More recently, with the gradual transformation of enterprises from product oriented to user oriented in the production, it is significant for formulating marketing strategies and product design to fully understand customers and their needs (Wu et al., 2020). At present, in-depth mining of user data and research on accurate user portrait are gradually deepening. Accurate user portrait technology has been widely used in the Internet, finance, retail, operators, advertising, and other industries (Liu and Du, 2020; Zhang et al., 2020). In terms of advertising, Google uses big data to locate high-value users, which helps mobile e-commerce app to locate high-value users and push advertisements and adopts flexible data tracking methods for users of different tag categories, so that brands can achieve accurate advertising push for target users (Shan, 2018). In terms of auxiliary business decision-making, eBay carries out business circle customer group analysis, product marketability analysis, store operation analysis, personalized consumption analysis, and customer loss analysis by integrating online and offline massive data, providing decision support for commercial real estate and comprehensive big data analysis and prediction for project parties and brands (Chen et al., 2021). Obviously, user portrait technology has become an effective method to improve customer service quality and customer experience, which is an important basis for integrating high-quality resources and realizing enterprise and user value (Pitner et al., 2012; Yu et al., 2017).

In the field of electric power research, it has become a rising research direction to build user portraits in different application scenarios according to the actual needs (Qiu et al., 2017). In the early stage of the study, traditional user portraits were mainly used in the marketing portal to build user electricity sensitivity or credit portraits to guide the electricity recovery work (Sanchez et al., 2008; Han et al., 2014; Ampimah et al., 2017). To meet the requirements of demand response, the methods of establishing electricity consumption behavior tag library and realizing the portrait of different types of users' electricity consumption behavior patterns were proposed in (Qiu et al., 2017) and (Zhong et al., 2018). With the gradual popularization of new energy (Yang et al., 2016; Yang et al., 2017; Yang et al., 2018; Yang et al., 2019a; Yang et al., 2019b; Yang et al., 2020), the load change and influence mechanism on the user side are becoming more and more complex. Therefore, the user portrait method is becoming more and more important.

The above pieces of literature are all portraits of the users who have been connected to the power grid, so as to formulate appropriate strategies to guide the corresponding users to change their electricity consumption behavior. However, there is rare research on the reported customers who are not connected to the power grid. In the distribution network planning work, reference (Lian et al., 2014) makes use of the complementarity between user loads to optimize the access decision reasonably, which can effectively improve the load distribution and utilization rate of power supply equipment. Moreover, the multidimensional analysis of user load can also provide effective guidance for the optimal scheduling and control of smart grid (Yang et al., 2015; Zhang et al., 2015; Xi et al., 2016; Zhang et al., 2016; Zhang et al., 2021). However, due to the lack of research on the load of customers who are not connected, it is limited to transfer the user load of the existing distribution network. For the more common scenario of business expansion, the data information support and optimization ability are insufficient.

In view of the above research results, the contributions of this paper can be summarized as follows:

1) Unlike the traditional single user portrait, this paper digs the power consumption behavior law of massive users, which constructs a more popularized and applied power consumption behavior portrait for industry.

For continuous load data, Lagrange interpolation (Criscuolo et al., 1984) is used to fill the vacancy and bad data, which can play a good repair effect. For the data points that need to be filled, m normal data points adjacent to the point are selected as samples to calculate the Lagrange interpolation formula: L ( t ) = ∑ i = 1 m x i ∏ j = 1 , j ≠ i m ( t − t j ) ∏ j = 1 , j ≠ i m ( t i − t j ) , (1) where t i and t j are the sampling time of the i and j data points in a day, x i means the load value of the i data point, and m is the number of sample points used to construct the polynomial. Finally, the approximate value of the missing value can be obtained by substituting the missing time t into L ( t ) .

To avoid the influence of the actual load value in the follow-up analysis process, the linear proportion normalization method is used to normalize the load data after repair, which can be expressed as x i ^ = x i x m a x , (2) where x i ^ represents the i -th load data point after normalization and x m a x is the peak load of the user.

To serve distribution network planning, the results of single user load characteristic analysis lack universality. Thus, it is necessary to explore the general law of massive users’ load changes.

Generally, the users are classified into various industries according to the National Economic Industry Classification national standard, forming a user group in each industry, which has similarity in the long-term development law of load. In this way, the summary of the law and subsequent application is facilitated. After the preliminary classification of users is completed, due to the different characteristics of user peaks and valleys, it is necessary to further carry out industry power mode analysis and feature extraction.

After the load analysis from the user level to the industry level, the load curves and related characteristics of the power users and industry modes are obtained, respectively, at the user level and the industry level, which is the description of the electricity consumption behavior of the industry from different angles and at different levels.

To enable the distribution network planners to grasp the electricity consumption of various industries intuitively, massive analysis results are sorted and visualized. Then, a panoramic portrait of industry electricity consumption behavior in a region is built. The form and content of the portrait are shown in Figure 2.

In the field of the classification problem, ensemble learning (EL) has gained the attention of a huge number of scholars, because it can make up for each single model's advantages (Wang et al., 2015). Among them, bagging (Hu et al., 2011) and boosting (Lu et al., 2006) are the most classic and widely used integration methods, which have their own characteristics in model generalization and model accuracy. Specifically, bagging builds multiple datasets in bootstrap way and trains multiple single models in parallel, integrating the output by Voting. The advantages of each model are integrated, preventing the overfitting phenomenon effectively. Boosting framework belongs to the mode of serial integration of models. The output of each model is used as the input of the lower model, and the minimum deviation is used as the loss function to continuously improve the accuracy of each model. Generally, the characteristics of the two integration methods are shown in Table 2.

To give full play to the advantages of the two integration methods, a model fusion framework of BB-stacking (bagging and boosting in stacking) based on the stacking method is proposed. It combines the two integration methods reasonably to realize the organic trade-off between high-precision and strong generalization performance, yielding the high-precision classification of customers.

Traditional stacking is composed of the base model in the lower layer and the metamodel in the upper layer. Firstly, the initial data is divided into subdatasets by k-fold, and then the subdatasets are input into the base model for training and classification. Then the output of the lower layer is reconstructed and sent to the upper model for training and classification. The BB-stacking model fusion algorithm proposed in this paper changes the lower base model layer into the integration layer on the basis of the two-tier structure of stacking, in which bagging and boosting methods are used to preliminarily integrate the original model and then transferred to the upper metamodel to coordinate the advantages of the two integration modes, obtaining a model with higher accuracy and more stable generalization ability.

In the traditional business process, customers provide limited information in the information filling process. To make the pattern classification of electricity consumption more accurate, this paper proposes an appropriate expansion of the electricity consumption information reported by the customers. In addition to the traditional information of electricity consumption type, electricity consumption location, and electricity consumption capacity, the peak electricity consumption time schedule, peak electricity consumption level estimation, and valley electricity consumption level estimation are added as the customer's electricity consumption information. Table 3 shows the demonstration of extended power consumption information of a customer.

According to the example of extended electricity consumption information of the reported customer in Table 3, the reference load curve of the reported customer is obtained by time simulation. According to the definition of the peak-valley characteristic index in Table 1, the peak-valley characteristic vector of the customer can be calculated as the input of the electricity consumption pattern classification model.

The next step is to calculate the load level of the customers. Based on the regional industry electricity consumption behavior portrait, the corresponding method is adopted to calculate the customers’ load according to their type of electricity consumption (Yaoyao et al., 2013). The specific calculation method is as follows:

1) Commercial, nonindustrial, and residential customers

It is recommended to use the load density method to calculate the customers' load because the load level of these customers is closely related to their building area. Assuming that the classification result of the customer is industry electricity consumption pattern C u , the customer accounting load can be computed by P h s = D u ¯ ⋅ S b z , (12) where S b z means the building area of the customer.

2) Industrial customers

Generally, the load level of industrial customers often depends on the capacity of production equipment and has weak correlation with the building area. It is recommended to use the utility rate method to calculate the customer load. Similarly, assuming that the classification result of the customer is industry electricity consumption pattern C u , the customer accounting load can be computed by P h s = ϵ u ¯ ⋅ P b z , (13) where ϵ u ¯ stands for the typical utility rate value of the electricity consumption pattern C u ; P b z means the building area of the customer.

After the customer load level estimation is completed, the load curve can be forecasted according to the electricity consumption pattern classification results of the customer in the industry. The forecasting load curve of the customer belonging to electricity consumption pattern C u is L y g , and the calculation formula is as follows: L y g = L u c ⋅ P h s , (14) where L u c is the typical load curve of industry electricity consumption pattern C u .

The customer load forecasting technology is based on the industry electricity consumption behavior portrait of the region. Through the limited electricity consumption information, customer load after access is estimated. To provide guidance for planners to make customer access decisions, the accuracy of the expected results needs to be evaluated. Assuming that the actual load curve of the customer is L s j , maximum absolute error ratio (MAER) and Euclidean distance (ED) are used to evaluate the accuracy of the forecasting results.

1) To directly reflect the maximum error level between the forecasting load curve and the actual load curve of customer, MAER is used to measure the error, which can be computed by

To verify the effectiveness of load forecasting technology for new customers, 141 metal products’ users with complete electricity consumption data are selected as hypothetical customers. Therefore, the corresponding characteristic vectors are extracted and input into the trained BB-stacking model for electricity consumption pattern classification. In this paper, the traditional decision tree model (Safavian and Landgrebe, 1991), GRNN model (Specht, 1991), and PNN model (Oh and Pedrycz, 2002) are used to compare with the BB-stacking model. The classification accuracy under different training sample sizes is shown in Figure 8.

It can be found that, in the training process, the BB-stacking classification model can achieve a better classification effect than other classification models in a small training sample size. When the training sample size increases, the BB-stacking classification model can maintain a more stable generalization ability. It is obvious that the BB-stacking classification model prevents the phenomenon of classification ability decline caused by underfitting, performing better classification effect and stability.

After the electricity consumption pattern classification of the customers is completed, the forecasting load curves of 141 customers are obtained through Eqs 12–14 combined with the typical electricity consumption characteristics of the industry. To compare the effectiveness of various methods, the similarity between the load curve forecast results and the actual load curves is evaluated by EM and MAER, and the comparative analysis is carried out from the following two aspects.

1) Comparative analysis of different clustering numbers

Considering that the different number of clusters selected by the clustering algorithm will lead to different electricity consumption patterns, which may affect the accuracy of the forecasting results of customers, this paper uses the GWO-FCM algorithm to test the case of cluster number c from 2 to 8. The mean value of EM and MAER under each clustering number is compared and shown in Figure 9.

2) Comparative analysis of different methods

To further study the effectiveness of the proposed customer load forecasting method, this paper compares the traditional load estimation method with the proposed method. In addition, the sectional load curve obtained by the segment simulation method in the process of electricity consumption pattern classification is compared with the proposed method in this paper.

The comparison between the load curve forecasting result of a certain customer and other methods is shown in Figure 10. It can be found that there is little difference between the maximum load level of the estimated load and the actual load curve, but there is no time sequence characteristic. To some extent, the segment simulation method reflects the time sequence characteristics of customers, but the difference is large when the load level changes. The overall trend of the load curve forecasting result is very close to the actual load curve, which can meet the optimal access requirements in the planning field considering the time sequence characteristics of customers’ load.

Comparing the error index of actual load curve and load curve forecasting results under different methods, the mean value comparison of test samples is shown in Table 5. It can be found that the proposed method performs best in both of the two evaluation indexes. Significantly, the accuracy is greatly improved compared with the traditional load estimation method.

Counting the two evaluation indexes of load estimation results of each hypothetical customer under different methods, two box diagrams are drawn to observe the distribution of sample indexes of each method, which are shown in Figures 11, 12. According to the box diagrams, the sample median level of the two error evaluation indexes of the traditional load estimation method is the highest, and the error fluctuation range is the widest; the results show that the segmented simulation method achieves the suboptimal effect in the two evaluation indexes, and the error fluctuation is also controlled in a small range. It is worth noting that the median of the evaluation index samples of the proposed method is the lowest, the falling range of the error index samples is the narrowest, and the fluctuation range of the algorithm performance is smaller, showing the best stability.