Electric power customer service index processing method and device, electronic equipment and readable storage medium

Abstract

The embodiment of the invention provides a power customer service index processing method and device, electronic equipment and a readable storage medium. The method comprises the following steps: automatically acquiring customer service data by modeling an operation path of a power business system, processing by using a multi-task learning index calculation model to obtain a multi-dimensional index, and extracting an index of a target area from the multi-dimensional index, an anomaly detection result and an anomaly trend prediction result can be obtained through an anomaly detection algorithm, and alarm pushing can be performed on the results based on index management and control requirements. In this way, delay and errors caused by manual operation are avoided, and the real-time performance and consistency of data collection are improved. Through an anomaly detection algorithm, the substandard condition of the county-level customer service indexes can be automatically identified, future trend prediction and risk early warning are carried out, and intelligent and efficient index management and control are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of electric power data processing, and in particular to an electric power customer service index processing method and device, electronic equipment and a readable storage medium. BACKGROUND

[0002] In the electric power industry, customer service index management is a core link to ensure the quality of power supply services, and high-quality management can timely discover and respond to the demand for maintenance and upgrading of electric power facilities. Under the traditional management mode, electric power enterprises need to monitor and manage customer service indexes through multiple electric power business systems such as marketing management platforms and power grid management platforms.

[0003] However, the electric power customer service index management still has the following deficiencies: the customer service index data is scattered, and needs to be manually exported and summarized by manually logging into the system, which has the problems of low efficiency, easy errors, poor real-time performance, etc. In addition, only historical data can be relied on for post-analysis, and there is a lack of automated identification and prediction means for non-compliant counties or index fluctuations, making it difficult to timely warn and intervene in abnormal situations. Therefore, the traditional manual management mode has been difficult to meet the efficient and accurate index management needs, and an intelligent solution based on process automation is urgently needed. SUMMARY

[0004] The embodiments of the present application provide an electric power customer service index processing method, device, electronic equipment and readable storage medium to solve the problem of low efficiency of electric power customer service index management.

[0005] In a first aspect, the embodiments of the present application provide an electric power customer service index processing method, comprising:

[0006] Modeling the operation path of the electric power business system, and automatically obtaining customer service data of the electric power business system based on the modeling results;

[0007] Processing the customer service data using a multi-task learning index calculation model to obtain multi-dimensional indexes;

[0008] Extracting the indexes of the target area from the multi-dimensional indexes, detecting the indexes of the target area through an anomaly detection algorithm to obtain an anomaly detection result, and processing the indexes of the target area through a time series prediction model to obtain an anomaly trend prediction result;

[0009] Alarm pushing is performed on at least one of the multi-dimensional indexes, the anomaly detection result and the anomaly trend prediction result.

[0010] In some possible implementation manners, the index calculation model includes a shared encoder layer and multiple branch decoder layers of different index types, and the processing of the customer service data using the multi-task learning index calculation model to obtain multi-dimensional indexes includes:

[0011] feature extraction is performed on the customer service data by the shared encoder to obtain a feature vector;

[0012] different dimensions are generated by a multi-type branch composed of the plurality of decoders;

[0013] The multi-type branch includes a complaint type index branch and a fault repair type index branch, the complaint type index branch adopts a binary cross-entropy loss function, and the fault repair type index branch adopts a multi-class cross-entropy loss function.

[0014] In some possible implementation manners, after the indexes of different dimensions are generated, the indexes of different dimensions are further verified, and the verification includes at least one of the following:

[0015] The indexes of different dimensions are cross-dimensionally logically verified by an index constraint rule library constructed based on a knowledge graph;

[0016] The dependency relationship between the indexes of different dimensions is verified by a graph structure algorithm;

[0017] The consistency between the current indexes and historical data is verified based on a time series of a sliding window.

[0018] In some possible implementation manners, an operation path of a power business system is modeled, and customer service data of the power business system is automatically obtained based on a modeling result, including:

[0019] The operation path of the power business system is modeled based on a finite state machine to obtain a collection path corresponding to a data collection operation;

[0020] The collection path is executed by setting a timing task to obtain the customer service data of the power business system.

[0021] In some possible implementation manners, when the timing task executes the collection path, a plurality of types of customer service data are captured from an interface of the power business system by using an interface element recognition method;

[0022] The interface element recognition method includes optical character recognition and coordinate positioning, and further includes a hierarchical analysis of dynamic elements in the interface by using an interface structure recognition model based on a convolutional neural network.

[0023] In some possible implementation manners, an anomaly detection algorithm includes a tree structure-based classifier and a density peak clustering detection mechanism, and the classifier adopts a weighted feature selection mechanism.

[0024] And / or, the time series prediction model adopts a multi-modal prediction architecture combining a long short-term memory network and a graph neural network.

[0025] In some possible implementations, an alert is pushed for at least one of the multi-dimensional indicators, the anomaly detection results, and the anomaly trend prediction results, including:

[0026] Monitor whether the anomaly detection results and / or the anomaly trend prediction results meet the preset alarm level;

[0027] If the conditions are met, then at least one of the multi-dimensional indicators, the anomaly detection results, and the anomaly trend prediction results will be pushed to the corresponding recipient according to the alarm level and the preset priority queue.

[0028] Secondly, embodiments of this application provide a power customer service indicator processing device, comprising:

[0029] The acquisition module is used to model the operation path of the power business system and automatically acquire customer service data of the power business system based on the modeling results.

[0030] The indicator processing module is used to process customer service data using a multi-task learning indicator calculation model to obtain multi-dimensional indicators.

[0031] The anomaly handling module is used to extract indicators of the target region from multi-dimensional indicators, detect the indicators of the target region through anomaly detection algorithms to obtain anomaly detection results, and process the indicators of the target region through a time series prediction model to obtain anomaly trend prediction results.

[0032] The alarm module is used to push alarms for at least one of the following: multi-dimensional indicators, anomaly detection results, and anomaly trend prediction results.

[0033] Thirdly, embodiments of this application provide an electronic device, including: a memory and a processor;

[0034] The memory stores the instructions that the computer executes;

[0035] The processor executes computer execution instructions stored in memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0036] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.

[0037] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0038] The power customer service indicator processing method, apparatus, electronic device, and readable storage medium provided in this application can automatically acquire customer service data by modeling the operation path of the power business system. Multi-dimensional indicators can then be obtained through processing using a multi-task learning indicator calculation model. Indicators for the target region can be extracted from these multi-dimensional indicators. Anomaly detection algorithms can be used to obtain anomaly detection results and anomaly trend prediction results. Based on indicator control requirements, alarms can be pushed to the multi-dimensional indicators, anomaly detection results, and anomaly trend prediction results. Path modeling enables efficient data linkage between multiple systems without manual intervention, avoiding delays and errors caused by manual operation and significantly improving the real-time performance and consistency of data collection. The anomaly detection algorithm can automatically identify non-compliance of district / county-level customer service indicators and, combined with a time series prediction model, predict future trends and provide risk warnings, thereby providing a basis for early intervention and optimization decisions, achieving intelligent and efficient indicator control. Attached Figure Description

[0039] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0040] Figure 1 A flowchart illustrating a method for processing electricity customer service indicators provided in this application embodiment;

[0041] Figure 2 A schematic diagram of the architecture of an index calculation model provided in an embodiment of this application;

[0042] Figure 3 This application provides a schematic diagram of a process for obtaining customer service data in an embodiment of the present application.

[0043] Figure 4 A schematic diagram of the structure of a power customer service indicator processing device provided in this application embodiment;

[0044] Figure 5 This is a schematic diagram of the structure of an electronic device provided in this application.

[0045] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0046] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0047] As the service level of the power industry continues to improve, the customer service system of power companies has gradually covered multiple aspects such as business acceptance, fault reporting, complaint handling, and electricity consultation. In order to ensure the quality of customer service and meet the service indicators required by regulatory agencies, power companies have generally established customer service indicator systems, such as average connection time, work order processing time, first-time resolution rate, and customer satisfaction.

[0048] However, under the current technological conditions, the management and control of power customer service indicators still faces the following shortcomings: At present, most customer service indicators rely on manual export, collation and statistics in multiple systems such as marketing management platforms and power grid operation and maintenance platforms. The process is cumbersome and time-consuming, and it is easy to miss or miscollect data, resulting in insufficient data integrity and timeliness. The existing management and control of customer service indicators mainly stays in the post-event analysis stage. There is a lack of automated identification and prediction methods for non-compliant districts and counties or indicator fluctuations, making it difficult to provide timely warnings and interventions for abnormal situations.

[0049] Based on this, this application proposes a technical concept that, through the deep integration of process automation and intelligent algorithms, constructs a full-chain automated management and control system encompassing cross-system data collection, multi-dimensional indicator calculation, anomaly identification, and trend prediction. Taking power customer service indicator management as an application scenario, this concept achieves standardized data collection, intelligent indicator calculation, real-time anomaly detection, and differentiated result delivery by modeling business operation paths, designing multi-task learning models, and introducing machine learning algorithms.

[0050] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0051] Figure 1 This is a flowchart illustrating a method for processing electricity customer service indicators provided in an embodiment of this application, as shown below. Figure 1 As shown, the method includes:

[0052] Step S101: Model the operation path of the power business system and automatically obtain customer service data of the power business system based on the modeling results.

[0053] The power business system can be a marketing management platform or a power grid management platform, or other systems related to power business customer service. Customer service data can include work orders such as customer complaints and fault repairs.

[0054] Operation path modeling can refer to abstracting the user's operation steps in the power business system (such as login, menu selection, form submission) into an executable path model, for example, describing the operation logic through a finite state machine (FSM) or decision tree.

[0055] For example, regarding the process of collecting customer service data from user operations on the power business system, operation path records can be obtained through log playback, operation trajectory collection tools, etc., and transformed into event sequences (such as "login → menu selection → form filling → submission → receipt generation"). A standardized operation path model can then be generated, supporting the simulation of user interaction with the power business system via script tools or RPA robots, automatically navigating and collecting data along the path. RPA (Robotic Process Automation) is a business process automation technology based on software robots and artificial intelligence (AI).

[0056] By abstracting business operation paths into standardized models, RPA robots or scripts can automatically execute data collection tasks according to preset logic, solving problems such as inconsistencies in manual operation paths and data omissions, and achieving efficient linkage of customer service data across systems.

[0057] Step S102: Process the customer service data using a multi-task learning indicator calculation model to obtain multi-dimensional indicators.

[0058] Among them, the multi-task learning indicator calculation model can refer to a machine learning model that adopts a shared encoder-multi-branch decoder architecture, which can be used to calculate relevant indicators of multiple dimensions (such as complaint indicators, fault repair indicators, and comprehensive service indicators) based on customer service data.

[0059] Figure 2 This is a schematic diagram of the architecture of an index calculation model provided in an embodiment of this application. Figure 2 As shown, the multi-branch decoder can include decoder branches responsible for calculating different types of indicators, such as fault repair indicator branches, complaint indicator branches, and comprehensive service indicator branches. The indicator calculation model can uniformly encode different types of customer service data (such as complaint records, fault repair records, etc.) through a shared encoder, extract feature vectors, and then use the decoders of each branch to calculate indicators, obtaining indicators in multiple dimensions such as complaint indicators, fault repair indicators, and comprehensive service indicators.

[0060] In some possible implementations, customer service data can be cleaned before being processed by a shared encoder. The data cleaning process may include: deduplicating data using a clustering-based deduplication mechanism; unifying data formats across systems through regularization expression matching and format mapping rules; and identifying and removing outliers using statistical methods and Bayesian inference models to form a raw indicator dataset that meets unified specifications.

[0061] Step S103: Extract indicators for the target region from multi-dimensional indicators, detect the indicators for the target region using an anomaly detection algorithm to obtain anomaly detection results, and process the indicators for the target region using a time series prediction model to obtain anomaly trend prediction results.

[0062] The target area can be a custom-defined geographical range. For example, the target area's metrics could be regional metrics for a specific district or county. Anomaly detection results can be used to determine whether regional customer service meets the required standards.

[0063] For example, after obtaining statistically derived multi-dimensional indicators using the indicator calculation model, core features can be extracted based on these indicators, including: average response time, work order processing rate, complaint resolution rate, and payment accuracy rate. To improve the robustness and accuracy of anomaly detection, external variables such as weather, holidays, power grid load changes, and regional user density can be added as contextual features. Finally, the indicator data is transformed into a time series matrix, forming a detection model corresponding to the three-dimensional tensor input anomaly detection algorithm of district / county × indicator × time.

[0064] For example, an anomaly detection algorithm may include a tree-based classifier and a density peak clustering detection mechanism, wherein the classifier employs a weighted feature selection mechanism. By introducing a weighted feature selection mechanism into the tree structure and combining it with the density peak clustering algorithm, the ability to identify low-frequency anomalies and boundary anomalies can be improved.

[0065] Tree-based classifier: Random Forest is used as the foundation. A weighted feature selection mechanism is introduced when splitting nodes. Indicators with high sensitivity to historical performance (such as response time) are given higher weights, while indicators with low volatility (such as billing accuracy) are given lower weights.

[0066] Density Peak Clustering (DPC) detection: Further clustering of initially identified suspicious samples (such as abnormal indicators) by calculating the local density of the samples and their relative position from the center to discover low-frequency anomalies (such as a sudden increase in the number of complaints in a short period of time) and boundary anomalies (such as processing rate fluctuations close to the threshold).

[0067] For example, the time series forecasting model employs a multimodal forecasting architecture combining Long Short-Term Memory (LSTM) networks and Graph Neural Networks (GNNs). LSTMs can capture the time dependencies of single indicators, especially short-term fluctuations and cyclical changes. GNNs can model the geographical and business relationships between different districts and counties, establishing an "indicator graph." The edge weights reflect the degree of business coupling between regions; for example, the number of repair requests from adjacent districts and counties often exhibits a linkage effect. The multimodal forecasting architecture concatenates the time series features output by the LSTM with the spatial topological features generated by the GNN in a fusion layer, dynamically allocating the contribution of different features through an attention mechanism, thereby improving the accuracy of the forecast. Finally, the time series forecasting model can output the predicted values ​​and confidence intervals of indicators for the next week or month, providing early warnings.

[0068] For example, anomaly detection models can be evaluated using F1-score and AUC metrics, while time series prediction models can be evaluated using root mean square error (RMSE) and mean absolute percentage error (MAPE). Furthermore, an adaptive learning mechanism can be used to automatically adjust weighted feature weights and model parameters periodically based on newly collected customer service indicator data, ensuring long-term effectiveness of detection and prediction performance.

[0069] For example, based on the anomaly detection results, suspicious data points can be marked as "correction samples" for the prediction model, and the prediction stability can be improved through retraining or online learning mechanisms. When the prediction curve continuously exceeds the threshold (e.g., the complaint rate is consistently higher than the average level by 30% in the coming week), an early warning can be automatically triggered, and the prediction results can be pushed to the system.

[0070] Through the above-mentioned anomaly detection and anomaly trend prediction, multi-dimensional intelligent identification of customer service indicators can be achieved: it can not only detect short-term sudden anomalies, but also capture long-term risk trends in advance, thereby providing the power customer service department with accurate and forward-looking control basis.

[0071] Step S104: Send an alert for at least one of the multi-dimensional indicators, anomaly detection results, and anomaly trend prediction results.

[0072] After completing the calculation of indicators, anomaly detection, and trend prediction, these results can be integrated and visualized. The RPA robot can automatically output a visual indicator management interface and issue reports to achieve automated and hierarchical indicator management and information push.

[0073] For data integration, the multi-dimensional indicators in step S102, the anomaly detection results and anomaly trend prediction results in step S103 can be uniformly aggregated to build an indicator data warehouse.

[0074] A three-layer data model is established in the data warehouse: Raw Data Layer (ODS): stores unprocessed multi-source indicator results; Processing Layer (DWD): performs partitioning, grading, and standardization of indicators, such as converting indicators from different districts and counties into percentages according to a unified dimension; Application Layer (DM): generates cross-regional and cross-time period comparative analysis tables according to application requirements, providing data support for visualization engine calls.

[0075] For the visual interface, a hierarchical visual layout can be adopted in the unified indicator management platform. This layout can specifically include:

[0076] Macro level: The overall customer service level index (such as national average response time and overall work order processing rate) is displayed through the overview dashboard.

[0077] Meso-level: The distribution of indicators in each region is displayed through a map heat map, and the color intensity reflects the quality of the indicators.

[0078] Micro level: Line charts and radar charts are used to display the changes in indicators and risk distribution in a single district or county.

[0079] The interface supports interactive operations, allowing users to quickly switch between time intervals, indicator dimensions, or regional ranges through filtering conditions.

[0080] For results delivery, the RPA robot automatically generates reports after the metrics visualization is generated. Reports support multiple formats including text and charts (PDF, Excel, and web-based dashboards). The robot features file encryption, directory archiving, and permission-based distribution to ensure a secure and controllable report distribution process. The RPA robot supports integration with third-party messaging platforms (such as email systems) to automatically complete scheduled push notifications.

[0081] By integrating the calculation and prediction results into a unified and visualized indicator management interface, the global status and regional differences of the indicators can be presented intuitively. At the same time, the RPA robot completes the output of results and timely notifications, reducing the workload of manual statistics and reporting, and realizing the automation and efficiency of indicator management.

[0082] In the above embodiments, customer service data can be automatically acquired by modeling the operation path of the power business system. Multi-dimensional indicators can then be obtained through processing using a multi-task learning indicator calculation model. Indicators for the target region can be extracted from these multi-dimensional indicators. Anomaly detection algorithms can then be used to obtain anomaly detection results and anomaly trend prediction results. Based on indicator control requirements, alarms can be pushed to the multi-dimensional indicators, anomaly detection results, and anomaly trend prediction results. Path modeling enables efficient data linkage between multiple systems without manual intervention, avoiding delays and errors caused by manual operation and significantly improving the real-time performance and consistency of data collection. Anomaly detection algorithms can automatically identify non-compliance of district and county-level customer service indicators and, combined with time series prediction models, predict future trends and provide risk warnings, thus providing a basis for early intervention and optimization decisions, achieving intelligent and efficient indicator control.

[0083] In one embodiment, such as Figure 3 As shown, the operation path of the power business system is modeled, and customer service data of the power business system is automatically obtained based on the modeling results, including:

[0084] Step S301: Model the operation path of the power business system based on the finite state machine to obtain the acquisition path corresponding to the data acquisition operation.

[0085] Step S302: Set the scheduled task execution path to obtain customer service data from the power business system.

[0086] When modeling and automatically collecting data, the typical customer service business scenarios of power business systems such as marketing management platforms and power grid management platforms can be broken down into processes, which may include the following steps:

[0087] S1. Business operation path identification:

[0088] We analyzed common business modules within the platform (such as new user installation applications, power outage reports, payment inquiries, work order processing, and complaint handling), and obtained the path records of customer service personnel in actual operations through log playback, operation trajectory collection tools, and human interviews. Then, we converted the collected path data into event sequences, such as: login → menu selection → form filling → submission → receipt generation.

[0089] S2. Path Modeling and Abstraction: Operation steps are modeled statefully using a Finite State Machine (FSM). Each interface / operation step is abstracted as a state node, and interface transitions and click events are modeled as state transitions. At the model level, a process pattern library is introduced to store standardized operation paths for various business scenarios, such as:

[0090] Power outage repair path = {identity verification → work order entry → fault classification → work order issuance → receipt generation};

[0091] The complaint handling process is as follows: {Complaint registration → Customer follow-up → Complaint analysis → Solution → Feedback archiving}.

[0092] For operation paths with multiple branches, decision trees can be used to describe the branching logic under different conditions, such as customer type, business nature, or urgency.

[0093] S3. Data Acquisition Task Configuration: Bind the modeled path to the RPA robot task script so that the RPA can automatically navigate to the target page and perform the data acquisition operation according to the path.

[0094] Then, the scheduling strategy for the data collection task can be set as follows:

[0095] Scheduled triggers: For example, collect the previous day's indicator data at 08:00 every day, and collect the summary data of the previous week every Monday.

[0096] Conditional triggering: When the customer service system experiences a sudden surge in work orders or when the number of complaints exceeds the threshold for three consecutive hours, temporary data collection will be triggered immediately.

[0097] Triggers can be implemented through a scheduling engine, which supports periodic scheduling defined by cron expressions and event listening mechanisms (such as exception alarms and data persistence events).

[0098] S4. Data Acquisition Task Initialization and Log Management: Before a task starts, the system generates a unique task ID and writes it to the task log table for subsequent task tracking and tracing. During the execution of each task, the execution time, execution status, data acquisition scope, and target platform are recorded to ensure that the task execution results can be statistically analyzed later.

[0099] By following the steps above, we can ensure the standardization of the operation path model, the automation and traceability of the data collection task, and lay a data foundation for subsequent indicator capture and cleaning.

[0100] In some possible implementations, when the scheduled task is executed during the data collection process, a method for identifying interface elements is used to extract various types of customer service data from the interface of the power business system.

[0101] The interface element recognition method includes optical character recognition and coordinate positioning, as well as an interface structure recognition model based on convolutional neural networks to perform hierarchical analysis of dynamic elements in the interface.

[0102] For example, the RPA robot can be woken up at set times (such as 0:00 and 12:00 every day) under the control of the dispatch center and receive task parameters, including the target platform URL (Uniform Resource Locator), login credentials, and data scraping template. If an emergency event (such as a batch power outage alarm) is detected on the marketing platform or power grid platform, the dispatch center will immediately trigger a temporary scraping task.

[0103] RPA robots can use automated login scripts to simulate users entering account passwords or calling the single sign-on (SSO) interface; when a graphic CAPTCHA is detected, a convolutional neural network CAPTCHA recognition model is called to automatically recognize the input.

[0104] Interface element recognition and interaction may include the following processes:

[0105] Static element recognition: Extract fixed text labels (such as "work order number" and "processing status") through OCR (Optical Character Recognition).

[0106] Dynamic element parsing: Employs a CNN (Convolutional Neural Network)-based interface structure recognition model to perform layered parsing of the page's DOM tree and extract dynamically loaded scrolling tables, drop-down menus, etc.

[0107] Complex interaction execution: The robot performs multiple operations based on the parsing results, such as "click the drop-down box → select the district / county → export the work order form", and records the operation path in the log.

[0108] For tabular data, the RPA robot can scroll through pages and crawl each row individually, ensuring complete acquisition of large-scale data. For chart-based metrics (such as response time distribution histograms), the Canvas Parsing algorithm is used to convert the graphical information back into raw values. All collected data is first stored in a memory cache, along with additional metadata (collection timestamp, data source platform, task ID).

[0109] After capturing data, the RPA robot can perform preprocessing operations such as format conversion on the data, and then store the preprocessed data into a specific database for later retrieval.

[0110] For example, the RPA robot can use a density clustering (DBSCAN)-based similarity detection method to calculate the Euclidean distance between the work order record and the "work order number", "customer ID" and "timestamp"; if the similarity threshold is >0.95, it is determined to be a duplicate and the latest record is retained.

[0111] For example, different date formats can also be identified through regular expressions. For instance, dates such as "2025 / 09 / 27" and "27-09-2025" can be uniformly converted into the standard format "YYYY-MM-DD" (year-month-day). Furthermore, cross-system mapping tables can be constructed, such as mapping "completed" on the marketing platform and "CLOSED" on the power grid platform to the unified label "work order closed".

[0112] For example, the 3σ principle can also be used to detect anomalies in numerical fields such as "average call duration" in customer service data. If the value exceeds the mean ± 3σ range, it is marked as an anomaly. The historical distribution is updated with posterior probabilities; if the probability of a certain indicator value is less than a threshold (e.g., 0.01), it is marked as an anomaly. Key-value pair mapping is used to compare the "completion status" and "closed-loop time" of the same work order on the marketing platform and the power grid platform. If discrepancies are found (e.g., the marketing platform shows it as completed, while the power grid platform still shows it as "processing"), the system records the discrepancies in the anomaly log for manual review.

[0113] Finally, all preprocessed indicator data is written to the indicator data warehouse in a unified field format (such as JSON format) and indexed to ensure fast retrieval and retrieval during subsequent execution step S102. For example, the primary key index is: work order number + timestamp; the secondary index is: district / county ID + indicator type.

[0114] In the above embodiments, interface element recognition technology and data cleaning algorithms are used to transform the original unstructured data into structured data, eliminating redundancy, missing items and outliers, ensuring the integrity and credibility of customer service indicator data, and laying a high-quality data foundation for subsequent statistical analysis and model calculation.

[0115] In one embodiment, the metric calculation model includes a shared encoder layer and multiple branch decoder layers with different metric types. The metric calculation model, using multi-task learning, processes customer service data to obtain multi-dimensional metrics, including:

[0116] Feature vectors are obtained by extracting features from customer service data through a shared encoder; different dimensional indicators are generated through multiple types of branches composed of multiple decoders.

[0117] The multiple branches include complaint-related indicator branches and fault repair indicator branches. The complaint-related indicator branches use a binary cross-entropy loss function, while the fault repair indicator branches use a multi-class cross-entropy loss function.

[0118] Reference Figure 2The model architecture shown can share an encoder (such as Bi-LSTM or ResNet) to extract unified features across tasks (such as work order number, processing status, timestamp, etc.), and the multi-branch decoder can compute different dimensional indicators in parallel, such as complaint rate, repair timeliness rate, etc.

[0119] For shared encoders, a unified input encoding layer can be used to extract features from indicator data from different sources (complaints, fault repair, comprehensive services, etc.). The temporal dependencies and multidimensional interaction features of the data are captured using a bidirectional recurrent neural network (Bi-LSTM) or a deep residual network (ResNet structure). A unified feature vector is output for use by downstream branch decoders.

[0120] For the branch decoder layer, the complaint-related indicator branch can be used to calculate timeliness and compliance indicators such as customer complaint acceptance, processing, and feedback. This branch can determine whether the standards are met through a classification network and classify and statistically analyze issues such as complaint processing delays and duplicate complaints. The fault repair indicator branch can process multi-dimensional information such as repair dispatch, on-site response, and work order closure. This branch outputs different labels based on work order type and processing time, such as "timely," "delayed," and "unprocessed," and calculates the overall compliance rate under the district / county and time period dimensions. The comprehensive service indicator branch can statistically analyze multi-dimensional customer evaluations and feedback to calculate the comprehensive satisfaction index.

[0121] In some possible implementations, a comprehensive indicator evaluation branch can be set at the decoder layer. This branch can aggregate the outputs of multiple task branches (such as complaint indicator branches and fault repair indicator branches) and assign differentiated weights to cross-task features based on an attention mechanism to form a comprehensive service level evaluation. For example, different indicators can be dynamically weighted based on the importance of the indicator calculation task and contextual changes. During peak periods or holidays, the weight of fault repair indicators is automatically increased to enhance the model's sensitivity to repair tasks.

[0122] It should be noted that the indicator calculation model does not operate with each branch operating independently. Instead, features interact between branches through the output of a shared layer. For example, the level of customer dissatisfaction reflected in the complaint indicator will affect the weighted calculation of overall satisfaction; and delays in fault repair may be positively correlated with an increase in the complaint rate.

[0123] In the above embodiments, the parallel computing capabilities of the multi-task learning model can significantly improve the computational efficiency and accuracy of multi-dimensional indicator results. The shared encoder-multi-branch decoder architecture, through unified feature extraction and parallel processing of task branches, avoids the inefficiency of traditional serial computation; the attention mechanism weighting of cross-task features (such as the impact of complaint indicators on overall satisfaction weight) enhances the correlation analysis between indicators. This technique enables the simultaneous generation of indicator calculation results (such as complaint rate and emergency repair timeliness rate) from multi-source power business systems, providing efficient and accurate computational support for multi-dimensional statistical analysis of power customer service indicators.

[0124] In one embodiment, after generating metrics of different dimensions, the method further includes validating the metrics of different dimensions, and the validation includes at least one of the following:

[0125] (1) Cross-dimensional logical verification of indicators of different dimensions is carried out by using an indicator constraint rule library built based on knowledge graph.

[0126] For example, a constraint rule base based on knowledge graphs can be established: a constraint library covering the entire customer service business chain can be constructed, such as: "Average response time ≤ maximum response time"; "Work order closed → complaint processing status must be completed". When the model output violates the logical constraints, the anomaly is automatically marked and the original data is backtracked for secondary verification.

[0127] (2) The dependency relationship between indicators of different dimensions is verified by graph structure algorithm.

[0128] Among them, graph structure algorithms can refer to algorithms that compare the relationships between nodes in a graph, such as dependency verification algorithms.

[0129] By comparing the dependencies between different indicators using a graph structure, such as "timely repair rate" and "average power outage duration," a consistent trend should be observed. If a contradiction occurs (e.g., a high timely repair rate but an excessively long power outage duration), an anomaly can be identified, and a consistency warning will be issued.

[0130] (3) Verify the consistency between the current indicator and historical data based on the time series of sliding windows.

[0131] For example, sliding windows with different time scales (such as 7 days and 30 days) can be set to compare current indicator results with historical data to detect any sudden anomalies. If short-term indicators are found to fluctuate sharply while long-term trends remain stable, it is determined to be a short-term anomaly rather than a structural problem. Fuzzy logic reasoning is used to dynamically adjust the judgment threshold based on historical trends and business scenarios. For example, during holidays, due to the large number of work orders, the response time threshold can be moderately relaxed; on ordinary workdays, a stricter threshold standard is applied.

[0132] The indicator calculation model using multi-task learning can not only achieve parallel calculation of multi-dimensional indicators, but also quickly discover potential deviations in the calculation results through logical verification and historical data comparison mechanisms, thereby realizing dynamic self-correction of indicator results and improving the accuracy and robustness of indicator control.

[0133] If the multi-dimensional indicator results output by the indicator calculation model pass the above verification, the indicator results can proceed to the next step of anomaly detection and trend prediction.

[0134] If the result fails the validation, the following actions can be triggered:

[0135] Write abnormal results to the log and mark the abnormality category (logical conflict / historical deviation / threshold abnormality); backtrack to the data cleaning module to check whether there are any omissions or errors in the original input; submit some questionable data for manual review to avoid misjudgment affecting the overall decision.

[0136] In the above embodiments, the credibility of multi-dimensional indicator results can be significantly improved by validating the indicators. The knowledge graph constraint rule base compares indicator dependencies through graph structure (such as the consistency between "work order closed-loop status and complaint handling status") and detects short-term anomalies and long-term trend deviations through sliding window time series comparison, avoiding errors in indicator results caused by calculation bias. This technical approach ensures that the indicator results (such as complaint rate and emergency repair timeliness rate) output by the multi-task learning model are logically consistent through logical conflict detection and historical trend verification, providing a reliable basis for the accurate management and control of power customer service indicators.

[0137] In one embodiment, an alert is pushed for at least one of the following: multi-dimensional indicators, anomaly detection results, and anomaly trend prediction results, including:

[0138] Monitor whether the anomaly detection results and / or anomaly trend prediction results meet the preset alarm levels; if so, push at least one of the multi-dimensional indicators, anomaly detection results, and anomaly trend prediction results to the corresponding recipients according to the alarm level and the preset priority queue.

[0139] The alarm level can be correlated with the degree of abnormality of the anomaly detection result or with the abnormal trend of the anomaly trend prediction result. For example, when the anomaly detection result is a severe anomaly, the alarm level is high; when the anomaly detection result is a moderate anomaly, the alarm level is medium; and when the anomaly detection result is a minor anomaly, the alarm level is low.

[0140] In this embodiment of the application, an alarm push can be made by combining event listening and priority scheduling for the indicators and the detection results of the indicators.

[0141] Event listening can refer to the immediate triggering of temporary scheduling events when the system detects abnormal indicators (such as the complaint rate continuously exceeding the threshold) or abnormal inflection points in the prediction curve.

[0142] Priority queues can refer to assigning notification tasks to different priority queues based on the severity of the exception. High-priority events are pushed immediately, while medium and low-priority events are pushed according to a timed strategy.

[0143] Specific scheduling strategies can be set according to the severity of the anomalies detected. For example, severe anomalies need to be pushed to the company's management and dispatch center immediately, medium anomalies can be pushed to the regional company manager on an hourly basis, and minor anomalies can be included in daily or weekly reports and received by the frontline customer service supervisor.

[0144] After an alert is pushed, the system can record the reading, confirmation, and feedback information of the recipient users and write it to a feedback database. By analyzing the feedback data, priority thresholds and push frequency can be dynamically adjusted, ensuring continuous improvement in indicator management through a closed-loop process of "indicator calculation—anomaly identification—prediction and early warning—visual output—feedback optimization." For example, if a certain type of anomaly is frequently ignored, its priority is lowered; if anomalies in a certain district or county repeatedly trigger major complaints, its push frequency is increased.

[0145] Figure 4 This application provides a schematic diagram of the structure of a power customer service indicator processing device, as shown below. Figure 4 As shown, the power customer service indicator processing device 400 provided in this embodiment includes:

[0146] The acquisition module 401 is used to model the operation path of the power business system and automatically acquire customer service data of the power business system based on the modeling results.

[0147] The indicator processing module 402 is used to process customer service data using a multi-task learning indicator calculation model to obtain multi-dimensional indicators.

[0148] The anomaly handling module 403 is used to extract indicators of the target area from multi-dimensional indicators, detect the indicators of the target area through anomaly detection algorithms to obtain anomaly detection results, and process the indicators of the target area through a time series prediction model to obtain anomaly trend prediction results.

[0149] Alarm module 404 is used to push alarms for at least one of the following: multi-dimensional indicators, anomaly detection results, and anomaly trend prediction results.

[0150] In some possible implementations, the indicator processing module 402 can also be used to: extract features from the customer service data through the shared encoder to obtain feature vectors; generate indicators of different dimensions through multiple types of branches composed of multiple decoders; wherein the multiple types of branches include complaint-type indicator branches and fault repair-type indicator branches, the complaint-type indicator branches adopt a binary cross-entropy loss function, and the fault repair-type indicator branches adopt a multi-class cross-entropy loss function.

[0151] In some possible implementations, the indicator processing module 402 can also be used to: perform cross-dimensional logical verification of indicators of different dimensions through an indicator constraint rule base built on a knowledge graph; verify the dependency relationship between indicators of different dimensions through a graph structure algorithm; and verify the consistency between the current indicator and historical data based on a sliding window time series.

[0152] In some possible implementations, the acquisition module 401 can also be used to: model the operation path of the power business system based on a finite state machine to obtain the acquisition path corresponding to the data acquisition operation; set a timed task to execute the acquisition path to obtain the customer service data of the power business system.

[0153] In some possible implementations, the alarm module 404 may also be used to: monitor whether the anomaly detection result and / or the anomaly trend prediction result meet a preset alarm level; if so, push at least one of the multi-dimensional indicators, the anomaly detection result and the anomaly trend prediction result to the corresponding recipient according to the alarm level and the preset priority queue.

[0154] The power customer service indicator processing device provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0155] Figure 5 This is a schematic diagram of the structure of an electronic device provided in this application. Figure 5 As shown, the electronic device 50 provided in this embodiment includes at least one processor 501 and a memory 502. Optionally, the device 50 further includes a communication component 503. The processor 501, memory 502, and communication component 503 are connected via a bus 504.

[0156] In a specific implementation, at least one processor 501 executes computer execution instructions stored in memory 502, causing at least one processor 501 to perform the above-described method.

[0157] The specific implementation process of processor 501 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0158] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0159] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0160] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0161] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0162] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0163] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0164] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0165] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0166] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0167] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0168] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0169] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0170] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A method for processing electricity customer service indicators, characterized in that, include: Model the operation path of the power business system, and automatically obtain customer service data of the power business system based on the modeling results; The customer service data is processed using a multi-task learning-based indicator calculation model to obtain multi-dimensional indicators; The indicators of the target region are extracted from the multi-dimensional indicators, the indicators of the target region are detected by the anomaly detection algorithm to obtain the anomaly detection result, and the indicators of the target region are processed by the time series prediction model to obtain the anomaly trend prediction result. An alert is pushed for at least one of the multi-dimensional indicators, the anomaly detection results, and the anomaly trend prediction results.

2. The method according to claim 1, characterized in that, The indicator calculation model includes a shared encoder layer and multiple branch decoder layers with different indicator types. The indicator calculation model, utilizing multi-task learning, processes the customer service data to obtain multi-dimensional indicators, including: The shared encoder is used to extract features from the customer service data to obtain a feature vector; By using multiple decoders to form multiple types of branches, indicators of different dimensions are generated; The multiple branches include a complaint-type indicator branch and a fault repair-type indicator branch. The complaint-type indicator branch uses a binary cross-entropy loss function, while the fault repair-type indicator branch uses a multi-class cross-entropy loss function.

3. The method according to claim 2, characterized in that, After generating metrics across different dimensions, the process also includes validating these metrics, which includes at least one of the following: By using a knowledge graph-based indicator constraint rule library, cross-dimensional logical verification is performed on indicators of different dimensions. The dependencies between metrics of different dimensions are verified using a graph structure algorithm. The consistency between current indicators and historical data is verified using time series analysis based on sliding windows.

4. The method according to any one of claims 1 to 3, characterized in that, The process of modeling the operation paths of the power business system and automatically acquiring customer service data of the power business system based on the modeling results includes: The operation path of the power business system is modeled based on the finite state machine to obtain the acquisition path corresponding to the data acquisition operation. Set a scheduled task to execute the data collection path and obtain customer service data from the power business system.

5. The method according to claim 4, characterized in that, When the scheduled task is executed on the data collection path, it uses an interface element recognition method to capture multiple types of customer service data from the interface of the power business system. The interface element recognition method includes optical character recognition and coordinate positioning, as well as an interface structure recognition model based on convolutional neural networks to perform hierarchical analysis of dynamic elements in the interface.

6. The method according to any one of claims 1 to 3, characterized in that, The anomaly detection algorithm includes a tree-based classifier and a density peak clustering detection mechanism, wherein the classifier employs a weighted feature selection mechanism. And / or, the time series prediction model adopts a multimodal prediction architecture that combines long short-term memory networks and graph neural networks.

7. The method according to any one of claims 1 to 3, characterized in that, The step of sending alerts for at least one of the multi-dimensional indicators, the anomaly detection results, and the anomaly trend prediction results includes: Monitor whether the anomaly detection results and / or the anomaly trend prediction results meet the preset alarm level; If the conditions are met, then at least one of the multi-dimensional indicators, the anomaly detection results, and the anomaly trend prediction results will be pushed to the corresponding recipient according to the alarm level and the preset priority queue.

8. A power customer service indicator processing device, characterized in that, include: The acquisition module is used to model the operation path of the power business system and automatically acquire customer service data of the power business system based on the modeling results. The indicator processing module is used to process the customer service data using a multi-task learning indicator calculation model to obtain multi-dimensional indicators. An anomaly handling module is used to extract indicators of the target region from the multi-dimensional indicators, detect the indicators of the target region through an anomaly detection algorithm to obtain anomaly detection results, and process the indicators of the target region through a time series prediction model to obtain anomaly trend prediction results. The alarm module is used to push alarms for at least one of the multi-dimensional indicators, the anomaly detection results, and the anomaly trend prediction results.

9. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1 to 7.

Images