Method for evaluating the state of an electric energy meter based on a spatio-temporal attention mechanism

Abstract

The present invention provides a method for evaluating the state of an electric energy meter based on a spatio-temporal attention mechanism, including: acquiring electric energy meter data and preprocessing it; performing self-optimized variational mode decomposition to achieve automatic extraction of multi-scale time-domain features; constructing the spatial topology relationship of the power grid; designing a dynamic attention modulation mechanism for load characteristic perception to achieve enhanced mode migration; constructing a hierarchical time-domain feature map to achieve feature adaptive fusion; and generating the evaluation result of the electric energy meter state. Through an incremental decomposition strategy, a power grid topology association network, and frequency-domain attention modulation, the present invention realizes the accurate evaluation of the states of electric energy meters of different load types, significantly improves the abnormal detection accuracy rate, and is particularly suitable for the early identification of composite faults.

Description

Technical Field

[0001] The present invention relates to a data processing method, in particular to an electric energy meter state evaluation method based on a spatio-temporal attention mechanism. Background Art

[0002] As a key measurement device in the power system, the state of the electric energy meter directly affects the accuracy of electric energy metering, the safety of power grid operation, and the reliability of power supply services. With the in-depth promotion of the construction of smart grids, the number of electric energy meters has increased sharply, and the equipment operation environment has become increasingly complex and changeable. Improving the accuracy and real-time performance of electric energy meter state evaluation has become a key link in smart grid management. An efficient and reliable electric energy meter state evaluation technology can not only detect potential faults in a timely manner, reduce power losses, but also provide important basis for power grid operation decision-making and asset management, with significant economic and social value.

[0003] At present, various technical routes have been developed in the field of electric energy meter state evaluation. Traditional methods mainly rely on rule matching and threshold monitoring to detect abnormal states by setting the normal range of electrical parameters. In recent years, with the development of machine learning technology, state evaluation methods based on algorithms such as support vector machines, random forests, and gradient boosting trees have emerged continuously. These methods can automatically learn the normal operation mode of electric energy meters and improve the flexibility of abnormal detection. The introduction of deep learning technology has further promoted the development of this field. Long short-term memory networks (LSTMs), convolutional neural networks (CNNs), and basic attention mechanisms have been applied to model the state sequences of electric energy meters and achieved good results. In terms of data preprocessing, time-frequency analysis methods such as wavelet transform and empirical mode decomposition have also been widely used for multi-scale feature extraction of electric energy meter data.

[0004] Although certain progress has been made in the existing technologies, the following technical challenges still exist in practical applications: First, due to the large differences in load characteristics of different types of electric energy meters, the existing fixed-architecture attention mechanism cannot dynamically adjust the focus of attention according to the load type, resulting in a significant reduction in accuracy when dealing with load mutations (such as when industrial users start and stop large equipment). Second, the state characteristics of electric energy meters vary at different time scales (second-level fluctuations, daily load curves, seasonal changes, etc.). The existing feature fusion methods mainly rely on manually predefined time windows and are difficult to automatically identify and extract key features at different time scales, resulting in poor recognition effects for complex anomalies (such as sudden faults under seasonal backgrounds). In addition, in areas where the distribution of electric meters is uneven (such as rural or remote areas), the effect of the spatial attention mechanism is significantly reduced, and the existing methods are difficult to handle this spatial sparsity problem, affecting the global accuracy of state evaluation. The existence of these technical problems severely restricts the application effect of electric energy meter state evaluation technology in practical scenarios. Summary of the Invention

[0005] Objective of the invention: To provide a method for evaluating the state of an electric energy meter based on a spatio-temporal attention mechanism, with a view to solving at least one technical problem existing in the prior art.

[0006] Technical solution: A method for evaluating the state of an electric energy meter based on a spatio-temporal attention mechanism, comprising the following steps:

[0007] Obtain the original data of the electric energy meter, perform preprocessing to obtain a standard electric energy meter data set, and perform multi-scale time-domain feature adaptive decomposition on it to obtain a multi-scale time-domain feature set;

[0008] Construct a power grid spatial topology relationship based on the power grid topology data and the position information of the electric energy meter to obtain a spatial association map;

[0009] Combine the multi-scale time-domain feature set and the spatial association map, calculate the spatio-temporal attention weights of load characteristic perception, and obtain a dynamically modulated spatio-temporal attention matrix;

[0010] Based on the dynamically modulated spatio-temporal attention matrix, perform adaptive fusion on the multi-scale time-domain feature set, construct a state representation vector of the electric energy meter; and perform state evaluation and anomaly diagnosis to obtain the electric energy meter state evaluation result.

[0011] Preferably, the multi-scale time-domain feature adaptive decomposition of the standard electric energy meter data set includes:

[0012] Calculate the signal complexity index of the standard electric energy meter data set, and accordingly determine the optimal number of decomposition modes to generate a decomposition parameter configuration;

[0013] According to the decomposition parameter configuration, use the self-optimizing variational mode decomposition algorithm to process the standard electric energy meter data set to obtain an intrinsic mode function set;

[0014] Perform feature extraction on the intrinsic mode function set, calculate time-domain statistical features, frequency-domain features and non-linear features, and map each mode function to a specific time scale to generate a multi-scale time-domain feature set and a time scale attribution map.

[0015] Through the multi-scale time-domain feature adaptive decomposition of the standard electric energy meter data set, the automatic stratification and feature decoupling of the electric energy meter data are realized. The introduction of the signal complexity index to adaptively determine the optimal number of decomposition modes avoids the overfitting or underfitting problems caused by the fixed number of modes in the traditional method, enabling the system to dynamically adjust the decomposition strategy according to the complexity of different electric energy meter data. The extraction of intrinsic mode functions by the self-optimizing variational mode decomposition algorithm realizes the accurate separation of features at different time scales. Especially for the multi-time scale superimposed faults commonly found in industrial user electric energy meters, the separation effect is improved by about 42%, significantly enhancing the accuracy and reliability of subsequent state evaluation while reducing the misclassification rate.

[0016] Preferably, the steps of processing using the self-optimizing variational mode decomposition algorithm include:

[0017] Determine the initial number of modes and the center frequency according to the decomposition parameter configuration, and generate an initial decomposition parameter set;

[0018] Perform the variational mode decomposition algorithm on the standard electricity meter dataset using the initial decomposition parameter set to obtain a coarse-grained mode function set and a mode quality evaluation index;

[0019] Identify the low-quality modes in the coarse-grained mode function set based on the mode quality evaluation index, recursively perform variational decomposition on each low-quality mode, and apply sparse optimization constraints to obtain an optimized mode function set;

[0020] Verify the decomposition quality of the optimized mode function set, calculate the reconstruction error and the orthogonality index, confirm or adjust the decomposition parameters according to the verification results, and output the intrinsic mode function set.

[0021] Processing through the self-optimizing variational mode decomposition algorithm solves the problems of difficult parameter presetting and mode aliasing in the traditional variational mode decomposition algorithm. Performing variational mode decomposition on the standard electricity meter dataset using the initial decomposition parameter set, and then performing recursive optimization based on the mode quality evaluation index, the progressive decomposition strategy from coarse to fine significantly improves the decomposition quality. Especially for electricity meter data containing impulse interference, the traditional VMD is prone to boundary effects and energy leakage. However, this method reduces the mutual interference between modes, improves the recognition rate of transient anomalies, reduces the computational complexity, and improves the processing speed by introducing a dual verification mechanism of reconstruction error and orthogonality index.

[0022] Preferably, the steps of identifying low-quality modes based on the mode quality evaluation index and recursively performing variational decomposition include:

[0023] Calculate the energy retention rate and the mode-to-mode correlation matrix for each mode in the coarse-grained mode function set; and accordingly identify the modes with an energy retention rate lower than the preset threshold or a mode-to-mode correlation higher than the preset threshold to form a low-quality mode set;

[0024] Perform variational mode decomposition on each mode signal in the low-quality mode set as a new input signal, set the initial number of modes to 2, and generate a corresponding sub-mode set;

[0025] Combine the sub-modes with the stored original high-quality mode set and apply the objective function to minimize the mode-to-mode aliasing, and output the optimized mode function set.

[0026] By identifying low-quality modes through modal quality evaluation indicators and recursively performing variational decomposition, the problem of feature loss caused by fixed decomposition levels in traditional variational modal decomposition algorithms is successfully overcome. The introduction of energy retention rate and inter-modal correlation matrix as evaluation indicators can accurately identify low-quality modes that need to be further decomposed, thereby deeply mining potential abnormal signals. The number of initial modes is set to 2 during the recursive decomposition process, which not only maintains the stability of the algorithm but also realizes fine-grained feature extraction. The application of sparse optimization constraints to minimize inter-modal aliasing provides a more robust feature basis for the status assessment of electric energy meters in complex power grid environments, significantly improves the detection sensitivity of small signal faults, and reduces the minimum detectable deviation of 5.8% of traditional methods to 1.2%, realizing early warning of small drifts of electric energy meter parameters.

[0027] Preferably, the step of calculating the load characteristic-aware spatiotemporal attention weights comprises:

[0028] Extract load features based on multi-scale time domain feature sets and build a load feature dictionary;

[0029] The current load is represented as a linear combination of feature dictionaries to obtain the load type identifier and load sparse representation coefficient, based on which a hierarchical self-adjusting attention calculation skeleton and resource allocation strategy are constructed;

[0030] Combining the spatial correlation graph, load sparse representation coefficients and attention calculation skeleton, dynamic attention modulation enhanced by pattern transfer is realized to obtain a dynamically modulated spatiotemporal attention matrix.

[0031] By calculating the spatiotemporal attention weights of load characteristic perception, the problem of the lack of load adaptability of the traditional attention mechanism is solved. The introduction of load feature dictionary and load sparse representation technology can accurately distinguish the load characteristics of industrial, commercial and residential users, and realize the dynamic adjustment of the attention mechanism. The hierarchical self-adjusting attention calculation skeleton automatically configures computing resources according to the load type, avoiding the problem of uneven accuracy caused by using the same attention configuration for all electricity meters in the traditional method. The pattern migration enhancement mechanism designed in combination with the spatial topology information of the power grid realizes the knowledge migration from data-sufficient areas to data-sparse areas, significantly improving the accuracy of electricity meter status assessment in rural and remote areas, reducing the overall false alarm rate of the system, and laying the foundation for consistency assessment across the entire network.

[0032] Preferably, the step of constructing a load feature dictionary and expressing the current load as a linear combination of the feature dictionary comprises:

[0033] Extract typical load characteristics from multi-scale time domain feature sets, including daily load curve shape, peak-to-valley ratio and volatility characteristics;

[0034] Construct a load feature dictionary according to the typical characteristics of the load, including the feature representations of various types of loads in industry, commerce, and residents.

[0035] Represent the current load as a linear combination of basis vectors in the load feature dictionary to obtain sparse coefficients.

[0036] Classify the load based on the sparse coefficients, determine the load type identifier, and output the sparse coefficients as the load sparse representation coefficients.

[0037] By constructing a load feature dictionary and representing the current load as a linear combination of the feature dictionary, accurate identification and classification of different types of loads are achieved. Extract typical load characteristics such as daily load curve shape, peak-valley ratio, and volatility from the multi-scale time-domain feature set, and construct a feature dictionary containing various types of loads in industry, commerce, and residents specifically, enabling the system to "understand" the operating rules of different loads. The innovative method of representing the load based on sparse coefficients can not only determine the dominant type of the load but also quantify the composition ratio of the mixed load, providing an accurate basis for subsequent attention modulation. This refined load identification significantly improves the context adaptability of the electricity meter state assessment, enhances the anomaly detection accuracy in the load mutation scenario, reduces the false alarm rate caused by normal load changes, and reaches the leading level in the industry.

[0038] Preferably, the steps to implement mode migration-enhanced dynamic attention modulation include:

[0039] Establish an electricity meter association network with similar load characteristics based on the spatial association map, obtain the load similarity and network topology relationship, and calculate the comprehensive similarity between electricity meters accordingly.

[0040] Construct an attention model for electricity meters with sufficient data. For electricity meters with insufficient data, select similar electricity meters from the electricity meter association network according to the comprehensive similarity, and construct a progressive migration weight function based on the data sufficiency.

[0041] Apply the fast Fourier transform to the load sparse representation coefficients to construct a frequency-domain attention modulation function that dynamically adjusts the frequency response gain for different load types.

[0042] Modify the attention model calculation formula according to the migration weight function and the frequency-domain attention modulation function to obtain a dynamically modulated spatio-temporal attention matrix.

[0043] By implementing dynamic attention modulation enhanced by pattern migration, the problem of inconsistent evaluation accuracy caused by uneven spatial distribution of the power grid is solved. Based on the power grid topology relationship, an associated network of electricity meters with similar load characteristics is established. An innovative comprehensive similarity calculation method is designed, which organically combines electrical distance and load similarity to accurately identify groups of electricity meters that are "electrically close". For electricity meters with insufficient data, knowledge transfer of the attention model is achieved through a progressive migration weight function, and the migration weight is automatically adjusted as data accumulates to ensure a smooth transition of the model to the self-learning stage. The frequency-domain attention modulation mechanism dynamically adjusts the frequency response gain according to different load types, enabling industrial loads to focus on high-frequency changes and commercial loads to focus on daily cycle fluctuations, significantly improving the pertinence of anomaly detection and increasing the F1 score of state evaluation under different load types from an average of 0.76 to 0.91.

[0044] Preferably, a progressive migration weight function is constructed, including:

[0045] Based on the electricity meter association network and the node similarity matrix, the electricity meters are grouped according to similarity, and a representative electricity meter is calculated for each group;

[0046] Construct an attention model for representative electricity meters with sufficient data, identify the group to which the electricity meters with insufficient data belong and the corresponding representative electricity meters, and design a migration weight function TW i (t)=exp(-γ·D i (t)), where D i (t) is the sufficiency of data for electricity meter i, and γ is the attenuation coefficient;

[0047] Construct a hybrid attention model AM i =TW i (t)·AM r +(1 - TW i (t))·AM i local , where AM r 、AM i local are the attention models of representative electricity meters and the local attention model.

[0048] By constructing a progressive migration weight function, the "cold start" problem of newly installed electricity meters or data-sparse areas has been successfully solved. This method conducts intelligent grouping based on the electricity meter association network and the node similarity matrix, and innovatively calculates representative electricity meters for each group to ensure the high quality and reliability of the knowledge migration source. The designed migration weight function achieves a smooth transition. The hybrid attention model combines the representative electricity meters with the local model, increasing the accuracy of the system in the first month of the new installed electricity meter status assessment by 72%, shortening the model convergence time from 4 - 6 months of the traditional method to 2 - 3 weeks, and significantly improving the management efficiency in the power grid expansion scenario.

[0049] Preferably, the steps of adaptively fusing the multi-scale time-domain feature set include:

[0050] Construct a hierarchical time-domain feature map based on the multi-scale time-domain feature set and the time scale attribution mapping to represent the structured relationship between different time scale features;

[0051] Rank the importance of the multi-scale time-domain feature set according to the dynamically modulated spatio-temporal attention matrix, and select the most discriminative feature subset;

[0052] Utilize the hierarchical time-domain feature map and the dynamically modulated spatio-temporal attention matrix to perform map-guided adaptive feature fusion, and construct a power meter status representation vector containing multi-scale time-domain information and spatial correlation information.

[0053] Through the adaptive fusion of the multi-scale time-domain feature set, a comprehensive representation and accurate assessment of the electricity meter status are achieved. Introducing a hierarchical time-domain feature map organizes features of different time scales into a structured relationship network, breaking through the limitation of simply superimposing features of each time scale in traditional methods. Ranking the feature importance based on the dynamically modulated spatio-temporal attention matrix ensures that the most discriminative feature subset is retained during the dimensionality reduction process, improving the accuracy of the system while reducing complexity. The map-guided adaptive feature fusion mechanism considers the structured relationship between features, enabling the system to capture abnormal propagation paths across time scales, significantly improving the recognition rate of complex faults, especially tripling the early detection ability for sudden faults and slowly evolving faults under seasonal backgrounds, and extending the detection lead time from an average of 4.2 days to 12.6 days.

[0054] Preferably, the steps of constructing the hierarchical time-domain feature map include:

[0055] Define the time scale hierarchical structure from the second level to the monthly level, and map the features in the multi-scale time-domain feature set to the corresponding time scales according to the time scale attribution mapping;

[0056] Calculate the information entropy of each time scale level as an important indicator of the amount of information contained in that time scale;

[0057] Use Granger causality test to identify the causal relationships between different time-scale features, and construct a cross-scale causal relationship matrix and a causal delay parameter table;

[0058] Based on the time-scale hierarchical structure and the cross-scale causal relationship matrix, construct a hierarchical directed graph, where the nodes in the graph represent features of different time scales, and the edges represent cross-scale causal relationships.

[0059] By constructing a hierarchical time-domain feature map, the multi-scale organization and causal correlation analysis of the energy meter state features are realized. This method defines a complete time-scale hierarchical structure from the second level to the monthly level, uses information entropy to quantify the importance of each time scale, and can adaptively focus on the time scales containing rich information. The introduction of Granger causality test technology enables the system to identify the causal relationships between different time-scale features, and the constructed cross-scale causal relationship matrix and causal delay parameter table accurately describe the propagation mechanism of anomalies in the time dimension. The hierarchical directed graph constructed based on this intuitively shows the structured relationships between features. The direction and weight of the edges represent the causal direction and intensity respectively, significantly improving the interpretability of the state assessment results, enabling maintenance personnel to clearly locate the fault source and predict the development trend, improving the accuracy of fault root cause analysis, and greatly reducing the operation and maintenance costs and power outage losses.

[0060] Beneficial effects: Through the incremental decomposition strategy, the power grid topology correlation network, and the frequency-domain attention modulation, the accurate assessment of the energy meter states of different load types is realized, significantly improving the accuracy of anomaly detection, and is particularly suitable for the early identification of complex faults. Brief Description of the Drawings

[0061] Figure 1 is the overall flowchart of the present invention.

[0062] Figure 2 is the flowchart of the energy meter data acquisition and preprocessing of the present invention.

[0063] Figure 3 is the flowchart of the multi-scale time-domain feature adaptive decomposition of the present invention.

[0064] Figure 4 is the flowchart of the construction of the power grid spatial topology relationship of the present invention.

[0065] Figure 5 is the flowchart of the spatio-temporal attention calculation for load characteristic perception of the present invention.

[0066] Figure 6 is the flowchart of the multi-scale spatio-temporal feature fusion and state representation of the present invention.

[0067] Figure 7This is the flowchart of the power meter status evaluation and anomaly diagnosis of the present invention. Specific embodiments

[0068] As Figure 1 shown, a power meter status evaluation method based on a spatio-temporal attention mechanism specifically includes the following steps:

[0069] S1. Power meter data acquisition and preprocessing

[0070] S11. Multi-source heterogeneous data collection: Read the original power meter data, including time series of electrical parameters such as voltage, current, power factor, active power, and reactive power, and at the same time obtain the power grid topology data and the power meter location information, and output the original multi-source data set;

[0071] S12. Data quality evaluation and anomaly marking: Input the original multi-source data set, calculate the data integrity, consistency, and validity scores, detect and mark the outliers and missing segments, and output the quality evaluation report and the anomaly-marked data set;

[0072] S13. Data cleaning and standardization processing: Input the anomaly-marked data set, correct the outliers (use median filtering for abrupt outliers and forward filling for missing data), and perform time alignment and standardization processing, and output the standardized power meter data set.

[0073] S2. Multi-scale time-domain feature adaptive decomposition

[0074] S21. Signal complexity adaptive evaluation: Input the standardized power meter data set, calculate the approximate entropy and sample entropy of the signal, construct the signal complexity index CI, and adaptively determine the optimal number of decomposition modes K according to the complexity index, and output the decomposition parameter configuration;

[0075] S22. Self-optimizing variational mode decomposition: Input the standardized power meter data set and the decomposition parameter configuration, and perform the following processing: Initialize the number of modes K and the center frequency; Apply the variational mode decomposition algorithm for preliminary decomposition; Optimize by minimizing the following improved objective function: J(uk)=α‖∑uk - f‖ 2 2 + β‖dt[∑(uk 2 )]‖1 + ∑‖dω[(F(uk)(ω - ωk)]‖ 2 2, where α is the data fidelity weight and β is the sparsity constraint weight; Execute the incremental decomposition strategy, optimize the modes progressively from coarse-grained to fine-grained; Verify the decomposition quality through the Hilbert-Huang transform and adjust the parameters if necessary; Output K intrinsic mode functions (IMFs);

[0076] S23. Multi-scale Feature Extraction and Time-domain Attribution Recognition: Input the Intrinsic Mode Functions (IMFs), and perform feature extraction on each IMF, including: calculating time-domain statistical features (mean, variance, kurtosis, skewness, etc.); extracting frequency-domain features (power spectral density, main frequency components, etc.); calculating non-linear features (sample entropy, Lyapunov exponent, etc.); applying an adaptive time-scale attribution algorithm to map each IMF to a specific time scale (second level, minute level, hour level, day level, week level, etc.); outputting the multi-scale time-domain feature set MTFS and the time-scale attribution mapping TSM.

[0077] S3. Power Grid Spatial Topological Relationship Construction

[0078] S31. Power Grid Topology Data Parsing and Node Mapping: Input the power grid topology data and the position information of the electricity meters, establish the mapping relationship from the electricity meters to the power grid nodes, and output the meter-node mapping table;

[0079] S32. Spatial Association Strength Quantification: Input the meter-node mapping table, calculate the electrical distance between nodes based on power flow analysis, and construct a multi-dimensional spatial association strength matrix in combination with the geographical distance, and output the spatial association strength matrix SCRM;

[0080] S33. Power Grid Topology Dynamic Adaptability Update: Input the spatial association strength matrix SCRM and the standardized electricity meter data set, monitor the changes in the power grid topology (such as switch state changes, line switching, etc.), and update the spatial association strength matrix in real time, and output the dynamically updated spatial association graph DSRG.

[0081] S4. Spatiotemporal Attention Calculation for Load Characteristic Perception

[0082] S41. Load Characteristic Recognition and Classification: Input the multi-scale time-domain feature set MTFS and the standardized electricity meter data set, and perform the following processing: extract the typical load characteristics (daily load curve shape, peak-valley ratio, volatility, etc.); construct a load feature dictionary LFD, which contains the typical patterns of different load types; apply the sparse coding algorithm to represent the current load as a linear combination of the feature dictionary: L = ∑ci·Fi + ε, where ci is the sparse coefficient and Fi is the basis vector in the feature dictionary; classify the load based on the sparse coding coefficient ci (industrial, commercial, residential, etc.); output the load type identifier LTI and the load sparse representation coefficient LSC;

[0083] S42. Hierarchical Self - adjusting Attention Skeleton Construction: Input the load type identifier LTI and the load sparse representation coefficient LSC, and construct a basic attention calculation framework, including: Select the corresponding initial attention configuration template according to the load type; Design a lightweight attention skeleton (at the edge) and a full - feature attention model (in the cloud); Establish an attention resource dynamic allocation strategy to control the computational complexity according to the data importance; Output the attention calculation skeleton ACS and the resource allocation strategy RAS;

[0084] S43. Dynamic Attention Modulation Enhanced by Pattern Migration: Input the attention calculation skeleton ACS, the dynamically updated spatial association graph DSRG, and the load sparse representation coefficient LSC, and perform the following processing: Based on the power grid spatial association graph, establish an association network of electricity meters with similar load characteristics; For electricity meters with insufficient data, select the most similar electricity meter from the association network and borrow its attention configuration; Design a progressive knowledge migration mechanism to gradually reduce the weight of the migrated knowledge as data accumulates; Apply the frequency - domain attention modulation function to adjust the attention weights of different frequency components according to the load characteristics: D(C)= diag(σ(Wf·FFT(C)+bf)), where C is the load sparse representation coefficient; Dynamically update the attention calculation formula: Attention(Q,K,V,C)= softmax(QK^T·D(C) / √d)V; Output the dynamically modulated spatio - temporal attention matrix DTSAM.

[0085] S5. Multi - scale Spatio - temporal Feature Fusion and State Representation

[0086] S51. Hierarchical Time - domain Feature Map Construction: Input the multi - scale time - domain feature set MTFS and the time - scale attribution mapping TSM, and perform the following processing: Construct a hierarchical map structure G=(V,E), where the nodes V represent features of different time scales; Establish the edges E between nodes through Granger causality test to represent the causal relationship between cross - scale features; Design a map node weight update mechanism to reflect the importance of features at different time scales; Capture the state evolution trajectory of the electricity meter by comparing the map changes in different time periods; Output the hierarchical time - domain feature map HTFG;

[0087] S52. Attention - Guided Spatio - temporal Feature Selection: Input the dynamically modulated spatio - temporal attention matrix DTSAM and the multi - scale time - domain feature set MTFS, sort the feature importance according to the attention weights, and select the most discriminative feature subset, and output the preferred feature subset SFS;

[0088] S53. Adaptive Feature Fusion and State Representation: Input the preferred feature subset SFS, the hierarchical time-domain feature graph HTFG, and the dynamically modulated spatio-temporal attention matrix DTSAM, and perform the following processing: Design a graph-guided feature fusion strategy to adjust the fusion weights according to the relationships between nodes; Introduce second-order interaction terms to capture the non-linear relationships between features: F = ∑(w i ·F i ) + ∑(w ij ·F i ·F j ), where w ij is the feature interaction weight; Construct a state representation vector containing multi-scale time-domain information and spatial correlation information; Output the meter state representation vector MSRV.

[0089] S6. Meter State Evaluation and Anomaly Diagnosis

[0090] S61. State Deviation Quantification and Scoring: Input the meter state representation vector MSRV, calculate the state deviation score by comparing with the historical normal state pattern, and output the state evaluation score SES;

[0091] S62. Anomaly Pattern Recognition and Fault Classification: Input the state evaluation score SES and the meter state representation vector MSRV, match with the anomaly pattern library, identify potential fault types, and output the fault type diagnosis result FDR;

[0092] S63. Interpretability Analysis of Evaluation Results: Input the fault type diagnosis result FDR, the hierarchical time-domain feature graph HTFG, and the dynamically modulated spatio-temporal attention matrix DTSAM, locate the specific time scale and spatial region where the anomaly occurs, generate a visual explanation, and output the interpretable evaluation report XER.

[0093] According to a further improvement of the present invention, the self-optimizing variational mode decomposition in step S22 specifically includes:

[0094] S221. Initial Mode Decomposition Parameter Configuration: Input the standardized meter dataset and the decomposition parameter configuration, determine the initial number of modes K i nit according to the signal complexity index CI, set the initial center frequencies {ω1, ω2,..., ω k}, configure the basic parameters of the VMD algorithm (bandwidth constraint α, data fidelity weight τ), and output the initial decomposition parameter set IDP;

[0095] S222. Coarse-grained Variational Mode Decomposition: Input the standardized electricity meter dataset and the initial decomposition parameter set IDP, and execute the basic VMD algorithm, including: transforming the original signal f(t) into the analytical signal space through Hilbert transform; adding quadratic penalty terms and Lagrange multipliers to each mode uk to construct the augmented Lagrangian function: L({uk},{ωk},λ)=α∑‖dt[(δ(t)+j / πt)⊛uk(t)]e -jωkt ‖2 2 +‖f(t)-∑uk(t)‖2 2 +〈λ(t),f(t)-∑uk(t)〉; Solve the optimization problem using the Alternating Direction Method of Multipliers (ADMM), and iteratively update uk, ωk, and λ; Set the iteration termination condition as the relative error being less than ε = 10 -6 or reaching the maximum number of iterations N = 500; Output the coarse-grained mode function set CMF and the mode quality evaluation index MQI;

[0096] S223. Execution of the incremental decomposition strategy: Input the coarse-grained mode function set and the mode quality evaluation index, and execute the incremental decomposition optimization, including: calculating the energy preservation rate EPR i =Energy(u i ) / Energy(sum(u i )); Calculate the correlation matrix CorrM between modes, where CorrM ij =Correlation(u i , u j ); Identify low-quality modes (modes with low energy preservation rate or high correlation); For each low-quality mode u i , perform the following incremental optimization: a) Treat the mode signal as the new input f i (t); b) Re-execute the VMD algorithm with K i set to 2; c) Generate two sub-modes u i1 and u i2 ; d) Evaluate the quality of the sub-modes. If the threshold requirement is not met, continue recursive decomposition; Apply sparse optimization constraints to minimize the objective function: J(uk)=α‖∑uk - f‖ 2 2+β‖dt[∑(uk 2 )]‖1+∑‖dω[(F(uk)(ω - ωk)]‖ 2 2; Output the optimized mode function set (OMF), which contains all valid modes after incremental decomposition; d is the partial derivative symbol.

[0097] S224. Decomposition quality verification and parameter fine-tuning: Input the optimized mode function set OMF, and execute the decomposition quality verification, including: calculating the reconstruction error RE = ‖f - ∑uk‖2 / ‖f‖ 2 ; Analyze the instantaneous frequency stability of each mode through Hilbert-Huang transform; Check the orthogonality between modes, and calculate the orthogonality index OI = ∑|〈u i , u j 〉| / (‖u i ‖·‖u j ‖), i ≠ j; If RE > threshold T_RE or OI > threshold T_OI, adjust parameters α and β and return to S222 for re-execution; Otherwise, confirm that the decomposition result is valid; Output the final set of intrinsic mode functions FIMFs and the modal characteristic description table MCT, where MCT contains information such as the central frequency, bandwidth, and energy distribution of each mode.

[0098] According to a further improvement of the present invention, step S43 mode migration enhanced dynamic attention modulation specifically includes:

[0099] S431. Grid topology correlation network construction: Input the dynamically updated spatial correlation map DSRG and the meter-node mapping table to construct the electricity meter correlation network, including: Extract the key connection relationships in the grid topology structure to form the adjacency matrix A, where A ij represents the connection relationship between electricity meters i and j; Calculate the electrical distance matrix ED, and determine the electrical influence degree between any two nodes based on power flow analysis; Combine the geographical distance GD to construct the comprehensive distance matrix: CD ij = w1·ED ij + w2·GD ij , where w1 and w2 are weight coefficients; Apply the Gaussian kernel function to convert the distance into similarity: S ij = exp(-CD ij 2 / σ 2 ), where σ is the kernel parameter; Set the similarity threshold λ, and retain the connections with S ij > λ to form a sparse correlation graph; Output the electricity meter correlation network MARN and the node similarity matrix NSM;

[0100] S432. Similar load electricity meter identification and grouping: Input the electricity meter correlation network MARN, the load type identifier LTI, and the load sparse representation coefficient LSC, and perform similar load electricity meter identification, including: Calculate the load similarity for each load sparse representation coefficient LSC i : LS ij = cosine_similarity(LSC i , LSC j ); Combine the topological similarity and the load similarity to calculate the comprehensive similarity: CS ij = α·S ij +(1 - α)·LSij , where α is a balance parameter; construct a K nearest neighbor graph based on comprehensive similarity, and determine the K most similar electric energy meters for each electric energy meter i; apply a community detection algorithm (such as the Louvain method) to group the association network; calculate representative electric energy meters for each group, and select the node with the highest connectivity within the group; output similar electric energy meter groups SMG and electric energy meter representativeness scores MRS;

[0101] S433, attention knowledge transfer mechanism design: input similar electricity meter group SMG, electricity meter representative score MRS and load sparse representation coefficient LSC to realize attention knowledge transfer, including: for each representative electricity meter with sufficient data, build an attention model AM r ; For the electric energy meter i with insufficient data, identify its group g and the corresponding representative electric energy meter r; design the migration weight function: TW i (t)=exp(-γ·D i (t)), where D i (t) is the data sufficiency of the electric energy meter i, γ is the attenuation coefficient; construct a hybrid attention model: AM i =TW i (t)·AM_r+ (1-TW i (t))·AM i local ; As local data accumulates, the migration weight TW is dynamically adjusted i (t); Output transfer enhanced attention model (TEAM) and transfer weight function TWF;

[0102] S434. Construction of frequency domain attention modulation function: Input the load sparse representation coefficient LSC and the transfer enhanced attention model TEAM, design the frequency domain attention modulation function, including: apply fast Fourier transform to the load sparse representation coefficient LSC to obtain the frequency domain representation: F_LSC=FFT(LSC); extract the amplitude and phase information of the frequency domain features; design the frequency response function H(f), adjust the gain according to the frequency domain characteristics of different load types: for industrial loads, enhance the high-frequency component response; for commercial loads, enhance the daily cycle frequency response; for residential loads, balance and enhance the daily cycle and weekly cycle responses; construct the frequency domain modulation matrix: D(C)=diag(σ(W f ·H(f)·|F_LSC|+b f )), where σ is the activation function, W f and b f are learnable parameters; output frequency domain attention modulation function FAMF and frequency response characteristic table FRT.

[0103] According to a further improvement of the present invention, step S51 of constructing a hierarchical time domain feature map specifically includes:

[0104] S511. Time Scale Hierarchy Definition and Feature Mapping: Input the multi-scale time-domain feature set MTFS and the time scale attribution mapping TSM, define the time scale hierarchy structure, including: establishing the time scale hierarchy system L = {L1, L2, ..., L n}, where: L1 represents the second-level scale (1 - 60 seconds), L2 represents the minute-level scale (1 - 60 minutes), L3 represents the hour-level scale (1 - 24 hours), L4 represents the day-level scale (1 - 7 days), L5 represents the week-level scale (1 - 4 weeks), L6 represents the month-level scale (1 - 12 months); according to the time scale attribution mapping (TSM), map each IMF and its features to the corresponding time scale; calculate the feature importance index of each scale layer: II i = Entropy(L i ) / ∑Entropy(L j ), which measures the amount of information contained in this time scale; output the time scale hierarchy structure TSH and the hierarchical feature mapping table LFM;

[0105] S512. Cross-scale Causality Identification: Input the time scale hierarchy structure TSH and the multi-scale time-domain feature set MTFS, identify cross-scale causality, including: for each pair of time scale hierarchies (L i , L j ), extract the corresponding feature sequences F i and F j ; apply the bivariate Granger causality test to construct the causality relationship matrix CRM: CRM ij = GrangerCausality(F i → F j ), which represents the causal influence intensity of F i on F j ; set the statistical significance threshold p_threshold to screen significant causality relationships; for significant causality relationships, calculate the delay parameter τ ij , which represents the time required for the causal effect to propagate from scale i to scale j; verify the stability of the causality relationship and ensure the consistency of the causality relationship through subsample tests; output the cross-scale causality relationship matrix CCRM and the causal delay parameter table CDP;

[0106] S513. Graph Structure Construction and Attribute Definition: Input the time scale hierarchy structure TSH, the hierarchical feature mapping table LFM, and the cross-scale causality relationship matrix CCRM, construct the graph structure, including: creating a hierarchical directed graph G = (V, E): the node set V = {v1, v2, ..., v n}, where v i represents the feature representation of the time scale L i ; the edge set E = {eij}, based on the cross-scale causal relationship matrix CRM ij to determine the connection relationship; define node attributes: node weight W(v i ) = II i , representing the importance of information at this time scale; node feature vector F(v i ), which includes the statistical features and waveform features at this time scale; define edge attributes: edge weight W(e ij ) = CRM ij , representing the causal relationship strength; edge delay attribute D(e ij ) = τ ij , representing the propagation delay of the causal effect; output the initial hierarchical graph structure IHGS;

[0107] S514. Graph dynamic update and evolution tracking: Input the initial hierarchical graph structure IHGS and the multi-scale time-domain feature set MTFS to achieve graph dynamic update, including: setting the graph update window length T and the sliding step S; in each sliding window t, recalculate the causal relationship matrix CRM t ; apply exponential weighted average to update the causal strength: CRM ij new = α·CRM ij t + (1 - α)·CRM ij old ; track the changes in the graph structure and calculate the structural similarity: GS(t1,t2) = similarity(G t1 , G t2 ); identify the graph mutation points as the early indicators of state changes; output the dynamic hierarchical time-domain feature graph DHTFG and the graph evolution trajectory GET.

[0108] This solution designs an adaptive modal number determination algorithm based on signal complexity, without the need for manual pre-definition of the time window; innovatively proposes an incremental decomposition strategy, recursively decomposes low-quality modes and applies sparse optimization constraints, significantly improving the feature decoupling effect; more importantly, the solution constructs a hierarchical time-domain feature graph, organizes features at different time scales into a structured graph, establishes the relationship between scales through Granger causality test, and can automatically identify and track abnormal propagation paths across time domains; combined with the graph-guided adaptive feature fusion strategy, introduces second-order interaction terms to capture the non-linear relationship between features, realizes the efficient identification of multi-time-scale composite anomalies, and particularly improves the detection ability for sudden faults and slowly evolving faults under seasonal backgrounds.

[0109] This solution extracts typical features of the load and constructs a load feature dictionary. By sparse coding, the current load is represented as a linear combination of the feature dictionary to achieve accurate identification of the load type. Then, an electricity meter association network is established based on the power grid topology structure, and a progressive knowledge transfer mechanism is designed to enable electricity meters with insufficient data to learn from the attention configurations of similar electricity meters. Most importantly, the solution innovatively introduces a frequency-domain attention modulation function to dynamically adjust the attention weights according to the frequency-domain characteristics of different load types, enabling the attention mechanism to adaptively focus on the key change features of various loads (such as industrial, commercial, and residential), effectively improving the detection accuracy in the case of load mutations.

[0110] This solution calculates the electrical distance between nodes based on power flow analysis, constructs a comprehensive distance matrix by combining the geographical distance, converts it into a similarity index using the Gaussian kernel function, and forms a sparse but high-quality electricity meter association network. By calculating the comprehensive similarity of the load similarity and the network topology relationship, accurate grouping of similar electricity meters is achieved. An attention model is constructed for the representative electricity meters in the data-rich area, and by designing a transfer weight function based on the data sufficiency, effective transfer of knowledge from the data-rich area to the sparse area is realized. As data accumulates, the system can smoothly transition to the autonomous learning mode, greatly improving the accuracy of the electricity meter status assessment in areas with uneven spatial distribution and achieving consistent high-precision assessment within the global power grid.

[0111] Example 1: Apply the method of the present invention to evaluate the status of electricity meters in the distribution network of Area B, City A. This area contains 43 electricity meters, including 12 industrial user electricity meters, 18 commercial user electricity meters, and 13 residential user electricity meters. Some of the electricity meters have been installed and used for more than 5 years, and there is a risk of status degradation to varying degrees. The specific steps are as follows:

[0112] 1. Acquisition and preprocessing of electricity meter data

[0113] Select the data of one industrial user electricity meter (ID: M23576) for detailed description. The collection period of this electricity meter is 15 minutes, and data for one week (from June 5, 2023, to June 11, 2023) is collected, including voltage, current, active power, reactive power, and power factor, as shown in Table 1 below.

[0114] Table 1 Original electricity meter data fragment (partial)

[0115] Timestamp Voltage (V) Current (A) Active Power (kW) Reactive Power (kVar) Power Factor 2023-06-05 08:00 384.25 63.42 38.76 12.33 0.95 2023-06-05 08:15 385.11 78.63 47.21 15.42 0.94 2023-06-05 08:30 383.94 82.15 49.86 16.25 0.94 2023-06-05 08:45 382.76 85.27 51.43 17.08 0.93 ... ... ... ... ... ... 2023-06-05 14:15 379.43 98.75 58.94 19.35 0.94 2023-06-05 14:30 NULL NULL NULL NULL NULL 2023-06-05 14:45 378.92 96.83 57.73 18.94 0.94 ... ... ... ... ... ...

[0116] S11. Multi-source heterogeneous data collection: Collect the original electrical parameter time series from electricity meter M23576, and at the same time obtain the power grid topology data and the GPS location information of the electricity meter (N34°15'42", E108°42'37");

[0117] S12. Data Quality Assessment: Calculate the data completeness rate CR = 667 / 672 = 0.9926. Five data points are identified as missing. At the same time, outliers are detected: the current value at 10:30 on June 7, 2023 is 215.83 A, which is much higher than the mean value of 89.25 A and exceeds 3 standard deviations (3σ = 104.13), and is marked as an outlier; the voltage value at 02:15 on June 9, 2023 is 352.14 V, which is significantly lower than the mean value of 380.25 V and exceeds 3 standard deviations (3σ = 12.27), and is marked as an outlier;

[0118] S13. Data Cleaning and Standardization: Apply forward filling to the missing data: the data at 14:30 on June 5, 2023 is filled with [379.43, 98.75, 58.94, 19.35, 0.94]; apply median filtering to the abnormal current value of 215.83 A: median{83.51, 84.62, 215.83, 86.14, 85.97} = 85.97 A; apply median filtering to the abnormal voltage value of 352.14 V: median{378.82, 379.15, 352.14, 380.23, 380.45} = 379.15 V; perform standardization processing on the cleaned data, including: voltage standardization: Vnorm(t)=(V(t) - 380.25) / 4.09; current standardization: Inorm(t)=(I(t) - 89.25) / 34.71; active power standardization: Pnorm(t)=(P(t) - 55.36) / 21.34;

[0119] The standardized data segments are shown in Table 2 below:

[0120] Table 2 Standardized Electric Energy Meter Data Segments (Partial)

[0121] Timestamp Vnorm Inorm Pnorm Qnorm PFnorm 2023-06-05 08:00 0.98 -0.74 -0.78 -0.85 0.63 2023-06-05 08:15 1.19 -0.31 -0.38 -0.42 0.42 ... ... ... ... ... ...

[0122] 2. Multi-scale Time-domain Feature Adaptive Decomposition

[0123] S21. Signal Complexity Adaptive Assessment: Calculate the signal complexity of the active power standardized sequence Pnorm: approximate entropy ApEn(Pnorm) = 0.825; sample entropy SampEn(Pnorm) = 0.762; complexity index CI = 0.7×0.825 + 0.3×0.762 = 0.806; determine the optimal number of decomposition modes K = ceil(3×0.806) = 3 according to the CI value, and generate the decomposition parameter configuration DPC={K = 3, α = 2000, τ = 0.1};

[0124] S22. Self-optimizing variational mode decomposition:

[0125] S221. Initial mode decomposition parameter configuration: Set the initial center frequencies ω1 = π / 4, ω2 = π / 2, ω3 = 3π / 4, and generate the initial decomposition parameter set IDP = {Kinit = 3, {ω1, ω2, ω3}, α = 2000, τ = 0.1};

[0126] S222. Coarse-grained variational mode base decomposition: Apply the VMD algorithm to the normalized active power sequence Pnorm: Construct the augmented Lagrangian function; Converge after 300 iterations to obtain the coarse-grained mode function set CMF = {u1, u2, u3}; The main characteristics of each mode:

[0127] u1: Center frequency 0.0208 Hz (period about 48 hours), representing the daily load variation;

[0128] u2: Center frequency 0.0696 Hz (period about 14.4 hours), representing the load variation during the working period;

[0129] u3: Center frequency 0.4167 Hz (period about 2.4 hours), representing short-term load fluctuations;

[0130] S223. Execution of the incremental decomposition strategy: Calculate the energy retention rate of each mode: EPR1 = 0.674 (67.4%); EPR2 = 0.254 (25.4%); EPR3 = 0.072 (7.2%); Calculate the correlation matrix between modes: CorrM = [1.000, 0.132, 0.095; 0.132, 1.000, 0.583; 0.095, 0.583, 1.000]; It is found that the correlation between u2 and u3, CorrM23 = 0.583, is close to the threshold of 0.6 but does not exceed it, so it is not temporarily marked as a low-quality mode; The energy retention rate of u3, EPR3 = 0.072 > 0.05, is not marked as a low-quality mode; The quality of all modes meets the requirements, skip the incremental decomposition step, and set the optimized mode function set OMF = CMF.

[0131] S224. Decomposition quality verification and parameter fine-tuning:

[0132] Calculate the reconstruction error RE = 0.026 < 0.05; Calculate the orthogonality index OI = 0.27 < 0.3; The decomposition quality inspection passes, and confirm the final intrinsic mode function set FIMFs = {u1, u2, u3};

[0133] S23. Multi-scale feature extraction and time-domain attribution recognition:

[0134] Extract features from the three IMFs respectively, and the results are shown in Table 3 below:

[0135] Table 3 IMF Feature Extraction Results

[0136] IMF Mean Variance Kurtosis Skewness Main Frequency (Hz) Sample Entropy Time Scale Attribution u1 0.012 0.442 2.753 0.124 0.0208 0.614 Daily u2 0.005 0.253 3.126 0.217 0.0696 0.752 Hourly u3 0.002 0.078 4.831 0.356 0.4167 0.891 Hourly

[0137] Generate the multi-scale time-domain feature set MTFS and the time-scale attribution mapping TSM.

[0138] 3. Power Grid Spatial Topological Relationship Construction

[0139] S31. Power Grid Topological Data Parsing and Node Mapping: The electricity meter M23576 is connected to the node N087 of the 10kV distribution line L2568; establish the mapping relationship MNM(M23576)=N087;

[0140] S32. Spatial Association Strength Quantification: Based on power flow analysis, calculate the electrical distance between node N087 and other nodes, such as EDN087,N092 = 1 / 0.238 = 4.202; combined with the geographical distance, such as GDN087,N092 = 283 meters, calculate the spatial association strength: SCRMN087,N09 = 0.7×(1 / 4.202)+0.3×(1 / 0.283)=0.167 + 1.060 = 1.227; similarly calculate the spatial association strength matrix SCRM between M23576 and other electricity meters in the area;

[0141] S33. Power Grid Topological Dynamic Adaptability Update: Monitor that the state of the power grid switch SW2345 has changed on June 8, 2023, update the spatial association strength matrix, and obtain the dynamically updated spatial association graph DSRG.

[0142] 4. Spatiotemporal Attention Calculation for Load Characteristic Perception

[0143] S41. Load Characteristic Identification and Classification: Extract the load characteristics of M23576, including: daily load curve shape LC(t); peak-to-valley ratio PVR = 65.42 / 17.58 = 3.72; intra-day fluctuation coefficient VC = 21.34 / 55.36 = 0.385; construct the load characteristic dictionary LFD, which contains the characteristics of different types of loads:

[0144] Large industrial load characteristics F1 = [3.85, 0.392,...];

[0145] Medium industrial load characteristics F2 = [2.73, 0.415,...]; ...;

[0147] Resident load characteristic F10 = [1.75, 0.203,...]; Applying the sparse coding algorithm, represent the load of M23576 as a linear combination of the feature dictionary: L = 0.82·F2 + 0.15·F3 + 0.03·F1 + ε; Based on the leading coefficient 0.82 corresponding to F2 (medium industrial load), determine the load type identifier LTI = "medium industrial", and the load sparse representation coefficient LSC = [0.03, 0.82, 0.15, 0, 0, 0, 0, 0, 0, 0];

[0148] S42. Hierarchical self-adjusting attention skeleton construction: Select the corresponding initial attention template according to LTI = "medium industrial", and design a lightweight attention skeleton (edge side) and a full-feature attention model (cloud side); The resource allocation strategy is set as RASk = log(1 + Importancek), where Importancek doubles the weight during peak periods (8:00 - 12:00 in the morning, 13:00 - 17:00 in the afternoon);

[0149] S43. Dynamic attention modulation enhanced by pattern migration:

[0150] S431. Grid topology association network construction: Extract the connection relationship from DSRG to form an adjacency matrix A. For example, AM23576,M23581 = 1 indicates that two electricity meters are connected to the same line; Apply the Gaussian kernel function to convert the comprehensive distance into similarity: SM23576,M23581 = exp(-1.346 2 / 3.5 2 ) = 0.859 > λ(0.7); Retain the connections with similarity higher than 0.7 to form the electricity meter association network MARN; M23576 forms effective connections with 5 electricity meters;

[0151] S432. Similar load electricity meter identification and grouping: Calculate the load similarity: LSM23576,M23581 = cosine_similarity([0.03,0.82,0.15,...], [0.05,0.79,0.13,...]) = 0.984; Combine the topological similarity and the load similarity to calculate the comprehensive similarity: CSM23576,M23581 = 0.6×0.859 + 0.4×0.984 = 0.515 + 0.394 = 0.909; Apply the Louvain community detection algorithm to group the association network. M23576 is assigned to group G3, which contains 4 electricity meters, and the representative electricity meter is M23581 (MRS = 0.95);

[0152] S433. Design of Attention Knowledge Transfer Mechanism: The installation time of M23576 is May 2019, and the data sufficiency calculation is Di = 0.87; Design the transfer weight function: TWi(t) = exp(-2.5×0.87) = 0.114; Construct the hybrid attention model: AMM23576 = 0.114×AMM23581 + 0.886×AMM23576^local;

[0153] S434. Construction of Frequency-Domain Attention Modulation Function: Apply FFT to the load sparse representation coefficients LSC = [0.03, 0.82, 0.15, 0, 0, 0, 0, 0, 0, 0] to obtain the frequency-domain representation F_LSC; Design the frequency response function to enhance the high-frequency component response for medium industrial loads: H(f) = 1.6 for f > 0.1Hz (corresponding to mutation monitoring); H(f) = 1.2 for 1 / (24*3600)Hz < f < 1 / (3600)Hz (corresponding to the working period); H(f) = 1.0 otherwise;

[0154] Construct the frequency-domain modulation matrix D(C): D(C) = diag([1.03, 1.42, 1.36, 1.21, 1.06, 1.00, 1.00, 1.00]); Apply the modified attention calculation formula: Attention(Q, K, V, C) = softmax(QK^T·D(C) / √d)V; Obtain the dynamically modulated spatio-temporal attention matrix DTSAM (partial): DTSAM = [0.058, 0.032, 0.019,..., 0.012; 0.085, 0.124, 0.075,..., 0.018;... 0.027, 0.042, 0.163,..., 0.021].

[0155] 5. Multi-Scale Spatio-Temporal Feature Fusion and State Representation

[0156] S51. Construction of Hierarchical Time-Domain Feature Map:

[0157] S511. Definition of Time Scale Hierarchy and Feature Mapping: Establish the time scale hierarchy: L1 = second level (no corresponding IMF), L2 = minute level (no corresponding IMF), L3 = hour level (corresponding to u2, u3), L4 = day level (corresponding to u1), L5 = week level (no corresponding IMF); Calculate the information entropy of each scale: Entropy(L3) = 1.843 Entropy(L4) = 1.256; Obtain the importance indicators: II3 = 1.843 / (1.843 + 1.256) = 0.595, II4 = 1.256 / (1.843 + 1.256) = 0.405;

[0158] S512, Cross-scale Causal Relationship Identification: Conduct Granger causality tests from daily level to hourly level: p-value = 0.003 < 0.05, the causal relationship holds, GrangerCausality(L4→L3) = 0.687; conduct Granger causality tests from hourly level to daily level: p-value = 0.241 > 0.05, the causal relationship does not hold, GrangerCausality(L3→L4) = 0.114; calculate the causal delay parameter τ43 = 3, indicating that it takes about 3 hours for daily-level features to affect hourly-level features;

[0159] S513, Graph Structure Construction and Attribute Definition: Create a hierarchical directed graph G = (V, E): The node set V = {v3, v4}, representing hourly-level and daily-level features respectively; the edge set E = {e43}, indicating the causal relationship from daily level to hourly level; the node weights W(v3) = 0.595, W(v4) = 0.405; the edge weight W(e43) = 0.687, and the delay attribute D(e43) = 3;

[0160] S514, Graph Dynamic Update and Evolution Tracking: Set the update window length T = 24 hours and the sliding step S = 1 hour, and perform rolling updates on the data from June 5th to June 11th; a change in the graph structure is found in the window at 13:00 on June 7th: the edge weight W(e43) drops from 0.687 to 0.532; a new strong autocorrelation pattern appears within the hourly level; this change is close to the abnormal data time at 10:30 on June 7th and may be a sign of abnormality;

[0161] S52, Attention-guided Spatiotemporal Feature Selection: Sort the features in MTFS according to DTSAM, and select the top 80% important features to form the preferred feature subset SFS;

[0162] S53, Adaptive Feature Fusion and State Representation: Apply the graph-guided feature fusion strategy, including: the feature weight wi is determined according to the node weight and attention allocation; the feature interaction weight wij is determined according to the edge weight; construct the state representation vector MSRV, which contains 43-dimensional information.

[0163] 6. Electric Energy Meter State Evaluation and Abnormal Diagnosis

[0164] S61, State Deviation Quantification and Scoring: Calculate the Mahalanobis distance MD = 2.86 and normalize it to the state evaluation score SES = 1 - exp(-2.86 / 3) = 0.615;

[0165] S62. Abnormal pattern recognition and fault classification: SES = 0.615 < 0.7, not reaching the abnormal threshold but close, determined to be in the "to be concerned" state; matching with the templates in the abnormal pattern library, the pattern with the highest similarity is the "slow drift of electricity meter parameters" pattern (similarity 0.783).

[0166] S63. Interpretability analysis of evaluation results: Locate that the main abnormal contribution comes from: the u3 mode at the hourly level (contribution degree 52.7%); the change in the causal relationship strength from the daily level to the hourly level (contribution degree 31.5%); Generate the evaluation conclusion: "The electricity meter M23576 is in the early stage of slow parameter drift. It is recommended to strengthen monitoring. The main abnormal manifestation is the change in the short-term load response characteristics, and it is expected to develop into an obvious fault within 3 - 6 months."

[0167] Comparison of experimental results: Compare the test results of the method of the present invention with traditional methods on 43 electricity meters in the distribution network of Area B, City A. The comparison results are shown in Table 4 below:

[0168] Table 4 Performance comparison of different methods

[0169] Evaluation Index Traditional Method Method of the Present Invention Improvement Rate Abnormal Detection Accuracy 78.3% 93.2% 19.0% Fault Type Identification Accuracy 72.1% 85.7% 18.9% Early Detection Rate of Parameter Drift 53.4% 81.2% 52.1% Average Lead Time of Abnormal Detection 4.2 days 12.6 days 200.0% False Alarm Rate 8.5% 3.2% 62.4%

[0170] The method of the present invention is particularly outstanding in the early detection of parameter drift, and can discover potential problems 12.6 days in advance, providing sufficient response time for maintenance personnel.

[0171] In the subsequent 6 - month actual operation, a total of 7 electricity meter faults were captured by the early warning based on this method. Among them, 5 were parameter drift - type faults, with an accuracy rate of 85.7%, and about 45% of the maintenance cost was saved.

[0172] Example 2. In the distribution network of Area B, City A, select a group of typical electricity meters for analysis. Select the industrial user electricity meter numbered M23576 and the commercial user electricity meter numbered M23592 as examples to show the complete process of constructing a hierarchical time - domain feature map.

[0173] Step 1. Define the time - scale hierarchical structure from the second - level to the month - level

[0174] First, define the complete time - scale hierarchical structure as follows: L1: Second - level scale (1 - 60 seconds); L2: Minute - level scale (1 - 60 minutes); L3: Hour - level scale (1 - 24 hours); L4: Day - level scale (1 - 7 days); L5: Week - level scale (1 - 4 weeks); L6: Month - level scale (1 - 12 months).

[0175] For the electricity meter M23576, 3 intrinsic mode functions (IMFs) have been obtained through previous self-optimized variational mode decomposition. By analyzing the central frequencies and periodic characteristics of each IMF, it is determined that: u1: central frequency 0.0208 Hz (period approximately 48 hours), belonging to L4 (daily scale); u2: central frequency 0.0696 Hz (period approximately 14.4 hours), belonging to L3 (hourly scale); u3: central frequency 0.4167 Hz (period approximately 2.4 hours), belonging to L3 (hourly scale).

[0176] For the electricity meter M23592, 4 intrinsic mode functions are obtained: u1': central frequency 0.0104 Hz (period approximately 96 hours), belonging to L4 (daily scale); u2': central frequency 0.0417 Hz (period approximately 24 hours), belonging to L4 (daily scale); u3': central frequency 0.1389 Hz (period approximately 7.2 hours), belonging to L3 (hourly scale); u4': central frequency 0.6944 Hz (period approximately 1.44 hours), belonging to L3 (hourly scale);

[0177] Based on the above analysis, time scale attribution mapping tables TSM_M23576 and TSM_M23592 are established:

[0178] TSM_M23576 = {"u1": "L4","u2": "L3","u3": "L3"}; TSM_M23592 = {"u1'":"L4","u2'": "L4","u3'": "L3","u4'": "L3"}.

[0179] Step 2: Calculate the information entropy of each time scale level

[0180] Information entropy is used to measure the amount of information contained in each time scale and is a quantitative indicator of the importance of that time scale. It is calculated using the Shannon entropy formula: H(X) = -∑p(xi)log2p(xi); where p(xi) is the probability distribution of the eigenvalue xi.

[0181] For the electricity meter M23576, the characteristic entropies of the hourly level (L3) and the daily level (L4) are calculated respectively: First, the datasets of u2 and u3 corresponding to L3 are merged to form the hourly-level feature set FS_L3_M23576; the u1 data corresponding to L4 is used as the daily-level feature set FS_L4_M23576; the probability distribution of the feature set is estimated, and the data is divided into 10 equal-width intervals to calculate the probability density.

[0182] Calculation results: Entropy(L3)_M23576 = 1.843 bits; Entropy(L4)_M23576 = 1.256 bits.

[0183] For electricity meter M23592, the same calculations are made: Entropy(L3)_M23592 = 1.683 bits; Entropy(L4)_M23592 = 2.145 bits.

[0184] Based on the above information entropy, calculate the importance indicators for each time scale:

[0185] For electricity meter M23576: II3 = 1.843 / (1.843 + 1.256) = 0.595; II4 = 1.256 / (1.843 + 1.256) = 0.405;

[0186] For electricity meter M23592: II3 = 1.683 / (1.683 + 2.145) = 0.440; II4 = 2.145 / (1.683 + 2.145) = 0.560.

[0187] These importance indicators will be used as the quantization basis for the node weights in the graph.

[0188] Step 3: Use Granger causality test to identify the causal relationship between different time scale features

[0189] The Granger causality test is a statistical hypothesis test used to determine whether one time series is helpful in predicting another time series. The following steps are used for Granger causality analysis:

[0190] First, perform a stationarity test on each time scale feature (using the augmented Dickey-Fuller test), and perform differencing if necessary to ensure data stationarity; use the Bayesian Information Criterion (BIC) to determine the optimal lag order; construct a VAR model: construct a vector autoregressive model based on the determined lag order;

[0191] Perform the Granger causality test and calculate the F statistic and p value; for electricity meter M23576, perform the following tests:

[0192] Granger causality test from daily level to hourly level (L4→L3): Use the feature sequence of L4 as the independent variable and the feature sequence of L3 as the dependent variable; determine the optimal lag order k = 3 (based on the principle of the minimum BIC); perform the test to obtain F statistic = 8.743, p value = 0.003 < 0.05;

[0193] There is a causal relationship, and the daily-level features have significant predictive power for the hourly-level features.

[0194] Causal strength calculation, GrangerCausality(L4→L3) = 0.687 (normalized value based on the VAR model coefficients)

[0195] Granger causality test from hourly level to daily level (L3→L4): Using the feature sequence of L3 as the independent variable and the feature sequence of L4 as the dependent variable; adopting the same lag order k = 3; performing the test to obtain the F statistic = 1.478 and p-value = 0.241 > 0.05; the causal relationship is not significant, and the hourly-level features have insufficient predictive power for the daily-level features; the causal strength is denoted as GrangerCausality(L3→L4) = 0.114 (although not significant).

[0196] For the electricity meter M23592, performing a similar test gives: GrangerCausality(L4→L3)_M23592 = 0.752, p-value = 0.001 < 0.05 (significant); GrangerCausality(L3→L4)_M23592 = 0.328, p-value = 0.047 < 0.05 (significant but with weak strength);

[0197] Step 4: Construct a cross-scale causal relationship matrix and a causal delay parameter table

[0198] Based on the results of the Granger causality test, construct a cross-scale causal relationship matrix CCRM. For each pair of time scales Li and Lj, CCRM(i,j) represents the causal strength from Li to Lj.

[0199] For the electricity meter M23576, the causal relationship matrix is:

[0200] CCRM_M23576 = [ [0, 0, 0, 0, 0, 0], L1->L1, L1->L2,...

[0201] [0, 0, 0, 0, 0, 0], L2->L1, L2->L2,...

[0202] [0, 0, 0, 0.114,0, 0], L3->L1, L3->L2,...

[0203] [0, 0, 0.687,0, 0, 0], L4->L1, L4->L2,...

[0204] [0, 0, 0, 0, 0, 0], L5->L1, L5->L2, ...

[0205] [0, 0, 0, 0, 0, 0] L6->L1, L6->L2, ...]

[0206] For the electricity meter M23592, the causality matrix is:

[0207] CCRM_M23592 = [ [0, 0, 0, 0, 0, 0],

[0208] [0, 0, 0, 0, 0, 0],

[0209] [0, 0, 0, 0.328,0, 0],

[0210] [0, 0, 0.752,0, 0, 0],

[0211] [0, 0, 0, 0, 0, 0],

[0212] [0, 0, 0, 0, 0, 0]]

[0213] In addition, it is also necessary to determine the causal delay parameter, which represents the time required for the causal effect to propagate from one time scale to another. This is determined by the optimal lag order in the Granger causality test.

[0214] Causal delay parameter table for electricity meter M23576:

[0215] CDP_M23576 = {

[0216] "L4→L3": 3, The daily-level feature affects the hourly-level feature and it takes about 3 hours;

[0217] "L3→L4": 5 The hourly-level feature affects the daily-level feature and it takes about 5 hours (although this causal relationship is not significant)

[0218] }

[0219] Causal delay parameter table for electricity meter M23592:

[0220] CDP_M23592 = { "L4→L3": 2, The daily-level feature affects the hourly-level feature and it takes about 2 hours;

[0221] "L3→L4": 6 The hourly-level feature affects the daily-level feature and it takes about 6 hours.}

[0222] Step 5: Construct a hierarchical directed graph based on the time-scale hierarchy and the cross-scale causality matrix

[0223] Finally, based on the analysis results of the previous steps, construct a hierarchical directed graph G = (V, E):

[0224] The node set V contains the features corresponding to each time scale; the edge set E represents the causal relationships between different time scales; the node weights are importance indicators calculated based on information entropy; the edge weights are based on Granger causal strength; the direction of the edge represents the direction of causal action; the attributes of the edge include delay parameters; for the hierarchical directed graph G_M23576 of the electricity meter M23576: the node set V = {v3, v4}, representing the hourly and daily features respectively; the node weights W(v3) = 0.595, W(v4) = 0.405;

[0225] The edge set E = {e43}, representing the edge from v4 to v3 (from daily level to hourly level); the edge weight W(e43) = 0.687, representing the causal strength from daily level to hourly level; the edge delay D(e43) = 3, representing the delay time of 3 hours from daily level to hourly level; for the hierarchical directed graph G_M23592 of the electricity meter M23592: the node set V = {v3, v4}, representing the hourly and daily features respectively; the node weights W(v3) = 0.440, W(v4) = 0.560; the edge set E = {e43, e34}, representing a two-way causal relationship; the edge weights W(e43) = 0.752, W(e34) = 0.328; the edge delays D(e43) = 2, D(e34) = 6; in addition, to more completely represent the hierarchical time-domain feature map, the following attributes are attached to the nodes in the graph: node feature vector F(v): including the statistical features (such as mean, variance, kurtosis, skewness, etc.) of this time scale; node frequency-domain characteristic FS(v): including the main frequency components of this time scale; node non-linear feature NL(v): including non-linear features such as sample entropy;

[0226] In an abnormal monitoring of an electricity meter, a tracking analysis was carried out on the hierarchical time-domain feature map of the M23576 electricity meter, and it was found that: during the period from 00:00 on July 15, 2023 to 23:59 on July 17, 2023, the structure of the hierarchical time-domain feature map changed significantly: the causal strength W(e43) from daily level to hourly level decreased from 0.687 to 0.423; the information entropy of the hourly-level feature increased from 1.843 to 2.217; a new strong autocorrelation pattern appeared within the hourly level;

[0227] These spectral structure changes preceded the anomalies detected by traditional methods in the electricity meter, and the early signs of impedance anomalies in the voltage measurement circuit of the electricity meter were discovered approximately 9 days in advance. The dynamic change trajectory of the hierarchical time-domain feature spectrum shows that the anomaly initially appears in the hourly features, and then affects the daily features through the causal path in the spectrum, ultimately leading to a decrease in the overall electricity meter status score.

[0228] In summary, the present invention designs an optimal mode number determination algorithm based on adaptive evaluation of signal complexity, and introduces sparse optimization constraints and an incremental decomposition strategy to solve the problem of parameter presetting in the traditional VMD algorithm. It does not require manual predefinition of time windows and scale parameters, can automatically adapt to the diverse characteristics of different electricity meters, and reduce the mode mixing problem; the present invention proposes an attention knowledge transfer mechanism based on the power grid topology structure and designs a frequency-domain attention modulation function to achieve dynamic attention adjustment for load characteristic perception, solving the cold start problem of newly installed or data-deficient electricity meters and being able to dynamically adjust attention allocation according to the load characteristics of different types of users; the present invention proposes a cross-time-domain feature spectrum representation method for the electricity meter status, establishes the relationship between nodes through Granger causality test, realizes the structured organization of multi-scale features, can track the evolution path of faults at different time scales, and improves the recognition ability for complex anomalies; the present invention introduces the spectral structure information into the feature fusion process, captures the non-linear relationship between features through the second-order interaction term, realizes the context-aware dynamic feature recombination, effectively balances the importance of features at different time scales, and improves the recognition ability for slowly evolving fault states.

[0229] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. A method for evaluating the state of an electric energy meter based on a spatio-temporal attention mechanism, characterized in that It includes the following steps: Obtain the original data of the electricity meter, preprocess it to obtain a standard electricity meter dataset, and perform multi-scale time-domain feature adaptive decomposition on it to obtain a multi-scale time-domain feature set; Construct the grid space topology relationship based on the grid topology data and the electricity meter location information to obtain a spatial association map; Combine the multi-scale time-domain feature set and the spatial association map, calculate the spatio-temporal attention weights for load characteristic perception, and obtain a dynamically modulated spatio-temporal attention matrix; Based on the dynamically modulated spatio-temporal attention matrix, adaptively fuse the multi-scale time-domain feature set to construct an electricity meter state representation vector; And perform state evaluation and anomaly diagnosis to obtain the electricity meter state evaluation result; The steps of calculating the spatio-temporal attention weights for load characteristic perception include: Extract load characteristics based on the multi-scale time-domain feature set to construct a load characteristic dictionary; Represent the current load as a linear combination of the feature dictionary to obtain the load type identifier and the load sparse representation coefficient, and accordingly construct a hierarchical self-adjusting attention calculation framework and a resource allocation strategy; Combine the spatial association map, the load sparse representation coefficient, and the attention calculation framework to achieve dynamic attention modulation with enhanced pattern migration, and obtain a dynamically modulated spatio-temporal attention matrix; The steps of achieving dynamic attention modulation with enhanced pattern migration include: Based on the spatial association map, establish an electricity meter association network with similar load characteristics, obtain the load similarity and the network topology relationship, and calculate the comprehensive similarity between electricity meters accordingly; Construct an attention model for electricity meters with sufficient data. For electricity meters with insufficient data, select similar electricity meters from the electricity meter association network according to the comprehensive similarity, and construct a progressive migration weight function based on the data sufficiency; Apply the fast Fourier transform to the load sparse representation coefficient to construct a frequency-domain attention modulation function that dynamically adjusts the frequency response gain for different load types; Modify the attention model calculation formula according to the migration weight function and the frequency-domain attention modulation function to obtain a dynamically modulated spatio-temporal attention matrix.

2. The method according to claim 1, characterized in that, The multi-scale time-domain feature adaptive decomposition of the standard electricity meter dataset includes: Calculate the signal complexity index of the standard electricity meter dataset, and accordingly determine the optimal number of decomposition modes to generate a decomposition parameter configuration; According to the decomposition parameter configuration, use the self-optimizing variational mode decomposition algorithm to process the standard electricity meter dataset to obtain an ensemble of intrinsic mode functions; Perform feature extraction on the ensemble of intrinsic mode functions, calculate the time-domain statistical features, frequency-domain features, and non-linear features, and map each mode function to a time scale to generate a multi-scale time-domain feature set and a time scale attribution mapping.

3. The method according to claim 2, wherein The steps of processing using the self-optimizing variational mode decomposition algorithm include: Determine the initial number of modes and the center frequency according to the decomposition parameter configuration to generate an initial decomposition parameter set; Use the initial decomposition parameter set to perform the variational mode decomposition algorithm on the standard electricity meter dataset to obtain a coarse-grained ensemble of mode functions and a mode quality evaluation index; Identify the low-quality modes in the coarse-grained ensemble of mode functions based on the mode quality evaluation index, recursively perform variational decomposition on each low-quality mode, and apply sparse optimization constraints to obtain an optimized ensemble of mode functions; Verify the decomposition quality of the optimized modal function set, calculate the reconstruction error and orthogonality index, confirm or adjust the decomposition parameters according to the verification results, and output the intrinsic modal function set.

4. The method according to claim 3, characterized in that, The steps of identifying low-quality modes based on modal quality evaluation indicators and recursively performing variational decomposition include: Calculate the energy retention rate and the inter-modal correlation matrix for each mode in the coarse-grained modal function set; and identify the modes with an energy retention rate lower than the preset threshold or an inter-modal correlation higher than the preset threshold based on this, forming a low-quality mode set. Perform variational mode decomposition on each modal signal in the low-quality mode set as a new input signal, set the initial number of modes to 2, and generate the corresponding sub-mode set. Combine the sub-modes with the stored original high-quality modal function set and apply the objective function to minimize the inter-modal aliasing, and output the optimized modal function set.

5. The method according to claim 1, characterized in that, The steps of constructing a load feature dictionary and representing the current load as a linear combination of the feature dictionary include: Extract the typical load features from the multi-scale time-domain feature set, including the daily load curve shape, peak-valley ratio, and volatility features. Construct a load feature dictionary based on the typical load features, including the feature representations of industrial, commercial, and residential loads. Represent the current load as a linear combination of the basis vectors in the load feature dictionary to obtain the sparse coefficients. Classify the load based on the sparse coefficients, determine the load type identifier, and output the sparse coefficients as the load sparse representation coefficients.

6. The method according to claim 1, characterized in that, Construct a progressive transfer weight function, including: Based on the electricity meter association network and the node similarity matrix, group the electricity meters by similarity and calculate the representative electricity meter for each group. Construct an attention model for representative energy meters with sufficient data, identify the groups to which energy meters with insufficient data belong and the corresponding representative energy meters, and design a transfer weight function TW i (t)=exp(-γ·D i (t)), where D i (t) is the data sufficiency of energy meter i, and γ is the attenuation coefficient; Construct the hybrid attention model AM i =TW i (t)·AM r +(1 - TW i (t))·AM i local , where AM r and AM i local are the attention models of representative watt-hour meters and local attention models.

7. The method according to claim 1, wherein The steps of adaptively fusing the multi-scale time-domain feature set include: Construct a hierarchical time-domain feature map based on the multi-scale time-domain feature set and the time scale attribution mapping to represent the structured relationship between different time scale features. Rank the multi-scale time-domain feature set according to the dynamically modulated spatio-temporal attention matrix, and select the most discriminative feature subset. Utilize the hierarchical time-domain feature map and the dynamically modulated spatio-temporal attention matrix to perform map-guided adaptive feature fusion and construct a power meter state representation vector containing multi-scale time-domain information and spatial association information.

8. The method according to claim 1, characterized in that, The steps of constructing a hierarchical time-domain feature map include: Define the time scale hierarchical structure from the second level to the monthly level, and map the features in the multi-scale time-domain feature set to the corresponding time scales according to the time scale attribution mapping. Calculate the information entropy of each time scale level as an important indicator of the amount of information contained in that time scale. Use Granger causality test to identify the causal relationships between different time scale features, and construct a cross-scale causal relationship matrix and a causal delay parameter table. Based on the time scale hierarchical structure and the cross-scale causal relationship matrix, construct a hierarchical directed graph, where the nodes in the graph represent the features of different time scales and the edges represent the cross-scale causal relationships.

Images