Intelligent fusion terminal full-temperature error self-calibration method based on digital twin migration

Abstract

The application provides a kind of intelligent fusion terminal full-temperature error self-calibration method based on digital twin migration, belongs to intelligent fusion terminal electric energy metering technical field. Including: selecting sample terminal to carry out full-temperature zone multi-working condition data acquisition;Adopt Gaussian process regression to establish group commonness error baseline model;Extract sample terminal normal temperature individual deviation feature and full-temperature individual deviation label;Based on XGBoost, train normal temperature deviation to full-temperature deviation migration mapping network;Normal temperature single-point self-checking is carried out to terminal to be shipped and normal temperature deviation feature is extracted;Normal temperature deviation feature is input into migration mapping network, and full-temperature individual deviation is predicted;Group baseline and individual deviation are algebraically superimposed, and full-temperature compensation coefficient table is reconstructed and programmed to terminal;When terminal runs, look-up table interpolation is completed according to real-time temperature to complete self-calibration.The application can reconstruct terminal full-temperature compensation curve according to normal temperature single-point data by the fusion of group commonness separation and individual deviation migration mapping.

Description

Technical Field

[0001] This invention relates to the field of intelligent fusion terminal power metering technology, and in particular to a method for full-temperature error self-calibration of intelligent fusion terminals based on digital twin migration. Background Technology

[0002] As a core device in the edge sensing layer of the power distribution Internet of Things (IoT), the intelligent converged terminal's internal sampling and measurement unit is responsible for energy metering and instantaneous quantity calculation. This unit typically consists of a current transformer (CT), a voltage transformer or sampling voltage divider network, an analog-to-digital converter (ADC), and a metering chip. Power industry standards and the JJG 596 metrological verification procedure require that the active energy metering error of the terminal be controlled within specified limits across the entire temperature range of -40℃ to 70℃ and under different load conditions.

[0003] However, due to the discreteness of semiconductor processes and the tolerance of passive devices, the core components of different terminals exhibit significant individualized temperature drift differences under wide temperature conditions, specifically as follows:

[0004] Inconsistent temperature drift of the ADC reference voltage source in sampling chips: The temperature coefficient of the built-in or external bandgap reference voltage source of the metering chip exhibits a Gaussian distribution across a batch of chips. Even within the same batch of chips, the deviation of the reference voltage from the nominal value varies slightly under extreme temperatures of -40℃ or 70℃, directly leading to inter-terminal dispersion of gain error;

[0005] Temperature coefficient of sampling resistor (TCR) discrepancy: The nominal TCR value of the precision voltage divider resistor in the voltage sampling circuit is usually between ±25ppm / ℃ and ±50ppm / ℃, but the individual deviation within the actual batch can reach more than ±10ppm / ℃. In the range of -40℃ to 70℃ (temperature difference of 110℃), the difference in resistance value drift between terminals is significantly amplified, leading to voltage channel ratio inconsistency problems;

[0006] Low-temperature permeability variation of current transformer core material: Current transformers use permalloy or nanocrystalline cores, whose initial permeability decreases nonlinearly at low temperatures (such as -40℃), and the magnitude of the decrease is strongly correlated with the micro-grain structure of the core heat treatment process. This makes it difficult to compensate for the phase difference (phase error) of different terminals under light load and low-temperature coupling conditions using a single preset curve. Summary of the Invention

[0007] In view of this, the present invention provides a self-calibration method for full-temperature error of intelligent fusion terminal based on digital twin transfer. By fusing group commonality separation and individual deviation transfer mapping, the full-temperature compensation parameters of the terminal can be reconstructed based on single-point data at room temperature. This effectively solves the problems of poor full-temperature consistency of batch terminals and low calibration efficiency of traditional full-temperature chambers caused by ignoring the discreteness of components in existing methods.

[0008] The technical solution adopted by the embodiments of the present invention to solve its technical problem is as follows:

[0009] The first aspect of this invention provides a method for full-temperature error self-calibration of an intelligent fusion terminal based on digital twin transfer, comprising:

[0010] Group sample full temperature range multi-condition data acquisition: Preset the full temperature range and multiple test conditions, collect the measurement error values ​​of the same model of intelligent fusion terminal at each temperature-condition combination point, and construct a group sample full temperature error dataset;

[0011] Population common error baseline modeling based on Gaussian process regression: The population error mean corresponding to each temperature-operating condition combination point is calculated based on the population sample full temperature error dataset to construct a training dataset. Gaussian process regression is used to establish a population common error baseline model. Based on the trained population common error baseline model, predictions are made for all operating conditions. The population common error baseline surface function and confidence interval are generated based on the prediction results.

[0012] Individual deviation feature extraction and transfer mapping dataset construction: Based on the population common error baseline model, extract the room temperature individual deviation feature vector and the extreme temperature individual deviation label vector at the preset extreme temperature node for each sample terminal, construct the transfer mapping dataset, and assign the terminals with significant individual deviations selected according to the confidence interval to the training set, validation set and test set.

[0013] Construction and training of the transfer mapping network: The transfer mapping network is constructed using the XGBoost supervised learning framework with extreme gradient boosting. The transfer mapping network is trained using the transfer mapping dataset to derive the full-temperature deviation feature based on the normal temperature deviation feature.

[0014] Room temperature single-point self-test and full temperature deviation migration prediction: The terminal to be shipped is only tested by a standard source under room temperature environment. The room temperature individual deviation feature vector of the terminal to be shipped is extracted and input into the migration mapping network to predict the full temperature deviation feature of the terminal to be shipped.

[0015] Full-temperature compensation coefficient reconstruction and terminal programming operation: Based on the full-temperature deviation characteristics and the common error baseline surface function of the group, the error compensation curve of the terminal to be shipped is reconstructed in the full temperature range, a full-temperature compensation coefficient lookup table is generated and programmed into the MCU of the terminal to be shipped. When the terminal to be shipped is running, the metering error self-calibration is completed by interpolating the table based on the real-time temperature.

[0016] Ideally, the collection of population sample data across multiple operating conditions and within a full temperature range includes:

[0017] N identical intelligent fusion terminals were selected as sample terminals and placed in a high and low temperature programmable test chamber within the temperature range [T]. min,T max [Set temperature point T according to gradient] i After each temperature point is kept constant and stabilized, a variety of typical operating conditions are output sequentially through a standard power source; the typical operating conditions include the output current, output voltage, and power factor of the standard power source.

[0018] The measurement error values ​​of each terminal at each temperature-operating condition combination point are calculated, and a group sample full-temperature error dataset is constructed.

[0019] Preferably, the population common error baseline modeling based on Gaussian process regression includes:

[0020] Construction of a common error dataset and training of a Gaussian process regression model: For each temperature-operating condition combination point, the arithmetic mean of the measurement errors of all sample terminals is calculated as the population error mean to construct a training dataset. The input feature vector includes the effective values ​​of temperature, current, and power factor after Z-score standardization. A Gaussian process regression model is established, using the Matérn 5 / 2 kernel function and introducing an automatic correlation determination ARD mechanism. Independent length scales are assigned to the three input dimensions of temperature, current, and power factor. A hyperparameter vector is defined, and hyperparameter optimization is performed by maximizing the logarithmic marginal likelihood function and using the L-BFGS quasi-Newton algorithm.

[0021] Population common error baseline surface generation and confidence interval construction: The trained Gaussian process regression model is used to predict any target operating point within the entire temperature range. The predicted mean is used as the population common error baseline value of the target operating point. The grid points covering the temperature range, current range and typical power factor are traversed to generate a complete population common error baseline surface and fit it as a population common error baseline surface function. At the same time, the confidence interval of the population baseline is constructed using the prediction variance. Terminals that exceed the confidence interval are identified as terminals with significant individual bias and are preferentially included in the training sample set of the transfer mapping network in subsequent steps.

[0022] Preferably, the construction of the group common error dataset and the training of the Gaussian process regression model include:

[0023] For each temperature-operating condition combination point i, calculate the arithmetic mean of the measurement errors of all sample terminals as the population error mean y. i Extract the feature vector of temperature-operating condition combination point i. Build a training dataset Where M is the total number of temperature-operating condition combination points, , , For the standardized temperature, current, and power factor;

[0024] Establish a Gaussian process regression model (GPR) and define the latent function. Prior to a zero-mean Gaussian process The observation model is ,in It is independent and identically distributed Gaussian noise. To represent the variance of the noise, the kernel function for Gaussian process regression is the Matérn 5 / 2 kernel function, and an automatic correlation determination ARD mechanism is introduced:

[0025] ;

[0026] ;

[0027] in For signal variance, , , These are the length-scale parameters for temperature, current, and power factor, respectively.

[0028] Define hyperparameter vector The GPR model is optimized by maximizing the logarithmic marginal likelihood function:

[0029] ;

[0030] In the formula, G represents the identity matrix, and K is an M×M covariance matrix; the elements in K It represents the covariance between eigenvectors.

[0031] Preferably, the generation of the common error baseline surface and the construction of the confidence interval include:

[0032] After hyperparameter optimization converges, the trained GPR model is used to analyze the temperature range [T]. min ,T max Any target working point within ] Make a prediction, predict the mean. With prediction variance They are respectively:

[0033] ;

[0034] ;

[0035] In the formula, This is the covariance vector between the test point and all other points;

[0036] To predict the mean As the common error baseline value of the group at the operating point*, it traverses the temperature range [T] min ,T max ], current range [I min ,Imax and typical power factor [φ] min ,φ max The grid points are fitted to generate a baseline surface function for the common error of the population. ;

[0037] Using prediction variance Construct confidence intervals for the population baseline:

[0038] ;

[0039] Terminals that exceed the confidence interval are marked as terminals with significant individual bias.

[0040] Preferably, the ambient temperature individual deviation feature vector is a five-dimensional vector, including voltage ratio difference deviation, current ratio difference deviation, active power error deviation, active power error deviation under rated operating conditions at ambient temperature, and active power error deviation under light load conditions; the extreme temperature individual deviation label vector is the individual deviation value of each preset extreme temperature node; the transfer mapping dataset is randomly divided into training set, validation set and test set according to proportion, wherein all terminals with significant individual deviations are preferentially and evenly distributed to the three sets;

[0041] The construction and training of the transfer mapping network includes:

[0042] A multi-output regression strategy was adopted to construct five independent XGBoost single-output regression sub-models, each corresponding to W preset extreme temperature nodes. The objective function of each XGBoost single-output regression sub-model consists of a squared error loss function and a regularization term.

[0043] Each XGBoost single-output regression sub-model was trained using the training set, the validation set was used for early stopping monitoring, and the test set was used for testing.

[0044] The objective function of the XGBoost single-output regression sub-model is:

[0045] ;

[0046] In the formula, The model predicts the value; N is the number of training samples, T is the total number of trees; Lt is the number of leaf nodes in the t-th tree. Let be the weight vector of the leaf nodes of the t-th tree. Leaf quantity penalty coefficient, To be related to the L2 regularization coefficient, is the L1 regularization coefficient.

[0047] Preferably, the single-point self-test at room temperature and the prediction of deviation migration at full temperature include:

[0048] Connect a standard power source at room temperature to perform a routine accuracy check on the terminal to be shipped, and collect the room temperature measurement error value.

[0049] The predicted value of the current temperature-operating point is calculated based on the baseline surface function of the common error of the group, and the difference is taken with the room temperature measurement error value to obtain the room temperature individual deviation feature vector. ;

[0050] Will Input the migration mapping network to obtain the full-temperature deviation characteristics of the terminal to be shipped.

[0051] Preferably, the full-temperature compensation coefficient reconstruction and terminal programming operation includes:

[0052] The predicted values ​​of each point in the full temperature range are calculated based on the common error baseline surface function of the group. Combined with the full temperature deviation characteristics, the corresponding sum values ​​are calculated one by one according to the temperature points.

[0053] Based on the correspondence between temperature points and sum values, an error curve is fitted, which serves as the error compensation curve for the terminal to be shipped across the entire temperature range.

[0054] The correction coefficients corresponding to each temperature node are calculated in reverse based on the error compensation curve, and a full-temperature compensation coefficient lookup table is generated. This table is then burned into the memory of the MCU of the terminal to be shipped through the communication interface. The MCU can then obtain the temperature of the sampling module in real time through the built-in temperature sensor, perform piecewise linear interpolation in the table based on the current temperature, and write the interpolation results into the corresponding correction register of the metering chip to complete the dynamic self-calibration of the full-temperature range power metering.

[0055] As can be seen from the above technical solution, the present invention provides a method for self-calibrating the full-temperature error of an intelligent fusion terminal based on digital twin transfer. The method includes: firstly, selecting sample terminals for multi-condition data acquisition across the entire temperature range; establishing a common error baseline model for the group using Gaussian process regression; extracting individual deviation features at room temperature and individual deviation labels for the full temperature range of the sample terminals; training a transfer mapping network from room temperature deviation to full temperature deviation based on XGBoost; performing a single-point self-test at room temperature on the terminal to be shipped and extracting room temperature deviation features; inputting the room temperature deviation features into the transfer mapping network to predict individual deviations at the full temperature range; algebraically superimposing the group baseline and individual deviations to reconstruct the full-temperature compensation coefficient table and burning it into the terminal MCU; and completing the self-calibration of measurement error by interpolating the table based on the real-time temperature during terminal operation. This invention, through the fusion of group commonality separation and individual deviation transfer mapping, can reconstruct the full-temperature compensation parameters of the terminal based on single-point data at room temperature, effectively solving the problems of poor full-temperature consistency of batch terminals and low efficiency of traditional full-temperature chamber calibration caused by ignoring the discreteness of components in existing methods. Attached Figure Description

[0056] Figure 1This is a flowchart of the full-temperature error self-calibration method for intelligent fusion terminals based on digital twin migration according to the present invention. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0058] This invention proposes a self-calibration method for full-temperature error of intelligent fusion terminals based on digital twin transfer. Its core lies in constructing a two-level cascaded architecture of group common baseline modeling and individual deviation transfer mapping. This method first establishes independent models for the two major components affecting the full-temperature measurement error of the terminal: the group systematic error determined by the measurement principle and common components, and the individual deviation caused by hardware discreteness. A batch of sample terminals are selected and subjected to comprehensive testing across the entire temperature range (-40℃ to 70℃) and multiple operating conditions (different currents and power factors). The collected multi-dimensional operating condition parameters, such as temperature, current, and power factor, are input into a Gaussian process regression model. Through probabilistic modeling, a group common error baseline surface containing mean prediction and confidence intervals is output, achieving statistical separation of common patterns and individual differences. Simultaneously, the deviation feature vectors of each sample terminal at a single point in normal temperature (25℃), such as voltage ratio difference and current angle difference, are extracted as input. The individual deviation values ​​of the terminal at extreme temperature nodes (-40℃, 70℃, etc.) are used as labels to train a transfer mapping network based on the XGBoost supervised learning framework. This network learns the nonlinear transfer function of "local deviation in normal temperature → global deviation in all temperature range", accurately capturing the coupling mapping relationship of peripheral hardware factors such as the discrete temperature coefficient of the sampling resistor, the difference in the permeability of the current transformer core, and the dispersion of the temperature drift of the ADC reference voltage source under temperature stress. In the mass production stage, the terminal to be shipped only needs to be calibrated once in a standard source environment in a normal temperature environment to extract its individual deviation features in normal temperature and input them into the trained transfer mapping network. The network automatically infers and predicts the individual deviation values ​​of the terminal at key nodes in the entire temperature range. The host computer algebraically superimposes the common baseline value of the group with the predicted individual deviation value to reconstruct the complete error compensation curve of the terminal in the entire temperature range, and then generates the gain correction coefficient and phase correction coefficient and burns them into the terminal MCU memory. During on-site operation, the MCU dynamically self-calibrates measurement errors by using piecewise linear interpolation to look up table values ​​based on real-time sampling values ​​from the built-in temperature sensor. This architecture achieves a technological breakthrough by jointly optimizing group baseline separation and migration feature mapping in stages. It reconstructs the full-temperature compensation curve from single-point data at room temperature, while also endowing the system with adaptive modeling capabilities for the batch discreteness of components, fundamentally solving the problem of consistent full-temperature measurement in batch terminals.

[0059] refer to Figure 1 As shown, the full-temperature error self-calibration method for intelligent fusion terminals based on digital twin migration provided by the present invention includes the following implementation steps:

[0060] Step S1, Collect data of the whole temperature range and multiple working conditions of the group sample: Preset the whole temperature range and multiple sets of test conditions, collect the measurement error values ​​of the same model of intelligent fusion terminal at each temperature-working condition combination point, and construct the whole temperature error dataset of the group sample.

[0061] Step S2, Modeling the Common Error Baseline of the Population Based on Gaussian Process Regression: Calculate the mean population error corresponding to each temperature-operating condition combination point based on the full temperature error dataset of the population sample, construct a training dataset, establish a common error baseline model of the population using Gaussian process regression, predict all operating conditions based on the trained common error baseline model of the population, and generate the common error baseline surface function and confidence interval of the population based on the prediction results.

[0062] Step S3, Individual Deviation Feature Extraction and Transfer Mapping Dataset Construction: Based on the population common error baseline model, extract the room temperature individual deviation feature vector and the extreme temperature individual deviation label vector at the preset extreme temperature node for each sample terminal, and construct the transfer mapping dataset.

[0063] Step S4, Construction and training of transfer mapping network: The transfer mapping network is constructed using the Extreme Gradient Boosting (XGBoost) supervised learning framework, and the transfer mapping network is trained using the transfer mapping dataset to derive the full-temperature deviation feature based on the normal temperature deviation feature;

[0064] Step S5, single-point self-test at room temperature and full-temperature deviation migration prediction: the terminal to be shipped is only tested by a standard source under room temperature, the room temperature individual deviation feature vector of the terminal to be shipped is extracted and input into the migration mapping network to predict the full-temperature deviation features of the terminal to be shipped.

[0065] Step S6, Full-temperature compensation coefficient reconstruction and terminal programming operation: Based on the full-temperature deviation characteristics and the common error baseline surface function of the group, the error compensation curve of the terminal to be shipped is reconstructed in the full temperature range, a full-temperature compensation coefficient lookup table is generated and programmed into the MCU of the terminal to be shipped. When the terminal to be shipped is running, the metering error self-calibration is completed by interpolating the table based on the real-time temperature.

[0066] Preferably, step S1, the collection of multi-condition data for the entire temperature range of the population sample, includes:

[0067] N (N≥50) identical intelligent fusion terminals were selected as modeling samples and placed in a high and low temperature controlled test chamber. The temperature range was [T...]. min ,T max [Set temperature points T according to gradient] i For example, within the temperature range of -40℃ to 70℃, temperature points are set according to a preset gradient (e.g., every 10℃ step), and each temperature point T... iAfter the temperature stabilizes, various typical operating conditions (including rated voltage / current, light load 0.1Ib, power factor 1.0 / 0.5L / 0.8C, etc.) are sequentially output through a standard power source. The host computer polls and collects the measured values ​​of voltage, current, active power, etc., output by the metering chips of each terminal, synchronously records the corresponding true values ​​of the standard source, records the metering error values ​​of each terminal at each temperature-operating condition combination point, and constructs a group sample full-temperature error dataset.

[0068] The metering error values ​​of each terminal at each temperature-operating condition combination point were calculated, and a group sample full-temperature error dataset was constructed. The metering error values ​​include active energy metering error, and the calculation formula is as follows:

[0069] (1)

[0070] in For measuring active power at the terminal, The standard source outputs active power.

[0071] Preferably, the population common error baseline modeling based on Gaussian process regression described in step S2 includes:

[0072] Step S21, Construction of the Population Common Error Dataset and Training of the Gaussian Process Regression Model: For each temperature-operating condition combination point, calculate the arithmetic mean of the measurement error of all sample terminals as the population error mean, construct the training dataset, and the input feature vector includes the effective values ​​of temperature, current and power factor after Z-score standardization; establish a Gaussian process regression model, select the Matérn 5 / 2 kernel function and introduce an automatic correlation determination ARD mechanism, and assign independent length scales to the three input dimensions of temperature, current and power factor; define the hyperparameter vector, and optimize the hyperparameters by maximizing the logarithmic marginal likelihood function and using the L-BFGS quasi-Newton algorithm.

[0073] The measurement error data of the N sample terminals collected in step S1 under multiple operating conditions (different currents and power factors) across the full temperature range (-40℃ to 70℃) are statistically summarized. For each temperature-operating condition combination point, the arithmetic mean of the measurement errors of all N sample terminals is calculated as the population error observation value for that point, and its standard deviation is calculated to characterize the sample dispersion. Z-score standardization is applied to the input features of each dimension under multiple operating conditions to eliminate the influence of numerical differences on model training.

[0074] ; ; (2)

[0075] The values ​​for temperature, current, and power factor under various operating conditions are provided. This represents the average values ​​of temperature, current, and power factor under all operating conditions. The standard deviations of temperature, current, and power factor under all operating conditions. These are the standardized values ​​for temperature, current, and power factors.

[0076] For each temperature-operating condition combination point i, calculate the arithmetic mean of the measurement errors of all sample terminals as the population error mean y. i (Arithmetic mean of measurement errors from N terminals) Extract the feature vector of temperature-operating condition combination point i. Build a training dataset Where M is the total number of temperature-operating condition combination points (M≥200).

[0077] Establish a Gaussian process regression model (GPR) and define the latent function. Prior to a zero-mean Gaussian process The observation model is ,in It is independent and identically distributed Gaussian noise. To represent the variance of the noise, the kernel function for Gaussian process regression is the Matérn 5 / 2 kernel function, and an automatic correlation determination ARD mechanism is introduced:

[0078] (3)

[0079] (4)

[0080] in For signal variance, , , These are length scale parameters for temperature, current, and power factor, respectively, representing the sensitivity of the output in each dimension;

[0081] Define hyperparameter vector The GPR model is optimized by maximizing the logarithmic marginal likelihood function:

[0082] (5)

[0083] In the formula, G represents the identity matrix, and K is the M×M covariance matrix. Here, M represents the noise variance, M represents the number of training samples, and K represents the elements in K. This represents the covariance between eigenvectors. Hyperparameter optimization employs the L-BFGS quasi-Newton algorithm, with a maximum of 200 iterations and a convergence tolerance of [missing information]. Initial signal variance Set to 1.0, noise variance Set as To avoid local optima, the initial value of the length scale is uniformly selected in the interval [0.1, 10].

[0084] Step S22, Generation of the common error baseline surface and construction of confidence intervals for the population:

[0085] After hyperparameter optimization converges, the trained GPR model is used to analyze the temperature range [T]. min ,T max Any target working point within ] Make a prediction, predict the mean. With prediction variance They are respectively:

[0086] (6)

[0087] (7)

[0088] In the formula, This is the covariance vector between the test point and all other points;

[0089] To predict the mean As the common error baseline value of the group at the operating point*, it traverses the temperature range [T] min ,T max ], current range [I min ,I max and typical power factor [φ] min ,φ max The grid points are fitted to generate a baseline surface function for the common error of the population. The temperature range is [-40℃, 70℃], the current range is [0.05Ib, 4Ib], and the typical power factor is [1.0, 0.5L].

[0090] Using prediction variance Construct confidence intervals for the population baseline:

[0091] (8)

[0092] In this invention, the confidence interval is a 95% confidence interval. Terminals exceeding this confidence interval are marked as terminals with significant individual bias. This confidence interval quantifies the statistical uncertainty of the mean error of the group sample at a given operating point. For any single terminal, if its measured error falls within this confidence interval, its error characteristics are considered to be within the normal fluctuation range of the group; if it exceeds the boundary of the confidence interval, the terminal is determined to have significant individual bias and needs to be included in the training sample set of the transfer mapping network for priority processing in subsequent step S3 to enhance the model's ability to fit extreme discrete samples. Simultaneously, this confidence interval is stored in the host computer database as a statistical criterion for terminal quality screening during mass production.

[0093] Preferably, the specific implementation of step S3, individual bias feature extraction and transfer mapping dataset construction, includes:

[0094] The Gaussian process regression population common baseline model trained in step S2 For each sample terminal j (j=1,2,…,N), extract its selected set from the normal temperature of 25℃ and other temperature nodes. The predicted baseline values ​​for the population at the following temperatures are given: -40℃, -20℃, 0℃, 55℃, 70℃. Under fixed typical operating conditions (e.g., rated current Ib, power factor 1.0), the population baseline values ​​for each node are calculated separately.

[0095] (9)

[0096] The measured error value of the terminal under the same working conditions The difference between the individual and the group baseline is defined as the individual deviation value:

[0097] (10)

[0098] The individual bias feature vector at room temperature (25℃) is extracted as the input feature vector of the transfer mapping network, forming a multidimensional feature vector:

[0099] (11)

[0100] in:

[0101] Indicates voltage channel ratio deviation; Indicates the current channel ratio deviation; This indicates the active power error deviation; This indicates the active power error deviation under light load conditions; This indicates the active power error deviation under inductive operating conditions;

[0102] Extract individual deviation values ​​from the preset extreme temperature node set Text, and construct the output label vector of the transfer mapping network:

[0103] (12)

[0104] As an extreme temperature individual bias label vector, it consists of the individual bias values ​​of each preset extreme temperature node; each component is represented as a percentage (%) and is Z-score standardized to eliminate the influence of differences in numerical ranges of different dimensions on model training.

[0105] By iterating through all N sample terminals, combining the input feature vector and the output label vector, a transfer mapping dataset is constructed. The dataset is randomly divided into training, validation, and test sets according to a preset ratio (e.g., 7:2:1). Terminals with significant individual biases are preferentially and evenly distributed among the three sets. During the data partitioning process, the statistical representativeness of the terminal error distribution within each subset is maintained. Stratified sampling is used to ensure that terminals with significant biases are covered in each subset, thereby enhancing the fitting and generalization ability of the subsequent transfer mapping network for terminals with large discreteness.

[0106] Step S4, the specific implementation of construction and training of the transfer mapping network, includes:

[0107] Step S41: XGBoost model architecture and multi-output regression strategy configuration:

[0108] A transfer mapping network based on the XGBoost (eXtreme Gradient Boosting) supervised learning framework is constructed. XGBoost iteratively adds weak learners (CART regression trees) to gradually fit the predicted residuals. For the input feature vector... (D=5, including room temperature voltage ratio deviation, current ratio deviation, active power deviation, light load deviation, and inductive deviation), output label vector (W=5, corresponding to the individual deviation values ​​of five temperature nodes: -40℃, -20℃, 0℃, 55℃, and 70℃).

[0109] A multi-output regression strategy is employed to construct five independent XGBoost single-output regression sub-models, each corresponding to W preset extreme temperature nodes. The objective function of each XGBoost single-output regression sub-model consists of a squared error loss function and a regularization term. The objective function of the XGBoost single-output regression sub-model is as follows:

[0110] (13)

[0111] In the formula, The model predicts the value; N is the number of training samples, T is the total number of trees; Lt is the number of leaf nodes in the t-th tree. Let be the weight vector of the leaf nodes of the t-th tree. Leaf quantity penalty coefficient, To be related to the L2 regularization coefficient, The L1 regularization coefficient is... This is the L1 regularization term. The specific hyperparameter configurations for each XGBoost sub-model are as follows: learning rate 0.05, maximum tree depth 4, subsampling ratio 0.8, column sampling ratio 0.8, minimum leaf weight sum 3, L2 regularization coefficient 1.0, L1 regularization coefficient 0.1, and maximum number of trees 200.

[0112] S42: Model Training and Performance Validation

[0113] The transfer mapping training set constructed using step S3 Model training and validation set Used for early stopping monitoring. The training process employs incremental gradient boosting: in the t-th iteration, based on the previous... The prediction residuals of the ensemble model are fitted to the t-th regression tree, and the optimal split point is searched using a greedy algorithm to minimize the objective function gain. The mean absolute percentage error (MAPE) on the validation set is evaluated every 10 training epochs. If the validation set error does not decrease for 20 consecutive epochs, training is terminated and the model is rolled back to the state corresponding to the optimal number of iterations. MAPE is defined as:

[0114] (14)

[0115] in The number of validation set samples is K=5, which represents the number of temperature nodes. To prevent division by zero smoothing constants.

[0116] After training, on the test set The prediction accuracy of the transfer mapping network is evaluated using the following three metrics:

[0117] Root mean square error:

[0118] (15)

[0119] Mean absolute error:

[0120] (16)

[0121] Coefficient of determination:

[0122] (17)

[0123] Each XGBoost single-output regression sub-model was trained separately using the training set, and the validation set was used for early stopping monitoring. The mean absolute percentage error (MAPE) on the validation set was evaluated every few training epochs; training was terminated if the MAPE did not decrease for several consecutive epochs. After training, the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (COP) were evaluated on the test set. Models were considered successful if they met preset thresholds. The early stopping strategy involved evaluating the MAPE on the validation set every 10 training epochs; training was terminated if the MAPE did not decrease for 20 consecutive epochs. The model acceptance criterion was the COP. If the mean absolute error of each temperature node is less than 0.05% and the value is greater than 0.85, the transfer mapping network is considered to have passed training and is ready for mass production deployment. If it does not meet the criteria, return to step S1 to supplement the number of sample terminals or adjust the hyperparameter configuration in step S41 to retrain.

[0124] Model weight solidification and deployment storage: After the model training is successful, the XGBoost sub-model weight parameters corresponding to the five temperature nodes are exported as binary model files. Each model file contains: decision tree structure information (split feature index, split threshold, leaf node weights) and hyperparameter configuration metadata. All model files are packaged and stored in a designated directory on the host computer calibration platform for use in the mass production stage (step S5).

[0125] Meanwhile, to facilitate table lookup compensation on the embedded end, in a preferred embodiment, the trained transfer mapping network is used to predict the full-temperature deviation of all terminals in the training set. After the input feature space is gridded, the corresponding full-temperature deviation prediction value is pre-calculated, a feature-deviation mapping lookup table is constructed and burned into the terminal MCU, further reducing the online inference computation overhead.

[0126] During mass production, each terminal device does not need to undergo a high / low temperature chamber before shipment. It only needs to be connected to a standard power source at room temperature (25℃) for a routine accuracy calibration, and its room temperature measurement error value needs to be collected. Step S5, which involves room temperature single-point self-checking and full-temperature deviation migration prediction, includes the following specific implementation steps:

[0127] A standard power source is connected at room temperature to perform a routine accuracy check on the terminal to be shipped, and the room temperature measurement error value is collected. The test conditions for room temperature single-point self-test include: rated voltage, rated current, and power factor of 1.0. The collected measurement error constitutes a five-dimensional room temperature individual deviation feature vector, which is standardized and input into five XGBoost sub-models for parallel inference, and outputs the individual deviation prediction values ​​of five temperature nodes.

[0128] The predicted value of the current temperature-operating point is calculated based on the common error baseline surface function of the group, and the difference is taken with the normal temperature measurement error value. That is, the normal temperature measurement error value of the terminal is substituted into the input. Subtracting the predicted values ​​at 25℃ yields the individual deviation feature vector at room temperature. ;

[0129] Will Input the XGBoost transfer mapping network trained in step S4, and the network will automatically infer and output the full-temperature deviation features of the terminal at key temperature nodes such as -40℃ and 70℃. The full-temperature deviation characteristics of the terminal to be shipped are obtained.

[0130] The specific implementation of step S6, which involves reconstructing the full-temperature compensation coefficient and programming the terminal, includes:

[0131] The host computer software uses the common error baseline surface function of the group. Calculate the predicted values ​​for each point across the entire temperature range, and combine them with the characteristics of the full-temperature deviation. Calculate the corresponding sum value for each temperature point;

[0132] Based on the correspondence between temperature points and sum values, an error curve is fitted, which serves as the error compensation curve for the terminal to be shipped across the entire temperature range.

[0133] The correction coefficients for each temperature node are calculated in reverse based on the error compensation curve (and written to the metering chip register), generating a full-temperature compensation coefficient lookup table. This table is then burned into the Flash memory of the terminal MCU to be shipped via the communication interface. When the terminal is running in the field, the MCU acquires the temperature of the data acquisition module in real time through its built-in temperature sensor. Based on the current temperature, it performs piecewise linear interpolation in the table and writes the interpolation results into the corresponding correction register of the metering chip, completing the dynamic self-calibration of energy metering across the entire temperature range.

[0134] When the component batch changes, only a small number of new batch samples need to be collected to incrementally fine-tune the migration mapping network, without repeating the full temperature range traversal test.

[0135] This invention provides a full-temperature error self-calibration system for an intelligent fusion terminal based on digital twin migration, comprising:

[0136] Computing terminal, used to execute Figure 1The proposed scheme specifically involves: 1) Performing multi-condition data acquisition across the entire temperature range of a group sample: Pre-setting the entire operating temperature range and multiple test conditions, collecting measurement error values ​​of the same model of intelligent fusion terminals at various temperature-condition combinations, and constructing a group sample full-temperature error dataset; 2) Modeling a common error baseline for the group based on Gaussian process regression: Calculating the mean group error corresponding to each temperature-condition combination point based on the group sample full-temperature error dataset, constructing a training dataset, establishing a common error baseline model using Gaussian process regression, predicting all conditions based on the trained common error baseline model, and generating the common error baseline surface function and confidence interval based on the prediction results; 3) Extracting individual deviation features and constructing a transfer mapping dataset: Based on the aforementioned common error baseline model, extracting the room-temperature individual deviation feature vector and... A transfer mapping dataset is constructed using the extreme temperature individual deviation label vectors at preset extreme temperature nodes. The transfer mapping network is then constructed and trained using the XGBoost supervised learning framework and the dataset, enabling the derivation of full-temperature deviation features based on room temperature deviation characteristics. For room temperature single-point self-testing and full-temperature deviation transfer prediction, the terminal to be shipped is only subjected to standard source verification under room temperature conditions. The room temperature individual deviation feature vectors of the terminal to be shipped are extracted and input into the transfer mapping network to predict the full-temperature deviation characteristics of the terminal. Finally, full-temperature compensation coefficient reconstruction and terminal programming are performed: based on the full-temperature deviation characteristics and the common error baseline surface function of the group, the error compensation curve of the terminal to be shipped is reconstructed across the entire temperature range, a full-temperature compensation coefficient lookup table is generated, and the table is programmed into the MCU of the terminal to be shipped.

[0137] The intelligent fusion terminal performs self-calibration of measurement errors by looking up and interpolating real-time temperature data during operation.

[0138] This invention provides an intelligent fusion terminal, which includes an MCU and deploys a full-temperature compensation coefficient lookup table. The full-temperature compensation coefficient lookup table is generated using the aforementioned method. The MCU obtains the temperature of the cross-collection module in real time through a built-in temperature sensor, performs piecewise linear interpolation in the table based on the current temperature, and writes the interpolation result into the corresponding correction register of the metering chip to complete the dynamic self-calibration of power metering across the entire temperature range.

[0139] The following is a specific embodiment for testing and evaluating the effectiveness of this solution:

[0140] Figure 1 The flowchart of the present invention is as follows: The method of the present invention is: a self-calibration method for full-temperature error of an intelligent fusion terminal based on digital twin transfer, the specific implementation process of which is as follows:

[0141] S1. Group Sample Data Acquisition Across Multiple Temperature Ranges and Operating Conditions: Eighty intelligent fusion terminals of the same model were selected as modeling samples, numbered 1 to 80, and placed in a high and low temperature controlled test chamber. The temperature setting range was -40℃ to 70℃, with 12 temperature points in 10℃ increments. After each temperature point reached the set value, it was kept at a constant temperature for 1 hour to ensure that the temperature of the data acquisition module inside the terminal was fully balanced with the ambient temperature inside the chamber. At each temperature point, the following six typical operating condition combinations were sequentially output through a standard power source:

[0142] Operating condition 1: U=220V, I=6A, PF=1.0;

[0143] Operating condition 2: U=220V, I=6A, PF=0.5L;

[0144] Operating condition 3: U=220V, I=1.5A, PF=1.0;

[0145] Operating condition 4: U=220V, I=1.5A, PF=0.5L;

[0146] Operating condition 5: U=220V, I=0.075A, PF=1.0;

[0147] Operating condition 6: U=220V, I=0.075A, PF=0.5L;

[0148] The host computer polls and reads the RMS voltage, RMS current, and active power measurements of each terminal via RS485 bus, synchronously records the true values ​​corresponding to the standard power source, and calculates the active power metering error e. Each terminal repeatedly samples three times at each temperature-operating condition combination point, and the average value is taken as the measured error value for that point. Finally, a group sample full-temperature error dataset is constructed, containing 80×12×6=5760 valid records.

[0149] S2. Baseline modeling of common population error based on Gaussian process regression:

[0150] The error data collected in step S1 is aggregated at operating point: For each temperature-operating point combination, the arithmetic mean y and standard deviation of the measurement errors of 80 terminals are calculated. Constructing the training dataset Where M = 576 (the unique combination number after averaging 12 temperature points × 6 operating conditions × 8 repetitions). The standardized parameter is: mean temperature. ℃, standard deviation ℃; average current A. Standard deviation A; Mean power factor Standard deviation .

[0151] Establish a Gaussian process regression model and define the latent function. Prior to a zero-mean Gaussian process The observation model is , The Matérn5 / 2 kernel function was selected, and an Automatic Correlation Determination (ARD) mechanism was introduced. The kernel function was defined, and the L-BFGS quasi-Newton algorithm was used for iterative solution. The maximum number of iterations was 200, and the convergence tolerance was [not specified]. The initial hyperparameters are set as follows: , , After optimization and convergence, the hyperparameter values ​​are obtained: , , The length scale of the temperature dimension The smallest value indicates that the error is most sensitive to temperature changes.

[0152] Generate the baseline surface and construct the confidence interval for the common error of the population, and use the trained GPR model to apply it to any target operating point across the entire temperature range. Make a prediction to predict the mean. This serves as the baseline value for the common error of the population at this operating condition. A complete common error baseline surface is generated by traversing grid points within the temperature range [-40℃, 70℃] with a step size of 5℃, current values ​​[0.075A, 1.5A, 6A], and power factor values ​​[1.0, 0.5L]. Simultaneously, construct a 95% confidence interval;

[0153] The confidence interval is stored in the host computer database. In this embodiment, under the conditions of -40℃, current 0.075A, and PF=0.5L, the population baseline prediction error is -0.79%, and the confidence interval half-width is 0.14%. Statistically, among the 80 sample terminals, 6 terminals had measured errors exceeding the lower limit of the confidence interval (<-0.93%) at this operating condition, and were marked as terminals with significant individual bias, and were preferentially included in the transfer mapping network training set in subsequent S3.

[0154] S3. Individual Bias Feature Extraction and Transfer Mapping Dataset Construction:

[0155] Based on the common baseline model of the population For each sample terminal j (j=1,2,…,80), extract its set of nodes at room temperature (25℃) and preset extreme temperatures. Predicted baseline values ​​for the population at the following temperatures: -40℃, -20℃, 0℃, 55℃, 70℃. Under fixed rated operating conditions of current In = 1.5A and power factor 1.0, calculate the population baseline values ​​for each node: ;

[0156] The measured error value of terminal j under the same operating conditions The difference from the group baseline value is defined as individual bias: ;

[0157] For each terminal, construct a five-dimensional input feature vector for the transfer mapping network. All components are expressed as percentages:

[0158] Voltage ratio deviation: Voltage measurement deviation at 25℃, U=220V, In=1.5A, PF=1.0;

[0159] Current ratio difference deviation: Same as the current measurement deviation under the above operating conditions;

[0160] Active power error deviation: Same as the active power error deviation under the above operating conditions;

[0161] Active power error deviation under light load conditions: active power error deviation at 25℃, U=220V, I=0.075A, PF=1.0;

[0162] Active power error deviation under inductive operating conditions: active power error deviation at 25℃, U=220V, I=1.5A, PF=0.5L;

[0163] Simultaneously construct a five-dimensional output label vector. ,correspond Individual deviation values ​​of five temperature nodes under rated operating conditions (1.5A, PF=1.0). Z-score normalization was performed on the feature vectors and label vectors of all 80 terminals, and the normalization parameters were obtained from the training set.

[0164] Combine feature vectors and label vectors to construct a transfer mapping dataset The samples were randomly divided into a training set (56 samples), a validation set (16 samples), and a test set (8 samples) in a ratio of 7:2:1. Stratified sampling was used to ensure that each subset contained at least one terminal with the aforementioned significant bias.

[0165] S4. Training of the individual bias transfer mapping network based on XGBoost:

[0166] S41: XGBoost Model Architecture and Multi-Output Regression Strategy Configuration

[0167] A transfer mapping network was constructed using XGBoost version 1.7.3. Five independent XGBoost single-output regression sub-models were built, corresponding to the five temperature nodes of -40℃, -20℃, 0℃, 55℃, and 70℃, respectively. The objective function of each sub-model was defined, and the hyperparameters were uniformly configured as shown in Table 1.

[0168] Table 1

[0169] S42: Model Training and Performance Validation

[0170] Each sub-model is trained separately using the training set, and the validation set is used for early stopping monitoring. The MAPE (Magnetic Performance Equation) of the validation set is evaluated every 10 training epochs; if it does not decrease for 20 consecutive epochs, early stopping is triggered. The MAPE calculation formula is: (18)

[0171] Each sub-model triggered early stopping within rounds 120-160. The test set evaluation results are shown in Table 2.

[0172] Table 2

[0173] The model training was deemed successful because the MAE of all temperature nodes was less than 0.05% and the R² was greater than 0.85.

[0174] S43: Model weight persistence and deployment storage:

[0175] The decision tree structures (split feature indices, split thresholds, leaf node weights) and hyperparameter configurations of the five sub-models were exported as JSON format, compressed into binary files, and stored in a specified directory on the host computer calibration platform. Simultaneously, to facilitate table lookup and compensation on the embedded end, the input feature space was gridded into 10 equal parts along each dimension. The trained transfer mapping network was used to pre-calculate the full-temperature deviation prediction value corresponding to each grid point, constructing a system of size [size missing]. The feature-deviation mapping lookup table is used for direct programming to the terminal MCU in subsequent step S6.

[0176] S5. Room Temperature Single-Point Self-Test and Full-Temperature Deviation Migration Prediction: During mass production, take one terminal to be shipped and connect it to a standard power source at a room temperature of 25℃. Test the following three characteristic operating conditions sequentially and collect the measurement errors:

[0177] Operating condition A (rated): U=220V, I=1.5A, PF=1.0;

[0178] The measured values ​​were: voltage ratio deviation +0.05%, current ratio deviation -0.02%, and active power error deviation -0.03%.

[0179] Operating condition B (light load): U=220V, I=0.075A, PF=1.0;

[0180] Measured: Active power error deviation +0.03%;

[0181] Operating condition C (inductive): U=220V, I=1.5A, PF=0.5L;

[0182] The measured active power error deviation was -0.05%.

[0183] The above five values ​​are used to construct a feature vector of individual deviation at room temperature:

[0184] (unit:%)

[0185] After Z-score normalization, the five XGBoost sub-models trained in step S4 are input. Each sub-model infers in parallel, outputting the predicted individual bias values ​​for the terminal at the five temperature nodes.

[0186] -40℃: Prediction bias -0.25%;

[0187] -20℃: Prediction bias -0.16%;

[0188] 0℃: Prediction bias -0.06%;

[0189] 55℃: Prediction bias +0.08%;

[0190] 70℃: Prediction bias +0.17%;

[0191] S6. Full-Temperature Compensation Coefficient Reconstruction and Terminal Programming: The host computer software algebraically adds the group common baseline value generated in step S2 to the individual deviation value predicted in step S5, based on temperature nodes. The reconstruction error values ​​for all temperature nodes and six operating conditions are calculated sequentially to form a full-temperature error surface. The correction coefficient of the metering chip is then calculated in reverse based on this error surface.

[0192] After the terminal is put into field operation, the MCU reads the terminal temperature every 30 seconds. When the temperature change exceeds ±2℃, the MCU performs bilinear interpolation based on the current temperature value in the compensation coefficient lookup table to obtain the corresponding correction coefficient, and transmits it to the metering chip through the communication interface to complete the self-calibration.

[0193] Compared with the prior art, the present invention has the following advantages:

[0194] (1) Breaking through the bottleneck of mass production calibration efficiency and eliminating dependence on high and low temperature chamber capacity. Traditional full-temperature compensation methods require cyclic testing of each terminal in a high and low temperature chamber with a full temperature gradient from -40℃ to 70℃, which takes 8 to 12 hours per unit. This invention only requires one standard source calibration at room temperature (25℃), and the compensation coefficient of the terminal in the full temperature range can be reconstructed through the inference of the previously trained transfer mapping network. The calibration time per unit is reduced to the minute level, the production capacity is significantly improved, and the equipment investment and energy consumption costs are effectively reduced.

[0195] (2) Fundamentally solves the problem of consistent metering across all temperatures for batch terminals. Existing group average curve compensation methods ignore the individual discreteness of components, resulting in high error rates for batch terminals under low-temperature light-load conditions of -40℃ and high-temperature heavy-load conditions of 70℃. This invention separates the common baseline of the group through Gaussian process regression and uses XGBoost transfer mapping network to learn the nonlinear transmission law of "local deviation at room temperature → global deviation at all temperatures". It accurately compensates for individualized error shifts caused by the discrete sampling resistor TCR, the difference in magnetic permeability of current transformer core, and the dispersion of ADC reference voltage source temperature drift, which significantly improves the metering consistency of batch terminals across the entire temperature range, and the first-pass calibration success rate can reach over 98%.

[0196] (3) The model is highly interpretable and easy to deploy in engineering. The confidence interval of the Gaussian process regression output provides a statistical basis for determining the significance of individual biases. The XGBoost model inference only requires tree traversal operations and does not require complex floating-point matrix operations. The compensation coefficients in the mass production stage are burned into the MCU in the form of a lookup table. Only piecewise linear interpolation is required during operation. There are no additional requirements for terminal hardware resources, making it suitable for various embedded metrology platforms.

[0197] (4) It has batch adaptive and full lifecycle evolution capabilities. When the component batch changes, only a small number of new batch samples need to be collected to incrementally fine-tune the migration mapping network, without repeating the full temperature range traversal test, resulting in low maintenance costs. At the same time, this method reserves an expansion interface for online error drift correction during terminal field operation, supporting continuous optimization of metrological accuracy throughout its entire lifecycle.

[0198] The above-disclosed embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of the invention. Those skilled in the art will understand that implementing all or part of the above-described embodiments and making equivalent changes in accordance with the claims of the present invention are still within the scope of the invention.

Claims

1. A method for full-temperature error self-calibration of an intelligent fusion terminal based on digital twin transfer, characterized in that, include: Group sample full temperature range multi-condition data acquisition: Preset the full temperature range and multiple test conditions, collect the measurement error values ​​of the same model of intelligent fusion terminal at each temperature-condition combination point, and construct a group sample full temperature error dataset; Population common error baseline modeling based on Gaussian process regression: The population error mean corresponding to each temperature-operating condition combination point is calculated based on the population sample full temperature error dataset to construct a training dataset. Gaussian process regression is used to establish a population common error baseline model. Based on the trained population common error baseline model, predictions are made for all operating conditions. The population common error baseline surface function and confidence interval are generated based on the prediction results. Individual deviation feature extraction and transfer mapping dataset construction: Based on the population common error baseline model, extract the room temperature individual deviation feature vector and the extreme temperature individual deviation label vector at the preset extreme temperature node for each sample terminal, construct the transfer mapping dataset, and assign the terminals with significant individual deviations selected according to the confidence interval to the training set, validation set and test set. Construction and training of the transfer mapping network: The transfer mapping network is constructed using the XGBoost supervised learning framework with extreme gradient boosting. The transfer mapping network is trained using the transfer mapping dataset to derive the full-temperature deviation feature based on the normal temperature deviation feature. Room temperature single-point self-test and full temperature deviation migration prediction: The terminal to be shipped is only tested by a standard source under room temperature environment. The room temperature individual deviation feature vector of the terminal to be shipped is extracted and input into the migration mapping network to predict the full temperature deviation feature of the terminal to be shipped. Full-temperature compensation coefficient reconstruction and terminal programming operation: Based on the full-temperature deviation characteristics and the common error baseline surface function of the group, the error compensation curve of the terminal to be shipped is reconstructed in the full temperature range, a full-temperature compensation coefficient lookup table is generated and programmed into the MCU of the terminal to be shipped. When the terminal to be shipped is running, the metering error self-calibration is completed by interpolating the table based on the real-time temperature.

2. The method for full-temperature error self-calibration of intelligent fusion terminal based on digital twin transfer as described in claim 1, characterized in that, The data collection for the population sample across the entire temperature range and under multiple operating conditions includes: N identical intelligent fusion terminals were selected as sample terminals and placed in a high and low temperature programmable test chamber within the temperature range [T]. min ,T max [Set temperature point T according to gradient] i After each temperature point is kept constant and stabilized, a variety of typical operating conditions are output sequentially through a standard power source; the typical operating conditions include the output current, output voltage, and power factor of the standard power source. The measurement error values ​​of each terminal at each temperature-operating condition combination point are calculated, and a group sample full-temperature error dataset is constructed.

3. The method for full-temperature error self-calibration of intelligent fusion terminals based on digital twin transfer as described in claim 2, characterized in that, The population common error baseline modeling based on Gaussian process regression includes: Construction of a common error dataset and training of a Gaussian process regression model: For each temperature-operating condition combination point, the arithmetic mean of the measurement errors of all sample terminals is calculated as the population error mean to construct a training dataset. The input feature vector includes the effective values ​​of temperature, current, and power factor after Z-score standardization. A Gaussian process regression model is established, using the Matérn 5 / 2 kernel function and introducing an automatic correlation determination ARD mechanism. Independent length scales are assigned to the three input dimensions of temperature, current, and power factor. A hyperparameter vector is defined, and hyperparameter optimization is performed by maximizing the logarithmic marginal likelihood function and using the L-BFGS quasi-Newton algorithm. Population common error baseline surface generation and confidence interval construction: The trained Gaussian process regression model is used to predict any target operating point within the entire temperature range. The predicted mean is used as the population common error baseline value of the target operating point. The grid points covering the temperature range, current range and typical power factor are traversed to generate a complete population common error baseline surface and fit it as a population common error baseline surface function. At the same time, the confidence interval of the population baseline is constructed using the prediction variance. Terminals that exceed the confidence interval are identified as terminals with significant individual deviations.

4. The method for full-temperature error self-calibration of intelligent fusion terminal based on digital twin transfer as described in claim 3, characterized in that, The construction of the common error dataset and the training of the Gaussian process regression model include: For each temperature-operating condition combination point i, calculate the arithmetic mean of the measurement errors of all sample terminals as the population error mean y. i Extract the feature vector of temperature-operating condition combination point i. Build a training dataset Where M is the total number of temperature-operating condition combination points, , , For the standardized temperature, current, and power factor; Establish a Gaussian process regression model (GPR) and define the latent function. Prior to a zero-mean Gaussian process The observation model is ,in It is independent and identically distributed Gaussian noise. To represent the variance of the noise, the kernel function for Gaussian process regression is the Matérn 5 / 2 kernel function, and an automatic correlation determination ARD mechanism is introduced: ; ; in For signal variance, , , These are the length-scale parameters for temperature, current, and power factor, respectively. Define hyperparameter vector The GPR model is optimized by maximizing the logarithmic marginal likelihood function: ; In the formula, G represents the identity matrix, and K is an M×M covariance matrix; the elements in K It represents the covariance between eigenvectors.

5. The method for full-temperature error self-calibration of intelligent fusion terminal based on digital twin transfer as described in claim 4, characterized in that, The generation of the common error baseline surface and the construction of confidence intervals for the population include: After hyperparameter optimization converges, the trained GPR model is used to analyze the temperature range [T]. min ,T max Any target working point within ] Make a prediction, predict the mean. With prediction variance They are respectively: ; ; In the formula, This is the covariance vector between the test point and all other points; To predict the mean As the common error baseline value of the group at the operating point*, it traverses the temperature range [T] min ,T max ], Current range [I min ,I max and typical power factor [φ] min ,φ max The grid points are fitted to generate a baseline surface function for the common error of the population. ; Using prediction variance Construct confidence intervals for the population baseline: ； Terminals that exceed the confidence interval are marked as terminals with significant individual bias.

6. The method for full-temperature error self-calibration of intelligent fusion terminal based on digital twin transfer as described in claim 5, characterized in that, The ambient temperature individual deviation feature vector is a five-dimensional vector, which includes the voltage ratio difference deviation, current ratio difference deviation, active power error deviation, light load active power error deviation, and inductive active power error deviation under rated operating conditions at ambient temperature; the extreme temperature individual deviation label vector is the individual deviation value of each preset extreme temperature node; the transfer mapping dataset is randomly divided into training set, validation set, and test set according to the proportion, wherein all terminals with significant individual deviations are preferentially and evenly distributed to the three sets. The construction and training of the transfer mapping network includes: A multi-output regression strategy was adopted to construct five independent XGBoost single-output regression sub-models, each corresponding to W preset extreme temperature nodes. The objective function of each XGBoost single-output regression sub-model consists of a squared error loss function and a regularization term. Each XGBoost single-output regression sub-model was trained using the training set, the validation set was used for early stopping monitoring, and the test set was used for testing. The objective function of the XGBoost single-output regression sub-model is: ; In the formula, The model predicts the value; N is the number of training samples, T is the total number of trees; Lt is the number of leaf nodes in the t-th tree. Let be the weight vector of the leaf nodes of the t-th tree. Leaf quantity penalty coefficient, To be related to the L2 regularization coefficient, The L1 regularization coefficient is... This is an L1 regularization term.

7. The method for full-temperature error self-calibration of intelligent fusion terminal based on digital twin transfer as described in claim 6, characterized in that, The single-point self-test at room temperature and the full-temperature deviation migration prediction include: Connect a standard power source at room temperature to perform a routine accuracy check on the terminal to be shipped, and collect the room temperature measurement error value. The predicted value of the current temperature-operating point is calculated based on the baseline surface function of the common error of the group, and the difference is made with the room temperature measurement error value to obtain the room temperature individual deviation feature vector. ; Will Input the migration mapping network to obtain the full-temperature deviation characteristics of the terminal to be shipped.

8. The method for full-temperature error self-calibration of intelligent fusion terminal based on digital twin transfer as described in claim 7, characterized in that, The full-temperature compensation coefficient reconstruction and terminal programming operation includes: The predicted values ​​of each point in the full temperature range are calculated based on the common error baseline surface function of the group. Combined with the full temperature deviation characteristics, the corresponding sum values ​​are calculated one by one according to the temperature points. Based on the correspondence between temperature points and sum values, an error curve is fitted, which serves as the error compensation curve for the terminal to be shipped across the entire temperature range. The correction coefficients corresponding to each temperature node are calculated in reverse based on the error compensation curve, and a full-temperature compensation coefficient lookup table is generated. This table is then burned into the memory of the MCU of the terminal to be shipped through the communication interface. The MCU can then obtain the temperature of the sampling module in real time through the built-in temperature sensor, perform piecewise linear interpolation in the table based on the current temperature, and write the interpolation results into the corresponding correction register of the metering chip to complete the dynamic self-calibration of the full-temperature range power metering.

Images