An intelligent electricity meter fault prediction method based on SVM-WS

Abstract

The present invention discloses an intelligent electric meter fault prediction method based on SVM-WS for fault prediction of intelligent electric meters. The method includes steps of preprocessing data, extracting index features to construct an SVM model, and using the constructed model to perform fault prediction of intelligent electric meters. The key is that after the step of extracting index features to construct the SVM model, the entropy weight method is used to calculate the weights of the SVM model, optimize the model weights, and after obtaining the improved SVM model, then perform the fault prediction of the intelligent electric meter. The present invention uses the SVM-WS method to analyze the ledger data, operation data, fault data, power consumption data, etc. of intelligent electric meters, and at the same time uses big data technology to construct a multi-source heterogeneous data set, calculate the sample weights between different features, construct a reasonable device portrait, perform reasonable status evaluation and fault prediction on intelligent electric meters, and achieve the effect of changing from passive repair to active operation and maintenance.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent meter fault prediction, and relates to a fault prediction method for intelligent meters, in particular to a method for transforming SVM using a weight incentive method and performing intelligent meter fault prediction. Background Art

[0002] With the substantial increase in the number of intelligent meter devices, the operation and management of the devices have become increasingly difficult. As the nerve endings of the smart grid, intelligent meters are characterized by a large number and wide distribution, playing an irreplaceable role in electric information collection and energy monitoring. At the same time, the reliability of intelligent meters is also related to the power consumption safety of ordinary households. Due to the difficulty in collecting a large amount of sample information and long-term user tracking records, it is difficult to predict the faults of electric energy meters. Therefore, it is urgent to implement fault prediction for intelligent meters. Establishing a set of scientific and effective fault prediction schemes is of great significance for the reliability design of electric energy meters and the use and replacement of power grid companies.

[0003] Most traditional fault predictions are based on human experience, and relevant logic scripts are written through programs for fault analysis and prediction. This method not only fails to guarantee the effectiveness, has a low degree of automation, but also overly relies on manual work. For example, professional personnel generally conduct on-site tests on meters periodically, making it difficult to comprehensively evaluate and predict the health status, operating status, fault analysis, and reliability assessment of intelligent meters under operating conditions, and unable to scientifically and effectively manage the entire life cycle of meters.

[0004] The patent with the publication number CN113011530A and the patent name of a method for predicting intelligent meter faults based on multi-classifier fusion also involves SVM, performs simple correlation analysis, and deletes some features. However, when mapping to a high-dimensional space, the data feature dimensions are too many, resulting in problems such as feature sparsity or dimensional explosion, and ultimately leading to inaccurate fault prediction when it cannot be solved. Summary of the Invention

[0005] The present invention proposes a method for predicting intelligent meter faults based on SVM-WS, using artificial intelligence algorithms to construct an intelligent meter fault prediction model. First, operating data, fault data, evaluation data, etc. are associated to generate a multi-source heterogeneous data set. Secondly, the SVM algorithm is used to initially construct an intelligent meter fault prediction model. Then, the weight incentive idea is used to perform weight incentive on the output of SVM to reduce the number of features in the high-dimensional space. Finally, an intelligent meter fault prediction model based on SVM-WS is obtained for fault analysis and prediction.

[0006] The technical solution of the present invention is an intelligent meter fault prediction method based on SVM-WS for fault prediction of intelligent meters. The method includes steps of preprocessing data, extracting index features to construct an SVM model, and using the constructed model to predict faults of intelligent meters. The key is that after the step of extracting index features to construct the SVM model, the entropy weight method is used to calculate the weights of the SVM model, optimize the model weights, and after obtaining the improved SVM model, the fault prediction of the intelligent meter is carried out.

[0007] The specific steps of optimizing the weights of the SVM model by the entropy weight method are as follows:

[0008] Optimize the weight of each feature in the SVM model, and set the excitation weight W i , and the optimization formula is: where E i is the information entropy of the i-th feature, f(x i ) is the non-linear mapping value of the i-th feature, k is a positive integer greater than 2, and i is a natural number.

[0009] In the data preprocessing step, the PCA algorithm is used for the feature data samples to reduce the fault dimension and delete the data samples with low data information content. The judgment steps for low data information content include:

[0010] a. Perform data standardization and unification processing on k indicators screened by feature engineering, where k is a positive integer greater than 2;

[0011] b. Perform linear transformation on the original sample data, calculate the maximum variance, and select the original samples with the maximum variance value less than the mean of the square root of the feature variance as the data samples with low data information content.

[0012] The specific steps of preprocessing the data are as follows:

[0013] 1). Process missing data;

[0014] 2). Process outliers;

[0015] 3). Data reduction, including dimension reduction and dimension transformation.

[0016] The beneficial effects of the present invention are as follows:

[0017] 1. By using the SVM-WS method, analyze the ledger data, operation data, fault data, power consumption data, etc. of intelligent meters, and at the same time use big data technology to construct a multi-source heterogeneous data set, calculate the sample weights between different features, construct a reasonable device portrait, and perform reasonable status evaluation and fault prediction on intelligent meters, achieving the effect of "from passive repair to active operation and maintenance".

[0018] 2. Utilize weight incentives and dimensionality reduction to optimize the traditional SVM algorithm, assign sample feature weights, and perform weight incentives before mapping features to high-dimensional space to avoid problems such as feature sparsity and feature explosion caused by too many dimensions in high-dimensional space. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a flow chart of the smart meter fault prediction method of the present invention.

[0020] Figure 2 It is a flow chart of the method for constructing the SVM model in the present invention. DETAILED DESCRIPTION

[0021] Related English or term explanations:

[0022] SVM: support vector machine. WS: weighted stimulus.

[0023] See Figure 1 The technical solution of the present invention includes three main steps: data preprocessing, SVM model construction, and weight excitation, which are as follows:

[0024] 1. Data preprocessing steps:

[0025] Data preprocessing mainly involves associating data from various business systems to form a wide table for modeling. Data cleaning and conversion must be performed from the following aspects:

[0026] (1) Missing data processing: Delete the variable. If the missing rate of the variable is high (greater than 80%), the coverage is low, and the importance is low, the variable can be directly deleted; fill in the statistics. If the missing rate is low (less than 65%) and the importance is low, fill in the missing data based on the data distribution. If the data conforms to a uniform distribution, fill in the missing data with the mean of the variable. If the data has a skewed distribution, fill in the missing data with the median.

[0027] (2) Outlier processing: Distance-based processing. By defining a proximity metric between objects, the distance between the outlier and the other objects is used to determine whether the outlier is far away from the other objects. Density-based processing. The local density of the outlier is significantly lower than that of most of the neighboring points. Using the clustering algorithm, small clusters far away from other clusters are discarded.

[0028] (3) Data reduction: Data reduction techniques are used to obtain a reduced representation of a dataset. They can scale the dataset while still closely preserving the integrity of the original data. This allows mining on the reduced dataset to be more efficient and produce the same analytical results. Common strategies include:

[0029] Dimensionality reduction. Data used for data analysis may contain hundreds of attributes, most of which are irrelevant to the mining task and redundant. Dimensionality reduction reduces the amount of data by removing irrelevant attributes while minimizing information loss.

[0030] Dimension transformation: Dimension transformation reduces existing data to a smaller dimension, ensuring the integrity of the data information as much as possible and greatly improving modeling efficiency.

[0031] In this step, it is also necessary to use the PCA algorithm to reduce the fault dimension of the feature data samples and delete data samples with low data information content. The steps for determining low data information content include:

[0032] Step 1: Data standardization. The k indicators screened by feature engineering are subjected to data standardization. The calculation formula is as follows:

[0033]

[0034] where X i is the i-th sample value, max(X i ) is the maximum value of X feature, max(X i ) is the minimum value of the X feature.

[0035] Step 2: Construct a linear transformation. Assuming the original data dimension is p, it can be written as:

[0036] X=(x1,x2,...,x p ) T ,

[0037] Where X represents a sample data. If the data set has N dimensions, the mean and variance of the sample are:

[0038]

[0039] After linear transformation, the results are as follows:

[0040]

[0041]

[0042] …

[0043]

[0044] Step 3: Solve for the maximum variance. The data after linear transformation is:

[0045] Y=(y1,y2,...,y p ) T ,

[0046] Its variance and covariance are:

[0047]

[0048] If the y variance reaches the maximum, the constraint optimization is:

[0049] max(μ T ∑μ).

[0050] The smaller the variance value, the less information the feature data contains; the larger the variance value, the more information the feature data contains. Feature data with a variance value less than the mean of the squared sum of the feature variance is deleted to reduce the dimensionality of the fault type.

[0051] 2. SVM model construction steps

[0052] See Figure 2 By preprocessing the data and integrating operational data, fault data, and ledger data, we constructed a multi-source heterogeneous dataset. We then combined specific business rules to extract relevant indicator features for constructing the SVM model. We then divided the dataset into training, validation, and test sets, preparing for model learning.

[0053] Step 1: Find the maximum hyperplane margin. The hyperplane formula is as follows:

[0054] W T x+b=0,

[0055] Combined with the expression of the hyperplane, calculate the distance from the sample point to the plane. Assume P(x1, x2, ..., x n ) is a point in the sample, where x i Represents the i-th characteristic variable, then the distance d from the point to the hyperplane is:

[0056]

[0057] Where ||W|| is the hyperplane norm and b is the constant intercept.

[0058] Step 2: Optimization model for maximum margin. Once the hyperplane is determined, all support vectors can be found and the margin can be calculated. Each hyperplane corresponds to a margin, and the goal of the calculation is to find the hyperplane corresponding to the largest value among all margins. Therefore, in mathematical terms, it is to determine w and b to maximize the margin. This is an optimization problem and its objective function can be written as:

[0059]

[0060] Where y represents the data point label.

[0061] Step 3: Lagrangian optimization solution. For the convenience of calculation, the objective function is equivalently replaced by:

[0062]

[0063] make Finding the partial derivative with respect to L gives the following result:

[0064]

[0065] Step 4: KKT assumptions. The KKT conditions are as follows:

[0066]

[0067] According to the KKT conditions, we can get:

[0068]

[0069] Therefore, in α * There is at least one So we can get:

[0070]

[0071] Step 5: Relax the variable constraints. For any training sample (x i *y i ), there is always α i =0, or y i (w i *x i +b)=1, but there is almost no completely linearly separable data. In order to solve this problem, the concept of "soft interval" is introduced, that is, some points are allowed not to meet the constraints, that is:

[0072] y i (W i , x i +b)≥1,

[0073] Using hinge loss, the original optimization problem becomes:

[0074]

[0075] where ξ i Each sample has a corresponding slack variable that represents the degree to which the sample does not meet the constraints. C is called the penalty parameter. The larger the C value, the greater the penalty for the classification.

[0076] Step 6: Nonlinear Optimization. For the nonlinear classification problem in the input space, it can be transformed into a linear classification problem in a certain - dimensional feature space through nonlinear transformation, and learn a linear support vector machine in the high - dimensional feature space. By using the kernel function to replace the inner product in the conventional calculation. Let the kernel function be K(x,z), then there exists a mapping from the input space to the feature space For any x, z in the input space, there is:

[0077]

[0078] In the dual problem of linear support vector machine learning, replacing the inner product with the kernel function K(x,z), the obtained solution is the nonlinear support vector machine, and the formula is:

[0079]

[0080] Step 7: Model Training. Input the training data, that is, the training data set T = {(x1, y1), (x2, y2), …(x N ,y N ), where x i ∈R, y i ∈{+1, -1}, i = 1, 2, …, N.

[0081] The output is the separating hyperplane and the classification decision function. Select an appropriate kernel function K(x,z) and penalty parameter C, and construct and solve the convex quadratic programming problem, the formula is:

[0082]

[0083] Finally, the optimal solution is obtained:

[0084]

[0085] III. Weight Incentive

[0086] Combined with the dimensionality reduction method, calculate the weights of the features of the SVM model, screen the feature factors with higher weights, optimize the model weights, and avoid problems such as dimensional sparsity when the SVM maps data from a low - dimensional space to a high - dimensional space. After feature screening and weight discrimination, the final SVM - WS model is obtained for intelligent meter fault prediction.

[0087] Weight Incentive. Optimize the SVM model weights and set the incentive behavior. The specific calculation formula is as follows:

[0088] where E i is the information entropy of the i - th feature, and the calculation formula is:

[0089]

[0090] Finally, SVM-WS is used for the fault prediction of smart meters.

[0091] In the present invention, the SVM algorithm is combined with ideas such as the entropy weight method and PCA dimensionality reduction to transform and upgrade the traditional SVM model. After feature screening and weight discrimination, the final SVM-WS model is obtained for the fault prediction of smart meters.

Claims

1. An intelligent meter fault prediction method based on SVM-WS, which is used for the fault prediction of intelligent meters. The method includes steps of preprocessing data, extracting index features to construct an SVM model, and using the constructed model to predict the faults of intelligent meters, and is characterized in that: After the step of extracting index features to construct the SVM model, the entropy weight method is used to calculate the weights of the SVM model and optimize the model weights. After obtaining the improved SVM model, the fault prediction of the smart meter is carried out; The specific steps of optimizing the weights of the SVM model by the entropy weight method are as follows: Optimize the weights of each feature in the SVM model and set the incentive weight W i , and the optimization formula is: , where is the information entropy of the i-th feature, is the non-linear mapping value of the i-th feature, k is a positive integer greater than 2, and i is a natural number.

2. The intelligent electric meter fault prediction method based on SVM-WS according to claim 1, wherein: In the data preprocessing step, the PCA algorithm is used to reduce the fault dimension of the feature data samples, and the data samples with low data information content are deleted. The judgment steps for low data information content include: a. Perform data standardization and unified processing on the k indicators screened by feature engineering, where k is a positive integer greater than 2; b. Perform a linear transformation on the original sample data, calculate the maximum variance, and select the original samples with the maximum variance value less than the mean of the square root of the feature variance as the data samples with low data information content.

3. The intelligent electric meter fault prediction method based on SVM-WS according to claim 1, wherein: The specific steps for preprocessing the data are as follows: 1). Process the missing data; 2). Process the outliers; 3). Data reduction, including dimensionality reduction and dimensionality transformation.

Images