Abnormal pattern data processing system driven by power marketing big data

Abstract

The present application relates to the technical field of data processing, and more particularly to an abnormal mode data processing system driven by power marketing big data, comprising a distributed collaborative acquisition module for constructing space-time aligned three-dimensional data flow, a multi-modal feature reconstruction module for separating periodic noise and quantifying environmental interference, an adversarial feature decoupling module for generating a purified feature vector set and a noise confidence index through orthogonal projection, a dynamic algorithm adaptation module for dynamically scheduling isolated forest algorithm, weighted distance measurement algorithm and sparse self-encoding clustering algorithm according to the noise confidence index, a behavior chain verification module for establishing a joint physical rule verification mechanism of environmental temperature threshold, load deviation degree and equipment state, and a closed-loop strategy engine module for adaptively adjusting the feature decoupling loss function weight according to the decision boundary offset, thereby effectively improving the accuracy and environmental adaptability of real power consumption anomaly identification in a complex noise environment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and in particular to an abnormal mode data processing system driven by power marketing big data. BACKGROUND

[0002] Power marketing big data is the process of collecting, processing and analyzing large-scale power consumption information data sets in the smart grid environment using advanced information technology in the power industry. The purpose is to optimize energy distribution, improve operational efficiency and improve customer service experience. By using complex algorithms to mine power consumption patterns, identify potential demand response opportunities, support accurate load forecasting and personalized marketing strategy development, it drives decision optimization, reduces network loss and enhances service reliability. In the development of modern power systems, power marketing big data plays a core role in information-driven, providing market trend insights and deep understanding of user behavior, and helping the industry evolve towards efficiency, intelligence and sustainability.

[0003] In the environment of power marketing big data, the abnormal mode data processing has the technical pain point of insufficient recognition accuracy under high-dimensional noise interference, especially in the massive power consumption sequence. Due to the influence of data sparsity and complex time sequence characteristics, it is difficult for the algorithm to accurately distinguish between real abnormal modes such as electricity fraud and normal behavior fluctuations. For example, in the analysis of user power consumption time series collected by smart meters, standard clustering models are sensitive to regular use changes caused by holidays or weather changes, resulting in false positive results, and normal users are incorrectly classified as abnormal, which weakens the reliability of load forecasting and the effectiveness of demand response strategy implementation. SUMMARY

[0004] In view of the deficiencies of the prior art, the present application provides an abnormal mode data processing system driven by power marketing big data. The present application solves the problem of insufficient recognition accuracy of abnormal modes under high-dimensional noise interference of power consumption time series, and cannot effectively distinguish between real electricity abnormality and normal dynamic fluctuations.

[0005] To solve the above technical problems, the specific technical solutions of the present application are as follows:

[0006] The abnormal mode data processing system driven by power marketing big data provided by the present application comprises:

[0007] A distributed collaborative collection module collects smart meter power time series data and environmental parameters, and outputs three-dimensional data streams including time, physical quantity and spatial dimension;

[0008] A multi-modal feature reconstruction module receives the three-dimensional data stream, applies a time window to segment and extract spectral features and environmental response features, and generates a load feature map;

[0009] The adversarial feature decoupling module receives the load feature map, performs feature space orthogonal projection to separate fluctuation feature vectors and noise confidence indicators, and outputs a purified fluctuation feature vector set;

[0010] The dynamic algorithm adaptation module receives the fluctuation feature vector set and noise confidence indicators, dynamically activates an optimization algorithm combination according to feature sparsity and noise confidence, and generates a set of abnormal suspicion points.

[0011] The behavior chain verification module receives the set of abnormal suspicion points, performs multi-dimensional verification to distinguish abnormal events from reasonable fluctuations, and outputs a list of verified abnormal events and normal dynamic fluctuation samples that pass verification.

[0012] The closed-loop strategy engine module receives the list of verified abnormal events and the normal dynamic fluctuation samples that pass verification, triggers marketing inspection instructions for events in the list of verified abnormal events, handles real abnormal electricity use, imports the normal dynamic fluctuation samples that pass verification into a historical feature library, adjusts periodic feature extraction parameters of the multi-modal feature reconstruction module through an incremental learning mechanism, adjusts loss function weights of the adversarial feature decoupling module according to a decision boundary offset.

[0013] Further, the distributed collaborative collection module of the power marketing big data-driven abnormal mode data processing system is configured to:

[0014] Deploy terminal devices at power grid nodes to periodically collect smart meter voltage and current instantaneous values, and obtain meteorological service parameters and power grid geographic coordinates.

[0015] The three-dimensional data stream includes a timestamp, a meter identifier, a power value, a temperature value, a humidity value, a longitude value, and a latitude value.

[0016] The periodically collected smart meter voltage and current instantaneous values and the obtained meteorological service parameters and power grid geographic coordinates are input into the multi-modal feature reconstruction module for segmentation.

[0017] Further, the multi-modal feature reconstruction module of the power marketing big data-driven abnormal mode data processing system is configured to:

[0018] Performing frequency spectrum feature extraction on the input three-dimensional data stream: applying fast Fourier transform to analyze daily periodic components and weekly and monthly periodic features to generate a standardized periodic feature spectrum; performing environmental response feature extraction: constructing a nonlinear correlation model of load values and temperature values to output a set of temperature sensitivity coefficients; and fusing the standardized periodic feature spectrum and the set of temperature sensitivity coefficients to generate a basic load feature map.

[0019] Identify the seasonal dependence mode in the basic load feature atlas, and add a seasonal identification factor to the atlas; input the load feature atlas with the seasonal identification factor into the anti-feature decoupling module to perform orthogonal projection processing.

[0020] Further, the anti-feature decoupling module of the power marketing big data driven abnormal mode data processing system is configured to:

[0021] The input load feature atlas is decomposed into a fluctuation feature subspace and a noise feature subspace by an encoder; a generator is used to reconstruct feature samples, and a discriminator is used to compare the reconstructed samples with the original samples to calculate the error amplitude; a fluctuation feature vector set containing 64 purified features and a noise confidence index are generated based on the error amplitude, wherein the noise confidence index is a continuous numerical value; and the fluctuation feature vector set and the noise confidence index are transmitted as output to the dynamic algorithm adaptation module.

[0022] Further, the dynamic algorithm adaptation module of the power marketing big data driven abnormal mode data processing system is configured to:

[0023] The feature sparsity, distribution kurtosis and noise confidence index of the input feature are monitored in real time; if the feature sparsity exceeds a set threshold, a sparse self-encoding clustering algorithm is called to process the feature; if the feature distribution kurtosis exceeds a set threshold, a weighted distance measurement algorithm is called to process the feature; if the noise confidence index reaches a preset critical value, the detection is performed after adjusting the depth parameter of the isolation forest algorithm; the suspicious points output by all algorithms are integrated to generate an abnormal suspicious point set marked with a floating radius decision boundary, and the abnormal suspicious point set marked with the decision boundary is output to the behavior chain verification module for multidimensional verification.

[0024] Further, the behavior chain verification module of the power marketing big data driven abnormal mode data processing system is configured to:

[0025] The current environmental temperature value, power grid instantaneous load value and power distribution equipment operating state code are obtained from real-time monitoring sources, the historical load baseline data of the target user are called, the power distribution equipment operating state code and the historical load baseline data are jointly analyzed, a dynamic constraint rule group is generated, and based on the dynamic constraint rule group, the deviation degree of the environmental temperature value, instantaneous load value and historical load baseline data, and whether the power distribution equipment operating state code satisfies the preset legality condition are detected.

[0026] If both conditions are met, the corresponding abnormal suspicious point is unmarked; the abnormal suspicious points that are not unmarked are summarized to generate a verification abnormal event list, and the verification abnormal event list is output to the closed-loop strategy engine module to drive the update of the marketing inspection strategy library.

[0027] Further, the power marketing big data driven abnormal mode data processing system disclosed by the application, the closed loop strategy engine module is configured to:

[0028] Based on the verification passing sample of the newly imported historical feature library, the periodic feature component phase angle is recalculated and the corresponding hash index structure is updated;According to the change amplitude of the recalculated periodic component phase angle, the loss function weight adjustment instruction of the adversarial feature decoupling module is generated by triggering the incremental learning mechanism;The loss function weight parameter of the adversarial feature decoupling module is updated by executing the adjustment instruction;The updated periodic component phase angle, hash index structure and loss function weight parameter are synchronized to the multi-modal feature reconstruction module and adversarial feature decoupling module, and the cross-module parameter collaborative refresh is completed.

[0029] Further, the power marketing big data driven abnormal mode data processing system disclosed by the application further comprises:

[0030] The noise confidence index is generated by the adversarial feature decoupling module and input into the dynamic algorithm adaptation module;

[0031] The dynamic algorithm adaptation module executes the noise confidence based optimization algorithm dynamic switching, including: presetting the noise confidence index threshold interval;

[0032] When the noise confidence index falls into the low noise interval, the isolated forest algorithm is activated to execute the core suspicious point identification;When the noise confidence index falls into the medium noise interval, the weighted distance measurement algorithm is activated to execute the anti-interference anomaly detection;

[0033] When the noise confidence index falls into the high noise interval, the sparse self-encoding clustering algorithm is activated to execute the feature dimension reduction and density clustering;

[0034] The selected optimization algorithm processes the fluctuation feature vector set according to the current input feature, and generates an abnormal suspicious point set.

[0035] Further, the power marketing big data driven abnormal mode data processing system disclosed by the application further comprises:

[0036] The behavior chain verification module outputs the verified passing normal dynamic fluctuation sample to the closed loop strategy engine module;

[0037] The closed loop strategy engine module imports the verified passing normal dynamic fluctuation sample into the historical feature library, and updates the periodic feature extraction parameter of the multi-modal feature reconstruction module through the incremental learning mechanism, so that the frequency spectrum feature extraction of the multi-modal feature reconstruction module adapts to the new sample data, and an operation mode of verifying the result driven feature reconstruction module parameter update is established.

[0038] Further, the power marketing big data driven abnormal mode data processing system disclosed by the application adjusts the loss function weight of the adversarial feature decoupling module according to the decision boundary offset, and the adjusting includes:

[0039] Real-time monitoring of the decision boundary offset marked by the abnormal suspicion point set generated by the dynamic algorithm adaptation module;

[0040] When the decision boundary offset falls into the preset high offset interval, the weight coefficient of the fluctuation feature reconstruction loss is reduced;

[0041] When the decision boundary offset falls into the preset low offset interval, the target weight of the noise discrimination loss is increased;

[0042] The dynamically adjusted loss function weight is synchronized to the output layer executor of the adversarial feature decoupling module.

[0043] Advantages of the application

[0044] The application eliminates collection noise by constructing a spatiotemporal alignment three-dimensional data stream through a distributed collaborative collection module, separates periodic background noise and quantifies environmental interference through a multi-modal feature reconstruction module, executes orthogonal projection to generate a purified feature vector set and a noise confidence index through an adversarial feature decoupling module, and reduces high-dimensional noise interference from the source of the feature; The dynamic algorithm adaptation module dynamically schedules the isolation forest algorithm, the weighted distance measurement algorithm and the sparse self-encoding clustering algorithm according to the noise confidence index, and realizes accurate differentiation of real abnormal events and environmental response fluctuations in combination with the multi-parameter physical rule joint detection mechanism of the behavior chain verification module; The closed-loop strategy engine module incrementally updates the periodic feature parameters based on the verification result, and adjusts the feature decoupling loss function weight according to the decision boundary offset, forming a closed-loop learning system of feature extraction, anomaly detection and feedback optimization, continuously improving the accuracy and environmental adaptability of abnormal identification in a complex noise environment, and effectively solving the technical problem of insufficient recognition accuracy of abnormal patterns and normal dynamic fluctuations in power consumption time series data. BRIEF DESCRIPTION OF DRAWINGS

[0045] In order to more clearly illustrate the technical solutions of the application, the following will briefly introduce the drawings needed in the embodiments. Obviously, for those skilled in the art, other drawings can also be obtained without creative labor on the basis of the drawings.

[0046] Figure 1 The system architecture diagram of the power marketing big data driven abnormal mode data processing system provided by the embodiment of the application. DETAILED DESCRIPTION

[0047] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions of the present application will be described clearly and completely below in combination with specific embodiments of the present application and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application. The technical solutions provided by the embodiments of the present application will be described in detail below in combination with the drawings. In order to better understand the objects of the present application, the present application will be further described in detail below.

[0048] Please refer to Figure 1 The present application provides an abnormal mode data processing system driven by power marketing big data, comprising:

[0049] A distributed collaborative acquisition module acquires intelligent electric meter power time series data and environmental parameters, and outputs three-dimensional data flow including time, physical quantity and spatial dimension;

[0050] A multi-modal feature reconstruction module receives the three-dimensional data flow, applies time window segmentation to extract frequency spectrum features and environmental response features, and generates a load feature map;

[0051] An adversarial feature decoupling module receives the load feature map, performs feature space orthogonal projection to separate fluctuation feature vectors and noise confidence indicators, and outputs a set of purified fluctuation feature vectors;

[0052] A dynamic algorithm adaptation module receives the set of fluctuation feature vectors and noise confidence indicators, dynamically activates an optimized algorithm combination according to feature sparsity and noise confidence, and generates a set of abnormal suspicion points;

[0053] A behavior chain verification module receives the set of abnormal suspicion points, performs multi-dimensional verification to distinguish abnormal events from reasonable fluctuations, and outputs a list of verified abnormal events and verified normal dynamic fluctuation samples;

[0054] A closed-loop strategy engine module receives the list of verified abnormal events and the verified normal dynamic fluctuation samples, triggers marketing inspection instructions for events in the list of verified abnormal events, handles real power abnormality, imports the verified normal dynamic fluctuation samples into a historical feature library, adjusts periodic feature extraction parameters of the multi-modal feature reconstruction module through an incremental learning mechanism, adjusts loss function weights of the adversarial feature decoupling module according to decision boundary offset.

[0055] The distributed cooperative acquisition module periodically acquires the instantaneous values of the voltage and current of the electric meter through the intelligent terminal device deployed at the nodes of the power grid, synchronously accesses the weather service interface to obtain the temperature and humidity parameters, and extracts the spatial coordinate data in combination with the power grid geographic information system. The module integrates multi-source heterogeneous data by using timestamp alignment technology to generate standardized three-dimensional data flow containing time dimension, physical quantity dimension and spatial dimension.

[0056] After receiving the three-dimensional data flow, the multi-modal feature reconstruction module first applies a sliding time window segmentation technique to divide the data units, performs a fast Fourier transform on each data unit to extract the daily periodic component and the weekly and monthly periodic feature spectrum. A load-temperature nonlinear regression model is simultaneously constructed to calculate the temperature sensitivity coefficient matrix. The periodic feature spectrum and the temperature sensitivity coefficient matrix are fused by tensor to generate a basic load feature spectrum. Seasonal identification factors are labeled in the basic spectrum by a seasonal pattern recognition algorithm to form a load feature spectrum with seasonal characteristics.

[0057] The adversarial feature decoupling module processes the load feature spectrum using an encoder-generator-discriminator architecture. The encoder performs orthogonal basis vector decomposition to separate the fluctuation feature subspace and the noise feature subspace; after the generator reconstructs the feature samples, the discriminator calculates the Euclidean distance between the original samples and the reconstructed samples as the error amplitude. Based on the error amplitude, a set of 64-dimensional purified fluctuation feature vectors and a continuous noise confidence index are generated.

[0058] The dynamic algorithm adaptation module monitors the sparsity distribution, kurtosis coefficient and noise confidence index of the input features in real time. When the feature sparsity exceeds the set threshold, the sparse auto-encoding clustering algorithm is activated; when the feature kurtosis exceeds the threshold, the weighted Mahalanobis distance measurement algorithm is called; and according to the threshold interval in which the noise confidence index is located, the tree depth parameter of the Isolation Forest algorithm is dynamically adjusted. The suspicious points output by each algorithm are processed by the decision boundary fusion module to generate a set of abnormal suspicious points with floating radius markers.

[0059] The behavior chain verification module calls the real-time state code of the distribution equipment, associates with the user's historical load baseline to build dynamic constraint rules. The environmental temperature value, load deviation rate and equipment state code are detected by the physical logic chain verification engine to check the compliance, and the suspicious points that meet the constraint conditions are unmarked as abnormal. The suspicious points that are not unmarked are aggregated to generate a list of verification abnormal events.

[0060] The closed-loop strategy engine module imports the normal samples that pass the verification into the historical feature library, triggers the incremental learning mechanism to recalculate the periodic feature phase angle and update the hash index structure. Based on the change amplitude of the phase angle, a loss function weight adjustment instruction is generated to adjust the weight proportion of the fluctuation feature reconstruction loss and the noise discrimination loss in the adversarial feature decoupling module. The updated periodic component parameters, hash index and loss weight are fed back to the feature reconstruction module and the feature decoupling module synchronously to complete the cooperative optimization of system parameters.

[0061] Each module forms a closed-loop technical path: three-dimensional data stream drives feature reconstruction, feature map decouples to generate purified feature vectors, dynamic algorithm adaptation generates a set of suspicious points, behavior chain verification outputs verification results, and the strategy engine module optimizes feature extraction and decoupling parameters according to the verification results. The system realizes the continuous evolution of processing capacity through an incremental learning mechanism, and solves the problem of identifying abnormal patterns and normal fluctuations in power marketing data.

[0062] Specifically, the distributed collaborative collection module of the power marketing big data driven abnormal pattern data processing system is configured to:

[0063] Deploy terminal devices at power grid nodes to periodically collect smart meter voltage and current instantaneous values, and obtain meteorological service parameters and power grid geographic coordinates;

[0064] The three-dimensional data stream includes a timestamp, a meter identifier, a power value, a temperature value, a humidity value, a longitude value, and a latitude value;

[0065] The periodically collected smart meter voltage and current instantaneous values and the obtained meteorological service parameters and power grid geographic coordinates are input into the multi-modal feature reconstruction module for segmentation.

[0066] The distributed collaborative collection module periodically collects voltage and current instantaneous value data of smart meters through intelligent monitoring terminal devices deployed at power grid substations and distribution areas, and the sampling frequency is synchronized with the power grid frequency period. The module obtains temperature value and humidity value environmental parameters in real time through a meteorological service interface, and calls an electric power geographic information system interface to extract longitude value and latitude value spatial coordinates of the installation location of the device. High-precision time synchronization protocol is used to align the time stamps of multi-source data, forming a structured three-dimensional data stream containing time stamp, meter identifier, power value, temperature value, humidity value, longitude value, and latitude value fields.

[0067] During the collection process, the voltage and current instantaneous values of the smart meter are transmitted to the edge computing node through power line carrier communication, and the power value is generated by integrating the instantaneous value. Meteorological service parameters are obtained through an application programming interface, and power grid geographic coordinates are called from a spatial database in real time. The three-dimensional data stream uses the timestamp as the primary index key to establish a data cube structure with time dimension, physical quantity dimension, and spatial dimension. The data cube is transmitted to the multi-modal feature reconstruction module through a message queue to provide standardized input for subsequent time window segmentation.

[0068] After receiving the three-dimensional data stream, the multi-modal feature reconstruction module first divides a sliding time window according to a timestamp sequence. The power value data in each window is subjected to fast Fourier transform to extract frequency spectrum features, and the temperature value and humidity value are used to construct an environmental response model. The longitude value and latitude value are used for spatial grid coding to realize spatial alignment of multi-dimensional features. The data cube structure guarantees the continuity of the time sequence, the integrity of the physical quantity, and the consistency of the spatial position, thereby laying a data foundation for feature reconstruction.

[0069] The application establishes a complete data acquisition chain: the terminal device layer realizes physical signal acquisition, the data transmission layer completes multi-source heterogeneous data fusion, and the data modeling layer generates standardized three-dimensional data streams. Each link adopts an industrial Internet of Things architecture and conforms to the power system data acquisition specification, thereby providing high-reliability input for subsequent feature processing.

[0070] Specifically, the multi-modal feature reconstruction module of the power marketing big data driven abnormal mode data processing system is configured to:

[0071] performing frequency spectrum feature extraction on the input three-dimensional data stream: applying fast Fourier transform to analyze daily periodic components and week-month periodic features to generate a standardized periodic feature spectrum; performing environmental response feature extraction: constructing a nonlinear correlation model of load values and temperature values to output a temperature sensitivity coefficient set; and fusing the standardized periodic feature spectrum and the temperature sensitivity coefficient set to generate a basic load feature spectrum.

[0072] identifying a seasonal dependence mode in the basic load feature spectrum and adding a seasonal identification factor to the spectrum; and inputting the load feature spectrum with the seasonal identification factor into the adversarial feature decoupling module to perform orthogonal projection processing.

[0073] After receiving the three-dimensional data stream output by the distributed collaborative acquisition module, the multi-modal feature reconstruction module first performs a frequency spectrum feature extraction operation. This operation applies a fast Fourier transform algorithm to process the power value time sequence to decompose daily periodic components and week-month periodic component features. A standardized periodic feature spectrum is generated through frequency spectrum energy normalization processing to eliminate amplitude deviation caused by different user power levels. The periodic feature spectrum represents the inherent regularity mode of the power load change over time.

[0074] An environmental response feature extraction operation is simultaneously performed to construct a nonlinear regression model of load values and temperature values. A kernel function mapping technology is used to establish the correlation between temperature changes and load fluctuations to output a temperature sensitivity coefficient set. The coefficient set contains sensitivity parameters of load changes in different temperature intervals, which quantifies the influence intensity of environmental factors on power consumption. The temperature sensitivity coefficient set and the standardized periodic feature spectrum have the same time dimension alignment benchmark.

[0075] The feature fusion stage performs tensor superposition operation on the normalized periodic feature spectrum and the temperature sensitivity coefficient set. The two heterogeneous features are mapped to a unified high-dimensional space through feature dimension expansion technology to generate a basic load feature graph. The graph fuses the time periodicity feature and the environmental response characteristic to form a feature benchmark model of user power consumption behavior. Each feature node in the graph contains a timestamp, a spatial coordinate and a multi-dimensional feature vector.

[0076] Based on the basic load feature graph, a seasonal pattern recognition operation is performed, and a time series pattern mining algorithm is used to analyze the seasonal dependence law in the historical data. The recognition result is marked as a seasonal identification factor in the form of a discrete classification variable and is attached to the corresponding time node of the basic load feature graph. The seasonal identification factor represents the feature offset of the power consumption mode in different seasonal periods and enhances the time context information of the feature graph.

[0077] Finally, a load feature graph with seasonal identification factors is generated and transmitted to the adversarial feature decoupling module through a data interface. The graph serves as input data for orthogonal projection processing, and its multi-dimensional feature structure meets the input requirements of the feature space decomposition algorithm. The periodic, environmental sensitivity and seasonal features fused in the feature graph provide a complete information base for subsequent feature decoupling.

[0078] The present application establishes a progressive relationship of feature processing: the original data extracts time period features through frequency spectrum transformation, quantifies the influence of external factors through environmental modeling, constructs a comprehensive representation model through feature fusion, and enhances the time dimension information through seasonal identification. The output of each link maintains the alignment of time and space dimensions, forming a structured feature evolution process.

[0079] Specifically, the adversarial feature decoupling module of the power marketing big data driven abnormal mode data processing system described in the present application is configured to:

[0080] The input load feature graph is decomposed into a fluctuation feature subspace and a noise feature subspace through an encoder; a generator is used to reconstruct the feature sample, and a discriminator is used to compare the reconstructed sample with the original sample to calculate the error amplitude; based on the error amplitude, a fluctuation feature vector set and a noise confidence index are generated, wherein the fluctuation feature vector set contains 64-dimensional purified features, and the noise confidence index is a continuous numerical value; the fluctuation feature vector set and the noise confidence index are transmitted as output to a dynamic algorithm adaptation module.

[0081] After receiving the load feature graph with seasonal identification factor, the anti-feature decoupling module first performs feature space orthogonal decomposition operation through an encoder network. The encoder uses orthogonal basis vector projection technology to map the input feature graph to two mutually exclusive subspaces: a fluctuation feature subspace representing the essential law of power consumption behavior, and a noise feature subspace reflecting environmental noise and random interference. The subspace decomposition process satisfies the orthogonality constraint condition, ensuring the completeness of feature decoupling.

[0082] The generator network reconstructs the feature sample based on the fluctuation feature subspace and reconstructs the load feature representation through deconvolution operation. The discriminator network simultaneously receives the original load feature graph and the reconstructed feature sample, and uses feature distance calculation algorithm to quantify the difference between the two. The difference value is the reconstruction error amplitude, which reflects the degree of loss of information integrity in the feature decoupling process.

[0083] The error amplitude data stream drives the feature purification mechanism, and the fluctuation feature subspace is processed by dimension compression to generate a set of fluctuation feature vectors. The vector set contains fixed-dimension purified feature vectors that eliminate high-dimensional noise interference in the original data. Simultaneously, the reconstruction error amplitude is converted into a normalized noise confidence index, which is a continuous numerical value representing the probability level of the current sample being contaminated by noise.

[0084] The final output is a data combination containing the fluctuation feature vector set and the noise confidence index, which is transmitted to the dynamic algorithm adaptation module through a high-speed data bus. The fluctuation feature vector set is the core feature input, and the noise confidence index is the algorithm selection auxiliary parameter, which together form the data basis of the downstream anomaly detection process. The feature dimension and data type strictly adapt to the input specification of the dynamic algorithm adaptation module.

[0085] The present application constructs a closed-loop verification mechanism for feature processing: spatial decomposition establishes the basis for feature separation, generation and discrimination structure to achieve decoupling effect verification, and error analysis drives the generation of purified features. The orthogonality constraint of the subspace guarantees the theoretical completeness of feature decoupling, and the feature distance calculation provides a quantifiable evaluation basis for decoupling effect. The output data combination contains both purified features and noise evaluation indicators, forming a multi-dimensional feature quality description system.

[0086] Specifically, the power marketing big data driven abnormal mode data processing system of the present application, the dynamic algorithm adaptation module is configured to:

[0087] The feature sparsity, distribution kurtosis and noise confidence index of the input features are monitored in real time; if the feature sparsity exceeds a set threshold, a sparse self-encoding clustering algorithm is called to process the features; if the feature distribution kurtosis exceeds a set threshold, a weighted distance measurement algorithm is called to process the features; if the noise confidence index reaches a preset critical value, the depth parameter of the isolation forest algorithm is adjusted and then detection is performed; all algorithm outputs are integrated to generate an abnormal suspect point set marked with a floating radius decision boundary, and the abnormal suspect point set marked with the decision boundary is output to a behavior chain verification module for multi-dimensional verification.

[0088] The dynamic algorithm adaptation module receives the fluctuation feature vector set and noise confidence index input in real time, and continuously monitors the sparsity distribution characteristics, statistical kurtosis characteristics and noise confidence values of the feature vectors through the feature analysis engine. The feature sparsity reflects the distribution density of effective information points in the feature space, the statistical kurtosis represents the steepness of the feature value distribution, and the noise confidence quantifies the degree of data interference. The monitoring process uses a sliding time window update mechanism to ensure the timeliness of parameter evaluation.

[0089] When the feature sparsity exceeds the preset threshold, the sparse self-encoding clustering algorithm processing procedure is activated. This algorithm extracts key features through dimension reduction in the encoding layer and performs density clustering analysis in the hidden layer space, effectively identifying abnormal clustering patterns in high-dimensional sparse features. The algorithm outputs a primary suspect point set labeled with a clustering outlier coefficient.

[0090] When the feature distribution kurtosis exceeds the preset threshold, the weighted distance measurement algorithm processing procedure is called. This algorithm is improved based on Mahalanobis distance calculation, and different dimension weight coefficients are given according to the feature variance distribution, enhancing the robustness of abnormal detection for non-Gaussian distribution features. The algorithm outputs a candidate abnormal point set with distance deviation.

[0091] When the noise confidence index reaches the preset critical value, the isolation forest algorithm depth adjustment mechanism is triggered. By dynamically increasing the depth parameter of the decision tree, the algorithm's inclusiveness for high-noise data is improved, reducing the false positive rate while maintaining the accuracy of anomaly detection. The adjusted isolation forest algorithm outputs an abnormal score set based on path length after detection.

[0092] The multi-algorithm output results are processed by the decision boundary fusion module, which uses fuzzy logic rules to integrate the suspect labels generated by different algorithms. The fusion process generates a dynamic floating radius decision boundary based on the clustering outlier coefficient, distance deviation and abnormal score. This boundary adaptively adjusts the abnormal judgment threshold range. Finally, an abnormal suspect point set marked with a floating radius is generated, each suspect point containing spatial coordinates, timestamp and abnormal probability value.

[0093] The abnormality suspicion point set is transmitted to the behavior chain verification module through a standardized interface, and the data format is adapted to the input specification of the multi-dimensional verification engine. The transmission process preserves the time and space dimension information, ensuring the traceability association of the suspicion point set and the original power data.

[0094] The application establishes a logical level of algorithm scheduling: the monitoring layer continuously quantifies feature characteristics, the decision layer triggers algorithm instances according to the quantification results, the execution layer outputs primary detection results, and the fusion layer generates unified judgment standards. The multi-algorithm parallel processing architecture realizes resource optimization allocation through feature attribute discrimination, and the floating boundary mechanism improves the environmental adaptability of abnormality judgment.

[0095] Specifically, the behavior chain verification module of the power marketing big data driven abnormal mode data processing system is configured to:

[0096] Obtain the current environment temperature value, the power grid instantaneous load value and the power distribution equipment operation state code from the real-time monitoring source, retrieve the historical load baseline data of the target user, jointly analyze the power distribution equipment operation state code and the historical load baseline data, generate a dynamic constraint rule group, and detect whether the deviation degree of the environment temperature value, the instantaneous load value and the historical load baseline data and the power distribution equipment operation state code satisfy the preset legality condition based on the dynamic constraint rule group;

[0097] If they are satisfied at the same time, the corresponding abnormal suspicion point is unmarked; the unmarked abnormal suspicion points are summarized to generate a verification abnormal event list, and the verification abnormal event list is output to the closed-loop strategy engine module to drive the update of the marketing inspection strategy library.

[0098] The behavior chain verification module first obtains the current environment temperature value, the power grid instantaneous load value and the power distribution equipment operation state code from the power grid monitoring system in real time. The target user load baseline data stored in the historical feature library is retrieved synchronously, which contains the typical power consumption mode under the same season identifier. The device operation state code contains device operation state flags, maintenance state identifiers and other discrete variables.

[0099] The power distribution equipment operation state code is associated and mapped with the historical load baseline data through the rule engine to generate a dynamic constraint rule group. The rule group establishes the corresponding relationship between the environment temperature threshold range, the load deviation degree permission interval and the device state permission condition, forming a quantifiable physical constraint logic chain. The rule parameters are dynamically adjusted according to the device type, adapting to the verification needs of different power distribution environments.

[0100] Performing multi-parameter joint detection based on dynamic constraint rule set: calculating the relative deviation degree of instantaneous load value and historical load baseline data; comparing whether the environmental temperature value is within the preset permitted range; verifying whether the device running state code conforms to the normal running identifier. The three detections need to meet the preset legality condition at the same time, which is defined as the temperature value being within the threshold range, the load deviation degree not exceeding the permitted interval, and the device state code having no abnormal identifier.

[0101] When the abnormal suspicion point correlation data completely meets the legality condition, the abnormal state mark of the point in the suspicion point set is removed. The mark removal operation updates the suspicion point state attribute, and records the environmental temperature value, load deviation degree and other basis parameters of the removal decision. The suspicion points that do not meet the condition retain the original abnormal mark and confidence score.

[0102] Finally, all abnormal suspicion points that are not removed are summarized, and a verification abnormal event list is generated in time and space dimensions. The list includes event timestamp, geographic coordinates, abnormal classification code and verification decision log. The verification abnormal event list is transmitted to the closed-loop strategy engine module through a data interface, triggering an incremental update operation of the marketing inspection strategy library.

[0103] The application constructs a traceable verification logic chain: real-time data and historical baseline provide verification basis, a dynamic rule engine generates computable constraint conditions, multi-parameter joint detection realizes cross verification, and a mark update mechanism guarantees state traceability. The verification result output structure adapts to the downstream strategy update demand, forming a complete abnormal event decision closed loop.

[0104] Specifically, the power marketing big data driven abnormal mode data processing system provided by the application, the closed-loop strategy engine module is configured to:

[0105] Based on the verification passed samples of the newly imported historical feature library, the phase angle of the periodic feature component is recalculated and the corresponding hash index structure is updated; according to the change amplitude of the recalculated periodic component phase angle, a loss function weight adjustment instruction of the adversarial feature decoupling module is generated by triggering an incremental learning mechanism; the adjustment instruction is executed to update the loss function weight parameters of the adversarial feature decoupling module; the updated periodic component phase angle, hash index structure and loss function weight parameters are synchronized to the multi-modal feature reconstruction module and adversarial feature decoupling module, and cross-module parameter collaborative refreshing is completed.

[0106] The closed-loop strategy engine module first receives the verification passed samples output by the behavior chain verification module, and incrementally imports them into the historical feature library storage. Based on the newly imported sample set, the phase angle parameters of the load periodic feature component are recalculated, and a frequency spectrum analysis algorithm is used to identify the feature phase shift. The hash index structure is updated synchronously, the feature index is reconstructed by an improved local sensitive hashing method, and the retrieval efficiency of similar samples is improved.

[0107] The recalculated periodic component phase angle is compared with the historical reference value, and the phase angle change amplitude is quantitatively calculated. When the change amplitude exceeds the preset sensitivity threshold, the incremental learning mechanism is triggered to generate a weight adjustment instruction. The instruction contains the weight adjustment coefficient of the fluctuation feature reconstruction loss function and the noise discrimination loss function in the adversarial feature decoupling module, and the adjustment coefficient dynamically generates positive and negative adjustment amounts according to the phase offset direction.

[0108] The loss function weight parameters of the adversarial feature decoupling module are updated when the adjustment instruction is executed, and the weight proportion of the fluctuation feature reconstruction loss term and the noise discrimination loss term in the neural network loss layer is adjusted through the parameter configuration interface. The weight updating process adopts the gradient smoothing algorithm to avoid parameter mutation and maintain the stability of model training.

[0109] The updated periodic component phase angle parameters, the optimized hash index structure and the adjusted loss function weight parameters are packaged as a parameter update package. Through the cross-module communication protocol, it is synchronously transmitted to the multi-modal feature reconstruction module and the adversarial feature decoupling module. The multi-modal feature reconstruction module receives the updated periodic component phase angle for correcting the feature extraction reference, and the hash index structure is used to optimize the feature retrieval process. The adversarial feature decoupling module loads the new loss function weight parameter to update the network structure, and realizes the parameter collaborative refresh of the two core processing modules.

[0110] The application constructs a closed-loop optimization mechanism: sample import updates feature library basic data, phase angle recalculation quantifies feature distribution changes, change amplitude triggers incremental learning decision, weight adjustment optimizes feature decoupling effect, and cross-module synchronization realizes system-level parameter update. Each link is connected through a standardized data interface to form a self-adaptive system optimization closed loop, ensuring continuous evolution of processing capacity.

[0111] Specifically, the power marketing big data driven abnormal mode data processing system provided by the application further comprises:

[0112] The noise confidence index is generated by the adversarial feature decoupling module and input into the dynamic algorithm adaptation module;

[0113] The dynamic algorithm adaptation module executes the dynamic switching of the optimization algorithm based on the noise confidence, including: presetting a noise confidence index threshold interval;

[0114] When the noise confidence index falls into the low noise interval, the isolated forest algorithm is activated to perform core suspect point identification; when the noise confidence index falls into the medium noise interval, the weighted distance measurement algorithm is activated to perform anti-interference anomaly detection;

[0115] When the noise confidence index falls into the high noise interval, the sparse self-encoding clustering algorithm is activated to perform feature dimension reduction and density clustering.

[0116] The selected optimization algorithm processes the fluctuation feature vector set according to the current input feature, and generates a set of abnormal suspicion points.

[0117] The noise confidence index generated by the anti-feature decoupling module is input into the dynamic algorithm adaptation module as a key control parameter. The index is a continuous numerical variable, quantitatively representing the severity of noise interference on the data. The dynamic algorithm adaptation module presets a discrete threshold interval for the noise confidence index, mapping the continuous value to three decision domains: low noise interval, medium noise interval, and high noise interval.

[0118] When the noise confidence index is in the low noise interval, the isolation forest algorithm processing procedure is activated. This algorithm is based on the random forest architecture to build isolated trees, and uses the path length principle to identify core abnormal suspicion points, and is suitable for scenarios with high data quality. The algorithm maintains the default tree depth parameter during execution, focusing on the identification of essential abnormal patterns.

[0119] When the noise confidence index enters the medium noise interval, the weighted distance measurement algorithm processing procedure is called. This algorithm introduces a dimension weighting mechanism, giving key dimensions higher weights through feature importance evaluation, enhancing the algorithm's robustness to local interference. The Mahalanobis distance variant is used in the distance calculation process to effectively suppress the influence of moderate-level noise on anomaly detection.

[0120] When the noise confidence index reaches the high noise interval, the sparse auto-encoding clustering algorithm processing procedure is triggered. This algorithm compresses the feature space through an encoder for dimension reduction, eliminating high-dimensional noise interference, and performs density clustering analysis in the hidden layer feature space. The algorithm prioritizes the robustness of feature representation before performing abnormal cluster detection.

[0121] Each activated optimization algorithm independently processes the input fluctuation feature vector set to generate a set of intermediate abnormal suspicion points. All intermediate results are input into the decision fusion engine, which uses time and space alignment techniques to integrate the output labels of different algorithms. After fusion, a unified set of abnormal suspicion points is generated, including floating radius decision boundary labels and abnormal probability scores.

[0122] The present application establishes a mapping logic between noise level and algorithm characteristics: in a low noise environment, efficient core anomaly recognition algorithms are used, in a medium noise environment, interference-resistant distance measurement algorithms are enabled, and in a high noise scenario, feature dimension reduction is performed first before detection. The noise confidence index serves as an algorithm selector, enabling dynamic adaptation of processing strategies to data quality. The final output is an integrated set of abnormal suspicion points, which is transmitted to the behavior chain verification module through a standardized interface.

[0123] Specifically, the power marketing big data driven abnormal pattern data processing system described in the present application further comprises:

[0124] The behavior chain verification module outputs the normal dynamic fluctuation sample of the verification pass to the closed-loop strategy engine module;

[0125] The closed-loop strategy engine module imports the normal dynamic fluctuation sample of the verification pass into a historical feature library, and updates the periodic feature extraction parameter of the multi-modal feature reconstruction module through an incremental learning mechanism, so that the spectral feature extraction of the multi-modal feature reconstruction module adapts to new sample data, and an operation mode of verification result driving feature reconstruction module parameter updating is established.

[0126] After the multi-dimensional verification of the behavior chain verification module is completed, the normal dynamic fluctuation sample of the verification pass is marked with a time stamp, a spatial coordinate and a load feature vector, and is transmitted to the closed-loop strategy engine module through a data pipeline. The sample contains reasonable fluctuation data verified by the physical constraint logic chain, and serves as an effective input for incremental learning of the system.

[0127] After receiving the sample, the closed-loop strategy engine module performs a historical feature library updating operation, and imports the new sample into a feature library storage partition in an append storage mode. The import process establishes an association index of the sample and a seasonal identification factor, and retains complete space-time dimension information. A feature library version management mechanism records data change states, and provides a data basis for incremental learning.

[0128] Based on the newly imported sample set, an incremental learning process is triggered, and the periodic feature extraction parameter of the multi-modal feature reconstruction module is recalculated. The phase angle parameter of the daily periodic component and the week-month periodic feature is updated through spectral recalibration technology, and the sliding window mean value method is used to smooth the parameter transition. The feature extraction parameter change amount is recorded through a hash digest, and a lightweight parameter update instruction is generated.

[0129] The updated periodic feature extraction parameter is synchronized to the multi-modal feature reconstruction module through a configuration interface, and adjusts the reference parameter of the spectral analysis unit of the multi-modal feature reconstruction module. This adjustment enables the feature reconstruction module to automatically adapt to the features of new sample data in subsequent processing, improves the identification accuracy of the daily periodic component and the week-month periodic feature capture capability. The parameter update adopts a hot deployment mechanism to ensure that the system continues to run without being affected.

[0130] The present application establishes a closed-loop learning mechanism: the verification result provides high-quality learning samples, the feature library incrementally stores and updates the data basis, the parameter recalculation realizes the optimization of the feature extraction model, and the hot deployment completes the dynamic upgrade of the system capability. The verification result and the feature extraction capability form a positive feedback loop, so that the system continuously adapts to the evolution of the power behavior mode, and improves the environmental adaptability of the abnormal detection.

[0131] Specifically, the power marketing big data driven abnormal mode data processing system disclosed by the present application adjusts the loss function weight of the adversarial feature decoupling module according to the decision boundary offset, which includes:

[0132] monitoring the decision boundary offset marked by the set of suspicious points generated by the dynamic algorithm adaptation module in real time;

[0133] when the decision boundary offset falls into a preset high offset interval, reducing the weight coefficient of the fluctuation feature reconstruction loss;

[0134] when the decision boundary offset falls into a preset low offset interval, increasing the target weight of the noise discrimination loss;

[0135] The dynamically adjusted loss function weight is synchronized to the output layer executor of the adversarial feature decoupling module.

[0136] The closed-loop strategy engine module receives the set of suspicious points output by the dynamic algorithm adaptation module in real time, and continuously monitors the decision boundary offset parameter marked in the set. The offset represents the distance change of the abnormal judgment boundary and the normal feature distribution, and the offset trend value is obtained through sliding window mean calculation. The monitoring process uses time series analysis technology to capture the dynamic evolution characteristics of the offset.

[0137] The discrete interval of the preset decision boundary offset is divided, and two key decision domains, high offset interval and low offset interval, are set. When the offset trend value is continuously in the high offset interval, a weight down instruction of the fluctuation feature reconstruction loss is generated. The weight down operation reduces the coefficient weight of the feature reconstruction term in the loss function, reduces the constraint strength of the model on the feature fidelity, and enhances the adaptability of the model to the distribution offset.

[0138] When the offset trend value is continuously in the low offset interval, a target weight up instruction of the noise discrimination loss is generated. The weight up operation increases the coefficient weight of the noise discrimination term in the loss function, strengthens the identification ability of the model to the noise feature, and improves the robustness of feature decoupling. The weight adjustment amount is dynamically calculated according to the magnitude of the offset deviation from the reference value.

[0139] The dynamically adjusted loss function weight parameter is synchronized to the output layer executor of the adversarial feature decoupling module through the configuration interface. The executor injects the new weight parameter into the neural network loss calculation layer, and updates the weight proportion of the fluctuation feature reconstruction loss and the noise discrimination loss in real time. Gradient smoothing technology is used for parameter update to avoid training shock and maintain the stability of model convergence.

[0140] The present application constructs a dynamic feedback control mechanism: boundary offset monitoring provides system state awareness, interval division establishes decision criteria, weight adjustment realizes model behavior control, and parameter synchronization completes closed-loop regulation. Through adaptive adjustment of the loss function weight, the feature decoupling module continuously optimizes the adaptability to the change of power data distribution, and guarantees the persistence of abnormal detection accuracy.

[0141] The application constructs a three-dimensional data stream of space-time alignment through a distributed cooperative acquisition module, and eliminates time dislocation noise of multi-source heterogeneous data; a multi-modal feature reconstruction module applies fast Fourier transform to separate periodic background noise and external interference factors such as environmental response modeling and quantization of temperature; an antagonistic feature decoupling module uses orthogonal projection technology to decompose the load feature spectrum into mutually exclusive fluctuation feature subspace and noise feature subspace, and generates a 64-dimensional purified feature vector set and a noise confidence index. This technical path realizes physical isolation and quantitative evaluation of noise components, and reduces the interference of high-dimensional noise on anomaly detection from the feature representation level.

[0142] A dynamic algorithm adaptation module dynamically schedules an optimization algorithm based on the noise confidence index: in a low-noise environment, an isolation forest algorithm is activated to capture core abnormal patterns, in a medium-noise scenario, a weighted distance measurement algorithm is used to enhance the anti-local interference ability, and in a high-noise condition, feature dimension reduction clustering is performed to suppress the influence of noise; a behavior chain verification module establishes joint constraint rules of environmental temperature threshold, load deviation degree permission interval and device state code, and removes the abnormal label when the three simultaneously satisfy the preset physical legality condition. This scheme realizes the accurate distinction between real abnormal events and reasonable environmental responses through dynamic algorithm adaptation and multi-dimensional physical rule verification.

[0143] A closed-loop strategy engine module incrementally imports normal samples that pass the verification into a historical feature library, triggers a spectrum recalibration to update periodic feature phase angle parameters; based on the decision boundary offset output by the dynamic algorithm adaptation module, the loss function weight coefficient of the antagonistic feature decoupling module is adjusted: in a high offset interval, the feature reconstruction loss weight is reduced to enhance the model adaptability, and in a low offset interval, the noise discrimination loss weight is increased to strengthen the feature decoupling ability; the updated periodic feature parameters, hash index and loss weight are fed back to the feature reconstruction module and the decoupling module, forming a closed-loop learning system of feature extraction, anomaly detection, result feedback and parameter optimization, continuously improving the recognition robustness in a complex noise environment.

Claims

1. A data processing system for abnormal patterns driven by big data in electricity marketing, characterized in that: include: The distributed collaborative acquisition module collects power time-series data and environmental parameters from smart meters, and outputs a three-dimensional data stream including time, physical quantities, and spatial dimensions. The multimodal feature reconstruction module receives the three-dimensional data stream, applies time window segmentation to extract spectral features and environmental response features, and generates a load feature map; The adversarial feature decoupling module receives the load feature map, performs orthogonal projection of the feature space to separate the fluctuation feature vector and the noise confidence index, and outputs the purified fluctuation feature vector set. The dynamic algorithm adaptation module receives the set of fluctuation feature vectors and the noise confidence index, and dynamically activates and optimizes the algorithm combination according to the feature sparsity and noise confidence to generate a set of suspected abnormal points. The behavior chain verification module receives the set of suspected abnormal points, performs multi-dimensional verification to distinguish between abnormal events and reasonable fluctuations, and outputs a list of verified abnormal events and samples of normal dynamic fluctuations that have passed verification. The closed-loop strategy engine module receives the list of verified abnormal events and the verified normal dynamic fluctuation samples. It triggers marketing audit instructions for events in the list of verified abnormal events, handles real electricity consumption anomalies, imports the verified normal dynamic fluctuation samples into the historical feature library, adjusts the periodic feature extraction parameters of the multimodal feature reconstruction module through an incremental learning mechanism, and adjusts the loss function weight of the adversarial feature decoupling module according to the decision boundary offset.

2. The abnormal pattern data processing system driven by big data in power marketing as described in claim 1, characterized in that, The distributed collaborative acquisition module is configured as follows: Terminal devices are deployed at power grid nodes to periodically collect instantaneous values ​​of voltage and current from smart meters, and to obtain meteorological service parameters and the geographical coordinates of the power grid. The three-dimensional data stream includes timestamps, meter identifiers, power values, temperature values, humidity values, longitude values, and latitude values; The periodically collected instantaneous values ​​of voltage and current from the smart meter, along with the acquired meteorological service parameters and power grid geographic coordinates, are input into the multimodal feature reconstruction module for segmentation.

3. The abnormal pattern data processing system driven by big data in power marketing as described in claim 2, characterized in that, The multimodal feature reconstruction module is configured as follows: Spectral feature extraction is performed on the input 3D data stream: Fast Fourier Transform is applied to analyze the daily periodic components and weekly / monthly periodic features to generate a standardized periodic feature spectrum; Environmental response feature extraction: Construct a nonlinear correlation model between load value and temperature value, and output a set of temperature sensitivity coefficients; By fusing the standardized periodic characteristic spectrum with the temperature sensitivity coefficient set, a basic load characteristic map is generated; Identify seasonal dependence patterns in the basic load feature map and add seasonality identification factors to the map; The load feature map with seasonality identifier is input into the adversarial feature decoupling module to perform orthogonal projection processing.

4. The abnormal pattern data processing system driven by big data in electricity marketing as described in claim 3, characterized in that, The adversarial feature decoupling module is configured as follows: The encoder decomposes the input load feature map into a fluctuation feature subspace and a noise feature subspace. The generator reconstructs the feature samples, and the discriminator compares the reconstructed samples with the original samples to calculate the error magnitude. Based on the error amplitude, a set of fluctuation feature vectors and a noise confidence index are generated, wherein the set of fluctuation feature vectors contains 64-dimensional purification features and the noise confidence index is a continuous value. The set of fluctuation feature vectors and the noise confidence index are transmitted as output to the dynamic algorithm adaptation module.

5. The abnormal pattern data processing system driven by big data in electricity marketing as described in claim 4, characterized in that, The dynamic algorithm adaptation module is configured as follows: Real-time monitoring of input feature sparsity, kurtosis, and noise confidence index; If the feature sparsity exceeds the set threshold, the sparse autoencoder clustering algorithm is called to process the features. If the kurtosis of the feature distribution exceeds a set threshold, the weighted distance metric algorithm is invoked to process the feature. If the noise confidence index reaches the preset threshold, the detection is performed after adjusting the depth parameters of the isolated forest algorithm. Integrate all the suspicious points output by the algorithms to generate a set of anomalous suspicious points that mark the decision boundary of the floating radius. Output the set of anomalous suspicious points marked with the decision boundary to the behavior chain verification module to perform multi-dimensional verification.

6. The abnormal pattern data processing system driven by big data in power marketing as described in claim 5, characterized in that, The behavior chain verification module is configured as follows: The system obtains the current ambient temperature value, the instantaneous load value of the power grid, and the operating status code of the power distribution equipment from the real-time monitoring source. It retrieves the historical load baseline data of the target user, analyzes the operating status code of the power distribution equipment together with the historical load baseline data, generates a dynamic constraint rule group, and detects the deviation of the ambient temperature value, the instantaneous load value and the historical load baseline data, as well as whether the operating status code of the power distribution equipment meets the preset legality conditions based on the dynamic constraint rule group. If all conditions are met, the corresponding suspicious point will be removed from the list. The system compiles a list of suspected anomalies that have not been de-marked, generates a list of verification anomalies, and outputs the list of verification anomalies to the closed-loop strategy engine module to drive the update of the marketing audit strategy library.

7. The abnormal pattern data processing system driven by big data in electricity marketing as described in claim 6, characterized in that, The closed-loop strategy engine module is configured as follows: Based on the newly imported historical feature library, the verification samples are used to recalculate the phase angle of the periodic feature components and update the corresponding hash index structure. Based on the recalculated phase angle change amplitude of the periodic component, the incremental learning mechanism is triggered to generate a loss function weight adjustment instruction for the adversarial feature decoupling module; The adjustment instruction is executed to update the weight parameters of the loss function of the adversarial feature decoupling module; The updated periodic component phase angle, hash index structure, and loss function weight parameters are synchronized to the multimodal feature reconstruction module and the adversarial feature decoupling module to complete cross-module parameter collaborative refresh.

8. The abnormal pattern data processing system driven by big data in electricity marketing as described in claim 7, characterized in that, Also includes: The noise confidence index is generated by the adversarial feature decoupling module and input into the dynamic algorithm adaptation module; The dynamic algorithm adaptation module performs dynamic switching of optimization algorithms based on noise confidence, including: preset noise confidence index threshold range; When the noise confidence index falls into the low noise range, the isolated forest algorithm is activated to perform core suspect identification. When the noise confidence index falls into the medium noise range, the weighted distance metric algorithm is activated to perform anti-interference anomaly detection. When the noise confidence index falls into the high noise range, the sparse autoencoder clustering algorithm is activated to perform feature dimensionality reduction and density clustering. The selected optimization algorithm processes the set of fluctuating feature vectors based on the current input features to generate a set of suspected anomalies.

9. The abnormal pattern data processing system driven by big data in electricity marketing as described in claim 8, characterized in that, Also includes: The behavior chain verification module outputs the verified normal dynamic fluctuation sample to the closed-loop strategy engine module; The closed-loop strategy engine module imports the verified normal dynamic fluctuation samples into the historical feature library and updates the periodic feature extraction parameters of the multimodal feature reconstruction module through an incremental learning mechanism, so that the spectral feature extraction of the multimodal feature reconstruction module adapts to the new sample data, and establishes an operation mode in which the verification results drive the parameter update of the feature reconstruction module.

10. The abnormal pattern data processing system driven by big data in power marketing as described in claim 9, wherein adjusting the loss function weights of the adversarial feature decoupling module according to the decision boundary offset further includes: Real-time monitoring of the decision boundary offset of the abnormal suspect point set marked by the dynamic algorithm adaptation module; When the decision boundary offset falls into the preset high offset range, the weight coefficient of the fluctuation feature reconstruction loss is reduced. When the decision boundary offset falls into the preset low offset range, the target weight of the noise discrimination loss is increased; The dynamically adjusted loss function weights are synchronized to the output layer actuator of the adversarial feature decoupling module.

Images