A smart electric energy meter life prediction method

Abstract

The present application relates to a kind of intelligent electric energy meter life prediction method, comprising the following steps: step 1, collect intelligent electric energy meter data and mark, establish data set;Step 2, life characteristic generation;Step 3, generate risk loss;Step 4, realize domain adaptive method based on MK-MMD;Step 5, construct the network model of MK-MMD domain adaptive method based on life characteristic generation and survival risk loss.The present application compared with prior art, with the following beneficial effects: service life characteristic generation method, based on statistical weibull distribution is generated the life characteristic containing rich life information, enhances the correlation between intelligent electric energy meter feature and its life, improves detection accuracy.2) propose survival risk loss, design survival risk loss to optimize loss function, guide model to carry out risk calculation to predicted life result, so that loss function meets actual application scene, better guide model training.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of data mining of artificial intelligence, and particularly relates to an MK-MMD domain adaptive intelligent electric energy meter life prediction method based on life feature generation and survival risk loss. BACKGROUND

[0002] The intelligent electric energy meter bears the important functions of ensuring user electricity safety, supporting intelligent power grid construction and assisting asset management. The intelligent electric energy meter has higher cost than the traditional electric energy meter, has the characteristics of large installation quantity, good market prospect and complex technology. The intelligent electric energy meter life prediction can prolong the service life of the intelligent electric energy meter as much as possible, realizes the maximum value of the intelligent electric energy meter, and how to accurately predict the life of the intelligent electric energy meter is one of the primary problems of the maximum value of the intelligent electric energy meter. SUMMARY

[0003] The application provides an intelligent electric energy meter life prediction method, which aims to generate life features with rich life information, calculate the survival risk of the prediction life result through a survival risk loss guide model, and realize more accurate intelligent electric energy meter life prediction by narrowing the domain distribution difference through a domain adaptive method.

[0004] The application is realized through the following technical means:

[0005] An intelligent electric energy meter life prediction method comprises the following steps:

[0006] Step 1, collect intelligent electric energy meter data and label, and establish a data set;

[0007] Step 2, life feature generation;

[0008] Step 3, survival risk loss;

[0009] Step 4, realize a domain adaptive method based on MK-MMD;

[0010] Step 5, construct a network model of the MK-MMD domain adaptive method based on life feature generation and survival risk loss.

[0011] The step 1 specifically comprises the following steps:

[0012] 1-1, extract intelligent electric energy meter related data from a power utilization information collection system.

[0013] 1-2, use a related method of data processing to clean and process the data;

[0014] 1-3, extract features from the voltage abnormal event table;

[0015] 1-4, merge table entries and generate label information;

[0016] The specific process of the step 1-1 includes the following: extracting the smart electric energy meter records with the same identifier from the smart electric energy meter asset information table D METER and the voltage abnormal event table EREC 10.

[0017] The specific process of the step 1-2 includes the following: respectively performing rule cleaning and missing value processing on the sample data extracted in the step 1-1.

[0018] The specific process of the step 1-3 includes the following: selecting a multi-table shared smart electric energy meter identifier METER ID, extracting the voltage abnormal event records of the smart electric energy meter from the processed EREC 10 table, and performing classified statistics on the abnormal event conditions in a time interval T, as the voltage abnormal event feature extraction result.

[0019] The specific process of the step 1-4 includes the following: merging the voltage abnormal event features and the smart electric energy meter asset information table features through the smart electric energy meter unique identifier METER ID, and then generating the remaining life label d of the sample. m -r m , d m is the scrap time of the mth sample of the smart electric energy meter, which is recorded in the smart electric energy meter asset information table, r m is the end time of the voltage abnormal event statistical record corresponding to the sample.

[0020] The step 2 specifically includes the following steps.

[0021] Step 2-1, dividing the original sample into different subsets according to the feature values.

[0022] Step 2-2, using maximum likelihood estimation to solve the Weibull distribution parameters of the sample subset for different subsets.

[0023] Step 2-3, generating the life characteristics of the batch based on the obtained Weibull distribution parameters.

[0024] In the step 2-1, the original feature values in the sample are discrete values, which are divided into different subsets according to the feature values.

[0025] The specific process of the step 2-2 includes the following: obtaining a likelihood function according to the sample subset and the maximum likelihood estimation principle, and the likelihood function is a function of the Weibull distribution parameters. Then, taking the logarithm of the likelihood function and taking the partial derivative to obtain a set of transcendental equations about the Weibull distribution parameters. Because it is difficult to obtain a formal solution of the transcendental equation, the Newton-Raphson algorithm of the numerical solution method of the nonlinear equation set is used to solve the distribution parameters η and β.

[0026] The life characteristic described in steps 2-3 is the average life under Weibull distribution, the formula is expressed as ηΓ(1+1 / β).

[0027] The step 3 comprises the following steps:

[0028] Step 3-1, using maximum likelihood estimation to obtain the Weibull distribution parameters of the overall life of the smart electric energy meter source domain sample (training set);

[0029] Step 3-2, determine the specific form of survival risk loss;

[0030] The specific process of the step 3-1 comprises the following: using the maximum likelihood estimation method described in step 2 to solve the Weibull distribution parameters η and β of the life of the smart electric energy meter source domain sample.

[0031] The specific process of the step 3-2 comprises the following: using the Weibull distribution parameters solved in step 3-1 to determine the form of the Weibull distribution survival risk function of the sample:

[0032]

[0033] In the formula, y c is the predicted life;

[0034] Further, the specific form of the survival risk loss suitable for the life prediction of the smart electric energy meter can be determined;

[0035]

[0036] In the formula, y c is the predicted life, y is the true life, J C (y c , y) is the prediction loss.

[0037] In the step 4, first, the source domain features and the target domain features of the source domain sample and the target domain sample are extracted using the feature extraction network, and then the domain distribution difference is calculated based on MK-MMD;

[0038] The specific process of the step 5 is as follows: first, the source domain and the target domain data generate life characteristics through the life characteristic generation module, and then the generated life characteristics and the original characteristics are combined and sent into the feature extraction network to extract features, and then the extracted features are sent into the domain adaptive module, and the MK-MMD method is used to calculate the domain difference loss of the source domain and the target domain features, and then the source domain features are sent into the prediction module to predict the life of the smart electric energy meter, the conventional prediction loss is calculated for the predicted value, and the predicted value is sent into the survival risk loss module to calculate the survival risk loss, and the three losses are superimposed as the total loss to guide the model parameter update, and more accurate life prediction is performed.

[0039] Compared with the prior art, the present invention has the following advantages:

[0040] 1) The lifespan feature generation method generates lifespan features containing rich lifespan information based on the statistical Weibull distribution, which enhances the correlation between smart energy meter features and their lifespan and improves detection accuracy.

[0041] 2) The concept of survival risk loss is proposed and designed to optimize the loss function. This guides the model to perform risk calculations on the predicted lifespan, making the loss function more consistent with real-world application scenarios and better guiding model training. Attached Figure Description

[0042] Appendix Figure 1 This is a network structure diagram of an MK-MMD domain adaptive smart meter lifetime prediction method based on lifetime feature generation and survival risk loss in the present invention.

[0043] Appendix Figure 2 This is the lifetime feature generation algorithm flow based on Weibull distribution in the method of this invention.

[0044] Appendix Figure 3 This is the flow of the MK-MMD domain adaptive algorithm generated based on survival risk loss and lifetime characteristics in the method of this invention. Detailed Implementation

[0045] The invention will be described in detail below through specific examples:

[0046] This invention relates to an adaptive smart meter lifetime prediction method based on lifetime feature generation and survival risk loss in the MK-MMD domain. The method includes the following steps:

[0047] Step 1: Collect and label data from smart meters to create a dataset;

[0048] Step 2, lifetime feature generation;

[0049] Step 3, survival risk and loss;

[0050] Step 4: Implement the domain adaptation method based on MK-MMD;

[0051] Step 5: Construct a network model based on the MK-MMD domain adaptive method of lifetime feature generation and survival risk loss.

[0052] Step 1, which involves collecting and labeling smart meter data to establish a dataset, specifically includes the following steps:

[0053] 1-1, extracting smart meter related data from Jiangsu Province part of the data of the electricity information collection system (including smart meter asset information table (D_METER), voltage abnormal event table (EREC_10)).

[0054] 1-2, using data processing related methods to clean and process data;

[0055] 1-3, extracting features from the voltage abnormal event table.

[0056] 1-4, merging table entries and generating label information.

[0057] The specific process of step 1-1 extracting smart meter related data from Jiangsu Province part of the data of the electricity information collection system (including smart meter asset information table (D_METER), voltage abnormal event table (EREC_10)) is as follows:

[0058] 1) Determine the smart meter samples common to multiple tables. All smart meter samples in the table have a unique identification field METER_ID. Determine the smart meter identifier set common to multiple tables through this field:

[0059] 2) According to the common smart meter identifier set determined above, extract related data from different tables respectively:

[0060] The specific process of step 1-2 using data processing related methods to clean and process data is as follows: rule cleaning and missing value processing are performed on the extracted data.

[0061] The specific process of step 1-3 extracting features from the voltage abnormal event table is as follows:

[0062] 1) Select the voltage table unique identifier METER_ID to be extracted;

[0063] 2) Extract the abnormal event records of this METER_ID from the voltage abnormal event table:

[0064] 3) The abnormal event record starting time point t, select a time interval T, and count the frequency of various abnormalities as the feature extraction result of the voltage abnormal event table of the voltage table:

[0065] Suppose we want to extract the voltage abnormal event features of a smart meter. First, determine the identifier METER_ID of the smart meter to be extracted; then, extract all voltage abnormal event records with METER_ID equal to this value from the voltage abnormal event table; then, determine an abnormal event record starting time point t, and take a time interval T. Suppose there are k records from time t to time t+T, which form Y, written as:

[0066] Y = {y1, y2,..., y k} (3)

[0067] EXCEPTION_SIGN is a three-bit Boolean variable, and all zeros are illegal states, so there are seven different abnormal condition records, and the abnormal events are recorded as follows:

[0068] f i (x = b i ) = a i (4)

[0069] In the formula, b i represents the i-th abnormal condition, a i represents the number of records of the i-th condition of the smart electric energy meter, and at the same time, the following conditions are met:

[0070]

[0071] Taking the description of each abnormal condition as a feature, the smart electric energy meter obtains seven feature columns of new record abnormal conditions, and the value is the number of records corresponding to each abnormal condition. Through the above operation, the feature extraction of the voltage abnormal event of the smart electric energy meter is completed.

[0072] The specific process of combining the table items and generating the label information of steps 1-4 is as follows:

[0073] 1) Use the identifier of the smart electric energy meter to combine the voltage table asset information table and the extracted features of the voltage table voltage abnormal event table into sample features:

[0074] 2) Generate sample labels. The sample label generation uses the scrap time (feature label is DESCARD_DATE) in the sample record minus the last cutoff time of the voltage abnormal event statistics, and the formula is as follows:

[0075] label m = d m -r m (6)

[0076] d m is the scrap time of the m-th sample of the smart electric energy meter, recorded in the smart electric energy meter asset information table, and r m is the end time of the voltage abnormal event statistics record, that is, t+T above.

[0077] The step 2 life feature generation specifically includes the following steps:

[0078] Step 2-1, divide the original sample into different subsets;

[0079] Step 2-2, using maximum likelihood estimation to solve the Weibull distribution parameters of the sample subsets for different subsets;

[0080] Step 2-3, generating the life characteristics of the batch based on the obtained distribution parameters.

[0081] The specific process of the step 2-1 of dividing the original sample into different subsets is as follows:

[0082] Suppose that the total sample of the smart electric energy meter is X = {f1, f2,..., f n}, wherein f n is the nth feature of the sample, the feature has m values, and the sample set is divided into different batches according to the different values, and m batches can be obtained. For example, in the current sample set, HARD_VER is a hardware feature of the sample, the feature has 3 feature values, the present application regards each feature value as a batch, and 3 batches can be obtained, the Weibull parameters of the batches are solved, and then the corresponding life characteristics are calculated on this basis.

[0083] After the batch division is completed, the data of the batch can be recorded as:

[0084]

[0085] In the formula, indicates the jth feature value of the ith feature, is a set of samples in the total sample whose ith feature value is equal to , and it is obviously a subset of the total sample.

[0086] Suppose is a set of k samples, the corresponding life label data is extracted to form a life distribution set for solving the Weibull parameters of the batch, and the set is regarded as Y and recorded as:

[0087] Y = {y1, y2,..., y k} (8)

[0088] y k is the life label data corresponding to the kth sample in the batch of the smart electric energy meter. The life label data is sorted in ascending order to obtain an ordered set T, which is recorded as:

[0089] T = {t1, t2,..., t k} (9)

[0090] The step 2-2 uses maximum likelihood estimation to solve the Weibull distribution parameters of the sample subsets for different subsets, and specifically includes the following steps:

[0091] The likelihood function is determined from the ordered set T:

[0092]

[0093] Take the log of this likelihood function:

[0094]

[0095] Take the partial derivatives of both sides with respect to β and η and set both equal to zero:

[0096]

[0097] Solving the above equations gives the estimates of β and η and Solving the above equations requires a numerical solution because the equations are transcendental and difficult to solve analytically. In this chapter, the Newton-Raphson algorithm for non-linear equations is used. The Newton-Raphson algorithm for solving a non-linear equation is as follows:

[0098] Suppose there is a set of non-linear equations:

[0099]

[0100] In this set of non-linear equations, at least one of the equations is a non-linear function of x. x = (x1, x2,..., x n ) T is the solution to the set of equations, x is a real number. The set of equations can be written as

[0101] F(x) = (f1(x), f2(x),..., f n (x)) T = 0 (14)

[0102] Suppose x * is the solution to the set of equations, x (a) represents the a-th iteration of x. Expanding f i in a first order Taylor series around x (a) , each component function can be approximated as f i (x (a) ) = 0, then:

[0103]

[0104] The overall simplification is:

[0105]

[0106] In detail, this can be written as:

[0107]

[0108] Further, the approximate solution of x * is obtained as the next iteration value.

[0109] x (a+1) = x (a) - [F'(x a )] -1 F(x a ) (18)

[0110] According to the above Newton-Raphson iterative algorithm method, the nonlinear equation group about the Weibull distribution parameters can be obtained:

[0111]

[0112] Thus, the Jacobi matrix of the distribution parameters is obtained:

[0113]

[0114] Each item in the matrix is as follows:

[0115]

[0116] Next, according to the Newton-Raphson iterative algorithm, the related parameters of the Weibull distribution can be solved after multiple iterations.

[0117] The step 2-3 generates the batch life characteristics based on the obtained distribution parameters, and specifically includes the following steps:

[0118] The above maximum likelihood estimation method is used to solve the Weibull distribution parameters, and the Weibull distribution parameters η and β are obtained. η is the scale parameter, and β is the shape parameter. The average life under the Weibull distribution reflects the life that each individual sample of the batch theoretically has, and is also the life characteristics generated in this chapter. Denoted as:

[0119]

[0120] The step 3 survival risk loss specifically includes the following steps:

[0121] Step 3-1, using maximum likelihood estimation to obtain the Weibull distribution parameters of the overall life of the smart electric energy meter source domain sample (training set);

[0122] Step 3-2, determine the specific form of the survival risk loss;

[0123] The step 3-1 uses maximum likelihood estimation to obtain the Weibull distribution parameters of the overall life of the smart electric energy meter source domain sample (training set), and specifically includes the following steps:

[0124] 1) Take out the life label of the source domain smart electric energy meter sample set as the set of solving Weibull distribution parameters;

[0125] 2) Use the maximum likelihood estimation method of step 2 to solve the corresponding Weibull distribution: η and β;

[0126] Said step 3-2 determines the specific form of the survival risk loss, and specifically includes the following steps:

[0127] 1) Determine the survival risk function according to the solved Weibull distribution parameters;

[0128]

[0129] 2) Write the survival risk loss function that integrates the survival risk function;

[0130]

[0131] Said step 4 realizes the domain adaptation method based on MK-MMD, and specifically includes the following steps:

[0132] Step 4-1, extract the source domain features and target domain features of the source domain samples (training set) and target domain samples (test set);

[0133] Step 4-2, calculate the domain distribution difference based on MK-MMD;

[0134] Said step 4-1 extracts the source domain features and target domain features of the source domain samples (training set) and target domain samples (test set), and specifically includes the following steps:

[0135] 1) Use the feature extraction network to extract the features of the source domain smart electric energy meter samples to form the source domain features;

[0136] 2) Use the feature extraction network to extract the features of the target domain smart electric energy meter samples to form the target domain features;

[0137] Suppose the feature extraction network is a k-layer network, and the k+1 layer is the output layer, so the network added with the domain adaptation layer becomes k+2 layers. The first k layers are still the feature extraction network, the k+1 layer is the domain adaptation layer, and the k+2 layer is the output layer. The extracted features are specifically described as follows:

[0138]

[0139] Where, X S is the source domain sample set, X T is the target domain sample set, F k+1 (x s ) represents the output of the source domain sample in the domain adaptation layer, F k+1 (x t) represents the output of the target domain sample at the domain adaptation layer, D S is the source domain feature, D T is the target domain feature.

[0140] Step 4-2 of the present application calculates the domain distribution difference based on MK-MMD, specifically comprising the following steps: using multiple kernel functions to map the features to different Hilbert kernel spaces, calculating the domain difference in multiple spaces respectively, and obtaining the domain difference based on the multi-kernel maximum mean difference (MK-MMD) by averaging and weighting.

[0141] Specifically, the maximum mean difference (MMD) is a widely used non-parametric method for measuring the distribution difference between the source domain and the target domain. When the MMD of the source domain and the target domain is small enough, they are considered to be identically distributed. The empirical estimate of MMD can be expressed as follows:

[0142]

[0143] φ(·) is a feature mapping with respect to k(d s ,d t ) = <φ(d s ), φ(d t )>). It can map the feature representation of the source domain and the target domain to the reproducing kernel Hilbert space (RKHS), where the distribution difference between the two domains can be calculated. The source domain d s ∈ D S and the target domain d t ∈ D T . k(d s ,d t ) is a kernel function, which can solve the MMD distance by kernel function, without knowing the mapping and feature space in advance.

[0144] Multi-kernel maximum mean difference (MK-MMD) is developed on the basis of MMD, which uses multiple kernels instead of a single kernel. Since the single kernel function inevitably fixes the selection of a set of basis functions of the Hilbert space, it leads to the direct connection between the application effect of MMD and the applicability of the kernel function to the problem, which is easily affected by the sensitivity of the kernel function and produces adverse results. MK-MMD uses different kernel functions to map features to different reproducing Hilbert kernel spaces, optimizing the high sensitivity of MMD to kernel functions. Therefore, the present application uses MK-MMD method to measure the domain distribution difference, which is a convex combination of multiple kernel functions, compared with MMD, which reduces the adverse effects of MMD selecting a single non-optimal kernel on the model performance.

[0145] Similarly, the empirical estimate of MK-MMD can be expressed as follows:

[0146]

[0147] This formula is similar to MMD. Only φ k (·) is a feature map about k k (d s ,d t ) = <φ k (d s ), φ k (d t )> is a feature map about k k (d s ,d t ) is defined as a convex combination of m kernels, thus optimizing the kernel selection, and the final kernel function is as follows:

[0148]

[0149] k i is the i-th kernel of MK-MMD, and α i is the weight of the kernel k i . The weights of all kernels should add up to 1.

[0150] The method to minimize the distance between the source domain and the target domain is to minimize the loss function:

[0151]

[0152] where J C (y c ,y) represents the prediction loss, y is the actual value, and y c is the predicted value. represents the distribution difference between the source domain X S and the target domain X T , and λ is a hyperparameter.

[0153] The step 5 constructs a network model of the MK-MMD domain adaptation method based on life feature generation and survival risk loss, specifically including the following steps: first, the source domain and target domain data pass through the life feature generation module to generate life features, and then the generated life features and original features are combined and sent into the feature extraction network to extract features, and then the extracted features are sent into the domain adaptation module, and the MK-MMD method is used to calculate the domain difference loss for the source domain and target domain features, then the source domain features are sent into the prediction module to predict the life of the smart electric energy meter, the prediction value is calculated for the conventional prediction loss, and the prediction value is sent into the survival risk loss module to calculate the survival risk loss, and the three losses are superimposed as the total loss to guide the model parameter update for more accurate life prediction.

[0154] Specifically, first, the smart meter X = {X S ,X T} sample is sent to the life feature generation module for life feature generation, and the generated life feature and the original feature are spliced and merged to obtain a data set containing rich life information, denoted as:

[0155] X L = {X SL ,X TL} (30)

[0156] The life label set Y corresponding to the source domain data sample:

[0157] Y S = {y1,y2,...,y m} (31)

[0158] The corresponding Weibull distribution parameters are solved using maximum likelihood estimation:

[0159]

[0160] Assuming that the feature extraction network is a k-layer network, the k+1 layer is the domain adaptive layer, and the output of the model is after the domain adaptive layer. The extracted smart meter sample features are sent to the domain adaptive layer to obtain:

[0161]

[0162] Where X SL is the source domain sample set, X ST is the target domain sample set, F k+1 (x sl ) represents the output of the source domain sample in the domain adaptive layer, F k+1 (x tl ) represents the output of the target domain sample in the domain adaptive layer, D SL is the source domain feature, and D TL is the target domain feature.

[0163] The loss function based on the survival risk loss and the domain adaptation is as follows:

[0164]

Claims

1. A method for predicting the lifespan of a smart energy meter, characterized in that: Includes the following steps: Step 1: Collect and label data from smart meters to create a dataset; Step 2, lifetime feature generation; Step 3: Generate risk loss; Step 4: Implement the domain adaptation method based on MK-MMD; Step 5: Construct the corresponding network model based on the domain adaptation method in Step 4 to realize the prediction of the lifespan of the electricity meter; Step 2 specifically includes the following steps: Step 2-1: Divide the original samples into different subsets according to their feature values, and each subset constitutes a batch; Step 2-2: For different subsets, use maximum likelihood estimation to solve for the Weibull distribution parameters of the sample subsets; Steps 2-3: Generate the lifetime characteristics of this batch based on the obtained Weibull distribution parameters; Step 3 includes the following steps: Step 3-1: Use maximum likelihood estimation to obtain the Weibull distribution parameters of the overall lifetime of the training set of smart energy meter source domain samples; Step 3-2: Determine the specific form of survival risk loss; In step 3-1, the maximum likelihood estimation method described in step 2 is used to solve for the Weibull distribution parameters that the source domain sample lifetime of the smart energy meter follows. and ; Step 3-2 determines the specific form of survival risk loss, and the specific process includes the following: using the Weibull distribution parameters obtained in step 3-1 to determine the form of the survival risk function that the sample follows: This allows for the determination of specific forms of survival risk loss applicable to the lifespan prediction of smart energy meters. ; in, Indicates the predicted loss. This is the actual value. It is a predicted value. For hyperparameters; The specific process of step 5 is as follows: a k-layer neural network is constructed for feature extraction. After the feature extraction network, a fully connected layer of layer k+1 is added as a domain adaptation layer. After the domain adaptation layer is the output layer of the model, which is used to realize data prediction. The operation process of the entire model is as follows: First, the source domain and target domain data are processed by the lifespan feature generation module to generate lifespan features. Then, the generated lifespan features and the original features are merged and fed into the feature extraction network to extract features. Then, the extracted features are fed into the domain adaptation module, and the MK-MMD method is used to calculate the domain difference between the source domain and target domain features. Next, the source domain features are fed into the prediction module to predict the lifespan of the smart energy meter. The predicted value is calculated using conventional prediction loss calculation. The predicted value is fed into the survival risk module to calculate the survival risk loss. The three losses are superimposed as the total loss to guide the model parameter update. Finally, the trained model is used to perform more accurate lifespan prediction.

2. The method for predicting the lifespan of a smart energy meter according to claim 1, characterized in that: Step 1 specifically includes the following steps: Step 1-1: Extract relevant data from smart meters from the electricity consumption information collection system; Steps 1-2: Clean and process the data using relevant data processing methods; Steps 1-3: Extract features from the voltage anomaly event table; Steps 1-4: Merge table entries and generate label information.

3. The method for predicting the lifespan of a smart energy meter according to claim 2, characterized in that: The smart meter related data mentioned in step 1-1 is the smart meter record with the same identifier METER_ID extracted from the smart meter asset information table D_METER and the voltage anomaly event table EREC_10: Steps 1-2 use relevant data processing methods to clean and process the data, specifically by performing rule-based cleaning and missing value processing on the sample data extracted in step 1-1. The specific process of feature extraction from the voltage anomaly event table in steps 1-3 includes the following: selecting a smart energy meter METER_ID identifier shared by multiple meters, extracting voltage anomaly event records of this smart energy meter from the processed EREC_10 table, and performing feature extraction at time intervals. The internal abnormal events are classified and statistically analyzed, and used as the feature extraction results of voltage abnormal events; The specific processes of steps 1-4 include the following: merging voltage anomaly event features and smart meter asset information table features using the smart meter's unique identifier METER_ID, and then generating the sample's remaining lifetime label. , It is the first smart energy meter The scrapping time of each sample This is the end time of the statistical record of voltage anomalies for this sample.

4. The method for predicting the lifespan of a smart energy meter according to claim 1, characterized in that: In step 2-1, the feature values ​​in the original sample are all discrete values, and they are divided into different subsets according to the feature values. Step 2-2, for different subsets, uses maximum likelihood estimation to solve for the Weibull distribution parameters of the sample subsets. The specific process includes: obtaining the likelihood function based on the sample subsets and the principle of maximum likelihood estimation; then, taking the logarithm and partial derivatives of the likelihood function to obtain a set of transcendental equations regarding the Weibull distribution parameters. Because it is difficult to find formal solutions to transcendental equations, the Newton-Raphson algorithm, a numerical solution method for nonlinear equations, is used to solve for the distribution parameters. and ; The lifetime characteristic described in steps 2-3 is the average lifetime under a Weibull distribution, and its mathematical form is: .

5. The method for predicting the lifespan of a smart energy meter according to claim 1, characterized in that: In step 4, the source domain features and target domain features of the source domain samples and target domain samples are first extracted, and then the domain distribution difference is calculated based on MK-MMD.

Images