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

Abstract

The invention discloses a transformer substation misoperation mode mining method for large-scale historical data, and relates to the technical field of transformer substation misoperation analysis and mode mining, and the method comprises the steps: building a time sequence dynamic data graph of transformer substation equipment, and carrying out the equipment state evaluation through combining the time sequence data and health state modeling of the equipment; performing cross-modal feature alignment and dynamic fusion, and extracting depth representation of multi-source data; and performing pattern recognition and risk assessment by using causal reasoning and interpretability analysis. According to the method, the health condition of the equipment can be monitored in real time by constructing the time sequence dynamic data graph and the health state of the equipment, and different types of data can be comprehensively utilized through cross-modal feature alignment and dynamic fusion and extraction of depth representation of multi-source data, so that comprehensive understanding of the equipment state is improved; and causal reasoning and interpretability analysis are combined, so that the causal relationship among various factors can be accurately identified, and further mode identification and risk assessment are carried out.

Description

Technical Field

[0001] The present invention relates to the technical field of substation misoperation analysis and pattern mining, and in particular to a substation misoperation pattern mining method oriented to large-scale historical data. Background Art

[0002] At present, the analysis of substation misoperation still mainly relies on the experience judgment and manual analysis of operation and maintenance personnel. This traditional analysis method relies too much on personal experience and lacks systematicity and replicability. As a result, it seems powerless in the face of the growing demand for large-scale data analysis. Especially in the complex substation operating environment, it is difficult to accurately grasp the complex relationship between various influencing factors relying solely on manual experience, which also limits the comprehensive assessment of system risks.

[0003] In terms of data analysis, existing technologies often focus too much on single-dimensional data detection, such as focusing only on abnormal equipment operating parameters. This one-sided analysis ignores the mutual influence of multiple dimensions such as environmental factors and human operations during substation operation, resulting in insufficient in-depth analysis of the correlation between equipment groups, making it difficult to timely detect and prevent complex faults caused by the interaction of multiple devices. In addition, existing methods rarely consider the aging phenomenon of equipment over time, lack a comprehensive health assessment of the equipment life cycle, and affect the scientific nature of preventive maintenance strategies and equipment renewal and transformation decisions.

[0004] More importantly, current analysis methods often treat each misoperation as an independent event, ignoring the possible inherent connections between different misoperations. This isolated analysis perspective makes it difficult for operation and maintenance personnel to discover the correlation patterns between misoperations, and cannot effectively predict and prevent possible cascading failures. This problem is particularly evident in the coordinated operation of large-scale substation groups, which may cause local problems to evolve 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 urgently, requiring the establishment of an effective data processing mechanism and an in-depth understanding of the operating laws behind the data. Summary of the invention

[0005] In view of the above-mentioned problems, the present invention is proposed.

[0006] Therefore, the technical problem solved by the present invention is that the existing substation misoperation analysis method is too dependent on the experience and judgment of operation and maintenance personnel, lacks systematic analysis of multi-dimensional data, and has insufficient assessment of equipment life cycle and change trends. It also has problems such as how to comprehensively consider the correlation between equipment to effectively identify and predict potential misoperation risks.

[0007] To solve the above technical problems, the present invention provides the following technical solutions: a method for mining substation error operation patterns based on large-scale historical data, including establishing a time-series dynamic data graph of substation equipment, and performing equipment status assessment by combining the equipment's time-series data and health status modeling; performing cross-modal feature alignment and dynamic fusion to extract deep representation of multi-source data; and using causal reasoning and interpretability analysis to perform pattern recognition and risk assessment.

[0008] As a preferred solution of the substation misoperation mode mining method for large-scale historical data of the present invention, wherein: the establishment of the time series dynamic data graph of the substation equipment includes collecting the substation equipment operation data and establishing the substation equipment operation data matrix, which is expressed as:

[0009] X∈R N×D

[0010] Among them, 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 characteristic dimension of the operation data collected by each device. Based on the substation equipment operation data matrix, the substation equipment association matrix is ​​established, and the Pearson correlation coefficient between the equipment is calculated, which is expressed as:

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

[0012] Where P is the substation equipment association 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. The associated equipment is determined by the adaptive threshold, which is expressed as:

[0013] θ=mean(P)+α*std(P)

[0014] Among them, θ is the adaptive threshold, mean(P) is the mean of the substation equipment association matrix P, α is an adjustable parameter, and 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 devices i and j are marked as associated devices, and the adjacency matrix A is generated to establish the network structure of the substation equipment spatial heterogeneous graph, including the equipment layer, function layer and system layer, which is expressed as:

[0015] A=[A 1 [i,j],A 2 [i,j],A 3 [i,j]]

[0016] Among them, A 1is the device layer adjacency matrix, A 2 is the function layer adjacency matrix, A 3 is the system-level adjacency matrix, A 1 [i,j] is the physical connection association between device i and device j at the device layer, A 2 [i,j] is the association between device i and device j at the functional layer, A 3 [i,j] is the interaction between device i and device j at the system level; 1 , A 2 and A 3 The concatenation is performed on the feature dimension, and the nonlinear transformation is performed through the fully connected layer FC. The meta-path weight matrix is ​​obtained through the softmax function. The meta-path weight matrix is ​​fused with the substation equipment association matrix to establish the fused substation equipment association matrix, which is expressed as:

[0017] P'=Q 1 A 1 +Q 2 A 2 +Q 3 A 3

[0018] Among them, P′ is the fused substation equipment association matrix, Q 1 is the device layer weight, Q 2 is the functional layer weight, Q 3 is the system layer weight; the cross-layer attention mechanism is introduced to extract and fuse the features of nodes in different layers through the multi-layer perceptron, and the node attention coefficient β of different layers is calculated according to the softmax function ij , update the node by weighted sum calculation, expressed as:

[0019]

[0020] in, is the updated feature of node i, h j is the feature vector of node j, β ij is the attention coefficient from node i to node j; construct a time window sequence, set the window size to 12 hours, capture the periodic changes of the device state, and calculate the attention score for any two time steps t and s, expressed as:

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

[0022] Among them, score is the attention score, W q , Wk is the learnable parameter matrix, W q , W k The dimension is d×D, d is 128, v is the 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 , expressed as:

[0023] γ ts =softmax(score(t,s))

[0024] The encoder-decoder structure based on graph convolution uses a multi-layer graph convolution network encoder to normalize the adjacency matrix. The decoder reconstructs the original running data features through linear transformation and activation function, which can be expressed as:

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

[0026] Where X′ is the reconstructed substation equipment operation data matrix output by the decoder, LeakyReLU is the nonlinear activation function, H^(L) is the output of the last layer of the encoder, and W d is the weight matrix of the decoder; optimize the parameters of the encoder and calculate the reconstruction loss, which is expressed as:

[0027]

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

[0029] As a preferred solution of the substation misoperation mode mining method for large-scale historical data described in the present invention, the equipment status assessment combined with the equipment's time series data and health status modeling includes collecting multi-dimensional time series operation feature data of substation equipment, including equipment operating parameters, environmental data, equipment aging indicators and historical operation records, and establishing equipment time series operation feature sequences, which are expressed as:

[0030] S∈R^(T′×D')

[0031] Among them, S is the equipment timing operation feature sequence, T′ is the time span of historical data, and D′ is the feature dimension of the multi-dimensional timing operation feature data collected by each device. The timing operation features of the equipment are processed by the Transformer encoder, and the sine-cosine position encoding method is used to capture the position information in the equipment timing operation feature sequence, which is expressed as:

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

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

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

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

[0036] MultiHead=Concat(head 1 ,...,head n )W O

[0037] Among them, head x is the output of the x-th attention head, Q, K, V are the query matrix, key matrix and value matrix respectively, and W x Q , W x K , W x V is the parameter matrix of the xth attention head, n is the number of attention heads set to 8, W O is the output projection matrix; construct the indicator association matrix, expressed as:

[0038] C∈R^(M×M)

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

[0040] Among them, C is the indicator association matrix, M is the total number of health indicators, x i 、x j are the feature vectors of the i-th and j-th health indicators respectively, and MultiHeadAttention is a multi-head attention mechanism. The graph attention network is introduced for deep feature extraction, which is expressed as:

[0041] δ ij =softmax(LeakyReLU(v' T [Wh i ||Wh j ]))

[0042]

[0043] Among them, h i 、h j is the feature representation of node i and node j, W is the feature transformation weight matrix; v' is the attention vector, simgoid is the activation function, δ ij is the attention weight of node i to node j.

[0044] As a preferred solution of the substation misoperation mode mining method for large-scale historical data described in the present invention, the cross-modal feature alignment includes establishing a contrastive learning model, selecting sample pairs using a contrastive learning framework, and selecting a sample pair strategy based on device features. When selecting a positive sample pair, different modal data of the same substation equipment 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 contrastive learning model is optimized according to the cosine similarity function, and the contrast loss function is calculated, which is expressed as:

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

[0046] Among them, L con is the contrast loss function, sim is the cosine similarity function, v i 、v j are the i-th and j-th eigenvectors 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; optimize the variational lower bound and construct the variational lower bound of the mutual information, expressed as:

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

[0048] Among them, I(A; B) is the mutual information between mode A and mode B, A and B are data of two different modes, f ε is the neural network mapping function determined by the parameter ε, a is the sample in modality A, b is the sample in modality B, b′ is the sample randomly sampled from the marginal distribution of modality B, E[f ε (a,b)] is the expected score of the positive sample pair, log(E[exp(f ε (a,b'))]) is the influence of negative sample pairs; the optimization goal is to maximize the sum of mutual information between all modal pairs, which is expressed as:

[0049] I'=max ε sum(I(A i ; B i ))

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

[0051] As a preferred solution of the substation misoperation mode mining method for large-scale historical data described in the present invention, the dynamic fusion includes integrating feature information from different modalities through dynamic weight allocation and cross-modal attention; the dynamic weight allocation includes generating a modal importance score by evaluating the quality, confidence and reliability of modal data using a multi-layer perceptron, which is expressed as:

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

[0053] Among them, s i is the modal importance score, quality i is the data quality score of the ith mode, ranging from [0,1], confidence i is the confidence score of the i-th mode, ranging from [0,1], reliability i is the reliability score of the i-th modality, with a value range of [0,1]. MLP is a multilayer perceptron with two hidden layers, and the hidden layer dimensions are 64 and 32 respectively. The normalized weight of the modality importance score is calculated by the softmax function, which is expressed as:

[0054] w i =softmax(s i )

[0055] Among them, w i is the normalized weight of the i-th mode; based on the normalized weight, the weighted fusion feature is calculated, expressed as:

[0056] z=sum(w i *h′ i )

[0057] Among them, z is the feature representation after weighted fusion, h′ i is the feature representation vector of the i-th modality; cross-modal attention includes cross-modal information interaction and calculation of the attention coefficient between modalities, which is expressed as:

[0058]

[0059] in, is the attention coefficient from modality i to modality j, W 1 , W 2 is the trainable transformation matrix, h′ j is the feature representation vector of the jth modality; based on the attention coefficient between modalities, the feature representation of cross-modal interaction is calculated, which is expressed as:

[0060]

[0061] Among them, z i is the updated representation of the i-th mode that incorporates information from other modes; the updated representations of each mode are adaptively fused and expressed as:

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

[0063] Among them, h′ fused is the final fusion feature representation, FC is the fully connected layer, [z i ,...,z m ] is the updated feature representation of m modalities.

[0064] As a preferred solution of the substation misoperation mode mining method for large-scale historical data described in the present invention, the use of causal reasoning includes constructing a causal graph using a structural causal model, quantifying the causal effects of influencing factors based on counterfactual reasoning; constructing a causal graph, establishing key variables and causal relationships, expressed as:

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

[0066] Where G is the causal graph, V′ is the set of key variable nodes in the substation system, and E′ is the set of directed edges between nodes. Based on the Bayesian information criterion, the causal graph is evaluated and the scoring function is calculated, which is expressed as:

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

[0068] Among them, Score(G) is the scoring function, BIC is the Bayesian information criterion, L(G) is the 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 scoring function, the optimal causal structure is generated, which is expressed as:

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

[0070] Among them, G * is the optimal causal structure; based on counterfactual reasoning, the average treatment effect is calculated and expressed as:

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

[0072] Among them, 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 the variable A′ to the value a′, and E[Y|do(A′=a′)] is the expectation of the misoperation risk indicator Y after artificially setting the variable A′ to the value a′. The backdoor criterion is used for causal inference to accurately estimate the causal effect from the observed data, which is expressed as:

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

[0074] Among them, P′(Y|do(A'=a')) is the conditional probability distribution of the misoperation risk indicator Y after the variable A' is artificially set to the value a', and C′ is the set of adjustment variables that meet the backdoor criterion; based on causal inference, the robustness test mechanism includes evaluating the potential impact of unobserved confounding factors on the causal effect estimate through sensitivity analysis, adopting a double robust estimation method, simultaneously modeling the processing allocation mechanism and the result generation mechanism, introducing confidence interval estimation, and characterizing the degree of uncertainty in the causal effect estimate.

[0075] As a preferred solution of the substation misoperation mode mining method for large-scale historical data described in the present invention, the explainability analysis includes constructing a three-layer attention visualization framework for key components in the spatial heterogeneous graph network structure of substation equipment. At the node level, the implicit representation of each node is scored by a multi-layer perceptron, which is expressed as:

[0076] μ l = softmax(MLP(p l ))

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

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

[0079] Among them, μ l′ is the importance score of the edge, p s 、p t are the 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, which is expressed as:

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

[0081] Among them, μ path is the importance score of the path, path is a complete path in the graph, and sum is the sum of the implicit representations of all nodes on the path; through decision rule extraction, the complex data pattern is transformed into an actionable operation and maintenance guide, and the CART decision tree is used for approximation. The training objective function of the decision tree is expressed as:

[0082] loss=MSE+λ*tree complexity

[0083] Among them, loss is the training objective function of the decision tree, MSE is the mean square sum of prediction errors, and tree complexity is the complexity of the decision tree, and λ is the balance factor. Based on the trained decision tree, a rule extraction method with importance threshold is designed, which is expressed as:

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

[0085] Among them, rules are the output decision rules, extract rules is to extract rules, tree is the trained decision tree, threshold=0.1 is the importance threshold set to 1; starting from the root node, traverse the decision tree and extract the decision rules of the nodes whose splitting gain is greater than the importance threshold threshold, including equipment load rules, temperature anomaly rules, operation sequence rules, environmental condition rules and equipment status combination rules; the equipment load rules include that if the transformer load rate exceeds 85% and lasts for more than 30 minutes, and the ambient temperature is higher than 35℃, there is a risk of equipment overload and the load needs to be transferred immediately; if the circuit breaker switching frequency exceeds twice 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 rules include that if the transformer oil temperature rise rate exceeds 2℃ / hour and the ambient temperature has been stable in the past 3 hours, there may be an internal fault in the transformer and an early warning needs to be initiated; when the busbar connection point temperature is 30 higher than the ambient temperature, ... ℃ or above, and the load current does not reach 80% of the rated value, there is a risk of poor contact; the operation sequence rules include that if multiple switchgear of the same circuit are operated continuously 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 in accordance with 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 that when the relative humidity exceeds 85% and the insulation monitoring value is lower than the standard threshold of 90%, there is a risk of insulation damage and monitoring needs to be strengthened; if thunderstorm weather occurs and the lightning protection equipment detection results are abnormal, the lightning protection emergency plan needs to be activated immediately; the equipment status combination rules include that when the status indicators of multiple related devices are abnormal at the same time and the time interval between the abnormalities is less than 5 minutes, there may be a risk of systemic failure; if the monitoring parameters of the main transformer and related circuit breakers fluctuate at the same time in a short period of time, and the fluctuation amplitude exceeds 50% of the normal value, the linkage protection mechanism needs to be activated.

[0086] Another object of the present invention is to provide a substation misoperation pattern mining system for large-scale historical data, which can establish a time-series dynamic data graph of substation equipment and combine the equipment's time-series data and health status modeling to perform equipment status assessment, thereby solving the problem that current substation misoperation analysis technology lacks systematic analysis of multi-dimensional data.

[0087] As a preferred solution of the substation misoperation pattern mining system for large-scale historical data described in the present invention, it includes: a dynamic graph network construction and feature learning module, a multimodal deep representation learning module, and a causal reasoning and interpretability analysis module; the dynamic graph network construction and feature learning module is used to establish a time-series dynamic data graph of substation equipment, and perform equipment status assessment in combination with the equipment's time-series data and health status modeling; the multimodal deep representation learning module is used to perform cross-modal feature alignment and dynamic fusion, and extract deep representation of multi-source data; the causal reasoning and interpretability analysis module is used to use causal reasoning and interpretability analysis to perform pattern recognition and risk assessment.

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

[0089] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a method for mining substation misoperation patterns oriented to large-scale historical data.

[0090] Beneficial effects of the present invention: The substation misoperation pattern mining method for large-scale historical data provided by the present invention can realize accurate evaluation of the equipment operation status by constructing a time-series dynamic data graph of the equipment, combining the time-series data and health status of the equipment, and effectively capturing changes in the equipment and the environment, thereby monitoring the health status of the equipment in real time and reducing the risk of misoperation caused by equipment failure. Through cross-modal feature alignment and dynamic fusion, deep representation of multi-source data is extracted, and different types of data can be comprehensively utilized to improve the comprehensive understanding of the equipment status. By effectively integrating multi-dimensional information, the model's recognition ability for the operating status of complex substation environments is enhanced, which helps to more accurately predict potential misoperations, optimize operation and maintenance decisions and resource allocation. Combined with causal reasoning and explainability analysis, the causal relationship between various factors can be accurately identified, thereby conducting in-depth pattern recognition and risk assessment, revealing the key factors affecting system stability, and then achieving effective risk prevention and control. By generating transparent decision-making basis to improve the decision-making level of operation and maintenance personnel, the present invention achieves better results in terms of efficiency, adaptability and reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0092] Figure 1 An overall flow chart of a method for mining substation misoperation patterns based on large-scale historical data provided by the first embodiment of the present invention.

[0093] Figure 2 A schematic diagram of modules of a substation misoperation pattern mining system for large-scale historical data provided by a third embodiment of the present invention. DETAILED DESCRIPTION

[0094] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0095] Example 1, reference Figure 1 , which is an embodiment of the present invention, provides a substation misoperation mode mining method for large-scale historical data, including:

[0096] S1: Establish a time series dynamic data diagram of substation equipment, and conduct equipment status assessment by combining the equipment’s time series data and health status modeling.

[0097] Furthermore, establishing a time-series dynamic data graph of substation equipment includes collecting substation equipment operation data and establishing a substation equipment operation data matrix, which is expressed as:

[0098] X∈R N×D

[0099] Among them, 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 characteristic dimension of the operation data collected by each device, including key operation parameters such as voltage, current, and temperature. Based on the substation equipment operation data matrix, the substation equipment association matrix is ​​established, and the Pearson correlation coefficient between the devices is calculated, which is expressed as:

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

[0101] Where P is the substation equipment association 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. The associated equipment is determined by the adaptive threshold, which is expressed as:

[0102] θ=mean(P)+α*std(P)

[0103] Among them, θ is the adaptive threshold, mean(P) is the mean of the substation equipment association matrix P, α is an adjustable parameter, and the value range of α is recommended to be between [0.8, 1.2], which is used to control the sparsity of the association edge, and 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 devices i and j are marked as associated devices, and the adjacency matrix A is generated to establish the network structure of the substation equipment spatial heterogeneous graph, including the equipment layer, function layer and system layer, which is expressed as:

[0104] A=[A 1 [i,j],A 2 [i,j],A 3 [i,j]]

[0105] Among them, A 1 is the device layer adjacency matrix, A 2 is the function layer adjacency matrix, A 3 is the system-level adjacency matrix, A 1 [i,j] is the physical connection association between device i and device j at the device layer, A 2 [i,j] is the association between device i and device j at the functional layer, A 3 [i,j] is the interaction between device i and device j at the system layer. In the network structure of the heterogeneous graph of substation equipment space, each layer describes the association characteristics of different dimensions. The equipment layer is mainly built based on the physical connection relationship. By analyzing the actual electrical connection method between devices, the underlying network topology is formed. In this layer, main devices such as transformers, circuit breakers, and disconnectors are used as nodes in the graph, and the physical connections between them constitute the edge relationship. The functional layer focuses on expressing the business function association between devices, reflecting the collaborative working relationship of devices during the operation of the substation. For example, devices belonging to a certain protection circuit have a close functional association, even if they may not be adjacent in 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 the electrical system and the secondary system.

[0106] A 1 , A 2 and A 3 The concatenation is performed on the feature dimension, and the nonlinear transformation is performed through the fully connected layer FC. The meta-path weight matrix is ​​obtained through the softmax function. The meta-path weight matrix is ​​fused with the substation equipment association matrix to establish the fused substation equipment association matrix, which is expressed as:

[0107] P'=Q 1 A1 +Q 2 A 2 +Q 3 A 3

[0108] Among them, P′ is the fused substation equipment association matrix, Q 1 is the device layer weight, Q 2 is the functional layer weight, Q 3 is the system layer weight; this adaptive weighting method can dynamically adjust the importance of information at different levels according to specific scenarios.

[0109] The cross-layer attention mechanism is introduced to extract and fuse the features of nodes in different layers through the multi-layer perceptron, and the attention coefficient β of nodes in different layers is calculated according to the softmax function. ij , update the node by weighted sum calculation, expressed as:

[0110]

[0111] in, is the updated feature of 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 to integrate information from different levels. In this way, each node can adaptively obtain the most relevant information from other levels according to the attention mechanism, thereby forming a more comprehensive and accurate node representation.

[0112] This multi-layer heterogeneous graph structure design can not only comprehensively characterize the multi-dimensional correlation between equipment in the substation, but also realize the effective fusion and interaction of information at different levels through the graph transformation network and cross-layer attention mechanism, providing a solid foundation for subsequent misoperation pattern recognition and risk prediction.

[0113] Construct a time window sequence, set the window size to 12 hours, capture the periodic changes of the device state, and calculate the attention score for any two time steps t and s, expressed as:

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

[0115] Among them, score is the attention score, W q , W k is the learnable parameter matrix, W q , W kThe dimension is d×D, d is 128, v is the 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 , expressed as:

[0116] γ ts =softmax(score(t,s))

[0117] The encoder-decoder structure based on graph convolution adopts a multi-layer graph convolution network encoder. The propagation rule of each layer takes into account the degree information of the node and is used to normalize the adjacency matrix. The decoder reconstructs the original running data features through linear transformation and activation function, which is expressed as:

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

[0119] Where X′ is the 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 the weight matrix of the decoder; optimize the parameters of the encoder and calculate the reconstruction loss, which is expressed as:

[0120]

[0121] Among them, L rec is the reconstruction loss, F is the Frobenius norm; L rec The loss function measures the Frobenius norm difference between the reconstructed feature matrix and the original feature matrix. By minimizing the loss function, a node representation can be obtained that retains the topological structure information and has good reconstruction ability.

[0122] It should be noted that the equipment status assessment based on the equipment time series data and health status modeling includes collecting multi-dimensional time series operation characteristic data of substation equipment, including key equipment operating parameters (voltage, current, temperature, etc.), environmental data (humidity, air pressure, etc.), equipment aging indicators (such as cumulative operating time, maintenance times) and historical operation records (such as frequency of erroneous operations), and establishing the equipment time series operation characteristic sequence, which can be expressed as:

[0123] S∈R^(T′×D')

[0124] Among them, s is the equipment timing operation feature sequence, T′ is the time span of historical data, and D′ is the feature dimension of the multi-dimensional timing operation feature data collected by each device. The timing operation features of the equipment are processed by the Transformer encoder, and the sine-cosine position encoding method is used to capture the position information in the equipment timing operation feature sequence, which is expressed as:

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

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

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

[0128] Through the multi-head self-attention mechanism, the long-range dependencies in the time series running feature sequence are captured and the attention heads are calculated, which is expressed as:

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

[0130] MultiHead=Concat(head 1 ,...,head n )W O

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

[0132] Construct the indicator association matrix, expressed as:

[0133] C∈R^(M×M)

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

[0135] Among them, C is the indicator association matrix, M is the total number of health indicators, x i 、x j are the feature vectors of the i-th and j-th health indicators respectively. MultiHeadAttention is a multi-head attention mechanism used to calculate the correlation strength between indicators.

[0136] In order to further improve the model's ability to model the relationship between health indicators, the graph attention network is introduced to perform deep feature extraction, which is expressed as:

[0137] δ ij =softmax(LeakyReLU(v' T [Wh i ||Wh j ]))

[0138]

[0139] Among them, h i 、h j is the feature representation of node i and node j, W is the feature transformation weight matrix; v′ is the attention vector, sigmoid is the activation function, 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 substation equipment is mainly used to perform basic processing on the operating data of substation equipment. It first calculates the correlation between devices, 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 related to whom" and "how this relationship changes over time."

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

[0142] Equipment health status modeling evaluates the operating status of each device based on the time series dynamic data graph and spatial heterogeneous graph of substation equipment. It uses time series characteristics and spatial relationships to analyze the operating parameters of the equipment and evaluate its health. This is like conducting a physical examination on each device and obtaining their "health report".

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

[0144] Furthermore, cross-modal feature alignment includes establishing a contrastive learning model, selecting sample pairs using a contrastive learning framework, and selecting sample pairs based on device features. When selecting positive sample pairs, different modal data of the same substation equipment are paired; these modal data include equipment 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 negative sample pairs, select the same modal data of different devices as negative samples; even if the data distribution of different devices is in the same modality, there should be differences. For example, two circuit breakers in different locations, even if they are both current data, should have different data characteristics due to their different roles in the power grid and operating environment.

[0146] Standardized temperature scaling is adopted, and the contrastive learning model is optimized according to the cosine similarity function to calculate the contrastive loss function, which is expressed as:

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

[0148] Among them, L con is the contrast loss function, sim is the cosine similarity function, v i 、v j are the i-th and j-th feature vectors in the positive sample pair, respectively. sum is the normalization function. τ is the temperature parameter set to 0.07. The choice of 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 and help learn more discriminative feature representations. k is the negative sample feature vector; optimize the variational lower bound and construct the variational lower bound of the mutual information, expressed as:

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

[0150] Among them, I(A; B) is the mutual information between mode A and mode B, A and B are data of two different modes, f ε is the neural network mapping function determined by the parameter ε, a is the sample in modality A, b is the sample in modality B, b′ is the sample randomly sampled from the marginal distribution of modality B, E[f ε (a,b)] is the expected score of the positive sample pair, log(E[exp(f ε (a,b'))]) is the influence of negative sample pairs; the optimization goal is to maximize the sum of mutual information between all modal pairs, which is expressed as:

[0151] I'=max ε sum(I(A i ; B i ))

[0152] Among them, I' is the optimization target, (A i ; B i ) is the i-th pair of modal data.

[0153] This optimization goal ensures the maximization of shared information between different modes while maintaining the unique information structure of each mode. By iteratively optimizing this objective function, the model can learn better feature representation, thereby improving the recognition accuracy of substation misoperation modes.

[0154] In practical applications, the present invention uses a small batch stochastic gradient descent method to optimize the model. In order to improve the training efficiency, a difficult sample mining strategy is adopted, and those sample pairs that are easily confused are given priority for training. At the same time, a regularization term is introduced to prevent the model from overfitting, ensuring that the learned feature representation has good generalization ability. This multimodal feature learning method based on contrastive learning and mutual information maximization provides a strong technical support for the accurate identification of substation misoperation 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 modal importance scores by evaluating modal data quality, confidence and reliability using a multi-layer perceptron, expressed as:

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

[0157] Among them, s i is the modal importance score, quality i is the data quality score of the ith mode, ranging from [0,1], confidence i is the confidence score of the i-th mode, ranging from [0,1], reliability i is the reliability score of the ith mode, with a value range of [0,1]. MLP is a multilayer perceptron with two hidden layers, and the hidden layer dimensions are 64 and 32 respectively. Data quality reflects the integrity and accuracy of the original data, confidence indicates the degree of certainty of the model in the prediction results of the mode, and reliability measures the stability of the mode in historical analysis.

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

[0159] w i =softmax(s i )

[0160] Among them, w i is the normalized weight of the i-th mode; based on the normalized weight, the weighted fusion feature is calculated, expressed as:

[0161] z=sum(w i *h′ i )

[0162] Among them, z is the feature representation after weighted fusion, h′ i is the feature representation vector of the i-th mode.

[0163] Cross-modal attention involves cross-modal information interaction, focusing on the interaction between different modalities, and calculating the attention coefficient between modalities, expressed as:

[0164]

[0165] in, is the attention coefficient from modality i to modality j, W 1 , W 2 is a trainable transformation matrix for feature projection, h′ j is the feature representation vector of the jth modality; based on the attention coefficient between modalities, the feature representation of cross-modal interaction is calculated, which is expressed as:

[0166]

[0167] Among them, z i is the updated representation of the i-th mode that incorporates information from other modes; the updated representations of each mode are adaptively fused and expressed as:

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

[0169] Among them, h′ fused is the final fusion feature representation, FC is the fully connected layer, [z i ,...,z m ] is the updated feature representation of m modalities.

[0170] This multi-level feature fusion scheme not only takes into account the differences in importance of different modes, but also fully models the mutual influence between modes. It can effectively improve the accuracy and robustness of substation misoperation pattern recognition. The scheme shows good adaptability in practical applications, especially in processing multi-source heterogeneous data.

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

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

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

[0174] Furthermore, the use of causal reasoning includes constructing a causal graph using a structural causal model to quantify the causal effects of influencing factors based on counterfactual reasoning; constructing a causal graph to establish the key variables and causal relationships, expressed as:

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

[0176] Among them, G is a causal graph, V′ is a set of key variable nodes in the substation system, including equipment status indicators, environmental parameters, operating behaviors, etc., and E′ is a set of directed edges between nodes, which is used to describe the causal relationship between variables.

[0177] Based on the Bayesian information criterion, the quality of the causal diagram is evaluated and the scoring function is calculated, which is expressed as:

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

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

[0180] By maximizing the scoring function, the optimal causal structure is generated, which is expressed as:

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

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

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

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

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

[0186] Among them, 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 the variable A′ to the value a′, and E[Y|do(A′=a′)] is the expectation of the misoperation risk indicator Y after artificially setting the variable A′ to the value a′. The backdoor criterion is used for causal inference to accurately estimate the causal effect from the observed data, which is expressed as:

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

[0188] Among them, P′(Y|do(A'=a')) is the conditional probability distribution of the misoperation risk index Y after the variable A' is artificially set to the value a', and C′ is a set of adjustment variables that meet the backdoor criterion, which is 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] In order to improve the reliability of causal inference, a robustness test mechanism is carried out. First, the potential impact of unobserved confounding factors on the causal effect estimate is evaluated through sensitivity analysis. Secondly, a double robust estimation method is adopted to simultaneously model the treatment allocation mechanism and the result generation mechanism. Even if one of the models is misspecified, a consistent causal effect estimate can still be obtained. Finally, a confidence interval estimate is introduced to characterize the degree of uncertainty in the causal effect estimate.

[0190] Through the above-mentioned causal analysis framework, the present invention can reveal the deep-seated causes of the risk of misoperation in substations and provide reliable decision-making support for risk prevention and control. For example, it can identify the specific contribution of factors such as equipment aging, environmental interference, and personnel fatigue to the risk of misoperation, so as to formulate targeted preventive measures. At the same time, this analysis method based on causal reasoning has strong interpretability, which helps operation and maintenance personnel understand and accept the analysis results.

[0191] It should be noted that the interpretability analysis includes building a three-layer attention visualization framework for the key components in the spatial heterogeneous graph network structure of substation equipment. At the node level, the importance of the implicit representation of each node is scored through a multi-layer perceptron, which is expressed as:

[0192] μ l =softmax(MLP(p l ))

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

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

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

[0196] Among them, μ l′ is the importance score of the edge, p s 、p tare the implicit representations of the source node and target node connected by the edge respectively; this mechanism can effectively identify the key interaction relationships between devices.

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

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

[0199] Among them, μ path is the importance score of the path, path is a complete path in the graph, and sum is the sum of the implicit representations of all nodes on the path. This hierarchical attention mechanism enables the model to explain the propagation path of misoperation risks at multiple levels from micro to macro.

[0200] By extracting decision rules, complex data patterns are transformed into actionable operation and maintenance guidance. The CART decision tree is used for approximation. The training objective function of the decision tree is expressed as:

[0201] loss=MSE+λ*tree complexity

[0202] Among them, loss is the training objective function of the decision tree, MSE is the mean square sum of prediction errors, and tree complexity is the complexity of the decision tree, and λ is the balance factor. Based on the trained decision tree, a rule extraction method with importance threshold is designed, which is expressed as:

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

[0204] Among them, rules are the output decision rules, extract rules To extract rules, tree is the trained decision tree, threshold=0.1 is the importance threshold set to 1; starting from the root node, traverse the decision tree and extract the decision rules of the nodes whose split gain is greater than the importance threshold threshold. Each rule contains a condition part (the value range describing the characteristics of the equipment status, operating environment, etc.) and a conclusion part (the corresponding 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 equipment load rules, temperature anomaly rules, operation sequence rules, environmental condition rules and equipment status combination rules; the equipment load rules include that if the transformer load rate exceeds 85% and lasts for more than 30 minutes, and the ambient temperature is higher than 35°C, there is a risk of equipment overload and immediate load transfer is required; if the circuit breaker switching frequency exceeds twice 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 rules include that if the transformer oil temperature rise rate exceeds 2°C / hour and the ambient temperature has been stable in the past 3 hours, there may be an internal fault in the transformer and an early warning needs to be activated; when the bus connection point temperature is more than 30°C higher than the ambient temperature and the load current does not reach 80% of the rated value, there is a risk of poor contact; the operation sequence rules include that if in If multiple switchgear of the same circuit are operated continuously within 10 minutes without executing the standard confirmation process, there is a risk of logical error in the operation; when the maintenance operation is not performed in accordance with 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 that when the relative humidity exceeds 85% and the insulation monitoring value is lower than the standard threshold of 90%, there is a risk of insulation damage and monitoring needs to be strengthened; if thunderstorm weather occurs and the lightning protection equipment detection results are abnormal, the lightning protection emergency plan needs to be activated immediately; the equipment status combination rules include that when the status indicators of multiple related devices are abnormal at the same time and the time interval between the abnormalities is less than 5 minutes, there may be a risk of systemic failure; if the monitoring parameters of the main transformer and related circuit breakers fluctuate at the same time within a short period of time, and the fluctuation amplitude exceeds 50% of the normal value, the linkage protection mechanism needs to be activated.

[0206] It should also be noted that the core task of the causal reasoning mechanism is to find out the cause chain behind the misoperation of the substation. It is like a professional accident investigator. By constructing a causal diagram, it analyzes the cause-effect relationship between various factors (such as equipment status, environmental conditions, personnel operation, etc.). It not only finds out "what caused the problem", but also quantitatively evaluates the impact of each factor. Through counterfactual reasoning, it can answer questions such as "what will happen if a certain condition changes", thereby providing specific guidance for preventing misoperation.

[0207] The main task of interpretable analysis is to convert complex analysis results into a form that is easy for people to understand. It is like an excellent translator. Through layered attention visualization, it converts the abstract model decision process into an intuitive graphical display, allowing operation and maintenance personnel to clearly see the risk source and transmission path. At the same time, it can also convert complex data patterns into simple and clear decision rules, allowing operation and maintenance personnel to know "what to do under what circumstances".

[0208] Embodiment 2 is an embodiment of the present invention, which provides a method for mining substation misoperation patterns based on large-scale historical data. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0209] In order to demonstrate the application effect of the method of the present invention in different scenarios, experiments were carried out in the scenarios of UHV substation, regional power grid level and 15 smart substations.

[0210] Experiment 1 analyzes the UHV substation and selects historical data from January 2015 to December 2023 as the experimental data set, involving 1,247 devices with a collection frequency of 100ms to ensure that the real-time dynamic status information of the substation equipment can be obtained. On this basis, a time-series dynamic data graph of the substation equipment is established. By constructing the equipment operation data matrix, a more realistic model training effect is ensured in actual operation.

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

[0212] During the model training phase, graph neural networks are used for feature learning, and model performance is improved by optimizing model parameters. During the training process, an attention mechanism is introduced to better capture the long-range dependencies between nodes and enhance the model recognition effect. In addition, an importance sampling strategy is adopted to prioritize training of sample pairs that are easily confused, so as to improve the generalization ability and robustness of the model.

[0213] In the later evaluation stage, the risk of misoperation of substation equipment is identified and warned based on the trained model. The output results of the model are compared and verified to ensure that the system can effectively identify potential risks and issue early warnings. The results of Experiment 1 show that the system's risk identification accuracy rate reaches 95.8%. At the same time, the average early warning time is 4.3 hours and the maximum early warning time is 12.5 hours, ensuring that 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 500kV to 35kV, involving 27 different types of equipment, and the deployment architecture includes 472 edge computing nodes, 12 regional coordination centers and 1 central control center to ensure efficient data processing and coordination. The results of Experiment 2 show that the system successfully identified 25 typical misoperation modes and 8 new modes, providing an important reference for improving the operation and maintenance level. The system has a fast response speed, and it only takes 127 milliseconds on average to give analysis results, showing efficient processing capabilities.

[0215] Experiment 3 focuses on the group optimization control of smart substations, covering 15 substations with a total installed capacity of 12,000MVA and covering 2.8 million users. It adopts a hierarchical control architecture, including the equipment layer, station control layer and group control layer. The equipment layer is responsible for real-time data collection, edge computing processing and local control execution to ensure rapid response. The station control layer performs state evaluation, strategy optimization and coordinated control to ensure the coordinated operation of each substation. The group control layer focuses on global optimization, resource scheduling and risk prevention and control to achieve overall optimization, and simultaneously implements adaptive control and collaborative decision-making strategies. Combined with load prediction, dynamic adjustment and multi-agent reinforcement learning, the flexibility and intelligence level of the system are further enhanced. The results of Experiment 3 show that the system has achieved an excellent performance of 99.8% equipment availability, an increase of 12.3% in energy utilization, and a decrease of 45.6% in maintenance costs. In terms of economic benefits, the annual savings reached 21.8 million yuan, and the investment recovery period was only 1.8 years, showing obvious economic value and social benefits, and effectively improving the operation efficiency and reliability of the power system.

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

[0217]

[0218] Experiment 1 conducted an in-depth analysis of the UHV substation. By using powerful hardware equipment, including high-performance CPUs and GPUs, Experiment 1 processed massive operating data for up to 8 years. Through the trained graph neural network model, the system achieved excellent results: the risk identification accuracy rate was as high as 95.8%, and it was able to issue an early warning 4.3 hours in advance, reserving sufficient processing time for operation and maintenance personnel. This is like installing an "intelligent early warning sentinel" for the substation, which can detect and remind possible problems early.

[0219] Experiment 2 was expanded to the regional power grid level, covering 236 substations of different voltage levels. It adopted a distributed architecture and, through the cooperation of edge computing nodes and regional collaborative centers, ensured processing efficiency and reduced communication costs. The system successfully identified 25 typical misoperation modes, including 8 new modes. These findings provided an important reference for improving the level of operation and maintenance. The system responded very quickly, with an average of only 127 milliseconds to provide analysis results, and it can operate stably even during peak hours.

[0220] Experiment 3 focused on the group optimization control of 15 smart substations. Through a three-layer control architecture, from real-time monitoring at the equipment layer to global optimization at the group control layer, the coordinated operation of the substation group was achieved. The innovative technologies used, such as multi-agent reinforcement learning and distributed consensus algorithms, made the system operation more intelligent and efficient, with obvious implementation effects: the equipment availability rate reached 99.8%, which could save 21.8 million yuan in costs each year, and the investment payback period was less than two years.

[0221] These three experiments formed a complete demonstration from point to surface: from a single UHV substation, to a regional power grid, and then to the optimized control of a substation group, proving that the method of the present invention has good adaptability and practical value in different scales and scenarios.

[0222] Example 3, reference Figure 2 , which is an embodiment of the present invention, provides a substation misoperation pattern mining system for large-scale historical data, including a dynamic graph network construction and feature learning module, a multimodal 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 the time-series dynamic data graph of substation equipment, and to perform equipment status assessment by combining the equipment's time-series data and health status modeling; the multimodal 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 interpretability analysis module is used to use causal reasoning and interpretability analysis to perform pattern recognition and risk assessment.

[0224] If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods of each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0225] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in conjunction with such instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in conjunction with such instruction execution systems, devices or apparatuses.

[0226] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering, or processing in another suitable manner as necessary, and then stored in a computer memory.

[0227] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc. It should be noted that the above embodiments are only used to illustrate the technical solution of the present invention and are not limited. Although the present invention is described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.

[0228] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for mining substation misoperation patterns based on large-scale historical data, characterized in that: include: Establish a time-series dynamic data graph of substation equipment, and conduct equipment status assessment by combining equipment time-series data and health status modeling; Perform cross-modal feature alignment and dynamic fusion to extract deep representation of multi-source data; Use causal reasoning and explainability analysis to perform pattern recognition and risk assessment.

2. The method for mining substation misoperation patterns based on large-scale historical data according to claim 1, characterized in that: The establishment of the time-series dynamic data graph of the substation equipment includes collecting the substation equipment operation data and establishing the substation equipment operation data matrix, which is expressed as: X∈R N×D Among them, 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 characteristic dimension of the operation data collected by each device; Based on the substation equipment operation data matrix, the substation equipment association matrix is ​​established, and the Pearson correlation coefficient between the equipment is calculated, which is expressed as: P[i,j]=cov(X[i,:],X[j,:]) / (stdX[i,:]*X[j,:]) Where P is the substation equipment association 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; The associated devices are determined by adaptive thresholds, which are expressed as: θ=mean(P)+α*std(P) Among them, θ is the adaptive threshold, mean(P) is the mean of the substation equipment association matrix P, α is an adjustable parameter, and 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 devices i and j are marked as associated devices, and the adjacency matrix A is generated to establish the network structure of the substation equipment spatial heterogeneous graph, including the equipment layer, function layer and system layer, which is expressed as: A=[A1[i,j],A2[i,j],A3[i,j]] 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 between device i and device j at the device layer, A2[i,j] is the association between device i and device j at the function layer, and A3[i,j] is the interaction association between device i and device j at the system layer; A1, A2 and A3 are concatenated in the feature dimension, and nonlinear transformation is performed through the fully connected layer FC. The meta-path weight matrix is ​​obtained through the softmax function. The meta-path weight matrix is ​​fused with the substation equipment association matrix to establish the fused substation equipment association matrix, which is expressed as: P'=Q1A1+Q2A2+Q3A3 Among them, P′ is the fused substation equipment association matrix, Q1 is the equipment layer weight, Q2 is the function layer weight, and Q3 is the system layer weight; The cross-layer attention mechanism is introduced to extract and fuse the features of nodes in different layers through the multi-layer perceptron, and the attention coefficient β of nodes in different layers is calculated according to the softmax function. ij , update the node by weighted sum calculation, expressed as: in, is the updated feature of node i, h j is the feature vector of node j, β ij is the attention coefficient from node i to node j; Construct a time window sequence, set the window size to 12 hours, capture the periodic changes of the device state, and calculate the attention score for any two time steps t and s, expressed as: score(t,s)=v T *tanh(W q *X t +W k *X s ) Among them, score is the attention score, W q , W k is the learnable parameter matrix, W q , W k The dimension is d×D, d is 128, v is the 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; Through the softmax function, the attention score is normalized to get the attention weight γ ts , expressed as: γ ts =softmax(score(t,s)) The encoder-decoder structure based on graph convolution uses a multi-layer graph convolution network encoder to normalize the adjacency matrix. The decoder reconstructs the original running data features through linear transformation and activation function, which is expressed as: X'=LeakyReLU(H^(L)W d ) Where X′ is the reconstructed substation equipment operation data matrix output by the decoder, LeakyReLU is the nonlinear activation function, H^(L) is the output of the last layer of the encoder, and W d is the weight matrix of the decoder; Optimize the encoder parameters and calculate the reconstruction loss, expressed as: Among them, L rec is the reconstruction loss and F is the Frobenius norm.

3. The substation misoperation pattern mining method for large-scale historical data according to claim 2 is characterized by: The equipment status assessment based on the equipment time series data and health status modeling includes collecting multi-dimensional time series operation characteristic data of substation equipment, including equipment operation parameters, environmental data, equipment aging indicators and historical operation records, and establishing equipment time series operation characteristic sequences, which are expressed as: S∈R^(T′×D') Among them, S is the equipment time series operation feature sequence, T′ is the time span of historical data, and D′ is the feature dimension of the multi-dimensional time series operation feature data collected by each device; The Transformer encoder is used to process the timing operation characteristics of the device, and the sine-cosine position encoding method is used to capture the position information in the timing operation feature sequence of the device, which is expressed as: PE(pos,2u)=sin(pos / 10000^(2u / o)) PE(pos,2u+1)=cos(pos / 10000^(2u / o)) Among them, PE(po,2u) is the encoding value of the even dimension at position pos, PE(pos,2u+1) is the encoding value of the odd dimension at position pos, pos is the position index in the time series running feature sequence, pos takes the value of [0,T'-1], o is the hidden layer dimension set to 256, u is the dimension index, u takes the value of [0,d / 2-1]; Through the multi-head self-attention mechanism, the long-range dependencies in the time series running feature sequence are captured and the attention heads are calculated, which is expressed as: head x =Attention(QW x Q ,KW x K ,VW x V ) MultiHead=Concat(head1,...,head n )W O Among them, head x is the output of the x-th attention head, Q, K, V are the query matrix, key matrix and value matrix respectively, and W x Q , W x K , W x V is the parameter matrix of the xth attention head, n is the number of attention heads set to 8, W O is the output projection matrix; Construct the indicator association matrix, expressed as: C∈R^(M×M) C[i,j]=MultiHeadAttention(x i ,x j ) Among them, C is the indicator association matrix, M is the total number of health indicators, x i 、x j are the feature vectors of the i-th and j-th health indicators respectively, and MultiHeadAttention is a multi-head attention mechanism; The graph attention network is introduced to perform deep feature extraction, which can be expressed as: δ ij =softmax(LeakyReLU(v' T [Wh i ||Wh j ])) Among them, h i 、h j is the feature representation of node i and node j, W is the feature transformation weight matrix; v' is the attention vector, simgoid is the activation function, δ ij is the attention weight of node i to node j.

4. The method for mining substation misoperation patterns based on large-scale historical data according to claim 3, characterized in that: The cross-modal feature alignment includes establishing a contrastive learning model, selecting sample pairs using a contrastive learning framework, and selecting a sample pair strategy based on device features. When selecting a positive sample pair, different modal data of the same substation equipment are paired; When selecting negative sample pairs, select the same modality data of different devices as negative samples; Standardized temperature scaling is adopted, and the contrastive learning model is optimized according to the cosine similarity function to calculate the contrastive loss function, which is expressed as: L con =-log(exp(sim(v) i ,v j ) / τ) / sum(exp(sim(v i ,v k ) / τ))) Among them, L con is the contrast loss function, sim is the cosine similarity function, v i 、v j are the i-th and j-th eigenvectors 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; Optimize the variational lower bound and construct the variational lower bound of mutual information, expressed as: I(A;B)≥E[f ε (a,b)]-log(E[exp(f ε (a,b'))]) Among them, I(A; B) is the mutual information between mode A and mode B, A and B are data of two different modes, f ε is the neural network mapping function determined by the parameter ε, a is the sample in modality A, b is the sample in modality B, b′ is the sample randomly sampled from the marginal distribution of modality B, E[f ε (a,b)] is the expected score of the positive sample pair, log(E[exp(f ε (a,b'))]) is the impact of negative sample pairs; The optimization objective is to maximize the sum of mutual information between all modal pairs, which can be expressed as: I'=max ε sum(I(A i ;B i )) Among them, I' is the optimization target, (A i ; B i ) is the i-th pair of modal data.

5. The method for mining substation misoperation patterns based on large-scale historical data according to claim 4, characterized in that: The dynamic fusion includes integrating feature information from different modalities through dynamic weight allocation and cross-modal attention; Dynamic weight assignment involves evaluating the quality, confidence, and reliability of modal data using a multilayer perceptron to generate a modal importance score, expressed as: s i =MLP(Concat[quality i ,confidence i ,reliability i ]) Among them, s i is the modal importance score, quality i is the data quality score of the ith mode, ranging from [0,1], confidence i is the confidence score of the i-th mode, ranging from [0,1], reliability i is the reliability score of the ith modality, with a value range of [0,1]. MLP is a multilayer perceptron with two hidden layers, and the hidden layer dimensions are 64 and 32 respectively; The normalized weight of the modality importance score is calculated by the softmax function, expressed as: w i =softmax(s i ) Among them, w i is the normalized weight of the i-th mode; Based on the normalized weights, the weighted fusion features are calculated and expressed as: z=sum(w i *h′ i ) Among them, z is the feature representation after weighted fusion, h′ i is the feature representation vector of the i-th mode; Cross-modal attention involves cross-modal information interaction and calculating the attention coefficient between modalities, expressed as: in, is the attention coefficient from modality i to modality j, W1 and W2 are trainable transformation matrices, and h′ j is the feature representation vector of the jth mode; Based on the attention coefficient between modalities, the feature representation of cross-modal interaction is calculated, which is expressed as: Among them, z i is the updated representation of the i-th modality incorporating information from other modalities; The updated modal representations are adaptively fused and expressed as: h′ fused =FC(Concat[z i ,...,from m ]) Among them, h′ fused is the final fusion feature representation, FC is the fully connected layer, [z i ,...,z m ] is the updated feature representation of m modalities.

6. The method for mining substation misoperation patterns based on large-scale historical data according to claim 5, characterized in that: The use of causal reasoning includes constructing a causal graph using a structural causal model and quantifying the causal effects of influencing factors based on counterfactual reasoning; Construct a causal diagram to establish key variables and causal relationships, expressed as: G=(V',E') Among them, G is a causal graph, V′ is the set of key variable nodes in the substation system, and E′ is the set of directed edges between nodes; Based on the Bayesian information criterion, the causal graph is evaluated and the scoring function is calculated, which is expressed as: Score(G)=BIC(G)=logL(G)-g / 2*loge Among them, Score(G) is the scoring function, BIC is the Bayesian information criterion, L(G) is the 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 scoring function, the optimal causal structure is generated, which is expressed as: G * =argmax G Score(G) Among them, G * is the optimal causal structure; Based on counterfactual reasoning, the average treatment effect is calculated and expressed as: ATE=E[Y|do(A'=1)]-E[Y|do(A'=0)] 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 the variable A′ to the value a′, and E[Y|do(A′=a′)] is the expectation of the misoperation risk indicator Y after artificially setting the variable A′ to the value a′; The backdoor criterion is used for causal inference to accurately estimate the causal effect from the observed data, which is expressed as: P'(Y|do(A'=a'))=sum(P(P|A'=a',C'=c')P(C'=c')) Among them, P′(Y|do(A′=a′)) is the conditional probability distribution of the misoperation risk index Y after the variable A′ is artificially set to the value a′, and C′ is the set of adjustment variables that meet the backdoor criterion; Based on causal inference, the robustness test mechanism includes evaluating the potential impact of unobserved confounding factors on the causal effect estimate through sensitivity analysis, adopting a double robust estimation method, simultaneously modeling the treatment allocation mechanism and the result generation mechanism, introducing confidence interval estimation, and characterizing the degree of uncertainty in the causal effect estimate.

7. The method for mining substation misoperation patterns based on large-scale historical data according to claim 6, characterized in that: The interpretability analysis includes building a three-layer attention visualization framework for the key components in the spatial heterogeneous graph network structure of substation equipment. At the node level, the importance of the implicit representation of each node is scored through a multi-layer perceptron, which is expressed as: μ l =softmax(MLP(p l )) Among them, μ l is the attention score of node l, p l is the implicit state vector of node l; At the edge level, based on the source node and the target node, the attention mechanism is jointly represented as: μ l′ =softmax(MLP([p s ;p t ])) Among them, μ l′ is the importance score of the edge, p s 、p t are the implicit representations of the source and target nodes connected by the edges respectively; At the path level, the overall importance is evaluated by aggregating the representations of all nodes on the path, expressed as: μ path =softmax(MLP(sum(p l for l in path))) Among them, μ path is the importance score of the path, path is a complete path in the graph, and sum is the sum of the implicit representations of all nodes on the path; By extracting decision rules, complex data patterns are transformed into actionable operation and maintenance guidance. The CART decision tree is used for approximation. The training objective function of the decision tree is expressed as: loss=MSE+λ*tree complexity Among them, loss is the training objective function of the decision tree, MSE is the mean square sum of prediction errors, and tree complexity is the complexity of the decision tree, λ is the balance factor; Based on the trained decision tree, a rule extraction method with importance threshold is designed, which is expressed as: rules=extract rules (tree,threshold=0.1) Among them, rules are the output decision rules, extract rules To extract rules, tree is the trained decision tree, threshold=0.1 is the importance threshold set to 1; Traverse the decision tree from the root node and extract the decision rules of the nodes whose split gain is greater than the importance threshold, including equipment load rules, temperature anomaly rules, operation sequence rules, environmental condition rules and equipment status combination rules; Equipment load rules include that if the transformer load rate exceeds 85% and lasts for more than 30 minutes, and the ambient temperature is higher than 35°C, there is a risk of equipment overload and load transfer is required immediately; If the switching frequency of the circuit breaker exceeds twice 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 rules include: if the transformer oil temperature rise rate exceeds 2°C / hour, and the ambient temperature has been stable for the past 3 hours, there may be an internal transformer fault and an early warning needs to be initiated; When the temperature of the busbar connection point is more than 30°C higher than the ambient temperature and the load current does not reach 80% of the rated value, there is a risk of poor contact; The rules for the sequence of operations include that if multiple switchgears in the same circuit are operated consecutively within 10 minutes without following the standard confirmation process, there is a risk of logical errors in the operation; When maintenance operations are not performed in accordance with the five prevention steps and weather conditions are poor on site, operations need to be suspended and reassessed; Environmental condition rules include that when relative humidity exceeds 85% and the insulation monitoring value is below the standard threshold of 90%, there is a risk of insulation damage and enhanced monitoring is required; If thunderstorms occur and the lightning protection equipment test results are abnormal, the lightning protection emergency plan needs to be activated immediately; The device status combination rule includes that when the status indicators of multiple associated devices are abnormal at the same time, and the time interval between the abnormalities is less than 5 minutes, there may be a risk of systemic failure; If the monitoring parameters of the main transformer and related circuit breakers fluctuate simultaneously within a short period of time, and the fluctuation amplitude exceeds 50% of the normal value, the linkage protection mechanism needs to be activated.

8. A system using the substation misoperation pattern mining method for large-scale historical data as claimed in any one of claims 1 to 7, characterized in that: Including dynamic graph network construction and feature learning module, multimodal deep representation learning module, causal reasoning and interpretability analysis module; 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 status by combining the equipment's time-series data and health status modeling; The multimodal deep representation learning module is used to perform cross-modal feature alignment and dynamic fusion to extract deep representations of multi-source data; The causal reasoning and explainability analysis module is used to perform pattern recognition and risk assessment using causal reasoning and explainability analysis.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the substation misoperation pattern mining method for large-scale historical data according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the substation misoperation pattern mining method for large-scale historical data according to any one of claims 1 to 7 are implemented.

Images