherein, is the feature of the updated node i, h j is the feature vector of node j, β ij is the attention coefficient from node i to node j; the cross-layer attention mechanism is introduced, which integrates information from different layers. In this way, each node can adaptively obtain the most relevant information in other layers according to the attention mechanism, so as to form a more comprehensive and accurate node representation.

[0112] This multi-layer heterogeneous graph structure design can not only comprehensively depict the multi-dimensional correlation between devices in the substation, but also realize the effective fusion and interaction of information at different levels through the graph conversion network and the cross-layer attention mechanism, which provides a solid foundation for subsequent misoperation mode identification and risk prediction.

[0113] The time window sequence is constructed, the window size is set to 12 hours, and the periodic change of the device state is captured. For any two time steps t and s, the attention score is calculated, which is represented as:

[0114] score(t,s)=v T *tanh(W q *X t +W k *X s )

[0115] wherein, score is the attention score, W q , W k are learnable parameter matrices, the dimensions of W q , W k are d×D, d is 128, v is a d-dimensional attention vector, tanh is the hyperbolic tangent activation function, T is the transpose operation of the attention vector v, X t is the device state feature vector at time step t, X s is the device state feature vector at time step s; the attention score is normalized by the softmax function to obtain the attention weight γ ts , which is represented as:

[0116] γ ts =softmax(score(t,s))

[0117] The encoder-decoder structure based on graph convolution adopts an encoder of a multi-layer graph convolution network, the propagation rule of each layer of which considers the degree information of nodes for normalizing an adjacency matrix, and the decoder reconstructs the original operation data features through linear transformation and an activation function, and is expressed as:

[0118] X' = LeakyReLU(H^(L)W d )

[0119] where X' is a reconstructed substation equipment operation data matrix output by the decoder, LeakyReLU is a nonlinear activation function to avoid the gradient vanishing problem, H^(L) is the output of the last layer of the encoder, and W d is a weight matrix of the decoder; the parameters of the encoder are optimized, and the reconstruction loss is calculated and is expressed as:

[0120]

[0121] where L rec is the reconstruction loss, and F is the Frobenius norm; the L rec loss function measures the Frobenius norm difference between the reconstructed feature matrix and the original feature matrix, and by minimizing the loss function, a node representation that both preserves the topological structure information and has good reconstruction capability can be obtained.

[0122] It should be noted that the device state evaluation in combination with the time series data and health state modeling of the equipment includes collecting multi-dimensional time series operation feature data of the substation equipment, including key operation parameters (voltage, current, temperature, etc.), environmental data (humidity, air pressure, etc.), equipment aging indicators (such as cumulative operation time, maintenance frequency), and historical operation records (such as the frequency of misoperation), establishing a device time series operation feature sequence, and is expressed as:

[0123] S∈R^(T′×D')

[0124] where s is the device time series operation feature sequence, T' is the time span of the historical data, and D' is the feature dimension of the multi-dimensional time series operation feature data collected for each device; the time series operation features of the equipment are processed by a Transformer encoder, a sine-cosine position encoding method is adopted to capture the position information in the device time series operation feature sequence, and is expressed as:

[0125] PE(pos,2u) = sin(pos / 10000^(2u / o))

[0126] PE(pos,2u+1) = cos(pos / 10000^(2u / o))

[0127] wherein PE(pos, 2u) is the encoding value of the even dimension at the position pos, PE(pos, 2u+1) is the encoding value of the odd dimension at the position pos, pos is the position index in the time-series running feature sequence, pos takes values from [0, T'-1], o is the hidden layer dimension set to 256, u is the dimension index, u takes values from [0, d / 2-1]; the position encoding can give each position in the sequence a unique encoding, so that the model can perceive the relative position relationship of the elements.

[0128] By using the multi-head self-attention mechanism, the long-range dependence relationship in the time-series running feature sequence is captured, and the attention head is calculated, represented as:

[0129] head x =Attention(QW x ,KW Q ,VW x K ) x V )

[0130] MultiHead=Concat(head1,...,head n )W O

[0131] wherein head x is the output of the xth attention head, Q, K, and V are the query matrix, the key matrix, and the value matrix, respectively, W x Q , W x K , and W x V are the parameter matrices of the xth attention head, used to project the input to a lower-dimensional feature space, n is the number of attention heads set to 8, and W O is the output projection matrix, used to integrate the outputs of the multi-head attention into a unified representation.

[0132] The index correlation matrix C is constructed and represented as:

[0133] C∈R^(M×M)

[0134] C[i,j]=MultiHeadAttention(x i ,x j )

[0135] wherein C is the index correlation matrix, M is the total number of health indicators, x i , x jare the feature vectors of the ith and jth health indicators, respectively, and MultiHeadAttention is a multi-head attention mechanism used to calculate the strength of the association between indicators.

[0136] To further enhance the model's ability to model the relationship between health indicators, a graph attention network is introduced for deep feature extraction, represented as:

[0137] δ ij = softmax(LeakyReLU(v T [Wh i ||Wh j ])

[0138]

[0139] where h i and h j are the feature representations of nodes i and j, W is the feature transformation weight matrix, v' is the attention vector, sigmoid is the activation function, and δ ij is the attention weight of node i to node j.

[0140] It should also be noted that the time series dynamic data graph of the substation equipment mainly serves to perform basic processing on the operation data of the substation equipment. It first calculates the correlation between devices and establishes the connection relationship between devices, and then considers the changes in the time dimension to form a dynamic graph structure. This step is like building a basic framework that records "who is associated with whom" and "how this association changes over time."

[0141] The substation equipment spatial heterogeneous graph network structure representation describes the relationship between devices from three different angles based on the time series graph. The device layer focuses on physical connections, the function layer focuses on business collaboration, and the system layer focuses on higher-level interactions. This is like adding more rich hierarchical information to the basic framework, helping operation and maintenance personnel to better understand the relationship between devices.

[0142] Device health state modeling is based on the time series dynamic data graph of the substation equipment and the spatial heterogeneous graph of the substation equipment to evaluate the operating condition of each device. It uses time series features and spatial relationships to analyze device operating parameters and assess their health level. This is like conducting a physical examination on each device to obtain their "health report."

[0143] S2: Perform cross-modal feature alignment and dynamic fusion to extract deep representations of multi-source data.

[0144] Further, the cross-modal feature alignment includes establishing a contrast learning model, selecting sample pairs in a contrast learning framework, and selecting sample pairs based on device features. When selecting a positive sample pair, different modal data of the same power substation device are paired. The modal data includes device operating parameters, environmental data, operation records, etc. For example, the current data of a circuit breaker is paired with its operation timing record, or the temperature data of a transformer is paired with its load data. This pairing method ensures that there is an inherent correlation between different modal data.

[0145] When selecting a negative sample pair, the same modal data of different devices is selected as a negative sample. The data distribution of different devices should be different even in the same modal, for example, two circuit breakers at different locations. Even if they are both current data, due to their different roles in the power grid and operating environments, the data features should also show differences.

[0146] Standardized temperature scaling is used, and the contrast learning model is optimized according to the cosine similarity function to calculate the contrast loss function, which is represented as:

[0147] L con =-log(exp(sim(v i ,v j ) / τ) / sum(exp(sim(v i ,v k ) / τ)))

[0148] Wherein, L con is the contrast loss function, sim is the cosine similarity function, v i and v j are the i-th and j-th feature vectors in the positive sample pair, sum is the normalization function, τ is the temperature parameter set to 0.07, and the selection of the temperature parameter has an important influence on the training effect of the model. A smaller τ value can make the model more sensitive to feature differences, which helps to learn more discriminative feature representations, and v k is the negative sample feature vector; optimize the variational lower bound, and construct the variational lower bound of mutual information, which is represented as:

[0149] I(A;B)≥E[f ε (a,b)]-log(E[exp(f ε (a,b'))])

[0150] Wherein, I(A;B) is the mutual information of modal A and modal B, A and B are data of two different modalities, f ε is a neural network mapping function determined by the parameter ε, a is a sample in modal A, b is a sample in modal B, b' is a sample randomly sampled from the edge distribution of modal B, and E[f ε[a,b)] represents the expected score of the positive sample pair, log(E[exp(f ε (a,b'))]) represents the influence of negative sample pairs; the optimization objective is to maximize the sum of mutual information among all modality pairs, expressed as:

[0151] I' = max ε sum(I(A i B i ))

[0152] Where I′ is the optimization objective, (A i B i ) represents the i-th pair of modal data.

[0153] This optimization objective ensures the maximization of information sharing among different modes while maintaining the unique information structure of each mode. By iteratively optimizing this objective function, the model can learn better feature representations, thereby improving the accuracy of identifying substation malfunction modes.

[0154] In practical applications, this invention uses mini-batch stochastic gradient descent to optimize the model. To improve training efficiency, a hard sample mining strategy is adopted, prioritizing the selection of easily confused sample pairs for training. At the same time, a regularization term is introduced to prevent model overfitting and ensure that the learned feature representation has good generalization ability. This multimodal feature learning method based on contrastive learning and mutual information maximization provides strong technical support for the accurate identification of substation malfunction modes.

[0155] It should be noted that dynamic fusion includes integrating feature information from different modalities through dynamic weight allocation and cross-modal attention; dynamic weight allocation includes generating modality importance scores using a multilayer perceptron by evaluating modality data quality, confidence, and reliability, as expressed as:

[0156] s i =MLP(Concat[quality i ,confidence i reliability i ])

[0157] Among them, s i For modal importance scores, quality i The data quality score for the i-th modality, with a value in the range [0,1], confidence i The confidence score for the i-th modality, with a value in the range [0,1], represents reliability. iThe reliability score of the i-th modality, with a value range of [0, 1], is calculated by a two-layer hidden layer multi-layer perceptron (MLP) with hidden layer dimensions of 64 and 32, respectively. The data quality reflects the integrity and accuracy of the original data, the confidence indicates the certainty of the model to the prediction result of the modality, and the reliability measures the stability performance of the modality in historical analysis.

[0158] The normalized weight of the modality importance score is calculated by a softmax function, denoted as:

[0159] w i =softmax(s i )

[0160] where w i is the normalized weight of the i-th modality. Based on the normalized weight, the weighted fused feature is calculated, denoted as:

[0161] z=sum(w i *h′ i )

[0162] where z is the weighted fused feature representation, and h′ i is the feature representation vector of the i-th modality.

[0163] The cross-modal attention includes cross-modal information interaction, focusing on the interaction between different modalities, and calculating the attention coefficient between modalities, denoted as:

[0164]

[0165] where a is the attention coefficient from modality i to modality j, W1 and W2 are trainable transformation matrices for feature projection, and h′ j is the feature representation vector of the j-th modality. Based on the attention coefficient between modalities, the cross-modal interaction feature representation is calculated, denoted as:

[0166]

[0167] where z i is the updated representation of the i-th modality incorporating information from other modalities; the updated modality representations are adaptively fused, denoted as:

[0168] h′ fused =FC(Concat[z i ,...,z m ])

[0169] where h′ fused is the final fused feature representation, FC is a fully connected layer, and [z i..., z m ] are the updated feature representations of m modalities.

[0170] This multi-level feature fusion scheme not only considers the importance differences of different modalities, but also fully models the mutual influence between modalities, which can effectively improve the accuracy and robustness of substation misoperation mode recognition. This scheme has good adaptability in practical application, especially when dealing with multi-source heterogeneous data.

[0171] It should also be noted that cross-modal feature alignment is mainly responsible for handling the matching problem between different types of data. Just like translation work, it converts data from different sources (such as device operating parameters, environmental data, operation records, etc.) into a "common language" that can be compared and understood with each other. The specific approach is to use contrastive learning to make relevant data closer and irrelevant data farther apart, while ensuring that the information shared between different types of data is fully preserved through mutual information maximization.

[0172] The dynamic fusion mechanism is responsible for intelligently combining different types of data that have been aligned. It is like a smart mixer that dynamically decides how to optimally mix the information by evaluating the importance, reliability, and relevance of each type of data in the current situation. It not only considers the individual importance of each type of data, but also considers the mutual influence between data, ultimately obtaining a comprehensive and more complete data representation.

[0173] S3: Use causal reasoning and explainability analysis for pattern recognition and risk assessment.

[0174] Furthermore, using causal reasoning includes constructing a causal graph using a structural causal model and quantifying the causal effect of influencing factors based on counterfactual reasoning; constructing a causal graph, establishing key variables and causal relationships, and representing them as:

[0175] G = (V', E')

[0176] Where G is the causal graph, V' is the set of key variable nodes in the substation system, including device state indicators, environmental parameters, and operation behavior, and E' is the set of directed edges between nodes, used to describe the causal relationships between variables.

[0177] Based on the Bayesian Information Criterion, the goodness of the causal graph is evaluated, and the score function is calculated, represented as:

[0178] Score(G) = BIC(G) = logL(G) - g / 2 * loge

[0179] wherein, Score(G) is a score function, BIC is a Bayesian information criterion, L(G) is a data likelihood function under the causal graph structure G, used to measure the fitting degree of the model to the observed data, g is the total number of model parameters, including the number of independent parameters in each node conditional probability table, e is the number of training samples, used to balance the relationship between model complexity and data size.

[0180] By maximizing the score function, the optimal causal structure is generated, denoted as:

[0181] G * =argmax G Score(G)

[0182] wherein, G * is the optimal causal structure.

[0183] In the specific implementation process, a heuristic search strategy is used for causal graph structure learning. First, a basic causal graph is initialized based on expert knowledge, and then the graph structure is gradually optimized through operations such as adding edges, deleting edges, and reversing edges. In order to improve search efficiency, an incremental update mechanism based on local score decomposition is introduced, which only needs to calculate the influence of the changed local structure on the total score.

[0184] Based on counterfactual reasoning, the average treatment effect is calculated, denoted as:

[0185] ATE=E[Y|do(A'=1)]-E[Y|do(A'=0)]

[0186] wherein, ATE is the average treatment effect, Y is the misoperation risk indicator, A' is the intervention variable to be evaluated, do(A'=a') is the intervention operation of artificially setting variable A' to value a', E[Y|do(A'=a')] is the expectation of misoperation risk indicator Y after artificially setting variable A' to value a'; the backdoor criterion is used for causal inference to accurately estimate the causal effect from the observed data, denoted as:

[0187] P'(Y|do(A'=a'))=sum(P(Y|A'=a',C'=c')P(C'=c'))

[0188] wherein, P'(Y|do(A'=a')) is the conditional probability distribution of misoperation risk indicator Y after artificially setting variable A' to value a', C' is a set of adjustment variables that satisfy the backdoor criterion, used to control the influence of confounding factors. By taking a weighted average of all possible C' values, an unbiased estimate of the intervention effect is obtained. In the substation scenario, C' usually includes key background variables such as environmental conditions, equipment age, and load level.

[0189] To improve the reliability of causal inference, a robustness checking mechanism is performed, first, the potential impact of unobserved confounding factors on causal effect estimation is evaluated through sensitivity analysis, second, a double robust estimation method is used to model the distribution mechanism and the result generation mechanism simultaneously, even if one of the models is specified incorrectly, a consistent causal effect estimation can still be obtained, finally, a confidence interval estimation is introduced to characterize the uncertainty degree of causal effect estimation.

[0190] Through the above causal analysis framework, the invention can reveal the deep causes of substation misoperation risk, and provide reliable decision support for risk prevention and control, for example, the specific contribution degree of factors such as equipment aging, environmental interference and personnel fatigue to misoperation risk can be identified, so that preventive measures can be developed accordingly, at the same time, this analysis method based on causal inference has strong interpretability, which helps operation and maintenance personnel to understand and accept the analysis results.

[0191] It should be noted that the interpretable analysis includes constructing a three-layer attention visualization framework for key components in the spatial heterogeneous graph network structure of the substation equipment, at the node level, the importance score of the hidden representation of each node is scored by a multilayer perceptron, represented as:

[0192] μ l =softmax(MLP(p l ))

[0193] Where μ l is the attention score of node l, which reflects the importance of different device nodes in misoperation risk assessment, and p l is the hidden state vector of node l.

[0194] At the edge level, the attention mechanism is jointly represented based on the source node and the target node, represented as:

[0195] μ l′ =softmax(MLP([p s ;p t ]))

[0196] Where μ l′ is the importance score of the edge, and p s , p t are the hidden representations of the source node and the target node connected by the edge respectively; this mechanism can effectively identify the key interaction relationship between devices.

[0197] At the path level, the overall importance is evaluated by aggregating the representations of all nodes on the path, represented as:

[0198] μ path =softmax(MLP(sum(p lfor l in path)))

[0199] where μ path is the importance score of path, path is a complete path in the graph, sum is the sum of the implicit representation of all nodes in the path; this hierarchical attention mechanism enables the model to explain the propagation path of operational risk from micro to macro levels.

[0200] Through decision rule extraction, complex data patterns are converted into operational operation guidance, CART decision tree is used for approximation, the training objective function of the decision tree is represented as:

[0201] loss=MSE+λ*tree complexity

[0202] where loss is the training objective function of the decision tree, MSE is the mean square error of the prediction error, tree complexity is the complexity of the decision tree, and λ is the balance factor; based on the trained decision tree, the rule extraction method of the importance threshold is designed, which is represented as:

[0203] rules=extract rules (tree,threshold=0.1)

[0204] where rules is the output decision rule, extract rules is the extracted rule, tree is the trained decision tree, and threshold=0.1 is the importance threshold set to 1; starting from the root node, the decision tree is traversed, and the decision rules of the nodes with split gain greater than the importance threshold threshold are extracted, each rule contains a condition part (describing the value range of device state, operating environment and other characteristics) and a conclusion part (corresponding to the risk assessment result), these rules are presented in the form of "if... then...", which is convenient for operation and maintenance personnel to understand and execute.

[0205] The decision rules include device load rules, temperature anomaly rules, operation sequence rules, environmental condition rules, and device state combination rules; the device load rules include, if the transformer load rate exceeds 85% and the duration exceeds 30 minutes, and the environmental temperature is higher than 35℃, there is a risk of device overload, and load transfer needs to be performed immediately; if the breaker switch frequency exceeds 2 times the normal value within 1 hour, and the operation interval is less than 15 minutes, there is a risk of misoperation, and manual confirmation is needed; the temperature anomaly rules include, if the transformer oil temperature rising rate exceeds 2℃ / hour, and the environmental temperature is stable in the past 3 hours, there may be an internal fault of the transformer, and a warning needs to be started; when the bus connection point temperature is higher than the surrounding temperature by 30℃ or more, and the load current does not reach 80% of the rated value, there is a risk of poor contact; the operation sequence rules include, if multiple switch devices of the same loop are operated continuously within 10 minutes, and the standard confirmation process is not performed, there is a risk of logical error in operation; when the maintenance operation is not performed according to the five-prevention steps, and the weather conditions on site are not good, the operation needs to be suspended and re-evaluated; the environmental condition rules include, when the relative humidity exceeds 85%, and the insulation monitoring value is lower than 90% of the standard threshold, there is a risk of insulation damage, and monitoring needs to be strengthened; if lightning weather occurs and the lightning protection equipment detection result is abnormal, the lightning protection emergency plan needs to be started immediately; the device state combination rules include, when the state indicators of multiple associated devices simultaneously occur abnormally, and the time interval of abnormal occurrence is less than 5 minutes, there may be a risk of systematic fault; if the monitoring parameters of the main transformer and the related circuit breaker fluctuate simultaneously within a short time, and the fluctuation amplitude exceeds 50% of the normal value, the linkage protection mechanism needs to be started.

[0206] It should also be noted that the core task of the causal reasoning mechanism is to find the reason chain behind the substation misoperation, which is like a professional accident investigator. By constructing a causal graph, it analyzes the causal relationship between various factors (such as device state, environmental condition, personnel operation, etc.). It not only finds out "what caused the problem", but also quantitatively evaluates the influence degree of each factor. Through counterfactual reasoning, it can answer the question "what if a certain condition changes, what will the result be", thereby providing specific guidance direction for preventing misoperation.

[0207] The main task of explainable analysis is to transform complex analysis results into a form that people can easily understand. It is like an excellent translator, which transforms abstract model decision-making process into intuitive graphical display through hierarchical attention visualization, so that operation and maintenance personnel can clearly see the risk source and propagation path. At the same time, it can also transform complex data patterns into simple and clear decision rules, so that operation and maintenance personnel know "what to do in what situation".

[0208] Embodiment 2, as an embodiment of the present application, provides a substation misoperation mode mining method for large-scale historical data. In order to verify the beneficial effects of the present application, economic benefit calculation and simulation experiments are carried out for scientific demonstration.

[0209] In order to reflect the application effect of the method in different scenarios, experiments are carried out in the scenarios of ultra-high voltage substations, regional power grids and 15 intelligent substations.

[0210] Experiment 1 analyzes the ultra-high voltage substation, selects the historical data from January 2015 to December 2023 as the experimental data set, involves 1247 devices, and the collection frequency is 100ms to ensure that the real-time dynamic state information of the substation device can be obtained. On this basis, the time sequence dynamic data graph of the substation device is established, and the device operation data matrix is constructed to ensure that more realistic model training effect can be obtained in actual operation.

[0211] In the data preprocessing stage, data cleaning technology is adopted to remove noise and incomplete data to improve the data quality of subsequent analysis. The correlation coefficient between devices is calculated to form a device correlation matrix. An adaptive threshold algorithm is applied to screen out important devices and form a device spatial heterogeneous graph network structure. This structure is crucial for subsequent multi-modal data feature extraction and cross-layer attention mechanism information fusion.

[0212] In the model training stage, graph neural network is used for feature learning. By optimizing the model parameters, the model performance is improved. During the training process, the attention mechanism is introduced to better capture the long-range dependence between nodes and enhance the model recognition effect. In addition, the importance sampling strategy is adopted to preferentially train those easily confused sample pairs to improve the generalization ability and robustness of the model.

[0213] In the later evaluation stage, the trained model is used to identify and warn the misoperation risk of substation devices. By comparing and verifying the model output results, it is ensured that the system can effectively identify potential risks and give early warning. The results of Experiment 1 show that the risk identification accuracy of the system reaches 95.8%, and the average early warning time is 4.3 hours, and the maximum early warning time is 12.5 hours, ensuring that the operation and maintenance personnel can take effective measures before potential risks occur.

[0214] Experiment 2 covers the regional power grid level, specifically covering 236 substations with voltage levels ranging from 500 kV to 35 kV, involving 27 different types of equipment, and deploying a structure including 472 edge computing nodes, 12 regional coordination centers, and 1 central control center to ensure efficient data processing and coordination. Experiment 2 results show that the system successfully identifies 25 typical misoperation modes and 8 new modes, providing important references for improving operational levels, and the system has a fast response speed, with an average of only 127 milliseconds required to give analysis results, demonstrating efficient processing capabilities.

[0215] Experiment 3 focuses on the group optimization control of smart substations, covering 15 substations with a total installed capacity of 12000 MVA, covering 2.8 million users, using a hierarchical control architecture including device layer, station control layer, and group control layer. The device layer is responsible for real-time data acquisition, edge computing processing, and local control execution to ensure fast response, the station control layer performs state assessment, strategy optimization, and coordinated control to ensure the coordinated operation of each substation, and the group control layer focuses on global optimization, resource scheduling, and risk prevention and control to achieve overall optimization, and implements adaptive control and collaborative decision-making strategies, combined with load prediction, dynamic adjustment, and multi-agent reinforcement learning, further enhancing the flexibility and intelligence level of the system. Experiment 3 results show that the system achieves excellent performance with a device availability rate of 99.8%, an energy utilization rate increase of 12.3%, and a maintenance cost reduction of 45.6%, with annual savings of 21.8 million yuan and an investment recovery period of only 1.8 years, demonstrating obvious economic value and social benefits, and effectively improving the operational efficiency and reliability of the power system.

[0216] Table 1 Key parameters and experimental results of Experiment 1

[0217]

[0218] Experiment 1 conducts in-depth analysis on ultra-high voltage substations. Experiment 1 uses powerful hardware equipment, including high-performance CPUs and GPUs, to process massive operation data for 8 years. Through the trained graph neural network model, the system achieves excellent results: the risk identification accuracy rate is as high as 95.8%, and it can give an early warning 4.3 hours in advance, leaving sufficient processing time for operation and maintenance personnel. This is like installing an "intelligent early warning sentinel" on the substation, which can detect and alert potential problems in advance.

[0219] Experiment 2 is expanded to the regional power grid level, covering 236 substations of different voltage levels, and adopts a distributed architecture, which cooperates with edge computing nodes and regional coordination centers to ensure processing efficiency and reduce communication costs. The system successfully identifies 25 typical misoperation modes, including 8 new modes, which provides an important reference for improving operation and maintenance levels. The system responds quickly, with an average response time of only 127 milliseconds, and can operate stably even during peak periods.

[0220] Experiment 3 focuses on the group optimization control of 15 intelligent substations, and through a three-layer control architecture, from real-time monitoring at the device layer to global optimization at the group control layer, it realizes the collaborative operation of the substation group. The innovative technologies such as multi-agent reinforcement learning and distributed consensus algorithm make the system run more intelligently and efficiently, and the implementation effect is obvious: the device availability rate reaches 99.8%, which can save 21.8 million yuan of cost per year, and the investment recovery period is less than two years.

[0221] The three experiments form a complete demonstration from point to surface: from a single ultra-high voltage substation, to a regional power grid, to the optimization control of a group of substations, proving that the method has good adaptability and practical value in different scales and scenarios.

[0222] Example 3, with reference to Figure 2 , provides a substation misoperation mode mining system for large-scale historical data, including a dynamic graph network construction and feature learning module, a multi-modal deep representation learning module, and a causal reasoning and explainability analysis module.

[0223] The dynamic graph network construction and feature learning module is used to establish a time-series dynamic data graph of substation equipment, and to evaluate the equipment state based on the combination of equipment time-series data and health state modeling. The multi-modal deep representation learning module is used for cross-modal feature alignment and dynamic fusion to extract deep representations of multi-source data. The causal reasoning and explainability analysis module is used for pattern recognition and risk assessment using causal reasoning and explainability analysis.

[0224] If the functions are implemented in software, the functions can be stored in or implemented as one or more instructions or code on a computer-readable medium. Computer-readable media include both computer storage media and communication media including any medium that facilitates transfer of a computer program from one place to another. A storage medium can be any available medium that can be accessed by a computer. By way of example, and not limitation, such computer-readable media can comprise RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and that can be accessed by a computer. Also, any connection is properly termed a computer-readable medium. For example, if the software is transmitted from a website, server, or other remote source using a coaxial cable, fiber optic cable, or twisted pair, then the coaxial cable, fiber optic cable, or twisted pair are included in the definition of medium. Disk and disc, as used herein, include compact disc (CD), laser disc, optical disc, digital versatile disc (DVD), and Blu-Ray® disc where disks usually reproduce data magnetically, while discs reproduce data optically with lasers. Combinations of the above should also be included within the scope of computer-readable media.

[0225] In other words, like a human driver of a vehicle, an autonomous vehicle can be programmed to follow traffic laws and to make decisions based on its environment. For example, an autonomous vehicle can be programmed to follow a speed limit, to stop at a stop sign, to yield to a pedestrian, to merge onto a highway, to change lanes, to park, and so on. In some embodiments, an autonomous vehicle can be programmed to follow traffic laws and to make decisions based on its environment using a machine learning algorithm. For example, an autonomous vehicle can be programmed to follow a speed limit, to stop at a stop sign, to yield to a pedestrian, to merge onto a highway, to change lanes, to park, and so on using a machine learning algorithm.

[0226] More specific examples (a non-exhaustive list) of the computer-readable medium include the following: an electrical connection (electronic) having one or more wires, a portable computer diskette (magnetic), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber (optical), and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable medium can also be paper or another suitable medium upon which the program is printed, as the program can be electronically captured, for example, via optical scanning of the paper or other medium, then compiled, interpreted or otherwise processed in a suitable manner, if necessary, and then stored in a computer memory.

[0227] It should be understood that portions of the present application can be implemented with hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented with software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, implementation can be with any or a combination of the following technologies, which are all well known in the art: a discrete logic circuit having logic gates for implementing logic functions upon an application of data signals, an application specific integrated circuit having appropriate combinational logic gates, a programmable gate array (PGA), a field programmable gate array (FPGA), etc. It should be understood that the foregoing embodiments are merely illustrative of the present application and are not to be used to limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, it will be apparent to those skilled in the art that various changes and modifications can be contributed to the present application without departing from the spirit and scope of the present application, and such changes and modifications should be encompassed within the scope of the appended claims.

[0228] It should be understood that the foregoing embodiments are merely illustrative of the present application and are not to be used to limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, it will be apparent to those skilled in the art that various changes and modifications can be contributed to the present application without departing from the spirit and scope of the present application, and such changes and modifications should be encompassed within the scope of the appended claims.

Claims

1. A substation misoperation mode mining method for large-scale historical data, characterized in that, The method comprises the following steps: Establishing a time sequence dynamic data graph of substation equipment, combining time sequence data and health state modeling of the equipment to evaluate the state of the equipment; Performing cross-modal feature alignment and dynamic fusion to extract deep representations of multi-source data; Using causal reasoning and explainability analysis to perform pattern recognition and risk assessment; The establishment of the time sequence dynamic data graph of the substation equipment comprises collecting substation equipment operation data, establishing a substation equipment operation data matrix, and being represented as: wherein, is a matrix of substation equipment operation data, is a set of real numbers, is the total number of equipment in the substation, is the feature dimension of the operation data collected for each equipment; Based on the substation equipment operation data matrix, a substation equipment correlation matrix is established, and the Pearson correlation coefficient between the equipment is calculated, which is represented as: wherein, is a substation equipment correlation matrix, is a device and a device Pearson correlation coefficient between, is a device and a device Covariance between eigenvectors, is a device and a device Standard deviation of eigenvectors; By using an adaptive threshold, the associated equipment is determined, which is represented as: wherein is an adaptive threshold, is a substation equipment association matrix is a mean value, is an adjustable parameter, is a substation equipment association matrix is a standard deviation; When the correlation coefficient is greater than a threshold , the device and the device are associated devices, and an adjacency matrix is generated, a substation device spatial heterogeneous graph network structure is established, including a device layer, a function layer and a system layer, and is represented as: wherein, is a device layer adjacency matrix, is a functional layer adjacency matrix, is a system layer adjacency matrix, is a device and a device a physical connection association at the device layer, is a device and a device an association at the functional layer, is a device and a device an interaction association at the system layer; The , and are spliced in the feature dimension, nonlinearly transformed through a full connection layer FC, and subjected to a softmax function to obtain an element path weight matrix. The element path weight matrix is fused with a substation device association matrix to establish a fused substation device association matrix, which is represented as: wherein, is the fused substation equipment correlation matrix, is the equipment layer weight, is the functional layer weight, is the system layer weight; The cross-layer attention mechanism is introduced, feature extraction and fusion of nodes in different layers are performed through a multi-layer perception, and a node attention coefficient of different layers is calculated according to a softmax function The updated node is calculated through weighted summation, and is represented as: wherein, is a feature of the updated node , is a feature vector of the node , is an attention coefficient from the node to the node ; A sequence of time windows is constructed with a window size of 12 hours to capture the periodic variation of the device state. For any two time steps and , the attention score is calculated and denoted as: wherein, is an attention score, , is a learnable parameter matrix, , has a dimension of , takes a value of 128, is a dimension of the attention vector, is a hyperbolic tangent activation function, is a transpose operation of the attention vector , is a device state feature vector at time step , is a device state feature vector at time step ; The attention score is normalized by a softmax function to obtain an attention weight is expressed as: Based on the graph convolutional encoder-decoder structure, a multi-layer graph convolutional network encoder is used to normalize the adjacency matrix, and the decoder reconstructs the original operation data features through linear transformation and activation function, which is represented as: wherein, is the reconstructed substation equipment operation data matrix output by the decoder, is a non-linear activation function, is the output of the last layer of the encoder, is the weight matrix of the decoder; The parameters of the encoder are optimized, and the reconstruction loss is calculated, which is represented as: wherein, is a reconstruction loss, is the Frobenius norm.

2. The substation misoperation pattern mining method for large-scale historical data according to claim 1, characterized in that: The combination of the time sequence data and the health state modeling of the equipment to evaluate the state of the equipment comprises collecting multi-dimensional time sequence operation feature data of the substation equipment, including equipment operation parameters, environmental data, equipment aging indicators and historical operation records, establishing a device time sequence operation feature sequence, which is represented as: wherein, is a sequence of equipment timing operation characteristics, is a time span of historical data, is a multi-dimensional time sequence operation characteristic data characteristic dimension collected for each equipment The time sequence operation feature of the device is processed by the Transformer encoder, and the sine-cosine position encoding method is used to capture the position information in the device time sequence operation feature sequence, which is represented as: wherein, is an even dimension encoding value at position , is an odd dimension encoding value at position , is a position index in a time series feature sequence, is valued as , is set to 256 for the hidden layer dimension, is a dimension index, is valued as ; Through the multi-head self-attention mechanism, the long-range dependence relationship in the time sequence operation feature sequence is captured, and the attention head is calculated, which is represented as: wherein, is the output of the th attention head, , , are the query matrix, the key matrix and the value matrix, respectively, , , is the parameter matrix of the th attention head, is set to 8, is the output projection matrix; An index correlation matrix is constructed, which is represented as: wherein, is an index association matrix, is the total number of health indicators, , are the eigenvectors of the first and the first health indicators, respectively, is a multi-head attention mechanism; A graph attention network is introduced to extract deep features, which is represented as: wherein, , is a feature representation of a node and a node , is a feature transformation weight matrix; is an attention vector, is an activation function, is an attention weight of a node to a node .

3. The substation misoperation pattern mining method for large-scale historical data according to claim 2, characterized in that: The cross-modal feature alignment comprises establishing a contrast learning model, selecting sample pairs using a contrast learning framework, and selecting sample pairs based on device feature selection strategies; When selecting a positive sample pair, the different modal data of the same substation equipment are paired; Standardized temperature scaling is used, and the contrast learning model is optimized according to the cosine similarity function to calculate the contrast loss function, which is represented as: in, To compare loss functions, The cosine similarity function is used. , These are the first positive samples in the positive sample pair. The and the first 1 eigenvector For normalization function, Set the temperature parameter to 0.

07. The feature vector of the negative sample; The variational lower bound is optimized, and the mutual information variational lower bound is constructed, which is represented as: wherein, is the mutual information of modalities and modalities , , is the data of two different modalities, is a neural network mapping function determined by parameters , is a sample in modality , is a sample in modality , is a sample randomly sampled from the edge distribution of modality , is the expected score of a positive sample pair, is the influence of a negative sample pair; The sum of the mutual information between all modal pairs is set as the optimization target, which is represented as: wherein, for optimizing the target, for the to modal data.

4. The substation misoperation pattern mining method for large-scale historical data according to claim 3, characterized in that: The dynamic fusion comprises integrating feature information from different modalities through dynamic weight distribution and cross-modal attention; Dynamic weight distribution comprises using a multi-layer perceptron to generate a modal importance score by evaluating the quality, confidence and reliability of the modal data, which is represented as: wherein, is a modal importance score, is a data quality score for the th modality, taking values in the range , is a confidence score for the th modality, taking values in the range , is a reliability score for the th modality, taking values in the range , is a two-layer perceptron with hidden layer dimensions 64 and 32, respectively. The normalized weight of the modal importance score is calculated by using a softmax function, which is represented as: wherein, is the normalized weight for the th modality; Based on the normalized weight, the features after weighted fusion are calculated, which are represented as: in, This is the feature representation after weighted fusion. For the first Feature representation vectors of each modality; Cross-modal attention comprises cross-modal information interaction, and the attention coefficient between modalities is calculated, which is represented as: wherein is an attention coefficient from a modality to a modality , , is a trainable transformation matrix, is a feature representation vector of the th modality; Based on the attention coefficient between modalities, the feature representation of cross-modal interaction is calculated, which is represented as: wherein, is an updated representation of the first modality that incorporates information from the other modalities; is an updated representation of the first modality that incorporates information from the other modalities; The updated modal representations are adaptively fused, which is represented as: wherein, is the final fused feature representation, is a fully connected layer, is is the updated feature representation of the modal.

5. The substation misoperation pattern mining method for large-scale historical data according to claim 4, characterized in that: The utilizing causal inference comprises constructing a causal graph by adopting a structural causal model, and quantifying a causal effect of an influencing factor based on counterfactual reasoning; The causal graph is constructed, the key variables and the causal relationship are established, and are expressed as: wherein, is a causal graph, is a set of key variable nodes in the substation system, is a set of directed edges between nodes; Based on the Bayesian information criterion, the causal graph is evaluated, and a scoring function is calculated, expressed as: in, For the scoring function, For Bayesian information criteria, In the cause-effect graph structure The likelihood function of the data below, This represents the total number of model parameters. This represents the number of training samples; By maximizing the scoring function, the optimal causal structure is generated, expressed as: wherein is the optimal causal structure; Based on counterfactual reasoning, the average treatment effect is calculated, expressed as: wherein, is the average treatment effect, is the risk of a misoperation, is the intervention variable to be evaluated, is the intervention operation that artificially sets the variable to the value , is the expectation of the risk of a misoperation after the intervention operation that artificially sets the variable to the value . The backdoor criterion is adopted for causal inference, and the causal effect is accurately estimated from the observed data, expressed as: wherein, is set to a value is set to a value post-misoperation risk indicator conditional probability distribution, set of tuning variables that satisfy the backdoor criterion; Based on the causal inference, the robustness test mechanism comprises evaluating the potential influence of unobserved confounding factors on the causal effect estimation through sensitivity analysis, simultaneously modeling the treatment allocation mechanism and the result generation mechanism, introducing confidence interval estimation, and characterizing the uncertainty degree of the causal effect estimation.

6. The substation misoperation pattern mining method for large-scale historical data according to claim 5, characterized in that: The explainability analysis comprises constructing a three-layer attention visualization framework for key components in a power substation equipment spatial heterogeneous graph network structure, at the node level, scoring the importance of the hidden representation of each node by a multilayer perceptron, expressed as: wherein, is the attention score for the node , is the hidden state vector for the node . At the edge level, the attention mechanism is jointly represented based on the source node and the target node, expressed as: wherein, is an importance score for the edge, , are implicit representations of the source node and the target node connected by the edge, respectively; At the path level, the overall importance is evaluated by aggregating the representations of all nodes on the path, expressed as: wherein, is the importance score for a path, is a complete path in the graph, is the sum over all nodes on the path of the latent representation. Through decision rule extraction, complex data patterns are converted into operational operation guidance, CART decision tree is adopted for approximation, and the training objective function of the decision tree is expressed as: wherein, is a training objective function for the decision tree, is a mean square sum of prediction errors, is a complexity of the decision tree, is a balancing factor; Based on the trained decision tree, the rule extraction method of the importance threshold is designed, expressed as: wherein, is the output decision rule, is the extraction rule, is the trained decision tree, is the importance threshold set to 1; traversing the decision tree from a root node to extract decision rules for nodes having a split gain greater than an importance threshold including device load rules, temperature anomaly rules, operation sequence rules, environmental condition rules, and device state combination rules; The equipment load rule comprises that if the transformer load rate exceeds 85% and the duration exceeds 30 minutes, and the ambient temperature is higher than 35℃, there is a risk of equipment overload, and load transfer needs to be performed immediately; If the breaker switch frequency exceeds 2 times of the normal value within 1 hour, and the operation interval is less than 15 minutes, there is a risk of misoperation, and manual confirmation is required; The temperature anomaly rule comprises that if the transformer oil temperature rising rate exceeds 2℃ / h, and the ambient temperature is stable in the past 3 hours, there may be an internal fault of the transformer, and the pre-warning needs to be started; When the bus connection point temperature is higher than the surrounding temperature by 30℃ or more, and the load current does not reach 80% of the rated value, there is a risk of poor contact; The operation sequence rule comprises that if multiple switch devices of the same loop are continuously operated within 10 minutes, and the standard confirmation process is not performed, there is a risk of logical error in the operation; When the maintenance operation is not performed according to the five-prevention steps, and the weather condition is not good, the operation needs to be suspended and re-evaluated; The environmental condition rule comprises that when the relative humidity exceeds 85% and the insulation monitoring value is lower than 90% of the standard threshold, there is a risk of insulation damage, and the monitoring needs to be strengthened; If lightning weather occurs and the lightning protection equipment detection result is abnormal, the lightning protection emergency plan needs to be started immediately; The equipment state combination rule comprises that when the state indicators of multiple associated devices simultaneously occur abnormally, and the time interval of the abnormality is less than 5 minutes, there may be a risk of systematic failure; If the monitoring parameters of the main transformer and related circuit breakers fluctuate simultaneously in a short time, and the fluctuation amplitude exceeds 50% of the normal value, the linkage protection mechanism needs to be started.

7. A system employing the substation misoperation pattern mining method for large-scale historical data according to any one of claims 1 to 6, characterized in that: The method comprises a dynamic graph network construction and feature learning module, a multi-modal deep representation learning module, a causal reasoning and explainability analysis module; The dynamic graph network construction and feature learning module is used to establish a time series dynamic data graph of the substation equipment, and to perform equipment state evaluation in combination with equipment time series data and health state modeling; The multi-modal deep representation learning module is used to perform cross-modal feature alignment and dynamic fusion, and to extract deep representations of multi-source data; The causal reasoning and explainability analysis module is used to perform mode recognition and risk assessment by using causal reasoning and explainability analysis.

8. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that, The processor executes the computer program to realize the steps of the substation misoperation mode mining method for large-scale historical data in any one of claims 1 to 6.

9. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the steps of the substation misoperation mode mining method for large-scale historical data in any one of claims 1 to 6.

Images