Dynamic return and supplement method and system based on electric energy measurement data reduction model

Abstract

The invention relates to the technical field of electric energy metering anomaly detection and diagnosis analysis, in particular to a dynamic return and compensation method and system based on an electric energy metering data restoration model, and the method designs a distortion traceability algorithm of multi-dimensional feature fusion and improved SVM (Support Vector Machine), and extracts the characteristics of load features, environment features, packet loss rate, sampling delay distortion and the like. And load fluctuation-dominated distortion, environment interference-dominated distortion and coupling-dominated distortion can be accurately positioned. For multi-source distortion restoration, a load self-adaption-environment compensation-coupling correction three-level restoration model is provided, and through differential compensation of different load types, dynamic calibration of environment parameters such as temperature and humidity, electromagnetic interference and the like and coupling coefficient matrix optimization, the measurement data restoration precision is stably maintained to be 98% or above. According to the invention, through the research on the metering data distortion traceability-coupling distortion data reduction-dynamic return and supplement calculation-return and supplement result closed-loop optimization method, the precise repair of metering errors and the scientific calculation of electric quantity return and supplement in a complex scene are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electric energy metering anomaly detection and diagnosis analysis, and particularly relates to a dynamic compensation method and system based on an electric energy metering data restoration model. BACKGROUND

[0002] With the transformation of the power system to the "source-grid-load-storage" interactive mode, the electric energy metering scene is facing the dual challenges of intensified load dynamic fluctuation in the power grid system and complex external environmental interference factors. The traditional metering data processing and power compensation method has significant limitations, including: 1. Severe distortion of metering data caused by load fluctuation (1) Intermittent load impact: the start-stop of large equipment of industrial users (such as steel and chemical industry) and the peak-valley switching of high-power home appliances (air conditioner, electric heating) of residential users will cause the current and power data collected by the metering terminal to have instantaneous peaks or sudden drops (the fluctuation amplitude can reach 2-5 times the rated value). The traditional metering model cannot completely capture the load fluctuation process due to the fixed sampling frequency, resulting in power integral calculation deviation, and the single-day electric energy metering error can reach 3%-8%.

[0003] (2) New energy grid connection interference: the output fluctuation of distributed photovoltaic and wind power will cause bidirectional flow of distribution network load. The traditional metering model is designed based on "one-way load", and the power direction judgment for the "load-power generation" alternating scene lags, which is easy to cause reverse flow leakage or positive flow overcounting. Especially when the load fluctuation and new energy output fluctuation are superimposed, the monthly electric energy metering error can be as high as 12%.

[0004] (3) Insufficient adaptation of load characteristics: the fluctuation laws of different types of loads (such as resistive, inductive, and capacitive) are significantly different. For example, residential load is periodic day and night, and industrial load is production batch. The traditional metering data processing uses a unified algorithm and does not dynamically adjust the analysis dimension according to the load type, resulting in a difference of more than 50% in repair accuracy of the same distortion problem in different load scenes.

[0005] 2. Accumulation of metering deviation caused by complex external environmental factors (1) Temperature and humidity influence: high temperature (>40℃) will cause the precision of the voltage transformer (PT) of the metering terminal to drift (error increases by 0.2%-0.5%), high humidity (>85% RH) will cause the insulation performance of the current transformer CT to decrease, resulting in current collection deviation; low temperature (<-10℃) will cause the operation speed of the metering chip to slow down, and the data sampling to be delayed. The superposition of the above environmental factors will cause the cumulative quarterly electric energy metering error to reach 5%-10%.

[0006] (2) Electromagnetic and physical interference: strong electromagnetic radiation in substations and industrial plants can interfere with the signal transmission of the metering terminal, leading to data packet loss or error, and in severe cases, the packet loss rate can reach 15%. External dust corrosion can cause poor contact of the metering terminal, causing intermittent interruption of the current signal, and the transmission of metering data such as linear interpolation does not consider the randomness of interference, and the completion error can reach 10%-20%.

[0007] (3) Coupling effect of environmental factors: In actual scenarios, load fluctuations and environmental disturbances often occur simultaneously, such as the simultaneous occurrence of a sudden increase in residential air conditioning load during summer high temperatures and metering terminal high-temperature drift. Traditional technologies do not establish a "load-environment" coupling distortion model, and only handle a single factor, resulting in a significant reduction in data restoration accuracy, which cannot meet the accuracy requirements of power withdrawal and compensation, and the national standard requires a withdrawal and compensation error of ≤2%.

[0008] The existing traditional power withdrawal and compensation method has significant shortcomings, including: (1) Traditional withdrawal and compensation method is based on a single factor: Traditional withdrawal and compensation methods are mostly based on artificial on-site calibration data, which can only reflect the metering deviation at the calibration time, and cannot cover the full-cycle deviation caused by the dynamic changes of power data under the influence of load fluctuations and complex environmental factors. The calculation error of the withdrawal and compensation power caused by this can reach 15%-30%.

[0009] (2) Traditional withdrawal and compensation method model parameters are set statically: Existing withdrawal and compensation models use fixed parameters and do not dynamically adjust with changes in load fluctuations, external environmental factors, and environmental disturbance intensity. During load peak and valley alternation and seasonal environmental changes, the withdrawal and compensation accuracy can decrease by more than 40%.

[0010] (3) Traditional withdrawal and compensation method lacks closed-loop verification and testing: The withdrawal and compensation results of traditional methods are not verified with grid dispatching data, which can easily cause mismatches between user-side withdrawal and compensation power and grid-side data, leading to power marketing disputes. SUMMARY

[0011] The purpose of the present application is to address the problem that existing traditional methods cannot handle data coupling distortion under the influence of load fluctuations and complex environmental factors, and to construct a dynamic withdrawal and compensation method based on a power metering data restoration model. The present application also discloses a dynamic withdrawal and compensation system based on a power metering data restoration model.

[0012] Technical solution: First, the present application provides a dynamic withdrawal and compensation method based on a power metering data restoration model, which includes: Constructing a multi-dimensional measurement data feature set: collecting load characteristics and determining the corresponding feature direction of the power flow, wherein the load characteristics include: load data characteristics, environmental data characteristics, and distortion data characteristics, which respectively refer to load characteristics under fluctuation, load characteristics under environmental interference, and load characteristics under both fluctuation and environmental interference; Based on the nonlinear characteristics of the load characteristics, a box counting method is used to construct a feature sensitivity factor, and the feature sensitivity factor is used to optimize the fitness function in the particle swarm optimization algorithm, and the optimal penalty factor and kernel function parameter combination are determined by combining the support vector machine algorithm, thereby realizing the classification of the load characteristics; Based on the classification results of the above load characteristics, the same type of resistive load data characteristics within a certain time range under each classification is collected, and the data is stored by day and then fitted according to the load curve to complete the missing data sampling. The inductive load data characteristics are compensated and corrected by identifying the power condition of the transient peak; A coupling correction coefficient, a reverse power correction amount, and an electric quantity that needs to be corrected due to current interruption caused by dust concentration are combined with multiple linear regression to establish a coupling distortion data restoration model to obtain the electric quantity that needs to be restored, thereby performing electric quantity recovery for different scenarios, and obtaining the total recovery amount.

[0013] Further, the method further comprises: Based on the distortion type, a basic weight is allocated and a weight coefficient is adjusted according to the severity of the distortion, and the final recovery electric quantity under the distortion condition is constructed; The final recovery electric quantity obtained is subjected to multi-dimensional verification, including: substation total meter verification, historical data verification, and similar user verification.

[0014] Further, the method further comprises: Based on the multi-dimensional verification result, a mapping relationship between the deviation level and the model parameter adjustment strategy is constructed, and the effect of iterative optimization is evaluated.

[0015] Further, it further comprises: The feature sensitivity factor is constructed based on the nonlinear characteristics of the load data characteristics, including: For a load data characteristic sequence containing a plurality of data points, according to the time resolution of the load data, a scale set for box counting is set, wherein each scale is the side length of a box; Using the maximum and minimum principle, the standardized load data characteristic sequence is mapped to a two-dimensional plane of time dimension-load dimension, thereby constructing a two-dimensional data point set; The two-dimensional data points are divided according to the side length of each box in the scale set, thereby obtaining a plurality of groups of corresponding box numbers; For the side length of each box, i.e. the current scale and the number of boxes, a double logarithmic transformation is performed to construct a linear graph for slope estimation. That is, the logarithmic scale is used as the horizontal axis and the logarithmic number of boxes is used as the vertical axis. Multiple sets of double logarithmic transformed data points are plotted on a rectangular coordinate system to form a double logarithmic scatter plot. Linear regression was performed on the double logarithmic scatter plot. Based on the box counting principle, the absolute value of the slope of the fitted line is the fractal dimension FD. The slope of the fitted line was calculated using the least squares method, and the obtained fractal dimension FD was then compared with the energy proportion of the high-frequency subband. Together as feature sensitivity factors Input parameters.

[0016] Furthermore, it also includes: The slope of the fitted line is calculated using the least squares method, and the obtained fractal dimension FD is compared with the energy ratio of the high-frequency subband. Together as feature sensitivity factors The input parameters include: slope Represented as: ;

[0017] in, m The total number of scales in the scale set. Table 1 The number of corresponding boxes at each scale For the first The logarithmic scale corresponding to the scale; The feature sensitivity factor Represented as: ,in, and All are weights.

[0018] Furthermore, it also includes: The optimization of the fitness function in the particle swarm optimization algorithm using the feature sensitivity factor includes: For each load data feature sample Calculate the individual's characteristic sensitivity factor and will As a weight for the sample classification results, the corrected classification accuracy is the weighted accuracy. , is represented as: ; in, As an indicator function, when the model predicts values With real labels If they match, use 1; otherwise, use 0. The sum of the sensitivities for all samples ensures that the weighted accuracy is still within the [0, 100%] interval; Calculate the average feature sensitivity of the entire load data set: ; According to The sensitivity adjustment coefficient is set according to the size , so that the more complex the data set is, The higher the weight of the dynamic logic; expressed as: ; To further punish the parameter combination that misclassifies high samples, a misclassification sensitivity penalty term is introduced, expressed as: ; Therefore, the final fitness function is obtained by integrating the above modules: .

[0019] Further, it also includes: The combination of support vector machine algorithm determines the optimal penalty factor and kernel function parameter combination, and then realizes the classification of load data features, including: The particle swarm optimization algorithm randomly generates M particles, calculates the corresponding of each particle, determines the initial global optimal particle and individual optimal particle; each particle adjusts the flight speed and position according to its historical optimal fitness value and global optimal fitness value, and the new position needs to satisfy constraint; When the number of iterations reaches the upper limit, or the global optimal fitness value changes by less than 0.1 for 10 generations in a row, stop iteration, and the corresponding to the global optimal particle is the optimal support vector machine parameter that adapts to the current load nonlinear characteristics, ensuring that the classification accuracy of the model on high samples is improved.

[0020] Further, it also includes: The same type of resistive load data features within a certain time range are collected, stored by day, and then fitted according to the load curve to complete the missing data sampling, including: First, perform data screening and extract historical data of the same type of resistive load in the past period, store by day, and set the sampling points; Take the current time to be repaired as the center to construct a window data set , where is the number of valid data days, , each is the load data sequence of the 24 hours of the th valid day; to The average and standard deviation of the load at each time point are calculated to obtain a 24-hour period reference curve and the fluctuation range , expressed as: ; wherein, is the load value at time point of the th valid day, reflecting the 24-hour period trend of the resistive load; If there is a missing value at time point in the current data to be repaired, denoted as , the missing value is filled according to the weighted correction of and the valid data collected on the current day, and the filling formula is expressed as: ; wherein, is a weight coefficient, which is dynamically adjusted according to the similarity between the current day and the historical period, to ensure that the filling value not only fits the historical trend, but also adapts to the actual load fluctuation of the current day.

[0021] Further, it further comprises: The power condition compensation correction of the inductive load data characteristics by identifying transient peaks, comprising: Collecting real-time current data of the inductive load, determining the sampling frequency, and calculating the current change rate of adjacent sampling points: ; wherein, T is the sampling period, is the current at time point ; when the inductive load starts, it will suddenly rise, and when it stops, it will suddenly drop; Extracting power data of the inductive load in normal operation, calculating the rated power , obtaining the power mean value in the normal operation period and the start-stop transient power correction coefficient, i.e. the start time coefficient and the stop time coefficient , which correspond to the compensation proportion of the start peak and the stop sudden drop, respectively; Correction at start time: the original peak power contains distortion components caused by magnetizing inrush current, and the correction formula is defined as: ; Correction at stop time: the stop power suddenly drops to 0 or below, and the correction formula is: ; wherein, is the stop sudden drop power.

[0022] Further, it further comprises: The reverse power correction amount is obtained by compensating and correcting the reverse power in the power flow direction, expressed as: wherein, is the reverse power, is the total power, including forward and reverse power, is the reverse correction coefficient.

[0023] Further, it also includes: The obtained reverse power correction amount and the power that needs to be corrected due to the current interruption caused by dust concentration determine the coupling distortion data restoration model to obtain the power that needs to be restored, including: The power that needs to be corrected due to the current interruption caused by dust concentration is expressed as: Wherein, is the interruption time, is the total time of the metering period, is the average power in the period before interruption; A coefficient matrix of the coupling of the electric energy metering data under the influence of load fluctuation and complex environmental factors is constructed, expressed as: ; wherein, Load fluctuation frequency (times / hour), Temperature change amount, is the electromagnetic interference intensity change amount, is the dust concentration change amount; The coupling distortion data restoration model is expressed as: .

[0024] Further, it also includes: The power is compensated for different scenarios, and the total compensation amount is obtained, including: In the short-term distortion scenario, the short-term distortion is the distortion of the distortion data characteristics less than 24 hours, and the compensated power is expressed as: ; In the long-term distortion scenario, the long-term distortion is the distortion of the distortion data characteristics reaching more than 24 hours, and the compensated power is compensated by dividing the sub-period, expressed as: ; wherein, is the number of sub-periods, is the restoration power and the measured power of the th sub-period; In the scenario of new energy grid connection, the reverse power compensation is calculated separately, and the formula is as follows: ; In combination with the above scenarios, the total compensation amount of the compensation power is obtained, expressed as: .

[0025] Further, it also includes: The base weight is allocated based on the distortion type, and the weight coefficient is adjusted combined with the severity of the distortion, to construct the final compensation power in the distortion condition, including: The distortion type of the distortion data feature is determined and the base weight is allocated, wherein the base weight corresponding to the load fluctuation dominant type is greater than that of the environmental interference dominant type and the coupling dominant type; The corresponding weight coefficient is determined based on the severity of the distortion data feature, and then the final compensation power is obtained, expressed as: ; wherein, The severity coefficients in the load fluctuation dominant type, the environmental interference dominant type, and the coupling dominant type are respectively, The base weights of the load fluctuation dominant type, the environmental interference dominant type, and the coupling dominant type are respectively.

[0026] Further, it also includes: The obtained final compensation power is subjected to multi-dimensional verification, including: substation total table verification, historical data verification, similar user verification, and line loss verification, wherein, The similar user verification includes: selecting a user group consistent with the target user in the same substation, determining the average compensation rate of the group, and then obtaining the deviation rate of the target user compensation rate and the average compensation rate ; If the deviation rate is less than the set threshold, it is determined that the current compensation power meets the consistency requirement of similar users, otherwise, it is determined that the compensation power does not meet the requirement; The historical data verification includes: extracting the historical same period power data of the target user, the historical same period power data is the power data corresponding to the time at least one year ago from the current time as the standard, calculating the fluctuation range of the historical same period power, and finally judging whether the restored power obtained according to the coupling distortion data restoration model is within the fluctuation range, if the restored power is within the fluctuation range, it is determined that the compensation power meets the historical trend; if it exceeds the range, it is necessary to recheck whether the load feature extraction is complete and whether the environmental compensation coefficient is suitable; The substation total table verification includes: calculating the theoretical power of the substation total table, determining the difference between the substation line theoretical loss and the difference between the substation total table reading measured total power, calculating the deviation rate of the user compensation total amount and the difference; The line loss verification includes: calculating the actual line loss based on the restored user total power and the substation line theoretical loss, and then calculating the deviation rate between the actual line loss and the line theoretical loss; If and If all are within the set threshold, it is determined that the power to be compensated is consistent with the grid-side data, and if any deviation rate exceeds the set threshold, the coefficient matrix of the electric energy metering data coupling needs to be adjusted again.

[0027] Further, it further comprises: The mapping relationship between the deviation level and the model parameter adjustment strategy is constructed based on the multi-dimensional verification result, comprising: If and exceed the standard, it is determined to be mild deviation, at which time the local parameters are fine-tuned, otherwise, It is determined to be moderate deviation, at which time the gradient descent algorithm is used to minimize the verification deviation sum , and the coupling coefficient of the coupling distortion data restoration model is iteratively updated ; when or the number of iterations is greater than or equal to 50 times, stop; otherwise, It is determined to be severe deviation, at which time the penalty factor and the kernel function parameter need to be re-optimized to ensure that the classification accuracy rate returns to above the set threshold.

[0028] On the other hand, the present application also provides a dynamic compensation system based on an electric energy metering data restoration model, which comprises: A data acquisition module for constructing a multi-dimensional metering data feature set: acquiring load characteristics and determining the corresponding tidal flow direction of the characteristics, wherein the load characteristics include: load data characteristics, environmental data characteristics and distortion data characteristics, the load characteristics, environmental data characteristics and distortion data characteristics respectively refer to load characteristics under fluctuation, load characteristics under environmental interference and load characteristics under fluctuation and environmental interference at the same time; A sensitivity factor construction module for constructing a feature sensitivity factor based on the nonlinear characteristics of the load characteristics using the box counting method, and using the feature sensitivity factor to optimize the fitness function in the particle swarm optimization algorithm, and combining the support vector machine algorithm to determine the optimal combination of the penalty factor and the kernel function parameter, and then realizing the classification of the load characteristics; A compensation module for acquiring the same type of resistive load data characteristics within a certain time range under each classification based on the classification results of the above load characteristics, and storing them by day and then fitting the load curve to complete the missing data sampling; and compensating and correcting inductive load data characteristics by identifying the power condition of transient peak; A dynamic compensation module for establishing a coupling correction coefficient, a reverse power correction amount and an electric quantity that needs to be corrected due to dust concentration causing current interruption to determine a coupling distortion data restoration model, obtaining the electric quantity that needs to be restored, and then compensating the electric quantity for different scenarios, and obtaining the total compensation amount.

[0029] Advantages: Compared with the prior art, the present application has the following advantages: Firstly, the present application solves the problem of measurement accuracy caused by multi-source distortion. The prior art can only handle single factor interference. The present application uses a three-level sub-model of "load self-adaptation, environment compensation, and coupling correction". Through differential compensation of different load types such as resistive, inductive, and capacitive, dynamic calibration of environmental parameters such as temperature, humidity, and electromagnetic interference, and optimization of the coupling coefficient matrix, it accurately covers the superimposed distortion of new energy bidirectional flow, load fluctuation, temperature and humidity, electromagnetic interference, and other factors. The reverse flow leakage error is reduced from 5%-12% to below 0.5%, the measurement data restoration accuracy is stable ≥98%, the voltage transient and data packet processing response time is <5ms, and the completion error is <3%, which is much better than the accuracy of traditional models. Secondly, the present application realizes scientific and fair dynamic compensation and scene adaptation. For complex scenes such as new energy grid connection, a scene-specific compensation strategy is designed, and short-term / long-term / reverse flow differential calculation is constructed. Combined with the total table of the transformer area, historical data, and threefold verification of similar users, the compensation error is controlled within 2%, which is better than the national standard limit of 5%, and the compensation accuracy of industrial users is improved by 40%. The measurement dispute rate is reduced from 60% to below 10%. At the same time, through the closed-loop optimization of "data linkage verification + parameter adaptive iteration", the load coefficient and environmental compensation coefficient of the model are automatically adjusted. When the load fluctuation frequency is 0.5-10 times / hour and the environmental temperature is -10℃-50℃, the model accuracy decay is <2%, and the model can adapt to new scenes such as energy storage and microgrid within 24 hours without manual re-development.

[0030] Thirdly, the present application greatly improves the operation and maintenance efficiency and comprehensive value. Compared with the traditional model with a manual verification period of 3-6 months, the present application extends the verification period of the measurement device to more than 2 years, reduces the manual maintenance workload by 60%, and shortens the fault positioning response time from 24 hours to 4 hours. Precise measurement data supports the increase of new energy consumption rate by 5-8 percentage points, reduces the cost of power grid measurement management by 25%-35%, and reduces line loss by 2-3 percentage points, promoting the transformation of traditional measurement to intelligent self-adaptation and adapting to more than 90% of measurement terminals nationwide, providing key technical support for the construction of new power systems. In summary, the present application provides a method for constructing an electric energy measurement data restoration model and a dynamic compensation method in the scenarios of power grid load fluctuation, such as industrial intermittent production load, residential peak-valley electricity fluctuation, new energy grid connection impact load, and complex environmental factors, such as temperature, humidity, strong electromagnetic interference, dust corrosion, voltage transient, and sudden rise. It is suitable for various electric energy measurement scenarios that exist in power grid enterprises, industrial users, commercial complexes, and other electric energy measurement scenarios that have the risk of measurement data distortion. It can realize accurate repair of electric energy measurement distorted data and scientific calculation of electric energy compensation, and ensure the fairness of power transaction and the standardization of measurement management. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 This is a flowchart of the dynamic compensation and refund method based on the power metering data restoration model described in an embodiment of the present invention; Figure 2 This is a schematic diagram of a dynamic compensation system based on an energy metering data restoration model, as described in an embodiment of the present invention. Detailed Implementation

[0032] To better understand the present invention, the technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0033] Example 1: As Figure 1 As shown, this embodiment provides an integrated technical solution for constructing an energy metering data restoration model and calculating a dynamic energy refund / refund method that can achieve "true source tracing - dynamic restoration - accurate refund / refund - closed-loop optimization". It solves problems such as reliable processing of energy metering data and fair energy refund / refund calculation in complex scenarios, significantly improving the accuracy of energy metering and operation and maintenance efficiency. It provides reliable data support for load management, renewable energy consumption, and electricity price reform in new power systems, and is applicable to various distributed energy metering networks such as residential, industrial, commercial, and renewable energy power plants. Specifically, it includes the following steps: S1 constructs a multi-dimensional metering data feature set: collects load features and determines the power flow direction of the corresponding features. The load features include: load data features, environmental data features and distortion data features. The load data features, environmental data features and distortion data features refer to load features under fluctuating conditions, load features under environmental disturbances, and load features under both fluctuating and environmental disturbances, respectively. S2 uses box counting to construct a feature sensitivity factor based on the nonlinear characteristics of load features, and uses the feature sensitivity factor to optimize the fitness function in the particle swarm optimization algorithm. It then combines the support vector machine algorithm to determine the optimal combination of penalty factor and kernel function parameters, thereby achieving the classification of load features. Based on the above classification results of load characteristics, S3 collects the data characteristics of resistive loads of the same type within a certain time range under each classification, stores them by day, and fills in the missing data according to the fitted load curve; and compensates and corrects the data characteristics of inductive loads by identifying the power conditions of transient peaks. S4 combines multiple linear regression to establish a coupling correction coefficient, a reverse power correction amount, and a power correction model for power interruption caused by dust concentration to determine the coupling distortion data restoration model, obtain the power to be restored, and then perform power refund and compensation for different scenarios to obtain the total refund and compensation amount.

[0034] Based on the steps described above, this embodiment provides a more detailed description, including the following: Aiming at the problems of metering data distortion, inaccurate calculation of power recovery and compensation, poor model adaptability and the like caused by superimposed scenes of load fluctuation and complex environmental factors, the present application proposes a method suitable for power recovery model construction and dynamic recovery and compensation under the influence of load fluctuation and complex environmental factors, realizes accurate recovery of metering data and scientific calculation of power recovery and compensation through multi-dimensional feature fusion, adaptive model construction and iteration, closed-loop recovery and compensation strategy design and verification, and guarantees the fairness of power transaction and the efficiency of electric energy metering management.

[0035] The technical scheme of the present application includes metering data distortion tracing, data coupling distortion recovery model construction under complex scenes, dynamic power recovery and compensation calculation, closed-loop verification and optimization of recovery and compensation results, and forms a complete process technical system including "distortion data tracing-data recovery-power recovery and compensation-verification iteration" and the like.

[0036] 1. Metering data distortion tracing method design Aiming at the problem that the traditional technology cannot distinguish the source of metering data distortion, a method of fusing multi-dimensional data feature extraction and different distortion type classifier design is proposed to realize accurate tracing of metering data distortion caused by load fluctuation and environmental interference.

[0037] (1) Multi-dimensional metering data feature set construction 1) Load data features: collect real-time current, active power and reactive power of the metering terminal , load fluctuation frequency, load type label, and use power factor to determine: when is resistive, is inductive at 0.8-0.95, is capacitive, and the direction of power flow is determined by the voltage-current phase difference, is forward, is reverse.

[0038] 2) Environmental data features: collect environmental temperature , relative humidity , electromagnetic interference intensity , dust concentration , voltage sag / step-up amplitude , and calculate the deviation value of each parameter from the rated working condition.

[0039] 3) Distortion data features: extract data loss packet rate , sampling delay , and power calculation deviation, which is calculated by the formula , wherein, is the measured power, ​For the standard source calibration power, the cumulative power deviation is calculated using the formula wherein, is the measured cumulative power, is the cumulative power of the standard.

[0040] (2) Data distortion classifier design 1) The nonlinear characteristics of power grid system load data (such as power consumption and power) are often manifested as periodic fluctuations, mutations, and complex patterns affected by environmental factors (such as seasons, weather, and events). When dealing with such "irregular" nonlinearities, the classic SVM is prone to overfitting or underfitting.

[0041] The particle swarm optimization (PSO) algorithm is used to optimize the penalty factor of the support vector machine SVM algorithm , and the value range of C is set to 1-10, and the value range of the kernel function parameter γ is set to 0.1-1. A particle swarm adaptive optimization method based on load feature orientation is proposed, which optimizes the ability of SVM to handle nonlinearities by dynamically adjusting the PSO search constraints, and solves the problem of low classification accuracy of the support vector machine SVM algorithm for nonlinear characteristics.

[0042] First, a feature sensitivity factor is constructed, based on the identified nonlinear feature pattern, a feature sensitivity factor is calculated. If the data shows severe high-frequency fluctuations in a certain time period, then will increase.

[0043] For a load data feature sequence containing N data points , according to the time resolution of the load data, the scale set of the box count is set , wherein, is the side length of the box, i.e. the "time-load" two-dimensional space scale corresponding to the load sequence, which covers small-scale capture of high-frequency fluctuations and large-scale coverage of overall trends. The scale in this embodiment does not refer to a single time or load unit, but refers to the side length of the square box used to cover the data points in the two-dimensional plane composed of "time" and "load value". The following section is a specific description of the concept of "time-load" two-dimensional space scale: Using the maximum and minimum principle, the normalized load data feature sequence is mapped to the "time -load value " two-dimensional plane, wherein, is the time dimension, is the load dimension, and a two-dimensional data point set is constructed. The two-dimensional plane is divided into boxes with a side length of The square box is uniformly divided to form a grid matrix, and the number of boxes in the time dimension is . Among them, , that is, the box length in the time dimension is proportional to the load dimension length, is the ceiling operation, and the number of boxes in the load dimension is , which ensures that the entire data point set is completely covered. The number of boxes containing at least one data point P is counted and denoted as .

[0044] For each in the scale set Repeat the above operation to obtain m groups of “scale - box number ” corresponding data, denoted as , to ensure that each group of data is not missed or repeated.

[0045] According to the core principle of box counting method, , the logarithmic transformation of scale and box number is performed respectively to construct a linear figure for slope estimation. With “logarithmic scale ” as the horizontal axis ( axis), and “logarithmic box number ” as the vertical axis ( axis), the group of double logarithmic transformed data points is plotted in the rectangular coordinate system to form a double logarithmic scatter plot.

[0046] Linear regression fitting is performed on the double logarithmic scatter plot, and the absolute value of the slope of the fitted straight line is taken as the fractal dimension FD. Assuming that the equation of the fitted straight line is: , where is the slope of the straight line, is the intercept.

[0047] According to the principle of box counting method , where is a proportional constant, it can be known that , that is, the fractal dimension is equal to the absolute value of the slope of the fitted straight line.

[0048] Then the least squares method is used to calculate the slope , (1) The fractal dimension FD obtained through the above process, together with the high-frequency sub-band energy ratio , is used as the input parameter of the feature sensitivity factor : (2), wherein, and are weights, and in this embodiment, the high-frequency subband energy is obtained by wavelet decomposition and high-frequency subband extraction, wherein, first, wavelet transform: using discrete wavelet transform or wavelet packet decomposition, the signal / image is decomposed into multi-scale, multi-direction subbands. For example: one-dimensional signal: decomposed into low-frequency approximation coefficients and high-frequency detail coefficients, such as Haar wavelet or Daubechies wavelet. Two-dimensional image: decomposed into low-frequency approximation, horizontal detail, vertical detail and diagonal detail four subbands by dwt2 function or PyWavelets library. Decomposition level: according to the signal length, the maximum decomposition level is selected, and each layer of decomposition further subdivides the low-frequency part, while the high-frequency part usually retains the details. Secondly, high-frequency subband energy calculation: the energy formula is the sum of the square of the high-frequency subband coefficients, which is: ; (3) wherein, is the coefficient value of the high-frequency subband at position , M × N is the subband size.

[0049] The PSO algorithm is guided by the fitness function to optimize the penalty factor of SVM C and the kernel function parameter γ, improving the classification accuracy of nonlinear load characteristics.

[0050] Feature sensitivity is a measure value, and the larger the value is, the more complex or obvious the nonlinear characteristics are.

[0051] It is defined as follows: ; (4) wherein, and are weights, determining the importance of different feature indicators.

[0052] This feature sensitivity is introduced into the fitness function of the particle swarm optimization (PSO) algorithm, and the performance of the SVM model in processing such characteristics is punished or rewarded.

[0053] Specifically, for each load sample , the individual feature sensitivity can be obtained by box counting method and wavelet energy proportion, and is taken as the weight of the sample classification result, and the classification accuracy is corrected to the weighted accuracy : ; (5) ​where, is an indicator function, taking 1 when the model prediction matches the true label , and 0 otherwise, is the sum of sensitivity for all samples, ensuring the weighted accuracy is still in the interval [0, 100%]; The average feature sensitivity of the entire load dataset is calculated: ; the sensitivity adjustment coefficient is set according to the size of , so that the more complex the dataset is, the higher the weight of the dynamic logic.

[0054] ; (6) To further punish parameter combinations that misclassify high samples, a misclassification sensitivity penalty term is introduced, which is expressed as: ; (7) where the coefficient 0.1 controls the punishment degree to avoid the penalty term being too large to overshadow the dominant role of accuracy; instead of , it is used to impose a more severe punishment on larger sample misclassification.

[0055] Integrating the above modules, the final fitness function is: ; (8) The PSO algorithm finds the parameter combination that maximizes by iteration.

[0056] Again, in order to more efficiently search within the target range ( ) and make full use of the information of "load feature orientation", the present invention also proposes to dynamically adjust the search boundary of PSO.

[0057] PSO randomly generates M particles, for each particle, calculates its corresponding , determines the initial global optimal particle and individual optimal particle; each particle adjusts its flight speed and position according to the fitness values of its own historical optimal and global optimal, and the new position needs to meet the constraints of ; when the iteration number reaches the upper limit, or the global optimal fitness value changes by less than 0.1 for 10 consecutive generations, the iteration is stopped, at this time the corresponding to the global optimal particle is the optimal SVM parameter that fits the nonlinear characteristics of the current load, ensuring that the classification accuracy of the model on high samples is improved.

[0058] ​​If a certain parameter region is found to have very high fitness under the current load characteristics, i.e. , the PSO algorithm can dynamically shrink the search range of its particle swarm and focus more particles on this potentially promising region.

[0059] On the contrary, if a certain region continues to perform poorly after a period of search, it can be temporarily excluded or reduced in search priority. If the current data exhibits unprecedented high nonlinearity, i.e. , the exploration range of certain regions can be appropriately expanded to prevent missing potential optimal solutions.

[0060] This dynamic boundary adjustment mechanism enables PSO to intelligently focus on the most promising parameter space, rather than evenly distributing computational resources across the entire preset range. It can more quickly discover the optimal SVM parameters for the current nonlinear load data, thereby improving efficiency and accuracy.

[0061] 2) Classification logic design: input the above load data characteristics with fluctuations, load data characteristics under environmental interference, and load distortion characteristics with both of the above, etc. to design output 3 types of distortion: ① load fluctuation dominant type; ② environmental interference dominant type; ③ coupling dominant type that meets the characteristics of the above two types.

[0062] 2. Coupling distortion data restoration model construction For the above multi-source distortion data metering restoration model construction, a "three-level sub-model coordination" data coupling distortion restoration method is proposed to solve the problem of traditional technology that cannot adapt to multi-load, multi-environmental interference, etc. under the scene of electric energy metering distortion.

[0063] That is, in this embodiment, after obtaining the classification data of the three types of distortion, the corresponding historical data is collected, i.e. the restoration of the corresponding model can be constructed. Specifically, historical data is collected for each type of classification data, such as collecting historical data for load fluctuation dominant type related data, which includes resistive load data and inductive load data.

[0064] (1) Load adaptive sub-model construction Resistive load repair: a "24-hour sliding window + periodic trend fitting" algorithm is proposed, which fits the load curve based on the historical data of the same time period in the past 3 months (excluding abnormal days), fills in the missing data, and the error is less than 2%. Specifically, first, the data is screened, and the historical data of the same type of resistive load in the past three months is extracted, stored by day, and 1440 sampling points are obtained per day (24 hours x 60 minutes = 1440), in units of kW. Abnormal daily data such as holiday load drop and abnormal fluctuation caused by equipment failure are removed by the "3σ rule", and valid data days ≥60 days are retained.

[0065] Secondly, taking the current time to be repaired as the center, such as 14:30 on June 10, a time-aligned 24-hour sliding window is constructed, which covers the 24-hour period from -12 hours to +11 hours, corresponding to the same 24-hour period of each valid day in the historical data, forming a window data set

[0066] For each time point (a total of 1440) in , the load mean and standard deviation are calculated to obtain the 24-hour cycle reference curve and the fluctuation range , the formula is:

[0067] Among them, is the load value of the th valid day at time , and reflects the 24-hour cycle trend of resistive load; If the current data to be repaired is missing at time , it is recorded as , then according to the valid data collected on the current day, such as the load values of the previous and subsequent 1 hour , a weighted correction is made to complete the formula: ; (11) Among them, is the weight coefficient, which can be dynamically adjusted according to the similarity between the current day and the historical cycle. When the similarity is high, , and when the similarity is low, , to ensure that the completed value not only fits the historical trend, but also adapts to the actual load fluctuation of the current day.

[0068] ​​​​​​​​​​Inductive load repair: A "transient feature extraction + power compensation" algorithm is proposed to identify the power spike at the start and stop time of the device through current rate of change determination. Specifically, when inductive loads (such as motors, transformers) start and stop, there is "magnetic inrush current", and the distortion is mainly power spike. Therefore, this embodiment adopts the method of identifying transient spikes and compensating and correcting.

[0069] Collect real-time current data of inductive load, sampling frequency 50Hz, i.e. every 20ms a sampling point, unit A, calculate the current rate of change of adjacent sampling points: ; (12) When the inductive load starts, It will rise sharply, and when it stops, it will drop sharply.

[0070] Build a power-based data compensation model and data distortion correction method. Extract the power data of the inductive load during normal operation, exclude the start and stop period, calculate the rated power , take the average power during normal operation, and at the same time obtain the start and stop transient power correction coefficient , respectively corresponding to the compensation ratio of start spike and stop sudden drop.

[0071] Correction at start time: original spike power contains distortion components caused by magnetic inrush current, and the correction formula is defined as: ; (13) Correction at stop time: stop power drops below 0 (negative deviation of measurement), and the correction formula is: ; (14) where, is the stop sudden drop power.

[0072] Two-way flow problem determination, through real-time flow direction judgment and reverse power correction technology design, define the reverse power The correction formula is as follows: (15) where, is the measured power, is the reverse power, is the total power, including forward and reverse power, and 0.02 is the reverse correction coefficient.

[0073] (2) Environment compensation sub-model construction For temperature and humidity compensation, based on the temperature and humidity characteristic curve of the measuring device, the current / voltage collection deviation in the system is corrected using the following formula:

[0074] wherein, is the temperature change value, is the humidity change value, 0.001, 0.0005 is the temperature and humidity influence coefficient of current, 0.0008, 0.0003 is the temperature and humidity influence coefficient of voltage.

[0075] For electromagnetic interference compensation, an LSTM time series completion algorithm is used to train the model using the first 10 minutes of normal data to complete the missing data. In this embodiment, the LSTM time series completion algorithm effectively handles long-term dependencies and nonlinear patterns of time series through the design of gating mechanism and cell state.

[0076] For dust concentration, set the current interruption caused by dust, and the power formula that needs to be corrected is: (18) wherein, is the interruption time, is the total time of the measurement period, is the average power in the last hour before interruption.

[0077] Voltage sag compensation coefficient is the sag amplitude.

[0078] (3) Coupling correction sub-model A coefficient matrix is constructed for the coupling of power measurement data under the influence of load fluctuations and complex environmental factors: Through multiple sets of historical data, a multiple linear regression is designed to establish the coupling correction coefficient:

[0079] wherein, is the load fluctuation frequency, unit is times / hour, temperature change, unit is ℃, is the electromagnetic interference intensity change, unit is dBμV / m, is the dust concentration change, unit is mg / m³; The above sub-models are integrated to restore the calculation of electric power: (20) In this embodiment, mainly includes corrected power and dust power. In fact, temperature compensation and electromagnetic compensation are also included in the above quantities, which are not easy to measure separately. Therefore, this embodiment introduces a multi-coupling compensation coefficient to solve the above problems.

[0080] 3. Dynamic retirement compensation power calculation ​A "scene algorithm + dynamic weight + multi-dimensional verification" restoration power refund method is proposed to solve the problems of traditional methods involving static refund, low accuracy, etc.

[0081] (1) Scene-based power refund algorithm Short-term distortion (<24 hours): use the difference method of "restored power - measured power": (21) Wherein, is the measured power under short-term distortion.

[0082] Long-term distortion (≥24 hours): according to "load peak and valley + environmental grade", the sub-period (peak time 8:00-22:00, valley time 22:00-8:00) is divided to calculate the refund in segments: (22) Wherein, is the number of sub-periods, is the restored power and measured power of the th sub-period.

[0083] New energy grid-connected scene: Calculate the reverse power refund separately , defined as follows:

[0084] Wherein, is the reverse power calculation power, is the reverse power measurement power.

[0085] According to the above description, the total refund amount of the refund power is: (24) (2) Dynamic weight distribution algorithm According to experience, the load fluctuation dominant coefficient is 0.4, the environmental interference dominant coefficient is 0.3, and the coupling dominant coefficient is 0.3. For the allocation of basic weights based on distortion type, this embodiment sets clear threshold intervals for each key performance indicator by combining "load fluctuation rate", "environmental parameter deviation", "data packet loss rate", etc., thereby objectively dividing the distortion degree into three levels, providing measurable and reproducible data support for the distribution of weight coefficients.

[0086] Specifically, according to the following table definition, the distortion degree type is determined:

[0087] Combined with the adjustment of weight coefficients based on distortion severity, the final refund power can be obtained: (25) Wherein, The severity coefficients in the three cases of load, environment, and coupling, respectively.

[0088] 3. Closed-loop verification and optimization of recovery results A closed-loop mechanism of "data linkage verification + parameter adaptive iteration" is proposed to solve the problem of long-term accuracy decline caused by the lack of iteration capability and fixed parameters in traditional models, ensuring that the model can maintain high recovery and compensation accuracy when the load type changes and the environmental conditions fluctuate.

[0089] Multi-dimensional verification, mainly including: Substation total table verification: the sum of user recovery and compensation power and the difference between the theoretical power and the measured power in the substation total table is less than or equal to 2%, otherwise the model parameters need to be adjusted. Historical data verification: the recovery and compensation power should be within ±15% of the user's historical power, otherwise manual review is triggered, such as on-site verification of metering devices. Same type user verification: the recovery rate of users with the same load type and environmental conditions in the same substation deviation ≤ 3%, to ensure the fairness of recovery and compensation.

[0090] Specifically, (1) Data linkage verification A three-dimensional verification system of "horizontal-vertical-grid side" is designed to comprehensively verify the rationality and accuracy of recovery results, avoiding misjudgment caused by single verification dimension. The specific design is as follows: 1) Horizontal verification, used for consistency verification of similar users Select users with the same load type, similar environmental conditions, and consistent metering device type in the same substation as the target user, calculate the average recovery rate of the group , recovery rate = recovery power / measured power × 100%, then calculate the deviation rate of the target user's recovery rate and the average recovery rate: (26) Among them, is the target user's recovery rate, , is the number of similar users.

[0091] Judgment standard: if , the recovery result meets the consistency requirement of similar users; if , trigger the "model parameter fine-tuning" process; if trigger the "manual review + model parameter retraining" process, such as confirming whether the target user's metering device has hardware failure on-site.

[0092] 2) Vertical verification, used for historical data trend verification Extract the same period electricity data of the target user in the past 12 months, such as the current selected time is September 2025, extract the data between September 2025 and September 2024, and calculate the fluctuation range of historical same period electricity , wherein the fluctuation range = historical same period electricity ± 15%, and then determine the "measured electricity + compensation electricity", i.e. the actual electricity after restoration Whether it is within the fluctuation range: (27) Special scene processing: if the user has load expansion, production process adjustment, etc. (which needs to be recorded and confirmed by the power marketing system), the historical fluctuation range is recalculated based on the adjusted load capacity. For example, if the capacity is expanded by 20%, the upper limit of the fluctuation range will be increased by 20%.

[0093] Determination criteria: if , the compensation result is consistent with the historical trend; if it exceeds the range, the load characteristic extraction needs to be checked again and the environmental compensation coefficient needs to be adjusted.

[0094] 4. Grid side verification, used to verify the total meter and line loss of the transformer area The sum of the compensation electricity of all users in the transformer area Two-way verification with grid side data: 1) Transformer area total meter verification: calculate the theoretical electricity of the transformer area total meter, which is the difference between the electricity of the upper grid gateway and the theoretical loss of the transformer area line And the measured total electricity of the transformer area total meter , that is: (28) Calculate the deviation of the total compensation amount of the user and the difference, that is: (29) 2) Line loss verification: based on the total electricity of the user after restoration , calculate the actual line loss , based on the line resistance, current calculation and line theoretical loss The deviation rate between them is: (30) 3) Determination criteria: if , the compensation result is consistent with the grid side data; if any deviation rate exceeds the above threshold setting, the coefficient matrix of the coupling correction sub-model needs to be adjusted again, such as optimizing the electromagnetic interference correction coefficient in the environmental compensation due to large line loss deviation.

[0095] 6. Parameter adaptive iterative adjustment Based on the 3D verification results, a mapping relationship between deviation levels and model parameter adjustment strategies is constructed. The gradient descent algorithm is used to achieve automatic iteration of model parameters, avoiding the subjectivity and lag of manual adjustments. The specific design process is as follows: (1) Model parameter adjustment strategy To dynamically adjust model parameters, a quantitative deviation assessment system based on "three-dimensional validation results" needs to be established. This is based on cross-sectional validation (deviation rate among similar users). ), longitudinal verification (historical data fluctuation range) and grid-side verification (transformer area total meter deviation rate) With line loss deviation rate The deviation level is determined by comprehensively assessing whether the specific indicators of these three dimensions exceed the standard and the degree of exceeding the standard. In a specific embodiment, a slight exceedance of only one dimension is defined as "mild deviation", an exceedance of two or more dimensions or a severe exceedance of one dimension is defined as "moderate deviation", and an exceedance of all three dimensions or an extreme abnormality in the key dimension is defined as "severe deviation".

[0096] For different levels of deviation, the core parameters of the coupled distortion data restoration model and the dynamic compensation / refund power calculation module are dynamically adjusted: 1) Slight deviation: Fine-tune using local parameters like Exceeding the limit (poor consistency among similar users): Adjust the load factor of the load adaptive sub-model. Adjust the step size to ±0.05. For issues such as large deviations in inductive users, this can be addressed. The value has been adjusted from 1.2 to 1.25. like Exceeding the standard (large deviation of the total meter in the distribution area): Adjust the temperature and humidity influence coefficient of the "environmental compensation sub-model". If the deviation is large in a high-temperature scenario, adjust the current temperature and humidity coefficient from 0.001 to 0.0012. Formula for adjusting parameters: ,in, The sign of the deviation is positive or negative; a positive deviation is +1, and a negative deviation is -1.

[0097] 2) Moderate deviation: Optimization using coupling coefficient matrix Based on the gradient descent algorithm, minimize the sum of validation bias. Iteratively update the coupling coefficients of the "coupled correction sub-model". : calculate Partial derivatives with respect to each coupling coefficient For example, the load fluctuation frequency coefficient is 0.01 and the temperature deviation coefficient is 0.005.

[0098] Update parameters along the negative direction of partial derivatives: Among them, learning rate (Ensure iterative stability; Iteration termination condition: when or the number of iterations ≥ 50 times.

[0099] 3) Severe deviation: retrain using model parameters Recruit "load-environment-distortion" sample data in the past 3 months, sample size ≥ 5000 groups, covering the characteristics of the current deviation scene, such as high dust and high fluctuation load; Based on the new sample, retrain and improve the SVM distortion classifier and the improved XGBoost error prediction model, and the specific steps are as follows: a. SVM classifier: re-optimize the penalty factor and the kernel function parameters , to ensure that the classification accuracy rate returns to more than 98%.

[0100] b. XGBoost error prediction model: adjust the tree depth (3-8 layers), learning rate (0.01-0.1), and ensure that the model generalization error is less than 2% through 5-fold cross-validation.

[0101] c. After retraining, test a small batch of samples, test sample size ≥ 100 groups, verify that the restoration accuracy is ≥ 98% and the compensation error is ≤ 2%, and then apply it to the actual scene.

[0102] After each parameter iteration, from three verification dimensions of horizontal, vertical, and grid side, the optimization effect is verified through index evaluation: a. Restoration accuracy improvement rate: , where is the metering data restoration accuracy.

[0103] b. Compensation error reduction rate: , where is the compensation power calculation error.

[0104] c. Verification deviation compliance rate: verification deviation compliance rate is a key indicator to measure the overall effectiveness of the three-dimensional verification system, which includes horizontal, vertical, and grid side. By quantifying the proportion of the number of verification dimensions that meet the standard, it is determined whether the compensation result and model parameters meet the requirements of actual application, providing clear triggering basis for subsequent parameter iteration.

[0105] Total number of verification dimensions : fixed at 3, i.e. horizontal verification (consistency of similar users), vertical verification (historical data trend), and grid side verification (total table and line loss of the area); Number of verification dimensions that meet the standard In the three-dimensional verification, the number of verification dimensions meeting the preset threshold requirement of the deviation rate, i.e., a single dimension deviation rate ≤ threshold value is determined as meeting the standard; Verification deviation meeting rate The ratio of the number of verification dimensions meeting the standard to the total number of verification dimensions is expressed in percentage.

[0106] The verification deviation meeting rate formula is defined as follows: (31) 7. Archiving mechanism construction The deviation reason, adjustment parameter and evaluation result of each iteration are stored in a blockchain, such as a consortium chain, to ensure traceability and non-tamperability of parameter adjustment, and to provide a reference for subsequent similar deviation scenarios.

[0107] (1) Precision stability: Through parameter self-adaptive iteration, the restoration accuracy can still be maintained at more than 98% when the load fluctuation frequency increases from 0.5 times / hour to 10 times / hour and the environmental temperature increases from -10℃ to 50℃, which is significantly improved compared with traditional static models.

[0108] (2) Operation and maintenance efficiency: No manual periodic parameter adjustment is required, and the iteration process is automatically completed, which can reduce the manual maintenance workload by more than 60% per year, and reduce the measurement dispute rate caused by parameter mismatch.

[0109] (3) Scene adaptability: The closed-loop mechanism can automatically learn new scene characteristics, such as newly added energy storage charging and discharging load and extreme rain weather, without the need to redevelop the model, thereby expanding the application range of the technical solution.

[0110] Embodiment two: The application also provides a restoration power model construction and dynamic compensation system suitable for load fluctuation and complex environmental factors, as shown in Figure 2 The system comprises: A data acquisition module for constructing a multi-dimensional metering data feature set: acquiring load characteristics and determining the corresponding feature flow direction, wherein the load characteristics include load data characteristics, environmental data characteristics and distortion data characteristics, and the load data characteristics, environmental data characteristics and distortion data characteristics refer to load characteristics under fluctuation, load characteristics under environmental interference and load characteristics under fluctuation and environmental interference at the same time; A sensitivity factor construction module for constructing a feature sensitivity factor based on the nonlinear characteristics of the load characteristics using a box counting method, and using the feature sensitivity factor to optimize the fitness function in the particle swarm optimization algorithm, and combining a support vector machine algorithm to determine the optimal combination of penalty factor and kernel function parameter, thereby realizing classification of the load characteristics; The compensation module is used for collecting the same type of resistive load data features in a certain time range under each classification based on the classification result of the load features, storing the same type of resistive load data features according to the fitting load curve, and completing the missing data sampling; the inductive load data features are compensated and corrected through the power condition of the transient peak; The dynamic compensation module is used for establishing a coupling correction coefficient, a reverse electric quantity correction amount, and a coupling distortion data restoration model for determining the electric quantity to be restored, and obtaining the total compensation amount by combining the multiple linear regression and the electric quantity compensation for different scenes.

[0111] Other technical features of the dynamic compensation system based on the electric energy metering data restoration model and the corresponding dynamic compensation method based on the electric energy metering data restoration model are similar to the embodiments of the present application, and will not be repeated here.

[0112] Although the preferred embodiments of the present application have been described, those skilled in the art can make additional changes and modifications to the embodiments once they know the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0113] Obviously, those skilled in the art can make various modifications and variations to the embodiments of the present application without departing from the spirit and scope of the embodiments of the present application. Thus, if these modifications and variations of the embodiments of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application also intends to include these modifications and variations.

Claims

1. A dynamic backfill method based on an electric energy metering data restoration model, characterized in that, The method comprises: Constructing a multi-dimensional measurement data feature set: collecting load characteristics and determining the corresponding feature flow direction, wherein the load characteristics include load data characteristics, environmental data characteristics, and distortion data characteristics, which refer to load characteristics under fluctuation, load characteristics under environmental interference, and load characteristics under both fluctuation and environmental interference, respectively; Based on the nonlinear characteristics of the load characteristics, a box counting method is used to construct a feature sensitivity factor, and the feature sensitivity factor is used to optimize the fitness function in the particle swarm optimization algorithm, and the optimal penalty factor and kernel function parameter combination is determined by combining the support vector machine algorithm, and then the classification of the load characteristics is realized; Based on the classification results of the above load characteristics, the same type of resistive load data characteristics within a certain time range under each classification is collected, and the data is stored by day and then fitted according to the load curve to complete the missing data sampling; the inductive load data characteristics are compensated and corrected by identifying the power condition of the transient peak; The coupling correction coefficient, the reverse power correction amount, and the power correction amount required for the current interruption caused by dust concentration are combined with multiple linear regression to determine the coupling distortion data restoration model, and the power to be restored is obtained, so that the power is restored for different scenarios, and the total restoration amount is obtained.

2. The dynamic backfill method based on electric energy metering data restoration model according to claim 1, characterized in that, The method further comprises: Assigning a basic weight to the classification results of the above load characteristics and adjusting the weight coefficient according to the severity of the distortion to construct the final restoration power under the distortion condition; The final restoration power obtained is subjected to multi-dimensional verification, including: substation total meter verification, historical data verification, and similar user verification.

3. The dynamic backfill method based on electric energy metering data restoration model according to claim 2, characterized in that, The method further comprises: Based on the multi-dimensional verification results, a mapping relationship between the deviation level and the model parameter adjustment strategy is constructed, and the effect of iterative optimization is evaluated.

4. The dynamic backfilling method based on the electric energy metering data restoration model according to any one of claims 1-3, characterized in that, The feature sensitivity factor is constructed based on the nonlinear characteristics of the load data characteristics using a box counting method, which comprises: For a load data characteristic sequence containing a plurality of data points, according to the time resolution of the load data, a scale set for box counting is set, wherein each scale is the side length of a box; Using the maximum and minimum principle, the standardized load data characteristic sequence is mapped to a two-dimensional plane of time dimension-load dimension, thereby constructing a two-dimensional data point set; The two-dimensional data points are divided according to the side length of each box in the scale set, thereby obtaining a plurality of box numbers corresponding to each box; The side length of each box, i.e. the current scale and the box number, are respectively subjected to double logarithmic conversion, thereby constructing a linear pattern for slope estimation, i.e. taking the logarithmic scale as the horizontal axis and the logarithmic box number as the vertical axis, and drawing the plurality of double logarithmic converted data points in the rectangular coordinate system to form a double logarithmic scatter plot; Linear regression fitting is performed on the double logarithmic scattered point diagram, according to the principle of box counting method, the absolute value of the slope of the fitting straight line is the fractal dimension FD, and the slope of the fitting straight line is calculated by using the least square method, and the obtained fractal dimension FD and the high frequency subband energy proportion are input parameters together as a feature sensitivity factor .

5. The dynamic backfill method based on electric energy metering data restoration model according to claim 4, characterized in that, The slope of the fitting straight line is calculated by using the least square method, and the obtained fractal dimension FD is compared with the high frequency sub-band energy proportion Commonly as the input parameter of the feature sensitivity factor , including: Slope is represented as: ; wherein, m is the total number of scales in the set of scales, Table the number of bins at the scale, is the log scale corresponding to the scale. The feature sensitivity factor is represented as: wherein, and are weights.

6. The dynamic backfill method based on electric energy metering data restoration model according to any one of claims 1-3, characterized in that, The fitness function in the particle swarm optimization algorithm is optimized using the feature sensitivity factor, which comprises: For each load data feature sample , the individual feature sensitivity factor of the sample is calculated , and the is used as the weight of the sample classification result to correct the classification accuracy to the weighted accuracy , which is expressed as: ； where, is an indicator function that takes 1 when the model prediction is consistent with the true label and 0 otherwise, is the sum of sensitivity for all samples, ensuring that the weighted accuracy is still in the [0, 100%] interval; The average feature sensitivity of the entire load dataset is calculated: ; according to the size of the sensitivity adjustment coefficient , the more complex the dataset, the higher the dynamic logic weight; represented as: ； To further penalize the parameter combinations that misclassify the high sample, a misclassification sensitivity penalty term is introduced is expressed as: ； Thus, integrating the above modules, the final fitness function is: .

7. The dynamic backfill method based on electric energy metering data restoration model according to claim 6, characterized in that, The optimal penalty factor and kernel function parameter combination is determined by combining the support vector machine algorithm, and then the classification of the load data characteristics is realized, which comprises: The particle swarm optimization algorithm randomly generates M particles, and calculates the corresponding [property name] for each particle. The initial globally optimal particle and the individual optimal particle are determined; each particle adjusts its flight speed and position based on its historical and globally optimal fitness values, and the new position must satisfy... Constraints; When the iteration number reaches the upper limit, or the global optimal fitness value changes less than 0.1 for 10 generations in succession, the iteration is stopped, and the global optimal particle corresponds to the global optimal solution That is, the parameters of the optimal support vector machine adapted to the nonlinear characteristics of the current load, ensuring that the classification accuracy of the model on high samples is improved.

8. The dynamic backfill method based on electric energy metering data restoration model according to any one of claims 1-3, characterized in that, The classification result of the load characteristics is used to collect the same type of resistive load data characteristics within a certain time range under each classification, store by day, and complete the missing data sampling according to the fitted load curve, including: First, data screening is performed, historical data of the same type of resistive load in a period of time is extracted, and sampling points are set after storing by day; At the current time of repair Build a window dataset centered on [the data type]. ,in, For the number of valid data days, Each For the first A 24-hour load data sequence for each valid day; right The mean load and standard deviation are calculated at each time point to obtain a 24-hour periodic baseline curve. With fluctuation range , is represented as: ; wherein, is the effective day at the time of the load value, reflects the 24-hour periodic trend of the resistive load; If the current data to be repaired is at time with missing data, denoted as , then according to the effective data collected on the current day is weighted and corrected, and the completion formula is expressed as: ; wherein, is a weight coefficient, which is dynamically adjusted according to the similarity between the current day and the historical period, to ensure that the completion value not only fits the historical trend, but also adapts to the actual load fluctuation of the current day.

9. The dynamic backfill method based on electric energy metering data restoration model according to any one of claims 1-3, characterized in that, The inductive load data characteristics are compensated and corrected by identifying the power condition of the transient peak, including: Collect real-time current data of inductive load, determine sampling frequency, calculate current change rate of adjacent sampling points: ; wherein T is the sampling period, is the current at time t, When the inductive load starts, it will rise sharply, and when it stops, it will drop sharply; Extract the power data of inductive load in normal operation, calculate the rated power , obtain the power mean value in normal operation period and the power correction coefficient in start-stop transient state, i.e. the start time coefficient and the stop time coefficient , both of which correspond to the compensation proportion of start peak and stop sudden drop respectively; Correction of the starting time: original peak power Including the distortion component caused by the magnetizing inrush, the correction formula is defined as: Correction of the stopping time: the stopping time power drops to 0 or below, the correction formula is: ; wherein, to stop the ramp-down power.

10. The dynamic backfill method based on electric energy metering data restoration model according to any one of claims 1-3, characterized in that, The reverse power correction amount is expressed as: wherein, is the reverse power, is the total power, including the forward and reverse power, is the reverse correction coefficient.

11. The dynamic backfill method based on electric energy metering data restoration model according to claim 10, characterized in that, According to the coupling distortion data restoration model, the electric quantity that needs to be restored is obtained, including: The electric quantity that needs to be corrected for the current interruption caused by dust concentration is expressed as: wherein, is the interruption time, is the total time of the metering period, is the average electric quantity in the period before interruption; A coefficient matrix of the electric energy metering data coupled under the influence of load fluctuation and complex environmental factors is constructed, denoted as: ; wherein, a load fluctuation frequency (times / hour), a temperature change amount, an electromagnetic interference intensity change amount, a dust concentration change amount; The coupled distortion data restoration model is represented as: .

12. The dynamic backfill method based on electric energy metering data restoration model according to claim 11, characterized in that, The electric quantity is compensated for different scenarios, and the total compensation quantity is obtained, including: In a short-term distortion scenario, the short-term distortion is a distortion condition of a distortion data feature less than 24 hours, and the power represented by the backfill is: ; In the long-term distortion scenario, the long-term distortion is a distortion condition of distortion data characteristics reaching more than 24 hours, and the power of the compensation is segmented and compensated by dividing the sub-period, which is expressed as: ; wherein, is the number of sub-periods, is the reduction power of the first sub-period and the measured power. In the scenario of new energy grid connection, the reverse electric quantity compensation is calculated separately, and the formula is as follows: ； In combination with the above scenario, the total amount of the returned and supplied power is obtained, which is expressed as: .

13. The dynamic backfill method based on electric energy metering data restoration model of claim 2, wherein, The classification result of the load characteristics is used to collect the same type of resistive load data characteristics within a certain time range under each classification, store by day, and complete the missing data sampling according to the fitted load curve, including: The distribution basis of the distortion type dominated by load fluctuation is greater than that of the environment interference dominant type and the coupling dominant type; The distortion severity degree based on the distortion data features determines the corresponding weight coefficient, and then obtains the final compensation power, expressed as: ; wherein, The severity coefficients under the three conditions of load fluctuation dominant type, environmental interference dominant type, and coupling dominant type are respectively, The basic weights of the load fluctuation dominant type, the environmental interference dominant type, and the coupling dominant type are respectively.

14. The dynamic backfill method based on electric energy metering data restoration model according to claim 2, characterized in that, The final compensation electric quantity obtained is verified in multiple dimensions, including: substation total meter verification, historical data verification, similar user verification, and line loss verification, wherein, The same kind of user check includes: selecting a user group consistent with the target user in the same area, determining the average refund rate of the group, and then obtaining the deviation rate of the target user refund rate and the average refund rate ; if the deviation rate is less than the set threshold, it is determined that the current refund power meets the same user consistency requirement, otherwise, it is determined that the refund power does not meet the requirement; The historical data verification includes: extracting the historical same period electric quantity data of the target user, the historical same period electric quantity data is the electric quantity data corresponding to the time at least one year ago from the current time as the standard, calculating the fluctuation range of the historical same period electric quantity, and finally judging whether the restored electric quantity obtained according to the coupling distortion data restoration model is within the fluctuation range. If the restored electric quantity is within the fluctuation range, it is determined that the compensated electric quantity conforms to the historical trend; if it is out of range, it needs to be rechecked whether the load characteristic extraction is complete and whether the environmental compensation coefficient is suitable; The total table check of the transformer area includes: calculating the theoretical electric quantity of the total table of the transformer area, determining the difference between the electric quantity of the gateway of the superior power grid and the difference between the theoretical loss of the transformer area line and the total measured electric quantity of the total reading of the transformer area, calculating the deviation rate of the total quantity of user withdrawal and compensation and the difference ; The line loss verification comprises: calculating actual line loss based on the restored total power consumption of users and the theoretical line loss of the transformer area, and calculating the deviation rate between the actual line loss and the theoretical line loss ; If and are within the set threshold, it is determined that the power supply and the grid-side data are consistent, and if any deviation rate exceeds the set threshold, the coefficient matrix of the power metering data coupling needs to be adjusted again.

15. The dynamic backfill method based on electric energy metering data restoration model according to claim 3, characterized in that, Based on the multi-dimensional verification result, a mapping relationship between the deviation level and the model parameter adjustment strategy is constructed, including: If and exceeds the standard, it is identified as a slight deviation, at which time the local parameters are fine-tuned, otherwise, identified as moderate bias, at which point the validation bias sum is minimized based on a gradient descent algorithm iteratively updating coupling coefficients of the coupled distortion data recovery model ; when or the number of iterations is ≥ 50 times, stop; otherwise, Severe bias is identified, and the penalty factor is re-optimized with kernel function parameters , ensuring that the classification accuracy is restored to above the set threshold.

16. A dynamic clawback system based on an electrical energy metering data restoration model, characterized by, The system includes: A data acquisition module is configured to construct a multi-dimensional measurement data feature set: collect load characteristics and determine the corresponding characteristic power flow direction, wherein the load characteristics include: load data characteristics, environmental data characteristics, and distortion data characteristics. The load data characteristics, environmental data characteristics, and distortion data characteristics refer to load characteristics under fluctuation, load characteristics under environmental interference, and load characteristics under both fluctuation and environmental interference, respectively. A sensitivity factor construction module is configured to construct a characteristic sensitivity factor based on the non-linear characteristics of the load characteristics using a box counting method, and to optimize the fitness function in a particle swarm optimization algorithm using the characteristic sensitivity factor, and to determine the optimal combination of penalty factor and kernel function parameter using a support vector machine algorithm, thereby realizing classification of the load characteristics. A compensation module is configured to collect the same type of resistive load data characteristics within a certain time range under each classification based on the classification result of the load characteristics, and to complete the missing data sampling according to the fitted load curve after storing by day. The inductive load data characteristics are compensated and corrected by identifying the power condition of the transient peak. The dynamic compensation module is configured to establish a coupling distortion data restoration model by combining a multiple linear regression to determine a coupling correction coefficient, a reverse electric quantity correction quantity, and an electric quantity that needs to be corrected due to a current interruption caused by dust concentration, to obtain an electric quantity that needs to be restored, to perform electric quantity compensation and restoration for different scenarios, and to obtain a total compensation and restoration quantity.

Images