Demand response potential customer intelligent recommendation system and method based on multi-dimensional feature fusion

Abstract

The invention discloses a demand response potential customer intelligent recommendation system and method based on multi-dimensional feature fusion, and belongs to the technical field of power demand side management. The system comprises a multi-source data acquisition module, a feature engineering module, a potential evaluation model and a value maximization recommendation module. Multi-dimensional information such as load characteristics, response willingness, industry characteristics and the like is creatively fused, high-precision potential evaluation is realized by adopting an improved random forest algorithm, and a recommendation list is generated in combination with long-term value prediction. According to the method, the problems of inaccurate evaluation, low recommendation value, cold start of new customers and the like of a traditional method are solved, and the accuracy and efficiency of demand response customer screening are remarkably improved. Experiments show that compared with a traditional method, the evaluation accuracy of the method is improved by more than 25%, and the average life cycle value of signed customers is improved by 30%.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of power demand side management, and relates to a demand response potential customer recommendation system and method, which is especially suitable for intelligent screening and recommendation of massive users under the background of new power systems to maximize the signing value and the compliance stability. The present application combines multi-dimensional load behavior characteristics, industry attributes and environmental sensitivity characteristics to construct a generalizable intelligent evaluation model, realizes the collaborative prediction of demand response potential and signing value and the optimal customer recommendation, and thus improves the accuracy and efficiency of demand response customer screening. BACKGROUND

[0002] With the improvement of the intelligent level of power systems, demand response as an important means to adjust the balance between power supply and demand has attracted widespread attention. However, there are still many problems in the screening and recommendation of potential demand response customers at present.

[0003] 1. Low efficiency of customer screening: the traditional method relies on manual experience to screen potential customers, which is low in efficiency and strong in subjectivity.

[0004] 2. Inaccurate response potential evaluation: existing evaluation methods are mostly based on a single dimension (such as load characteristics or historical response rate), which is difficult to fully reflect the actual response potential of customers.

[0005] 3. Signing value not fully considered: the current recommendation system focuses on customers with high response potential, and lacks comprehensive consideration of the long-term economic value and compliance stability of customers, which easily leads to low overall value of signed customers.

[0006] 4. New customer cold start problem: for new users who lack historical participation records, existing methods are difficult to accurately evaluate their potential, and the recommendation system lacks support for new customers, causing resource waste or missing high-value users.

[0007] In view of the above problems, the present application proposes a demand response potential customer intelligent recommendation system and method based on multi-dimensional feature fusion, which realizes the accurate identification and recommendation of high-value customers by innovatively combining multi-dimensional information such as load characteristics, response willingness and industry characteristics. SUMMARY

[0008] The technical problem to be solved by the present application is how to efficiently and accurately screen out the customer group with high response potential and the maximum signing value from massive power users, especially to solve the problems of inaccurate new customer evaluation, insufficient multi-dimensional feature fusion and lack of long-term value consideration. The specific technical scheme is as follows: the present application provides a demand response potential customer intelligent recommendation system based on multi-dimensional feature fusion, which comprises:

[0009] To solve the problems of low precision of customer screening, non-maximization of signing value and new customer cold start in the prior art, embodiments of the present application propose a demand response potential customer recommendation system and method based on multi-dimensional feature fusion. The core of the embodiments of the present application is to fuse multi-dimensional load behavior characteristics, industry attributes and environmental sensitivity characteristics, build a generalizable intelligent evaluation model, and realize the collaborative prediction of demand response potential and signing value and optimal customer recommendation.

[0010] According to an aspect of the present application, a multi-source data acquisition module is provided, comprising: the module is used for acquiring multi-source heterogeneous data including historical load data, user participation records, industry attributes and external environment information. Through the interface with the power system SCADA (Supervisory Control And Data Acquisition System), AMI (Advanced Metering Infrastructure) and EMS (Energy Management System), the system obtains real-time user 15-minute resolution load curve (96 daily power points), annual cumulative electricity consumption, maximum demand, time-of-use electricity price and peak load contribution and other dynamic and static power parameters. At the same time, through the interface of the demand response management platform, the participation data of the user's previous DR activities are obtained, such as the number of registrations, the number of actual responses, the signing capacity, the actual response amount, the response income, the response completion rate and other key indicators; from the industrial and commercial registration system or enterprise information system, the industry category, installed capacity, main equipment type and registered capital and other industry attribute information of the user are obtained. In addition, the module acquires temperature, humidity, wind speed and other environmental time series data through the weather data API, and synchronously obtains peak-valley electricity price period price and event calendar information such as holidays and weekends from the market trading platform, to build the data source of environmental sensitivity characteristics. After multi-source data acquisition, the module performs time alignment, missing value interpolation (such as linear interpolation or neighborhood mean filling) and outlier detection (using outlier smoothing method based on quartile or clustering) on the original time series data, to ensure data integrity and reliability, and provide high-quality input for subsequent feature extraction.

[0011] According to an aspect of the present application, a feature engineering module is provided, which extracts and fuses multi-dimensional features based on multi-source data to characterize the participation willingness and electricity consumption behavior of users. In one aspect, the STL (local regression trend seasonal decomposition) algorithm is applied to the historical load time series of each user for time series decomposition, which splits the load signal into long-term trend components and intra-day seasonal components, and calculates their average amplitude, standard deviation and peak-valley difference, etc. statistical indicators to fully characterize the long-term trend and short-term fluctuations of user load, thereby forming the load behavior characteristics. On the other hand, an improved participation willingness RFM model is constructed to quantify the user's inclination to participate in demand response. After Min-Max normalization of all RFM+ completion rate four-dimensional vectors, the actual demand is weighted and summed to obtain the comprehensive participation willingness score of each user. In addition, the user's industry is One-hot encoded in terms of industry attributes, and the response benchmark weight of different industries is preset. In terms of environmental sensitivity, the Spearman correlation coefficient is calculated between the time series data of temperature and electricity price and the load sequence to generate weather sensitivity and price sensitivity features. For each of the above feature sets, the entropy weight method is used to adaptively calculate the weight of each index to objectively reflect the information quantity difference of each feature. Finally, the TOPSIS method is used to evaluate the distance between the load behavior feature vector and the positive and negative ideal solutions to generate the load feature score. The participation willingness score and the load feature score are integrated, and the multi-dimensional input feature vector of the user is constructed according to the preset weight fusion, which is used for subsequent potential assessment model training and prediction.

[0012] According to an aspect of the present application, a potential assessment model is provided, which quantitatively evaluates the demand response potential of each user. First, for a single user's load sequence, the Calinski-Harabasz (CH) clustering criterion is used to automatically determine the optimal cluster number, the specific process being: input the normalized load time series of a single user and set the initial cluster number k, then use the k-means algorithm to generate clustering centers of different k values, calculate the inter-class dispersion matrix B and the intra-class dispersion matrix W, and finally determine the optimal cluster number k by

[0013]

[0014] The k value that maximizes the CH value is selected as the optimal clustering number, and a k-means clustering algorithm is used to extract typical working day and typical rest day load curve patterns of the user. The clustering mechanism can adaptively generate k typical day curve patterns according to the load difference of the user, and provide a robust benchmark load model for users in different industries and production modes. Subsequently, the extracted typical load patterns, RFM willingness to participate scores, industry attribute encodings and environmental sensitivity features and other N-dimensional features are input into an improved random forest model for training. In the improved random forest algorithm, a dynamic feature weight mechanism is introduced based on the Bagging idea: during the training process, the sampling probability of each feature is dynamically adjusted according to the performance of each sub-decision tree on the out-of-bag (OOB) samples, so that the features with higher influence in a specific industry background are preferentially selected. This mechanism enables the model to automatically focus on the key features of different user groups, improving the evaluation accuracy and model generalization ability. At the same time, the model has a built-in uncertainty quantification module: by summarizing the prediction distribution of the potential score from all sub-trees, the mean and standard deviation of the output potential score are calculated, and further a confidence interval (such as a 95% confidence interval) is given to quantify the uncertainty of the prediction result, assisting the operation personnel in evaluating the credibility of the prediction. In the cold start scenario of new users, the module also integrates a transfer learning component: a general model across industries or regions is pre-trained, and when a new user with insufficient historical response data appears, the model is fine-tuned on the initial load data of the new user, so that the response potential of the new user can be quickly evaluated without a large amount of local labeled data. Through the above methods, the potential evaluation model of the application significantly outperforms traditional single models in actual verification, improving the prediction accuracy while generating interpretable potential scores.

[0015] According to one aspect of the application, a value maximization recommendation module is provided, which includes: the module ranks and recommends users comprehensively according to the potential evaluation results and long-term value prediction. First, a subscription value prediction model based on the XGBoost algorithm is constructed to predict the potential cumulative economic benefits of users within a specified contract period (such as three years). The model takes multiple features such as potential score, historical electricity consumption scale, industry weight coefficient, flexible rate of adjustable equipment, and historical participation benefits as input, uses k-fold cross-validation and early stopping strategy to optimize model parameters, and ensures the accuracy and robustness of the prediction. After training, the system generates a standardized subscription value score for each user. Subsequently, an alpha-beta-gamma weighted recommendation score mechanism is used to comprehensively rank users: the system sets weights alpha, beta and gamma according to business needs, and calculates the final score by weighted summation of the potential score, standardized subscription value score and compliance stability (i.e. completion rate in the RFM model) of the user, i.e.

[0016]

[0017] By flexibly adjusting the weight proportions of a, b, and g, the operator can trade off between high potential, high yield, and high stability, and prefer different customer types. The system sorts all users in descending order according to the comprehensive score, and automatically generates a "priority signing list" and a "candidate list". To support decision-making, this module also generates a detailed customer portrait report, including customer potential radar chart, industry comparison chart, and risk-value two-dimensional analysis atlas, etc., which visualizes the potential and value characteristics of each customer, helping the operation team to make precise invitation and signing decisions. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 The overall architecture diagram of the demand response potential customer intelligent recommendation system based on multi-dimensional feature fusion provided by the embodiments of the present application shows the data flow and calling relationship between the multi-source data acquisition module, the feature engineering module, the potential evaluation model, and the value maximization recommendation module.

[0019] Figure 2 The feature extraction logic diagram in the embodiments of the present application shows the specific generation process of time series decomposition, RFM features, industry coding, and environmental sensitivity features.

[0020] Figure 3 The user sample scoring diagram in the embodiments of the present application shows the user sample scoring data used for random forest algorithm training.

[0021] Figure 4 The out-of-bag error of the random forest algorithm training process in the embodiments of the present application decreases with the number of trees. DETAILED DESCRIPTION

[0022] The present application provides a demand response potential customer intelligent recommendation system based on multi-dimensional feature fusion, which is characterized by fusing multi-dimensional load behavior, industry attributes, and environmental sensitivity features, constructing a generalizable intelligent evaluation model, and realizing the collaborative prediction of demand response potential and signing value and optimal customer recommendation. The following will combine the specific embodiments of the present application with the accompanying drawings to further illustrate the present application. Figure 1 to the accompanying drawings Figure 4 and specific embodiments, the specific implementation manner of the present application will be described in detail.

[0023] As Figure 1As shown, the system in this embodiment processes data from left to right through the "Multi-Source Data Acquisition Module," "Feature Engineering Module," "Potential Assessment Model," and finally the "Value Maximization Recommendation Module." In the Multi-Source Data Acquisition Module, the system acquires real-time user load data (96 sampling points daily) at a 15-minute resolution via SCADA / AMI / EMS interfaces, including annual cumulative electricity consumption, maximum demand, time-of-use pricing, and peak load contribution. Simultaneously, it obtains user participation records (such as registration count, response count, contracted capacity, actual response volume, response revenue, and completion rate) through the demand response management platform interface. Regarding external environmental information, the module calls meteorological data APIs to obtain time-series data such as temperature, humidity, and wind speed, and crawls peak-valley electricity prices from the power market trading platform, combined with information such as holidays and important event calendars. For industry attribute data, the module can obtain static characteristics such as the user's industry category, installed capacity, and main equipment types and quantities from business registration or industry statistics systems. All raw data undergoes time alignment, missing value imputation (e.g., linear interpolation or neighborhood mean), outlier detection, and cleaning after collection to form a cleaned, high-quality dataset, laying the foundation for subsequent feature construction.

[0024] like Figure 2 As shown, this example illustrates the detailed flow of data from multiple sources to fused features within the "Feature Engineering Module." In this module, the STL algorithm is first used to decompose the 96-point / day time series of each user's historical load, separating the long-term trend component and the intraday seasonal component. The mean, standard deviation, and peak-to-valley difference of both are then calculated to comprehensively characterize the user's load behavior. Simultaneously, based on user participation records from the DR platform, a four-dimensional vector of the participation intention RFM model is calculated for each user: R represents the user's average response volume in demand response, F represents the number of times the user participates in DR within a year, and M represents the effectiveness of the user's response, i.e., the ratio of response volume to contracted volume. This three-dimensional vector is then Min-Max normalized and weighted to obtain the user's comprehensive participation intention score. Regarding industry attributes, the user's industry is uniquely encoded, and different industry response benchmark weights are preset. Regarding environmental sensitivity, Spearman correlation coefficients are calculated between temperature and electricity price series and load series to obtain temperature sensitivity and price sensitivity features. Next, in this module, the entropy weight method is used to calculate objective weights for all extracted load features, obtaining a weight allocation that reflects the contribution of information entropy. The TOPSIS method is then used to calculate the distance between each user's load feature vector and the positive and negative ideal solutions, generating a load feature score. Finally, the multidimensional feature vectors (including load feature scores, participation intention scores, industry coding, and environmental sensitivity, etc.) are integrated and normalized in the feature engineering module before being passed to the potential assessment model module.

[0025] likeFigure 3 As shown, in the "potential model training" stage, the system inputs the high-dimensional features synthesized in step two into the improved random forest model. Here, first, multiple decision trees are constructed using Bagging bootstrap sampling in the training set; then the feature sampling probability of each tree is dynamically adjusted according to its error on the OOB data, so that the trees with better performance give higher weights to key features, reflecting the industry's adaptive dynamic weight mechanism. In addition, the model aggregates the prediction results of each sub-tree and outputs the mean and standard deviation of the potential score, forming a quantitative confidence interval to assist operators in further judging the result reliability. The process of the out-of-bag error decreasing with the number of trees is shown in FIG. 6. Figure 4 As shown, it can be used for model stability analysis. In the cold start scenario, the system loads the existing industry model through the transfer learning mechanism and fine-tunes it based on the new user load data, without the need for historical response records, to obtain the preliminary potential score.

[0026] After completing the potential assessment, the system enters the contract value prediction and multi-objective sorting recommendation stage. First, the feature vector generated in the previous step is input into the XGBoost model to predict the potential cumulative economic benefits of each user within a specified contract period (e.g., three years), generating a standardized contract value score. The training process automatically optimizes the model structure and parameters through cross-validation and early stopping strategies, balancing generalization ability and training efficiency. After completing the value prediction, the system uses the a-b-g weighted recommendation score mechanism for multi-objective sorting: according to the business weights a, b, and g set by the operator, the user's potential score, contract value score, and compliance stability index (completion rate) are weighted and summed to obtain a comprehensive sorting score. Finally, the system sorts all users in descending order based on the comprehensive score, automatically generating a "priority signing customer list" and a "candidate list". At the same time, a visual portrait report is generated for each customer, including a potential radar chart, an industry comparison chart, and a risk-value two-dimensional analysis chart, which details the customer's potential, value, and risk characteristics. The operation team can make precise invitation and signing decisions based on this, thereby improving overall signing efficiency and compliance stability. In summary, the potential customer recommendation system and method based on multi-dimensional feature fusion provided by the present application significantly improves data sufficiency, evaluation accuracy, and recommendation effectiveness through the introduction of multi-source data, refined feature engineering, and innovative model algorithms, achieving the coordinated optimization of demand response potential and signing value, and providing an efficient and interpretable customer screening solution for demand side management.

Claims

1. A demand response potential customer intelligent recommendation system based on multi-dimensional feature fusion, characterized in that, Comprise: Multi-source data collection module for obtaining user historical load data, response records, industry attributes and environmental data; Feature engineering module for extracting load features, response willingness features, industry features and environmental sensitivity features; Potential assessment model using an improved random forest algorithm, including dynamic weight adjustment and uncertainty quantification functions; Value maximization recommendation module for multi-objective optimization and long-term value prediction.

2. The system of claim 1, wherein, The improved random forest algorithm includes: Dynamic weight adjustment mechanism automatically adjusts the weights of different features according to the user's industry; Uncertainty quantification module outputs the confidence interval of potential assessment; Transfer learning component to solve the new customer cold start problem.

3. The system of claim 1, wherein, The response willingness feature in the feature engineering module is calculated using an improved RFM model, including: R: represents the user's flat response amount in demand response; F: represents the number of times the user participates in demand response; M: represents the effectiveness of the user's response, i.e. the ratio of response amount to contract amount.

4. A demand response potential customer intelligent recommendation method based on multi-dimensional feature fusion, characterized in that, Comprise the following steps: Collect multi-source user data; Extract multi-dimensional features and fuse them; Use the improved random forest algorithm to assess response potential; Predict long-term user value; Generate a recommendation list.

5. The method of claim 4, wherein, The improved random forest algorithm uses industry-specific feature weights during training and quantifies the uncertainty through bootstrap sampling.

Images