Box-type substation state monitoring and early warning method based on artificial intelligence

Abstract

The invention discloses a box-type substation state monitoring and early warning method based on artificial intelligence, relates to the technical field of intelligent power grids, and aims to solve the problems of missing report, false report and response lag caused by the fact that an existing static threshold ignores multi-physical coupling and a depth model highly depends on scarce fault samples. According to the scheme, sliding window kernel density estimation is carried out on a multi-channel time sequence signal, a dynamic coupling matrix is constructed through recursion Copula decomposition, a three-level threshold surface is generated through time-varying quantile regression, abnormal samples and graph attention network extraction state representation are generated in combination with a conditional variation auto-encoder, lightweight recursion pruning is carried out, and the dynamic coupling matrix is obtained. An abnormal score is generated through a multilayer Bayesian network and particle filtering, a multi-step risk trend is discriminated through a Gaussian kernel derivative slope, and finally unscented Kalman filtering is used for smoothing and online threshold correction; according to the method, the detection sensitivity and the early warning recall rate of the box-type substation to the transient coupling fault are remarkably improved, the response speed is improved, and the false alarm frequency is effectively reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of smart grid, and more particularly to a box-type substation state monitoring and early warning method based on artificial intelligence. BACKGROUND

[0002] With the rapid development of smart grid, as a key node in the distribution network, the operation state monitoring and early warning capability of the box-type substation is directly related to the power supply reliability and equipment safety. The traditional monitoring method based on threshold and experience rule is difficult to cope with the dynamic changes of multi-variable coupling and working condition, and gradually evolves to the direction of digital twin, deep learning and other artificial intelligence-based technologies to realize real-time modeling, accurate prediction and intelligent early warning of the box-type substation equipment state.

[0003] The current box-type substation monitoring and early warning method mainly relies on the collaborative application of multi-modal sensor network and intelligent algorithm. Through the deployment of temperature, current, vibration and other sensors and combined with wireless communication technology, real-time acquisition and transmission of equipment operation data are realized, and some schemes also introduce convolutional neural network to extract image features to identify equipment abnormal state, or predict equipment aging trend through time series analysis model, combined with cloud big data analysis to realize hierarchical early warning. At the same time, the digital twin platform is used to construct virtual mapping to assist operation and maintenance personnel in fault location and decision optimization. In the existing patent literature, such as patent CN116231874A, an attribute matrix analysis method is introduced to determine the state interval by calculating the mean and variance of the multi-station homomorphism data set, realizing static threshold division; patent CN119401655A integrates deep learning algorithm and digital twin technology on this basis, using three-dimensional simulation scene to predict and display equipment health state; in addition, CN116776650A proposes a multi-level front auxiliary end processing architecture and dynamic simulation mapping unit to cooperatively complete overall simulation prediction and fault probability evaluation.

[0004] However, since the attribute matrix reconstruction only calculates the mean and variance of each parameter in the same state to determine the early warning interval, the mean / variance analysis only reflects the distribution characteristics of a single physical quantity in the "steady state", ignoring the coupling dynamics between multiple physical quantities such as voltage, current, temperature (e.g. temperature rise causing electrical parameter drift or harmonic and temperature interaction). When the system experiences rapid transients (such as short circuit impact, switching transient), single variable statistical interval is difficult to capture the coupling effect in time, which easily leads to the risk of missing or delayed reporting of faults. In addition, the selection of "boundary value" of this method depends on the distribution of historical data. Once the running environment (load mode, seasonal temperature) changes, the original threshold value will be invalid, which is easy to produce false alarm. Secondly, the iTransformer prediction framework based on the digital twin scene often needs a large number of labeled samples for each hardware state (normal, slight abnormal, serious abnormal) in the actual engineering scene; however, real fault events are rare, and the positive and negative sample ratio is seriously unbalanced. In the case of insufficient or skewed samples, the model is prone to overfitting to the historical fault mode, and the sensitivity to new or composite faults is insufficient, reducing the prediction recall rate. At the same time, the digital twin scene rendering + Transformer forward calculation requires extremely high computing power on the edge side or mobile terminal, resulting in high page refresh delay and difficulty in meeting the on-site rapid response. SUMMARY

[0005] In view of the deficiencies of the prior art, the application discloses a box-type substation state monitoring and early warning method based on artificial intelligence, aiming to solve the problems in the background art.

[0006] In order to achieve the above technical effects, the application adopts the following technical solutions: A box-type substation state monitoring and early warning method based on artificial intelligence, comprising: Step 1, performing non-parametric kernel density estimation on the voltage, current, temperature, Gas pressure and harmonic multi-channel time series signal data collected by each box-type substation through a sliding window mechanism, and constructing a joint distribution sample set; Step 2, using a recursive Gaussian Copula model to decompose the edge and joint correlation of the joint distribution sample set, and obtaining a dynamic Copula correlation matrix; Step 3, based on the dynamic Copula correlation matrix, constructing a three-dimensional threshold surface family through quantile regression to generate normal, early warning and alarm three-level time-varying thresholds; Step 4, based on the standardized time series data of step 1, constructing a variational autoencoder to learn the hidden variables of the normal working condition feature subspace, and obtaining diversified synthetic abnormal samples; Step 5, inputting the diversified synthetic abnormal sample and historical protection action log, switch device alarm record into a time sequence diagram network, extracting a switch cabinet state feature vector by using a node and time attention mechanism, and obtaining a feature vector sequence; Step 6, iteratively pruning the feature vector sequence by using a lightweight recursive pruning algorithm, eliminating time redundancy nodes, and obtaining a state prediction sequence; Step 7, probabilistically fusing the state prediction sequence and the three-dimensional threshold surface family, calculating a comprehensive abnormal score curve through a Bayesian network, and judging a multi-step risk trend based on a curve slope; Step 8, based on the comprehensive abnormal score curve and the multi-step risk trend, extracting a filtered threshold drift vector by using an unscented Kalman filter smoothing method, and feeding back and adjusting a tail dependence parameter of quantile regression in step 3 to dynamically update a quantile surface.

[0007] Based on the above technical solutions, the positive and beneficial effects of the present application are: The present application can capture coupling effects and identify abnormalities immediately when the system experiences short-circuit impact, switch mutation and other fast transients by decomposing the high-dimensional joint distribution of multi-channel voltage-temperature-harmonic signals into a dynamic Copula correlation matrix and constructing a time-varying threshold surface family of quantile mapping, thereby eliminating the risk of single-variable threshold false negatives or false positives. The threshold surface is further self-corrected through multi-sliding window voting and exponential weighted smoothing, which can adjust the early warning boundary in real time according to the load mode and seasonal temperature fluctuations, avoiding false positives caused by the failure of historical static thresholds.

[0008] Meanwhile, the present application introduces a graph network representation of synthetic abnormalities and federated contrast learning based on variational autoencoder to address the problem of fault sample scarcity and class imbalance, filling in the gaps of multi-hardware state data and improving the sensitivity of the model to new and composite faults, thereby maintaining high recall rate under different operating conditions without relying on massive annotations. Combined with iterative pruning and lightweight recursive reasoning, as well as filtering and trend correction by unscented Kalman filter, the overall scheme can ensure the accuracy of deep coupling detection while achieving online inference and adaptive threshold update at a level of seconds, effectively meeting the rapid response needs of box-type substations and reducing the cost of false alarms and missed alarms in operation and maintenance. BRIEF DESCRIPTION OF DRAWINGS

[0009] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, a brief introduction to the drawings needed in the embodiments or prior art description will be given below. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative labor. Figure 1A step flow chart for the present application; Figure 2 A working principle diagram of the recursive Gaussian Copula model of the present application; Figure 3 An example diagram of the three-dimensional threshold surface family of the present application; Figure 4 A structural example diagram of the variational autoencoder of the present application; Figure 5 A working process step diagram of the lightweight recursive pruning algorithm of the present application. DETAILED DESCRIPTION

[0010] To make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions of the present application will be described clearly and completely below in conjunction with embodiments. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0011] To facilitate the understanding of the present application, a kind of box-type substation state monitoring and early warning method based on artificial intelligence disclosed in the embodiments of the present application is introduced in detail, please see Figure 1 A step schematic diagram of a kind of box-type substation state monitoring and early warning method based on artificial intelligence shown in the figure, the method steps include: Step 1, the voltage, current, temperature, Gas pressure and harmonic multichannel time sequence signal data collected by each box-type substation are subjected to non-parametric kernel density estimation, and a joint distribution sample set is constructed; it should be noted that the "sliding window mechanism" in the present application is different from the traditional fixed window technology, which refers to the segmentation and interception of multichannel data in a continuous time domain in an overlapping manner. The "kernel density estimation" in the present application is different from the parameter estimation method, which emphasizes the non-parametric probability density function reconstruction capability and can capture high-order distribution characteristics. The "joint distribution sample set" in the present application is different from the independent edge sample, which represents the joint statistical behavior between multiple physical quantities.

[0012] In the actual box-type substation field, the step 1 obtains the phase voltage time sequence output by the voltage transformer in the distribution room or GIS cabinet, the phase current time sequence output by the outgoing current transformer, the temperature curve of the winding temperature sensor, the pressure curve of the SF6 gas pressure transmitter and the harmonic amplitude sequence output by the partial discharge detection sensor or harmonic analysis module. All sensors are connected to the local RTU or edge computing node, and after synchronization and alignment at a frequency of 8 ms or higher through IEC61850 sampling protocol, they are transmitted to the monitoring server through Ethernet or optical fiber bus.

[0013] Specifically, the sliding window mechanism in the application is composed of the following processes: when a new sampling packet arrives, if the difference from the last window start time is greater than the step size S, the window is triggered to move; the ratio of the step size S to the window width L is usually set to be between 0.2 and 0.5 to capture the mutation while maintaining statistical stability, and can be determined by the actual situation, which is not limited. Each window intercepts a multi-channel synchronous sampling vector, recording a local data block containing voltage, current, temperature, gas pressure and harmonic values.

[0014] Subsequently, based on the local window sample set, multi-dimensional bandwidth matrix adaptation is performed. First, Z score standardization is performed on each channel sample, and the formula is: wherein is the original channel sample, , are the mean and standard deviation of the channel sample, respectively. In the application, "standardization" is different from normalization, and emphasizes centralization and scale unification.

[0015] After standardization, the sample covariance matrix Σ of each channel is calculated, which is used to describe the linear correlation between different physical quantities. The basic bandwidth vector is initialized based on the Silverman rule , and the formula is: wherein represents the number of physical quantity channels, is the number of samples in the window, is the median of the standard deviation of each channel. It should be noted that in the application, "median" is different from average, and is more resistant to the influence of outliers.

[0016] Next, the regularized diagonal bandwidth matrix is constructed, which is the standard deviation vector of each channel sample, wherein the coefficient α∈(0, 1] is used to control the global smoothing strength, which can be adjusted online in the range of 0.5 to 0.9 according to the transient detection requirement. The diagonal bandwidth matrix is different from the full matrix bandwidth, which focuses on edge scale control to reduce the computational complexity of non-parametric estimation.

[0017] After the bandwidth matrix is generated, a multi-dimensional non-parametric kernel density estimation algorithm is used to reconstruct the probability density of the local sample set. The core formula is: wherein is the kernel function, ( ,…, ), represents the point to be estimated; The kernel function in the present application is a Gaussian kernel, which is selected as the i-th diagonal element of the bandwidth matrix H to capture the continuity of the distribution of each channel. It should be noted that the "kernel function" in the present application is different from the window function, which emphasizes the weight weighting characteristic. The i-th dimension value of the j-th sample.

[0018] After the window ends, the kernel density estimation result is the joint probability density function in the time window. Through the iterative processing of the overlapping windows, a time-continuous joint probability density function sequence can be generated. This sequence constitutes a joint distribution sample set, which will be referenced in the subsequent steps for Copula modeling and threshold generation. It should be noted that the "joint density function sequence" in the present application is different from the independent density estimation, which emphasizes the time continuity and multi-channel correlation characteristics.

[0019] In implementation, the implementation mode can be deployed in an edge computing node in a box-type substation monitoring system, and cooperates with Docker containerized deployment to perform seamless calculation on real-time data streams. In addition, when there is insufficient data or abnormal collection, a missing value imputation mechanism can be triggered to complete data completion using sample moving average or polynomial interpolation method, thereby ensuring the completeness of the input of kernel density estimation.

[0020] Step 2, using a recursive Gaussian Copula model, the edge and joint correlation of the joint distribution sample set is decomposed to obtain a dynamic Copula correlation matrix; for details, please refer to Figure 2 The working principle of the recursive Gaussian Copula model is as follows: the joint distribution sample set is mapped to a standard normal edge through a probability integral transformation, and the Gaussian Copula parameters are iteratively estimated by calling the expectation maximization algorithm on the mapped data with a fixed forgetting factor in a sliding time window; after each iteration, the Pearson correlation coefficient between channels is calculated based on the estimated covariance matrix, and the principal coupling direction vector and the corresponding eigenvalue are extracted by using spectral decomposition on the covariance matrix, and are updated according to the time stamp to form a dynamic Copula correlation matrix reflecting transient shock, harmonic temperature coupling and current temperature cross-correlation effects.

[0021] It should be noted that the "joint distribution sample set" in the present application is different from the traditional single-point statistical quantity set, and specifically refers to the time-sequential joint probability density function output by the sliding window kernel density estimation. Each of the d monitoring channels (voltage, current, temperature, harmonics, etc.) represents the statistical dependence within the window width L. This step designs a dynamic modeling mechanism for the internal physical coupling effects of the box-type substation (such as the temperature rise lag caused by current heating, the mutual interference of harmonic distortion and heat dissipation efficiency). Its specific implementation is realized through three layers of operations: probability integral transformation, recursive parameter estimation and spectral analysis. Three key parameters need to be preset in the process: forgetting factor β (typical value 0.95, used to control the decay rate of historical data, the specific value can be adjusted according to the stability of the device working condition), convergence threshold ε (such as , determines the termination condition of the EM algorithm), the number of main coupling directions k (default value 3, to capture the dominant coupling modes such as current-temperature, harmonic-temperature, and the like, which can be flexibly set between 1 and the number of channels d according to the diagnostic needs in actual application).

[0022] First, the probability integral transformation and the standard normal mapping are performed, as a possible implementation, the joint distribution of each time window t The marginal cumulative distribution function is calculated channel by channel: (i is the channel index, and the integral calculation discretization step is recommended to be 1% of the data range), the original signal is converted into a [0, 1] uniformly distributed variable ; the Gaussian marginal data is obtained by mapping through the inverse standard normal function , forming a standard normal data set . This operation eliminates the Gamma distribution characteristics (shape parameter k≈1.8) of the temperature channel and the Gumbel distribution characteristics (skewness γ≈1.2) of the current channel, so that the correlation modeling focuses on the covariance structure. The technical essence lies in constructing the Gaussian Copula basis-the joint distribution of the standard normal data set 𝐙 is defined by the covariance matrix .

[0023] Then the recursive expectation maximization algorithm is started to estimate iteratively. In the initialization, the optimal covariance matrix of the previous window and the unit matrix are combined with a weight: (the forgetting factor β (0.95) controls the strength of historical information inheritance here). In the EM cycle, the expectation step calculates the conditional expectation of the log-likelihood ; the maximization step solves the update equation: where n is the number of samples in the window. When the Frobenius norm difference (convergence threshold ) of the adjacent iteration matrices is less than , the output is terminated. It should be noted that the "recursive" mechanism in this application is different from the traditional batch EM algorithm, which realizes continuous adaptation of parameters through β: for example, β=0.95 means that the historical information weight decays by 5% per window, which can track the slow change of the correlation structure caused by seasonal temperature drift or gradual load change, while the ability to capture sudden transient events is maintained through window memory.

[0024] Finally, a dynamic Copula correlation matrix is ​​constructed. A two-track analysis is performed on the covariance matrix: one track calculates the Pearson correlation coefficient for channel pairs (i,j). To form the basic correlation matrix =[ ]d×d, and secondly, perform spectral decomposition. , where Λt=diag( ,⋯, () is a diagonal matrix with eigenvalues ​​sorted in descending order. The first k eigenvectors are selected. and its eigenvalues As the main coupling direction.

[0025] It should be specifically defined that, in this application, "transient shock" refers to a sudden change in the modulus of the eigenvector and an eigenvalue proportion exceeding a threshold. / ∑ >0.3) represents the sudden correlation mode, such as the current-harmonic correlation jump caused by circuit breaker operation within 5 milliseconds; while "harmonic-temperature coupling" specifically refers to the nonlinear cross-correlation between harmonic components (such as the 5th harmonic channel h5) and the temperature channel T, which is quantized by the product of eigenvector components; "current-temperature cross-correlation" is different from the steady-state thermal model, and the dynamic delay effect it describes (such as the current peak leading the temperature rise by 2-5 minutes) needs to be analyzed by... Time-shifted cross-correlation analysis extraction. The dynamic Copula correlation matrix is ​​output as a structured data volume: When analyzing the physical meaning, the coupling type can be predefined based on the eigenvector component patterns: for example =[0.1,0.8,0.3,−0.5] corresponds to the temperature rise correlation dominated by the current channel (current component 0.8). =[0.7,0.1,−0.6,0.4] maps harmonic-temperature coupling (harmonic component 0.7 and temperature component 0.4 are inversely correlated). This type of pattern matching algorithm can use cosine similarity and other measures, and the specific implementation method is not limited.

[0026] This step is deeply integrated with the overall monitoring and early warning method: dynamic correlation matrix { } serves as a device status fingerprint input to the downstream AI model. When Correlation coefficients of key channels (e.g., current-temperature) Breaking through historical fluctuation zone When an anomaly warning is triggered; main feature direction The continuous offset (cumulative angular change > 10∘ / month) indicates thermal aging of the insulation material; while the coupling mode {( Matching with the preset fault database (e.g.) (Approaching [0, 0.9, 0.1, −0.4] indicates cooling failure) to support root cause diagnosis. It should be noted that the "dynamic Copula correlation matrix" in this application does not only contain correlation coefficients, but rather integrates the original correlation matrix, main coupling directions, and timestamps into a ternary structure. Its spectral decomposition features overcome the shortcomings of traditional correlation coefficients in characterizing nonlinear correlations, thus providing deep learning models with input features possessing explicit physical meaning. The flexibility in technical implementation is reflected in: the forgetting factor β can be set to 0.98 for highly stable equipment (such as indoor stations in constant temperature rooms) to enhance memory, while 0.90 is recommended for heavy-duty outdoor equipment to improve sensitivity; the number of main directions k can be set to 1 in basic monitoring and extended to d in refined fault tree analysis; feature meaning mapping allows for custom channel weight templates combined with the equipment physical model, enhancing the interpretability of industrial scenarios.

[0027] In implementation, recursive Gaussian Copula modeling removes marginal distribution interference through probability integral transformation, and recursive EM based on the forgetting factor achieves dynamic parameter tracking. Combined with spectral analysis to capture the main coupling direction, it jointly solves the correlation modeling problem under non-Gaussian, dimensional heterogeneous, and operating condition drift conditions of multi-channel signals. Its output dynamic correlation matrix transforms implicit physical effects such as transient shocks and harmonic-temperature coupling into quantifiable features, enabling subsequent AI early warning models to inherit the physical traceability of white-box models while maintaining black-box learning capabilities, systematically improving the sensitivity and diagnostic confidence of transformer substation condition assessment.

[0028] Step 3: Based on the dynamic Copula correlation matrix, construct a quantile regression model. Figure 3 The three-dimensional threshold surface family shown generates time-varying thresholds for normal, early warning, and alarm levels. Specifically, the input for this step is a dynamic Copula correlation matrix extracted within a specified sliding window based on a recursive Gaussian Copula model. This matrix is ​​obtained by nonparametric kernel density estimation and high-dimensional feature fusion of multi-channel voltage, current, winding oil temperature, SF6 gas pressure, and partial discharge signals, and is used to construct a four-dimensional feature tensor. The temporal dimension of this tensor corresponds to the monitoring time point, the channel dimension corresponds to different sensor paths, the coupling factor dimension corresponds to the correlation coefficient entries output by the Copula model, and the spatial topology dimension corresponds to the topology index of the busbars, switchgear, and insulation chambers inside the prefabricated substation. To adapt to different monitoring scenarios and network structures, the "spatiotemporal coordinate system" in this application specifically refers to a data organization method that combines time-series labels with physical location encoding, which is different from traditional schemes that are based solely on time or spatial separation.

[0029] Regarding parameter settings, the system predefines a set of quantization quantization points Q = {0.05, 0.50, 0.95} corresponding to the normal operation boundary, early warning trigger boundary, and alarm critical boundary, respectively. The tensor B-spline basis function set B consists of multi-order three-dimensional B-spline basis functions. The number of basis function nodes can be flexibly adjusted according to the internal structural complexity and sampling density of the prefabricated substation. The control points and node vectors corresponding to each basis function need to be optimized through historical data statistical analysis during the system initialization phase. The penalty coefficient λ is responsible for weighting the L2 norm of the basis function coefficient matrix C to prevent excessive oscillations during the high-dimensional fitting process. The entropy regularization coefficient α, based on the information entropy maximization criterion, uniformly constrains the coefficient distribution. The second-order smoothing constraint coefficient β uses a second-order difference penalty term to smooth the neighborhood changes of the coefficient matrix. The value range and initial settings of the above coefficients can be determined through experimental simulation in the digital twin environment, or fine-tuned based on on-site operating experience. The specific values ​​can be determined according to the actual situation, and no limitation is imposed.

[0030] In practice, the input feature tensor T is first expanded using a tensor B-spline basis to obtain the basis function coefficient tensor C, and the tensor expansion error is ensured to be below a preset threshold. Solving for the condition using the least squares method The optimal coefficient C is obtained to achieve local accurate fitting of multidimensional coupling characteristics. Unlike one-dimensional spline interpolation, the tensor B-spline basis functions described in this scheme adopt a multi-order tensor product form, which makes the three-dimensional surface have C2 continuity between basis function nodes, and can effectively adapt to the complex coupling characteristics of temperature field, air pressure field and electric field inside the box-type substation.

[0031] Secondly, for each quantile q∈Q, an asymmetric weighted absolute bias loss function is used. Constructing the quantile regression objective function ,in This represents the actual coupling degree after Copula transformation at time t. The values ​​represent the quantile function reconstructed based on coefficient C. The "asymmetric absolute bias loss function" differs from conventional absolute error; it automatically adjusts the weights of positive and negative residuals through a q-bias term to achieve accurate fitting of the specified quantile. This process requires gradient calculation for each data point at each time step and accumulating the time-series residuals in the objective function to ensure a smooth transition of temporal features.

[0032] To further suppress local overfitting at quantile surfaces and improve model robustness, this implementation scheme... Introducing an entropy regularization term based on this. and second-order difference smoothing term To form an enhanced objective function The entropy regularization term helps control the density of the coefficient distribution and prevents extreme values ​​from dominating the model response; the second-order difference smoothing term constrains the local curvature of the surface by penalizing the differences between adjacent coefficients. It should be noted that the "entropy regularization" in this application differs from the traditional cross-entropy loss, focusing on the diversity constraints based on the coefficient distribution, and has better robustness to actual sudden anomalies.

[0033] In the optimization solution stage, by optimizing each enhanced objective function... The gradient is obtained by taking the partial derivative. ,in The system uses the Adam optimization algorithm to iteratively update C, and evaluates the prediction error and overfitting degree of the current surface on the validation set after each iteration. When the error decreases below a preset threshold... or significant overfit index Stop iterating when an upward trend is observed. The parameters for the "Adam optimization algorithm" here include the learning rate η, momentum coefficients β1 and β2, etc., which can be referenced from the Transformer model parameter settings and fine-tuned according to the on-site sampling frequency and computing resources. No restrictions are imposed on this.

[0034] Once the optimization process converges, the system will obtain the optimized result. Mapping back to 3D B-spline basis functions and to Min-Max normalization is performed to ensure that the surface values ​​are distributed within the range [0,1], thereby avoiding excessive differences in the magnitude of values ​​between different quantiles. The normalization formula is as follows: ,in and These represent the minimum and maximum values ​​on the surface, respectively.

[0035] In the subsequent real-time threshold mapping stage, the system maps the real-time value of the online coupling degree d to a normalized surface by solving a least-squares fitting problem. Specifically, this involves solving... and output Real-time generation Three-level threshold set. It is worth noting that in this application, "least square fitting" can use either ordinary least squares or weighted least squares algorithm, and can be configured according to the on-site computing power and algorithm stability, without any limitation.

[0036] After the threshold set is generated, the three-level time-varying threshold is immediately passed to the Bayesian network fusion module in step 7 for anomaly scoring and early warning determination. When the real-time coupling degree value exceeds the threshold for M consecutive times... And within the time window Δτ, when the logical consistency check passes, the system triggers a high-priority alarm command and generates an SNMP Trap or Modbus write register command to send to the operation and maintenance host computer, thereby completing the overall early warning closed loop.

[0037] Step 4: Based on the standardized time-series data from Step 1, a variational autoencoder is constructed to learn latent variables in the feature subspace of normal operating conditions, thereby obtaining diverse synthetic anomalous samples. The variational autoencoder generates diverse synthetic anomalous samples based on the batch-normalized multidimensional feature vectors of normal operating conditions and the corresponding operating condition labels. The working principle of the variational autoencoder is as follows: In the encoder, the operating condition label vector and the original features are fused and mapped to the mean and variance parameters of latent variables through a cascaded fully connected layer, and latent vectors are generated by sampling through a reparameterization method. In the decoder, the temperature curve, pressure fluctuation, and discharge time series are reconstructed in parallel through a multi-layer deconvolutional network and skip connections. During the reconstruction process, a self-attention mechanism is used to focus on key channels. During training, a composite loss function that aligns the reconstruction error with the prior latent variables is used as the optimization objective, and multi-scale interpolation sampling is performed based on the distribution differences of latent variables after each iteration to generate diverse synthetic anomalous samples.

[0038] For details, please refer to Figure 4 In this method, the variational autoencoder's network structure and data flow for learning the feature subspace under normal operating conditions and generating abnormal samples are as follows: First, the standardized multidimensional feature vectors from the prefabricated substation and their corresponding operating condition labels are input into the encoder module. After concatenation, the fused vector is mapped to the parameters of a Gaussian distribution in the latent space—namely, the mean and log-variance of the latent variable distribution—through two fully connected layers and batch normalization and activation operations. Finally, the encoder parameterizes this Gaussian distribution into specific latent vectors through reparameterized sampling.

[0039] Subsequently, the latent vectors enter the decoder module, first passing through a fully connected layer and being reshaped into multi-channel tensors. Then, they sequentially pass through two sets of deconvolutions and skip connections between the corresponding encoder layers, ensuring that high-frequency pulse information and low-frequency trend signals are transmitted together. Each deconvolution is followed by a self-attention channel focusing unit to emphasize key channels such as sudden temperature rises, pressure drops, or partial discharge spikes. The decoder ultimately outputs a reconstructed data sequence with the same dimension as the input features. This reconstruction result is used to calculate the reconstruction error with the original data. Furthermore, after training, it can be used in the latent space to generate diverse synthetic anomaly samples through multi-scale interpolation and reparameterization operations, thereby enhancing the subsequent prediction model's ability to identify rare or compound fault modes.

[0040] Step 4 involves deploying the system to an edge GPU node or a high-performance server in the operations and maintenance center, and connecting it to the standardized five-channel time-series feature vector sequence and corresponding operating condition labels generated in Step 1. The triggering condition is when the normal operating condition dataset accumulates to a preset threshold (e.g., no less than 10,000 window data points), or when training begins after the data warm-up phase is completed during the first run following deployment. The specific conditions can be determined based on the actual situation, and there are no restrictions on this.

[0041] The input consists of batch-normalized feature vectors. and working condition labels Composition, in which The voltage, current, winding oil temperature, SF6 gas pressure and partial discharge signals collected by the RTU / SCADA system of the box-type substation at a fixed sampling rate are fused after wavelet decomposition, multi-scale Fourier transform and maximum likelihood kernel density estimation to ensure that each channel has a zero-mean unit variance distribution. Multiple normal operating condition categories are represented using One-Hot encoding, such as full load, light load, and seasonal delayed start. It should be noted that "batch normalization" in this application differs from online normalization; it refers to performing normalization independently within each training batch to avoid statistical offsets across batches.

[0042] The encoder network consists of a prior mapping layer, a conditional union layer, and a parameter output layer. The prior mapping layer... The composite structure maps X_norm to a 512-dimensional intermediate representation; the conditional union layer combines this representation with... After concatenation, input the Dense(256)-ReLU structure to generate fused features. The parameter output layer is based on The latent variable mean vectors are output through two Dense(128) networks respectively. Sum of logarithmic variance vector , where L is the hidden dimension, and the recommended value range is 16–64. It can be determined according to the training complexity, and there is no limit to this.

[0043] The latent variable sampling unit adopts the reparameterization method according to the formula ε follows a standard normal distribution N(0,I) and is sampled using a pseudo-random number generator. During training, a Z-transform operation is automatically triggered in each iteration to ensure the continuity of the latent space. The module also introduces a multi-scale interpolation sampling strategy. After a fixed training period P, for the current batch μ and σ... 2 Perform linear interpolation and quadratic Lagrange interpolation to generate M sets of interpolation weights. To produce This involves a multimodal sampling set in a specific form, thereby generating anomalous sample diversity. It should be noted that the "multi-scale interpolation" in this application differs from ordinary noise perturbation; it achieves this by adjusting the interpolation weights. Discrete division within the range [0.5, 1.5] is used to achieve interpolation transformation.

[0044] The decoder network employs a combination of skip connections and deconvolution. First, the latent vector z is input into the Dense(256)-ReLU layer to recover the intermediate tensor dimension, and then compared with the corresponding layer of the encoder. Feature fusion is achieved through the Concatenate operation; subsequently, it is achieved through a three-layer process. The progressive upsampling network will restore the fused tensor to its original form. The shape, and the last layer, outputs synthesized samples through Dense (the number of columns is equivalent to the number of channels) and Tanh activation. To improve the reconstruction quality of critical channels, a self-attention module is inserted after the second TransposeConv layer in the network to compute... Later based on The weighted V generates an attention map, which is then added to the original layer output.

[0045] The training strategy employs a compound loss method. ,in This refers to the mean square error within the batch. The standard KLD divergence, with weights for both. The initial learning rate is 0.05, and it decreases linearly to 0.01 after every 10 epochs of training. The AdamW optimizer is used, and the initial learning rate is... When the training set loss does not decrease for more than 10 consecutive rounds, a cosine annealing learning rate adjustment mechanism is triggered, reducing the learning rate to η×0.1. The batch size is BatchSize=64, which can be adjusted according to the computing power and memory of the edge device without limitation.

[0046] The sampling trigger condition is during training when the validation set reconstruction error... The relative decrease is less than the threshold after K=5 consecutive iterations. When the value is 0.001 (e.g., 0.001) and the validation set KLD does not show an increasing trend, the system automatically starts the multi-scale interpolation sampling module to generate... Group latent vector sampling and batch output via decoder The synthesized sample set is then filtered by the statistical boundary verification module (based on the Chi-square test or quantile bias test) before being added to the outlier sample pool. It should be noted that the "statistical boundary verification module" can use the Kolmogorov-Smirnov test, Mann-Whitney U test, or autocorrelation coefficient test; the specific method can be selected according to system requirements and is not limited thereto. This strategy differs from simple noise perturbation, as it can generate rich samples covering both small and large anomalies along the principal component direction. The sampling size (e.g., 100 groups per round) and interpolation dimension can be dynamically configured according to the scarcity of actual fault samples and are not limited thereto.

[0047] The output is a diverse set of synthetic anomaly samples that have been screened, including three representations: physical quantity time series, spectral features, and wavelet coefficients. These can be stored as HDF5 or Parquet format files and used in subsequent steps to train the time series prediction model and anomaly classification model, significantly improving the detection rate and robustness for rare fault conditions.

[0048] The aforementioned synthetic abnormal samples, along with real historical faults and normal samples, are used by the graph neural network and pruning algorithms in subsequent steps 5 and 6. This effectively compensates for the scarcity of fault samples, improves the model's ability to identify multiple types, combinations of types, and novel fault precursors, and achieves the technical effect of high recall and low false alarm rate in the overall monitoring and early warning process of prefabricated substations.

[0049] Additionally, it should be noted that in this module, "latent variable prior distribution" specifically refers to a multidimensional Gaussian distribution assumed to be standard normal in the initial training state, which contrasts with the posterior distribution automatically learned in the variational autoencoder architecture; "reconstruction error" differs from traditional pixel-level MSE, emphasizing multi-scale fusion evaluation in the TTFT (time-frequency-time peak value) domain; and "self-attention module" differs from ordinary attention mechanisms, only weighting the intermediate feature channels of decoding.

[0050] Step 5: Input the diverse synthetic anomaly samples, historical protection action logs, and switch device alarm records into a time-series graph network, and extract the switch cabinet status representation vector using a node and time-based attention mechanism to obtain the representation vector sequence; Specifically, at the prefabricated substation site, the time-series graph network module deployed on the edge computing node or operation and maintenance platform receives the synthetic anomaly samples, historical protection action logs, and switchgear alarm records output from step 4. The trigger condition is the arrival of any new window data or an update to the protection action log; this notifies the module to begin processing via the internal event bus, thus ensuring attention is paid to every state change. It should be noted that in this application, "graph network" differs from a general time-series model, specifically referring to a structured neural network built based on physical topology, used for information interaction at both the temporal and node dimensions.

[0051] First, for the primary circuit structure of the prefabricated substation involved in the current monitoring time, the connection relationships between primary equipment such as protection and control devices, circuit breakers, and surge arresters are read from the system configuration database. Each device or measuring point is abstracted as a graph node. Based on the electrical connectivity between devices and the branching relationship of busbars within the GIS cabinet, an adjacency matrix A is generated. The adjacency matrix is ​​stored in binary or weighted form. The weights can be preset according to the busbar conductor resistance and mutual inductance coupling coefficient. The specific weights can be adjusted according to the actual equipment parameters, and there are no restrictions on this. It should be noted that the "adjacency matrix" in this application is different from a simple topology table, emphasizing electrical connectivity weights and mutual inductance coupling.

[0052] In parallel, based on the time-series entries retrieved from the protection action log and alarm record database, the timestamp and event type of each record are extracted, the time difference Δt between any two timestamps is calculated, and the time similarity matrix S is filled accordingly. In this application, the term "time similarity matrix" is distinct from the static time index and is specifically used to represent the impact of event time intervals on the model's attention allocation. Similarity calculation can employ an exponential decay function: in , For logs or alerts, use the timestamp. This is the time decay factor, and its numerical range can be... Internal adjustments are made to reflect the impact of events at the microsecond or second level, without any limitations. Through this function, more recent protection or alarm events receive higher weight in subsequent attention calculations.

[0053] After completing the A and S constructions, the module sends the feature vectors of each node at the current moment (including synthetic anomaly sample features, protection device status codes, circuit breaker switch position codes, alarm level codes, etc.) and their corresponding time embeddings (such as the time offset encoding relative to the grid operating time in cosmic standard time) to the linear mapping layer. This linear layer consists of a trainable weight matrix. The structure is mapped to a vector of query Q, key K, and value V, respectively: Where X is the concatenated input matrix of nodes or time steps, with dimensions consistent with the number of nodes and time step. It should be noted that the "query / key / value mapping" in this application differs from traditional attention, emphasizing the independent learning of node and time step information.

[0054] Next, node attention is calculated along the node dimension: for the query vectors of all nodes at the same time. With key vector The pairwise inner product yields the unnormalized fractional matrix. in The key vector dimension is used for scaling; subsequently, the row vectors are scaled. Apply the Softmax operation to generate node attention weights. This weight distribution ensures that nodes with tight connections or stronger mutual coupling receive greater attention. It should be noted that the "node attention weight" in this application differs from flat attention, emphasizing physical connectivity and protecting coupling. The following is an excerpt of code: #Q_node,K_node:[num_nodes,d_k] E_node=Q_node@K_node.T / sqrt(d_k) alpha_node=softmax(E_node,axis=1) # Row normalization After calculating the attention weights at the node level, the processing flow for attention at the time step begins. The system first stacks the node feature sequences of the same graph network at different sampling times into a matrix in chronological order. and through the same linear mapping layer as the node processing. Generate time-space query, key, and value vectors. Then, calculate the attention score between each pair of time-space indices p and q. And perform soft maximum normalization on each row to obtain the time-matter focus weights. This weight emphasizes the importance of historical vectors representing the time proximity of protection actions and alarm events in the current moment's representation. The time decay feature in the mapping function and the time similarity matrix are mutually verified to ensure that the weight reflects both physical proximity and event temporal correlation.

[0055] Node focus weight Pay attention to weight at all times The joint attention coefficient tensor is calculated using the outer product method. This coefficient quantifies the association between section i at time p and section j at time q on a four-dimensional index. Subsequently, the system combines this joint coefficient with the value vector. Weighted aggregation is achieved by multiplying element-wise and summing along both the node and time dimensions: H is the weighted representation output matrix, with dimensions corresponding to the original number of nodes and the time step. It should be noted that the "joint attention coefficient" in this application is different from cascaded or serial attention, highlighting the simultaneous capture of spatiotemporal relationships.

[0056] The aggregation junction H undergoes residual connection and layer normalization processing. Specifically, the original input XH is first element-wise summed, and then the summed tensor is normalized using a standard layer to generate the final switchgear state representation vector sequence. This sequence becomes the input to the lightweight pruning and Bayesian scoring modules in subsequent steps. It should be noted that "layer normalization" in this application differs from batch normalization, emphasizing channel standardization of individual sample sequences.

[0057] Within the overall monitoring and early warning framework, the node-moment attention mechanism module, through the aforementioned process, not only integrates information on electrical connectivity and mutual inductance coupling between nodes but also comprehensively considers the impact of historical events at different sampling times. The resulting representation vector sequence can highly condense the complex dynamic signals of switchgear and busbar chambers in prefabricated substations. Subsequent modules such as lightweight pruning, Bayesian fusion, and trend discrimination all rely on this sequence to provide accurate and real-time feature support in response to fault precursors such as short-circuit impacts, sudden increases in partial discharge, or insulation leaks, ultimately achieving a closed loop for second-level intelligent early warning and operation and maintenance decision-making.

[0058] Step 6: Iteratively prune the representation vector sequence using a lightweight recursive pruning algorithm to remove temporally redundant nodes and obtain the state prediction sequence; for details, please refer to [link to relevant documentation]. Figure 5 The lightweight recursive pruning algorithm works as follows: For each representation vector and its adjacent time vectors, the cosine similarity change rate is calculated, and a time importance score sequence is generated; the score sequence is sub-segmented based on a binary recursive partitioning strategy, and the sub-segmentation is further refined when the sub-sequence length exceeds a set window; within each sub-sequence, the scores at each time point are comprehensively evaluated based on the local magnitude range and average gradient, and the node with the lowest contribution is identified as redundant using an adaptive threshold; after removing the redundant nodes layer by layer from the inside out, the similarity change rate is recalculated and the time importance score sequence is updated; in the recursive pruning loop, the segmentation depth and threshold coefficient are dynamically adjusted through online information gain and sequence reconstruction error; the output is a state prediction sequence containing only high information preservation nodes.

[0059] It should be noted that the "recursive pruning" in this application is different from the traditional fixed-window downsampling. It can adaptively remove redundant nodes while preserving the characteristics of fault precursors. The "online information gain" is different from offline evaluation. It reflects the impact of pruning decisions on the accuracy of sequence reconstruction in real time. The "adaptive threshold" is different from the static threshold. It can be dynamically adjusted according to sequence changes and model feedback.

[0060] In implementation, firstly, the algorithm uses the multi-channel representation vector sequence output in step 5. As input, each vector It incorporates fused information from switchgear status codes, partial discharge characteristics, and temperature and pressure dynamics. The trigger condition is when the input length reaches the minimum processing threshold (e.g., no less than 60 time-point vectors) or the operation and maintenance system issues a pruning command. Afterward, the algorithm enters the initial score calculation phase. It first calculates the cosine similarity between the vector at each time point in the sequence and its preceding and following vectors, and then calculates the absolute value of the difference between two adjacent similarity scores to form the original time importance score sequence. This score intuitively reflects the magnitude of the state change at time t compared to t−1 and t+1, and can capture abrupt changes such as short-circuit impacts, switching transients, or partial discharge pulses. It should be noted that the "cosine similarity" in this application differs from Euclidean distance, and is particularly suitable for sensitive detection of angle changes in high-dimensional state vectors.

[0061] After the scoring sequence is generated, the algorithm uses a recursive binary partitioning strategy to process long sequences. Initially, the entire scoring sequence is considered as a first-level subsequence. If the length of this subsequence exceeds a preset upper limit (e.g., 50 time points) and the recursion depth does not exceed the maximum depth (e.g., 4 levels), it is divided into two equal parts, and the above check and partitioning are repeated for each end of the subsequence. Otherwise, the current subsequence is retained. This process can be fine-tuned according to actual computing power and real-time requirements, and there are no limitations on these. After the subsequence partitioning is completed, each subsequence can be regarded as a set of time points with similar dynamic characteristics within a continuous time period, which helps to preserve localized fault precursor signals.

[0062] Next, the algorithm evaluates the local information density of the score values ​​within the subsequence. First, it calculates the range of the score for that segment. The average gradient (the average of the absolute values ​​of the differences between adjacent scores) is used to quantify the overall change magnitude and fluctuation frequency of the interval. Then, an adaptive pruning threshold is automatically generated based on the statistical characteristics of the scores in this segment: the average score of this segment is subtracted from a certain multiple of the score standard deviation (the multiple can be initially set to 1.0 and fine-tuned later based on information gain feedback), which serves as the threshold for judging inefficient time nodes. It should be noted that the "adaptive pruning threshold" in this application differs from the fixed pruning threshold, emphasizing the dynamic differentiation ability between fault precursors and stable segments.

[0063] For each score value in this subsequence A node is marked as redundant when its value is less than the adaptive threshold and the state reconstruction error gain at the corresponding time (the reconstruction residual difference calculated and cached during the previous pruning or graph neural prediction) is also lower than the preset error threshold. This "reconstruction error gain" represents the change in the output error of the graph attention representation or subsequent prediction model before and after removing this node, which is different from the absolute error and emphasizes the impact of the pruning operation on the overall performance.

[0064] After marking redundant nodes, the algorithm removes them layer by layer from the inside out: first, it removes them in the deepest subsequence, then it goes back up to the parent subsequence to repeat the removal, and then it recalculates the cosine similarity of the remaining nodes and updates the time importance score, driving the possible next round of recursive pruning, until the length of all subsequences no longer exceeds the limit or the maximum recursive layer has been reached.

[0065] After each round of recursion, the algorithm calculates an online information gain metric, which is the ratio of the improvement in reconstruction error of the overall sequence before and after pruning, as well as the compression ratio (the ratio of the two can be regarded as an evaluation value of pruning efficiency). This metric is then used to dynamically adjust parameters such as the maximum subsequence length, recursion depth, and pruning threshold multiple to achieve a balance between information retention and sequence compression. It should be noted that "online information gain" in this application is different from traditional information gain, specifically referring to the immediate impact of pruning on the reconstruction or prediction performance of downstream models.

[0066] The iterative pruning and dynamic parameter adjustment process described above continues until the algorithm terminates under the condition that all subsequences are within the preset length or no more nodes that can be removed are identified. Ultimately, the algorithm outputs a "state prediction sequence" containing only the moments with high information content. Its length is typically compressed to 20% to 30% of the original sequence, but it fully preserves key fault precursor features such as short-circuit impacts, harmonic temperature coupling abrupt changes, and partial discharge pulses.

[0067] In the overall monitoring and early warning method, this lightweight sequence is passed to the Bayesian fusion and risk trend determination module in step 7. Due to the significant reduction in the number of nodes, the computational load of subsequent particle filtering and Monte Carlo sampling calculations is significantly reduced, and the system response latency can be shortened to less than 1 / 3 of the original, thus meeting the real-time requirement of second-level early warning for prefabricated substations. At the same time, the pruning strategy based on adaptive parameter adjustment and online information gain feedback also ensures that the system can maintain high robustness and high recall under diverse operating conditions such as different load modes, seasonal temperatures, and equipment aging levels.

[0068] Step 7: Perform probability fusion between the state prediction sequence and the three-dimensional threshold surface family, calculate the comprehensive anomaly score curve through a Bayesian network, and determine the multi-step risk trend based on the curve slope; Specifically, when the lightweight state prediction sequence generated in step 6 arrives at the risk fusion module, the system first uses the timestamp as the key and calls the threshold mapping function for each time vector in the sequence to calculate its spatial position in the three-dimensional threshold surface family to the 5%, 50%, and 95% quantiles. The mapping process uses the least squares fitting method to measure the shortest perpendicular distance between the vector to be tested and each threshold surface. This shortest distance represents the deviation of the state at that time from the normal, warning, and alarm levels. Subsequently, in order to convert the deviation into a probabilistic "likelihood index," the system applies a normalization function to the three distance values ​​respectively. The formula is the reciprocal of the distance divided by the sum of the distances, or a soft cosine normalization method can be used. Here, the normalization strategy can be adjusted online according to the historical alarm experience threshold smoothing coefficient. The specific method can be determined according to the actual situation and is not limited thereto. It should be noted that the "likelihood probability vector" in this application is different from the posterior probability in statistics, emphasizing the mapping of geometric deviation to risk propensity.

[0069] After obtaining the three-dimensional likelihood vector at each time step, the module needs to combine them with the historical dynamic Copula correlation matrix to reflect the impact of the current coupling mode on the reliability of the early warning. The system constructs a multi-layer Bayesian network using these nodes as random variables across three types of nodes: busbar chamber temperature, switchgear status coding, and SF6 insulation structure. The first layer of the network represents the coupling state layer of the three types of physical variables; the second layer represents the likelihood input layer for each quantile; and the third layer is the comprehensive risk scoring layer. For each parent-child node relationship, the system pre-generates a conditional probability table based on the historical fault database and Copula matrix statistics. The values ​​of these entries can be estimated using the EM algorithm or expert experience method, and online EM fine-tuning can be used to adapt to equipment aging or environmental changes when necessary. It should be noted that the "multi-layer Bayesian network" in this application differs from simple naive Bayes, focusing on hierarchical relationships of multiple physical quantities and quantile mapping dependencies.

[0070] After the model structure is prepared, a particle filter algorithm is used to perform recursive posterior updates. The system first generates several particle samples in the node space based on the initial prior distribution and assigns them equal weights. Then, each time a new likelihood vector is received, the weights of each particle are updated according to its likelihood probability generated in the network. After weight normalization, resampling is performed to overcome particle degradation. This process continues until the entire sequence is processed. This real-time recursive process ensures that the posterior distribution of network nodes always reflects the latest state and coupling information. It should be noted that the "particle filter" in this application differs from the Kalman filter because it has stronger adaptability to nonlinear and non-Gaussian distributions, making it very suitable for fault early warning scenarios in complexly coupled prefabricated substations.

[0071] After weight updates and resampling, the system fuses the likelihoods of multi-channel quantile surfaces into a continuous anomaly score curve using Monte Carlo sampling. This curve, with time on the horizontal axis and risk score on the vertical axis, visually presents the evolution trend of the prefabricated substation's status from normal to alarm. The score at each point on the continuous curve is the comprehensive risk measure under the current posterior distribution, requiring no additional normalization. It should be noted that the "anomaly score curve" in this application differs from simple threshold scoring, emphasizing a time-series comprehensive evaluation driven by multi-layer Bayesian and particle filtering.

[0072] After acquiring the continuous anomaly rating curve, the system extracts its first-order slope using an adaptive Gaussian kernel derivative operator. This operator first dynamically estimates the bandwidth parameter of the Gaussian kernel based on the local variation characteristics of the current rating curve. The bandwidth is generated by multiplying the local standard deviation of the curve within a sliding window by a scaling factor, thus adaptively responding to different risk evolution rates. After selecting the bandwidth, a kernel-weighted derivative operation is performed on the curve at each time point: the curve value is convolved with the gradient function of the Gaussian kernel in the local neighborhood to obtain the slope value at each time step. It should be noted that the "Gaussian kernel derivative operator" in this application differs from ordinary numerical differencing, as it can smooth noise and highlight the true trend of sudden risk increases.

[0073] After extracting the slope sequence, the system detects segments with continuously positive slope signs, calculates the duration of these positive segments, and further calculates the radius of curvature within each segment—using the reciprocal of the absolute value of the second derivative of the slope sequence to assess the curvature of the curve and quantify the swiftness of risk evolution. When the radius of curvature is less than a preset threshold (e.g., 50 time units) and the length of the positive segment exceeds a fixed sliding window (e.g., 3 time units), it is determined to be a short-term fault propagation trend. It should be noted that in this application, "radius of curvature" differs from simple curvature, emphasizing the acceleration information of the risk curve at the micro-time level.

[0074] After trend determination is completed, the system enters the logical consistency verification stage: the risk propagation determination result must maintain logical truth value across multiple consecutive sliding windows to avoid false alarms caused by occasional noise. Specifically, each trend determination is binarized and then ANDed with a queue of three windows in length. The verification passes only when all determinations in the queue are "propagation". The length of this "fixed-duration sliding window" and the confidence threshold can be flexibly configured according to the operation and maintenance strategy and are not limited thereto.

[0075] Once the verification is successful, the system immediately constructs an early warning event message, encapsulating the early warning level, timestamp, key node identifiers, and risk indicators into a message conforming to the SNMP Trap protocol, or converting it into a Modbus write register instruction, specifying the target register address and early warning code. The message is then sent to the upper-level operation and maintenance system via the substation communication module (such as IEC61850 MMS or Modbus / TCP gateway). This communication method, unlike manual intervention, ensures that alarms can be quickly captured at the network level and presented on the operation and maintenance visualization platform, while simultaneously being written to the local log database, triggering the fault response process.

[0076] In the overall prefabricated substation condition monitoring and early warning method, the implementation of step 7 ensures seamless integration from high-dimensional coupling characteristics to multi-level threshold mapping, probabilistic fusion, and trend determination, achieving accurate identification and second-level early warning for various risk modes such as short-circuit impacts, sudden increases in equipment partial discharge, and precursors to insulation leakage. Through highly adaptive bandwidth selection, first-order slope kernel derivative, radius of curvature determination, and logical consistency verification, as well as an event delivery mechanism directly compatible with IEC61850 and Modbus protocols, the system ensures high robustness and low false alarm rate under diverse field conditions, and provides an interpretable risk evolution path for operation and maintenance decisions.

[0077] Furthermore, the method for determining the short-term fault propagation trend in step 7 is as follows: For continuous abnormal scoring curves, a Gaussian kernel bandwidth is adaptively selected at each sampling point based on the scoring variance within the local window. The scoring curve is then subjected to a Gaussian first-order derivative convolution operation using this Gaussian kernel bandwidth to obtain a point-by-point slope sequence. Within this point-by-point slope sequence, continuous positive segments are identified using a sliding decision algorithm, and the radius of curvature for each segment is calculated based on its second derivative. The calculation formula is as follows: in, This represents the value of the continuous anomaly score curve at time t; and These represent the first and second derivatives of the scoring curve with respect to time, respectively, and are used to reflect changes in slope and curvature. The incremental scaling factor represents the amplification of the effect on the radius of curvature when the slope of the scoring curve increases. When the length of any segment exceeds the preset time window and its minimum radius of curvature is lower than the preset threshold, it is considered a short-term fault propagation trend.

[0078] In actual operation of prefabricated substations, step 7 continuously receives the continuous anomaly scoring curve generated by a multi-layer Bayesian network and particle filtering. When the curve is updated to a new sampling point or the historical window rolls, the short-term fault propagation judgment process is triggered. In this application, the "continuous anomaly scoring curve" is different from static threshold scoring, emphasizing the dynamic risk measure obtained by comprehensive evaluation of multi-channel quantile likelihood and coupling effect; the "short-term fault propagation trend" is different from long-term trend prediction, focusing on the detection of risk evolution inflection points from seconds to minutes.

[0079] The judgment process first performs adaptive Gaussian kernel bandwidth selection on the scoring curve for each sampling time. The system then extracts a number of scoring points (e.g., ±5 time points) centered on the current time using a sliding window approach, and calculates the variance σ of the scoring values ​​within this window. 2 The bandwidth h is set to k·σ, where the coefficient k (e.g., 1.2) can be fine-tuned to adjust the sensitivity to mutations according to the test scenario. The specific value can be determined based on the actual situation and is not limited thereto. It should be noted that the "adaptive bandwidth" in this application differs from the fixed bandwidth; it can automatically balance smoothness and resolution in stable and rapidly changing segments.

[0080] Once the bandwidth is determined, the system convolves the scoring curve with the first derivative of the Gaussian kernel function to extract the slope. In engineering implementation, mature signal processing library functions can be called, such as the derivative mode of gaussian_filter1d in SciPy, or the convolution can be performed directly using the following pseudocode: #scores: Continuous anomaly score sequence #h: Adaptive bandwidth at the current time kernel=gaussian_derivative_kernel(h,order=1) slopes=convolve(scores,kernel,mode='same') Where gaussian_derivative_kernel is based on the formula One-dimensional convolution kernels are generated to ensure the numerical stability and noise resistance of smooth first-order derivative operations.

[0081] After the slope sequence is calculated, the system scans and detects all continuous positive segments along the time axis, that is, it identifies the longest continuous segment with a slope value > 0 and a positive value in adjacent time intervals. The start and end times, length, and local maximum slope are recorded. The segment length is measured by the number of sampling points or the corresponding actual time length (e.g., 3s, 5s), and can be flexibly configured according to the operation and maintenance strategy. The specific length can be determined according to the actual situation and is not limited. It should be noted that the "continuous positive segment" in this application is different from a simple positive interval, emphasizing the continuity before and after the inflection point and the minimum duration requirement.

[0082] For each identified positive segment, the system further calculates the radius of curvature based on the second derivative of the slope sequence. In the discrete implementation, the radius of curvature R is calculated by taking the difference between the slope sequences `slopes` to form the second derivative `curv=diff(slopes)`, and then using the following pseudocode: curv2=convolve(slopes,gaussian_derivative_kernel(h,order=2),mode='same') radius=1.0 / (abs(curv2)+epsilon) The epsilon is used to avoid division by zero, and is typically taken in the range of 1e-6 to 1e-4 to ensure that the curvature increases sharply at a critical inflection point. A small radius of curvature indicates that the curve has made a significant inflection point at that moment, often corresponding to a sudden increase in the rapid spread of faults.

[0083] Thus, for each positive segment, the system has a segment length L and a minimum radius of curvature. Two core indicators. When any segment satisfies... and At that time, it was determined to be a short-term fault propagation trend, among which (e.g., 3s or 3 sampling points) and (For example, a curvature radius of 50ms) can be pre-set through experiments based on the characteristics of the on-site power grid and operation and maintenance experience, or it can be fine-tuned online after deployment; there are no limitations on this. It should be noted that the "dynamic threshold" in this application is different from the static threshold, and can adaptively respond according to the sampling rate and operating load conditions.

[0084] After determining the propagation trend, the system combines logical consistency verification: the Boolean result of each trend determination is written to a queue of length M (e.g., 3), and a logical AND operation is performed in the queue. Only when all M determination results are true is the robustness of the short-term fault propagation trend confirmed to prevent false alarms caused by noise or occasional peaks. This mechanism can be implemented using a simple sliding window counter or queue structure without additional complex calculations.

[0085] Finally, upon successful verification, the module triggers an early warning event. The system encapsulates the early warning message, including key parameters such as trigger time, section start and end, maximum slope, and minimum radius of curvature, and generates SNMP Trap messages or Modbus write register commands according to the IEC61850 MMS or Modbus / TCP protocols commonly used in prefabricated substations. The issued operation can call the substation communication gateway API and write the message to the operation and maintenance database and visualization platform, while simultaneously recording local logs for offline auditing.

[0086] Step 8: Based on the comprehensive anomaly score curve and multi-step risk trend, the unscented Kalman filter smoothing method is used to extract the filtered threshold drift vector, and the tail dependency parameters of the quantile regression in step 3 are adjusted accordingly to dynamically update the quantile surface.

[0087] Specifically, after the system completes multi-layer Bayesian fusion, particle filtering, and Monte Carlo sampling in step 7, and obtains the continuous anomaly score curve and multi-step risk trend determination results, it enters the unscented Kalman filter (UKF) smoothing and threshold drift correction stage. In this application, the "unscented Kalman filter" differs from the extended Kalman filter. When processing nonlinear system state estimation, it does not rely on Jacobian matrix linearization. Instead, it automatically generates a set of representative σ points in the state space and weights them for transmission, thus more accurately capturing the nonlinear evolution of the score curve and threshold residuals. In this application, the "threshold drift vector" differs from simple time difference, emphasizing the extraction of the systematic mismatch between the score curve and the threshold surface through a state-space model, used for online updating of the quantile regression tail dependency parameters.

[0088] Step 8 involves deploying the hardware on the operations and maintenance center or edge GPU nodes. The trigger condition is to start immediately whenever a new sampling point is updated in the continuous anomaly scoring curve and the multi-step trend determination enters the "proliferation" state. Alternatively, it can be run in batches at fixed intervals (e.g., 1 second). The specific timing can be determined based on the communication latency and operations and maintenance needs of the prefabricated substation; there are no restrictions on this. Upon entering this stage, the system first constructs a state-space model for UKF: [The system then uses the current value of the scoring curve...] Treating it as an observation, the residuals generated by the quantile regression model at the previous time step are considered as observations. As the state variable to be estimated, the system maps the discrete-time index t to the continuous running time based on the actual deployment resolution and time step to ensure that state updates and observation updates remain synchronized.

[0089] In state-space modeling, the state transition function is defined as follows: ,in The process noise represents the natural fluctuation of the quantile residuals without additional input; the observation function is defined as follows: ,in The mapping function of the theoretical threshold surface to the score under the action of quantile residuals can be realized by the surface evaluation after reconstruction by tensor B-spline basis functions. The observation noise represents the random fluctuations in the particle filter score itself. It should be noted that in this application, "process noise" and "observation noise" are different from environmental noise, corresponding to the inherent uncertainty of the model and the score calculation error, respectively.

[0090] During UKF implementation, the system first estimates based on the current state. Covariance Matrix 2n+1 σ points are automatically generated, the number of σ points being determined by the state dimension n=1, and the weights of the σ points adopt the standard UKF weighting formula. The generated σ points are mapped to the prediction space through the state transition equation, forming a set of predicted σ points; the predicted mean and covariance are summed using weighted averages to obtain the prior state estimate. with prior covariance Subsequently, the predicted σ point is mapped to the observed σ point through the observation function, and then the observed mean is calculated from it. With covariance and state-observation covariance And calculate the Kalman gain accordingly. Kalman gain combined with newly arrived observations Correcting the prior state yields the posterior state estimate. With posterior covariance After that, the system will and Persistent storage is provided for use in the next iteration.

[0091] After completing the UKF state estimation, the system will calculate the posterior state covariance. Mean of prior residuals Combined into a "threshold drift vector" This vector represents the systematic shift trend and uncertainty between the current scoring curve and the established threshold surface. It should be noted that the "threshold drift vector" in this application is different from the simple residual, emphasizing that it includes both the drift magnitude and the drift confidence.

[0092] After obtaining the drift vector, the system enters the tail dependency parameter update process. First, based on the excess distance of the original samples of each quantile surface relative to the threshold surface (i.e., the positive offset of the sample point from the surface), the local increment of the loss function for quantile regression is constructed. The system calculates the loss function with respect to the tail dependency parameters in the tensor B-spline mapping coefficient space using a finite difference method. Partial derivatives of: Increment After applying a small perturbation, the quantile surface is reconstructed using the tensor B-spline module, and the excess distance loss for the corresponding sample is recalculated. The difference between the two is divided by... This yields an approximate gradient value. In practical deployments, this process can be replaced by an automatic differentiation framework (such as TensorFlow or PyTorch) to replace finite differences with instantaneous gradient computation, thereby improving efficiency and accuracy.

[0093] After gradient calculation is complete, the system uses gradient descent or an adaptive learning rate optimizer (such as Adam) to perform an update on the tail-dependent parameters: Learning rate It can be initially set to 0.01 and dynamically decreased or increased based on the online convergence speed and model stability. It should be noted that in this application, the "tail-dependent parameter" is distinct from the overall Copula-related parameters, specifically referring to the term affecting the coupling strength of extreme quantile surfaces.

[0094] After the parameters are updated, the system calls the tensor B-spline mapping module to reconstruct the 3D quantile surface family with the new tail-dependent parameters: the updated parameters are injected during the B-spline basis function expansion process. This is used to adjust the curvature of each quantile in the high-value region, thereby enhancing the response sensitivity in the extreme-risk region while maintaining the smoothness of the median risk region. For the multi-physical quantity coupling peak region of the main discharge chamber temperature, SF6 pressure and partial discharge characteristics, the spatiotemporal filling of the new threshold surface can be completed by micro-batch resampling (batchsize 10–20) after surface update, which is usually completed within 50ms.

[0095] In implementation, this step forms a closed-loop self-correction mechanism through UKF σ-point prediction, state update and drift vector extraction, and online gradient optimization of tail-dependent parameters. This ensures that the quantile threshold surface continuously synchronizes with the rapidly changing multi-physical coupling state of the prefabricated substation. This closed loop not only solves the problem of traditional static threshold failure but also compensates for the insufficient response of digital twin and Transformer methods to real-time extreme coupling, comprehensively improving the system's early warning accuracy and response speed. Furthermore, it achieves end-to-end collaboration in the overall intelligent monitoring and early warning process, from data acquisition, coupled modeling, threshold generation, dynamic feedback to model adaptation.

[0096] While specific embodiments of the present invention have been described above, those skilled in the art should understand that these specific embodiments are merely illustrative. Those skilled in the art can omit, substitute, and modify the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, combining the above method steps to perform substantially the same function and achieve substantially the same result using substantially the same method falls within the scope of the present invention. Therefore, the scope of the present invention is defined only by the appended claims.

Claims

1. A method for condition monitoring and early warning of prefabricated substations based on artificial intelligence; characterized in that: include: Step 1: Collect voltage, current, and temperature data for each prefabricated substation using a sliding window mechanism. Nonparametric kernel density estimation was performed on gas pressure and harmonic multi-channel time series signal data to construct a joint distribution sample set; Step 2: Using a recursive Gaussian Copula model, perform marginal and joint correlation decomposition on the jointly distributed sample set to obtain the dynamic Copula correlation matrix; Step 3: Based on the dynamic Copula correlation matrix, construct a family of three-dimensional threshold surfaces through quantile regression to generate three levels of time-varying thresholds: normal, early warning, and alarm. Step 4: Based on the standardized time series data from Step 1, construct a variational autoencoder to learn latent variables in the feature subspace of normal operating conditions and obtain diverse synthetic abnormal samples. Step 5: Input the diverse synthetic anomaly samples, historical protection action logs, and switch device alarm records into a time-series graph network, and extract the switch cabinet status representation vector using a node and time-based attention mechanism to obtain the representation vector sequence; Step 6: Iteratively prune the representation vector sequence using a lightweight recursive pruning algorithm to remove temporally redundant nodes and obtain the state prediction sequence. Step 7: Perform probability fusion between the state prediction sequence and the three-dimensional threshold surface family, calculate the comprehensive anomaly score curve through a Bayesian network, and determine the multi-step risk trend based on the curve slope; Step 8: Based on the comprehensive anomaly score curve and multi-step risk trend, the unscented Kalman filter smoothing method is used to extract the filtered threshold drift vector, and the tail dependency parameters of the quantile regression in step 3 are adjusted accordingly to dynamically update the quantile surface.

2. The method for condition monitoring and early warning of prefabricated substations based on artificial intelligence according to claim 1, characterized in that: In step 1, the sliding window mechanism is used to perform nonparametric kernel density estimation on the collected data as follows: S1. Using a time window with a step size S and a width L, extract synchronous sampling vectors from the multi-channel time-series data to generate data including voltage, current, temperature, and... A local window sample set of gas pressure and harmonic values, wherein the step size S is less than the width L; S2. Adapt the multi-dimensional bandwidth matrix based on the statistical characteristics of data from each channel within the window; the adaptation method is as follows: 1) Perform Z-score normalization on the data of each channel within the window; 2) Calculate the covariance matrix of the standardized sample. ; 3) Initialize the base bandwidth according to Silverman rules. The initialization formula is: in, For the number of channels, The number of samples within the window. This represents the median of the standard deviations for each channel. 4) Combine the covariance matrix and the baseline bandwidth to construct a regularized diagonal bandwidth matrix. To control the local fitting range of kernel density estimation, the formula is as follows: in, Represents the covariance matrix The square root vector of the diagonal elements is used to convert the variance of each channel into a standard deviation scale; Used to control the global smoothing intensity; S3, Based on Regularized Diagonal Bandwidth Matrix Perform nonparametric kernel density estimation on the local window sample set to obtain the joint probability density function of the current window; S4. Generate a set of time-continuous joint probability density functions through overlapping windows iteratively, which serves as the basic sample set.

3. The method for condition monitoring and early warning of prefabricated substations based on artificial intelligence according to claim 1, characterized in that: The recursive Gaussian Copula model maps the joint distribution sample set to the standard normal margin through probability integral transformation, and calls the expectation-maximization algorithm to iteratively estimate the Gaussian Copula parameters on the mapped data with a fixed forgetting factor within a sliding time window. After each iteration, the Pearson correlation coefficient between channels is calculated based on the estimated covariance matrix. The main coupling direction vector and corresponding eigenvalues ​​are extracted from the covariance matrix by spectral decomposition. The dynamic Copula correlation matrix, which reflects the effects of transient impact, harmonic temperature coupling and current-temperature cross-correlation, is updated according to the timestamp.

4. The method for condition monitoring and early warning of prefabricated substations based on artificial intelligence according to claim 1, characterized in that: In step 3, the method for constructing the three-dimensional threshold surface family by quantile regression is as follows: a high-dimensional feature tensor is constructed in the spatiotemporal coordinate system based on the dynamic Copula correlation matrix. A quantile regression model is constructed by expanding the tensor B-spline basis function and combining it with the penalty term. An asymmetric weighted absolute bias loss is used to fit the 5%, 50%, and 95% quantile surfaces. Overfitting of the surface is eliminated by parameter smoothing constraints and entropy regularization. On the obtained normalized three-dimensional threshold surface family, the real-time coupling degree value is mapped to the three quantile surfaces of normal, early warning, and alarm respectively by the mapping method.

5. The method for condition monitoring and early warning of prefabricated substations based on artificial intelligence according to claim 1, characterized in that: In step 4, the variational autoencoder generates diverse synthetic abnormal samples based on the batch-normalized normal condition multidimensional feature vector and the corresponding condition label; the working principle of the variational autoencoder is as follows: In the encoder, the fused working condition label vector and the original features are mapped to the latent variable mean and variance parameters through a cascaded fully connected layer, and latent vectors are generated by sampling through a reparameterization method. In the decoder, temperature curves, pressure fluctuations, and discharge timing are reconstructed in parallel through multi-layer deconvolutional networks and skip connections. During the reconstruction process, a self-attention mechanism is used to focus on key channels. During training, a composite loss function that aligns the reconstruction error with the prior latent variables is used as the optimization objective. After each iteration, multi-scale interpolation sampling is performed based on the differences in the distribution of latent variables to generate diverse synthetic anomalous samples.

6. The method for condition monitoring and early warning of prefabricated substations based on artificial intelligence according to claim 1, characterized in that: In step 5, the process of the node and time-time attention mechanism extracting the switch cabinet state representation vector includes: A node adjacency matrix is ​​constructed based on the primary circuit connectivity of the prefabricated substation, and a time similarity matrix is ​​constructed based on the timestamp of the protection action log and the alarm record interval. The input node features and time-series embeddings are linearly mapped to generate query, key, and value vectors, respectively. Attention scores are calculated along the node dimension using node queries and node keys, and then normalized to obtain node attention weights. Attention scores are calculated along the time dimension using time queries and time keys, and then normalized to obtain time attention weights; The node attention weight is multiplied by the time attention weight to form a joint attention coefficient, and the node time value vector is weighted and aggregated using this coefficient. The weighted aggregation result is output as a sequence of switch cabinet state representation vectors through residual connection and layer normalization.

7. The method for condition monitoring and early warning of prefabricated substations based on artificial intelligence according to claim 1, characterized in that: The working process of the lightweight recursive pruning algorithm includes: For each representation vector, calculate the cosine similarity rate of change with its neighboring time vectors and generate a time importance score sequence; The scoring sequence is subsequenced based on a binary recursive partitioning strategy. When the length of the subsequence exceeds the set window, the partitioning is further refined. Within each subsequence, the scores at each time point are comprehensively evaluated based on the local magnitude range and the average gradient, and the node with the lowest contribution is identified as redundant using an adaptive threshold. After removing redundant nodes layer by layer from the inside out, the similarity change rate is recalculated and the time importance score sequence is updated. In the recursive pruning loop, the segmentation depth and threshold coefficient are dynamically adjusted by online information gain and sequence reconstruction error; The output contains only the state prediction sequence of the high-information-preserving nodes.

8. The method for condition monitoring and early warning of prefabricated substations based on artificial intelligence according to claim 1, characterized in that: The working process of step 7 includes: The vector at each time step in the lightweight state prediction sequence is projected onto a family of three-dimensional threshold surfaces. The perpendicular distance from the vector to the 5%, 50%, and 95% quantile surfaces is calculated using the least squares method and normalized to a likelihood probability vector. Based on the aforementioned likelihood probability vector and historical dynamic Copula correlation matrix, the busbar chamber temperature, switchgear status, and A multilayer Bayesian network is constructed on the nodes of the insulating structure, and a conditional probability table is defined. The multilayer Bayesian network is recursively updated using a particle filter algorithm, and a continuous anomaly score curve is generated by fusing multi-channel quantile likelihoods through Monte Carlo sampling. The first-order slope is extracted by applying the adaptive Gaussian kernel derivative operator to the continuous anomaly scoring curve, and the short-term fault propagation trend is determined by the length of the continuous positive slope segment and the radius of curvature threshold. The fault trend is subjected to a logical consistency check within a fixed-duration sliding window. If the logic is true for three consecutive windows, an early warning event is triggered. The warning instructions are encoded as SNMP Trap and Modbus write register instructions, and sent to the upper-level operation and maintenance system through the communication module of the box-type substation.

9. The method for condition monitoring and early warning of prefabricated substations based on artificial intelligence according to claim 8, characterized in that: The method for determining the short-term fault propagation trend in step 7 is as follows: For the continuous abnormal scoring curve, the Gaussian kernel bandwidth is first adaptively selected at each sampling point based on the scoring variance within the local window. The Gaussian kernel bandwidth is then used to perform Gaussian first derivative convolution operation on the scoring curve to obtain a point-by-point slope sequence. In the point-by-point slope sequence, continuous positive segments are identified using a sliding judgment algorithm, and the radius of curvature for each segment is calculated based on the second derivative. The calculation formula is as follows: in, This represents the value of the continuous anomaly score curve at time t; and These represent the first and second derivatives of the scoring curve with respect to time, respectively, and are used to reflect changes in slope and curvature. The incremental scaling factor represents the amplification of the effect on the radius of curvature when the slope of the scoring curve increases. When the length of any segment exceeds the preset time window and its minimum radius of curvature is lower than the preset threshold, it is considered a short-term fault propagation trend.

10. The method for condition monitoring and early warning of prefabricated substations based on artificial intelligence according to claim 1, characterized in that: The working method of step 8 is as follows: After obtaining the continuous anomaly score curve, a state-space model is constructed. The score curve value is used as the filtered observation, and the quantile residual at the previous time step is used as the filtered state variable. σ-point generation and state prediction and observation update are performed using unscented Kalman filtering. The threshold drift vector is extracted from the filtered state error covariance. Subsequently, based on the excess distance between the threshold drift vector and the original sample on each quantile plane, the partial derivative of the quantile regression loss function with respect to the tail dependency parameter is calculated using the finite difference method to obtain the gradient value. Finally, the tail dependency parameter is adjusted according to the gradient descent algorithm, and the updated parameter is used to reconstruct the third-order quantile surface in the tensor B-spline basis function, maintaining the threshold family and the busbar chamber of the box-type substation. Real-time coupling matching of air pressure and partial discharge characteristics.

Images