© The Author(s), published by EDP Sciences, 2025

1 Introduction

Electricity metering is the basis of electricity market operation. Electric energy is a kind of commodity, and the trading behavior of commodity should be subject to market supervision. However, due to its own characteristics, it can not be charged before consumption like traditional commodities, and its entire sales and use behavior is completed at the same time. The non-storable nature of electricity poses difficulties for conventional market supervision and pricing techniques because of grid limits, network effects, real-time supply and demand balancing, and highly controlled pricing. Unlike commodities like oil or metals, these variables complicate price fixing and control, leading to regional price disparities. Whether the electricity energy charging is reasonable is directly related to the economic interests of the electricity production department, the sales and supply department and the electricity customers. With the beginning of a new round of market-oriented reform of China’s energy industry, power grid enterprises will face greater pressure from all aspects of society, and to improve the quality of power supply service and ensure the high quality of power supply has become a top priority for power grid enterprises. Power grid businesses will face pressure to innovate in infrastructure and technology due to the growing demands of customers and regulatory agencies for sustainability and dependability. To raise the calibre of their services, they will spend money on automated systems, smart grid technologies, predictive maintenance, and more precise monitoring systems. Power grid businesses may enhance consumer transparency and dependability under the current reform conditions by integrating renewable energy sources, digitizing their operations, and introducing demand response programs and customer-centric services. The power measurement data is not only the basic condition for the smooth development of the marketing work of the power grid company, but also the indispensable basic data for the operation and data analysis of the power grid. Power grid businesses can use voltage stability, customer feedback, energy loss reduction, and reliability measures to measure changes in the quality of the power supply. These measurements, which monitor voltage swings, power outages, complaint resolution rates, and energy loss reduction, show advancements in theft prevention and grid efficiency. And the metering equipment belongs to the power secondary equipment, if there is a serious failure, it will also lead to the safe operation of its associated primary equipment is affected, resulting in power grid security accidents, resulting in power supply interruption. Inaccurate power measurement can result in operational inefficiencies, revenue losses, and grid imbalance. Blackouts, equipment damage, and imbalances in supply and demand can all result from it. Unfair billing and increased operating expenses might result from theft concealed by inaccurate billing. Measurement errors may potentially hamper grid optimization attempts. However, from the reality point of view, due to the influence of human factors and non-human factors, the stability and accuracy of the operation of the electric energy metering device have been tested more and more.

Zhuang et al. reduces the data to two dimensions using PCA (Principal Component Analysis) and then uses the Local Outlier Factor (LOF) algorithm. Based on the local density deviation of a data point relative to its neighbors, the Local Outlier Factor (LOF) method finds abnormalities. For it to function, the density surrounding a data point must be measured and compared to that of its neighbors. Because of its non-parametric, unsupervised detection and local sensitivity, LOF benefits power-use datasets with changing data density and big, unlabeled datasets. To identify outlier objects (abnormal power data that are different from normal power data). It can be seen that most of the current machine learning methods used in the measurement anomaly diagnosis are relatively simple [1]. Several factors are used in power metering anomaly identification, including energy consumption patterns, peak demand, power factor, load variability, temporal aspects, and user profile data. These offer context for finding irregularities in particular user segments and aid in recognizing departures from typical consumption, possible theft, or meter manipulation. Coma-Puig et al. have compared many common machine learning methods, such as Bayesian classifier, KNN (K-Nearest Neighbor), SVM (Support Vector Machine), etc. It is found that with sufficient calibration samples, the accuracy rate of these different methods can reach 90% [2]. Zhang et al. proposed an anomaly diagnosis model based on deep neural network, which can directly input the data collected by the original sensor without feature selection and signal processing [3]. A model’s performance, scalability, learning curve, efficiency, and resource management must all be assessed through incremental data testing. Contrasting performance with various data sizes helps find plateaus and gains, maximizes resource utilization, and finds overfitting and underfitting. The model can process the original data of the sensor with high dimensions, and reduce the work of feature selection through data segmentation and network classification. It reduces the difficulty of training, but it requires enough correctly calibrated anomaly samples. Kou and Cui used the characteristics of recurrent neural network processing time series data and decision tree for fault diagnosis of automatic meter reading system, and integrated recurrent neural network and decision tree through traditional Adaboost method to improve the performance of fault diagnosis [4]. When many machine learning techniques are combined and work well together, feature engineering is superior, and there is enough amount and quality of data to enable efficient training, prevent overfitting, and capture various elements of the data. Neural networks paired with density and distance measures provide multiple benefits for anomaly identification. They improve each other’s sensitivity and robustness to noise, making it possible to identify outliers in dense data areas. Divina et al. applied the integration method to short-term load prediction by integrating multiple weak learners to achieve a robust integration model [5]. The robustness of models against noise or incomplete data may be improved by employing strategies like k-fold cross-validation, outlier identification and removal, imputation for missing data, data augmentation, and ensemble approaches. These techniques can lower bias, eliminate noisy data points, and improve generalization.

Liu et al. proposed a network traffic anomaly detection model, quantified network traffic feature attributes by entropy, and then used k-means algorithm to cluster abnormal traffic to obtain the detection model [6]. Wu and Zhang uses an improved k-means clustering algorithm to compare the density of each object with the average density of all objects [7]. If the density of the data object is less than the average density of the data set, and the distance from the object to its cluster center is greater than the average distance from all objects in the cluster to the cluster center, The object is considered to be an abnormal object. Ren et al. combined with the fault location inference knowledge base, proposed a feature information representation method, and built a secondary device fault location model by using the characteristics of recurrent neural network, which is good at processing timing information [8].

Both machine learning and deep learning methods can achieve more satisfactory results when the number of labeled samples is sufficient, the sample distribution is uniform, and the sample noise is small [9]. Large datasets are needed for training deep learning models like Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM), which work well with sequence or picture data. Smaller datasets or ones lacking intrinsic temporal or spatial features are not suited for them because of their high computing cost and interpretive complexity. Because of its interpretability and balance, SVM and Random Forest are better suited for datasets. Thirusubramanian, through enhanced algorithms and anomaly detection, artificial intelligence (AI) has greatly improved financial fraud detection in Internet of Things contexts. This work aims to develop reliable fraud detection models by investigating the technique, datasets, and assessment metrics needed for adaptive learning, which incorporates automated reaction mechanisms and regular retraining [10]. Naresh assesses the performance of four machine learning models – Gradient Boosting, Random Forest, Polynomial Regression, and Linear Regression – to forecast e-commerce trends, enhance inventory management, refine marketing strategies, and improve demand prediction accuracy using a hybrid approach [11]. Naresh, a framework for hybrid machine learning, has been created to identify instances of financial fraud in the digital economy, namely on e-commerce sites. The system continually monitors and adjusts to new fraud behaviour by utilizing neural networks, decision trees, and support vector machines to increase detection accuracy and reliability [12]. However, in the scenarios where the data samples are insufficient and the data cannot be expanded by traditional data enhancement methods, the above methods have problems such as low model classification accuracy and poor performance. Therefore, this paper proposes a multi-model fusion of power metering data anomaly detection method for the power metering anomaly problem.

The innovation of this paper is as follows: firstly, the advantages and disadvantages of the machine learning algorithms SVM and RandomForest are analyzed, and the problems of high dimensional data, poor anti-noise ability and poor fitting ability exist in the process of checking the anomaly of the energy measurement data by using a single model are pointed out [13]. In order to further improve the detection performance, a multi-model fusion method for anomaly detection of electric energy measurement data is obtained by summing the probability results calculated by different models according to different weights. Then, by designing incremental model, the performance of different models in different data amounts is analyzed, and the effect of model fusion is proved. Finally, the abnormal data set is established, the criterion is set, the performance of the fusion model is compared with that of the single model, and the corresponding research conclusion is drawn.

2 Electric energy metering machine learning algorithm

2.1 SVM

SVM (Support Vector Machine) is a type of classification problem designed to solve specific data vectors, the algorithm is essentially to find the obvious support vectors between the vectors involved in the computation, which can be utilized to delineate the obvious blocks between the data [14]. The purpose of this paper is to classify whether or not the electricity measurement is abnormal, and the SVM’s excellent ability to utilize the samples can still be useful even if the sample size of users is small enough.$$\begin{array}{l}\underset{\omega ,b}{\min }\frac{1}{2}{\left|\right|\omega \left|\right|}^2\\ \mathrm{s.t}.y_i\left({\omega }^T{x}_i+b\right)\ge 1,i=\mathrm{1,2},\dots ,m.\end{array}$$

In the above formula, the SVM essentially optimizes the distance of its support vector boundary while ensuring that the predicted class is the same as the actual class. Even if the data is not linearly separable in the existing space, it can be mapped to a higher dimensional space by designing a kernel function to find support vectors.

From the above formula, it can be seen that SVM is strictly based on the user data vector itself to find support vectors, but it is difficult to discretely data the data in the optimal form under the actual use of user electricity, which leads to the obvious fluctuation of SVM performance [15]. In addition, corresponding to its excellent performance in the case of small sample size, SVM consumes an exaggerated amount of resources in large-scale training, so the simple SVM algorithm is difficult to adapt to complex power consumption data. This paper will make up for its shortcomings in vector expression and data sample size by combining the following two algorithms, so that the algorithm can be applied to different scenarios.

2.2 Random forest

RandomForest uses an integrated structure with a decision tree as the base model, which is an integrated learning algorithm statistical theory based on a decision tree, which utilizes Bootsrap sampling to draw k different samples from the initial sample, here the training sample and the initial sample require the same sample size, and then builds a corresponding decision tree model for each of the k samples to obtain the corresponding k classification results, and use the decision tree to model each sample one by one [16]. The Random Forest approach trains many decision trees on various data subsets through ensemble learning. Random selections of training data and features are used, together with bootstrapping, to generate a variety of models and an ensemble learning procedure to combine predictions and increase resilience and accuracy by offsetting individual tree faults. The influence of individual noisy characteristics is lessened by Random Forest’s averaged predictions from several decision trees, making it resilient to noise. Nevertheless, dimensionality reduction methods such as PCA are required when dealing with high-dimensional data.$$\begin{array}{l}\mathrm{Ent}\left(D\right)=-\sum _{k=1}^{\left|y\right|}\mathrm{}{p}_k{\log }_2{p}_k\\ \mathrm{Gain}\left(D,a\right)=\mathrm{Ent}\left(D\right)-\sum _{v=1}^V\mathrm{}\frac{\left|D^v\right|}{\left|D\right|}\mathrm{Ent}\left(D^V\right).\end{array}$$

In the formula, Ent(D) is the information entropy of the current node, p is the statistical probability, Gain(D, a) is the learning information gain, D is the current node, and a is the partition dimension.

The integration of random forest is also the basis for better anti-noise ability than a single decision tree. Compared to a single decision tree, Random Forest produces more stable results by reducing noise influence, selecting random features for each split, averaging predictions to smooth out noisy predictions, and building multiple decision trees using bootstrapped data samples. Therefore, under the premise of obvious noise of actual industrial data, the adoption of random forest can combine the advantages of decision tree and avoid the troubles caused by data itself [17]. Of course, the advantage of random forest lies in its ability to generalize missing data. Considering the statistical errors of electric meter data and the uncertainties such as special power failure, such errors can be offset as much as possible by adding random forest. But at the same time, random forest usually does not have a particularly obvious advantage in the case of high dimensions.

3 A multi-model fusion algorithm for anomaly detection of electrical energy metering data

The idea of multi-model fusion algorithm for anomaly detection of electric energy measurement data constructed in this paper is to sum the probability results calculated by different models according to different weights. The multi-model fusion strategy enhances anomaly detection performance in intricate datasets such as power theft by combining predictions from machine learning models, such as SVM and RF. This approach strengthens each model’s unique shortcomings while improving overall robustness, accuracy, and generalisability to increase effectiveness. The multi-model fusion approach combines the advantages of each model to enhance detection performance. It provides a range of viewpoints, improved accuracy through model voting or averaging, and resilience to noise and fluctuating data. Random Forest employs ensemble diversity to handle complicated datasets, whereas SVM concentrates on well-defined decision boundaries.$$\begin{array}{l}S={\omega }_1{M}_{\mathrm{svm}}\left(X\right)+{\omega }_2{M}_{\mathrm{rdf}}\left(X\right)\\ \mathrm{s.t}.\sum _i\mathrm{}{\omega }_i=1.\end{array}$$

In the formula, S is the final output probability, M_svm(X) is the probability output of SVM algorithm, M_rdf(X) is the probability output of Random Forest algorithm, and ω is its corresponding weight. The best algorithms for anomaly detection are SVM and Random Forest because of their interpretability, resilience against outliers, efficacy in high-dimensional environments, and unambiguous decision bounds. SVM works well with complicated datasets and has good outlier handling capabilities. Random Forest is easier to perceive and comprehend than Neural Networks and Gradient Boosting.

And balance these differences by weight, try to make the single model play their own advantages, and pull the balanced data from the classification boundary to the obvious classification interval, so that the fuzzy data of the original single model can be corrected [18]. In order to train the model below, the basic training process built in this paper is shown in Figures 1 and 2. Effective preprocessing, including managing missing data, outlier identification, and cleaning, guaranteed the model’s success. Automated scripts were utilized to impute missing values, identify outliers, and guarantee dataset integrity using mean/median imputation, k-nearest neighbours, and Random Forest techniques.

Fig. 1 SVM model basic construction training process.

Fig. 2 Basic training flow of random forest model.

To verify the fit of the model to the data for different data volumes, this paper proposes an incremental learning approach so as to obtain the minimum requirement of the data volume in order to determine the weights of the model fusion in real electricity consumption data with different data volumes [19].

4 Experimental analysis

4.1 Build the dataset

All the data sets used in this experiment were collected from the public database of the laboratory, and part of them consisted of the electricity consumption data of 1000 users during 2021–2023, and the collection frequency was half an hour. Because all the data were collected with the consent of the users, the data were benign users, and there was no stealing of electricity. According to the generation method of electric theft data, electricity consumption examples of electric theft users are generated to assist the experiment for analysis. The unbalanced situations in electricity theft detection datasets sometimes call for creating synthetic data to enhance model training. Creating realistic synthetic samples entails modelling tampered meter data and adding anomalous patterns using methods like SMOTE or GANs. The specific generation method is as follows:$$\begin{array}{cc}{h}_1\left(x_t\right)=\alpha x_t,& \alpha =\mathrm{random}\left(\mathrm{0.1,0.8}\right)\end{array}$$ $$\begin{array}{cc}{h}_2\left(x_t\right)={\beta }_1{x}_t,& {\beta }_t=\{\begin{array}{cc}0& \mathrm{start}-\mathrm{time}<t<\mathrm{end}-\mathrm{time}\\ 1 \mathrm{else}& \end{array}.\end{array}$$

Using the above formula, multiply the data sample by random values in the same range, so that the smart meter does not record the measurement value or the record value is 0 for a randomly selected time, where the start time is from 0 to 23 minus the minimum duration, which is 4 h in this formula [20]. Finally, the time of the sample data is inverted within 24 h.

After the data is collected, the process as shown in Figure 3 is used to process the data. Combined with the power theft data generation method, the data for network training and verification is generated. The 535 benign data set samples were divided into 458 samples for training and 77 samples for testing. One sample was randomly drawn every 7 consecutive days for the test set and the other 6 training sets. Again, for each abnormal load, 535 samples are generated. To sum up, the test set contains 3287 samples with a scaling range [−1, 1]. Min-max normalization is a scaling technique that prevents dominance in algorithms and assures uniform feature contribution to a model’s learning process by scaling data into a predetermined range, usually [0, 1]. It is appropriate for datasets and models that are sensitive to data magnitude.

Fig. 3 Data processing flowchart.

4.2 Evaluation criteria

In this paper, we use the Receiver Operating Characteristic (ROC) curve as an evaluation criterion for the performance of the model, FPR is the horizontal axis of the curve, True Positive Rate (TPR) is the vertical axis of the curve [21]. The formula is as follows:$$\begin{array}{c}\mathrm{FPR}=\frac{\mathrm{FP}}{(\mathrm{FP}+\mathrm{TN})}\\ \mathrm{TPR}=\frac{\mathrm{TP}}{(\mathrm{TP}+\mathrm{TN})}\end{array}$$

The above formulas are the horizontal and vertical coordinates of the ROC curve, and FP, TN, and TP represent the negative samples of wrong, negative samples of right, and positive samples of right, respectively. AUC (Area Under Curve) represents the area under the ROC curve. AUC value is used to judge the strengths and weaknesses of classifiers, i.e., prediction models:$$\mathrm{AUC}=\{\begin{array}{ll}=1& \mathrm{Predictive} \mathrm{model} \mathrm{perfection}\\ \in \left(\mathrm{0.5,1}\right)& \mathrm{Have} \mathrm{predictive} \mathrm{value}\\ =0.5 \mathrm{Same} \mathrm{as} \mathrm{random} \mathrm{guess}& \\ <0.5 \mathrm{Worse} \mathrm{than} \mathrm{random} \mathrm{guess}& \end{array}.$$

4.3 Result analysis

4.3.1 Single model analysis

Before analyzing the fusion model, first analyze the performance of the single model on the data set using the incremental method. Due to their susceptibility to scale, SVMs have speed, feature scaling, and scalability limits. Fusion models minimize overfitting in noisy data and strike a compromise between SVM’s advantages and Random Forest’s effectiveness. By generalizing to previously unknown data, this model enhances performance. Scaling, normalization, feature selection, and addition of additional features were among the feature engineering approaches used to optimize the dataset for Random Forest and SVM. Methods including correlation analysis and recursive feature elimination were applied to lower dimensionality, increase classification accuracy, and find anomalies. The incremental accuracy rate of SVM model is shown in Figure 4, and the incremental accuracy rate of Random Forest model is shown in Figure 5.

Fig. 4 Increment accuracy rate of SVM model.

Fig. 5 Incremental accuracy rate of random forest model.

From Figure 4, it can be seen that the SVM model can still achieve ideal results under the condition of training with only 10% training data to 90% training data, indicating that the SVM is still competent for basic objectives under the condition of small samples, and is suitable for work when there is not enough user data in the initial stage of the establishment of the sampling equipment [22]. Generally, SVM is relatively stable when the amount of data increases, that is to say, when the amount of data increases slightly, SVM will not become a burden, and can still provide strong support for the fusion model.

From Figure 5, the random forest model achieves the same training effect as SVM when only 10% of the training data is trained to 90% of the training data, so it can be analyzed that the dimensionality information of the extracted data is still within the scope of the decision tree and will not cause serious noise resulting in unadaptability [23]. At the same time, it is found that the training effect is slightly decreased when the data is 40%. It can be seen that some samples tend to have a big impact on the model. Therefore, when the data is not perfect enough, the random forest cannot independently perform the recognition task, and auxiliary models are still needed. The fusion model exceeds individual models in accuracy because it combines the robust classification of Random Forest with the exact decision bounds of SVM. However, it requires more computing power because two different models must be used for prediction and training. Nevertheless, because of its ensemble variety, it may generalize more effectively than individual models.

4.3.2 Result analysis of multi-model fusion

Since SVM and random forest algorithm with relatively stable performance are not sensitive to perturbations of data set changes, they are equipped with 1/2 and 1/3 weights respectively. For the first time, the performance of the project with such an allocation exceeded the historical best of the two single models, as shown in Table 1.

The historical best accuracy of SVM, random forest and fusion model is 88.1%, 87.7% and 90.4%, respectively. The reason may be that there are many adjacent data in the data vector, that is, purely from the data law, the model is close to the probability judgment of the existing data. Therefore, in the case of using more model voting, the relatively confident model can balance the doubtful data to a large extent [24]. In settings with dense populations, model voting can lead to less sensitive detection, outcomes that are too smooth, and bias in favour of the dominant class. Higher false negatives come from its tendency to ignore minute variations over smooth data and favour the dominant class unfairly. For example, the probability of the existing data being divided into positive and negative samples on SVM is 0.51 and 0.49, but such probability is significantly separated in random forest due to the existence of large information gain. For example, the probability of positive and negative samples is 0.2 and 0.8, so if only two models are considered, they will eventually be divided into negative samples in the actual environment. The actual result will be more inclined to the confident model, so when the two models are trained as much as possible, the tolerance of the worse samples will be lower, so as to achieve further optimization of the actual effect. The actual data also proves the feasibility of the idea designed in this paper. The optimal AUC values of the three models are shown in Table 2, and the ROC curve of the fusion model in actual conditions is shown in Figure 6.

Fig. 6 ROC curve of fusion model.

The ROC curve of the fusion model in the actual situation has the best performance, with AUC reaching 0.9653 and TPR exceeding 0.64 in the case of zero False Positive Threshold (FPT). The comprehensive performance is better than that of other individual models.

In order to further verify the performance of the fusion model presented in this paper, random verification was carried out on the data of normal and abnormal power users. Single user data included two groups of normal and abnormal power users, including the peak, normal, valley, load, line loss, voltage, current and power factor of electricity consumption [25]. The two groups of data are determined by the experienced staff as A group of normal user A and a household as abnormal user B. The two groups of data are respectively determined by the three models described above. The determination results are shown in Table 3.

As can be seen from the data in the table, for the correct data, the three models are judged correct, but for the known abnormal customer data, RandomForest judged to be normal, and the staff according to the experience of judgment is contrary to the situation, so immediately arrange the work plan, go to the user A and B, the actual scene investigation, after the scene situation and data carefully compared, found that user A is a normal power customer, user B is an abnormal power customer, and through the previous analysis of the highest rate of the fusion model proposed in this article, two different cases of the user are judged correct, so selected as the optimal model in line with the actual needs of a single user to determine the situation.

5 Power measurement data anomaly correction scheme

The main causes of abnormal power measurement data are human factors and non-human factors. For the abnormal power measurement data detected, the following programs can usually be adopted to correct it. Machine learning techniques that handle noise and unpredictability in power measurement data include Random Forests and SVM. Random Forests filter out noise by averaging forecasts from numerous decision trees, SVMs utilize margin optimization and kernel functions to detect outliers, and Deep Learning models such as LSTM may find trends in data on power usage.

5.1 Inspection and detection of electric energy metering devices

The inspection and testing of electric energy metering devices mainly includes the pre-operation verification and filing of electric energy meters and mutual inductors [26]. The first verification and periodic verification after operation, fault handling, verification work after the replacement of main equipment. The specific content of the above work is clearly stipulated in the “JJG448-Electric Energy Measurement Technical Management Regulations”, which should be carried out in strict accordance with the requirements. This work is the most direct means to check and determine the abnormal measurement.

5.2 Electric energy remote acquisition system data analysis and early warning

The electric energy remote acquisition system has been fully built in the power grid enterprise, and has the ability to achieve 100% full collection. Through the comparative analysis between the collected data and the limit value, the early warning of the electric energy can be realized.

5.3 Power inspection work

Electricity inspection work is divided into periodic and temporary, periodic work is generally arranged and executed according to the user’s voltage level and power consumption, temporary work is arranged and carried out quickly because of sudden safety accidents, major activities and received reports [27]. Compared with the previous single means of looking for abnormal power customers, with the help of the data early warning of the electric energy remote acquisition system, the efficiency of the power inspection has been greatly improved.

5.4 Processing of abnormal power consumption behavior and charge catch-up

For the abnormal power consumption behavior verified on site, the staff will analyze the cause of the anomaly and determine whether it is man-made or non-human anomaly [28]. If the anomaly is caused by humans, people should be investigated for responsibility, take photographs on the scene, fill in the relevant documents for accident identification, and be signed by the responsible person, sent to the electricity accounting team for electricity loss recovery; If it is not man-made anomaly, it should first eliminate the fault point, ensure the operation of the main equipment, carry out remedial measures in a timely manner, restore the normal operation of the measuring device, minimize the scope of abnormal data marketing, analyze the anomaly afterwards, and carry out targeted investigation and rectification of the same equipment according to the anomaly.

6 Conclusion

To sum up, the abnormal measurement of power in the power grid is directly related to the actual economic interests of both sides of the power supply. How to accurately and quickly locate the metering point where the abnormal measurement occurs in the abnormal detection of power measurement data is the key to improve the solution and processing of the entire power measurement anomaly. Based on the advantages and disadvantages of SVM and Random Forest big data analysis methods, this paper adopts machine learning method to propose a multi-model fusion abnormal detection method of electric energy metering data, which deeply integrates the two machine learning algorithms, improves the discrimination and positioning accuracy of abnormal electricity consumption data, and verifies the performance of the model by using actual user data. The verification results show that the fusion model has the best performance in the actual situation.

In this paper, machine learning algorithm is used to give an anomaly detection method of electric energy measurement data, although certain research results have been achieved, but the problem of electric energy measurement anomaly involves two different fields of technology and management, and in practical application, the two are interrelated and inseparable, and this paper mainly discusses this problem from the technical aspect, the practical application effect is still unknown. With the construction of the ubiquitous power Internet of Things, the power metering device is still the key to data collection, with the upgrade of the terminal equipment, the collected data will be diversified, bringing more detailed analysis of user behavior.

Funding

No funds, grants were received by any of the authors.

Conflicts of interest

The authors report no declarations of interest.

Data availability statement

All data generated or analyzed during this study are included in the manuscript.

Author contribution statement

All author have contributed to the design and methodology of this study, the assessment of the outcomes and the writing of the manuscript.

References

All Tables

All Figures

	Fig. 1 SVM model basic construction training process.
In the text

	Fig. 2 Basic training flow of random forest model.
In the text

	Fig. 3 Data processing flowchart.
In the text

	Fig. 4 Increment accuracy rate of SVM model.
In the text

	Fig. 5 Incremental accuracy rate of random forest model.
In the text

	Fig. 6 ROC curve of fusion model.
In the text