Service optimization method and device for high-value power customers

Abstract

The application relates to the field of information technology and provides a service optimization method and device for high-value power customers, which comprises the following steps: determining the correlation coefficient of power customer characteristics by using a correlation coefficient matrix and eliminating redundant characteristics; marking high-value power customers according to the dimension-reduced power customer characteristic set; obtaining the dynamic behavior data of the high-value power customers from the high-value customer group to obtain a dynamic characteristic vector of the behavior of the high-value power customers; outputting a behavior prediction result by a long short-term memory network model for the dynamic characteristic vector; obtaining the complaint records and interaction logs of the high-value power customers to mark complex demands; determining priority service objects by using a weighted scoring method; and generating an individualized service strategy for the priority service objects by using a recommendation algorithm. The application realizes accurate identification, behavior prediction and service optimization of the high-value power customers, effectively improves the service quality and satisfaction of the power customers, and provides technical support for fine operation and differentiated service of power enterprises.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information technology, and in particular to a service optimization method and device for high-value power customers. BACKGROUND

[0002] Power supply enterprise customer demand analysis is a core research field for optimizing service quality and improving customer satisfaction in the power industry, and its importance lies in its direct relationship with the operational efficiency and market competitiveness of the enterprise. With the advancement of power market reform, customer demand is increasingly diversified, and demand analysis has become a key link in promoting precise service and resource allocation. However, many current solutions still remain at the qualitative analysis level, relying on manual experience or simple statistics, and are difficult to fully capture the complexity and dynamic changes of customer behavior. This results in the enterprise being unable to accurately identify high-value customers or quickly respond to their core demands when faced with massive customer data, and service efficiency and customer experience are significantly constrained. The limitations of existing methods mainly lie in the lack of depth of data processing and the single dimension of analysis. Traditional methods often classify customers based on a single indicator such as electricity consumption or complaint frequency, ignoring the multi-dimensional characteristics of customer behavior and their potential relevance. This type of rough analysis is difficult to adapt to the individualization trend of customer demands and cannot provide scientific decision-making basis for the enterprise. In this field, the core challenge lies in how to effectively integrate the multi-dimensional data of customers and convert it into operational quantitative indicators. Specifically, the heterogeneity of customer characteristics, the dynamic nature of behavior, and the complexity of demands are the three major technical factors. Since customer characteristics cover electricity habits, payment records, and other aspects, the data sources are scattered and the formats are different, making integration difficult; customer behavior changes over time, and static models are difficult to reflect their true needs; the complexity of demands involves multiple factors such as emotions and scenarios, making it difficult to simply classify. These unresolved technical factors lead to unique challenges for the enterprise in identifying high-value customers and optimizing service strategies, and there is an urgent need for systematic quantitative methods. Therefore, how to multi-dimensionally value customer characteristics, behavior, and demands, and build a scientific evaluation system to accurately identify high-value customers and improve their satisfaction, has become a key problem in the field of power supply enterprise customer demand analysis. SUMMARY

[0003] In order to achieve the purpose of the present application, in a first aspect, the present application provides a service optimization method for high-value power customers, mainly comprising:

[0004] The multi-dimensional data is acquired from the power consumption records, payment data and complaint logs of power customers, and the multi-dimensional data is preprocessed to obtain a power customer dataset with unified data structure. For the power customer dataset with unified data structure, the correlation coefficients between the features of the power customers are calculated by using the corr function of the pandas library, a Pearson correlation coefficient matrix is constructed, and the principal component feature vectors of the power customers are generated based on the Pearson correlation coefficient matrix by using the principal component analysis method to obtain the power customer feature set after dimension reduction. According to the power customer feature set after dimension reduction, the eigenvalues of the power customers are calculated by calculating the Euclidean norm of the principal component feature vectors of the power customers, and it is determined that the power customer is a high-value power customer if the eigenvalue exceeds a preset threshold T1. The power consumption and complaint frequency of the high-value power customer are acquired from the power customer dataset, the time series analysis technology is used to calculate the change rates of the power consumption and complaint frequency in a continuous time window to obtain the dynamic feature vector of the high-value power customer. For the dynamic feature vector of the high-value power customer, a multi-dimensional feature matrix of the high-value power customer is constructed according to the time sequence, the multi-dimensional feature matrix is input into a pre-trained long short-term memory network model, and a behavior prediction result of the high-value power customer is output. The complaint records and interaction logs of the high-value power customer are acquired from the power customer dataset, the complaint records and interaction logs of the high-value power customer are input into a preset sentiment analysis model for sentiment analysis to obtain a sentiment analysis result of the high-value power customer, a sentiment analysis feature vector is constructed by using the sentiment analysis result and the behavior prediction result, the sentiment analysis feature vector is input into a pre-constructed decision tree model, a demand complexity classification is performed, and a demand complexity classification result is obtained. The power consumption, behavior dynamics and demand complexity classification result of the high-value power customer are comprehensively scored by using a weighted scoring method, and a service priority classification result of the high-value power customer is determined based on the comprehensive score. If the service priority classification result of the high-value power customer is a priority service object, a personalized service strategy for the high-value power customer is generated by using a recommendation algorithm, if the expected response time of the personalized service strategy is less than a preset threshold T3, it is determined that the personalized service strategy is an executable strategy, and a final service optimization scheme is obtained.

[0005] In a second aspect, the present application also provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any of the above methods when executing the computer program.

[0006] The technical scheme provided by the embodiments of the present application can have the following beneficial effects:

[0007] The application discloses a service optimization method and device for high-value power customers. The method identifies high-value customers by integrating customer power consumption, payment and complaint data, using principal component analysis and clustering algorithm. Further, time series analysis and long short-term memory network model are used to predict customer behavior changes, and long short-term memory network model is used to analyze customer demand complexity. Finally, a quantitative evaluation system is constructed based on a weighted scoring method to generate personalized service strategies for priority service objects. The application realizes accurate identification, behavior prediction and service optimization of high-value customers, effectively improves customer service quality and satisfaction, and provides technical support for fine operation and differentiated services of power enterprises. BRIEF DESCRIPTION OF DRAWINGS

[0008] Figure 1 The flowchart of the service optimization method for high-value power customers of the application.

[0009] Figure 2 The structural schematic block diagram of the computer device of the embodiment of the application. DETAILED DESCRIPTION

[0010] In order to make the purpose, technical scheme and advantages of the application clearer, the application will be described in detail below in combination with the drawings and specific embodiments.

[0011] As Figure 1 , the service optimization method for high-value power customers of the embodiment can specifically include:

[0012] S101, obtaining multi-dimensional data from power customer power consumption records, payment data and complaint logs, and preprocessing the multi-dimensional data to obtain a power customer data set with unified data structure.

[0013] Multi-dimensional data is obtained from power consumer electricity consumption records, payment data and complaint logs, structured fields and unstructured fields in the multi-dimensional data are separated using a Pandas library to obtain an initial data set. The initial data set is processed using the Pandas library, if missing values are detected, the numerical type field is filled by the mean filling method, and the category type field is filled by the mode filling method to obtain a filled data set. For the filled data set, if the value of the numerical type field is detected to be out of the preset numerical threshold range, the abnormal value out of the numerical threshold range is removed to obtain a cleaned data set. The cleaned data set is standardized using a StandardScaler, different formats of fields are converted into a preset structure using a unified coding rule to obtain a standardized data set. Features including customer identification, power consumption, payment amount and complaint frequency are extracted from the standardized data set, a multi-dimensional feature vector is generated using a groupby function of the Pandas library to obtain an aggregated data set. The aggregated data set is classified using a K-means clustering algorithm, customer groups are divided according to power consumer electricity consumption behavior and complaint behavior to obtain a classified power consumer data set. The classified power consumer data set is analyzed using an Apriori algorithm to determine the electricity consumption mode and complaint mode correlation relationship between the power consumer groups to obtain a data structure unified power consumer data set.

[0014] Exemplarily, in the power customer electricity service scenario, the processing and analysis of multi-dimensional data can effectively mine the power customer behavior patterns and improve the service quality. Exemplarily, multi-dimensional data is obtained from the power customer electricity records, payment data and complaint logs. The power customer electricity records include structured fields such as power customer identification and monthly electricity consumption, and unstructured fields such as notes; the payment data includes payment amount and payment time; the complaint logs include complaint content and time. The fields are separated using the Pandas library, the structured fields such as electricity consumption and payment amount are directly extracted, and the unstructured fields such as complaint text are converted into category labels through keyword extraction. For example, the record of power customer A includes electricity consumption of 500 kWh, payment amount of 600 yuan, and complaint content of “unstable voltage”. After separation, the initial data set is obtained, including the structured fields of electricity consumption, payment amount and the unstructured converted complaint category of “voltage problem”. In a possible implementation manner, the missing values of the initial data set are processed. If the numerical type field such as electricity consumption is missing, the mean value is used for filling. For example, the electricity consumption of power customer A is missing in a certain month, and the mean value of 400 kWh of other months of power customer A is used for filling. If the category type field such as complaint category is missing, the mode is used for filling, such as filling as the value of “no complaint” if it appears most frequently. The data set after filling is complete, which is beneficial to subsequent analysis and ensures data consistency. Specifically, when cleaning the data set, the abnormal values of the numerical type field are detected. The preset electricity consumption threshold range is 50 to 1000 kWh, and if the electricity consumption of power customer A is 2000 kWh, it is determined as an abnormal value and is removed. The data set after cleaning is more consistent with the actual business scenario, and the data reliability is improved. Preferably, the numerical type fields such as electricity consumption and payment amount are converted into a standard distribution with a mean value of 0 and a variance of 1 by using StandardScaler. The category type field such as payment status is uniformly coded into a numerical value, such as 1 for “paid” and 0 for “not paid”. The standardized data set has a uniform format, which is convenient for subsequent algorithm processing. For example, after standardization, the fields of different dimensions can be directly compared, which improves the clustering effect, unifies the dimension and format, and is convenient for algorithm processing. Features are extracted and a multi-dimensional feature vector is generated, the groupby function of Pandas is used to aggregate according to the power customer identification, and the average electricity consumption, total payment amount and complaint frequency are calculated. For example, the feature vector of power customer A is average electricity consumption of 450 kWh, total payment amount of 7200 yuan and complaint frequency of 2 times. The aggregated data set highlights the power customer behavior patterns and provides a basis for clustering. In an embodiment, the K-means clustering algorithm is used to classify the aggregated data set. K=3 is set to divide the power customer groups into three categories of high electricity consumption and high complaint, low electricity consumption and no complaint, and medium electricity consumption and occasional complaint. For example, power customer A is divided into the high electricity consumption and high complaint group, which reflects its behavior characteristics. It can be understood that the Apriori algorithm is used to analyze the power customer behavior patterns. The mining results show that the high electricity consumption power customers are often associated with “voltage problem” complaints, and the low electricity consumption power customers are mostly “no complaint”.For example, continuous features such as power consumption, payment amount, complaint frequency, etc. are divided into high / medium / low categories according to business rules, and the power customer group labels generated by K-means clustering are retained (such as high power consumption and high complaint). The group label of each power customer is combined with the discretized features to form a transaction (such as "high power consumption + high power consumption + high complaint frequency"). The Apriori algorithm is used to set the minimum support, and the high-frequency co-occurrence item set containing power customer groups, power consumption features, and complaint features is extracted. Based on the frequent item set, the minimum confidence is set to filter the strong association rules between power customer groups and power consumption / complaint patterns (such as "high power consumption power customer group → high complaint frequency"). The rules are structured according to support, confidence and other indicators to form a data structure unified power customer dataset containing power customer behavior association relationships, and the result is obtained: 80% of high power consumption power customers have voltage-related complaints, revealing the association between power consumption and complaint types. The data structure unified power customer dataset provides a basis for enterprise optimization services, such as strengthening voltage stability for high power consumption power customers. It should be noted that the above process forms a complete analysis chain through data cleaning, standardization, feature extraction, clustering and association analysis. Each step supports each other to ensure data quality and analysis accuracy, ultimately helping enterprises understand power customer behavior, optimize resource allocation, and improve service quality.

[0015] In S102, for the data structure unified power customer dataset, the correlation coefficient between the features of the power customers is calculated using the corr function of the pandas library, a Pearson correlation coefficient matrix is constructed, and a principal component analysis method is used to generate a principal component feature vector of the power customers based on the Pearson correlation coefficient matrix, to obtain a reduced power customer feature set.

[0016] The power customer features are obtained from the data structure unified power customer dataset, the correlation coefficient between the features of the power customers is calculated using the corr function of the pandas library, and a Pearson correlation coefficient matrix is constructed, wherein the elements of the Pearson correlation coefficient matrix reflect the correlation coefficients between different features of the power customers. According to a preset correlation coefficient threshold, if the absolute value of the correlation coefficient corresponding to any two features exceeds the preset correlation coefficient threshold, the two features are determined as redundant features, the feature pair composed of the two features is marked, the var function of the numpy library is used to calculate the variance, and the feature with larger variance is retained, and the redundant features are removed, to obtain a reduced feature set. For the reduced feature set, a principal component analysis method is used to extract principal component features to generate a principal component feature vector, to obtain a reduced power customer feature set.

[0017] Exemplarily, power customer features are obtained from a data structure unified power customer dataset, which contains fields such as power customer identification, average power consumption, total payment amount, complaint frequency, etc. The correlation coefficient between the features of the power customer is calculated using the corr function of the pandas library to construct a Pearson correlation coefficient matrix, which measures the linear correlation between the features. The Pearson correlation coefficient ranges from -1 to 1, and an absolute value close to 1 indicates strong correlation. For example, the dataset contains power customer A's average power consumption of 450 kWh, total payment amount of 7200 yuan, and complaint frequency of 2 times. Through the corr function calculation, the correlation coefficient between the average power consumption and the total payment amount is 0.85, indicating a high correlation between the two. In one possible implementation, the preset correlation coefficient threshold is 0.8, and if the absolute value of the correlation coefficient of any two features exceeds 0.8, it is determined that the feature is redundant. For example, the correlation coefficient 0.85 between the average power consumption and the total payment amount exceeds the correlation coefficient threshold, and is marked as a redundant feature pair. The var function of the numpy library is used to calculate the variance of the two, assuming that the variance of the average power consumption is 10000, in units of kWh square, and the variance of the total payment amount is 5000, in units of yuan square. The average power consumption with larger variance is retained, and the total payment amount is removed to obtain a reduced feature set. The reduced dataset reduces redundant information and improves subsequent analysis efficiency. Specifically, for the reduced feature set, principal component analysis is used to extract principal components to generate principal component feature vectors. Principal component analysis combines the original features into new orthogonal features through linear transformation, retaining the main information of the data. For example, the reduced feature set contains average power consumption and complaint frequency, and after applying principal component analysis, two principal component feature vectors are generated. The first principal component feature vector may mainly reflect power consumption behavior, and the second principal component feature vector reflects complaint tendency. The dimensionality of the reduced feature vector is reduced, for example, from 10 dimensions to 2 dimensions, retaining 90% of the information variance. Preferably, the reduced power customer feature set can be used for subsequent power customer behavior analysis. For example, the feature vector of power customer A after dimensionality reduction is [2.5, -0.3], reflecting its high power consumption and moderate complaint tendency. Compared with the original high-dimensional data, the reduced feature set is easier to visualize and model, improving analysis efficiency. It can be understood that the above process forms a logically rigorous feature optimization chain through correlation analysis, feature removal, and dimensionality reduction. Correlation analysis identifies redundant features, variance comparison ensures that features with large information amounts are retained, and principal component analysis further compresses the data dimension, improving data processing efficiency and analysis accuracy.

[0018] S103、According to the reduced power customer feature set, the Euclidean norm of the principal component feature vector of the power customer is calculated to obtain the feature value of the power customer, and it is judged whether the feature value exceeds a preset threshold T1. If yes, the power customer is determined to be a high-value power customer.

[0019] The principal component feature vector of the power customer is obtained from the dimension-reduced power customer feature set, and the characteristic value of the power customer is obtained by calculating the Euclidean norm of the principal component feature vector of the power customer. If the characteristic value does not exceed the preset threshold T1, the power customer is marked as an ordinary customer. If the characteristic value exceeds the preset threshold T1, the customer is marked as a high-value power customer, and a high-value customer group is obtained.

[0020] Exemplarily, the dimension-reduced power customer feature set contains the principal component feature vector of the power customer, which is used for further analysis of the value of the power customer. The Euclidean norm is calculated by squaring and square-rooting the sum of the components of the vector, obtaining a single characteristic value, quantifying the overall strength of the power customer features. Specifically, the Euclidean norm of the feature vector [2.5, -0.3] of the power customer A is calculated by The calculated feature value is approximately 2.52. This feature value can be used to distinguish between different types of electricity customers, simplifying subsequent classification. In one possible implementation, a preset threshold T1 is used to distinguish between ordinary and high-value electricity customers. The setting of T1 is based on business needs, such as determining the median or average of the feature values ​​through historical data analysis as a reference. Assuming T1 is set to 2.0, if the Euclidean norm of the electricity customer's feature value does not exceed 2.0, it is marked as an ordinary electricity customer; if it exceeds 2.0, it is marked as a high-value electricity customer. For example, electricity customer A's feature value of 2.52 exceeds T1, so it is marked as a high-value electricity customer; electricity customer B's feature vector is [1.2, 0.4], and the calculated feature value is approximately 1.26, which does not exceed T1, so it is marked as an ordinary electricity customer. This classification method clearly distinguishes electricity customer groups by quantifying feature strength, facilitating targeted management by enterprises. It should be noted that the calculation of the Euclidean norm is simple and efficient, and suitable for the comprehensive evaluation of high-dimensional feature vectors. Compared to other distance metrics, the Euclidean norm assigns equal weights to each component, making it suitable for orthogonal features generated by principal component analysis. For example, the feature vector of electricity customer C is [3.0, 1.0], with a feature value of approximately 3.16, significantly exceeding T1, indicating that its electricity consumption behavior or complaint tendency is outstanding, consistent with the characteristics of a high-value electricity customer. Preferably, the classification based on feature values ​​can be refined in conjunction with business scenarios. For example, a high-value electricity customer may be an electricity customer with high electricity consumption or high complaint frequency, requiring priority attention. In one embodiment, the identification of high-value electricity customer groups can be further combined with electricity customer profiles. For example, electricity customer D has a feature value of 3.5, marking it as a high-value electricity customer, and its feature vector [3.2, -0.8] shows high electricity consumption and low complaint tendency, suitable for recommending customized electricity plans; electricity customer E has a feature value of 2.8, and its feature vector [1.5, 2.0] reflects medium electricity consumption and high complaint tendency, requiring optimization of service experience. Understandably, this classification logic forms a tight chain from feature vectors to feature values, and then to electricity customer segmentation, ensuring that the classification results are interpretable and practical. In one embodiment, the threshold T1 can be dynamically adjusted to adapt to different business objectives. For example, initially, T1 is set to 2.0 to screen core high-value electricity customers, and later lowered to 1.8 to expand the target group. Electricity customer F has a feature value of 1.9, initially marked as an ordinary electricity customer, but may later be reclassified as a high-value electricity customer due to T1 adjustments. This flexibility enhances the adaptability of the classification method. Exemplarily, the high-value electricity customer group obtained through classification can be used for targeted marketing or resource allocation, enhancing the company's responsiveness to key electricity customers.

[0021] S104. Obtain the electricity consumption and complaint frequency of the high-value electricity customers from the electricity customer dataset, and use time series analysis technology to calculate the rate of change of electricity consumption and complaint frequency within a continuous time window to obtain the dynamic feature vector of the high-value electricity customers.

[0022] The power consumption and complaint frequency of the high-value power customers are obtained from the power customer data set, time series data of the power consumption and complaint frequency are extracted, and a sequence data set containing timestamps is generated. For the sequence data set, a sliding window method is adopted, the window size is set to 7 days, the sliding step is 1 day, the change rate of the power consumption and complaint frequency in the continuous time window is calculated, and the change rate sequence is obtained. Through the change rate sequence, the dynamic feature vector of the high-value power customer is constructed.

[0023] Exemplarily, the power consumption and complaint frequency of high-value power customers are obtained from the power customer dataset to generate a time series dataset. The time series data is indexed by timestamps, recording the power consumption and complaint frequency of power customers at specific time points. For example, the power consumption data of high-value power customer A may be 1000 kWh on January 1, 2025, and 1050 kWh on January 2, and the complaint frequency may be 0 times on January 1 and 1 time on January 2. The time series dataset is in daily units, containing the power consumption and complaint frequency of consecutive dates, forming structured sequence data. This dataset provides a basis for subsequent analysis. In one possible implementation, the time series data is processed using a sliding window method, with a window size of 7 days and a step size of 1 day. The sliding window extracts the power consumption and complaint frequency data within 7 days by sliding daily, and calculates the change rate thereof. The change rate reflects the fluctuation trend of the power consumption or complaint frequency within the time window. For example, the power consumption of power customer A from January 1 to January 7 increases from 1000 kWh to 1100 kWh, with a change rate of about 2.86% per day; the complaint frequency increases from 0 times to 2 times, with a change rate of about 0.29 times per day. The change rate sequence is in time window units, recording the change rate of power consumption and complaint frequency of each window, forming a new feature sequence. It should be noted that the generation of the change rate sequence needs to consider the stationarity of the data. The power consumption of the power customer may be affected by seasons or holidays, and the complaint frequency may fluctuate due to the quality of service. The sliding window method smooths short-term fluctuations by fixing the window size, highlighting the trend changes. For example, the power consumption change rate of power customer B from January 8 to January 14 is stable at 1.5%, and the complaint frequency change rate is 0.1 times, reflecting that the power consumption behavior is stable but the complaint tendency increases slightly. This method ensures that the change rate sequence can capture dynamic behavior patterns. Specifically, a dynamic feature vector of the high-value power customer is constructed based on the change rate sequence. The dynamic feature vector takes the change rate of each time window as an element, describing the dynamic characteristics of the power customer in power consumption and complaint behavior. For example, the feature vector of power customer A in a certain window may be [2.86, 0.29], indicating a power consumption growth rate of 2.86% and a complaint frequency growth rate of 0.29 times; the feature vector of power customer B is [1.5, 0.1]. The dynamic feature vector integrates multi-dimensional data, reflecting the time dimension characteristics of the power customer behavior, facilitating subsequent analysis of the behavior trend of the power customer. In one embodiment, the construction of the dynamic feature vector can adjust the window size in combination with business requirements. For example, the window size can be expanded to 14 days to capture longer-term trends, or shortened to 3 days to highlight short-term fluctuations. The feature vector of power customer C in a 14-day window may be [1.8, 0.05], reflecting stable growth in power consumption and few complaints. This flexibility makes the dynamic feature vector adaptable to different business scenarios. Preferably, the application of the dynamic feature vector can support predictive analysis of power customer behavior.For example, a high electricity consumption growth rate for customer A may indicate increased future demand, allowing the company to adjust its power supply plan in advance; a rising complaint rate for customer B may foreshadow service problems, requiring timely intervention. This method captures behavioral changes through dynamic features, improving the accuracy of electricity customer management. Understandably, the above process, from time series extraction to dynamic feature vector construction, forms a logically rigorous analytical chain. The time series dataset ensures the temporal continuity of the data, the sliding window method uncovers trends, and the dynamic feature vector integrates multi-dimensional information, providing companies with actionable insights into electricity customer behavior. For instance, the dynamic feature vector of customer A can be used to predict their electricity demand and optimize resource allocation; customer C's low complaint rate suggests high satisfaction, allowing for priority recommendations of value-added services. This method, through progressive analysis, enhances the understanding and application of the behavior of high-value electricity customers.

[0024] S105. For the dynamic feature vectors of high-value electricity customers, construct a multi-dimensional feature matrix of high-value electricity customers according to the time series. Input the multi-dimensional feature matrix into a pre-trained long short-term memory network model and output the behavior prediction results of high-value electricity customers. The long short-term memory network model is constructed using a deep learning framework and trained using the historical multi-dimensional feature matrix as input.

[0025] The training of the long short-term memory network model comprises: obtaining a historical dynamic feature vector of a historical high-value power customer, arranging the historical dynamic feature vector in a time sequence order to form a historical multi-dimensional feature matrix containing a time dimension, wherein each row of the historical multi-dimensional feature matrix corresponds to a feature vector of a time window, and each column corresponds to different time sequence features. The historical multi-dimensional feature matrix is divided into a training set, a validation set and a test set in time sequence, wherein the training set is used for long short-term memory network model parameter learning, the validation set is used for hyperparameter tuning and overfitting judgment, and the test set is used for final model performance evaluation. A deep learning framework is used to build a long short-term memory network model, which comprises at least one long short-term memory network model hidden layer and one long short-term memory network model fully connected output layer, wherein the long short-term memory network model hidden layer is used to capture long-term dependencies in time series, and the long short-term memory network model fully connected output layer is used to output predicted customer behavior change trends. The training set is used to iteratively train the long short-term memory network model, the mean square error is used as the loss function, the Adam adaptive optimization algorithm is selected as the optimizer, the validation set is introduced in the training process to monitor the loss value of the model on non-training data in real time, the early stopping mechanism is used to avoid overfitting, and the training is automatically terminated when the validation set loss does not decrease within a preset number of rounds. The completed long short-term memory network model is used to predict the test set data to obtain the predicted value of the user behavior change trend, and the error between the predicted value and the actual value is calculated. The calculated error value is compared with a preset error threshold, if the error value is less than or equal to the preset error threshold, it is determined that the model prediction result is valid, and the training is considered to be completed, if the error value is greater than the preset error threshold, the model optimization process is triggered, the hyperparameters of the long short-term memory network model are adjusted, and the input of the training set, the validation set and the test set is re-executed until the error value meets the preset condition, wherein the hyperparameters include the number of hidden layer neurons, the time step and the learning rate.

[0026] Exemplarily, the historical dynamic feature vectors of high-value power customers are extracted from the power customer dataset, assuming that the feature vector of power customer A contains the power consumption rate of change and the complaint frequency rate of change. Each vector corresponds to a 7-day sliding window, for example, the vector from January 1 to January 7, 2024 is [2.5, 0.2], indicating that the daily average growth of power consumption is 2.5% and the daily average increase of complaint frequency is 0.2. In chronological order, the feature vectors from January 1 to December 31, 2024 form a historical multi-dimensional feature matrix, with each row being a 7-day window vector and each column being a power consumption rate of change or a complaint frequency rate of change. The matrix size is 358 rows by 2 columns, and the time dimension is completely preserved. In one possible implementation, the historical multi-dimensional feature matrix is divided into a training set, a validation set, and a test set in chronological order. Assuming that the data from January 1 to October 31, 2024 is the training set, which accounts for about 70%, i.e. 250 rows, used to learn the power customer behavior pattern by the long short-term memory network model; November 1 to November 30 is the validation set, accounting for about 10%, i.e. 36 rows, used to optimize the hyperparameters; December 1 to December 31 is the test set, accounting for about 20%, i.e. 72 rows, used to evaluate the model performance. This division ensures the continuity of the time series and avoids data leakage. Specifically, the long short-term memory network model is constructed based on a deep learning framework. The model contains one hidden layer with 128 neurons to capture the long-term dependencies of power consumption and complaint frequency, such as seasonal power consumption peaks or complaint concentration periods; the fully connected output layer outputs the predicted rate of change for the next time window, such as [2.7, 0.3]. The hidden layer filters key information through the forget gate and the input gate, preserving long-term trends and being suitable for handling complex time-series patterns of power customer behavior. It should be noted that the training process uses the training set to iteratively optimize the model, with mean squared error as the loss function to measure the deviation between the predicted value and the actual value. The Adam optimizer is suitable for unstable fluctuations in time-series data due to its adaptive learning rate adjustment. The validation set monitors the loss value in real time, for example, after 10 rounds of training, the validation set loss decreases from 0.05 to 0.02, and if it does not decrease for 5 consecutive rounds, the early stopping mechanism terminates the training to avoid overfitting. In one embodiment, the test set is used to predict the power customer behavior trend. Assuming that the actual rate of change of power customer A from December 1 to December 7, 2024 is [2.8, 0.25], the model predicts [2.75, 0.27], with an error of [0.05, 0.02]. The preset error threshold is 0.1, and the error meets the conditions, so the model prediction is valid. If the error exceeds the threshold, for example, the prediction is [3.0, 0.4], the error is [0.2, 0.15], then adjust the hyperparameters, such as increasing the hidden layer neurons to 256 or reducing the learning rate to 0.0005, and retrain. Preferably, the hyperparameter adjustment is flexible to meet business needs. For example, the time step is set to 7, matching the sliding window size to capture short-term trends; if the business needs to predict long-term trends, it can be increased to 14.The prediction result of the power customer B shows that the electricity consumption growth rate increases from 1.5% to 2.0%, prompting the enterprise to plan power supply in advance. This method accurately captures the dynamic behavior of power customers through a multi-dimensional feature matrix and a long short-term memory network model. It can be understood that the above process forms a rigorous analysis chain from matrix construction to model prediction. The division of training set, validation set and test set ensures the generalization ability of the model, the design of hidden layer excavates the timing rule, and the error evaluation and hyperparameter optimization guarantee the prediction accuracy. It can be understood that when the model training is completed, a multi-dimensional feature matrix of high-value power customers is constructed, the multi-dimensional feature matrix is input into the long short-term memory network model, and the behavior prediction result of the high-value power customer is output, for example, the low complaint prediction result of the power customer A prompts high satisfaction, and the enterprise can preferentially recommend value-added services to improve the stickiness of the power customer.

[0027] S106, obtain the complaint record and the interaction log of the high-value power customer from the power customer data set, input the complaint record and the interaction log of the high-value power customer into a pre-set sentiment analysis model for sentiment analysis to obtain a sentiment analysis result of the high-value power customer, construct a sentiment analysis feature vector through the sentiment analysis result and the behavior prediction result, input the sentiment analysis feature vector into a pre-constructed decision tree model to perform a complexity classification of the appeal, and obtain a complexity classification result of the appeal, wherein the complexity classification result of the appeal is one of a high complexity appeal result, a medium complexity appeal result and a low complexity appeal result, the sentiment analysis model is obtained by training a BERT-base model based on historical complaint records and interaction logs, and the decision tree model is trained using the sentiment analysis feature vector.

[0028] Obtain the complaint record and the interaction log corresponding to the high-value customer in the power customer data set with unified data structure, merge the data according to the customer ID and the time stamp, and generate an original text data set. Use regular expressions to remove special symbols and garbled characters in the original text, filter irrelevant words through a stop word table, and output a standardized text data set. Input the text data in the standardized text data set into a pre-set Word2Vec model of the Gensim tool, fix the output dimension to 300 dimensions, and output a 300-dimensional text feature vector. Train a BERT-base sentiment analysis model according to historical complaint records and historical interaction logs, input the 300-dimensional text feature vector into the trained model to obtain a sentiment analysis result of the high-value power customer. Standardize the behavior prediction result and the sentiment analysis result by standard deviation standardization, combine the standardized sentiment analysis result and the behavior prediction result to construct a sentiment analysis feature vector. According to the decision tree pre-trained according to the historical sentiment analysis feature vector, input the sentiment analysis feature vector into the decision tree to obtain a complexity classification result of the appeal including a high complexity appeal result, a medium complexity appeal result and a low complexity appeal result.

[0029] Exemplarily, in the power customer data processing, obtaining high-value customer complaint records and interaction logs needs to ensure the uniformity of data structure. The complaint records contain customer ID, timestamp, complaint content, and the interaction log records customer ID, timestamp, and interaction channel such as telephone or online consultation. The data merging is sorted by customer ID and timestamp to generate an original text data set. The original data set may contain thousands of records, such as 10 complaints and 50 interaction logs of power customer A. This merging preserves the time sequence relationship, facilitating subsequent analysis of power customer behavior. In a possible implementation, regular expressions are used to clean the text data. Assuming that the complaint text of power customer A contains special symbols “#￥%” and garbled characters “&%x”, the regular expressions can match and delete these characters, and only keep Chinese, English, and numbers. The stop word table filters irrelevant words such as “de” and “is”. For example, the original text “bill #￥% has a mistake &%x” is cleaned to “bill has a mistake”. The standardized text data set corresponds to each cleaned record in a line, such as the 50 records of power customer A forming a pure text list. This processing ensures that the text is suitable for subsequent feature extraction. Specifically, the text data in the standardized text data set is input into the preset Word2Vec model of the Gensim tool, and the output dimension is fixed to 300 dimensions, outputting a 300-dimensional text feature vector. Assuming that the text data set of power customer A contains 500 words, Word2Vec maps words such as “bill” and “complaint” to 300-dimensional vectors through context. For example, a 300-dimensional vector is generated for a complaint “bill has a mistake” of power customer A. It can be understood that in the Word2Vec model, 300 dimensions are a commonly used dimension that balances calculation efficiency and semantic expression capability (too low loses information, too high increases redundancy), which is sufficient to capture complex semantic relationships. This vector representation captures semantic relationships, facilitating sentiment analysis. It should be noted that the BERT-base sentiment analysis model is trained based on historical complaints and interaction logs to identify text sentiment tendencies. The input of “bill abnormal” of power customer A into the model outputs a negative sentiment score of 0.8, reflecting dissatisfaction. The 300-dimensional vector is further input into BERT to enhance the accuracy of sentiment analysis. For example, behavior prediction results such as electricity consumption growth rate 2.5% and sentiment analysis results such as negative sentiment 0.8 need to be standardized. Standard deviation standardization converts data to a distribution with a mean of 0 and a standard deviation of 1. For example, the growth rate 2.5% is standardized to 1.2, and the sentiment 0.8 is standardized to 0.9, constructing a sentiment analysis feature vector [1.2, 0.9]. It can be understood that the sentiment analysis feature vector is input into a pre-trained decision tree, and a complaint complexity classification is output. The decision tree learns classification rules based on historical vectors, for example, power customers with high electricity growth and high negative sentiment may have high complexity complaints. The vector [1.2, 0.9] of power customer A is predicted to be a medium complexity complaint, prompting the enterprise to prioritize processing. In a possible implementation, the complaint complexity classification guides resource allocation.High complexity needs to be followed by a dedicated person, medium complexity arranges regular support, and low complexity is solved by self-service. For example, the low complexity of power customer B is quickly solved by online guidance, improving efficiency. This classification logic ensures the rationality of resource allocation and improves the satisfaction of power customers.

[0030] In S107, the use of weighted scoring method is used to comprehensively score the use of electricity, behavior dynamics and complexity of appeal classification results of high value power customers, and the service priority classification results of high value power customers are determined based on the comprehensive score. If the comprehensive score is higher than the preset comprehensive score threshold T2, it is determined as a priority service object, otherwise it is determined as a common service object.

[0031] In S107, the use of weighted scoring method is used to comprehensively score the use of electricity, behavior dynamics and complexity of appeal classification results of high value power customers, and the service priority classification results of high value power customers are determined based on the comprehensive score. If the comprehensive score is higher than the preset comprehensive score threshold T2, it is determined as a priority service object, otherwise it is determined as a common service object.

[0032] Exemplarily, in the high-value power customer data processing, the data structure unified power consumption data, behavior dynamicity record and appeal complexity classification result are obtained, and it is necessary to ensure the consistency of data fields. The power consumption data includes customer IB, timestamp, monthly power consumption, and the unit is kilowatt-hour; the behavior dynamicity record includes customer IB, timestamp, behavior type such as payment frequency or consultation times; the appeal complexity classification result includes customer IB, timestamp, complexity level such as high, medium and low. The data is merged to generate a comprehensive data set with customer IB and timestamp as the key. For example, power customer A has a power consumption of 500 kilowatt-hours on February 1, 2024, and consults the electricity fee online on the same day, and the appeal complexity is medium, and after merging, a unified record is formed. This merging method ensures the completeness of the time sequence and facilitates subsequent analysis. Specifically, the power consumption, behavior dynamicity and appeal complexity values in the comprehensive data set need to be defined with clear mapping rules. The power consumption directly takes the kilowatt-hour value, such as 500; the behavior dynamicity can be quantified by the behavior frequency, for example, power customer A consults 2 times a month, which is mapped to a dynamicity score of 2; the appeal complexity is assigned according to the level, such as high for 3, medium for 2 and low for 1, and the medium complexity of power customer A is assigned a value of 2. This quantification method unifies different types of data into numerical values, facilitating weighted calculation. In one possible implementation, a weight distribution table is pre-set according to business needs. The power consumption weight is 0.4, reflecting the core service demand; the behavior dynamicity weight is 0.3, reflecting the power customer activity; and the appeal complexity weight is 0.3, highlighting the service priority. For example, power customer A has a power consumption of 500, a behavior dynamicity of 2 and an appeal complexity of 2, and the comprehensive score is 500*0.4+2*0.3+2*0.3=201.2. This weighting method balances multi-dimensional data and comprehensively reflects the value of power customers. Preferably, the comprehensive score is compared with a pre-set threshold T2, and T2 is set to 150, reflecting the boundary of priority service. The score of power customer A 201.2 is higher than T2, and is marked as a priority service object; if the score of power customer B is 120, which is lower than T2, it is marked as a normal service object. The priority service object can be arranged for follow-up, and the normal service object is processed through the standard process. For example, power customer B is given priority response due to high power consumption and frequent consultation, and the electricity fee doubt is quickly solved to improve satisfaction. It can be understood that the service priority classification result guides resource allocation. The high power consumption and complex appeal of the priority service object indicate that it is sensitive to service and needs to be responded quickly to maintain loyalty. For example, the consultation record of power customer A shows that he pays attention to electricity fee transparency, and priority processing can reduce the risk of complaint. The low power consumption and simple appeal of the normal service object such as power customer B are suitable for self-service, optimizing resource utilization. In one embodiment, the comprehensive data set supports dynamic adjustment of weights. For example, the power consumption weight can be increased to 0.5 in the peak season to highlight the power demand; and the behavior dynamicity weight can be increased to 0.4 in the off-season to focus on power customer interaction. This flexibility adapts to business changes and ensures that the classification result meets the actual needs.For example, the high electricity consumption of power customer A in the peak season further improves its priority and obtains more efficient service. It should be noted that the quantification of behavior dynamic record needs to be refined in combination with business scenarios. For example, the record of payment delay can reduce the dynamic score, reflecting the risk; frequent consultation can increase the score, reflecting the activity. After the comprehensive quantification of the on-time payment and multiple consultations of power customer A, the dynamic score is more accurate. This multi-faceted analysis ensures that the scoring system is comprehensive and the classification result is reliable. For example, the electricity consumption of power customer A is 800 kWh, the monthly consultation is 3 times, and the complexity of the appeal is high. The score is 800 x 0.4 + 3 x 0.3 + 3 x 0.3 = 321.8, which is much higher than T2, and is marked as a priority service object. Its high electricity consumption and complex appeal prompt the need for personal follow-up and quick response to enhance the trust of power customers. This multi-dimensional evaluation method balances the weight to generate a service priority classification result that effectively guides the service strategy, improves operational efficiency and power customer experience.

[0033] In S108, if the service priority classification result of the high-value power customer is a priority service object, a recommendation algorithm is used to generate a personalized service strategy for the high-value power customer. If the expected response time of the personalized service strategy is less than a preset threshold T3, it is determined as an executable strategy, and a final service optimization scheme is obtained.

[0034] If the service priority classification result of the high-value power customer is a priority service object, the customer ID, electricity consumption, behavior dynamic record, and appeal complexity classification result data of the high-value power customer are obtained, and the SQL query is used to extract the customer demand and historical service record to generate a customer demand dataset. According to the customer demand dataset, a user-based collaborative filtering algorithm is used to generate a personalized service strategy, and a preliminary strategy set is obtained. The expected response time of each strategy in the preliminary strategy set is extracted, and a linear regression model is used to quantify the expected response time to obtain a response time dataset. If the expected response time in the response time dataset is lower than the preset threshold T3, it is marked as an executable strategy, and the SQL query is used to generate an executable strategy set. According to the executable strategy set, a quicksort algorithm is used to sort the strategies by response efficiency to obtain an optimized service strategy sequence. The highest priority strategy in the optimized service strategy sequence is extracted, and the SQL query is used to associate it with the customer ID and customer demand to generate a final service optimization scheme.

[0035] Exemplarily, when obtaining the customer ID and attribute data of the priority service object from the service priority classification result, the required information can be extracted from the database through an SQL query. Assuming that the customer relationship management system stores the customer ID, power consumption, behavioral dynamics, and appeal complexity attributes, the SQL query can be designed to filter the customer ID and related fields from the priority service object table. For example, the query statement can extract the ID C001 of power customer A, the power consumption of 600 kWh, the number of queries of 4 times, and the appeal complexity of high from the priority_customers table. This query ensures accurate data extraction and provides a basis for subsequent analysis. In a possible implementation manner, when generating the power customer demand dataset, the SQL query can combine the power customer demand table and the service record table. Assuming that the demand table records that power customer A submitted a bill dispute on April 5, 2025, and the service record table shows that the service time of the past three services is 30 minutes, 45 minutes, and 20 minutes, respectively. The query associates the two tables through the power customer ID and the timestamp to form a dataset containing the power customer ID, the demand type, and the service time. This association retains the time sequence relationship between the demand and the service, facilitating the analysis of the power customer preferences. Specifically, based on the power customer demand dataset, a personalized service strategy is generated by using a user-based collaborative filtering algorithm. Collaborative filtering analyzes the behaviors of similar power customers to recommend a strategy suitable for C001. For example, power customer C002 has similar power consumption and frequent inquiries as C001, and the one-on-one follow-up strategy of C002 can be recommended to C001. The preliminary strategy set includes one-on-one service and self-service query guidance. This algorithm uses the similarity between power customers to improve the pertinence of the strategy. Preferably, the expected response time of each strategy is extracted from the preliminary strategy set, such as 1 hour for one-on-one service and 0.5 hour for self-service query. A linear regression model is used to quantify the response time to generate a response time dataset. Linear regression predicts the response time based on historical data, such as 0.9 hours for the response time of one-on-one service of C001. The model considers factors such as service type and power customer activity to ensure accurate prediction. This quantification provides a basis for strategy screening. It should be noted that if the response time is lower than a preset threshold T3, such as 1 hour, it is marked as an executable strategy. The SQL query filters the strategies with a response time lower than 1 hour from the database, such as the one-on-one service and self-service query of C001, to generate an executable strategy set. This screening ensures that the strategy is efficient and meets the needs of the priority service object. In an embodiment, the executable strategy set is sorted by response efficiency using a quicksort algorithm. The quicksort algorithm sorts the self-service query of 0.5 hours in front of the one-on-one service of 0.9 hours with the response time as the key field to generate an optimized service strategy sequence. This sorting ensures that the most efficient strategy is executed first, improving the service response speed. It can be understood that the highest priority strategy, such as self-service query, is extracted from the optimized service strategy sequence.The SQL query associates it with the customer ID and demand, such as the electricity bill detail demand of C001, and binds it with the self-service query strategy to generate the final service optimization scheme. The scheme includes pushing a self-service query link to C001 to quickly solve the electricity bill question. This association ensures that the strategy accurately matches the power customer demand, improving service efficiency and power customer satisfaction.

[0036] Reference Figure 2 In the embodiments of the present application, a computer device is also provided, which can be a server, and the internal structure thereof can be as shown in Figure 2 The computer device includes a processor, a memory, a network interface and a database connected through a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The database of the computer device is configured to store data such as customer complaint records and interaction logs. The network interface of the computer device is configured to communicate with external terminals through network connection. The computer program is executed by the processor to implement the service optimization method for high-value power customers of any of the above embodiments.

[0037] Those skilled in the art can understand that Figure 2 the structure shown in the above embodiments is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied.

[0038] The above description is only the preferred embodiments of the present application and the explanation of the applied technical principles. Those skilled in the art should understand that the scope of the invention involved in the present application is not limited to the technical solutions formed by the specific combinations of the above technical features, and should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the concept of the present application. For example, the above features are replaced with the technical features disclosed in the present application (but not limited to) having similar functions to form technical solutions.

Claims

1. A service optimization method for high-value power customers, characterized by, The method comprises: The multi-dimensional data is obtained from the power consumption records, payment data and complaint logs of power customers, the structured fields and unstructured fields in the multi-dimensional data are separated using a Pandas library, the numerical missing values are filled by mean value, the category missing values are filled by mode value, the abnormal values exceeding a preset numerical threshold are removed, and then the standardization processing is performed, a plurality of features are extracted from the standardized data, a multi-dimensional feature vector is generated by using a groupby function of the Pandas library, an aggregated data set is obtained, a K-means clustering algorithm is used to classify the aggregated data set, the power customers are divided into groups according to the power consumption behaviors and complaint behaviors of the power customers, a classified power customer data set is obtained, an Apriori algorithm is used to analyze the classified power customer data set, the correlation between the power consumption patterns and complaint patterns of the power customer groups is determined, and a data structure unified power customer data set is obtained; for the data structure unified power customer data set, a correlation coefficient between the features of the power customers is calculated by using a corr function of the pandas library, a Pearson correlation coefficient matrix is constructed, a principal component analysis method is used to generate a principal component feature vector of the power customers based on the Pearson correlation coefficient matrix, and a reduced power customer feature set is obtained; according to the reduced power customer feature set, the Euclidean norm of the principal component feature vector of the power customers is calculated to obtain a feature value of the power customers, and it is judged whether the feature value exceeds a preset threshold T1, if yes, the power customer is determined as a high-value power customer; the power consumption and complaint frequency of the high-value power customer are obtained from the power customer data set, a time series analysis technology is used to calculate the change rates of the power consumption and complaint frequency in a continuous time window, and a dynamic feature vector of the high-value power customer is obtained; for the dynamic feature vector of the high-value power customer, a multi-dimensional feature matrix of the high-value power customer is constructed according to the time sequence, the multi-dimensional feature matrix is input into a pre-trained long short-term memory network model, and a behavior prediction result of the high-value power customer is output; the complaint records and interaction logs of the high-value power customer are obtained from the power customer data set, special symbols and random codes in the original text are removed by using a regular expression, irrelevant words are filtered through a stop word table, and a standardized text data set is output, the text data in the standardized text data set is input into a Word2Vec model of a pre-set Gensim tool, the output dimension is fixed to 300 dimensions, a 300-dimensional text feature vector is output, the 300-dimensional text feature vector is input into a BERT-base sentiment analysis model trained according to historical complaint records and historical interaction logs, a sentiment analysis result of the high-value power customer is obtained, a sentiment analysis feature vector is constructed by the sentiment analysis result and the behavior prediction result, the sentiment analysis feature vector is input into a pre-constructed decision tree model, a complexity classification of appeal is performed, and a complexity classification result of appeal is obtained; a weighted scoring method is used to comprehensively score the power consumption, behavior dynamics and complexity classification result of appeal of the high-value power customer, and a service priority classification result of the high-value power customer is determined based on the comprehensive score.If the service priority classification result of the high-value power customer is a priority service object, a recommendation algorithm is used to generate a personalized service strategy for the high-value power customer, and if the predicted response time of the personalized service strategy is less than a preset threshold T3, it is determined as an executable strategy, and a final service optimization scheme is obtained.

2. The method of claim 1, wherein, The power customer data set unified in the data structure, the correlation coefficient between the features of the power customer is calculated by using the corr function of the pandas library, the Pearson correlation coefficient matrix is constructed, the principal component feature vector of the power customer is generated based on the Pearson correlation coefficient matrix, and the power customer feature set after dimension reduction is obtained, comprising: From the data structure unified power customer data set, the correlation coefficient between the features of the power customer is calculated by using the corr function of the pandas library, the Pearson correlation coefficient matrix is constructed, wherein the elements of the Pearson correlation coefficient matrix reflect the correlation coefficient between different features of the power customer; According to the preset correlation coefficient threshold, if the absolute value of the correlation coefficient corresponding to any two features exceeds the preset correlation coefficient threshold, it is determined that the feature is redundant, the feature pair composed of the two features is marked, the variance is calculated by using the var function of the numpy library, the feature with larger variance is retained, and the redundant feature is removed, to obtain the simplified feature set; For the simplified feature set, the principal component feature is extracted by using the principal component analysis method to generate the principal component feature vector, and the power customer feature set after dimension reduction is obtained.

3. The method of claim 1, wherein, The characteristic value of the power customer is obtained by calculating the Euclidean norm of the principal component feature vector of the power customer according to the power customer feature set after dimension reduction, whether the characteristic value exceeds the preset threshold T1 is judged, if yes, the power customer is determined as a high-value power customer, comprising: The principal component feature vector of the power customer is obtained from the power customer feature set after dimension reduction, and the characteristic value of the power customer is obtained by calculating the Euclidean norm of the principal component feature vector of the power customer; If the characteristic value does not exceed the preset threshold T1, it is marked as an ordinary customer; If the characteristic value exceeds the preset threshold T1, the customer is marked as a high-value power customer, and a high-value customer group is obtained.

4. The method of claim 1, wherein, The power consumption and complaint frequency of the high-value power customer are obtained from the power customer data set, the time series analysis technology is used to calculate the change rate of the power consumption and complaint frequency in the continuous time window, and the dynamic feature vector of the high-value power customer is obtained, comprising: The power consumption and complaint frequency of the high-value power customer are obtained from the power customer data set, the time series data of the power consumption and complaint frequency are extracted, and a sequence data set containing time stamps is generated; For the sequence data set, the sliding window method is used, the window size is set to 7 days, and the sliding step is 1 day, the change rate of the power consumption and complaint frequency in the continuous time window is calculated, and the change rate sequence is obtained; The dynamic feature vector of the high-value power customer is constructed through the change rate sequence.

5. The method of claim 1, wherein, The training of the long short-term memory network model comprises: The historical dynamic feature vector of the historical high-value power customer is obtained, the historical dynamic feature vector is arranged in time sequence order to form a historical multi-dimensional feature matrix containing time dimension, wherein each row of the historical multi-dimensional feature matrix corresponds to a feature vector of a time window, and each column corresponds to different time sequence features; The long short-term memory network model is constructed by using a deep learning framework, and the model comprises at least one long short-term memory network model hidden layer and one long short-term memory network model fully connected output layer, wherein the long short-term memory network model hidden layer is used to capture long-term dependencies in time series, and the long short-term memory network model fully connected output layer is used to output a predicted customer behavior change trend; The long short-term memory network model is iteratively trained using a training set, the mean square error is used as a loss function, the Adam adaptive optimization algorithm is selected as an optimizer, a validation set is introduced in the training process to monitor the loss value of the model on non-training data in real time, an early stopping mechanism is used to avoid overfitting, and the training is automatically terminated when the validation set loss does not decrease within a preset number of rounds; The completed long short-term memory network model is used to predict the test set data, and a predicted value of the user behavior change trend is obtained, and the error between the predicted value and the actual value is calculated; The calculated error value is compared with a preset error threshold, if the error value is less than or equal to the preset error threshold, it is determined that the model prediction result is valid, and the training is considered to be completed, if the error value is greater than the preset error threshold, a model optimization process is triggered, the hyperparameters of the long short-term memory network model are adjusted, and the input of the training set, the validation set and the test set is re-executed until the error value meets the preset condition.

6. The method of claim 1, wherein, The complaint records and interaction logs of the high-value power customers are obtained from the power customer data set, the complaint records and interaction logs of the high-value power customers are input into a preset sentiment analysis model for sentiment analysis to obtain sentiment analysis results of the high-value power customers, a sentiment analysis feature vector is constructed by the sentiment analysis results and the behavior prediction results, the sentiment analysis feature vector is input into a pre-constructed decision tree model, a complexity classification of appeal is performed to obtain a complexity classification result of appeal, including: The complaint records and interaction logs of the high-value customers corresponding to the data structure unified power customer data set are obtained, the data is merged according to the customer ID and the time stamp, and the original text data set is generated; The behavior prediction results and the sentiment analysis results are standardized by standard deviation standardization, and a sentiment analysis feature vector is constructed by combining the standardized sentiment analysis results and the behavior prediction results; The decision tree is pre-trained according to the historical sentiment analysis feature vector, the sentiment analysis feature vector is input into the decision tree, and a complexity classification result of appeal including a high complexity appeal result, a medium complexity appeal result and a low complexity appeal result is obtained.

7. The method of claim 1, wherein, If the service priority classification result of the high-value power customer is a priority service object, a recommendation algorithm is used to generate a personalized service strategy for the high-value power customer, if the predicted response time of the personalized service strategy is less than a preset threshold T3, it is determined as an executable strategy, and a final service optimization scheme is obtained, including: If the service priority classification result of the high-value power customer is a priority service object, the customer ID, the power consumption, the behavior dynamic record and the complexity classification result data of the high-value power customer are obtained, and the customer demand data set is generated by extracting the customer demand and the historical service record using SQL query. According to the customer demand dataset, a personalized service strategy is generated by using a user-based collaborative filtering algorithm to obtain a preliminary strategy set; The expected response time of each strategy is extracted from the preliminary strategy set, and a linear regression model is used to quantify the expected response time to obtain a response time dataset; If the expected response time in the response time dataset is lower than a preset threshold T3, it is marked as an executable strategy, and a SQL query is used to generate an executable strategy set; According to the executable strategy set, the strategies are sorted by response efficiency using a quicksort algorithm to obtain an optimized service strategy sequence; The highest priority strategy is extracted from the optimized service strategy sequence, and a SQL query is used to associate it with the customer ID and customer demand to generate a final service optimization scheme. 8.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-7. The processor executes the computer program to implement the steps of the method of any one of claims 1 to 7.

Images