A smart grid power distribution line fault prediction analysis method

Abstract

The present application relates to a kind of smart grid distribution line fault prediction analysis method, this method is through distributed sampling and accurate spatial orientation, the real-time load of each sampling point of distribution line, electrical parameter, environmental information and component aging state are synchronously collected and standardized processing, then based on multi-time scale dynamic characteristic and aging characteristic extraction, realize the automatic generation of structured, hierarchical feature library and scene label.Utilize label-driven historical model parameter migration and meta-learning mechanism, quickly adapt to new working condition and fine-tune model weight, fuse multi-time scale features for contribution weighting, finally improve the accuracy and robustness of fault prediction.The system also continuously optimizes model performance through real-time A / B comparison and automatic parameter switching, has data tracing and result interpretation capability, significantly enhances the timeliness, reliability and intelligent level of power distribution network fault diagnosis.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of "smart grid fault prediction and adaptive model", in particular to a smart grid distribution line fault prediction analysis method. BACKGROUND

[0002] At present, the fault prediction and early warning technology of the smart grid distribution line has become an important research direction in the field of safe operation and maintenance of the distribution system. With the continuous progress of intelligent sensing and data fusion technology, methods based on machine learning and deep learning models gradually replace traditional fault monitoring methods based on rules or physical thresholds, and become the mainstream development trend. The existing technology mainly focuses on the fusion modeling of multi-source data, uses historical operation monitoring data, equipment aging parameters and environmental variables to build a data-driven fault prediction model, and improves the recognition accuracy of typical working conditions and known fault types through time series analysis, electrical feature discrimination and state quantity statistics. This kind of method mostly uses convolutional neural network, time series recurrent neural network and fusion type end-to-end learning model architecture, and focuses on the generalization ability and batch training effect of the model in the scene of large sample and synchronous data richness.

[0003] Although the existing smart distribution grid fault prediction technology has certain online learning and incremental fine-tuning ability, it still has the technical shortcomings of insufficient model adaptability to new scenes, low utilization rate of small sample data, and rigid feature weight distribution when facing highly dynamic and complex working condition environment. These defects seriously restrict the ability of the smart grid system to deal with sudden operating states, autonomous emergency adjustment and improve fault warning accuracy, and have become the core bottleneck restricting the application promotion and high-reliability fault diagnosis of the industry. SUMMARY

[0004] The present application provides a smart grid distribution line fault prediction analysis method, which aims to solve the problems or one of the problems mentioned in the background technology.

[0005] The smart grid distribution line fault prediction analysis method provided by the present application specifically comprises:

[0006] S1: Collecting the operation indexes and environmental sensing data of a plurality of sampling points in the distribution line, the operation indexes including real-time load, voltage, current, harmonic and environmental temperature and humidity, climate variables, recording the unique identification of the components of each sampling point and the corresponding aging state information.

[0007] S2: Denoising, outlier rejection and normalization processing are performed on the collected multi-source operation indexes and environmental sensing data to generate a high-quality data set, and the unique identification of each data item associated with the components and the aging state information are maintained.

[0008] S3: Based on the normalized data set containing the aging state of the components, multi-time scale features are extracted, including short-period dynamic operation features and long-period aging features, to form a continuous feature vector group, and the equipment and sampling point position to which each feature vector belongs are identified.

[0009] S4: Determine whether there is a working condition change in the real-time operation data stream, if the change amplitude of the dynamic operation feature exceeds the preset threshold, then based on the sampling point environment data and the feature vector, the corresponding dynamic scene label is automatically generated, and the label node includes load fluctuation, extreme climate or abnormal power grid event information.

[0010] S5: Based on the generated dynamic scene label, the fault prediction model parameters and feature importance weights under the similar working condition in the history are migrated to the current new scene to initialize the fault prediction model weight vector under the current scene.

[0011] S6: For the new working condition corresponding to the current dynamic scene label, the feature weight of the fault prediction model is incrementally fine-tuned through the few-shot meta-learning algorithm with the limited new sampling sample and its feature vector as input, realizing the adaptive calibration of the model to the feature distribution of the new working condition.

[0012] S7: For the adaptively calibrated fault prediction model, the dynamic running feature weight and the component aging feature weight are integrated, the feature contribution degree is dynamically weighted according to the dynamic scene label, and the final multi-time scale fusion prediction factor is formed.

[0013] S8: The multi-time scale fusion prediction factor is taken as input to drive the model to output the power distribution line fault prediction result of the corresponding scene, and feature importance explanation information is generated for each target sampling point.

[0014] S9: Real-time monitoring of the prediction accuracy and stability of the adaptively calibrated model, A / B comparative analysis of the prediction results and the original model, if the short-time accuracy and stability of the new model are better than those of the original model, then automatically switch to the new model, otherwise roll back to the original model parameter state, and record the tuning process log for subsequent performance optimization.

[0015] The intelligent power grid distribution line fault prediction analysis method provided by the application has the following beneficial effects:

[0016] The application designs a distributed sampling point space positioning and multi-channel acquisition scheme, realizes deep binding of operation indicators, environmental perception, device health and spatial information, guarantees the structured, consistent and traceable of multi-source data, and provides a solid foundation for feature extraction and subsequent modeling. Compared with the existing prediction method which only relies on single point or a small amount of operation parameters, the multi-dimensional data fusion and spatial distribution processing of the application significantly improves the data integrity and feature coverage, so that the model can perceive the multi-point dynamic changes of the power grid.

[0017] By employing noise suppression, outlier removal, and normalized chain-like data cleaning, combined with the extraction and splicing of multiple features across multiple time scales (short-cycle dynamics and long-cycle degradation), this invention effectively overcomes the problem of data heterogeneity, such as different sampling periods and different component health states. Compared with traditional methods that only use a single time window or static features for modeling, this invention greatly enriches the expression of model input.

[0018] The innovative dynamic scene label generation and label transfer mechanism can capture sudden changes in operating conditions (such as drastic load fluctuations, extreme weather, or abnormal power grid events) in real time, and efficiently transfer model parameters and weights from similar historical scenarios to new operating conditions, achieving automatic adaptation of weights and feature space. Compared with traditional model training methods that require a large number of new samples or manual parameter adjustment, this solution directly shortens the parameter adjustment lag and achieves near real-time model response capability.

[0019] This invention integrates few-sample meta-learning and feature enhancement techniques. With limited labeled samples for new working conditions, it enables rapid pre-calibration and incremental fine-tuning of model feature weights, avoiding problems such as decreased model generalization ability or severe false alarms caused by insufficient samples. Compared to existing purely supervised large-sample learning schemes, this technology significantly improves prediction accuracy with only a fraction of the sample size; when rapidly deployed in new scenarios, the model's false alarm rate is significantly reduced, and the average response time for detecting sudden faults is significantly shortened.

[0020] By employing a dynamic weighted algorithm based on feature contribution, dynamic operation and component aging characteristics are fused according to scenario label weights to form multi-timescale predictive factors, enabling real-time evaluation of various influencing factors under dynamic and complex scenarios. This mechanism effectively ensures that the model consistently outputs the optimal feature weight distribution when faced with sudden load changes, aging equipment, and extreme environments, thereby improving model interpretability and early warning sensitivity. Attached Figure Description

[0021] Figure 1 This is the main flowchart of a method for predicting and analyzing faults in smart grid distribution lines.

[0022] Figure 2 This is a sub-flowchart of a method for predicting and analyzing faults in smart grid distribution lines.

[0023] Figure 3 This is another sub-flowchart of a method for predicting and analyzing faults in smart grid distribution lines. Detailed Implementation

[0024] 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.

[0025] 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.

[0026] As attached Figure 1 As shown, this application provides a method for predicting and analyzing faults in smart grid distribution lines, specifically including:

[0027] S1: Collect operating indicators and environmental sensing data from multiple sampling points in the power distribution line. The operating indicators include real-time load, voltage, current, harmonics, ambient temperature and humidity, and climate variables. Record the unique identifier of each component and its corresponding aging status information at each sampling point.

[0028] S2: Denoise, outlier removal, and normalization are performed on the collected multi-source operating indicators and environmental perception data to generate a high-quality dataset, while maintaining the unique identifier of the associated component and aging status information for each data item.

[0029] S3: Based on the normalized dataset containing the aging status of components, extract multi-timescale features, including short-cycle dynamic operation features and long-cycle aging features, to form continuous feature vector groups, and identify the device and sampling point location to which each feature vector group belongs.

[0030] S4: Determine whether there are changes in operating conditions in the real-time operating data stream. If the change in dynamic operating characteristics exceeds a preset threshold, then automatically generate corresponding dynamic scene labels based on the environmental data and feature vectors of the sampling points. The label nodes include information on load fluctuations, extreme weather, or abnormal power grid events.

[0031] S5: Based on the generated dynamic scene labels, the parameters and feature importance weights of the fault prediction model under similar historical working conditions are transferred to the current new scene to initialize the fault prediction model weight vector under the current scene.

[0032] S6: For the new working conditions corresponding to the current dynamic scene label, with a limited number of new sampled samples and their feature vectors as input, the feature weights of the fault prediction model are incrementally fine-tuned through the few sample meta-learning algorithm to achieve adaptive calibration of the model to the feature distribution of the new working conditions.

[0033] S7: For the adaptively calibrated fault prediction model, the dynamic operation feature weights and component aging feature weights are combined, and the feature contribution of the dynamic scene label is dynamically weighted to form the final multi-time scale fusion prediction factor.

[0034] S8: Using multi-timescale fusion prediction factors as input, the model outputs the corresponding scenario's power distribution line fault prediction results, and generates feature importance explanation information for each target sampling point.

[0035] S9: Monitor the prediction accuracy and stability of the model after adaptive calibration in real time, perform A / B comparison analysis between the prediction results and the original model. If the new model has better short-term accuracy and stability than the original model, automatically switch to the new model; otherwise, roll back to the original model parameter state and record the tuning process log for subsequent performance optimization.

[0036] Step S1: Collect operational indicators and environmental sensing data from multiple sampling points in the power distribution line. The operational indicators include real-time load, voltage, current, harmonics, ambient temperature and humidity, and climate variables. Record the unique identifier of each component and its corresponding aging status information at each sampling point. Specifically, this includes:

[0037] S1.1: Distributed positioning of sampling points on the power distribution line is performed. The spatial coordinate information of each physical sampling location is identified by the intelligent sampling unit to generate spatial positioning data of the sampling points, laying the spatial index foundation for subsequent data association and multi-point feature fusion.

[0038] The data input for distributed sampling point location includes the pre-set physical topology of the power distribution line, the initial values ​​of the installation coordinates of each sampling point, and the initialization parameters of the on-site intelligent sampling unit.

[0039] Distributed positioning algorithms (such as RTK high-precision positioning or UWB real-time positioning, with parameters including base station coordinates, reference beacon frequency, and unique sampling point number) are used to accurately determine the physical spatial coordinates of the sampling points.

[0040] Furthermore, through the sampling point location data acquisition interface (parameters: sampling unit ID, coordinate frequency, time synchronization code), the intelligent sampling unit automatically identifies and uploads the spatial coordinate information of each physical sampling point in real time, thereby obtaining a high-resolution spatial positioning data stream.

[0041] A spatial coordinate calibration method (parameters: GIS map reference, terrain compensation model) is adopted to realize the automatic mapping between sampling point positioning data and power grid geographic reference system, and generate a distributed spatial positioning calibration matrix.

[0042] Furthermore, by using a location encoding conversion algorithm (parameters: spatial index code, device unique identifier), the spatial coordinates of the sampling points are converted into spatial positioning data objects in a unified format, generating a searchable spatial index for subsequent multi-source data association operations.

[0043] A hierarchical storage and spatial data caching mechanism is adopted to partition and sequentially index the real-time generated sampling point spatial positioning data, ensuring the real-time performance and integrity of spatial positioning data in high-frequency sampling scenarios.

[0044] Through the chain-like processing described above, the data output by the distributed positioning algorithm is transformed into structured spatial positioning data of sampling points through multi-step spatial encoding, calibration and indexing, thus laying the spatial indexing foundation for subsequent multi-dimensional data association of sampling points and multi-point feature fusion.

[0045] For example, in a smart grid power distribution line scenario, the power distribution line is 15 kilometers long, with 128 smart sampling units deployed along the line. A UWB real-time positioning system is used, with a base station spacing of 500 meters and a spatial resolution of 0.1 meters. Each sampling point corresponds to a unique ID. The sampling points use the UWB positioning module to measure their spatial coordinates (x, y, z) in real time based on base station signals. For example, the spatial location of sampling point number 12 is (120.35, 15.60, 3.2) m, and a spatial index code "P012-12035015600320" is generated through a location encoding conversion algorithm. The spatial coordinates of all sampling units are uploaded in batches to a GIS server for geographic mapping, with a calibration error of less than 0.12 meters. Finally, the system generates 128 spatial positioning data objects for sampling points, providing a large-capacity spatial index table for subsequent binding of electrical operating parameters, environmental perception, and unique equipment identification information. In complex working conditions, the auxiliary algorithm refreshes the spatial coordinate data in real time every 10 seconds to ensure the accuracy of the spatial data of the sampling points and the ability to associate multi-source data in dynamic environments. It has been verified that the spatial positioning data missing rate is less than 0.1% and the spatial association accuracy is higher than 99.99%, which lays a high-quality spatial data foundation for subsequent feature fusion and prediction model training.

[0046] S1.2: Based on the spatial positioning data of sampling points, the current real-time load, voltage, current, harmonics and other operating indicators are collected by the multi-channel operating condition sensing interface to obtain a multi-source electrical operating parameter dataset of the power distribution line, providing time series raw data for dynamic state feature extraction.

[0047] Based on the spatial positioning data of the sampling points, the spatial coordinate index of each sampling point is selected as the acquisition object and input to the multi-channel working condition sensing interface.

[0048] A multi-channel real-time operating condition acquisition method (parameter settings include: number of acquisition channels, signal type, sampling frequency, spatial index code, and data buffer queue length) is adopted to realize the synchronous acquisition of electrical operating parameters of each spatial positioning sampling point.

[0049] Furthermore, through signal separation and channel classification algorithms (parameters: channel identifier ID, load sampling channel, voltage sampling channel, current sampling channel, harmonic sampling channel), the real-time load, voltage, current, harmonic and other raw signal data from each channel are classified and integrated according to spatial coordinates and sampling timestamps to form a multi-source synchronous electrical parameter raw data block.

[0050] Furthermore, a high-frequency data buffering and sequence sorting algorithm (parameters: unique ID of sampling point, sampling period T, sliding window length L) is adopted to perform time-series segmentation, sequence encoding and real-time labeling on the acquired multi-channel real-time sampled raw signals, ensuring the integrity and continuity of the original data time sequence, and effectively supporting the subsequent dynamic state feature extraction.

[0051] Furthermore, through the initial screening of abnormal data and the signal integrity detection algorithm (parameters: signal amplitude threshold A, sampling packet loss rate λ, time synchronization threshold τ), the signal integrity, loss status and abrupt abnormal samples are identified in real time for the electrical operating parameter acquisition sequence, physical link noise or signal jumps are eliminated, and a high-confidence original electrical operating parameter sequence is output.

[0052] A unified data structure encapsulation algorithm (parameters: spatial index code, unique identifier of sampling point, signal type, acquisition timestamp, data value) is used to standardize the field definitions and format the real-time load, voltage, current, and harmonic data acquired from the above multi-channel acquisition, generating a structured multi-source electrical operation parameter dataset.

[0053] Through a four-level chain processing of high-channel real-time acquisition, multi-channel signal separation, data buffering and processing, and anomaly screening, the spatial positioning data of sampling points and the original data of electrical parameters from multiple fields are efficiently integrated into a dynamic state time series input, providing high-quality and time-complete basic data for subsequent dynamic operating condition feature extraction steps.

[0054] For example, in a real-world power distribution line scenario, assuming a single line is 15 kilometers long, 128 sampling points are deployed in space, each sampling point is equipped with a 4-channel sensing interface (corresponding to load, voltage, current, and harmonics respectively), the acquisition cycle is set to 1 second / time, and the signal storage length is 20 seconds.

[0055] The system inputs spatial positioning index codes. For example, if the sampling point “P012-12035015600320” is within the acquisition window, the load channel acquires real-time load of 890A, voltage of 230V, current of 875A, and harmonic distortion rate of 2.8%.

[0056] The data buffer stores data in a 128-sample-point × 4-channel × 20-frame format, achieving zero packet loss. An outlier screening algorithm is applied to automatically remove outliers with amplitude fluctuations exceeding 4 times the standard deviation, ultimately outputting a high-confidence sequence of raw electrical parameters, formatted as 128 × 4 × 20 data blocks to meet subsequent feature extraction requirements.

[0057] In this embodiment, through high-channel parallel acquisition and intelligent anomaly screening, the integrity of the acquired operating parameters reaches 99.98%, the abnormal data rejection rate is 0.02%, the acquisition latency is less than 80ms, and each data item contains spatial positioning, signal type, unique ID and timestamp, so as to provide multi-source time series raw data for subsequent steps, supporting high-resolution modeling and intelligent fault prediction under dynamic operating conditions.

[0058] S1.3: Utilize environmental sensor nodes to synchronously acquire environmental perception data such as temperature, humidity, and climate around each sampling point, and form an environmental perception data stream associated with electrical operating parameters through the environmental information acquisition interface, serving as an environmental impact factor in dynamic operating condition changes.

[0059] S1.4: Automatically read the unique identifier of each component at each sampling point, and use an embedded recognition algorithm to index and bind the recognition result with the spatial positioning information, operating indicators and environmental perception data of the sampling point to realize the primary key association of multidimensional data.

[0060] S1.5: For sampling points with unique identifiers of bound components, call the device health monitoring interface to obtain the aging status information of the components in real time, including aging level, historical degradation parameters and expected life indicators, generate device aging status data and form a structured multi-source acquisition dataset with the above spatial, electrical and environmental data.

[0061] S1.6: Perform timestamp uniform calibration and group coding on the aforementioned structured multi-source acquisition dataset to generate a highly consistent time series acquisition matrix across sampling points, so that subsequent steps can efficiently segment, filter, and perform dynamic and aging feature extraction.

[0062] Step S2: The collected multi-source operational indicators and environmental sensing data are denoised, outlier removed, and normalized to generate a high-quality dataset, while maintaining the unique identifier and aging status information of each associated component. Specifically, this includes:

[0063] S2.1: Perform noise suppression processing based on the Kalman filter algorithm on the collected power distribution line operation indicators and environmental perception data to remove random measurement noise in the original data stream, obtain multi-source sampling data after preliminary noise suppression, retain the unique identifier and aging status information of components, and provide more stable data input conditions for the subsequent outlier discrimination process.

[0064] For the power distribution line operation indicators and environmental perception data in the aforementioned structured multi-source acquisition dataset, the input conditions include calibrated time series raw data, spatial location information of sampling points, unique identifiers of components, and aging status characteristic fields.

[0065] The Kalman filter algorithm (parameter configuration: initial value of state-space model, measurement noise covariance R, process noise covariance Q, observation matrix H and initial estimation error) is used to perform dynamic noise suppression processing on the operational index data (such as real-time load, voltage, current, harmonics) and environmental perception data (including temperature, humidity and climate variables) of each sampling point, so as to achieve optimal linear estimation and filtering of random measurement noise in the original data.

[0066] Furthermore, the Kalman filter is solved for multidimensional parameters using a recursive state update method, and noise suppression is achieved using the following state estimation and observation update formulas:

[0067] , , ,in, This is the estimated state value after the Kth step update. These are actual measured values. For Kalman gain, Let I be the covariance at step K, I be the identity matrix, R be the measurement noise covariance, and H be the observation matrix.

[0068] An independent channel Kalman filter is used for the environmental sensing signal, and the parameter settings are optimized according to the noise standard deviation of each channel to ensure effective dynamic suppression of non-stationary noise in temperature, humidity and climate data.

[0069] Furthermore, by comparing the residuals of the noise suppression results with the original data, the noise covariance parameter of the Kalman filter is dynamically adjusted to achieve adaptive optimization of local noise for different sampling points and different component environments.

[0070] Furthermore, the noise-suppressed multi-source sampling data, with the spatial positioning information of the sampling points and the unique identifier of the components as the primary keys, performs a data field reorganization operation, retains the unique identifier of the equipment and aging status characteristics, and realizes the synchronous association between the noise-suppressed data and the health status of the equipment.

[0071] Furthermore, for the dataset output by the Kalman filter, a noise suppression contribution evaluation algorithm (parameters: original signal variance, noise residual mean, signal recovery rate) is used to generate noise suppression performance indicators for each sampling point and each component, and data sampling windows with insufficient suppression effect are screened for secondary filtering or data backfilling.

[0072] The Kalman filter algorithm is used to perform dynamic noise suppression processing on multi-source sampling data of power distribution lines, while strictly preserving the unique identifiers and aging characteristics of components. This achieves noise filtering of the original data stream and association with the primary key field, providing a stable and reliable data input foundation for subsequent outlier detection and high-confidence data generation.

[0073] For example, in a scenario of noise suppression for distribution lines in a smart grid, the input data consists of electrical operation index data from 128 sampling points × 4 channels × 20 frames, and environmental temperature, humidity, and climate sensing data from 128 points. The Kalman filter parameters for each channel are set to: measurement noise covariance R = 0.03, process noise covariance Q = 0.008, and the state estimate is initialized to the first frame after power-on for each sampling point. The standard deviation of the Kalman filter noise for environmental data is set to 0.02. The Kalman filter performs noise suppression on the current measurements from the 128 sampling points. The original mean variance is 0.12; after filtering, the mean signal noise residual is reduced to 0.016, and the signal recovery rate is increased to 99.8%. Each data point retains associated fields, such as the sampling point number "P012-12035015600320", the component unique identifier "E012XGH-202405", the aging status "L3", and the timestamp "202406151210". The noise suppression performance evaluation indicators show a signal recovery rate greater than 98.5%, a noise reduction ratio of 8.5, and a data stream latency of less than 120ms after filtering. For samples with poor filtering performance, the system automatically optimizes the R and Q parameters a second time, further reducing the mean residual to 0.008, achieving high-confidence noise suppression data and supporting the data integrity and accuracy of subsequent anomaly detection and normalization processing.

[0074] S2.2: Apply the 3σ rule and the isolated forest anomaly detection algorithm to the multi-source sampling data after Kalman filtering to identify and remove waveform mutation points, data with physical and logical inconsistencies, and abnormal environmental variable samples, so as to obtain high-confidence multi-source sampling data with outliers removed, and ensure that each data item is consistent with the unique identifier of the component and aging status information of the sampling point.

[0075] S2.3: Based on the high-confidence multi-source sampling data after outlier removal, the max-min standardization and Z-score normalization algorithms are used to perform a unified scale transformation on all variables, including real-time load, voltage, current, harmonics, ambient temperature and humidity, and climate characteristics, to obtain normalized multi-source operating indicators and environmental data, and simultaneously associate them with the corresponding component identifiers and aging status to form a standardized structured dataset.

[0076] S2.4: Further perform field integrity verification and correlation mapping on the normalized structured dataset to ensure that each data entry contains the unique identifier of the component, aging status characteristics, and spatial location of the sampling point. The component health status metadata is fused with the spatiotemporal sampling data to form a data object that can be used for subsequent multi-timescale feature extraction.

[0077] S2.5: Input the spatially correlated standardized multi-source dataset into the data version control system, use time labels and sampling sequence segmentation to generate high-quality segmented data blocks that can be used for historical partitions and current real-time data streams, and provide a traceable data foundation for subsequent dynamic scene label generation and long-cycle aging feature mining.

[0078] Step S3: Based on the normalized dataset containing the aging status of components, extract multi-timescale features, including short-cycle dynamic operation features and long-cycle aging features, to form continuous feature vector groups, and identify the device and sampling point location to which each feature vector group belongs. Specifically, this includes:

[0079] S3.1: For the normalized dataset containing unique component identifiers and aging status information, a time-scale data scheduling strategy is adopted to group short-cycle (e.g., minute-level, hour-level) dynamic operating indicators and long-cycle (e.g., week-level, month-level) degradation status parameters into blocks to construct a multi-time-scale data mapping relationship, providing a hierarchical data source for the subsequent feature extraction process.

[0080] S3.2: For the short-cycle dynamic operation data after each grouping, use time-series feature extraction algorithms (such as time-series statistical feature analysis and Fourier transform) to calculate key operation feature parameters, including time-series statistical features such as load fluctuation amplitude, current peak value, and harmonic mean value, and generate a short-cycle dynamic operation feature matrix, which serves as the time window input for subsequent feature vector groups.

[0081] For the grouped short-cycle dynamic operation data, the input conditions include the time series of electrical operation indicators that have been normalized and outliers have been removed, involving variables such as real-time load, voltage, current, and harmonics, and synchronously associated with the spatial location of the corresponding sampling points and the unique identifier of the components.

[0082] The time series statistical feature analysis method (parameters: sampling window length T, data sampling frequency F, statistical type including mean, variance, maximum value, minimum value, peak value, amplitude, rate of change) is used to calculate the feature parameters of short-cycle time series data for each sampling point, so as to realize the quantitative characterization of periodic fluctuations and abnormal changes in the operating status.

[0083] Furthermore, key operational characteristic parameters are calculated using the following time-series statistical formulas:

[0084] ; ; ; ; ;in, For the first Electrical performance data at each point in time. The number of samples within the window. The mean, For variance, For fluctuation amplitude, Peak value, Let T be the rate of change per unit time, and T be the length of the statistical window.

[0085] Furthermore, a Fourier transform algorithm is employed (parameter: Fast Fourier Transform window length). Sampling interval This paper analyzes the frequency domain of short-cycle electrical operation index sequences, decomposes the spectral distribution characteristics of each signal, and extracts frequency domain features such as harmonic mean and spectral energy ratio. The Fourier transform formula is:

[0086] ,in, For the frequency domain Term coefficient, is the original time-domain sampled value, and N is the total number of sampled points.

[0087] Furthermore, based on Fourier spectrum analysis, the harmonic mean index is calculated: ,in, M represents the harmonic mean, and M is the selected harmonic order range (e.g., 2nd to 5th order). This represents the amplitude of the corresponding FFT spectrum.

[0088] By combining time-series statistical feature analysis with Fourier transform, the aforementioned short-cycle dynamic operating data is transformed into a feature matrix. The matrix dimensions include multi-dimensional features such as load fluctuation amplitude, peak current, voltage change rate, and harmonic mean, which serve as the short-cycle time-series input for the feature vector group.

[0089] Using the aforementioned feature extraction algorithm and statistical formula, the grouped short-cycle dynamic operation data is processed to produce a short-cycle dynamic operation feature matrix, which enables quantitative characterization of the operation characteristics of each sampling point and each component in the minute to hour-level operation state, providing a standardized data foundation for the subsequent synthesis of feature vector groups and model input.

[0090] For example, in the scenario of monitoring distribution lines in an urban smart grid, time-series statistical feature analysis is performed on the normalized current time-series data of 128 sampling points (sampling frequency F=10Hz, statistical window length T=60 minutes, sample size N=36000) to calculate the mean within the window. ,variance Fluctuation amplitude peak rate of change Applying Fast Fourier Transform (FFT) to the same time-series data (FFT window length) Extract the amplitude values ​​of the 2nd to 5th harmonics to obtain the harmonic mean. These features are combined to form a 128×8 dimensional short-cycle dynamic operation feature matrix, with each row corresponding to a set of sampling points and unique component identifiers. For different types of sampling points (such as main lines and branch lines), the statistical window and FFT parameters are dynamically adjusted to adapt the operation features of each point. Verification shows that this step effectively extracts various short-cycle key features during power operation and provides highly consistent standard feature data input for subsequent aging feature association and model input. The output feature matrix is ​​directly used in subsequent feature splicing and model training, achieving high-precision and robust operation feature representation.

[0091] S3.3: For the aging status parameters of long-cycle components after each grouping, based on aging trend modeling algorithms (such as exponential smoothing and piecewise linear regression), calculate aging status characteristic values ​​such as insulator failure probability, connector impedance increase rate, and equipment service life to form a long-cycle aging feature matrix. Gradually correlate it with the short-cycle dynamic operation characteristics of the same sampling point to improve feature coordination capability.

[0092] For the normalized dataset containing unique component identifiers and aging status information, a long-period data grouping strategy is adopted to determine the degradation status parameter block of each sampling point and the corresponding component within a long-term window such as a week or month.

[0093] An aging trend modeling algorithm is used to perform long-cycle state time series modeling for major components such as insulators and connectors. The input parameters include the aging level of the components in each cycle (such as L0-L4), historical degradation cumulative parameters (such as degradation coefficient, failure history frequency), expected life index and actual service life.

[0094] Furthermore, the aging parameter sequence of the components is modeled using the exponential smoothing method, and the aging index smoothing calculation is performed using the following formula:

[0095] ,in, This is the aging index after smoothing the current cycle. This is the original aging data. α is the smoothing coefficient (0 < α < 1). This is the smoothed value from the previous period.

[0096] Furthermore, to address the accelerated degradation and staged abrupt changes in aging characteristics, a piecewise linear regression algorithm is used to partition and model the aging time series with periodic points as breakpoints. Parameters include the length of each periodic segment. Segmentation start point Piecewise regression weight matrix The piecewise degradation rate was calculated using least squares fitting.

[0097] Using a failure probability calculation method, the stage failure probability of key components such as insulators and connectors is inferred based on the modeled degradation trend and historical failure data. The formula is as follows: ,

[0098] in, This is the degradation rate coefficient. This refers to the actual service life.

[0099] For the connector impedance increase rate, based on the historical impedance measurement data series, the annual impedance growth rate is extracted using a slope calculation formula: ,in, The rate of increase of impedance, This is the current periodic impedance value. This is the initial impedance value. This refers to the service life.

[0100] Furthermore, the aging state characteristic values ​​of the components obtained above (such as the insulator failure probability) are... Connector impedance increase rate Equipment service life (etc.) The structured matrix is ​​assembled according to the sampling points and the unique identifiers of the components to generate a long-cycle aging feature matrix.

[0101] By using an association mechanism, the long-cycle aging feature matrix and the short-cycle dynamic running feature matrix under the same sampling point are bound by primary key index to form a feature collaborative mapping relationship, which enhances the collaborative ability between aging features and running features and creates conditions for subsequent feature splicing and high-dimensional input of the model.

[0102] By modeling aging trends, calculating probabilities, and using piecewise regression algorithms, the standardized data processed in the previous step is effectively transformed into a long-cycle component aging feature matrix, thereby achieving quantitative expression and synergistic fusion of aging state characteristics.

[0103] For example, in the long-term operation and maintenance scenario of urban smart grid distribution lines, historical degradation data of insulators is collected monthly, and the aging level for each cycle is input. The smoothing coefficients are [1.2, 1.36, 1.54, 1.77, 2.05]. Calculate the monthly smoothed aging index sequentially. The values ​​are 1.2, 1.296, 1.396, 1.513, and 1.633, respectively. These are the initial impedance values ​​for connector impedance measurements. Current 5th year impedance Then the rate of increase in impedance for / year. A degradation rate coefficient is set for the insulator failure probability. / year, service life The failure probability is [number] years. These feature values ​​are assembled into a long-cycle aging feature matrix. Each sampling point is associated with a unique component identifier and a set of [monthly aging smoothing value, impedance increase rate, service life, failure probability], forming a 128×4-dimensional long-cycle aging feature matrix. Verification shows that the associated feature matrix can be seamlessly concatenated with the short-cycle operation matrix using the primary key, significantly improving the joint modeling capability of operation-aging features and providing high-dimensional input at multiple time scales for model inference and dynamic operating condition adaptability analysis.

[0104] S3.4: For each sampling point and its corresponding unique component identifier, perform a feature mapping assembly operation to combine the short-cycle dynamic operation feature matrix with the long-cycle aging feature matrix to generate a continuous feature vector group containing the unique device identifier, sampling point location information, dynamic operation features and aging features, providing a full variable carrier for the adaptive model input.

[0105] The input conditions include a normalized structured multi-source dataset, which contains spatial location information of sampling points, unique device identifiers, short-cycle dynamic operation feature matrices, and long-cycle component aging feature matrices.

[0106] The feature mapping assembly algorithm is used (parameter: short-period feature matrix). Long-period characteristic matrix Sampling point location field Unique Identifier for Components This enables synchronization of primary key indexes with short-cycle dynamic features and long-cycle aging features.

[0107] Furthermore, through a feature concatenation method (parameters: concatenation mode "horizontal expansion", feature field type encoding), the features at each sampling point are concatenated. and According to the device's unique identifier and sampling point positioning Perform field-level concatenation to obtain continuous feature vectors. :

[0108] ,in, For the first Short-cycle dynamic characteristics, For the first Both are long-term aging characteristics, encoded separately by time window and periodic block, and both are associated with a specified primary key.

[0109] Furthermore, through the feature vector integrity verification method (parameter: total number of fields) Primary key integrity threshold After verification and splicing Each entry must contain all necessary feature types and primary key fields. Missing fields will be automatically filled with NULL values ​​or specified default values ​​to ensure the consistency of the vector group structure.

[0110] Furthermore, using the vector batch processing interface, continuous feature vectors under all sampling points and device unique identifiers are processed. The adaptive model preprocessing module is imported in batches and partitioned by timestamp and spatial coordinates to achieve batch loading of features.

[0111] Through the aforementioned feature mapping assembly and splicing algorithm, the short-cycle dynamic operation feature matrix and the long-cycle component aging feature matrix are efficiently fused under the primary key to generate a structured continuous feature vector group. This provides a high-dimensional carrier of all variables for the subsequent adaptive fault prediction model input, realizing the collaborative expression of operation features and aging features.

[0112] S3.5: Based on feature vector groups, use feature effectiveness evaluation methods (such as principal component analysis and feature importance score calculation) to analyze the contribution of each feature under historical fault data. Group and label continuous feature vector groups according to the unique identifier of the equipment and the location of the sampling point to form a labeled feature library with traceability and data integrity, so as to support the subsequent dynamic scene label generation and feature weight interactive adjustment.

[0113] Step S4: Determine whether there are changes in operating conditions in the real-time operating data stream. If the change in dynamic operating characteristics exceeds a preset threshold, automatically generate corresponding dynamic scene labels based on the environmental data and feature vectors of the sampling points. The label nodes include information on load fluctuations, extreme weather, or abnormal power grid events. Specifically, this includes:

[0114] S4.1: For the normalized feature vector group containing the aging status of components, perform dynamic operation feature change monitoring based on the sliding window algorithm, and extract professional parameters such as real-time load change rate, voltage fluctuation standard deviation, and instantaneous current deviation in each time window as dynamic operation feature measurement to quantitatively characterize the dynamic state of the current working condition.

[0115] Dynamic operation feature changes are monitored on the normalized feature vector group containing the aging status of components. The input conditions include the structured high-dimensional feature vector group obtained by the previous steps. The feature fields include windowed operation indicators such as real-time load, voltage, current, and harmonics, as well as spatial positioning, unique equipment identifier, and aging status parameters.

[0116] Using the sliding window algorithm (parameter: window length) Window step size Feature sampling frequency In the time series data stream, feature data is segmented in an equally spaced step manner to ensure high spatiotemporal resolution monitoring of the dynamic running data of each sampling point.

[0117] Furthermore, through statistical calculations of feature data within the window, core dynamic operating characteristic metrics such as real-time load change rate, voltage fluctuation standard deviation, and instantaneous current deviation are calculated. For the real-time load change rate, the following formula is used for quantitative calculation: ,in, For load data at the end of the window, For the initial load data of the window, This is the window length.

[0118] For the standard deviation of voltage fluctuations, the windowed variance formula is used: ,in, For the first Voltage values ​​at each sampling point The average voltage within the window. This represents the number of samples within the window.

[0119] For instantaneous current deviation, it is calculated as the difference between the latest sample within the window and the mean value at the previous moment: ,in, For the current sample at the end of the window, This is the average current value of the previous window.

[0120] Furthermore, by automatically performing structured encoding on the time-series statistical results of each window, a dynamic operational feature measurement set including the aforementioned indicators is generated, thereby achieving accurate quantification of the dynamic state of each sampling point.

[0121] Furthermore, the dynamic operational characteristics monitored within each time window are synchronously bound to the spatial location of the sampling point and the unique identification field of the equipment to form a dynamic characteristic monitoring data block with primary key traceability.

[0122] By using a sliding window dynamic change monitoring algorithm, the normalized feature vector set generated in the previous steps is transformed into a time-series dynamic operating feature metric set, which enables a quantitative characterization of the current operating conditions of each sampling point of the power distribution line, providing a high-confidence input for subsequent environmental factor fusion and operating condition change determination.

[0123] For example, the sliding window length is set for the current and load time-series data of 128 sampling points in the urban central area power distribution network. Minutes, window step Minutes, sampling frequency For sampling point numbered "P010-12035815600339", the first and last loads in the first window were respectively and Based on this, the load change rate is calculated. The average voltage sample within this window Sample standard deviation Current at the end of the window , compared with the mean of the previous window Therefore, instantaneous deviation The aforementioned window statistical indicators are integrated to generate a structured dynamic operational feature measurement data block, which is then bound to the primary key of the sampling point and the unique identifier of the component. After batch statistics, a 128×N-dimensional dynamic operational feature monitoring matrix is ​​formed for all sampling points across all time windows. System verification results show that this processing chain can complete the monitoring of dynamic operational feature changes for all samples within 5 minutes. Window parameters and run configurations can be flexibly adjusted according to actual business scenarios, and the output conversion accuracy is better than 98%, with no missing or misjudged segmentation boundaries. The above process effectively improves the abnormal operating condition capture rate and dynamic state recognition sensitivity, laying a technical foundation for subsequent environmental factor fusion, operating condition change judgment, and dynamic scene label generation.

[0124] S4.2: Obtain environmental perception data of sampling points corresponding to the above dynamic operation feature vector group, and use the environmental state fusion algorithm to calculate the correlation between temperature and humidity values, climate type, sudden meteorological events and the geographical location of sampling points, and generate environmental factor feature groups for the same time window.

[0125] The sampling point environmental perception data corresponding to the dynamic running feature vector group is obtained. The input conditions include normalized temperature and humidity sensor data, climate type classification results, sudden meteorological event signals and sampling point geographical coordinates.

[0126] An environmental perception data synchronization retrieval method (parameters: unique identifier of sampling point E_id, time window T_win, spatial location P_loc) is adopted to synchronize the temperature and humidity, climate type, sudden meteorological event identifier and geographical location of each sampling point within the current time window, forming the original environmental perception dataset.

[0127] Furthermore, through the environmental feature standardization algorithm (parameter: maximum and minimum normalization interval [a]), min ,a max Climate type code L weather Meteorological event intensity level S event This achieves a unified scale conversion for different types of environmental data. For temperature and humidity normalization, the following formula is used:

[0128] ,in, To standardize environmental elements, These are the original environmental values. These represent the minimum and maximum values ​​of the element within the historical interval.

[0129] Furthermore, through an environmental state fusion algorithm (parameters: normalized temperature and humidity value E), norm Climate type code L weather Meteorological event intensity S event Sampling geographic coordinates P loc The environmental elements are subjected to spatiotemporal stratification and weighting. Based on historical meteorological distributions in the comprehensive database, a weighted fusion formula is used:

[0130] ,in, For the first Environmental factor characteristic values ​​at sampling points The weighting coefficients for the multi-source environment are determined by fitting historical data.

[0131] Furthermore, for sudden meteorological events, an event space influence radius algorithm is adopted (parameter: event occurrence coordinates P). event Radius of influence R event Sampling point coordinates P {loc,j} To determine whether a sampling point is located within the area affected by a meteorological event, the following formula is used:

[0132] .

[0133] Furthermore, the above-mentioned environmental features are structured and summarized according to sampling points and time windows to generate environmental factor feature groups for each sampling point within the same time window, providing highly consistent environmental input for subsequent joint determination with dynamic running feature measurement and scene label generation.

[0134] By using an environmental state fusion algorithm, the synchronously retrieved temperature and humidity, climate type, meteorological emergencies and sampling point geographical locations are fused together and transformed into environmental factor feature groups with spatiotemporal hierarchy and primary key index, so as to realize the comprehensive quantification and structured expression of the environmental impact of power distribution lines under dynamic operating conditions.

[0135] For example, in the rainy season operation and maintenance scenario of an urban smart distribution network, for sampling point number "P010-12035815600339", the time window is from 13:00 to 13:15 on June 10, 2023, the average temperature is 25.8°C, the average humidity is 91%, and the historical temperature range is [a min ,a max The temperature range is [12.5, 38.7], and the humidity range is [27, 97]. The normalized formula is used to calculate the temperature. ,

[0136] humidity .

[0137] Current climate type code L weather =3 (moderate rainfall), meteorological event intensity S event =2 (Thunderstorm Warning), the event occurred 2km away from this sampling point, the event influence radius was set to 3km, the sampling point was determined to be in the meteorological event influence area, and the event intensity was valid.

[0138] Using weighting coefficients w temp =0.25, w hum =0.35, w clim =0.20, w event =0.20, calculate the characteristic value of environmental factors according to the fusion formula:

[0139] .

[0140] Finally, the structured output environmental factor feature group includes the sampling point primary key "P010-12035815600339", the time window label "2023061013:00-13:15", the environmental factor feature value "1.443", the normalized temperature and humidity value "0.491 / 0.914", the climate type "3", the event intensity "2", and the geospatial determination result "valid".

[0141] Batch processing of 128 sampling points takes less than 12 seconds, and the formation rate of environmental factor feature groups reaches 100%. The above example output effectively improves the spatial coupling capability between dynamic operating conditions and environmental data, realizes the environmental perception accuracy and adaptive capability of the power grid fault prediction model in highly dynamic and complex scenarios, and supports the technical foundation for abnormal operating condition judgment and scene label generation.

[0142] S4.3: Based on the characteristic change amplitude threshold judgment mechanism, the change amplitude index is calculated one by one for the acquired dynamic operation characteristic measurement and environmental factor characteristic group. When any index exceeds the preset abnormal operating condition threshold, the operating condition change judgment is triggered and the triggering factor (such as severe load fluctuation, extreme weather) is marked.

[0143] S4.4: For data windows identified as having changed operating conditions, the environmental factor group of sampling points and the dynamic operating characteristic change factor are integrated, and a scene label generation algorithm is used to automatically generate dynamic scene labels. The label nodes include, but are not limited to, load fluctuation scene labels, extreme weather scene labels, and abnormal power grid event labels. Each label is associated with a corresponding characteristic change factor to ensure that the label output is interpretable.

[0144] S4.5: The dynamic scene labels and their corresponding triggering factors are used as output results. The feature factors, judgment thresholds and triggering conditions of the scene label generation process are recorded to provide scene-specific input basis for the parameter transfer of the subsequent fault prediction model and the fusion of feature weights at multiple time scales, forming a hierarchical scene label data stream.

[0145] Step S5: Based on the generated dynamic scene labels, the parameters and feature importance weights of the fault prediction model under similar historical working conditions are transferred to the current new scene to initialize the fault prediction model weight vector in the current scene. Specifically, this includes:

[0146] S5.1: Classify and identify dynamic scene tags. Based on the tag content such as load fluctuation, extreme weather or abnormal power grid events contained in the tag nodes, use the scene tag clustering algorithm to retrieve the set of tags with the highest similarity to the current dynamic scene tags in the historical operating condition database to obtain the candidate operating condition tag sequence.

[0147] S5.2: For candidate working condition label sequences, model parameters are mapped. The label weight similarity calculation method is used to perform similarity weighted filtering on the set of fault prediction model parameters and the set of feature importance weight vectors saved under historical working conditions, so as to obtain model parameters and feature weight recombination under similar scenarios.

[0148] The collected candidate operating condition label sequence is used as input data, which includes dynamic scene label nodes and their corresponding characteristic attributes such as load fluctuations, extreme weather, and abnormal power grid events.

[0149] The label weight similarity calculation method is adopted. Based on the fault prediction model parameter set and feature importance weight vector set saved in the historical working condition database, the label content, label components and feature weights are defined as the basis for similarity measurement.

[0150] By vectorizing the label content, historical working condition labels and current candidate working condition labels are encoded into high-dimensional label feature vectors, and the similarity between label vectors is calculated using the cosine similarity formula:

[0151] Where A is the current dynamic scene label feature vector, and B is the historical working condition label feature vector. Let A and B be the vector dot product, where A and B are the vector magnitudes.

[0152] Furthermore, a similarity score is assigned to each group of labels, combined with the parameter vectors already saved by the fault prediction model under historical operating conditions. With feature importance weight vector The parameter set is filtered based on a weighted screening formula:

[0153] ,in, Assign a weighted similarity score between the current scene feature components and the historical model feature vectors. The typical configuration range is for the weighting coefficients of label and feature similarity. .

[0154] Furthermore, by sorting and filtering by score, the highest-scoring sets of historical scene model parameters are selected. Reorganization of feature rights This forms a candidate parameter mapping set.

[0155] A similarity threshold filtering mechanism is adopted, and a preset threshold is set. Only the score is retained. Greater than The parameter weights are reorganized to ensure that the distribution of the transfer model parameters matches the current scenario.

[0156] By calculating label weight similarity and filtering parameter mapping, we obtain historical model parameter sets and feature importance weight reorganizations suitable for the current new scenario, thus preparing weight mapping inputs for subsequent scenario label migration.

[0157] For example, in the scenario of "extreme thunderstorm weather" in the western urban area of ​​the distribution network system, the dynamic scenario labels that occurred at 13:00 on June 10, 2023 are input: severe load fluctuation (code 1), weather thunderstorm (code 4), and abnormal temperature and humidity (code 7). The historical operating condition database stores 256 similar scenario labels and fault model parameter groups from the past three years.

[0158] Label vectorization is used to generate the current scene label feature vector. , with historical scene tags Calculate cosine similarity .

[0159] Dynamic operation characteristics of the current scene Features of historical models Calculate feature similarity Weighting balance coefficient Score based on this:

[0160] Set a similarity threshold Two historical scenario parameter groups with scores greater than 0.95 were selected and output, representing the model parameters for the "historical thunderstorm high load" condition. and feature rights reorganization Model parameters for "load under historical thunderstorms" and feature rights reorganization Upon reviewing the migration records, both sets of parameters demonstrated a fault identification accuracy rate exceeding 95% in past thunderstorm scenarios.

[0161] Through the above-mentioned label weight similarity calculation and mapping filtering chain, the historical scene model parameters and feature weights with the highest matching degree are accurately transferred to the current new working condition scene and encoded, and the data blocks that can be used for subsequent parameter transfer and weight reorganization structure transformation are output, ensuring the relevance and reliability of parameter transfer.

[0162] S5.3: The selected historical scene model parameters and feature importance weights are recombined, and a structural mapping transformation is performed based on the label migration algorithm. The dimensions and structure of the parameters and weight groups are adjusted to adapt to the data feature distribution under the current dynamic scene label, so as to realize the initial scene migration of the weight vector.

[0163] For the reorganization of historical scene model parameters and feature importance weights, the input conditions are the set of historical parameters obtained by filtering through S5.2 label weight similarity and the data feature distribution corresponding to the current dynamic scene label.

[0164] The label transfer algorithm is used to perform structural mapping transformation on the historical model parameter set and feature weight recombination. The mapping algorithm parameters include feature field dimension, weight vector length, and target scene distribution matrix.

[0165] Furthermore, through scene label structure parsing, the historical scene label vectors and the current target label are aligned in dimensions and matching rules are set to determine the mapping index relationship and structure transformation matrix.

[0166] The parameter dimension reconfiguration method is used to project the feature weight vector of the historical parameter set onto the feature space of the current new scene, and the following transformation formula is executed:

[0167] ,in, The weight vector after the transfer. For the tag migration structure mapping matrix, This is the historical parameter weight vector.

[0168] Furthermore, for feature importance reweighting, a weight distribution remapping method is used to calculate the feature contribution adjustment coefficient under different scenarios, and the following weighting formula is applied:

[0169] ,in, For the first Weights after class feature transfer For the current scenario, the first Class feature contribution adjustment coefficient Weights for similar features in historical scenarios.

[0170] Through the above formula chain, the dimensional transformation and component weighting of the model parameter weighting are realized under the label migration structure, and the data block that has initially completed the scene transformation is output.

[0171] By adopting a data structured coding method, the final migration results are bound to the primary key according to the data feature distribution of the current scene, so as to achieve structural consistency between the parameter group and the target scene.

[0172] By using a label transfer algorithm and dimensional structure reconfiguration, historical model parameters and feature weights are recombined and transformed into preliminary transfer weight vectors suitable for the current new working conditions, achieving the expected technical effect of adaptive transfer of working condition-level parameters.

[0173] For example, under the scenario of "severe thunderstorm fluctuations" in the distribution network, fault model parameters from historical "high thunderstorm load" scenarios are obtained through S5.2 filtering. and feature weights Current scene label structure mapping matrix .

[0174] The weight vector after the migration is:

[0175] Among them, the weight of the second type of feature is enhanced due to the high matching between the current scene label and the historical scene features; the weight of the third type of feature is further reweighted due to the electrical mutation factors in the thunderstorm scene.

[0176] Using characteristic contribution adjustment coefficient , The weights after migration are calculated as follows: , .

[0177] Ultimately, the model parameter transfer results for the current "severe thunderstorm fluctuation" scenario are generated. With feature weights

[0178] Each parameter is perfectly matched to the feature space structure of the target scene.

[0179] In system performance verification, the model weight vector after the above migration was used in the online fault prediction task. The fault identification sensitivity in the high-load section of thunderstorms was improved to 97%, the average false alarm rate was less than 4%, and the key feature weights were associated with the scene label primary key, which met the high reliability requirements of scene adaptability.

[0180] S5.4: The model parameters obtained from the initial migration are reweighted, and specific environmental variables and component aging status features in the current dynamic scene label are integrated. Through weighted fusion processing, the scene adaptability is further corrected, and the initial weight vector of the fault prediction model suitable for the current sampling point is output.

[0181] S5.5: The weight vector of the fused fault prediction model is used as the startup configuration parameter of the fault prediction model in the current new scenario to initialize the model running state, provide a parameter basis for subsequent few-sample adaptive fine-tuning and online prediction, and form a parameter migration record log for traceability and performance evaluation.

[0182] Step S6: For the new working condition corresponding to the current dynamic scene label, using a limited number of newly sampled samples and their feature vectors as input, the feature weights of the fault prediction model are incrementally fine-tuned through a few-sample meta-learning algorithm to achieve adaptive calibration of the model to the feature distribution of the new working condition. Specifically, this includes:

[0183] S6.1: Collect a limited number of new samples for new operating conditions under the current dynamic scene label, and use the feature vector containing real-time operation features and component aging features as input factors to construct a data subset for developmental training, so as to ensure that the input samples have multi-time scale information coverage.

[0184] S6.2: Based on the feature vector of the newly sampled sample, feature enhancement techniques (such as feature aggregation and multi-scale feature expansion) are used to expand the expression space of the input data, enhance the original features into a high-dimensional feature representation that can be used for few-sample learning, and generate a feature enhancement matrix that supports the input of meta-learning algorithms.

[0185] Feature enhancement techniques are applied to the feature vectors of newly sampled samples. Input conditions include multi-timescale feature vectors acquired in real-time by the acquisition system, covering short-cycle dynamic operation features and long-cycle component aging features. A feature aggregation algorithm (parameters: window length, spatial correlation weight) is used to achieve high-dimensional concatenation of feature groups from multiple sampling points and multiple variables. Furthermore, a multi-scale feature expansion method (parameters: time resolution grouping, indicator diversity filtering) is used to construct feature subsets for different time periods and generate a cross-scale feature expansion matrix, thereby increasing the dimensionality of the original feature vectors.

[0186] A feature transformation algorithm (such as a multilayer autoencoder, parameters: hidden layer dimension, activation function) is used to perform nonlinear mapping on the concatenated high-dimensional feature vector group, thereby improving the feature representation capability and removing redundant correlations. Principal component analysis (PCA) is then performed on the high-dimensional feature space through principal component screening (parameter: contribution rate threshold 0.95) to select and reconstruct the feature dimensions that best represent the working condition distribution, retaining effective features for subsequent meta-learning algorithms.

[0187] Furthermore, normalization techniques (such as Z-score normalization, parameters: mean, standard deviation) are employed to standardize each dimension of the feature enhancement matrix, eliminating data distribution bias and ensuring the equivalence of different features in the few-shot learning algorithm. A feature consistency verification module (parameters: primary key index, spatiotemporal label) is used to verify the primary key integrity of the enhanced high-dimensional feature matrix, ensuring that all sample data are associated with dynamic scene labels, spatial positioning of sampling points, and unique component identifiers.

[0188] Through the continuous processing of feature aggregation, multi-scale expansion, transformation, filtering, normalization and primary key verification, the output is a feature enhancement matrix suitable for the input of the meta-learning algorithm, thereby maximizing the expressive power of the sample feature space under the new working conditions.

[0189] For example, under the new operating condition of "severe fluctuations due to thunderstorms" in the power distribution network, the newly collected samples cover 5 sampling points, each point containing 12 types of features including real-time load, current, harmonics, and insulator aging level. The sampling window length is set to 15 minutes, and the spatial correlation weight is 0.2. The feature aggregation module concatenates the 12-dimensional features of each sampling point into a 60-dimensional initial input.

[0190] Multi-scale feature expansion was employed, merging minute-level features and hourly-level aging parameters to obtain a 120-dimensional feature matrix. A multi-layer autoencoder with a hidden layer dimension of 32 and ReLU activation function was used. After mapping transformation, it outputs a 32-dimensional high-expression feature representation. The principal component selection contribution rate threshold was 0.95, retaining a 24-dimensional subset of principal features. After normalization, all features were standardized to a mean of 0 and a variance of 1.

[0191] Feature consistency verification ensures that all samples are bound to spatial identifiers, device IDs, and scene labels, such as sampling point P1 corresponding to a thunderstorm scene and insulator number IN098. A high-quality feature enhancement matrix (N×24, where N is the number of samples) is ultimately formed. This matrix is ​​then input into subsequent meta-learning algorithms (such as MAML) to achieve effective feature transfer and representation for new operating condition samples. Application results show that, under the new scenario of severe thunderstorm fluctuations, the expanded feature enhancement matrix improves the fault identification rate by 10% after model adaptive calibration, and the feature contribution distribution is more significantly prominent in key dynamic and aging synergistic factors, supporting efficient model transfer and performance improvement.

[0192] S6.3: Input the feature enhancement matrix into a sample meta-learning algorithm (such as Model Independent Meta-Learning, MAML) to transfer and initialize the existing feature weight parameter set of the fault prediction model, so as to realize the rapid adaptive transfer of feature weights under new working conditions.

[0193] S6.4: For the feature weight parameters after migration initialization, perform incremental fine-tuning based on the label information of the newly sampled samples, and use mini-batch optimization algorithms (such as gradient update method) to dynamically update the feature weights of the fault prediction model to ensure that the weight distribution fully reflects the current sample features.

[0194] S6.5: Combining the fine-tuned feature weight parameter group and scene label nodes (such as load fluctuations and extreme weather), the feature contribution dynamic calculation module performs hierarchical adjustment on the weights of various prediction factors to generate a feature weight calibration vector that highly matches the distribution of the new working conditions, which is used to support subsequent prediction and inference.

[0195] Step S7: For the adaptively calibrated fault prediction model, its dynamic operating feature weights and component aging feature weights are combined, and dynamically weighted according to the feature contribution of the dynamic scene label to form the final multi-time-scale fusion prediction factor. Specifically, this includes:

[0196] S7.1: Obtain the dynamic operating feature weight set of the adaptively calibrated fault prediction model as input conditions, take the fault prediction model parameter weight vector as the execution object, and use the feature merging and fusion algorithm to generate dynamic operating feature weight results containing multi-dimensional real-time load, voltage, current, harmonics and other variables, providing basic data for subsequent feature weighting operations.

[0197] S7.2: Obtain the aging feature weight set of the fault prediction model components after adaptive calibration as input conditions, take the aging state parameter weight corresponding to the unique identifier of the component as the execution object, and use the aging feature weight mapping algorithm to generate the aging feature weight data of the corresponding component, providing key health factors for comprehensive feature weighting.

[0198] S7.3: Obtain the dynamic scene label of each sampling point as input condition, execute the feature contribution dynamic weighting algorithm, based on the dynamic running feature weight result obtained in the previous step and the component aging feature weight data, use the label nodes such as load fluctuation, extreme climate, and abnormal power grid events in the scene label as weighting parameters, calculate the weighted contribution of each feature weight in the current scene, and generate the feature contribution weighting matrix.

[0199] Using the adaptively calibrated fault prediction model's dynamic running feature weight set and the component aging feature weight set as input conditions, and combined with the dynamic scene label data generated in real time at each sampling point, a dynamic weighting algorithm for feature contribution is executed to achieve quantitative fusion processing of feature weights driven by operating conditions.

[0200] The scene label parsing module (parameters: node type load fluctuation, extreme weather, abnormal power grid event) is used to extract the dynamic scene label data associated with each sampling point by node to obtain the scene label weight vector. Each component represents the label strength of nodes in different scenarios.

[0201] Furthermore, through the feature contribution calculation algorithm (parameter: dynamically running feature weight set) Component aging feature weight set For each dynamic running feature and aging feature, weighted processing is performed based on the scene label weights. For the first... The formula for calculating the scenario-weighted contribution of each dynamic operating feature is as follows: ,in The correlation coefficient between features and scene label nodes is obtained by pre-setting or online learning based on the statistical correlation between features and labels.

[0202] For the The formula for calculating the scenario-weighted contribution of each aging feature is as follows: ,in This represents the correlation coefficient between aging features and scene label nodes.

[0203] Through the weighted calculations described above, the contribution of each dynamic operating feature and aging feature under the current operating condition is obtained. Furthermore, the weighted contributions of all features are aggregated to generate a feature contribution weighted matrix. This is used for subsequent multi-timescale predictor fusion.

[0204] By using a dynamic weighted algorithm based on feature contribution, the feature weights obtained in the previous step are fused into a feature contribution weighted matrix that reflects the working condition scenario label-driven feature contribution. This enables the quantification of the contribution of various fault prediction features in complex dynamic scenarios and outputs standardized weighted results.

[0205] For example, under the dynamic scene labels of "severe thunderstorm fluctuations" and "aging insulators", the dynamic operating feature weight vector of sampling point P1 is: These correspond to load, current, harmonics, and voltage fluctuations, respectively. The component aging characteristic weight vector is: These correspond to the insulator failure probability and the connector impedance increase rate, respectively. Scene label weight vector ,in Corresponding to the intensity of thunderstorm scenarios, This corresponds to high-aging scene intensity. Assume the dynamic feature scene correlation coefficients obtained through scene association training are as follows: , The contribution of the first dynamic feature is calculated as follows:

[0206] Following a similar approach, the weighted contribution of all features is calculated, and the final output is the weighted matrix of feature contribution of sampling point P1. Under this operating condition, the contribution of aging features is significantly enhanced, providing key indicators for subsequent prediction, inference, and explanation of fault causes. Practical application results show that in scenarios with severe thunderstorm fluctuations and aging equipment, after the model integrates the contribution weighted matrix, the fault prediction accuracy increases to 92%. The key feature weight distribution is highly matched with the scene label nodes, verifying the effectiveness of the dynamic weighting mechanism for feature contributions.

[0207] S7.4: Based on the feature contribution weighting matrix and the existing feature weight set, a multi-timescale predictive factor fusion algorithm is executed. With short-cycle dynamic operation features and long-cycle component aging features as input parameters, a joint weight optimization strategy is used to fuse feature weights of different time scales to generate the final multi-timescale fusion predictive factor, providing highly interpretable input for the model's output fault prediction results.

[0208] S7.5: Output multi-timescale fusion prediction factors as feature inputs for subsequent fault prediction model inference, and record the dynamic scene label and feature weight distribution corresponding to each fusion prediction factor for subsequent feature importance interpretation and model performance tracking, thereby improving the transparency and credibility of fault prediction factors.

[0209] Step S8: Using multi-timescale fused prediction factors as input, the model outputs the corresponding scenario's power distribution line fault prediction results, and generates feature importance explanation information for each target sampling point. Specifically, this includes:

[0210] S8.1: Using multi-timescale fused prediction factors as input conditions, the fault prediction model after adaptive calibration is called, and a high-dimensional feature vector-driven deep inference mechanism is adopted to perform fault probability distribution inference for each target sampling point, thereby obtaining the preliminary power distribution line fault probability output under the corresponding dynamic working condition scenario.

[0211] S8.2: Based on the fault probability output obtained through fault prediction model inference, the feature contribution attribution algorithm is applied to decompose and quantify the weighted contribution of each feature component in the multi-timescale fusion prediction factor, forming a feature importance distribution weight set for each target sampling point, which serves as the input for interpretive analysis.

[0212] S8.3: For the feature importance distribution weight set of the target sampling points, combined with dynamic scene labels and historical prediction experience base, perform feature contribution aggregation and visualization transformation, and map the quantified feature importance data into interpretable label output to generate fault cause explanation information for operation and maintenance support.

[0213] The feature importance distribution weight set, multi-timescale fusion predictor, and inferred dynamic scene label data for each target sampling point are used as input conditions.

[0214] A feature contribution aggregation algorithm (parameters: feature importance distribution weight set, dynamic scene label strength vector, historical prediction experience database mapping table) is adopted to realize multi-dimensional aggregation processing of fault feature contributions, and to merge and integrate feature contributions from different dimensions and time scales.

[0215] Furthermore, through the visualization conversion module (parameters: feature contribution aggregation matrix, scene label descriptor, historical experience interpretation template), the quantitative data of feature contribution is transformed into interpretable label output, and the significance of the contribution of each feature or feature group in the current scene is expressed by standardized labels and graphical symbols.

[0216] Furthermore, through the fault cause tracing algorithm (parameters: historical prediction experience base, feature contribution aggregation result, scene label classification model), the algorithm automatically retrieves the historical fault cause labels most relevant to the current sampling point scene and feature combination, and identifies key influencing factors and fault types.

[0217] Furthermore, by using a label mapping output method, the aggregated explanatory label group is bidirectionally bound to the quantitative feature contribution set to generate fault cause explanation data for each target sampling point, including: main fault impact features, scene triggering factors, feature contribution ranking, and historical experience descriptions.

[0218] By using feature contribution aggregation and visualization transformation algorithms, the feature importance distribution and scene label content obtained in the previous step are transformed into a set of fault cause labels with high interpretability, thereby achieving a visual analysis effect of fault features for operation and maintenance support.

[0219] For example, under conditions of severe thunderstorm fluctuations and aging insulators, the feature contribution weighting matrix of target sampling point P1 is (0.288, 0.123, 0.158, 0.264, 0.544, 0.239), where feature 0.544 corresponds to the contribution of insulator failure probability, and 0.288 corresponds to the contribution of load fluctuation. Historical experience database retrieval shows that this combination has a 93% correlation with "lightning aging failure" type faults. The system automatically generates fault cause explanation labels: "Aging insulators significantly increase fault risk (contribution 54%), severe thunderstorm load fluctuations are a secondary triggering factor (contribution 29%), it is recommended to pay attention to insulator insulation strength replacement and load dynamic stability control." The output label format is structured JSON, including contribution ranking, scenario triggering factors, and historical experience reference explanations, effectively supporting intelligent operation and maintenance reports and fault source tracing analysis.

[0220] S8.4: The failure probability output and feature importance interpretation information generated by the prediction model are formatted, integrated, and encoded to generate a structured prediction result data package that includes a unique identifier for the target sampling point, a prediction scenario label, the model failure probability result, and the feature weight interpretation. This data package serves as a standardized input for downstream operation and maintenance decisions and continuous model validation.

[0221] S8.5: Taking structured prediction result data packets as input, it automatically connects to the distribution network intelligent operation and maintenance system to realize real-time reporting and early warning push of prediction information, and further records the fault prediction and feature interpretation results in the traceability database, providing a complete traceable data chain for subsequent model retraining and model credibility evaluation.

[0222] Step S9: Real-time monitoring of the prediction accuracy and stability of the adaptively calibrated model; A / B comparison analysis of the prediction results with the original model; if the new model has better short-term accuracy and stability than the original model, automatically switch to the new model; otherwise, roll back to the original model parameter state, and record the tuning process log for subsequent performance optimization. Specifically, this includes:

[0223] S9.1: Continuously collect the distribution line fault prediction results output by the adaptively calibrated fault prediction model online, monitor its prediction accuracy and stability indicators based on real-time data stream, and obtain short-term performance evaluation parameters during the actual operation of the model, including indicators such as prediction accuracy score and stability fluctuation rate, to provide basic performance factors for subsequent comparative analysis.

[0224] S9.2: Using an A / B comparison experiment mechanism, based on the prediction results of the adaptive calibration model and the original fault prediction model under the same power distribution line operating conditions, the fault prediction accuracy, false alarm rate and stability standard deviation parameters of the two sets of models are calculated within the same time window to obtain a model performance comparison factor for mutual reference, so as to realize the performance significance analysis of different models under dynamic operating conditions.

[0225] S9.3: When the comparative analysis results show that the prediction accuracy score and stability volatility of the adaptive calibration model are better than those of the original model, the model primary weight switching operation is automatically executed. Based on the model performance factor, the adaptive calibration model is set as the current primary fault prediction model, and the updated model weight parameter set is output to ensure that the system can apply the optimal prediction capability in a timely manner under sudden operating conditions.

[0226] S9.4: If the A / B comparison analysis results show that the short-term performance of the adaptive calibration model is inferior to that of the original model, the model parameter rollback mechanism will be automatically executed to restore the parameter weight state of the original fault prediction model, downgrade the substandard adaptive calibration model to a backup model, and output the model rollback operation log to ensure the stable and reliable operation of the fault prediction system.

[0227] S9.5: Periodically record the execution logs of the whole process model monitoring, A / B comparison analysis, weight switching and parameter rollback operations. Generate model performance traceability data blocks based on the optimization process logs and store them in the optimization database in the form of structured log data. Support subsequent model performance evaluation, optimization tracking and source analysis of model optimization history under dynamic running conditions.

[0228] 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.

[0229] 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.

[0230] 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 smart grid distribution line fault prediction analysis method, specifically comprising: S1: Collecting operation indexes and environmental perception data of multiple sampling points in the distribution line, recording the unique identification and corresponding aging state information of each sampling point; S2: Processing the collected multi-source operation indexes and environmental perception data of the distribution line to improve the quality; S3: Based on the data set that has been normalized and contains the aging state information of the components, systematically extracting the multi-time scale features of the distribution line; S4: Identifying the change state of the working condition in the real-time operation data stream of the distribution line, and realizing the automatic generation of dynamic scene labels through the joint analysis of dynamic operation feature vectors and sampling point environmental perception data; S5: Based on the dynamic scene labels generated by real-time monitoring, migrating the fault prediction model parameters and feature importance weight set stored in the historical database under similar working conditions to the current new working condition scene; S6: Using limited new sampling samples and their feature vectors, incrementally fine-tuning the feature weights of the fault prediction model through a few-shot meta-learning algorithm; S7: Fusing and weighting the core weight set of the self-adaptive calibrated fault prediction model, fully integrating dynamic operation feature weights and component aging feature weights; S8: Taking the multi-time scale fusion prediction factor as input to drive the fault prediction model to output the distribution line fault prediction result under the corresponding dynamic working condition scene.

2. The method of claim 1, wherein, The operation indexes in step S1 include real-time load, voltage, current, harmonic and environmental temperature and humidity, climate variables.

3. The method of claim 1, wherein, The quality improvement processing in step S2 specifically includes noise suppression, outlier rejection and data normalization.

4. The method of claim 1, wherein, The systematic extraction of multi-time scale features of the distribution line in step S3 includes short-period dynamic operation features and long-period aging features.

5. The method of claim 1, wherein, After step S8, S9 is further included: monitoring the prediction effect and stability of the fault prediction model that has completed self-adaptive calibration in the actual running environment in real time, and comparing the model performance by using A / B comparison analysis method.

6. The method of claim 1, wherein, The step S3 specifically includes: Using a time-scale data scheduling strategy to construct a multi-time scale data mapping relationship for the normalized data set containing the unique identification and aging state information of the components; For each grouped short-period dynamic operation data, calculate the key operation feature parameters to generate a short-period dynamic operation feature matrix; For each grouped long-period component aging state parameter, calculate the aging state feature value based on the aging trend modeling algorithm to form a long-period aging feature matrix; For each sampling point and its corresponding component unique identification, perform feature mapping assembly operation to generate a continuous feature vector group; Based on the feature vector group, use the feature effectiveness evaluation method to analyze the contribution of each feature under the historical fault data to form a labeled feature library.

7. The method of claim 1, wherein, The step S4 specifically includes: Based on the normalized feature vector group containing the aging state of the components, perform dynamic operation feature change monitoring to extract the real-time load change rate, voltage fluctuation standard deviation, and current instantaneous deviation index; Obtain the sampling point environmental perception data corresponding to the dynamic operation feature vector group to generate the same time window environmental factor feature group; Based on the feature change amplitude threshold decision mechanism, the working condition change is determined and the trigger factor is marked; For the data window determined as working condition change, the sampling point environment factor group and the dynamic running feature change factor are fused to generate a dynamic scene label; The dynamic scene label and its corresponding trigger factor are taken as output results to provide scene-based input basis for subsequent fault prediction model parameter migration.

8. The method of claim 1, wherein, The step S5 specifically includes: Classify and identify the dynamic scene label, and retrieve a historical working condition label set with the highest similarity to the current dynamic scene label; Map the model parameters for the candidate working condition label sequence, and screen the model parameters and feature weight set in the similar scene; Convert the structure mapping based on the label migration algorithm using the screened historical scene model parameters and feature importance weight set; Fuse the specific environmental variables and component aging state features in the current dynamic scene label with the model parameter weight set obtained by preliminary migration, and perform weighted fusion; The fault prediction model weight vector after fusion processing is taken as the start-up configuration parameter of the fault prediction model in the current new scene.

Images