Electricity customer power failure demand deep mining method facing customer service

Abstract

The invention relates to a customer service-oriented deep mining method for power failure demands of electricity customers. The method comprises the following steps: collecting a text transferred by customer service hotline call voice data, an online consultation chatting record and power failure repair work order details, and synchronously calling a customer historical power consumption file; performing keyword extraction and semantic analysis on the text data, and distinguishing customer surface demands and deep appeals; judging whether the customer is a high-sensitivity group in combination with a historical file; the method comprises the following steps of: disassembling factors such as time characteristics, coverage range and duration estimation of power failure through scene dimensions, and mining potential appeals and deep pain points which are not directly expressed by a user; and aiming at high-priority requirements, a customer service background resource library is linked to generate a matching scheme. According to the method, the problem of fragmentation of user appeal information is effectively solved, and no omission of core dimensions such as special electric equipment types and influence ranges is ensured; the subjectivity of manual judgment is avoided, and the requirements of high urgency and large-range influence are preferentially handled.

Description

Technical Field

[0001] This invention belongs to the field of residential electricity technology, specifically a method for in-depth mining of power outage needs of electricity customers for customer service purposes. Background Technology

[0002] User power outage demand mining is a method that utilizes big data analytics and artificial intelligence to deeply understand users' potential needs and behavioral patterns during power outages from multi-source heterogeneous data. This process goes beyond traditional passive repair responses; it proactively integrates massive amounts of information, including historical power outage records, user repair work orders, electricity consumption behavior data, geographic information, social media sentiment, and even weather forecasts. Through data mining and machine learning algorithms, it constructs user profiles and power outage impact assessment models. Its core objective is to accurately identify the differentiated demands of different user groups regarding their sensitivity to power outages, tolerance levels, and timeliness of power restoration. This provides data-driven decision support for optimizing emergency repair resource scheduling, developing personalized communication strategies, and providing value-added services such as emergency power supplies, ultimately improving power supply reliability and customer satisfaction. It is a crucial step in the transformation of modern power companies from simply providing "electricity" to providing "good electricity."

[0003] However, existing technologies lack standardized and structured extraction mechanisms when handling power outage requests from customers. User requests are easily fragmented and key dimensions are easily missed. Priority assessment relies on subjective human judgment, and high-urgency requests, such as those related to medical equipment, may be delayed. Resource scheduling requires manual cross-departmental communication, resulting in low decision-making efficiency. The follow-up process lacks targeted verification indicators, leading to many invalid inquiries and an incomplete service loop, which can easily result in unreasonable allocation of emergency resources and insufficient user satisfaction. Summary of the Invention

[0004] To address the aforementioned problems, this invention provides a method for in-depth analysis of power outage needs of electricity customers in customer service settings. The technical solution is as follows:

[0005] A method for in-depth mining of power outage needs of electricity customers for customer service purposes includes the following steps:

[0006] S1: Collect transcripts of customer service hotline calls, online chat logs, and power outage repair work order details, while simultaneously retrieving customers' historical electricity usage records. This data will serve as the core material for S2's requirement layering analysis and provide the original reference for S5's requirement verification.

[0007] S2: Extract keywords and perform semantic analysis on the text data from S1 to distinguish between surface-level and deeper customer needs; combine historical records to determine whether customers belong to highly sensitive groups. The stratification results will be directly pushed to S3 for scenario-based requirement modeling;

[0008] S3: Based on the layered requirements of S2, a three-dimensional mapping framework of "power outage scenario - user role - hidden pain point" is built: By breaking down the time characteristics, coverage, and estimated duration of power outages through the scenario dimension, and combining the core electricity dependence attributes of user roles, the system can uncover potential demands and deep pain points that users have not directly expressed. The requirement list generated by this framework will serve as the core basis for prioritizing S4;

[0009] S4: Referring to the scenario-based requirement list in S3, prioritize requirements using a three-dimensional scoring method: "urgency - scope of impact - resource availability"; for high-priority requirements, generate matching solutions by linking with the customer service backend resource library. The ranking results will guide the follow-up verification direction in S5;

[0010] S5: The customer service team conducts follow-up calls to high-priority needs sorted by S4: for example, asking elderly customers living alone "whether they need our assistance in contacting the community to provide emergency lighting", to verify whether the pain points captured by S3 are accurate; the follow-up results are synchronized to S1 to update customer files, and at the same time fed back to S3 to optimize the three-dimensional mapping framework, forming a closed loop of demand mining.

[0011] In a preferred embodiment, in step S1, after the user's voice request is transcribed into text, the system maps the natural language description into structured fields containing core dimensions such as user ID, region, request type, type of special electrical equipment, scope of impact, power outage duration, and contact information through keyword matching and semantic parsing. Subsequently, the system automatically verifies the completeness of the fields, such as confirming whether the address is within the service coverage area and whether the special equipment type belongs to a highly dependent category. If any are missing, a completion mechanism is triggered, and finally, original request entries that conform to the specifications are generated as the input basis for subsequent scenario-based processing.

[0012] In a preferred embodiment, in step S2, based on the text data such as customer service call speech transcription, online consultation records, and historical electricity consumption records collected in S1, an algorithm integrating TF-IDF keyword extraction and a pre-trained BERT semantic model is used to perform demand stratification and sensitive group identification. First, the weight values ​​of keywords in the text are calculated using TF-IDF to quickly locate the explicit surface demands expressed by customers. Then, the text is converted into a high-dimensional semantic vector using the BERT model, and cosine similarity is calculated with a pre-set deep demand semantic vector template library to uncover the deep demands not directly stated in the text. At the same time, sensitive group feature matching rules are constructed, and features such as electricity consumption type, past repair frequency, and special user identifiers in the historical records are converted into numerical vectors. These vectors are compared with the sensitive group feature library using a weighted feature matching algorithm to identify highly sensitive groups. The stratified demand results and sensitive group labels are directly pushed to S3 for scenario-based demand modeling.

[0013] In a preferred embodiment, in step S2, the deep-level appeal semantic matching formula accurately uncovers the deep-level appeals that customers have not directly expressed by combining semantic vector comparison with keyword weights, specifically as follows:

[0014] ,

[0015] In the formula:

[0016] Sim deep This indicates the semantic matching degree of the deeper appeal, with a value ranging from 0 to 1.

[0017] V text BERT semantic vectors representing customer text data;

[0018] V template A semantic vector template representing a pre-set, deep-seated demand;

[0019] ||V text || represents the magnitude of the client's text semantic vector;

[0020] ||V template || represents the magnitude of the deep-level appeal template vector;

[0021] W tfidf This represents the TF-IDF weights of keywords related to deeper appeals in the text;

[0022] Formula for matching the features of sensitive groups:

[0023] ,

[0024] In the formula:

[0025] Score sen This represents the matching score for sensitive group features, with a value ranging from 0 to 1.

[0026] F user,i This represents the numerical value of the i-th sensitive feature in the user's history file;

[0027] W sen,i This represents the weight of the i-th sensitive feature;

[0028] k represents the total number of feature items in the sensitive group feature library;

[0029] Demand stratification comprehensive score formula:

[0030] ,

[0031] In the formula:

[0032] Score layer This represents the overall score for demand stratification, with a value ranging from 0 to 1.

[0033] Sim deep Indicates the semantic matching degree of deeper demands;

[0034] W surface This indicates the weight value of keywords related to surface demand.

[0035] In a preferred embodiment, step S3 employs an algorithm combining feature engineering and semantic mapping to construct a three-dimensional framework. First, feature extraction is performed on the structured data output from S2: the power outage scenario dimension is broken down into temporal features, spatial features, and expected duration features, with each feature assigned a standardized weight value, transforming it into a scenario feature vector; the user role dimension extracts core electricity dependency attributes, including production continuity requirements, living security levels, and dependence on special equipment, transforming them into numerical dependency vectors. Subsequently, a feature cross-validation algorithm is used to perform matrix operations on the scenario feature vectors and role dependency vectors, generating a preliminary pain point association matrix. Matrix elements represent the potential pain point association strength under different scenario-role combinations.

[0036] In a preferred embodiment, in step S3, based on the pain point association matrix, a semantic similarity dictionary is constructed using a pre-trained implicit demand semantic mapping model and a corpus of electricity usage scenarios. The association strength values ​​in the matrix are then converted into corresponding implicit pain point tags. Simultaneously, the confidence level of each tag is calculated, and tags with confidence levels higher than a preset threshold are selected to form a structured demand list. This list will directly serve as the core input for priority ranking in step S4, ensuring that implicit pain points with high relevance and high confidence are processed first.

[0037] In a preferred embodiment, in step S3, the scene feature quantization formula is:

[0038] ,

[0039] In the formula:

[0040] S represents the comprehensive score of scene features, 0≤S≤1;

[0041] w T Indicates the weight of time features;

[0042] T represents the standardized value of the time characteristic, indicating the degree of impact of the power outage occurrence time;

[0043] w P Indicates spatial feature weights;

[0044] P represents the standardized value of spatial characteristics, indicating the impact range of the power outage area;

[0045] w D Indicates the weights of persistent features;

[0046] D represents the standardized value of the persistence feature, indicating the degree of impact of the expected duration of the power outage;

[0047] User role dependency formula

[0048] ,

[0049] In the formula:

[0050] R represents the overall score of user role dependency, 0≤R≤1;

[0051] w M Indicates production dependency weight;

[0052] M represents the standardized value of production dependence, which indicates the degree to which production and operation depend on electricity consumption;

[0053] w L Indicates the weight of basic living security;

[0054] L represents the standardized value for basic living needs, indicating the basic requirements for electricity supply in daily life.

[0055] w Sd Indicates the special equipment dependency weight;

[0056] S d This indicates the standardization value that special equipment depends on, representing the degree of dependence on special electrical equipment;

[0057] Formula for the correlation of hidden pain points:

[0058] ,

[0059] In the formula:

[0060] C is the pain point relevance of the scene-role combination, 0≤C≤1;

[0061] S is a scene feature vector with 1×n dimensions;

[0062] The transpose of the RT user role dependency vector, with an n×1 dimension;

[0063] The confidence formula for hidden pain points is:

[0064] ,

[0065] In the formula:

[0066] Confidence represents the confidence level of hidden pain points, 0≤Confidence≤1;

[0067] α represents the relevance weight;

[0068] β represents the semantic similarity weight;

[0069] Sim represents the semantic similarity value, which is the matching result from a pre-trained corpus.

[0070] In a preferred embodiment, in step S4, the system first reads the scenario-based demand list output in S3 and assigns a three-dimensional value to each demand: the urgency dimension is assigned a value based on the power dependence level system, with medical equipment dependence demand assigned 10 points, general lighting demand assigned 5 points, and non-essential power demand assigned 3 points; the impact range dimension is assigned a value based on the number of users covered or the size of the area, with the entire building assigned 8 points and a single household assigned 3 points; the resource accessibility dimension is assigned a value linked to real-time data from the emergency resource database, with an emergency generator vehicle being available and reachable within 3 kilometers assigned 9 points, and temporary lighting equipment having sufficient inventory and a delivery distance ≤ 1 kilometer assigned 9 points. Then, a weighted summation model is used to calculate the total score, with the weights allocated as follows: urgency 60%, impact range 30%, and resource accessibility 10%. Taking the demand for a complete power outage in Building 3 of XX Community involving ventilator dependence as an example, its total score is 10 × 0.6 + 8 × 0.3 + 9 × 0.1 = 9.3 points.

[0071] In a preferred embodiment, in step S4, the system sorts all requests in descending order of total score, defining the top 20% as high-priority requests. For such requests, the system triggers a resource database linkage mechanism, retrieves the real-time dispatch table of emergency power generation vehicles, the temporary equipment inventory ledger, and the delivery route planning module, generates a matching list containing resource dispatch paths, estimated arrival times, and backup plans, and pushes it to the customer service specialist's workbench in the corresponding region. At the same time, the system marks the core indicators for the follow-up verification of the request, such as the restoration status of power supply to medical equipment and the timeliness of resource arrival, and synchronizes them to the subsequent follow-up task system.

[0072] In a preferred embodiment, in step S5, the system sorts all requests in descending order of total score, defining the top 20% as high-priority requests. For such requests, the system triggers a resource database linkage mechanism, retrieves the real-time dispatch table of emergency power generation vehicles, the temporary equipment inventory ledger, and the delivery route planning module, generates a matching list containing resource dispatch paths, estimated arrival times, and backup plans, and pushes it to the customer service specialist's workbench in the corresponding region. At the same time, the system marks the core indicators for follow-up verification of the request, such as the status of power restoration for medical equipment and the timeliness of resource arrival, and synchronizes them to the subsequent follow-up task system.

[0073] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0074] (i) In this invention, a standardized demand structure extraction mechanism effectively solves the problem of fragmented user demand information, ensuring that core dimensions such as the type of special electrical equipment and the scope of impact are not overlooked. The three-dimensional priority evaluation model is based on objective data, avoiding the subjectivity of human judgment, and prioritizing the handling of demands with high urgency and wide-ranging impact. In particular, demands related to medical equipment can quickly enter the resource matching stage, significantly improving the accuracy of emergency response. At the same time, the automated resource linkage mechanism directly connects to the backend resource database to generate the optimal scheduling plan, reducing the time cost of manual communication and decision-making, greatly improving the efficiency of demand processing, and ensuring that emergency resources can quickly reach the user groups that need them most.

[0075] (II) In this invention, the targeted follow-up verification design allows customer service personnel to focus on core issues during follow-ups, such as the status of power restoration for medical equipment and the timeliness of resource delivery. This avoids ineffective inquiries and enhances user awareness and satisfaction. Furthermore, the entire process forms a complete service loop. From demand extraction to assessment, matching, and follow-up verification, the output of each step guides the execution of subsequent steps, ensuring consistency and traceability of service quality. Simultaneously, it promotes the rational allocation of emergency resources, avoids resource waste, and provides reliable support for the continuous optimization of the power emergency service system. Attached Figure Description

[0076] Figure 1 This is a flowchart illustrating a method for in-depth mining of power outage needs of electricity customers for customer service purposes, according to the present invention. Detailed Implementation

[0077] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0078] Reference Figure 1 A method for in-depth mining of power outage needs of electricity customers for customer service purposes, the method includes the following steps:

[0079] S1: Collect transcripts of customer service hotline calls, online chat logs, and details of power outage repair orders, while simultaneously retrieving customers' historical electricity usage records. This data will serve as the core material for S2's requirement layering analysis and will also provide the original reference for S5's requirement verification.

[0080] S2: Extract keywords and perform semantic analysis on the text data from S1 to distinguish between surface-level and deeper customer needs; combine historical records to determine whether customers belong to highly sensitive groups. The stratification results will be directly pushed to S3 for scenario-based requirement modeling.

[0081] S3: Based on the layered requirements of S2, a three-dimensional mapping framework of "power outage scenario - user role - hidden pain point" is built: By breaking down the time characteristics, coverage, and estimated duration of power outages at the scenario level, and combining the core electricity dependence attributes of user roles, the system can uncover potential demands and deep pain points that users have not directly expressed. The requirement list generated by this framework will serve as the core basis for prioritizing S4.

[0082] S4: Referring to the scenario-based requirement list in S3, prioritize requirements using a three-dimensional scoring method: "urgency - scope of impact - resource availability." For high-priority requirements, generate matching solutions by linking with the customer service backend resource library. The ranking results will guide the follow-up verification direction in S5.

[0083] S5: The customer service team conducts follow-up calls to high-priority needs sorted by S4: for example, asking elderly customers living alone "whether they need our assistance in contacting the community to provide emergency lighting", to verify whether the pain points captured by S3 are accurate; the follow-up results are synchronized to S1 to update customer files, and at the same time fed back to S3 to optimize the three-dimensional mapping framework, forming a closed loop of demand mining.

[0084] In step S1, after the user's voice request is transcribed, the system uses keyword matching and semantic parsing to map the natural language description into structured fields containing core dimensions such as user ID, region, request type, type of special electrical equipment, scope of impact, power outage duration, and contact information. Subsequently, the system automatically verifies the completeness of the fields, such as confirming whether the address is within the service coverage area and whether the special equipment type belongs to a highly dependent category. If any is missing, a completion mechanism is triggered, ultimately generating original request entries that conform to the specifications, which serve as the input basis for subsequent scenario-based processing.

[0085] In step S2, based on the text data such as customer service call transcripts and online consultation records collected in S1, as well as historical electricity consumption records, an algorithm that integrates TF-IDF keyword extraction and a pre-trained BERT semantic model is used to perform demand stratification and sensitive group identification. First, the weight values ​​of keywords in the text are calculated using TF-IDF to quickly locate the explicit surface demands expressed by customers. Then, the BERT model is used to convert the text into a high-dimensional semantic vector, and cosine similarity is calculated with a pre-set deep demand semantic vector template library to uncover the deep demands not directly stated in the text. At the same time, sensitive group feature matching rules are constructed. Features such as electricity consumption type, past repair frequency, and special user identifiers in the historical records are converted into numerical vectors. A weighted feature matching algorithm is used to compare these vectors with the sensitive group feature library to identify highly sensitive groups. The stratified demand results and sensitive group labels are directly pushed to S3 for scenario-based demand modeling.

[0086] In step S2, the formula for semantic matching of deep-level requests combines semantic vector comparison with keyword weights to accurately uncover the deeper-level requests that customers have not directly stated. Specifically:

[0087] ,

[0088] In the formula:

[0089] Sim deep Indicates the semantic matching degree of the deep appeal (value ranges from 0 to 1).

[0090] V text BERT semantic vectors representing customer text data;

[0091] V template A semantic vector template representing a pre-set, deep-seated demand;

[0092] ||V text || represents the magnitude of the client's text semantic vector;

[0093] ||V template || represents the magnitude of the deep-level appeal template vector;

[0094] W tfidf This represents the TF-IDF weights of keywords related to deeper appeals in the text;

[0095] Sensitive Group Feature Matching Score Formula

[0096] ,

[0097] In the formula:

[0098] Score sen This represents the matching score for sensitive group features, with a value ranging from 0 to 1.

[0099] F user,i This represents the numerical value of the i-th sensitive feature in the user's history file;

[0100] W sen,i This represents the weight of the i-th sensitive feature;

[0101] k represents the total number of feature items in the sensitive group feature library;

[0102] Demand Stratification Comprehensive Score Formula

[0103] ,

[0104] In the formula:

[0105] Score layer This represents the overall score for demand stratification, with a value ranging from 0 to 1.

[0106] Sim deep Indicates the semantic matching degree of deeper demands;

[0107] W surface This indicates the weight value of keywords related to surface demand;

[0108] In step S3, a three-dimensional framework is constructed using an algorithm combining feature engineering and semantic mapping. First, features are extracted from the structured data output in S2: the power outage scenario dimension is broken down into temporal features, spatial features, and expected duration features, with each feature assigned a standardized weight value, transforming it into a scenario feature vector; the user role dimension extracts core electricity dependency attributes, including production continuity requirements, living security levels, and dependence on special equipment, transforming them into numerical dependency vectors. Then, a feature cross-validation algorithm is used to perform matrix operations on the scenario feature vectors and role dependency vectors, generating a preliminary pain point association matrix. Matrix elements represent the potential pain point association strength under different scenario-role combinations.

[0109] In step S3, based on the pain point association matrix, a semantic similarity dictionary is constructed using a pre-trained implicit demand semantic mapping model and a corpus of electricity usage scenarios. The association strength values ​​in the matrix are then converted into corresponding implicit pain point labels. Simultaneously, the confidence level of each label is calculated, and labels with confidence levels higher than a preset threshold are selected to form a structured demand list. This list will directly serve as the core input for priority ranking in step S4, ensuring that implicit pain points with high relevance and high confidence are processed first.

[0110] In step S3, the scene feature quantization formula

[0111] ,

[0112] In the formula:

[0113] S represents the comprehensive score of scene features, 0≤S≤1;

[0114] w T Indicates the weight of time features;

[0115] T represents the standardized value of the time characteristic, indicating the degree of impact of the power outage occurrence time;

[0116] w P Indicates spatial feature weights;

[0117] P represents the standardized value of spatial characteristics, indicating the impact range of the power outage area;

[0118] w D Indicates the weights of persistent features;

[0119] D represents the standardized value of the persistence feature, indicating the degree of impact of the expected duration of the power outage;

[0120] User role dependency formula:

[0121] ,

[0122] In the formula:

[0123] R represents the overall score of user role dependency, 0≤R≤1;

[0124] w M Indicates production dependency weight;

[0125] M represents the standardized value of production dependence, which indicates the degree to which production and operation depend on electricity consumption;

[0126] w L Indicates the weight of basic living security;

[0127] L represents the standardized value for basic living needs, indicating the basic requirements for electricity supply in daily life.

[0128] w Sd Indicates the special equipment dependency weight;

[0129] S d This indicates the standardization value that special equipment depends on, representing the degree of dependence on special electrical equipment;

[0130] Formula for the correlation of hidden pain points:

[0131] ,

[0132] In the formula:

[0133] C is the pain point relevance of the scene-role combination, 0≤C≤1;

[0134] S is a scene feature vector with 1×n dimensions;

[0135] The transpose of the RT user role dependency vector, with an n×1 dimension;

[0136] The confidence formula for hidden pain points is:

[0137] ,

[0138] In the formula:

[0139] Confidence: Confidence level of hidden pain points, 0≤Confidence≤1;

[0140] α represents the relevance weight;

[0141] β represents the semantic similarity weight;

[0142] Sim represents the semantic similarity value, which comes from the matching results of a pre-trained corpus;

[0143] In step S4, the system first reads the scenario-based requirement list output in S3 and assigns values ​​to each requirement across three dimensions: The urgency dimension is assigned based on the power dependence level system, with medical equipment dependence requirements assigned 10 points, general lighting requirements assigned 5 points, and non-essential power requirements assigned 3 points; the impact range dimension is assigned based on the number of users covered or the size of the area, with an entire building assigned 8 points and a single household assigned 3 points; the resource accessibility dimension is assigned in conjunction with real-time data from the emergency resource database, with an emergency generator vehicle available and reachable within 3 kilometers assigned 9 points, and temporary lighting equipment having sufficient inventory and a delivery distance ≤ 1 kilometer assigned 9 points. Then, a weighted summation model is used to calculate the total score, with weights allocated as follows: urgency 60%, impact range 30%, and resource accessibility 10%. Taking the requirement of a complete power outage in Building 3 of XX Community involving ventilator dependence as an example, its total score is: 10 × 0.6 + 8 × 0.3 + 9 × 0.1 = 9.3 points.

[0144] In step S4, the system sorts all requests in descending order of total score, defining the top 20% as high-priority requests. For these requests, the system triggers a resource database linkage mechanism, retrieves the real-time dispatch table of emergency power generation vehicles, the temporary equipment inventory ledger, and the delivery route planning module, and generates a matching list containing resource dispatch paths, estimated arrival times, and backup plans, which is then pushed to the customer service specialist's workbench in the corresponding region. At the same time, the system marks the core indicators for follow-up verification of this request, such as the status of power restoration for medical equipment and the timeliness of resource arrival, and synchronizes them to the subsequent follow-up task system.

[0145] In step S5, the system sorts all requests in descending order of total score, defining the top 20% as high-priority requests. For these requests, the system triggers a resource database linkage mechanism, retrieves the real-time dispatch table of emergency power generation vehicles, the temporary equipment inventory ledger, and the delivery route planning module, and generates a matching list containing resource dispatch paths, estimated arrival times, and backup plans, which is then pushed to the customer service specialist's workbench in the corresponding region. At the same time, the system marks the core indicators for follow-up verification of this request, such as the status of power restoration for medical equipment and the timeliness of resource arrival, and synchronizes them to the subsequent follow-up task system.

[0146] From the above, we can conclude that:

[0147] This invention effectively solves the problem of fragmented user demand information through a standardized demand structure extraction mechanism, ensuring that core dimensions such as the type of special electrical equipment and the scope of impact are not overlooked. The three-dimensional priority evaluation model, based on objective data, avoids the subjectivity of human judgment, allowing high-urgency and wide-ranging demands to be prioritized, especially those related to medical equipment, which can quickly enter the resource matching stage, significantly improving the accuracy of emergency response. Simultaneously, the automated resource linkage mechanism directly connects to the backend resource database to generate optimal scheduling plans, reducing the time cost of manual communication and decision-making, greatly improving demand processing efficiency, and ensuring that emergency resources can quickly reach the users who need them most.

[0148] In this invention, a targeted follow-up verification design allows customer service personnel to focus on core issues during follow-ups, such as the status of power restoration for medical equipment and the timeliness of resource delivery. This avoids ineffective inquiries and enhances user awareness and satisfaction. Furthermore, the entire process forms a complete service loop. From demand extraction to assessment, matching, and follow-up verification, the output of each step guides the execution of subsequent steps, ensuring consistent and traceable service quality. Simultaneously, it promotes the rational allocation of emergency resources, avoids resource waste, and provides reliable support for the continuous optimization of the power emergency service system.

[0149] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0150] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for in-depth mining of power outage needs of electricity customers for customer service operations, characterized in that, Includes the following steps: S1: Collect the transcripts of customer service hotline calls, online chat logs, and details of power outage repair work orders, and simultaneously retrieve customers' historical electricity usage records; S2: Extract keywords and perform semantic analysis on the text data of S1 to distinguish between the customer's surface needs and deeper demands; combine historical records to determine whether the customer is a highly sensitive group; S3: Based on the layered requirements of S2, build a three-dimensional mapping framework of power outage scenario - user role - hidden pain point: By breaking down the time characteristics, coverage and duration of power outage occurrence through the scenario dimension, and combining the core electricity dependence attributes of user roles, we can explore the potential demands and deep pain points that users have not directly expressed. S4: Referring to the scenario-based requirement list in S3, the requirements are sorted using a three-dimensional scoring method of "urgency - scope of impact - resource availability"; for high-priority requirements, a matching solution is generated by linking the customer service backend resource library. S5: The customer service team conducts follow-up calls to the high-priority requests sorted by S4 to verify the accuracy of the pain points captured by S3; the follow-up results are synchronized to S1 to update customer profiles, and at the same time fed back to S3 to optimize the three-dimensional mapping framework, forming a closed loop of demand mining.

2. The method for in-depth mining of power outage needs of electricity customers for customer service purposes as described in claim 1, characterized in that, In step S1, after the user's voice request is transcribed, the system maps the natural language description into structured fields containing user ID, region, request type, special electrical equipment type, scope of impact, power outage duration, and contact information through keyword matching and semantic parsing. The system automatically verifies the completeness of the fields, and if any are missing, it triggers a completion mechanism to finally generate original request entries that conform to the specifications, which serve as the input basis for subsequent scenario-based processing.

3. The method for in-depth mining of power outage needs of electricity customers for customer service purposes as described in claim 1, characterized in that, In step S2, based on the customer service call transcripts, online consultation records, and historical electricity consumption data collected in S1, an algorithm that integrates TF-IDF keyword extraction and a pre-trained BERT semantic model is used to carry out demand stratification and sensitive group identification. The specific process includes: By calculating the weight values ​​of keywords in the text using TF-IDF, we can quickly locate the explicit needs of customers. We use the BERT model to transform the text into a high-dimensional semantic vector and perform cosine similarity calculation with a pre-set deep-needs semantic vector template library to uncover the deep-needs not directly stated in the text. We construct sensitive group feature matching rules, transforming electricity usage type, past repair frequency, and special user identification features in historical files into numerical vectors. We then compare these vectors with a sensitive group feature library using a weighted feature matching algorithm to identify highly sensitive groups.

4. The method for in-depth mining of power outage needs of electricity customers for customer service purposes as described in claim 1, characterized in that, In step S2, the deep-level appeal semantic matching formula combines semantic vector comparison with keyword weights to uncover deeper appeals that customers have not directly stated, specifically: ， In the formula: Sim deep Indicates the semantic matching degree of deeper demands; V text BERT semantic vectors representing customer text data; V template A semantic vector template representing a pre-set, deep-seated demand; ||V text || represents the magnitude of the client's text semantic vector; ||V template || represents the magnitude of the deep-level appeal template vector; W tfidf This represents the TF-IDF weights of keywords related to deeper appeals in the text; Formula for matching the features of sensitive groups: ， In the formula: Score sen Indicates the score for matching the characteristics of sensitive groups; F user,i This represents the numerical value of the i-th sensitive feature in the user's history file; W sen,i This represents the weight of the i-th sensitive feature; k represents the total number of feature items in the sensitive group feature library; Demand stratification comprehensive score formula: ， In the formula: Score layer This indicates the overall score for demand stratification; Sim deep Indicates the semantic matching degree of deeper demands; W surface This indicates the weight value of keywords related to surface demand.

5. The method for in-depth mining of power outage needs of electricity customers for customer service purposes as described in claim 1, characterized in that, In step S3, a three-dimensional framework is built using an algorithm that combines feature engineering and semantic mapping. First, feature extraction is performed on the structured data output from S2: the power outage scenario dimension is decomposed into time features, spatial features, and expected duration features. Each feature is assigned a standardized weight value and transformed into a scenario feature vector. Core electricity dependency attributes are extracted from the user role dimension, including production continuity requirements, living security level, and dependence on special equipment, and converted into numerical dependency vectors. Then, through a feature cross-validation algorithm, the scenario feature vectors and role dependency vectors are used to perform matrix operations to generate a preliminary pain point association matrix. The matrix elements represent the potential pain point association strength under different scenario-role combinations.

6. The method for in-depth mining of power outage needs of electricity customers for customer service purposes as described in claim 1, characterized in that, In step S3, based on the pain point association matrix, a semantic similarity dictionary is constructed using a pre-trained implicit demand semantic mapping model and a corpus of electricity usage scenarios. The association strength values ​​in the matrix are converted into corresponding implicit pain point labels. At the same time, the confidence level of each label is calculated, and labels with confidence levels higher than a preset threshold are selected to form a structured list of needs. This list will be directly used as the core input for priority ranking in S4 to ensure that implicit pain points with high relevance and high confidence are processed first.

7. The method for in-depth mining of power outage needs of electricity customers for customer service purposes as described in claim 1, characterized in that, In step S3, the scene feature quantization formula ， In the formula: S represents the comprehensive score of scene features; w T Indicates the weight of time features; T represents the standardized value of the time characteristic; w P Indicates spatial feature weights; P represents the standardized value of spatial features; w D Indicates the weight of persistent features; D represents the standardized value of the persistent feature; User role dependency formula: ， In the formula: R represents the overall score of user role dependency; w M Indicates production dependency weight; M represents the standardized value of production dependence; w L Indicates the weight of basic living security L represents the standardized value of living security; w Sd Indicates the special device dependency weight S d This indicates that special equipment depends on standardized values; Formula for the correlation of hidden pain points: ， In the formula: C is the pain point relevance of the scene-role combination, 0≤C≤1; S is a scene feature vector with 1×n dimensions; The transpose of the RT user role dependency vector, with an n×1 dimension; The confidence formula for hidden pain points is: ， In the formula: Confidence represents the level of confidence in hidden pain points; α represents the relevance weight; β represents the semantic similarity weight; Sim represents the semantic similarity value, which is the matching result from a pre-trained corpus.

8. The method for in-depth mining of power outage needs of electricity customers for customer service purposes as described in claim 1, characterized in that, In step S4, the system first reads the scenario-based requirement list output in S3 and assigns a three-dimensional value to each requirement: the urgency dimension is assigned a value based on the power dependence level system, with medical equipment dependence requirements assigned 10 points, general lighting requirements assigned 5 points, and non-essential power requirements assigned 3 points; the impact range dimension is assigned a value based on the number of users covered or the size of the area, with the entire building assigned 8 points and a single household assigned 3 points; the resource accessibility dimension is assigned a value linked to real-time data from the emergency resource database, with an emergency generator vehicle being available and reachable within 3 kilometers assigned 9 points, and temporary lighting equipment having sufficient inventory and a delivery distance of ≤1 kilometer assigned 9 points; then, a weighted summation model is used to calculate the total score, with the weights allocated as follows: urgency 60%, impact range 30%, and resource accessibility 10%.

9. The method for in-depth mining of power outage needs of electricity customers for customer service purposes as described in claim 1, characterized in that, In step S4, the system sorts all requests in descending order of total score, and defines the top 20% as high-priority requests. For such requests, the system triggers the resource database linkage mechanism, retrieves the real-time dispatch table of emergency power generation vehicles, the temporary equipment inventory ledger and the delivery route planning module, generates a matching list containing resource dispatch path, estimated arrival time and backup plan, and pushes it to the customer service specialist workbench of the corresponding region.

10. The method for in-depth mining of power outage needs of electricity customers for customer service purposes as described in claim 1, characterized in that, In step S5, the system sorts all requests in descending order of total score, and defines the top 20% as high-priority requests. For such requests, the system triggers the resource database linkage mechanism, retrieves the real-time dispatch table of emergency power generation vehicles, the temporary equipment inventory ledger and the delivery route planning module, generates a matching list containing resource dispatch path, estimated arrival time and backup plan, and pushes it to the customer service specialist workbench of the corresponding region.

Images