Multi-terminal Cooperative Correction System for Power Distribution Units Based on Adversarial Learning

Abstract

This invention belongs to the field of power distribution optimization, and particularly relates to a multi-terminal collaborative correction system for power distribution units based on adversarial learning. The system includes an acquisition module that achieves differentiated data acquisition and anomaly identification by strengthening the acquisition adjustment model; a preprocessing unit that uses a distributed filtering model under a federated framework for noise reduction and sparsification; and a self-calibration module that predicts temperature changes using a distributed LS-SVR generator, predicts voltage and current drift using a particle swarm optimization BP neural network discriminator, and uses a comprehensive adversarial loss function for adversarial training to dynamically adjust the real-time demand allocation coefficient so that its deviation from the true value meets a threshold. This application significantly improves acquisition efficiency, correction accuracy, and adaptability to complex load dynamic changes.

Description

Technical Field

[0001] This invention belongs to the field of power distribution optimization, and particularly relates to a multi-terminal collaborative correction system for power distribution units based on adversarial learning. Background Technology

[0002] In applications where 5G communication and smart power systems are deeply integrated, the communication power distribution unit needs to monitor the status through voltage, current, and temperature sensors and rely on self-calibration algorithms to ensure data accuracy in order to meet the stringent requirements of power output stability, response speed, and energy conversion efficiency. However, under complex operating conditions, sensor data is susceptible to multi-source interference, especially temperature drift, which causes the measured value to deviate from the true value due to temperature differences within the module. Multi-source noise, such as high-frequency switching actions, couples to the sensing circuit, damaging signal integrity and affecting the accurate distribution and adjustment during the power distribution process. Therefore, how to ensure the real-time performance and accuracy of sensor data under multi-source interference, and how to separate switching noise from the effective signal so that the response delay and distribution accuracy of power distribution meet the target requirements, has become a key research focus. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention proposes a multi-terminal collaborative correction system for power distribution units based on adversarial learning. This system includes a data acquisition module that uses an enhanced acquisition adjustment model to achieve differentiated data acquisition and anomaly identification; a preprocessing unit that employs a distributed filtering model under a federated framework for noise reduction and sparsification; and a self-calibration module that predicts temperature changes using a distributed LS-SVR generator, predicts voltage and current drift using a particle swarm optimization-optimized BP neural network discriminator, and uses a comprehensive adversarial loss function for adversarial training to dynamically adjust the real-time demand allocation coefficient, ensuring that its deviation from the true value meets a threshold. This application effectively solves the allocation accuracy problem under temperature drift and multi-source noise interference by combining an enhanced acquisition adjustment model with a generative adversarial algorithm, ensuring that power allocation accurately matches load demand.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A multi-terminal collaborative correction system for power distribution units based on adversarial learning includes: an acquisition module; the acquisition module includes an acquisition unit and a preprocessing unit.

[0006] The acquisition unit is used to perform differentiated acquisition and anomaly identification on each target distribution sub-terminal and distribution terminal of the power distribution unit according to the preset enhanced acquisition adjustment model, and obtain multi-source acquisition data corresponding to the distribution terminal and the target distribution sub-terminal with anomalies. The enhanced acquisition adjustment model is constructed by combining the historical anomaly frequency ratio of each target distribution sub-terminal, the duration of the corresponding anomaly type, and the actual demand allocation coefficient of each target distribution sub-terminal with the reinforcement learning algorithm. The actual demand allocation coefficient is constructed by the ratio of the actual load demand corresponding to all target distribution sub-terminals.

[0007] The preprocessing unit is used to perform differentiated filtering, noise reduction, and sparsification processing on the multi-source collected data based on the anomalies identified by each target allocation sub-end, and to call the distributed filtering model constructed according to the federated algorithm framework and filtering algorithm to obtain the preprocessed multi-source correction parameters and sparsity error loss. The sparsification processing is implemented based on the contribution of each variable in the multi-source collected data to the correction of the real demand allocation coefficient under the anomaly type corresponding to each target allocation sub-end.

[0008] Specifically, the multi-source acquisition data includes the total voltage, current, and first temperature data corresponding to the distribution end, as well as the actual demand distribution coefficient, and the measured distribution voltage, current, and second temperature offset information sequence of the corresponding target distribution sub-end under the corresponding abnormal type for each target distribution sub-end; the power distribution unit multi-end collaborative correction system also includes a self-calibration module; the self-calibration module includes an error prediction unit;

[0009] The error prediction unit is used to obtain the predicted temperature change of each target sub-terminal to the distribution terminal and the average temperature change of the distribution terminal based on the offset between the real-time voltage measured at each target sub-terminal in the multi-source correction parameters and the corresponding actual demand voltage, combined with the generator constructed by the preset distributed LS-SVR model. It also obtains the temperature change prediction loss based on the first temperature data and the average temperature change.

[0010] Specifically, the self-calibration module also includes a distribution offset prediction unit; the distribution offset prediction unit obtains the predicted voltage drift and current drift for each target distribution sub-terminal based on a preset deviation prediction parameter space combined with a discriminator constructed by a BP neural network optimized by particle swarm optimization; at the same time, it obtains the measured voltage and current offset based on the actual load demand and measured distribution voltage and current information of each target distribution sub-terminal; and obtains the power distribution prediction deviation loss based on the measured voltage and current offset and the predicted voltage drift and current drift.

[0011] The deviation prediction parameter space includes the predicted temperature change, the second temperature offset information sequence, the anomaly type, the average drift coefficient of similar historical anomalies, the line impedance temperature coefficient, the voltage temperature coefficient, and the current temperature coefficient; the anomaly types include overvoltage, undervoltage, and overcurrent.

[0012] Specifically, the self-calibration module also includes an adversary calibration unit;

[0013] The adversarial calibration unit constructs a comprehensive adversarial loss function based on power allocation prediction deviation loss, temperature change prediction loss, and filtering error loss. This comprehensive adversarial loss function is fed back to the discriminator and used to perform adversarial training in combination with a preset adversarial loss function threshold. The adversarial training results that meet the adversarial loss function threshold are then used to adjust the real-time demand allocation coefficients so that the real-time demand allocation coefficients and the actual demand allocation coefficients meet the preset allocation deviation threshold.

[0014] Specifically, the enhanced acquisition adjustment model is constructed by combining the first acquisition adjustment sub-model of one allocation end and the target acquisition adjustment sub-models corresponding to M target allocation sub-ends with the anomaly identification sub-model in a federated framework;

[0015] The implementation process of enhancing the data acquisition and adjustment model includes:

[0016] Obtain the actual demand allocation coefficient and the preset allocation end sampling coefficient. Initialize the data acquisition configuration parameters of the first acquisition adjustment sub-model based on the allocation end sampling coefficient. At the same time, initialize the data acquisition configuration parameters of the corresponding target acquisition adjustment sub-model based on the actual demand allocation coefficient.

[0017] The initial first acquisition adjustment sub-model and the target acquisition adjustment sub-model are used to synchronously acquire data, and the data acquired by the allocation end is analyzed. When the real-time demand allocation coefficient and the first temperature data of the allocation end change, the sampling parameters of the allocation end are adjusted based on the real-time first temperature data and the preset first sampling adjustment function. The first sampling adjustment function is constructed by combining the change in the first temperature data and the sampling configuration parameters of the allocation end with the convolution kernel function.

[0018] Specifically, the implementation process of the enhanced acquisition and adjustment model also includes:

[0019] While analyzing the data collected at the allocation end, the anomaly identification sub-model corresponding to each target allocation sub-end is invoked to extract and identify anomalies in the data collected by each target allocation sub-end within a preset time period. When the change value of the principal contribution variable corresponding to each anomaly type is greater than the preset anomaly threshold within the preset time period, the current target allocation sub-end is determined to be abnormal and the corresponding anomaly type is obtained. The principal contribution variable corresponding to each anomaly type is obtained through principal component algorithm based on the anomaly identification accuracy and the corresponding extracted anomaly variable features.

[0020] Based on the corresponding anomaly type and the preset anomaly type-data acquisition parameter configuration library, combined with a deep search algorithm, the optimal data acquisition configuration parameters corresponding to the target allocation sub-end are obtained. The initialized data acquisition configuration parameters are updated and adjusted. After every N anomaly processing operations, the anomaly type, data acquisition configuration parameters, and recognition effect data are stored in the anomaly type-data acquisition parameter configuration library. The data acquisition configuration parameters of the same type of anomaly are updated through a clustering algorithm.

[0021] Specifically, the data acquisition configuration parameters include sampling frequency, sampling window width, and data resolution; the sampling window width is obtained by combining the duration of the corresponding anomaly type, the anomaly recognition rate, and the sliding window with the simulation algorithm.

[0022] The implementation process of enhancing the acquisition and adjustment model also includes:

[0023] When there are more than two or more target allocation sub-ends with anomalies and the data acquisition configuration parameters are adjusted, and the current resources cannot meet the needs of all the adjusted target allocation sub-ends, a conflict arbitration mechanism is constructed based on the real-time demand allocation coefficient and the degree of harm of the corresponding anomaly to the corresponding target allocation sub-end, and the data acquisition resources of each target allocation sub-end are adjusted.

[0024] Specifically, the process of differentiated filtering noise reduction and sparsity processing includes:

[0025] Based on the federated algorithm framework and the anomaly type features of the target allocation sub-ends, a matching algorithm is used to match a local filtering sub-model that is suitable for the anomaly type of each target allocation sub-end. At the allocation end, a federated averaging algorithm is used to aggregate the filtering parameters of each target allocation sub-end to obtain a distributed filtering model. The anomaly type features of the target allocation sub-ends are associated with the anomaly types identified by the enhanced acquisition and adjustment model.

[0026] Based on the data collected from each target allocation sub-end and allocation end, multi-source collected data after differential filtering and noise reduction is obtained through the corresponding local filtering sub-model and cross-sub-end coupling noise collaborative suppression.

[0027] Specifically, the process of collaborative suppression of cross-sub-end coupling noise is as follows:

[0028] Based on the noise spectrum features extracted from the local filtering sub-model of each target allocation sub-end, the cosine similarity of the noise spectrum features between each target allocation sub-end is calculated using the cosine function to obtain the noise correlation degree.

[0029] A target sub-end group with a noise correlation greater than the correlation threshold is obtained. Based on the maximum noise energy frequency and noise coupling strength within the target sub-end group, the center frequency and bandwidth of the band-stop filter are simultaneously optimized to achieve collaborative suppression of cross-end coupled noise.

[0030] Specifically, the process of differentiated filtering noise reduction and sparsification also includes:

[0031] Based on the multi-source collected data after differential filtering and the actual demand allocation coefficient, the comprehensive contribution of each variable is obtained by calculating the sensitivity of each variable to the correction of the actual demand allocation coefficient, mutual information entropy and effective participation rate.

[0032] Based on the multi-source acquisition data after differential filtering and noise reduction corresponding to each target allocation sub-end, combined with a sparse attention network, and using the comprehensive contribution of all variables as the attention sparse weight, the sparsed multi-source acquisition data is obtained.

[0033] The data sparsity process is evaluated based on the deviation between the real-time demand allocation coefficient and the actual demand allocation coefficient. The evaluation results are then fed back to the sparse attention network to adjust the attention sparsity weights in real time until the deviation between the real-time demand allocation coefficient and the actual demand allocation coefficient meets the corresponding allocation deviation threshold in real time.

[0034] Compared with the prior art, the beneficial effects of the present invention are:

[0035] This invention addresses the shortcomings of existing technologies by strengthening the acquisition and adjustment model and dynamically optimizing acquisition parameters by combining historical anomaly characteristics of the target allocation sub-end with demand allocation coefficients. This achieves differentiated data acquisition, improving anomaly identification efficiency and resource allocation rationality. A distributed filtering model built on a federated framework matches and adapts filtering algorithms to target allocation sub-ends with different anomaly types. Combined with cross-sub-end coupling noise collaborative suppression technology, it accurately reduces noise while protecting data privacy, significantly improving the quality of multi-source acquired data. The self-calibration module integrates a distributed LS-SVR generator and a particle swarm optimized BP neural network discriminator. Through adversarial training, it integrates multi-dimensional loss functions to achieve high-precision prediction and correction of voltage and current drift, effectively reducing demand allocation deviation. A sparsity processing and dynamic weight adjustment mechanism based on variable contribution compresses data volume while ensuring the integrity of key information. The arbitration mechanism in the event of resource conflicts minimizes the overall harm. This application significantly improves the acquisition efficiency, correction accuracy, and operational robustness of the intelligent power allocation unit in multi-sub-end and multi-interference scenarios, and can stably adapt to complex load dynamic changes. Attached Figure Description

[0036] Figure 1 This is a block diagram of the multi-terminal collaborative correction system for power distribution units based on adversarial learning according to the present invention.

[0037] Figure 2 This is a simplified diagram of the distributed power distribution unit of the present invention. Detailed Implementation

[0038] Please see Figure 1 and Figure 2 The present invention provides an embodiment of a multi-terminal collaborative correction system for power distribution units based on adversarial learning, comprising: an acquisition module and a self-calibration module;

[0039] The acquisition module is used to perform differentiated acquisition and anomaly identification on each target allocation sub-end and allocation end according to the enhanced acquisition adjustment model to obtain multi-source acquisition data. Then, the multi-source correction parameters and sparse error loss are obtained through noise reduction and sparsification processing using a distributed filtering model. In this embodiment, the allocation end is a power supply, such as a battery or mains power, and the power allocation unit is used for power allocation and optimization; the target allocation sub-ends are the loads to be powered, and the total number of target allocation sub-ends is M.

[0040] The self-calibration module is used to predict temperature, voltage, and current drift and calculate losses. It adjusts the real-time demand allocation coefficient through adversarial training and feeds back abnormal information to fine-tune the acquisition configuration parameters of the enhanced acquisition model.

[0041] It should be further noted that the acquisition module in this embodiment includes an acquisition unit and a preprocessing unit;

[0042] The acquisition unit is used to perform differentiated acquisition and anomaly identification on each target distribution sub-terminal and distribution terminal of the power distribution unit according to a preset enhanced acquisition adjustment model, and obtain multi-source acquisition data corresponding to the distribution terminal and the target distribution sub-terminal with anomalies. The enhanced acquisition adjustment model is constructed by combining reinforcement learning algorithm with the historical anomaly frequency ratio of each target distribution sub-terminal, the duration of the corresponding anomaly type, and the actual demand allocation coefficient of each target distribution sub-terminal. The multi-source acquisition data includes the total voltage, current, and first temperature data and the actual demand allocation coefficient corresponding to the distribution terminal, as well as the measured distribution voltage, current, and second temperature offset information sequence of the target distribution sub-terminal under the corresponding anomaly type. The actual demand allocation coefficient is constructed by the ratio of the actual load demand corresponding to all target distribution sub-terminals. It should be further noted that the distribution terminal and each target distribution sub-terminal in this embodiment can switch roles. For example, the battery corresponds to the target distribution sub-terminal when charging and the distribution terminal when discharging. The anomaly types include, but are not limited to, overvoltage, undervoltage, and overcurrent.

[0043] It should be further explained that the enhanced acquisition adjustment model in this embodiment is constructed by combining a first acquisition adjustment sub-model of one allocation end and target acquisition adjustment sub-models corresponding to M target allocation sub-ends with an anomaly identification sub-model using a federated framework; it should also be explained that the first acquisition adjustment sub-model and target acquisition adjustment sub-model in this embodiment are constructed by a deep policy network; the anomaly identification sub-model is constructed by a random forest algorithm;

[0044] It should be further explained that the implementation process of the enhanced acquisition adjustment model in this embodiment includes:

[0045] Obtain the actual demand allocation coefficient and the preset allocation end sampling coefficient. Initialize the data acquisition configuration parameters of the first acquisition adjustment sub-model based on the allocation end sampling coefficient. At the same time, initialize the data acquisition configuration parameters of the corresponding target acquisition adjustment sub-model based on the actual demand allocation coefficient.

[0046] The initial first acquisition adjustment sub-model and the target acquisition adjustment sub-model are used to synchronously acquire data, and the data acquired by the allocation end is analyzed. When the real-time demand allocation coefficient or the first temperature data of the allocation end changes, the sampling parameters of the allocation end are adjusted based on the real-time first temperature data and the preset first sampling adjustment function. The first sampling adjustment function is constructed by combining the change in the first temperature data and the sampling configuration parameters of the allocation end with the convolution kernel function.

[0047] It should be further explained that the construction process of the first sampling adjustment function in this embodiment includes:

[0048] Based on the real-time first temperature data collected by the distribution terminal and the historical first temperature data, the change in the first temperature data is obtained by calculating the difference between the two.

[0049] Based on the current sampling frequency, sampling window width, and data resolution preset by the allocation end, a set of sampling configuration parameters containing the above parameters is constructed through parameter integration;

[0050] Based on the rate of change and duration of the first temperature data change, a convolution kernel is designed by combining a Gaussian function and a rectangular function. The size of the convolution kernel is dynamically determined according to the absolute value of the first temperature data change. The larger the absolute value of the first temperature data change, the larger the size of the convolution kernel.

[0051] Based on the change in the first temperature data and the set of sampling configuration parameters, the product of corresponding elements of the change in the first temperature data and the set of sampling configuration parameters is processed by convolution operation to obtain the correlation weight matrix between temperature and sampling parameters.

[0052] Based on the maximum weight value of the correlation weight matrix, the sampling parameter adjustment coefficient is obtained through a linear mapping operation with preset coefficients, where the coefficients of the linear mapping operation are preset fixed values.

[0053] Based on the initial sampling frequency and initial sampling window width of the allocation end, the target sampling frequency is obtained by multiplying the initial sampling frequency by 1 and the sum of the sampling parameter adjustment coefficient. The target sampling window width is obtained by multiplying the initial sampling window width by 1 and the sum of half of the sampling parameter adjustment coefficient. The adjusted target sampling parameters of the allocation end are then obtained.

[0054] Based on preset upper and lower limits of sampling parameters, a clamping operation is used to constrain the range of the target sampling parameters to obtain the final sampling adjustment parameters. The upper and lower limits of the sampling parameters include the minimum and maximum sampling frequencies and the minimum and maximum sampling window widths. It should be further noted that the clamping operation in this embodiment is based on preset upper and lower limits of sampling parameters, and applies a forced range constraint to the dynamically adjusted target sampling parameters: if the target sampling frequency is lower than the lower limit, it is set to the minimum value; if it is higher than the upper limit, it is set to the maximum value. The target sampling window width is subject to the same constraint logic. This operation ensures that the parameters are always within the hardware safety range and the algorithm's effective range, avoiding hardware overload or data overflow caused by over-limit sampling, ensuring that transient signals are captured at the lowest sampling frequency during drastic temperature changes, while suppressing redundant data at the highest sampling frequency under stable operating conditions. Combined with reasonable constraints on the sampling window width, this optimizes the integrity of anomaly capture and resource utilization, providing compliant and stable parameter input for subsequent processing.

[0055] The first sampling adjustment function dynamically integrates temperature change characteristics with sampling configuration parameters to achieve adaptive adjustment of sampling parameters at the allocation end. This enables the sampling frequency and window width to intelligently match temperature fluctuations, ensuring sampling accuracy and improving data integrity during drastic temperature changes, while avoiding redundant sampling during stable temperatures and reducing resource consumption.

[0056] It should be further explained that the process of adjusting the sampling parameters at the allocation end in this embodiment includes:

[0057] Based on the real-time demand allocation coefficient collected by the allocation terminal and the preset demand allocation coefficient, the difference between the two is compared to see if it exceeds the preset threshold. At the same time, based on the real-time first temperature data and the average temperature data of the most recent three sampling periods, the temperature is compared to determine whether there has been a significant change, and an adjustment trigger signal is obtained.

[0058] Based on the adjustment trigger signal, the real-time first temperature data is processed by calling the temperature feature extraction module to obtain the temperature change rate, the duration of temperature change, and the temperature fluctuation amplitude.

[0059] Based on the current sampling parameters of the allocation end, including the current sampling frequency, the current sampling window width, and the current data resolution, a set of sampling parameters is formed through parameter encapsulation;

[0060] Based on the temperature change rate, the duration of temperature change, and the amplitude of temperature fluctuation, the convolution kernel type and size of the first sampling adjustment function are determined through a preset mapping rule, wherein the larger the temperature change rate, the larger the convolution kernel size.

[0061] Based on the set of sampling parameters and temperature feature parameters, the correlation weight matrix between the sampling parameters and temperature features is obtained through the convolution operation of the first sampling adjustment function;

[0062] Based on the correlation weight matrix, the maximum weight value is obtained through a feature extraction algorithm and used as the sampling parameter adjustment coefficient.

[0063] Based on the sampling parameter adjustment coefficients, the current sampling frequency is adjusted using a linear adjustment rule, the current sampling window width is adjusted using a nonlinear adjustment rule, and the current data resolution is adjusted using a step-wise adjustment rule to obtain the initially adjusted sampling parameters;

[0064] Based on the maximum sampling frequency, minimum sampling frequency, maximum sampling window width, minimum sampling window width, and selectable data resolution levels supported by the distribution hardware, the initially adjusted sampling parameters are verified through range constraint processing to obtain the final adjusted sampling parameters.

[0065] Based on the final adjusted sampling parameters, the sampling control register of the allocation end is updated through the configuration interface to complete the real-time update of the sampling parameters, and the adjustment time and parameter values ​​before and after the adjustment are recorded.

[0066] While analyzing the data collected at the allocation end, the anomaly identification sub-model corresponding to each target allocation sub-end is invoked to extract and identify anomalies in the data collected by each target allocation sub-end within a preset time period. When the change value of the principal contribution variable corresponding to each anomaly type is greater than the preset anomaly threshold within the preset time period, the current target allocation sub-end is determined to be abnormal and the corresponding anomaly type is obtained. The principal contribution variable corresponding to each anomaly type is obtained through principal component algorithm based on the anomaly identification accuracy and the corresponding extracted anomaly variable features.

[0067] It should be further explained that the process of obtaining the corresponding exception type in this embodiment includes:

[0068] Based on the typical duration of historical anomaly data statistics, a preset time length is obtained by setting a fixed time interval and the average of the average duration of anomalies, which serves as the time window for anomaly identification.

[0069] Based on common fault modes of power distribution units, an enumeration is used to define a set of abnormal types, including overvoltage, undervoltage, and overcurrent. Each type is associated with a corresponding monitoring variable, such as overvoltage being associated with a voltage variable and overcurrent being associated with a current variable.

[0070] Based on the measured voltage, current and temperature offset information collected by the target sub-terminal within a preset time period, the characteristic sequences of each variable, including the rate of change, peak value and fluctuation amplitude, are extracted through time-frequency domain analysis.

[0071] Based on the historical training data of the anomaly identification sub-model, the contribution weight of each variable feature to the anomaly identification result is calculated, and the contribution weight is positively correlated with the identification accuracy to obtain the importance ranking of the variable features.

[0072] Based on the variable features ranked by importance, dimensionality reduction is performed using the principal component algorithm. The first principal component with a cumulative contribution rate of ≥85% is retained, and its corresponding original variable is determined as the main contributing variable for this anomaly type.

[0073] Based on the initial and final values ​​of the main contributing variable within a preset time period, the change value of the main contributing variable is obtained by calculating the difference between the two.

[0074] Based on the minimum effective change value of the main contributing variable in similar historical anomalies, an anomaly threshold is set by increasing a preset percentage, which serves as the critical value for judging anomalies.

[0075] Based on the change value of the main contribution variable and the preset abnormal threshold, by comparison calculation, if the change value is greater than the abnormal threshold, it is determined that there is an abnormality in the current target allocation sub-terminal, and the type of the abnormality is the abnormality type associated with the main contribution variable. For example, the main contribution variable of voltage corresponds to overvoltage / undervoltage, and the main contribution variable of current corresponds to overcurrent.

[0076] This process uses the principal component algorithm to accurately locate the core influencing variables of each anomaly type. By comparing the change values ​​within a preset time window with the anomaly threshold, it achieves accurate determination of the anomalies and types of the target sub-ends, improves the targeting and accuracy of anomaly identification, and provides a clear basis for subsequent parameter adjustment and correction.

[0077] Based on the corresponding anomaly type and a preset anomaly type-data acquisition parameter configuration library, combined with a deep search algorithm, the optimal data acquisition configuration parameters for the corresponding target allocation sub-end are obtained. The initialized data acquisition configuration parameters are updated and adjusted. After every N anomaly processing operations, the anomaly type, data acquisition configuration parameters, and recognition effect data are stored in the anomaly type-data acquisition parameter configuration library, and the data acquisition configuration parameters for the same type of anomaly are updated using a clustering algorithm. It should be further noted that the data acquisition configuration parameters in this embodiment include sampling frequency, sampling window width, data resolution, and anomaly threshold correction coefficient. It should also be noted that, among the data acquisition configuration parameters, the sampling frequency controls the number of acquisitions per unit time, balancing the capture of rapidly changing signals with resource consumption, improving the real-time anomaly capture capability, and avoiding redundancy. The sampling window width defines the anomaly analysis time range, matches the duration of the anomaly type, and ensures complete capture of anomaly features to improve recognition accuracy. The data resolution determines the acquisition accuracy, captures small parameter offsets, provides support for the correction of the demand allocation coefficient, and reduces invalid redundancy. The anomaly threshold correction coefficient dynamically adjusts the judgment threshold to adapt to the working conditions of the target allocation sub-end, reducing false alarms and missed alarms to improve recognition reliability. These parameters work together to optimize the quality and efficiency of data acquisition, providing accurate data support for anomaly identification and system correction.

[0078] It should be further noted that the sampling window width in this embodiment is obtained by combining the duration of the corresponding anomaly type, the anomaly recognition rate, and the sliding window with the simulation algorithm.

[0079] It should be further explained that the process of implementing the depth search algorithm in this embodiment includes:

[0080] Based on the anomaly type of the target allocation sub-end, the sampling frequency range, sampling window width baseline value, data resolution selectable level and anomaly threshold correction coefficient range corresponding to the anomaly type are obtained by querying the preset anomaly type-data acquisition parameter configuration library, which serves as the initial parameter space for depth search.

[0081] Based on the dimensional characteristics of the parameter space, the parameter space is decomposed into a continuous subspace and a discrete subspace through hierarchical partitioning. A progressive search strategy is adopted, first searching the continuous subspace and then matching the discrete subspace. The dimensional characteristics of the parameter space include that the sampling frequency and sampling window width are continuous values, while the data resolution and the outlier threshold correction coefficient are discrete values.

[0082] Based on the historical duration of anomaly types and the preset anomaly recognition rate target, a candidate set of sampling window widths is generated through a sliding window and simulation algorithm, which is used as a constraint condition for the continuous subspace.

[0083] Based on the initial data acquisition configuration parameters, the step size is adjusted by setting parameters to construct the path generation rules for depth search. It should be further noted that in this embodiment, the sampling frequency step size changes dynamically with the anomaly frequency ratio, and the higher the frequency ratio, the smaller the step size.

[0084] Based on the mapping relationship between data acquisition configuration parameters and anomaly identification results, an evaluation function is constructed by integrating anomaly detection rate, data redundancy rate and false alarm rate, which serves as the basis for value judgment in the search process.

[0085] Based on the evaluation function value, the parameter space is traversed through a depth-first strategy. At the same time, a dynamic pruning mechanism is introduced. When the evaluation function value of a branch is lower than a preset proportion of the current optimal value, the search of that branch is terminated to reduce invalid search paths.

[0086] Based on the parameter combination corresponding to the maximum value of the evaluation function obtained during the search process, the optimal data acquisition configuration parameters are obtained by verifying against hardware performance limitations; hardware performance limitations include, but are not limited to, maximum sampling frequency, resolution support level, etc.

[0087] Based on the difference between the optimal data acquisition configuration parameters and the initialization parameters, the initialization parameters are iteratively adjusted using a parameter smoothing transition algorithm until they completely match the optimal data acquisition configuration parameters; in this embodiment, the parameter smoothing transition algorithm updates gradually at a rate of 5% per iteration.

[0088] This process uses a deep search algorithm with hierarchical search and dynamic pruning to accurately locate the optimal data acquisition configuration parameters by combining anomaly characteristics and hardware constraints. This enables adaptive updates of the data acquisition configuration parameters, reducing data redundancy while ensuring a high anomaly recognition rate and improving the acquisition efficiency and adaptability of the target allocation sub-end under different anomaly scenarios.

[0089] When there are more than two or more target allocation sub-ends with anomalies and the data acquisition configuration parameters are adjusted, and the current resources cannot meet the needs of all the adjusted target allocation sub-ends, a conflict arbitration mechanism is constructed based on the real-time demand allocation coefficient and the degree of harm of the corresponding anomaly to the corresponding target allocation sub-end. The data acquisition resources of each target allocation sub-end are adjusted to minimize the total degree of harm.

[0090] It should be further explained that the detailed implementation process of the conflict arbitration mechanism in this embodiment includes:

[0091] Based on the real-time demand allocation coefficient and anomaly type of each target allocation sub-end, the comprehensive priority weight of each target allocation sub-end is obtained by pre-setting a basic hazard weight for the anomaly type and combining the weighted sum of the real-time demand allocation coefficient and the basic hazard weight.

[0092] It should be further explained that the process of obtaining the comprehensive priority weight in this embodiment includes:

[0093] Based on the anomaly type of the target allocation sub-end, by analyzing the equipment damage rate, downtime and maintenance cost caused by different anomaly types in historical fault data, the basic hazard weight of each anomaly type is obtained through the evaluation algorithm. Among them, the basic hazard weight of the overcurrent anomaly type is higher than that of the overvoltage anomaly type, and the basic hazard weight of the overvoltage anomaly type is higher than that of the undervoltage anomaly type.

[0094] The real-time demand allocation coefficient of the target allocation sub-end is used as the demand influence weight, where the higher the real-time demand allocation coefficient, the greater the demand influence weight.

[0095] Based on the basic hazard weight and the demand impact weight, the comprehensive hazard level value of the target allocation sub-end is obtained through weighted summation, which serves as the comprehensive priority weight of the corresponding target allocation sub-end.

[0096] Based on the adjusted data acquisition configuration parameters of each abnormal sub-end, the resource requirement value of each target allocation sub-end is calculated through a hardware resource consumption model, and the total resource requirement is obtained by summing them up. It should be further explained that this embodiment, based on the data acquisition configuration parameters of the target allocation sub-end, establishes a mapping relationship between parameters and resource consumption by analyzing the correlation data between parameters and hardware resources (bandwidth, storage, and computing units) consumption during historical acquisition processes, thus obtaining a hardware resource consumption model. This model is used to accurately calculate the resource requirements of each abnormal target allocation sub-end, providing a basis for comparing the total resource requirement with the maximum carrying capacity, supporting the triggering of the conflict arbitration mechanism and resource allocation, realizing reasonable resource scheduling, avoiding hardware overload, and ensuring the stability and efficiency of data acquisition. The mapping relationship between parameters and resource consumption is specifically as follows: as the sampling frequency increases, bandwidth consumption increases linearly; as the resolution increases, storage usage increases stepwise.

[0097] The maximum resource capacity value of the allocation end is obtained based on the hardware parameter manual, such as the maximum bandwidth and maximum storage rate of the ADC. By comparing the total resource demand with the maximum resource capacity value, if the total resource demand exceeds the maximum resource capacity value, conflict arbitration is triggered.

[0098] The abnormal target allocation sub-ends are sorted in descending order based on the comprehensive priority weight, and priority levels are divided by setting priority level thresholds.

[0099] Basic resource quotas are allocated based on priority levels, and the initial allocation sum is obtained by accumulating the basic resource quotas.

[0100] Based on the difference between the maximum resource carrying capacity and the initial total allocation, resources are supplemented by additional allocation according to the comprehensive priority weight ratio until available resources are exhausted or all target allocation sub-ends reach the required value.

[0101] Based on the final allocated resource value, the actual data acquisition configuration parameters of each target allocation sub-end are deduced through the resource-parameter mapping rule, and parameters exceeding the allocated resources are compressed and adjusted; the resource-parameter mapping rule is a linear mapping between bandwidth and sampling frequency, and a stepwise mapping between storage and resolution;

[0102] Anomaly identification is performed based on the adjusted data acquisition configuration parameters. The anomaly miss rate of each target allocation sub-end is calculated, which is the ratio of the number of missed detections to the total number of anomalies. If the miss rate of the high-priority target allocation sub-end is >5%, 10% of the resources are compressed from the low-priority target allocation sub-end and supplemented to the high-priority target allocation sub-end to dynamically balance the resource allocation.

[0103] Based on the target allocation sub-end priority, resource allocation results, and anomaly identification effect of each conflict arbitration, a conflict case library is formed through log recording, providing data support for subsequent optimization of basic hazard weights and quota ratios.

[0104] This process dynamically allocates limited resources by comprehensively prioritizing them. While ensuring the accuracy of data collection for high-priority target allocation sub-ends (high demand, high risk), it minimizes the performance loss of low-priority target allocation sub-ends, thereby reducing the overall anomaly risk to a minimum. Furthermore, through dynamic adjustments and case library iterations, it continuously optimizes the rationality and adaptability of resource conflict handling.

[0105] This process utilizes multiple technologies to achieve efficient and precise control of power distribution system data acquisition and anomaly handling. The enhanced acquisition and adjustment model integrates the distribution end and target distribution sub-end models based on a federated framework, combining deep policy networks and random forest algorithms. It achieves differentiated acquisition based on historical anomaly frequency, duration, and demand ratio, making data acquisition more closely aligned with the actual operating conditions of each target distribution sub-end and improving the relevance and effectiveness of the raw data. The first sampling adjustment function dynamically adjusts the sampling frequency and window width of the distribution end by fusing temperature change characteristics with sampling parameters. This ensures sampling accuracy to capture complete features during drastic temperature changes and reduces redundant sampling to lower resource consumption during stable conditions, balancing data quality and system load. Anomaly identification uses principal component analysis to locate core influencing variables. Combined with a preset time length and anomaly threshold comparison, it accurately determines the anomaly type, providing a clear basis for subsequent parameter optimization and improving the relevance of anomaly response. The deep search algorithm accurately locates the optimal acquisition configuration parameters from the parameter space for different anomaly types, ensuring high recognition rates while reducing data redundancy and enhancing the adaptability of the target distribution sub-end to diverse anomaly scenarios. The conflict arbitration mechanism allocates resources based on a comprehensive priority weighting system, combines a hardware resource consumption model to handle resource competition, prioritizes the allocation of high-priority targets to sub-end data collection needs, and continuously optimizes the allocation strategy through dynamic adjustments and case library iterations. This minimizes the overall harm when resources are limited, ensuring stable system operation under complex and concurrent scenarios. These progressively advanced technologies form a closed loop from data collection and anomaly identification to resource allocation, comprehensively improving the system's collection efficiency, anomaly handling accuracy, and resource utilization rationality.

[0106] The preprocessing unit is used to perform differentiated filtering, noise reduction, and sparsification processing on the multi-source collected data based on the anomalies identified by each target allocation sub-end, and to call the distributed filtering model constructed according to the federated algorithm framework and filtering algorithm to obtain the preprocessed multi-source correction parameters and sparsity error loss. The sparsification processing is adjusted according to the contribution of each variable in the multi-source collected data to the correction of the real demand allocation coefficient under the anomaly type corresponding to each target allocation sub-end.

[0107] It should be further explained that the differential filtering noise reduction and sparsification process in this embodiment includes:

[0108] Based on the federated algorithm framework and the anomaly type characteristics of the target allocation sub-ends, a matching algorithm is used to match a local filtering sub-model adapted to its anomaly type for each target allocation sub-end. At the allocation end, a federated averaging algorithm is used to aggregate the filtering parameters of each target allocation sub-end to obtain a distributed filtering model. The anomaly type characteristics of the target allocation sub-ends are associated with the anomaly types identified by the enhanced acquisition and adjustment model. It should be further noted that the local filtering sub-models here are constructed from different filtering algorithms. This embodiment analyzes the historical anomaly sampling information of each target allocation sub-end to obtain the corresponding anomaly type, and then matches it with the established anomaly type-filtering algorithm information database based on past anomaly information to obtain the local filtering sub-model constructed by the filtering algorithm corresponding to the current time of each target allocation sub-end. The local filtering sub-models include, but are not limited to, Gaussian low-pass filtering, Kalman filtering, etc.

[0109] Based on the data collected from each target allocation sub-end and allocation end, multi-source collected data after differential filtering and noise reduction is obtained through the corresponding local filtering sub-model and cross-sub-end coupling noise collaborative suppression.

[0110] It should be further explained that the process of cooperative suppression of cross-sub-end coupling noise in this embodiment is as follows:

[0111] Based on the noise spectrum features extracted from the local filtering sub-model of each target allocation sub-end, the cosine similarity of the noise spectrum features between each target allocation sub-end is calculated using the cosine function to obtain the noise correlation degree.

[0112] A target sub-end group with a noise correlation greater than the correlation threshold is obtained. Based on the maximum noise energy frequency and noise coupling strength within the target sub-end group, the center frequency and bandwidth of the band-stop filter are simultaneously optimized to achieve collaborative suppression of cross-sub-end coupled noise.

[0113] It should be further explained that the detailed process of achieving cooperative suppression of cross-sub-end coupling noise in this embodiment includes:

[0114] Step 1: Based on the noise spectrum features extracted from the local filter sub-model of each target allocation sub-end, calculate the cosine similarity of the noise spectrum features between the target allocation sub-ends using the cosine function to obtain the noise correlation degree, and classify the target allocation sub-ends with a correlation degree greater than the preset correlation degree threshold into the same cooperative suppression group.

[0115] Step 2: Based on the noise spectrum characteristics of each target sub-end in the collaborative suppression group, calculate the noise energy value at each frequency point using the energy integration algorithm, compare the noise energy values ​​of all sub-ends in the group at the same frequency point, and obtain the frequency corresponding to the maximum noise energy value in the group as the center frequency optimization benchmark.

[0116] Step 3: Based on the preset noise injection scheme within the cooperative suppression group, specifically: inject a noise signal of a preset frequency into any target sub-end, measure the noise response amplitude at that frequency at other target sub-ends within the cooperative suppression group using a spectrum analyzer, calculate and normalize the ratio of the noise response amplitude to the injected amplitude, and obtain the noise coupling strength within the group.

[0117] Step 4: Based on the frequency corresponding to the maximum noise energy value within the group, and combined with the noise energy proportion of each target sub-terminal within the collaborative suppression group at that frequency point, obtain the target center frequency of the band-stop filter through weighted summation, where the weight is the energy proportion of each target sub-terminal; the noise energy proportion is: the ratio of the energy of a target sub-terminal at that frequency to the total energy within the group;

[0118] Step 5: Based on the noise coupling strength, set the base bandwidth, and adjust the base bandwidth through a linear mapping rule to obtain the target bandwidth; the linear mapping rule is: for each increase in coupling strength, the bandwidth increases by a preset value;

[0119] Step 6: Based on the target center frequency and target bandwidth, send update commands to the band-stop filters of all target allocation sub-ends in the cooperative suppression group through the parameter synchronization protocol to uniformly update the filter parameters;

[0120] Step 7: Based on the noise monitoring data of each target sub-end in the group after parameter update, evaluate the suppression effect by calculating the noise energy attenuation rate (the ratio of the difference in noise energy before and after adjustment to the energy before adjustment). If the attenuation rate is lower than the preset standard, repeat steps 4 to 6 to re-optimize the parameters.

[0121] Based on the center frequency, bandwidth, and corresponding suppression effect of each optimization, a parameter optimization case library is formed through log recording, providing a reference for parameter initialization in similar noise scenarios.

[0122] This process optimizes the center frequency and bandwidth of the band-stop filter synchronously, enabling the filter to be precisely aligned with the dominant noise frequency within the group and to adapt to the coupling strength. This achieves coordinated suppression of cross-target sub-end coupled noise, significantly reducing noise interference to multi-target sub-end data acquisition and improving the signal quality and operational stability of the power distribution system.

[0123] Based on the multi-source collected data after differential filtering and the actual demand allocation coefficient, the comprehensive contribution of each variable is obtained by calculating the sensitivity of each variable to the correction of the actual demand allocation coefficient, mutual information entropy and effective participation rate.

[0124] It should be further explained that the sensitivity in this embodiment is based on the historical value sequence of each variable in the multi-source acquisition data after differential filtering and the corresponding demand allocation coefficient correction value sequence. By fixing other variables and changing only the value of the target variable, the ratio of the change in the demand allocation coefficient correction value to the change in the target variable is calculated to obtain the sensitivity of the variable to the demand allocation coefficient correction.

[0125] Mutual information entropy is based on the sequence of variable values ​​and the sequence of correction values ​​of demand allocation coefficients. By statistically analyzing the joint probability distribution and their respective marginal probability distributions, the mutual information value of the two is calculated using the information entropy formula. It reflects the degree of nonlinear correlation between the variable and the correction result, and is normalized as the mutual information entropy.

[0126] The effective participation rate is based on the number of times a variable is included in the correction model during the historical correction process. It is obtained by calculating the ratio of the number of times the variable actually participates in the correction operation to the total number of corrections. The higher the ratio, the more often the variable is relied upon in the actual correction.

[0127] This embodiment accurately identifies key variables for correction through multi-dimensional evaluation, avoiding evaluation bias caused by a single indicator. This allows subsequent sparsification processing based on comprehensive contribution to retain core information while eliminating redundant variables, ensuring the accuracy of demand allocation coefficient correction while compressing data volume and improving system data processing efficiency.

[0128] Based on the multi-source acquisition data after differential filtering and noise reduction corresponding to each target allocation sub-end, combined with a sparse attention network, and using the comprehensive contribution of all variables as the attention sparse weight, the sparsed multi-source acquisition data is obtained.

[0129] The data sparsity process is evaluated based on the deviation between the real-time demand allocation coefficient and the actual demand allocation coefficient. The evaluation results are fed back to the sparse attention network to adjust the attention sparsity weights in real time until the deviation between the real-time demand allocation coefficient and the actual demand allocation coefficient meets the corresponding allocation deviation threshold in real time.

[0130] It should be further explained that the specific process of implementing the sparse attention network in this embodiment includes:

[0131] Step 1: Based on the multi-source acquisition data after differential filtering and noise reduction at each target allocation sub-end, convert it into a three-dimensional tensor through dimensional expansion, and use it as the input feature of the sparse attention network;

[0132] Step 2: Based on the comprehensive contribution of all variables, normalize them to the interval between 0 and 1 to obtain the initial attention sparse weight matrix, where the dimension matches the variable dimension of the input feature.

[0133] Step 3: Based on the multi-head attention mechanism of sparse attention network, the input features are multiplied by the initial attention sparse weight matrix to obtain the attention weighted features of each variable; by setting a dynamic sparsity threshold, the attention weighted features are masked, where the features of variables with weights lower than the sparsity threshold are set to zero to obtain the initial sparsified features.

[0134] Step 4: Based on the initial sparsity features, feature fusion is performed through the feedforward neural network layer of the network to output the sparsed multi-source acquisition data; the feedforward neural network layer includes linear transformation and activation function;

[0135] Step 5: Based on the real-time demand allocation coefficient and the actual demand allocation coefficient, obtain the deviation change value by calculating the absolute value of the difference between the two; compare the deviation change value with the preset allocation deviation threshold. If the deviation change value exceeds the allocation deviation threshold, a positive adjustment signal is generated; otherwise, a negative adjustment signal is generated.

[0136] Step 6: Based on the adjustment signal, update the initial attention sparse weight matrix of the sparse attention network using the gradient descent algorithm. The positive adjustment signal strengthens the weight of high-contribution variables, while the negative adjustment signal weakens the weight of low-contribution variables. At the same time, the weight update amplitude is constrained, and a single update does not exceed the preset proportion of the current weight.

[0137] Step 7: Based on the updated attention sparse weight matrix, repeat steps 3 to 6 for the next round of sparsification and evaluation adjustment until the deviation between the real-time demand allocation coefficient and the actual demand allocation coefficient continuously meets the allocation deviation threshold.

[0138] Based on the sparsification results and deviation changes of the attention weight matrix in each round, a weight adjustment case library is formed through log recording, providing a reference for attention weight initialization in similar load scenarios.

[0139] This process uses the overall contribution as the attention weight and combines it with a dynamic adjustment mechanism to enable the sparse attention network to accurately retain the key variable features for correcting the demand allocation coefficient, eliminate redundant information, and ensure correction accuracy while significantly compressing the amount of data. By optimizing the weights in real time through deviation feedback, it ensures that the sparsity processing always adapts to dynamic changes in the load, thereby improving the system's data processing efficiency and correction stability.

[0140] This embodiment's distributed filtering model uses a federated framework to match local filtering sub-models adapted to different anomaly types for each sub-end. Combined with a federated averaging algorithm to aggregate parameters, it achieves differentiated noise reduction, making the filtering more closely match the anomaly characteristics of the target allocation sub-end. Cross-sub-end coupling noise collaborative suppression divides the system into collaborative suppression groups through noise correlation analysis, simultaneously optimizing the center frequency and bandwidth of the band-stop filter to accurately suppress dominant noise within the group, significantly reducing signal interference from multiple target allocation sub-ends and improving data purity. Comprehensive contribution calculation evaluates variables from multiple dimensions, including sensitivity, mutual information entropy, and effective participation rate, avoiding bias from single indicators and accurately identifying core variables. A sparse attention network uses comprehensive contribution as initial weights, removes redundant features through dynamic threshold masks, and integrates core information using a feedforward neural network. Simultaneously, based on the deviation feedback between real-time and actual demand allocation coefficients, gradient descent dynamically adjusts the weights, ensuring that high-contribution variables are prioritized. This significantly compresses the data volume while consistently meeting the correction threshold. These technologies form a closed loop, improving data processing efficiency and ensuring the accuracy and stability of demand allocation coefficient correction, providing high-quality data support for the precise control of power distribution systems.

[0141] It should be further noted that the self-calibration module in this embodiment includes an error prediction unit, an allocation offset prediction unit, an adversary calibration unit, and a feedback adjustment unit.

[0142] The error prediction unit is used to obtain the predicted temperature change of each target sub-terminal to the distribution terminal and the average temperature change of the distribution terminal based on the offset between the real-time voltage measured at each target sub-terminal in the multi-source correction parameters and the corresponding actual demand voltage, combined with the generator constructed by the preset distributed LS-SVR model. It also obtains the temperature change prediction loss based on the first temperature data and the average temperature change.

[0143] It should be further explained that the construction and training process of the distributed LS-SVR model in this embodiment includes:

[0144] Based on the number and distribution of target allocation sub-ends, by independently dividing the computing units according to the target allocation sub-ends, a local LS-SVR sub-model is configured for each target allocation sub-end, forming the basic architecture of the distributed LS-SVR model;

[0145] Based on the offset between the real-time voltage and the actual demand voltage of each target sub-terminal in the multi-source correction parameters, combined with the historical temperature change, line impedance value and anomaly type label of the target sub-terminal, the input feature vector of the local LS-SVR sub-model is obtained by feature concatenation.

[0146] Based on the historical input feature vector of each target sub-end and the corresponding actual temperature change, the kernel function type and penalty parameters of the local LS-SVR sub-model are determined by the particle swarm optimization algorithm to complete the local model training; among them, the target sub-end corresponding to the overcurrent anomaly type adopts the RBF kernel, and the target sub-end corresponding to the overvoltage or undervoltage anomaly type adopts the polynomial kernel, and the initial value of the penalty parameter is set based on the load level of the target sub-end.

[0147] Based on the trained local LS-SVR sub-model, the real-time input feature vector is input into the model to obtain the predicted temperature change of each target allocation sub-end to the allocation end.

[0148] Based on the real-time demand allocation coefficients of each target allocation sub-terminal, a weighted summation operation is performed using the weights constructed by the real-time demand allocation coefficients to fuse the predicted temperature changes of all target allocation sub-terminals and obtain the average temperature change corresponding to the allocation sub-terminal. It should be further noted that the real-time demand allocation coefficients in this embodiment are obtained by the ratio of the voltage and current data of all target allocation sub-terminals measured in real time.

[0149] Based on the first temperature data and the average temperature change at the distribution end, a loss function is constructed by calculating the deviation between the two, and the loss value is fed back to each local LS-SVR sub-model.

[0150] Based on the loss value, the parameters of each local LS-SVR sub-model are updated collaboratively using the federated averaging algorithm, and the distributed LS-SVR model is iteratively optimized until the loss value is lower than the preset loss threshold. It should be further noted that only the parameter gradients are passed during the collaborative update process, not the original data.

[0151] This process achieves independent prediction and global fusion of the temperature impact of each target allocation sub-end through a distributed architecture. It combines dynamic kernel functions and federated optimization to improve prediction accuracy, providing accurate data support for temperature drift prediction. At the same time, it reduces data transmission volume to ensure privacy, helps the system achieve accurate self-calibration through temperature loss assessment, and improves the stability of power allocation.

[0152] The power allocation offset prediction unit, based on a preset deviation prediction parameter space and a discriminator constructed by a BP neural network optimized by particle swarm optimization, obtains the predicted voltage drift and current drift for each target sub-terminal. At the same time, based on the actual load demand and measured distribution voltage and current information of each target sub-terminal, it obtains the measured voltage and current offset. Based on the measured voltage and current offset and the predicted voltage drift and current drift, it obtains the power allocation prediction deviation loss.

[0153] It should be further explained that the specific process of implementing the particle swarm optimization BP neural network in this embodiment includes:

[0154] Based on a pre-defined deviation prediction parameter space, feature normalization is used to map each parameter to the same order of magnitude, resulting in an input feature matrix. The deviation prediction parameter space includes predicted temperature change, a second temperature offset information sequence, anomaly type, the average drift coefficient of similar historical anomalies, line impedance temperature coefficient, voltage temperature coefficient, and current temperature coefficient. The predicted temperature change is a pre-estimate of the temperature change at the target sub-terminal, reflecting the potential impact of temperature fluctuations on voltage and current drift. The second temperature offset information sequence is the time-series data of the temperature offset of the target sub-terminal relative to a reference value, used to capture the dynamic characteristics of temperature changes. Anomaly type refers to specific anomalies such as overvoltage and overcurrent, used to distinguish the differentiated impact of different anomaly scenarios on drift. The average drift coefficient of similar historical anomalies is the average level of voltage and current drift in similar past anomalies, providing a historical benchmark for current prediction. The line impedance temperature coefficient characterizes the law of line impedance change with temperature, reflecting the influence of temperature on line transmission characteristics. The voltage temperature coefficient reflects the degree of voltage change with temperature and is directly related to the temperature sensitivity of voltage drift. The current temperature coefficient reflects the magnitude of current change with temperature and is directly related to the temperature sensitivity of current drift.

[0155] The number of nodes in the input layer of the BP neural network is determined based on the dimension of the input feature matrix. The number of nodes in the output layer is determined to be 2 based on the output target (voltage drift and current drift). The hidden layer structure is dynamically adjusted according to the anomaly type (3 hidden layers for overcurrent anomaly type, and 2 hidden layers for overvoltage and undervoltage anomaly types). The number of nodes in each layer is 1.5 times the number of nodes in the input layer, thus constructing the initial BP neural network framework.

[0156] Based on the particle swarm optimization algorithm, the initial weights and thresholds of the BP neural network are encoded as particle position vectors, with the particle dimensions being the number of weights from the input layer to the hidden layer, the number of weights between hidden layers, the number of weights from the hidden layer to the output layer, and the number of thresholds for each layer.

[0157] Based on historical datasets (including input feature matrices and corresponding measured voltage and current drift), a fitness function is constructed by calculating the mean square error between predicted and measured values. The particle swarm optimizes the algorithm by minimizing the fitness value.

[0158] Based on the global optimal position vector of particle swarm optimization, the values ​​are assigned to the weights and thresholds of the BP neural network through parameter decoding to obtain the BP neural network optimized by particle swarm optimization, which serves as the core prediction model of the discriminator.

[0159] Based on the real-time deviation prediction parameters of the target allocation sub-terminal, forward propagation is performed through the input discriminator. After processing by the hidden layer activation function, the target allocation sub-terminal corresponding to the overcurrent anomaly type adopts the ReLU function, and the target allocation sub-segment corresponding to the overvoltage and undervoltage anomaly types adopts the Sigmoid function. The predicted voltage drift and current drift are obtained by linear mapping of the output layer.

[0160] Based on the deviation between the predicted results and the measured drift in each batch, the discriminator parameters are fine-tuned through the backpropagation algorithm. The particle swarm optimization parameters are used as initial values, and backpropagation is used for local optimization. The model is iteratively updated until the prediction error stabilizes within the preset range.

[0161] The adversarial calibration unit constructs a comprehensive adversarial loss function based on power allocation prediction deviation loss, temperature change prediction loss, and filtering error loss. This comprehensive adversarial loss function is fed back to the discriminator and used for adversarial training in combination with a preset adversarial loss function threshold. The adversarial training results that meet the adversarial loss function threshold are then used to adjust the real-time demand allocation coefficients so that the real-time demand allocation coefficients and the actual demand allocation coefficients meet the preset deviation threshold.

[0162] The feedback adjustment unit is used to feed back the anomaly type and duration of each target allocation sub-end during the adjustment process to the enhanced acquisition adjustment model for real-time fine-tuning training, and to update and adjust the configuration parameters of information acquisition corresponding to each target allocation sub-end in real time.

[0163] This process achieves precise self-calibration of the power distribution system through multi-level technical collaboration, significantly improving operational stability and control accuracy. The error prediction unit adopts a distributed LS-SVR model, configuring a local LS-SVR sub-model for each target distribution sub-end. It takes input voltage offset, historical temperature changes, line impedance, and anomaly type as inputs, and optimizes the kernel function and penalty parameters through particle swarm optimization. Combined with the federated averaging algorithm, it updates parameters in a coordinated manner, accurately predicts the temperature change of each target distribution sub-end to the distribution end and calculates the loss. It retains the characteristics of the target distribution sub-ends and achieves global fusion, providing a reliable basis for temperature drift calibration. The offset prediction unit uses a particle swarm optimization BP neural network as the discriminator, inputs the offset prediction parameter space, dynamically adjusts the hidden layer structure, optimizes the weight threshold, and combines backpropagation to fine-tune the parameters, accurately predicting the voltage and current drift. The offset loss is obtained by comparing with the measured values. The multi-dimensional features and dynamic network structure improve the targeting and accuracy of the drift prediction. The adversarial calibration unit integrates the temperature change prediction loss, power allocation prediction offset loss, and filtering error loss to construct a comprehensive adversarial loss function. Through feedback training, it balances the weights of each loss and drives the real-time demand allocation coefficient adjustment, ensuring that the deviation from the true value is stable within the threshold. The adversarial mechanism strengthens the robustness of the calibration. The feedback adjustment unit feeds back the anomaly type and duration to the enhanced acquisition adjustment model, fine-tuning the acquisition parameters in real time, forming a prediction, calibration, and feedback closed loop, enabling the system to dynamically adapt to changes in operating conditions. This application not only improves the accuracy of temperature, voltage, and current drift prediction, but also ensures the real-time calibration accuracy of the demand allocation coefficient through adversarial training and closed-loop feedback, significantly enhancing the adaptability and operational stability of the power allocation system.

[0164] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments under the guidance of the present invention without departing from the spirit and scope of the claims. All of these variations are within the protection scope of the present invention.

Claims

1. A power distribution unit multi-terminal collaborative correction system based on adversarial learning, characterized in that, include: Acquisition module and self-calibration module; the acquisition module includes an acquisition unit and a preprocessing unit; The acquisition unit is used to perform differentiated acquisition and anomaly identification on each target distribution sub-end and distribution end of the power distribution unit according to a preset enhanced acquisition adjustment model, so as to obtain multi-source acquisition data corresponding to the distribution end and the target distribution sub-end with anomaly. The enhanced acquisition and adjustment model is constructed based on the historical anomaly frequency ratio of each target allocation sub-end, the duration of the corresponding anomaly type, and the actual demand allocation coefficient of each target allocation sub-end, combined with a reinforcement learning algorithm; the actual demand allocation coefficient is constructed from the actual load demand ratio of all target allocation sub-ends. The preprocessing unit is used to call the distributed filtering model constructed according to the federated algorithm framework and filtering algorithm to perform differentiated filtering, noise reduction and sparsification processing on the multi-source collected data based on the anomalies identified by each target allocation sub-end, so as to obtain the preprocessed multi-source correction parameters and sparsity error loss. The sparsity processing is achieved based on the contribution of each variable in the multi-source collected data to the correction of the actual demand allocation coefficient under the anomaly type corresponding to each target allocation sub-end. The multi-source acquisition data includes the total voltage, current, and first temperature data corresponding to the allocation end, as well as the actual demand allocation coefficient, and the measured allocation voltage, current, and second temperature offset information sequence of the corresponding target allocation sub-end under the corresponding abnormal type for each target allocation sub-end; the self-calibration module includes an error prediction unit; The error prediction unit is used to obtain the predicted temperature change of each target sub-terminal to the distribution terminal and the average temperature change of the distribution terminal based on the offset between the real-time voltage measured at each target sub-terminal in the multi-source correction parameters and the corresponding actual demand voltage, combined with the generator constructed by the preset distributed LS-SVR model, and to obtain the temperature change prediction loss based on the first temperature data and the average temperature change.

2. The adversarial learning based power distribution unit multi-terminal collaborative correction system of claim 1, wherein, The self-calibration module also includes a distribution offset prediction unit. The distribution offset prediction unit obtains the predicted voltage drift and current drift for each target distribution sub-terminal based on a preset deviation prediction parameter space combined with a discriminator constructed by a BP neural network optimized by particle swarm optimization. At the same time, based on the actual load requirements and measured distribution voltage and current information of each target distribution sub-terminal, it obtains the measured voltage and current offset. Based on the measured voltage and current offset and the predicted voltage drift and current drift, it obtains the power distribution prediction deviation loss. The deviation prediction parameter space includes the predicted temperature change, the second temperature offset information sequence, the anomaly type, the average drift coefficient of similar historical anomalies, the line impedance temperature coefficient, the voltage temperature coefficient, and the current temperature coefficient; the anomaly type includes overvoltage, undervoltage, and overcurrent.

3. The adversarial learning based power distribution unit multi-terminal collaborative correction system of claim 2, wherein, The self-calibration module also includes an adversary calibration unit; The anti-sub-calibration unit constructs a comprehensive anti-loss function based on power distribution prediction bias loss, temperature change prediction loss and filtering error loss, feeds the comprehensive anti-loss function back to the discriminator, performs anti-training combined with a preset anti-loss function threshold, and adjusts the real-time demand distribution coefficient using an anti-training result that meets the anti-loss function threshold, so that the real-time demand distribution coefficient and the real demand distribution coefficient meet a preset distribution deviation threshold.

4. The adversarial learning based power distribution unit multi-terminal collaborative correction system of claim 3, wherein, The reinforcement collection adjustment model is constructed by combining a first collection adjustment sub-model of a distribution end and M target collection adjustment sub-models and abnormality identification sub-models corresponding to target distribution sub-ends in a federated framework; The implementation process of the reinforcement collection adjustment model includes: Obtaining a real demand distribution coefficient and a preset distribution end sampling coefficient, initializing data collection configuration parameters of the first collection adjustment sub-model based on the distribution end sampling coefficient, and initializing data collection configuration parameters of the corresponding target collection adjustment sub-model based on the real demand distribution coefficient; Synchronizing data collection using the initialized first collection adjustment sub-model and target collection adjustment sub-model, analyzing the data collected at the distribution end, adjusting the distribution end sampling parameters based on the real-time first temperature data combined with a preset first sampling adjustment function when the real-time demand distribution coefficient of the distribution end changes; the first sampling adjustment function is constructed by combining the first temperature data change and the distribution end sampling configuration parameters with a convolution kernel function.

5. The adversarial learning based power distribution unit multi-terminal collaborative correction system of claim 4, wherein, The implementation process of the reinforcement collection adjustment model also includes: While analyzing the data collected at the distribution end, calling the abnormality identification sub-model corresponding to each target distribution sub-end to extract abnormal features and identify abnormalities from the data collected at each target distribution sub-end within a preset time length, and determining that the current target distribution sub-end is abnormal and obtaining the corresponding abnormal type when the change value of the main contribution variable corresponding to each abnormal type is greater than a preset abnormal threshold within the preset time length; the main contribution variable corresponding to each abnormal type is obtained by principal component algorithm according to the abnormality identification accuracy and the corresponding extracted abnormal variable features; Based on the corresponding abnormal type and the preset abnormal type-data collection parameter configuration library combined with a deep search algorithm, the optimal data collection configuration parameters corresponding to the target distribution sub-end are obtained, the initialized data collection configuration parameters are updated and adjusted, and when each abnormality is processed N times, the abnormal type, data collection configuration parameters and identification effect data are stored in the abnormal type-data collection parameter configuration library, and the data collection configuration parameters of the same type of abnormality are updated by clustering algorithm.

6. The adversarial learning based power distribution unit multi-end collaborative correction system of claim 5, wherein, The data collection configuration parameters include sampling frequency, sampling window width and data resolution; the sampling window width is obtained by combining the duration of the corresponding abnormal type, the abnormality identification rate, a sliding window and a simulation algorithm. The implementation process of the reinforcement collection adjustment model also includes: When there are more than two target allocation sub-ends with abnormality and the current resources cannot meet all the adjusted target allocation sub-ends after the data acquisition configuration parameter adjustment, a conflict arbitration mechanism is constructed according to the real-time demand allocation coefficient and the corresponding abnormality of the corresponding target allocation sub-ends, and the data acquisition resources of each target allocation sub-end are adjusted.

7. The adversarial learning based power distribution unit multi-terminal collaborative correction system of claim 6, wherein, The process of the differential filtering denoising and sparsification processing includes: Based on the federal algorithm framework and the target allocation sub-end abnormality type characteristics, a local filtering sub-model suitable for the abnormality type of each target allocation sub-end is matched through a matching algorithm, and a federal average algorithm is used to aggregate the filtering parameters of each target allocation sub-end at the allocation end to obtain a distributed filtering model; the target allocation sub-end abnormality type characteristics are associated with the abnormality type identified by the reinforcement collection adjustment model; Based on the data collected by each target allocation sub-end and the allocation end, the corresponding local filtering sub-model and cross-sub-end coupled noise are cooperatively suppressed to obtain multi-source collected data after differential filtering denoising.

8. The adversarial learning based power distribution unit multi-end collaborative correction system of claim 7, wherein, The process of the cross-sub-end coupled noise cooperative suppression is: Based on the noise spectrum characteristics extracted by the local filtering sub-model of each target allocation sub-end, the cosine similarity of the noise spectrum characteristics between each target allocation sub-end is calculated through a cosine function to obtain a noise correlation degree; A target allocation sub-end group with a noise correlation degree greater than a correlation degree threshold is obtained, and the center frequency and bandwidth of the band-stop filter are simultaneously optimized based on the maximum noise energy frequency and noise coupling strength in the target allocation sub-end group to cooperatively suppress the cross-sub-end coupled noise.

9. The adversarial learning based power distribution unit multi-end collaborative correction system of claim 8, wherein, The process of the differential filtering denoising and sparsification processing further includes: Based on the multi-source collected data after differential filtering and the real demand allocation coefficient, the sensitivity, mutual information entropy, and effective participation rate of each variable to the real demand allocation coefficient correction are calculated to obtain the comprehensive contribution degree of each variable; Based on the multi-source collected data after differential filtering denoising corresponding to each target allocation sub-end, a sparse attention network is combined, and the comprehensive contribution degree of all variables is used as the attention sparsity weight to obtain the multi-source collected data after sparsification processing; The deviation change value of the real-time demand allocation coefficient and the real demand allocation coefficient is used to evaluate each round of data sparsification process, and the evaluation result is fed back to the sparse attention network to adjust the attention sparsity weight in real time until the deviation of the real-time demand allocation coefficient and the real demand allocation coefficient meets the corresponding allocation deviation threshold in real time.

Images