Early warning method for abnormal monitoring of power equipment

Abstract

The invention relates to an early warning method for abnormal monitoring of power equipment, and the method comprises the steps: integrating the operation state, type identification and connection topology of the equipment, dynamically constructing the interaction influence and topological evolution characteristics between the equipment, and fusing a graph neural network of an attention mechanism with a meta-learning framework, thereby achieving the early warning of the abnormal monitoring of the power equipment. And key node identification, abnormal propagation path mining and future abnormal trend scoring are realized. Meanwhile, through comparison of scores and historical data, efficient early warning can be achieved by self-adaption of a threshold value, and model parameters are optimized according to actual error feedback. According to the scheme, the accuracy rate, the environment adaptability and the model generalization ability of abnormal trend prediction among multiple devices in a complex power system are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of "intelligent monitoring of power equipment and multi-device topology evolution perception and anomaly trend prediction technology", and in particular to an early warning method for power equipment anomaly monitoring. Background Technology

[0002] With the continuous expansion and intelligent upgrading of power systems, multi-device anomaly monitoring and early warning have become fundamental to ensuring the safe and stable operation of the power grid. Current technological approaches primarily employ time-series analysis or static multi-device relationship modeling, combining equipment physical parameters, real-time operating status, and fixed topology information to predict anomaly trends in individual devices or a limited number of interconnected devices. Mainstream solutions include traditional statistical models (such as ARIMA and VAR), machine learning regression and classification algorithms, and the increasingly popular graph neural networks and deep time-series models. These technologies have achieved significant results in power equipment anomaly detection, single-device trend analysis, and risk assessment of anomaly propagation in some interconnected devices. However, with the increasing diversity of equipment types in power systems, frequent changes in operating environments, and the dynamic occurrence of equipment connection / deconnection events, existing technologies face new challenges.

[0003] Most existing technical solutions treat the relationships between devices as a fixed topology or static network structure. They analyze the spatial propagation trend of device anomalies based on static connectivity after collecting, normalizing, and denoising device operating parameters. This type of method is applied in substations, transmission lines, and local centralized control scenarios, and is suitable for power systems with relatively simple equipment types and relatively stable operating environments. However, facing the increasing heterogeneity of power system equipment, the rapid popularization of new intelligent data acquisition terminals, and the common occurrence of "dynamic device access—topology change—anomaly propagation path change," this static technical solution suffers from insufficient generalization ability and poor adaptability. On the one hand, the model parameters are tightly coupled with the topology structure; changes in equipment type or physical structure will lead to a significant decrease in model prediction performance. On the other hand, traditional methods often ignore the dynamic changes in the influence intensity between time-varying devices during the topology evolution process, making it difficult for the model to capture sudden changes in anomaly propagation paths caused by the access of new equipment or changes in key nodes. Summary of the Invention

[0004] This application provides an early warning method for monitoring abnormalities in power equipment, aiming to solve one of the problems or issues of the existing technology mentioned in the background section.

[0005] This application provides a method for early warning of power equipment anomalies, specifically including:

[0006] S1: Collect operating status data of multiple devices in the power system, including voltage, current, temperature, load rate, device type identification and connection topology between devices, to form a multi-dimensional device operating status time series dataset.

[0007] S2: Normalize and denoise the time series data of the multidimensional device operating status to eliminate the impact of sensor measurement errors and environmental noise on data quality.

[0008] S3: Based on the device type identifier and connection topology relationship, construct the device interaction influence matrix to describe the dynamic association strength between devices in the anomaly propagation path.

[0009] S4: The interaction influence matrix between the devices is incrementally updated using a sliding time window to dynamically capture the evolution process of device connection relationships and form a topological evolution feature sequence.

[0010] S5: Input the topology evolution feature sequence and the normalized equipment operation status data into the attention-based graph neural network model to identify key topology change nodes and abnormal propagation paths.

[0011] S6: Based on the anomaly propagation path and equipment status characteristics output by the graph neural network model, generate an equipment anomaly trend score to quantify the anomaly risk level of the equipment within a future time window.

[0012] S7: Based on device type identification and topology change characteristics, a lightweight meta-learning framework is constructed to enable the model to automatically adjust parameters to adapt to the new device combination structure when a small number of new devices are connected to the sample.

[0013] S8: Compare and analyze the abnormal trend score of the device with the historical score data to determine whether the current score exceeds the preset warning threshold. If it does, trigger the warning output.

[0014] S9: Based on the error between the actual early warning results and the model's predicted score, dynamically adjust the training weights of the graph neural network model and the meta-learning framework to improve the model's generalization ability under different device configurations.

[0015] The method for early warning of power equipment anomalies provided in this application has the following beneficial effects:

[0016] 1) Significantly enhances model generalization ability and improves adaptability to new devices and changing environments. Compared to traditional modeling methods that fix the relationships between devices as a static graph structure, this invention models the connections and interactions between devices as a time-varying structure through a dynamic topology evolution modeling mechanism. Utilizing a sliding time window and incremental update algorithm, the invention can promptly capture and express the dynamic evolution characteristics of device topology in actual operation, providing the model with accurate structural change information. Combined with a lightweight meta-learning framework, the model can quickly complete parameter adaptation with only a small amount of new device data, significantly improving the model's generalization ability and transferability when new devices are added or the operating environment changes, and significantly enhancing the system's coverage and applicability to complex real-world scenarios.

[0017] 2) The accuracy of anomaly propagation path identification is significantly improved. This is achieved by introducing graph neural networks based on attention mechanisms (such as...).

[0018] This invention (GAT) can adaptively identify key topology change nodes and abnormal propagation paths. By fusing multimodal inputs with device operating status and topology evolution characteristics, and in conjunction with a multi-head attention mechanism, the model's accuracy in identifying key nodes and abnormal paths is significantly improved. This more effectively prevents the failure to detect or misjudge abnormal propagation, enhances the reliability of abnormal trend prediction, and optimizes the system's early warning performance.

[0019] 3) Achieving high accuracy and stability in trend prediction models. The invention proposes using joint feature representation and GRU temporal modeling methods to perform fine-grained modeling of the temporal evolution trend of abnormal equipment states, and integrates multi-head attention and fully connected networks to output trend scores. Through normalization and adaptive calibration algorithms, the scoring bias problem caused by equipment heterogeneity is effectively solved, resulting in a more balanced score distribution. The combination of score analysis and dynamic threshold adjustment further improves the stability and sensitivity of abnormal trend scoring under signal and noise conditions, reducing false alarms and missed alarms. Attached Figure Description

[0020] Appendix Figure 1 This is the main flowchart of an early warning method for monitoring abnormalities in power equipment.

[0021] Appendix Figure 2 This is a sub-flowchart of an early warning method for monitoring abnormalities in power equipment.

[0022] Appendix Figure 3 This is another sub-flowchart of a method for early warning of abnormal monitoring of power equipment. Detailed Implementation

[0023] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0024] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed. In addition, examples of various specific processes and materials are provided in this invention, but those skilled in the art will recognize the application of other processes and / or the use of other materials.

[0025] As attached Figure 1 As shown, this application provides an early warning method for monitoring power equipment anomalies, specifically including:

[0026] S1: Collect operating status data of multiple devices in the power system, including voltage, current, temperature, load rate, device type identification and connection topology between devices, to form a multi-dimensional device operating status time series dataset.

[0027] S2: Normalize and denoise the time series data of the multidimensional device operating status to eliminate the impact of sensor measurement errors and environmental noise on data quality.

[0028] S3: Based on the device type identifier and connection topology relationship, construct the device interaction influence matrix to describe the dynamic association strength between devices in the anomaly propagation path.

[0029] S4: The interaction influence matrix between the devices is incrementally updated using a sliding time window to dynamically capture the evolution process of device connection relationships and form a topological evolution feature sequence.

[0030] S5: Input the topology evolution feature sequence and the normalized equipment operation status data into the attention-based graph neural network model to identify key topology change nodes and abnormal propagation paths.

[0031] S6: Based on the anomaly propagation path and equipment status characteristics output by the graph neural network model, generate an equipment anomaly trend score to quantify the anomaly risk level of the equipment within a future time window.

[0032] S7: Based on device type identification and topology change characteristics, a lightweight meta-learning framework is constructed to enable the model to automatically adjust parameters to adapt to the new device combination structure when a small number of new devices are connected to the sample.

[0033] S8: Compare and analyze the abnormal trend score of the device with the historical score data to determine whether the current score exceeds the preset warning threshold. If it does, trigger the warning output.

[0034] S9: Based on the error between the actual early warning results and the model's predicted score, dynamically adjust the training weights of the graph neural network model and the meta-learning framework to improve the model's generalization ability under different device configurations.

[0035] Step S1: Collect operating status data of multiple devices in the power system, including voltage, current, temperature, load rate, device type identification, and inter-device connection topology, to form a multi-dimensional device operating status time series dataset. Specifically, this includes:

[0036] S1.1: Perform data acquisition operations on multi-source sensing devices such as smart meters, temperature sensors, and load monitoring devices deployed in the power system to obtain real-time operating parameters such as voltage, current, temperature, and load rate.

[0037] In the operating environment of a power system, a multi-source sensing unit with smart meters, temperature sensors, and load monitoring devices as its core serves as the data acquisition endpoint. The input conditions are the voltage, current, temperature, and load rate signals output in real time by each acquisition node.

[0038] A periodic polling acquisition method is adopted (parameters: sampling period of 1 second, sampling accuracy of not less than 0.1 units) to achieve unified time base data acquisition for each sensing unit, ensuring the time alignment consistency of multi-dimensional operating parameters.

[0039] Furthermore, a high-fidelity conversion of analog signals to digital signals is achieved through an analog-to-digital conversion processing algorithm (parameters: 16-bit resolution, sampling rate satisfying the Nyquist sampling theorem), and quantized data sequences of voltage U, current I, temperature T, and load rate L are obtained.

[0040] Furthermore, the drift correction of the acquisition channel is achieved through a sensor self-calibration algorithm (parameter: zero drift compensation coefficient is calculated based on historical mean deviation), and a calibrated instantaneous operating parameter value vector is generated.

[0041] Furthermore, by adding timestamps (parameter: uniformly adopting the UTC standard time format), the data of each parameter is precisely bound to the system's global clock, resulting in an initial record set of running parameters with a time-series index.

[0042] Through the above multi-layer algorithm processing, the original sampled signal from the previous step is transformed into a multi-source operating state parameter matrix that has been quantized, corrected and has time labels, thereby achieving the unification and traceability of the collected data.

[0043] For example, in a power grid substation, 12 smart meters, 8 temperature sensors, and 5 load monitoring devices are installed. The sampling period is set to 1 second. Within each sampling period, the voltage range of 0 to 500 volts, the current range of 0 to 2000 amps, the temperature range of -40 to 120 degrees Celsius, and the load rate range of 0 to 150% are simultaneously collected. The collected signals are converted by a 16-bit A / D converter, with quantization step sizes of [missing information]. Fu, install, Celsius and Percentage. After correction using the zero-drift compensation algorithm, the voltage measurement deviation decreased from 0.8 volts to 0.05 volts, and the temperature deviation decreased from 0.6 degrees Celsius to 0.05 degrees Celsius. All collected data, after being appended with UTC timestamps accurate to the second, formed a 25×4 multidimensional operating parameter matrix, which served as the input baseline data for subsequent device type identification binding and topology relationship analysis. In a ten-hour operation test, this processing method achieved a data packet loss rate of less than 0.1%, significantly improving the integrity and timeliness of the early warning model training data.

[0044] S1.2: Based on the equipment type coding rules, the collected equipment operating parameters are attached with equipment type identifiers to distinguish different types of power equipment such as transformers, circuit breakers, and capacitors, thereby constructing equipment heterogeneity characteristic information.

[0045] Based on the collected, quantized, corrected, and time-stamped multi-source operating status parameter matrix, a device type coding rule mapping algorithm is adopted (parameters: type code length is 4 bits, coding space supports 2). 4 A unique identifier is used to attach a corresponding device type identifier vector to each row of operating parameter records.

[0046] Furthermore, a unique mapping relationship between device name and device type code is achieved through a hash mapping generation method (parameters: 256 hash buckets, open chaining for collision resolution), and a device identifier mapping table dataset is obtained.

[0047] Furthermore, a type verification algorithm based on regular expression matching (parameters: regular expression pattern matching length is 4, character set is [0-9A-F]) is used to verify the device type code format and generate a compliance verification result vector.

[0048] Furthermore, through an index merging processing algorithm (parameters: primary key index is the device unique number, foreign key index is the type code), the association and merging of the running parameter matrix and the type identifier mapping table are realized, and a multi-dimensional running status parameter matrix with a device type identifier field is generated.

[0049] Furthermore, by using a feature extension coding method (parameters: the category vector adopts One-Hot coding, and the coding dimension is equal to the number of device type categories), the device type identifier is transformed from a discrete symbol vector to a sparse feature vector, and a device heterogeneity feature matrix is ​​generated.

[0050] By using the above-mentioned equipment type encoding and feature construction processing methods, the time-tagged operating status data obtained in the previous step is transformed into a dataset that simultaneously contains the heterogeneous features of equipment types, thereby achieving the technical effect that different types of power equipment can be distinguished and aggregated in subsequent analysis.

[0051] For example, for the operating status parameter matrix of 25 acquisition nodes in a substation, a 4-digit hexadecimal type encoding rule is adopted, where transformer code is 1001, circuit breaker code is 1002, and capacitor code is 1003. A unique mapping relationship is established in the equipment identification mapping table. For capacitor records appearing in the original data, it is verified whether their type code format conforms to the pattern [0-9A-F]{4}. If the match is successful, it is marked as compliant. Through primary and foreign key index merging, the four columns of operating parameters (voltage U, current I, temperature T, and load rate L) are concatenated with the type code column to obtain a 5-column equipment operating status matrix. One-Hot encoding is used to map type code 1001 to a [1,0,0] vector, 1002 to a [0,1,0] vector, and 1003 to a [0,0,1] vector. The expanded matrix has 8 columns, realizing the explicit representation of type information in the model input vector. This verifies that it improves type discrimination and model convergence speed in topology analysis and abnormal pattern recognition. During the verification process, the type mapping and encoding processing took less than 0.5 seconds, and the type code misjudgment rate was 0.0%, fully meeting the real-time requirements of online monitoring.

[0052] S1.3: The connection topology data between devices is obtained synchronously through the SCADA system and the topology identification module. Based on the graph structure representation method, a device connection adjacency matrix is ​​generated to describe the physical connection and energy flow relationship between devices.

[0053] S1.4: Align the operating parameters with device type identifiers with the adjacency matrix using timestamps to generate structured device status data tuples with time indexes, thus forming a time-series dataset of device operating status.

[0054] S1.5: Perform data integrity verification and outlier removal on the time series dataset, and fill in missing data based on the sliding window mechanism to ensure the continuity and consistency of multi-dimensional device operation status data.

[0055] Step S2: Normalize and denoise the time series data of the multidimensional device operating status to eliminate the impact of sensor measurement errors and environmental noise on data quality. Specifically, this includes:

[0056] S2.1: Normalize the collected raw equipment operating status data such as voltage, current, temperature, and load rate. Use the min-max normalization method to unify the numerical range of each dimension feature to eliminate the influence of the difference in the dimensions of different sensors on subsequent modeling, and obtain the normalized multidimensional equipment operating status feature vector.

[0057] Based on the multidimensional device operation status time series data after type identification binding and timestamp alignment, the min-max normalization method (parameter: each feature is scaled independently, target range is [0,1]) is adopted to achieve a unified mapping of different physical quantities and eliminate the influence of feature magnitude differences on the model convergence speed and stability.

[0058] Furthermore, through a numerical scanning algorithm (parameter: floating-point precision 10), - Calculate the minimum and maximum values ​​of each feature dimension in the full sample set, and generate feature boundary vectors (min', max') to form the interval basis for normalization transformation.

[0059] Furthermore, for any eigenvalue x i The following standardized relation is used for mapping:

[0060]

[0061] Where min and max are the minimum and maximum values ​​of the feature within the acquisition period, respectively, x i x represents the original eigenvalues. i ' represents the normalized eigenvalues.

[0062] Furthermore, through vectorized batch processing, all feature dimensions are simultaneously mapped to the [0,1] interval to generate a normalized multidimensional device operating status feature matrix, while keeping the time index synchronized with the device type identifier.

[0063] Furthermore, a normalization verification algorithm (parameter: maximum allowable normalization error 1×10^{-8}) is used to check whether the mapped data is distributed within the target interval [0,1]. If out-of-bounds values ​​are found, boundary truncation correction is performed to ensure the numerical stability and usability of the normalized output.

[0064] By using a min-max normalization process, the original sensor data with heterogeneous dimensions is transformed into dimensionless feature vectors with uniform dimensions, thereby achieving a scaled representation of the multidimensional features of the equipment's operating status and providing numerically balanced input data for subsequent wavelet denoising and feature extraction stages.

[0065] For example, in a substation, the original range of voltage characteristics with a sampling period of 1 second is [210.5, 249.8] volts, the original range of current characteristics is [320, 865] amps, the temperature range is [32.6, 58.4] degrees Celsius, and the load rate range is [67.4, 132.9]%. When performing min-max normalization, the voltage of 230.2 volts is calculated as follows:

[0066]

[0067] Similarly, the normalized value for a current of 600 amps is 0.5196, the normalized value for a temperature of 40.0 degrees Celsius is 0.2857, and the normalized value for a load rate of 100.0% is 0.5315. After batch processing, all elements of the output matrix are within the range of [0,1], boundary detection is passed, and the values ​​are stable, effectively improving the subsequent S2.2 wavelet threshold denoising algorithm's ability to suppress multi-dimensional feature collaborative noise.

[0068] S2.2: Based on the normalized multidimensional equipment operating state feature vector, a wavelet threshold denoising algorithm is used to filter the time series data to suppress sensor measurement errors and transient environmental noise interference, and to obtain the denoised equipment state time series after wavelet coefficient reconstruction.

[0069] S2.3: Perform sliding window segmentation processing on the denoised equipment state time series after wavelet coefficient reconstruction, extract statistical features within each time window, including mean, variance, skewness and kurtosis, to enhance the local dynamic representation capability of equipment operating status and generate a time window statistical feature sequence.

[0070] S2.4: Based on the statistical feature sequence of the time window, principal component analysis (PCA) is used to reduce the dimensionality of the features, extract the principal component features whose variance contribution rate exceeds the preset threshold, so as to reduce feature redundancy and retain the status information of key equipment, and generate a low-dimensional feature representation matrix.

[0071] S2.5: Standardize the low-dimensional feature representation matrix so that each principal component feature satisfies a zero-mean unit variance distribution to adapt to the input requirements of the subsequent graph neural network model and obtain the standardized equipment operating status feature input tensor.

[0072] Step S3: Based on the device type identifier and connection topology, construct an inter-device interaction influence matrix to describe the dynamic association strength between devices in the anomaly propagation path. For example... Figure 2 As shown, it specifically includes:

[0073] S3.1: Perform graph structure modeling processing on the equipment type identification information and physical connection topology data to generate an initial graph structure composed of equipment nodes and edge relationships, which serves as the basic topology framework for constructing the interaction influence matrix.

[0074] S3.2: Based on the device type identification information, the device similarity measurement algorithm is used to calculate the functional coupling degree between different device nodes to obtain the device functional similarity weight matrix, which serves as the static attribute input of the interaction influence matrix.

[0075] Based on the input device type identification information, a device function similarity measurement algorithm is adopted (parameters: weight combination coefficient is set to 0.5 static + 0.5 structural, type feature space dimension is n). t This allows for the preliminary calculation of the functional coupling degree between different device nodes.

[0076] Furthermore, by using a type encoding vectorization processing method (parameter: one-hot encoding dimension equal to the total number of device types N), a dense vector representation of the device type identifier is achieved, and the device type feature matrix is ​​obtained.

[0077] Furthermore, a cosine similarity calculation method is adopted (parameter: vector norm ε = 1 × 10^-12 to prevent zero vector anomalies) to calculate the cosine of the angle between the functional features of the equipment and generate the basic similarity matrix C between equipment types, as shown in the following formula:

[0078]

[0079] Where V(i) is the type feature vector of device i, and ε is a small constant to prevent the denominator from being zero.

[0080] Furthermore, this is combined with a domain prior rule database (parameters: total number of rule entries R, rule matching threshold θ). r =0.7), using a rule-based matching scoring algorithm, the values ​​in the similarity matrix C are non-linearly weighted according to the historical coupling relationship of device functions to generate a corrected functional similarity matrix F, the formula of which is:

[0081] F i,j =S i,j ×W i,j

[0082] Among them, W i,j Weight the score for rule matching.

[0083] Furthermore, a Gaussian radial basis function (parameter: bandwidth σ = 1.0) is used to perform kernel mapping smoothing on the functional similarity matrix F, making the weights of high-similarity device pairs more concentrated at 1, and the weights of low-similarity pairs closer to 0, thereby generating the device functional similarity weight matrix G, with the formula as follows:

[0084]

[0085] Through the above process of calculating the weight of device function similarity, the device type feature data in the previous step is transformed into static attribute inputs in the interaction influence matrix, thereby realizing the quantification of functional coupling relationship on static topology and providing a foundation for dynamic state correlation fusion.

[0086] For example, in a power system containing five types of equipment (transformers, circuit breakers, capacitors, instrument transformers, and voltage transformers), the equipment type coding dimension is set to 5, and the generated type feature matrix is ​​M×5 (M is the total number of equipment, 20). For transformers and instrument transformers, the cosine similarity calculated by one-hot coding is 0.6. A priori rule of "transformer-instrument transformer commonly connected in series" is matched in the rule base, with a rule score weight of 0.85, and the similarity is corrected to 0.51. After Gaussian radial basis kernel mapping, the weight value is 0.78, ultimately forming the static functional weights at the corresponding positions in the G matrix. This achieves the goal of quantifying the degree of functional coupling between different types of equipment, and significantly improves the accuracy of topology evolution perception between equipment when subsequently fused with the dynamic correlation matrix. In this process, for different equipment pairings, if no rule weight is matched, the original cosine similarity is maintained. Gaussian smoothing suppresses noisy weak correlations, effectively reducing the risk of misjudgment. The final generated G matrix can be directly input into the subsequent weighted fusion algorithm and combined with the dynamic matrix to obtain the comprehensive interaction influence matrix.

[0087] S3.3: Perform correlation analysis on the normalized equipment operating status data, and use the Pearson correlation coefficient to calculate the dynamic state correlation coefficient between equipment to form a dynamic interaction strength matrix, which serves as the time-varying attribute input of the interaction influence matrix.

[0088] Based on the equipment operating status feature input tensor after S2.5 standardization, the Pearson correlation coefficient calculation method (parameters: bivariate zero-mean processing, correlation coefficient range [-1,1]) is used to realize the dynamic state correlation measurement between pairs of equipment.

[0089] Furthermore, using a time series segmentation processing method (parameter: window length L is consistent with S2.3), the device operating status feature vector X is extracted pairwise within each time window. it With X jtIt also performs mean centering operation to generate a mean-free feature sequence to eliminate the influence of dimensional and mean shift on correlation calculation.

[0090] Furthermore, using the Pearson correlation coefficient formula, the dynamic state correlation coefficient r between device i and device j within the current time window k is calculated. ij (k):

[0091]

[0092] Where, μ i With μ j These are the average operating status characteristics of devices i and j within this window, respectively.

[0093] Furthermore, an M×M dynamic state correlation matrix R is generated through matrix operations. k The row and column indices correspond to different devices, and the matrix elements are the correlation coefficient values ​​of the corresponding device pairs, so as to achieve the overall quantification of the correlation of high-dimensional multi-device states.

[0094] Furthermore, a sliding window aggregation algorithm (parameter: window overlap rate of 50%) is used to perform time smoothing on the dynamic state correlation matrix of continuous time windows, generating a time-varying smooth matrix sequence to suppress the interference of transient spikes on dynamic attribute inputs.

[0095] Through the above Pearson correlation calculation and time smoothing process, the normalized equipment status data is transformed into a dynamic interaction intensity matrix that reflects the linkage between the operating status of equipment, providing a mathematically accurate quantitative basis for the time-varying attribute input of the interaction influence matrix.

[0096] For example, in a monitoring scenario involving three key pieces of equipment—transformer T1, circuit breaker B2, and capacitor C3—the sampling period is 1 second, and the sliding window length is set to 60 seconds. Within a certain time window, the average voltage of T1 is 0.65, the average current of B2 is 0.72, and the average temperature of C3 is 0.48. Taking T1 and B2 as examples, their mean vectors are [-0.10, 0.05,] and [0.08, -0.12,] respectively. Substituting these values ​​into the above formula yields a correlation coefficient of 0.82, indicating a significant positive correlation between their operating states within the current window. As the sliding window progresses over time, each window generates an R matrix. After smoothing with a 50% overlap rate, the resulting dynamic state correlation sequence effectively reveals the state coupling change trajectory of T1 and B2 at different operating stages in anomaly propagation analysis. When fused with the static functional similarity weight matrix, it provides a high-confidence time-varying correlation input for generating the comprehensive interaction influence matrix in S3.4.

[0097] S3.4: Based on the static equipment functional similarity weight matrix and the time-varying equipment state correlation coefficient matrix, a weighted fusion algorithm is used to perform fusion processing to generate a comprehensive equipment interaction influence matrix, which characterizes the potential strength of abnormal propagation paths between equipment.

[0098] Based on the static device function similarity weight matrix G output by S3.2 and the dynamic state correlation matrix R generated by S3.3, a weighted fusion algorithm is adopted (parameters: both static weight coefficient α and dynamic weight coefficient β are set as adjustable variables, satisfying α+β=1) to achieve unified quantification of static and dynamic correlation strength.

[0099] Furthermore, by using a matrix element weighted fusion calculation method, the static weight values ​​at corresponding positions are linearly combined with the dynamic correlation coefficients to generate a comprehensive interaction strength matrix M, as shown in the following formula:

[0100] M i,j =α×G i,j +β×R i,j

[0101] Among them, M i,j This represents the combined interactive influence weight between device i and device j.

[0102] Furthermore, a weight normalization method is adopted (parameter: L1 norm normalization is performed for each row) to ensure that the sum of the comprehensive interaction weights of each device node is 1, so as to avoid model bias caused by differences in units.

[0103] Furthermore, to suppress the interference of abnormal spike noise on the fusion results, a bilateral filter (parameters: spatial domain standard deviation σ = 0.5, value domain standard deviation σ = 0.1) is used to smooth matrix M while maintaining matrix symmetry, so as to reflect the interaction consistency under undirected topological conditions.

[0104] Furthermore, the confidence index of each matrix element (defined as the product of the consistency between static and dynamic values) is calculated, and low-confidence elements are marked for removal in the sparsity processing stage of S3.5.

[0105] By weighted fusion and confidence calculation, the static functional coupling data and dynamic operating status correlation data of the first two sub-steps are transformed into a comprehensive interaction influence matrix that characterizes the potential strength of abnormal propagation paths between devices, providing multi-dimensional consistent input features for subsequent topology evolution modeling.

[0106] For example, in a power system containing 5 devices (T1, T2, B1, C1, M1), the T1-T2 corresponding elements of the static functional similarity matrix G are 0.78, and the T1-T2 corresponding elements of the dynamic correlation matrix R in the current window are 0.82. Setting α = 0.6 and β = 0.4, the values ​​are substituted into the fusion formula for calculation:

[0107] M(T1,T2)=0.6×0.78+0.4×0.82=0.468+0.328=0.796

[0108] After L1 normalization, the value is adjusted to 0.312, and the confidence level after bilateral filtering is 0.64, which is higher than the preset confidence threshold of 0.5. Therefore, this edge is retained. After the above processing, the fusion matrix not only preserves the main propagation path with high confidence, but also improves the dual adaptability of the topological description to mutation and persistence features, effectively reducing redundant calculations in subsequent topological evolution analysis.

[0109] S3.5: Perform sparsification on the integrated equipment interaction influence matrix, and use a threshold filtering algorithm to remove low-weight edge connections in order to optimize the graph structure complexity and highlight key anomaly propagation paths, forming a sparse interaction influence matrix for subsequent topology evolution modeling.

[0110] Step S4: Incrementally update the inter-device interaction influence matrix using a sliding time window to dynamically capture the evolution of device connection relationships and form a topological evolution feature sequence. For example... Figure 3 As shown, it specifically includes:

[0111] S4.1: Perform structured encoding on the connection topology between devices and the device type identifier to obtain the initial connection topology vector representation between devices, which serves as the input basis for sliding time window processing.

[0112] Given that the comprehensive interaction influence matrix after sparsification and the device type identification data are available, a structured feature encoding method (parameter: the node feature dimension is set to the concatenation length of the device type encoding dimension d and the connection feature dimension l) is used to realize the vectorized representation of the connection topology between devices.

[0113] Furthermore, through adjacency matrix densification and index mapping algorithms (parameters: densification fill value is set to 0, index mapping rules are based on device unique ID hash function), the physical connection relationships are uniformly aligned in the vector coordinate space, and the initial connection topology basis matrix A0 is obtained.

[0114] Furthermore, for the device type identification data, a one-hot encoding method is adopted (parameter: total number of categories is N). type Map each device type to a string of length N. type The dense vectors are used to form the device type feature matrix T.

[0115] Furthermore, a feature concatenation operator (parameter: row vectors in column concatenation order [T, A0]) is used to merge the device type features with the corresponding device's connection weights in the initial topology basis matrix to generate the device's initial topology vector v.i .

[0116] Furthermore, using a vector normalization algorithm (parameter: L2 norm normalization is used to avoid inconsistent feature scaling caused by differences in connectivity), the standardized result u of the initial topology vector for each device is calculated. i The formula is as follows:

[0117]

[0118] Where L is the vector length, v i [k] represents the value of device i in the k-th feature dimension.

[0119] Through the above encoding and normalization processes, the connection topology and type characteristics between devices are fused into an initial connection topology vector set in a unified numerical space, thereby standardizing the input format for subsequent sliding time window processing.

[0120] For example, in a power system containing four types of equipment (transformers, circuit breakers, capacitors, and voltage transformers), N type The value is 4, and the total number of nodes is 10. The adjacency matrix A0 is 10×10 dimensional. The type vector of transformer T1 is given, and its corresponding row connection weight is given. The feature concatenation operator is used to concatenate the type vector and connection weights into an initial topological vector v of length 14. T1 After L2 norm normalization, u is obtained. T1 The resulting standardized topology vector set of 10 devices serves as direct input for the S4.2 sliding window construction, achieving stable feature representation and comparability under varying device connection scales, effectively supporting the robustness of subsequent topology evolution feature extraction.

[0121] S4.2: Construct a sliding time window based on historical device connection status data, and set the window length and sliding step parameters to achieve local time window coverage of dynamic changes in device connection relationships.

[0122] Based on the standardized initial connection topology vector set obtained by S4.1, a sliding window generation algorithm (parameters: window length L and sliding step size S are independently adjustable) is used to achieve temporal local coverage of historical device connection status data.

[0123] Furthermore, a window index mapping mechanism is used (parameter: using timestamp hash as the first index of the window, mapping rule is as follows). This ensures that the translation of the sliding window on the time axis satisfies the requirements of coverage integrity and no omissions.

[0124] Furthermore, a window data slicing method is employed (parameters: index range [idx×S, idx×S+L-1]) to extract the corresponding initial connection topology vector subset within each window, generating a local topology sample matrix W. k .

[0125] Furthermore, a window boundary alignment algorithm (parameters: boundary zero-padding strategy or boundary copying strategy are optional) is used to pad the length of the starting or ending window samples that are less than L, ensuring the consistency of the window matrix size.

[0126] Furthermore, a window overlap rate control strategy is implemented (parameter: overlap rate r is defined as...). The combined values ​​of L and S are dynamically adjusted to optimize the balance between the temporal resolution and computational load in capturing topology changes.

[0127] By constructing a sliding time window, historical device connection status data is divided into multiple local matrix sequences with time order, thereby achieving local time window coverage of dynamic changes in device connection relationships and providing a strictly time-divided input benchmark for the detection of connection status change events in S4.3.

[0128] For example, in a power system containing 20 devices, the monitoring cycle is 1 second, L is 60, S is 30, and the overlap rate is... Within a window of 1 to 60 seconds, a 20×14 matrix W1 is generated based on the collected standardized initial connectivity topology vectors. Within a window of 31 to 90 seconds, W2 is generated. The two matrices have 50% sample overlap in adjacent portions, ensuring a smooth transition in topology evolution. In the last 60 seconds of data, a boundary replication strategy is used to pad the final vector configuration to the full window length, ultimately forming a sequence of K window matrices. This provides high-resolution, structurally uniform time-slice data input for subsequent dynamic topology evolution modeling. In actual operation, this configuration enabled the capture of anomalous connectivity mutations within 30 seconds, while reducing computational load by approximately 48% compared to a full-coverage scan, thus improving real-time processing efficiency.

[0129] S4.3: Detect events that change the connection status between devices within a sliding time window, and generate a sequence of topology change events based on events such as connection establishment, disconnection, and weight changes, as an incremental input signal for topology evolution modeling.

[0130] The local topological sample matrix W within the window output by S4.2 k As input, a differential comparison algorithm (parameter: using an absolute difference threshold τ to eliminate small fluctuations) is used to perform element-wise comparison of the topology vectors at the beginning and end of the window to detect changes in connection state.

[0131] Furthermore, an event classification rule set (parameters: categories include connection establishment, connection termination, and weight change; a secondary threshold τ is set for the weight change category) is defined. w (Used to distinguish between significant changes and minor variations), it classifies the event types of the detection results and generates an event label matrix E. k .

[0132] Furthermore, using an event weight calculation method (parameter: weight is defined as the product of the magnitude of change and the static functional similarity weight), the event weight value of a single edge is calculated using the following formula:

[0133] ω i,j =|ΔA i,j |×G i,j

[0134] Where, ΔA i,j G is the difference between the first and last adjacency matrices of the window. i,j This represents the static functional similarity weight.

[0135] Furthermore, a time-series event aggregation algorithm is used (parameter: time merging threshold τ). t Set to twice the window step size, merge events that occur at similar times and on the same node to reduce event fragmentation and build a structurally complete topology change event sequence S. k .

[0136] Furthermore, a sequence sorting and index mapping algorithm is adopted (parameters: priority is sorted by event type and event weight, and the same category is arranged according to time sequence), for S k Each event is assigned a unique event ID and sorting index so that it can be directly referenced in subsequent S4.4 incremental graph update steps.

[0137] Through the above differential detection, event classification, weight calculation, temporal aggregation and index mapping processing, the local topological state differences within the window are transformed into a structured and quantifiable sequence of topological change events, achieving the goal of dynamically capturing the evolution of device connection relationships and providing accurate incremental input signals for incrementally updating the dynamic interaction influence matrix.

[0138] For example, in a power system containing 10 devices, the sliding window length L = 60 seconds, the step size S = 30 seconds, the absolute difference threshold τ = 0.05, and the weight change threshold τ w =0.1. In window 5, the connection weight between nodes T1 and B2 changes from 0.35 to 0.62, with a difference ΔA. T1,B2 =0.27, higher than τ w This is determined to be a significant weight change event; the corresponding static functional similarity weight G T1,B2 =0.81, substitute into the formula to calculate the event weight ω T1,B2=0.2187. During the timing aggregation phase, if the time interval between this event and another connection establishment event of T1-B2 within the adjacent window is less than 60 seconds (τ... t If the events are multiple, they are merged into a single multi-attribute event record. After index mapping, an event ID EID_045 is generated and stored in S. w5 In practice, this method has been verified to quickly generate detection results even in scenarios where connection weights gradually change but may still lead to changes in abnormal propagation paths. It reduces computation time by approximately 45% compared to the full-scan comparison method, while maintaining an event detection accuracy of over 96%.

[0139] S4.4: An incremental graph update algorithm is used to dynamically reconstruct the interaction influence matrix between devices within the window. The edge weights in the adjacency matrix are updated based on the topology change event sequence to generate the dynamic interaction influence matrix under the current time window.

[0140] S4.5: Time series splicing of the dynamic interaction influence matrix output by the continuous sliding window is performed to form a device topology evolution feature sequence, which is used as the dynamic topology perception input for the subsequent graph neural network model.

[0141] Step S5: Input the topology evolution feature sequence and the normalized equipment operating status data into a graph neural network model based on an attention mechanism to identify key topology change nodes and anomaly propagation paths. Specifically, this includes:

[0142] S5.1: Perform feature concatenation processing on the normalized equipment operation status data and topological evolution feature sequence to construct a unified multimodal input tensor for subsequent graph neural network modeling.

[0143] S5.2: Construct a dynamic adjacency matrix based on the connection topology between devices. This adjacency matrix is ​​updated in real time according to the topology evolution feature sequence, reflecting the time-varying characteristics of the interaction influence between devices.

[0144] Based on the topology evolution feature sequence and the connection topology between devices, a dynamic adjacency matrix construction algorithm (parameters: total number of nodes N, initial adjacency matrix A0) is used to realize time-varying correlation modeling of network structure.

[0145] Furthermore, the adjacency matrix is ​​updated using a function (parameter: the update rule is based on the topological evolution feature sequence F). t The edge weight changes in the matrix A (from the previous time window) will be used to determine the adjacency matrix A. t-1 The change matrix ΔA corresponding to the time step of the feature sequence t Perform element-wise weighted summation to obtain the adjacency matrix A at the current time step. t The calculation formula is:

[0146] A t =At-1 +ΔA t

[0147] In this matrix, the elements correspond to the edge weights between devices; positive changes indicate enhanced connectivity, while negative changes indicate weakened connectivity.

[0148] Furthermore, an edge-weighted smoothing filtering algorithm (parameter: smoothing coefficient α ranges from (0,1)) is used to perform an exponentially weighted moving average on the updated adjacency matrix to suppress occasional noise fluctuations, as shown in the following formula:

[0149] A t =αA t +(1-α)A t-1

[0150] Furthermore, by performing normalization processing on the smoothed adjacency matrix (parameter: using symmetric normalization), the calculation formula is as follows:

[0151] A t' =D(A t ) -1 / 2 A t D(A t ) -1 / 2

[0152] Where D is the degree matrix, and the diagonal elements are the connectivity degrees of the corresponding nodes.

[0153] Furthermore, by utilizing a time index mapping mechanism (parameter: time step synchronized with topology changes), the normalized adjacency matrix at each time step is bound to the multimodal input tensor at the corresponding time step, forming a dynamic graph structure input dataset.

[0154] By using a dynamic adjacency matrix real-time update algorithm, the topological evolution feature sequence is transformed into a set of adjacency matrices with time-varying characteristics, thereby enabling graph neural networks to dynamically perceive device interaction relationships.

[0155] For example, in a power system with a total number of nodes N = 10, the initial adjacency matrix A0 is a symmetric matrix with edge weights ranging from 0 to 1 and a smoothing coefficient α = 0.7. The change matrix ΔA at a certain time t... t The elements at nodes 1 and 3 are 0.2, the elements at nodes 4 and 5 are -0.15, and all other elements are 0. Substituting these values ​​into the update formula yields the unsmoothed A. tSubsequently, an exponentially weighted moving average was applied, increasing the edge weight of nodes 1-3 from 0.6 to 0.74, and decreasing the edge weight of nodes 4-5 from 0.5 to 0.395. After symmetric normalization, the node degree changes were mapped to standardized weight ratios, ensuring consistency of feature dimensions across networks of different sizes. The resulting normalized adjacency matrix was then matched and bound to the device state feature vectors at corresponding times, and input into the GAT model, enabling real-time topology updates and accurate identification of key nodes in scenarios with new device access.

[0156] S5.3: Input the multimodal input tensor and dynamic adjacency matrix into the graph attention network (GAT), and calculate the importance weight of each node in the current topology through the multi-head attention mechanism to identify key topology change nodes.

[0157] Based on the unified multimodal input tensor and dynamic adjacency matrix generated in the previous step, the multi-head attention mechanism of Graph Attention Network (GAT) (parameters: number of attention heads H, output feature dimension d of each head, adjacency sparsity threshold τ) is used to calculate the importance weight of nodes under the current topology.

[0158] Furthermore, through a linear feature transformation function (parameter: weight matrix W) h Given a dimension of d×F (where F is the input feature dimension), the input feature vector of each node is projected onto the multi-head attention space to obtain multiple sets of feature subspace representations x. i,h .

[0159] Furthermore, an attention coefficient calculation function is used (parameter: learnable parameter vector a). h The dimension is 2×d, the activation function is LeakyReLU, and the negative slope α=0.2). The unnormalized attention score e is calculated based on the feature representations of node i and its neighbor j. i,j,h :

[0160] e i,j,h =LeakyReLU(a h ·[x i,h ||x j,h ])

[0161] Where, x i,h ||x j,h This represents the vector concatenation operation.

[0162] Furthermore, a masked Softmax normalization algorithm is used (parameter: the mask matrix is ​​based on the current dynamic adjacency matrix A). t (Assigning -∞ to non-neighbor locations) and the unnormalized coefficient e i,j,h Transform into normalized attention weights α i,j,h :

[0163]

[0164] Furthermore, the feature vectors of each neighbor node are weighted and summed according to the normalized attention weights, and then concatenated or averaged among the multiple heads to obtain the comprehensive significance representation z. i .

[0165] Through the above processing method, the time-varying features of multimodal input and dynamic graph structure are mapped into a set of node-level importance weight vectors, thereby supporting the accurate identification of key topology change nodes.

[0166] For example, in an instance where the total number of nodes N=8, the input feature dimension F=16, and the number of attention heads H=4, the weight matrix W for each attention head is... h The dimension configuration is 8×16, and after linear transformation, each node obtains a feature vector of length 8. LeakyReLU activation (negative slope 0.2) is used to calculate the unnormalized score. For example, the score of nodes 1 and 3 in the first attention head is 0.85, and after Softmax normalization, the corresponding weight is 0.41. Summarizing the results of the four attention heads, the final comprehensive significance score of node 3 is 0.78, which is higher than the set key node threshold of 0.7, and it is determined to be a key topology change node. This method achieves a key node identification accuracy of over 95% in dynamic topology scenarios, improving by approximately 7 percentage points compared to the single-head attention model, and maintains stability under conditions of sudden changes in node load.

[0167] S5.4: Based on the node importance weights output by the graph attention network, and combined with the historical patterns of anomaly propagation paths between devices, a path weighted aggregation algorithm is used to extract the key path features of the anomaly propagation paths.

[0168] S5.5: Perform nonlinear transformation and normalization on the key path features to generate a node-level anomaly propagation intensity vector, which serves as the input feature for the subsequent anomaly trend scoring module.

[0169] Step S6: Based on the anomaly propagation path and device status characteristics output by the graph neural network model, generate a device anomaly trend score to quantify the anomaly risk level of the device within a future time window. Specifically, this includes:

[0170] S6.1: The anomaly propagation path feature vector and the device state feature vector output by the graph neural network model are concatenated and fused to construct a joint feature representation that includes topological evolution information and device operating status.

[0171] S6.2: Based on the joint feature representation, a gated recurrent unit (GRU) is used to model the feature evolution in the time dimension in order to extract the evolution trend features of the abnormal state of the equipment in the time series.

[0172] Based on the joint feature representation tensor of the input, a gated recurrent unit (GRU) network (parameters: hidden state dimension H, time step length T, number of layers L) is used to model the temporal dependency of the feature sequence in the time dimension.

[0173] Furthermore, by setting the update gate z of the GRU t Calculation function (parameter: weight matrix W) z U z With bias vector b z The activation function is Sigmoid, which adaptively controls the information retention ratio between the current input features and the historical hidden states. Its calculation formula is:

[0174] z t =σ(W z x t +U z h t-1 +b z )

[0175] Furthermore, by resetting the gate r t Calculation (parameter: weight matrix W) r U r With bias vector b r The activation function is Sigmoid, which adjusts the degree to which the current input forgets the historical hidden state. Its calculation formula is:

[0176] r t =σ(W r x t +U r h t-1 +b r )

[0177] Furthermore, according to the candidate hidden state calculation formula (parameter: weight matrix W) h U h With bias vector b h The activation function is tanh. The current candidate hidden state is generated by combining the results of the reset gate's filtering of the previous hidden state. The calculation formula is:

[0178]

[0179] Furthermore, by updating the gate and combining it with the weighted combination of the previous hidden state and the current candidate hidden state, the hidden state h at the current time step is obtained. t The calculation formula is:

[0180]

[0181] Furthermore, by iterating the above calculation process along the time axis, a GRU loop is executed on the device anomaly feature sequence of length T to obtain a hidden state sequence H containing time evolution information. seq .

[0182] Through dynamic adjustment of the gating mechanism, this step effectively transmits the topology and equipment operating status association information from the previous moment to the current moment, while suppressing irrelevant disturbances, thereby capturing and smoothly modeling the dynamic characteristics of abnormal states.

[0183] For example, in a power equipment anomaly monitoring instance with 10 nodes, a time step of 12, and a hidden state dimension of 64, the dimension of the input joint feature tensor is 12×10×128, and the GRU parameters are initialized using a normal distribution with a mean of 0 and a standard deviation of 0.05. The average value of the update gate under steady-state conditions is 0.72, indicating a high retention rate of historical information. The average value of the reset gate decreases to 0.25 at the topology abrupt change time step, achieving rapid decay of dependence on the old structure. After 12 iterations, the hidden state sequence exhibits orderly changes before and after the abrupt change point. The highlighted positions correspond to the significant impact of key topology changes on the equipment state. This hidden state sequence will serve as the input for the subsequent S6.3 multi-head attention mechanism, providing temporal evolution feature support for anomaly trend prediction.

[0184] S6.3: Perform multi-head attention mechanism calculation on the evolution trend features to identify the device nodes and their time steps that have a key impact during the anomaly propagation process, and generate weighted high-order anomaly propagation impact features.

[0185] S6.4: Based on the aforementioned high-order anomaly propagation impact characteristics, a fully connected neural network is used for nonlinear mapping processing to generate a preliminary score value for the device anomaly trend. This score value reflects the probability intensity of the device experiencing anomalies within a future time window.

[0186] S6.5: The preliminary score is normalized and calibrated by combining the equipment type identifier and historical score distribution information to eliminate equipment heterogeneity and score bias issues, and generate the final equipment anomaly trend score.

[0187] Step S7: Based on device type identification and topology change characteristics, a lightweight meta-learning framework is constructed to enable the model to automatically adjust parameters to adapt to the new device combination structure when a small number of new devices are added to the sample. Specifically, this includes:

[0188] S7.1: Extract and structure the features of equipment type identifiers, topology change characteristics between equipment and historical equipment operating status data to generate equipment context embedding vectors.

[0189] S7.2: Based on the device context embedding vector, construct a lightweight meta-learning parameter initialization network, and use a task-aware parameter generation mechanism to generate an initial model parameter set for newly accessed devices.

[0190] S7.3: Perform a fast inner-layer optimization based on gradient descent on a small amount of operational status sample data collected after the new device is connected, in order to obtain local parameter update results that adapt to the current device combination structure.

[0191] For a small amount of operational status sample data after the new equipment is connected, a batch gradient descent optimization method (parameters: learning rate η, batch size B, number of iterations N) is used to achieve rapid inner-layer optimization of model parameters for the current equipment combination structure.

[0192] Furthermore, the mean squared error loss function L between the model's predicted output and the true label is calculated, where...

[0193] To achieve quantitative assessment of the prediction error for this batch of samples.

[0194] Furthermore, based on the backpropagation algorithm (parameters: gradient decay factor β, optimizer type Adam), the gradient vector of the loss function with respect to each parameter θ of the model is calculated. It also generates gradient adjustment terms for parameter updates.

[0195] Furthermore, the formula is updated using gradient descent.

[0196]

[0197] The initial model parameters are rapidly optimized in one or more steps to obtain local parameter update values ​​θ that are adapted to the new device access conditions. ' .

[0198] Furthermore, for a very small number of training samples, a gradient accumulation strategy (parameter: accumulation step size K) is adopted to balance the bias and variance of parameter updates, so as to improve the stability of inner layer optimization and reduce the adverse effects on outer layer generalization performance.

[0199] Through the aforementioned rapid inner-layer optimization process, a small number of operational state samples after the new device is connected are transformed into a local optimization parameter set compatible with the current device combination structure, providing adaptive optimization support for subsequent outer-layer parameter updates, and achieving rapid convergence and performance maintenance of the model under new topology changes.

[0200] For example, in a power equipment combination consisting of 10 transformers and 5 circuit breakers, 2 new transformers were added. 50 operational status samples of each new device were collected over 5 minutes, with a feature dimension of 128. The batch size B = 10, learning rate η = 0.01, and number of iterations N = 5. The Adam optimizer (β1 = 0.9, β2 = 0.999) was used for fast inner-layer optimization. The mean squared error decreased by 15% after a single iteration. After 5 updates, the prediction accuracy improved by 8.3% compared to the unoptimized model, verifying the effectiveness of this inner-layer optimization in improving the adaptability of the new topology under low-sample conditions.

[0201] S7.4: Based on the local parameter update results and model prediction error, perform outer layer parameter update based on higher-order derivatives to optimize the generalization performance of the meta-learning parameter initialization network.

[0202] S7.5: Initialize the network output with the optimized meta-learning parameters and output it to the graph neural network model, so that the model can quickly adapt to the new device combination structure, thereby improving the accuracy and stability of abnormal trend prediction.

[0203] Step S8: Compare and analyze the device abnormal trend score with historical score data to determine whether the current score exceeds a preset warning threshold. If it does, trigger a warning output. Specifically, this includes:

[0204] S8.1: Perform normalization and alignment processing on the abnormal trend score of the equipment and the historical score data of the corresponding equipment to eliminate the influence of the difference in scoring scale on the comparison results and obtain the normalized sequence of equipment trend scores for subsequent threshold judgment.

[0205] S8.2: Based on the normalized sequence of equipment trend scores, a sliding window is used to calculate the deviation between the current score and the historical score window mean, so as to quantify the degree of change in the equipment operating status and generate a score deviation feature vector.

[0206] S8.3: Input the scoring deviation feature vector into the dynamic threshold adjustment model, combine the equipment type identification and operating condition information, calculate the dynamic warning threshold of the current equipment under the current score, and generate adaptive threshold parameters.

[0207] The scoring deviation feature vector is input into the dynamic threshold adjustment model, and an adaptive regression method based on multi-dimensional feature fusion (parameters: feature weight vector w, regularization coefficient λ) is used to realize the dynamic modeling of scoring threshold calculation under different equipment operating conditions.

[0208] Furthermore, by performing category encoding on the scoring deviation feature vector and the corresponding device type identifier, a low-dimensional device category feature vector is generated using one-hot encoding and embedding mapping, which serves as the prior structural input to the threshold prediction model.

[0209] Furthermore, by combining real-time acquired operating condition information (including ambient temperature, load rate, running time, etc.), normalization and standardization are used to transform it into an operating condition feature vector with zero mean and unit variance distribution, and to ensure that this feature is comparable to the equipment category feature in numerical space.

[0210] Furthermore, the scoring deviation feature vector, equipment category feature vector, and operating condition feature vector are concatenated column-wise to form a comprehensive feature matrix X, which is then input into the regression function T of the dynamic threshold adjustment model. The dynamic early warning threshold T is calculated using the following formula. d :

[0211] T d =f(X;w,λ)

[0212] Where w is the feature weight vector and λ is the regularization coefficient.

[0213] Furthermore, during the model training phase, a weighted least squares regression optimization method is adopted. By minimizing the weighted squared error between the predicted threshold and the historical best threshold, w and λ are dynamically adjusted to keep the model's sensitivity to key features optimal.

[0214] Through the above dynamic threshold adjustment process, the scoring deviation feature vector generated in the previous step is transformed into an adaptive threshold parameter that can adapt to different equipment types and operating conditions, thereby achieving a unified and accurate early warning triggering standard across equipment and operating conditions.

[0215] For example, in a power substation environment, 10 transformers and 6 circuit breakers operate simultaneously, with external temperatures ranging from 15℃ to 32℃ and load rates fluctuating between 45% and 87%. In practical applications, the scoring deviation feature vector obtained in the previous sub-step has a dimension of 12. The equipment category features are transformed into a 5-dimensional vector through one-hot encoding and embedding mapping, and the operating condition features are normalized to obtain an 8-dimensional vector. The total dimension of X formed by concatenating these three features is 25. The initial feature weights w are set to an equal-weight distribution, with a regularization coefficient λ = 0.1. A weighted least squares regression model is trained using historical normal state scoring data, ensuring that the mean square error between the predicted threshold and the historical optimal threshold is controlled within 0.02. During the testing period, when the equipment type changes from transformer to circuit breaker and the operating condition load rate rises to 85%, the dynamic threshold automatically adjusts from 0.68 to 0.74, effectively improving the anomaly detection sensitivity under high load conditions while avoiding false alarms under low load conditions. The adaptive threshold parameter is directly invoked in the subsequent S8.4 sub-step, realizing an early warning triggering mechanism that is highly matched with changes in device status.

[0216] S8.4: Compare the current score in the normalized sequence of equipment trend scores with the adaptive threshold parameter. If the current score exceeds the dynamic warning threshold, generate an equipment abnormality warning signal.

[0217] S8.5: Based on the equipment abnormality warning signal, execute the warning output control logic, encapsulate the warning information into a standardized warning message package, and send it to the operation and maintenance monitoring platform through the communication interface to complete the warning triggering and information push.

[0218] Step S9: Based on the error between the actual early warning results and the model's predicted score, dynamically adjust the training weights of the graph neural network model and the meta-learning framework to improve the model's generalization ability under different device configurations. Specifically, this includes:

[0219] S9.1: Quantify the error between the actual early warning result and the model prediction score to obtain error sequence data. This error sequence data serves as input to the feedback optimization module to evaluate the degree of deviation of the model prediction under the current equipment configuration.

[0220] S9.2: Based on the error sequence data, perform sliding window statistical analysis to extract the temporal correlation features of the errors and generate a dynamic feature vector of the errors. This vector is used to describe the temporal variation of the model prediction error as the equipment state evolves.

[0221] S9.3: Perform feature fusion processing on the aforementioned error dynamic feature vector and equipment type identifier to identify the influencing factors of equipment configuration on model prediction error, and generate an equipment-related error sensitivity matrix. This matrix is ​​used to quantify the uncertainty distribution of different types of equipment in model prediction.

[0222] S9.4: Based on the aforementioned device-related error sensitivity matrix, gradient backpropagation optimization is performed on the node weights and edge weights of the graph neural network model to adjust the model's response sensitivity to key topology change nodes. This optimization process employs an adaptive learning rate algorithm to improve the model's prediction stability under different device topologies.

[0223] S9.5: Incrementally update the fast-adaptive parameter set in the meta-learning framework, generating a meta-parameter update direction vector based on the aforementioned error dynamic feature vector and the device-related error sensitivity matrix. This vector guides the model to quickly adjust parameters when new devices are added, adapting to the new device combination structure and improving the model's generalization ability.

[0224] S9.6: The optimized graph neural network model is fused with the updated meta-learning framework to generate an adaptive enhanced trend prediction model. This model has the ability to adapt to changes in device type, load conditions, and environmental factors, thereby improving the accuracy and stability of multi-device correlation anomaly trend prediction.

[0225] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0226] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0227] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for early warning of abnormal monitoring of power equipment, specifically including: S1: Collect operating status data of multiple devices in the power system. The operating status data includes voltage, current, temperature, load rate, device type identification and connection topology between devices, and form a multi-dimensional device operating status time series dataset. S2: Normalize the time series data of the multidimensional equipment operating status; S3: Construct an interaction influence matrix between devices based on device type identification and connection topology; S4: The interaction influence matrix between devices is incrementally updated using a sliding time window to dynamically capture the evolution process of device connection relationships and form a topological evolution feature sequence. S5: Input the topological evolution feature sequence and the normalized equipment operating status data into the attention-based graph neural network model; then, based on the anomaly propagation path and equipment status features output by the graph neural network model, generate equipment anomaly trend score data; S6: Based on device type identification and topology change characteristics, a lightweight meta-learning framework is constructed to enable the model to automatically adjust parameters to adapt to the new device combination structure when a small number of new devices are connected to the sample. S7: Compare and analyze the abnormal trend score of the device with the historical score data to determine whether the current score exceeds the preset warning threshold. If it does, trigger the warning output.

2. The method for early warning of power equipment anomalies according to claim 1, characterized in that, It also includes step S8: Based on the error between the actual early warning results and the model prediction score, dynamically adjust the training weights of the graph neural network model and the meta-learning framework to improve the model's generalization ability under different device configurations.

3. The method for early warning of power equipment anomalies according to claim 1, characterized in that, Step S1 specifically includes: S1.1: Perform data acquisition operations on multi-source sensing devices such as smart meters, temperature sensors, and load monitoring devices deployed in the power system to obtain real-time operating parameters such as voltage, current, temperature, and load rate; S1.2: Based on the equipment type coding rules, the collected equipment operating parameters are attached with equipment type identifiers to distinguish different types of power equipment such as transformers, circuit breakers, and capacitors, thereby constructing equipment heterogeneity characteristic information; S1.3: The connection topology data between devices is synchronously obtained through the SCADA system and the topology identification module. Based on the graph structure representation method, a device connection adjacency matrix is ​​generated to describe the physical connection and energy flow relationship between devices. S1.4: Align the operating parameters with device type identifiers with the adjacency matrix with timestamps to generate structured device status data tuples with time indexes, so as to form a time series dataset of device operating status; S1.5: Perform data integrity verification and outlier removal on the time series dataset, and fill in missing data based on the sliding window mechanism to ensure the continuity and consistency of multi-dimensional device operation status data.

4. A method for early warning of power equipment anomalies according to claim 1 or 3, characterized in that, In step S1, the sampling period of the operating status data is 1-10 seconds, the sampling accuracy is not less than 0.1 units, and the sensing signal is appended with a UTC timestamp after 16-bit analog-to-digital conversion and self-calibration to form a multi-dimensional operating parameter matrix with traceability attributes.

5. The method for early warning of power equipment anomalies according to claim 1, characterized in that, In step S3, the dynamic matrix of the interaction influence matrix is ​​obtained by segmenting the time window based on the Pearson correlation coefficient. The window length is 10-120 seconds, with 50% window overlap. The correlation matrix is ​​then subjected to time smoothing and L1 normalization.

6. The method for early warning of power equipment anomalies according to claim 1, characterized in that, In step S4, the topology evolution feature sequence is based on the differential comparison algorithm to detect changes in the topology state at the beginning and end of each window, classifies and generates connection establishment, disconnection, and weight change events, and calculates the event weights by multiplying the absolute difference with the static function weights, and forms a structured output of topology change events through temporal aggregation and sequence indexing.

7. The method for early warning of power equipment anomalies according to claim 1, characterized in that, Step S5 specifically includes: S5.1: Perform feature concatenation processing on the normalized equipment operation status data and topological evolution feature sequence to construct a unified multimodal input tensor for subsequent graph neural network modeling; S5.2: Construct a dynamic adjacency matrix based on the connection topology between devices. This adjacency matrix is ​​updated in real time according to the topology evolution feature sequence to reflect the time-varying characteristics of the interaction influence between devices. S5.3: Input the multimodal input tensor and dynamic adjacency matrix into the graph attention network (GAT), and calculate the importance weight of each node under the current topology through the multi-head attention mechanism to identify key topology change nodes; S5.4: Based on the node importance weights output by the graph attention network, and combined with the historical patterns of anomaly propagation paths between devices, a path weighted aggregation algorithm is used to extract the key path features of the anomaly propagation path. S5.5: Perform nonlinear transformation and normalization on the key path features to generate a node-level anomaly propagation intensity vector, which serves as the input feature for the subsequent anomaly trend scoring module.

8. A method for early warning of power equipment anomalies according to claim 1 or 7, characterized in that, In step S5, the graph neural network anomaly trend prediction module uses input features to match and bind with the dynamic graph adjacency matrix, uses a multi-head graph attention mechanism to determine the importance of nodes and key anomaly propagation paths, achieves anomaly propagation intensity vector output through node weight aggregation, and combines a gated recurrent network to model temporal correlation.

9. The method for early warning of power equipment anomalies according to claim 1, characterized in that, In step S6, the meta-learning adaptive framework includes a task-aware parameter generation network based on device context embedding. Batch gradient descent inner layer optimization is used for the B=5-50 running samples collected after the new device is connected. The optimizer is Adam, the learning rate is between 0.001 and 0.05, and gradient accumulation is used to improve the parameter stability under a small number of samples.

10. The method for early warning of power equipment anomalies according to claim 1, characterized in that, In step S7, the trigger warning output module calculates the sliding window deviation between the normalized score and the historical score distribution, and combines the unique hot coding of equipment type and the normalized working condition characteristics to generate an adaptive threshold in real time using a regression model. The feature weights and regularization parameters are dynamically adjustable, and the module automatically responds to changes in equipment category and working condition.

Images