Intelligent reply method for electric power marketing inspection work order

Abstract

The invention discloses an intelligent reply method for an electric power marketing inspection work order. The method comprises the following steps: S1, obtaining work order data of a current inspection work order and a multi-dimensional capability index of a current business person to construct a state vector; s2, inputting the state vector into a pre-trained hierarchical reinforcement learning model, and outputting a decision instruction containing a high-level strategy instruction and a bottom-layer reply action; s3, determining a target interaction mode from a plurality of preset interaction modes based on the high-level strategy instruction; s4, displaying a corresponding reply auxiliary information set on a user interface based on different target interaction modes, and receiving actual operation of a user; the auxiliary information set is generated based on bottom-layer reply actions; and S5, calculating a reward value based on an actual operation result, and updating the hierarchical reinforcement learning model by using the reward value. According to the invention, capability states and work order characteristics of business personnel can be dynamically sensed, and real personalized intelligent assistance is provided.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power marketing, and particularly relates to an intelligent reply method for a power marketing inspection work order. BACKGROUND

[0002] In the field of power marketing inspection, efficiently and accurately processing and replying to various inspection work orders is a key link to guarantee power supply service quality and standard operation. At present, technical solutions in the industry aimed at improving work order processing efficiency mainly include two ways: one is a process automation driven solution, such as a robot process automation technology, which automatically completes the distribution, state tracking and archiving of work orders by simulating manual operation, and its core capability lies in the accurate execution of predetermined rules and structured processes; the other is a content generation driven solution, such as a general question and answer technology based on generative artificial intelligence, which parses user questions through a natural language processing module and generates coherent text replies using a pre-trained large language model, and its advantage lies in the general language understanding and generation capability.

[0003] However, the above two types of technical paths have limitations in dealing with the reply scene of professional power marketing inspection work orders. First, the process automation driven solution is essentially a script execution of preset rules, which cannot understand the deep semantics of the work order content and completely lacks the ability to analyze and make decisions on complex and variable inspection topics, with low intelligence level. Second, although the content generation driven solution can generate fluent text, its model lacks deep encoding of knowledge, standard system and business logic in the field of power marketing, and cannot perceive and adapt to the professional ability state of different business personnel, resulting in hidden dangers in the professional, accuracy and risk control of its reply. What is particularly important is that existing technologies generally lack an interactive way to quantitatively evaluate the quality of the reply process and the ability growth of business personnel, which cannot realize personalized intelligent assistance, and is also difficult to optimize its decision-making ability through continuous learning, so as to meet the high requirements of reply quality, compliance and personnel empowerment in actual business.

[0004] A power marketing inspection work order intelligent management method based on robotic process automation is disclosed in Chinese Patent Publication No. CN119204973A. The user logs in through the sequentially arranged login marketing system module; the work order generation module generates inspection work orders based on system inspection rules and captures relevant topic data; the work order assignment module assigns work orders according to inspection topics and processing personnel; the work order reminder module sends reminder information after sorting the work orders; the review module gives corresponding rollback conditions and rollback reasons for different approval results; and finally, the archive module is archived. The invention cannot understand the deep semantics of the work order content and does not have analysis and decision-making capabilities; and the system inspection rules in the invention exist in the form of hard coding or configuration tables, lack depth coding of power marketing field knowledge and standard system, and cannot handle situations outside the rules; at the same time, the invention cannot perceive and adapt to the professional ability of different business personnel. SUMMARY

[0005] The purpose of the present application is to provide an intelligent reply method for power marketing inspection work orders, which can dynamically perceive the ability state of business personnel and the characteristics of work orders by constructing multi-dimensional quantitative business indicators and hierarchical reinforcement learning models, and provide truly personalized intelligent assistance. The present application not only solves the problem of automatic processing, but also through continuous learning and optimization mechanism, makes the system have the ability to grow together with the business personnel, significantly improves the intelligent level and comprehensive efficiency of work order processing under the premise of ensuring processing standardization and risk control.

[0006] The technical solution of the present application is as follows: On the one hand, the present application provides an intelligent reply method for power marketing inspection work orders, comprising the following steps: S1: obtaining work order data of the current inspection work order and multi-dimensional ability indicators of the current business personnel, and constructing a state vector based on the two; S2: inputting the state vector into a pre-trained hierarchical reinforcement learning model, and outputting decision instructions containing high-level strategy instructions and low-level reply actions from the hierarchical reinforcement learning model; S3: determining a target interaction mode from a plurality of preset interaction modes based on the high-level strategy instructions; different interaction modes correspond to different display granularity and automation degree of reply assistance information; S4: based on different target interaction modes, displaying a corresponding set of reply assistance information on the user interface and receiving actual operations of the user; the set of assistance information is generated based on the low-level reply action; S5: calculating a reward value based on the result of the actual operation, and updating the hierarchical reinforcement learning model using the reward value.

[0007] Preferably, the multi-dimensional capability index of the service personnel includes risk sensitivity, rectification effectiveness, specification reference accuracy and timeliness coefficient. The risk sensitivity is calculated as:

[0008] Wherein: represents the total number of risk points actively identified by the service personnel in historical work orders; represents the total number of risk points determined by the system to exist in historical work orders; represents the one-time pass rate of historical high-risk work orders of the service personnel; represents the benchmark pass rate, i.e. the average pass rate of high-risk work orders of all service personnel; and are weight coefficients, which are set by domain experts' experience and calibrated by historical data regression analysis; The rectification effectiveness is calculated as:

[0009] Wherein: represents the recurrence rate of similar work orders handled by the service personnel; represents the average score of operability of the proposed rectification measures, which is evaluated by the system according to the specificity, checkability and integrity of the measures, with a score range of 0 to 1; and are weight coefficients; The specification reference accuracy is calculated as:

[0010] Wherein: represents the number of work orders in which the service personnel correctly referenced the specification clauses; represents the total number of work orders handled by the service personnel; represents the relevance score of the referenced clauses to the work order problems, which is calculated by the system based on a semantic similarity model, with a score range of 0 to 1; and are weight coefficients; The timeliness coefficient is calculated as:

[0011] Wherein: represents the benchmark processing time, i.e. the average processing time of similar work orders by all service personnel; represents the average processing time of the service personnel; represents the on-time completion rate of urgent work orders; and is a weight coefficient.

[0012] Preferably, the state vector is constructed based on the current inspection work order data and the multi-dimensional capability indicators of the current service personnel, and specifically is: vectorizing the work order data: for numerical work order data features, directly as numerical vectors; for categorical work order data features, converted into fixed-length vectors through one-hot encoding; for text work order data features, extracted into semantic vectors through a pre-trained language model; the multi-dimensional capability indicator values including risk sensitivity, rectification effectiveness, specification reference accuracy, and timeliness coefficient are directly taken as numerical vectors; the work order feature vector and the multi-dimensional capability indicator vector are spliced to form the state vector.

[0013] Preferably, the hierarchical reinforcement learning model adopts a double-network coupling architecture, composed of a high-level policy network and a bottom-level action network. The high-level policy network is a fully connected neural network, including an input layer, a hidden layer, and an output layer; the input layer has the same dimension as the state vector; the high-level policy network contains three hidden layers, each with 512, 256, and 128 neurons, respectively, using a ReLU activation function to introduce a nonlinear transformation; the output layer uses a Softmax activation function, and the output dimension is consistent with the number K of preset interaction modes, with each output node corresponding to the probability distribution of a high-level policy instruction; The bottom-level action network is a modular network set, containing K independent sub-networks, each corresponding to a high-level policy instruction; each sub-network is a fully connected neural network, containing an input layer, a hidden layer, and an output layer; the input layer receives the same state vector, and the hidden layer structure is dynamically adjusted according to the complexity of the corresponding high-level policy instruction; the output layer has the same dimension as the number of candidate bottom-level reply actions under the corresponding high-level policy instruction, and uses a Softmax activation function to output the Q value of each action; The high-level policy network and the bottom-level action network are connected through a policy selector, which selects the corresponding bottom-level action sub-network according to the probability distribution output by the high-level policy network, forming a complete decision-making path.

[0014] Preferably, the generation process of the policy instruction in step S2 is divided into two stages: policy making and action selection. In the policy making stage, the state vector is input into the high-level policy network, and after forward propagation calculation, the policy probability distribution P is obtained, and the policy selector selects the current high-level policy instruction according to the probability distribution P using the ε-greedy strategy ; the policy selection formula is:

[0015] wherein: The exploration rate parameter is gradually decayed during the training process; In the action selection stage, based on the selected high-level policy instruction , the corresponding low-level action sub-network is activated, the state vector is input into the activated low-level action sub-network, and the Q value estimation of all candidate reply actions is obtained, and the action selector selects a specific low-level reply action according to the Q value distribution using the same ε-greedy strategy ; The final output decision instruction is a binary tuple ( , ), wherein represents the high-level policy instruction, which defines the interaction mode that the system should adopt; represents the low-level reply action, which specifies the specific reply content or operation.

[0016] Preferably, the interaction mode in step S3 includes a guided learning mode, an efficiency optimization mode, and an expert autonomous mode; In the guided learning mode, the system provides the highest granularity of information display; the user interface will display the complete reasoning process of the work order analysis, including the basis for problem positioning, detailed interpretation of relevant specification clauses, item-by-item explanation of risk points, and comparison of advantages and disadvantages of each alternative scheme; the automation level of the guided learning mode is the lowest; In the efficiency optimization mode, the system uses medium-granularity information display; the user interface focuses on key decision points, displays the filtered alternative schemes and corresponding core advantage and disadvantage comparison, and omits the reasoning process; the automation level of the efficiency optimization mode is medium; In the expert autonomous mode, the system uses the minimum granularity of information display; the user interface only presents the final recommended scheme and key indicators, omitting all intermediate processes and auxiliary explanations; the automation level of the expert autonomous mode is the highest.

[0017] Preferably, in step S5, a multi-dimensional composite reward function is used to calculate the reward value based on the matching degree of the user's actual operation and the decision instruction, and the multi-dimensional composite reward function includes operation matching degree reward, review result reward, risk coverage degree reward, specification reference accuracy reward, and efficiency reward; The calculation of the operation matching degree reward is as follows:

[0018] wherein: represents the feature vector of the user's actual operation; represents the feature vector of the system recommended action; These are the matching degree weight coefficients; the feature vector contains encoded information in three dimensions: operation type, operation object, and operation parameters. The audit result reward The calculation is expressed as:

[0019] in: This indicates the work order approval rate. Indicates the benchmark pass rate. This indicates the rate of work order returns for modification. and These are the weighting coefficients; The risk coverage reward The calculation is expressed as:

[0020] in: This indicates the number of risk points identified in the response. This indicates the total number of risk points identified by the system. This indicates the quality score of the risk response measures. and These are the weighting coefficients; The accuracy reward for the reference of the standard The calculation is expressed as:

[0021] in: The accuracy score of the cited clauses is indicated. This indicates the relevance score between the cited clauses and the work order issue. and These are the weighting coefficients; The efficiency reward The calculation is expressed as:

[0022] in: Indicates the baseline processing time. Indicates the actual processing time. This indicates the on-time completion rate of emergency work orders. and These are the weighting coefficients; The calculation of the reward value is expressed as follows:

[0023] in: This is the reward value.

[0024] On the other hand, the present invention provides an intelligent response system for power marketing inspection work orders, including a state vector construction module, a decision instruction generation module, an interaction pattern recognition module, an interaction execution module, and a model optimization module; The state vector construction module is used to obtain the work order data of the current inspection work order and the multi-dimensional capability indicators of the current business personnel, and construct a state vector based on the two. The decision instruction generation module is used to input the state vector into a pre-trained hierarchical reinforcement learning model, and the hierarchical reinforcement learning model outputs decision instructions containing high-level policy instructions and low-level response actions. The interaction pattern recognition module is used to determine the target interaction pattern from a set of preset interaction patterns based on the high-level strategy instructions; different interaction patterns correspond to different granularities and degrees of automation in displaying response assistance information. An interactive execution module is used to display a corresponding set of response assistance information on the user interface and receive the user's actual operation based on different target interaction modes; the set of assistance information is generated based on the underlying response action. The model optimization module is used to calculate the reward value based on the result of the actual operation, and to update the hierarchical reinforcement learning model using the reward value.

[0025] In another aspect, the present invention also provides an electronic device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method as described in any embodiment of the present invention.

[0026] In another aspect, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in any embodiment of the present invention.

[0027] Compared with the prior art, the present invention has the following technical effects: 1. This invention replaces the traditional single evaluation dimension with a multi-dimensional evaluation method. By dynamically calculating professional indicators such as risk sensitivity and rectification effectiveness, it achieves a precise quantitative assessment of the competence of business personnel.

[0028] 2. This invention reduces the dimension of action space in complex scenarios through collaborative decision-making between a high-level policy network and a low-level action network; the modular network architecture enables precise mapping between policies and actions, improving the accuracy of decision-making and response speed.

[0029] 3. This invention optimizes user experience and operational efficiency by dynamically adjusting the granularity of information display and the degree of automation. Attached Figure Description

[0030] Figure 1 This is an overall flowchart of the intelligent response method for electricity marketing audit work orders described in this invention; Figure 2 This is the overall architecture diagram of the hierarchical reinforcement learning model described in this invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments of the present application and with reference to the accompanying drawings.

[0032] Example 1 This embodiment provides an intelligent response method for electricity marketing audit work orders. (See attached document.) Figure 1 As shown, it includes the following steps: S1: Obtain the current inspection work order data and the multi-dimensional capability indicators of the current business personnel, and construct a state vector based on the two. Specifically, the work order data is obtained by using a script integrated in the browser, combined with a computer vision model, to perform real-time rendering analysis of the current power marketing inspection work order interface, dynamically identifying form elements such as input boxes, drop-down menus, and text fields in the interface. The filled work order number, user name, and subject content data are extracted from these elements and structured into a work order data object. The work order data object includes numerical features such as urgency, categorical features such as inspection topic and problem type, and textual features such as problem description. This method allows for general data acquisition from the graphical interface without direct access to the backend database.

[0033] The multi-dimensional capability indicators are numerical indicators calculated based on the historical work order processing records of business personnel, used to quantify their business processing capabilities. As a preferred embodiment, the multi-dimensional capability indicators for business personnel include: risk sensitivity, which measures the ability of business personnel to identify and respond to potential business risks in work orders; rectification effectiveness, which measures the actual effectiveness of the rectification measures proposed by business personnel; standard reference accuracy, which measures the ability of business personnel to accurately reference industry standards and company policies; and timeliness coefficient, which measures the ability of business personnel to quickly process work orders while ensuring quality.

[0034] Risk sensitivity The calculation is expressed as:

[0035] in: This represents the total number of risk points proactively identified by business personnel in historical work orders; This indicates the total number of risk points identified by the system in historical work orders; This represents the one-time pass rate of high-risk work orders in the history of business personnel, which is the ratio of the number of high-risk work orders passed on the first try to the total number of high-risk work orders. This represents the baseline pass rate, which is the average pass rate for high-risk work orders across all business personnel. and The weighting coefficients are set using the experience of domain experts and calibrated using historical data regression analysis.

[0036] The effectiveness of the rectification The calculation is expressed as:

[0037] in: This indicates the recurrence rate of similar work orders handled by business personnel, which is the ratio of the number of recurring work orders to the total number of work orders processed. The average score represents the feasibility of the proposed corrective measures. The score is evaluated by the system based on the specificity, verifiability, and completeness of the measures, and ranges from 0 to 1. and These are the weighting coefficients.

[0038] The precision of the specification reference The calculation is expressed as:

[0039] in: This indicates the number of work orders in which business personnel correctly cited the standard clauses; This indicates the total number of work orders processed by the sales staff; The score represents the relevance of the cited clause to the work order issue, calculated by the system based on a semantic similarity model, with a score ranging from 0 to 1. and These are the weighting coefficients.

[0040] The time efficiency coefficient The calculation is expressed as:

[0041] in: This represents the baseline processing time, which is the average time taken for all business personnel to process similar work orders. This indicates the average processing time for sales personnel; This indicates the on-time completion rate of emergency work orders, which is the ratio of the number of emergency work orders completed on time to the total number of emergency work orders. and These are the weighting coefficients.

[0042] As a preferred implementation method in this embodiment, the state vector is constructed based on the work order data of the current inspection work order and the multi-dimensional capability indicators of the current business personnel as follows: Work order data is vectorized as follows: For numerical work order data features such as urgency, they are directly used as numerical vectors; for categorical work order data features such as inspection topics, they are converted into fixed-length vectors through one-hot encoding; for textual work order data features such as problem descriptions, they are extracted into semantic vectors through a pre-trained language model. Multi-dimensional capability indicators, including risk sensitivity, rectification effectiveness, standard reference accuracy, and timeliness coefficient, are directly used as numerical vectors. The work order feature vector is concatenated with the multi-dimensional capability index vector to form a state vector. Represented as:

[0043] in: Represents a work order data vector. This represents a multi-dimensional capability indicator vector.

[0044] S2: Input the state vector into a pre-trained hierarchical reinforcement learning model, and have the hierarchical reinforcement learning model output a decision instruction containing high-level policy instructions and low-level response actions.

[0045] As a preferred embodiment of this practice, the hierarchical reinforcement learning model employs a dual-network coupling architecture, such as... Figure 2 As shown, it consists of a high-level policy network and a low-level action network, and hierarchical decision-making is achieved through dynamic routing.

[0046] The high-level policy network is a fully connected neural network, including an input layer, hidden layers, and an output layer; the dimension of the input layer is the same as the dimension of the state vector; the high-level policy network contains three hidden layers, with 512, 256, and 128 neurons in each hidden layer, respectively, and uses the ReLU activation function to introduce a nonlinear transformation; the output layer uses the Softmax activation function, and the output dimension is the same as the number of preset interaction modes K, with each output node corresponding to a probability distribution of a high-level policy instruction.

[0047] The underlying action network is a modular network collection containing K independent sub-networks, each sub-network corresponding to a high-level policy instruction; each sub-network is a fully connected neural network, containing an input layer, a hidden layer, and an output layer; the input layer receives the same state vector, and the structure of the hidden layer is dynamically adjusted according to the complexity of the corresponding high-level policy instruction; the dimension of the output layer is consistent with the number of candidate underlying response actions under the corresponding high-level policy instruction, and the Q-value of each action is output using the Softmax activation function.

[0048] The high-level policy network and the low-level action network are connected by a policy selector. The policy selector selects and activates the corresponding low-level action sub-network based on the probability distribution output by the high-level policy network, thus forming a complete decision path.

[0049] In a preferred embodiment of this invention, the generation process of the strategy instruction in step S2 is divided into two stages: strategy formulation and action selection. During the policy formulation phase, the state vector is input into the high-level policy network. After forward propagation, the policy probability distribution P is obtained. The policy selector selects the current high-level policy instruction based on the probability distribution P using an ε-greedy policy. The strategy selection formula is:

[0050] in: The exploration rate parameter gradually decays during the training process.

[0051] During the action selection phase, based on the selected high-level policy instructions... The corresponding underlying action subnetwork is activated, and the state vector is input into the activated underlying action subnetwork to obtain the Q-value estimates of all candidate response actions. The action selector selects the specific underlying response action according to the same ε-greedy strategy based on the Q-value distribution. The final output decision instruction is a binary tuple ( , ),in It represents high-level policy instructions and defines the interaction mode that the system should adopt; This indicates the underlying response action, specifying the specific response content or operation.

[0052] Furthermore, the training process of the hierarchical reinforcement learning model is divided into two stages: pre-training and online learning. During the pre-training phase, a training dataset is constructed using historical work order processing records, and network parameters are initialized through supervised learning. The high-level policy network is trained using the cross-entropy loss function to learn the implicit policy selection patterns in the samples. The low-level action network is trained using the mean squared error loss function to learn the value estimation of action selection under each policy.

[0053] During the online learning phase, a deep Q-learning algorithm is employed to continuously optimize network parameters using temporal difference error. Loss function. Defined as:

[0054] in: Indicates the current network parameters. Indicates the target network parameters. As a discount factor, For instant rewards, For the next state, For the next action.

[0055] The reward function of the hierarchical reinforcement learning model is designed as a multi-dimensional composite reward function, which forms a comprehensive reward signal through weighted summation. This model breaks sample correlation through an experience replay mechanism, uses a target network to stabilize the learning process, and employs a gradient descent algorithm to iteratively optimize network parameters, gradually improving decision quality.

[0056] The hierarchical architecture proposed in this embodiment effectively reduces the complexity of the action space by decomposing the decision-making process, while maintaining the flexibility of decision-making, enabling the model to adapt to the complex scenario requirements of power marketing audit work order processing.

[0057] S3: Based on the high-level strategy instructions, determine the target interaction mode from multiple preset interaction modes; different interaction modes correspond to different granularities and degrees of automation in displaying response auxiliary information.

[0058] As a preferred embodiment of this example, the interaction modes in step S3 include guided learning mode, efficiency optimization mode and expert self-management mode, each mode corresponding to a specific application scenario and user ability level.

[0059] The first interaction mode is a guided learning mode, suitable for business personnel whose skills are still developing, or for scenarios involving high-risk and complex work orders. In this mode, the system will provide maximum guidance and support to ensure the standardization and accuracy of business processing.

[0060] The second interaction mode is the efficiency optimization mode, which is designed for business personnel with intermediate business skills and is activated when handling routine work orders. In this mode, the system balances guidance and efficiency.

[0061] The third interaction mode is the expert autonomy mode, which is designed for highly skilled professionals and is activated when handling simple or highly repetitive work orders. In this mode, the system grants users maximum operational autonomy to fully leverage the advantages of automation and significantly improve work efficiency.

[0062] The system determines the target interaction mode by mapping high-level policy instructions to interaction modes. The policy probability distribution output by the high-level policy network is parsed by the policy decoder and converted into specific mode selection instructions.

[0063] The determination process follows the pattern matching principle: the system calculates the similarity between high-level policy instructions and predefined interaction pattern features, selecting the pattern with the highest matching degree as the target interaction pattern. The matching degree calculation formula is: ,in For the first The degree of matching between various interaction modes This represents the probability distribution vector output by the high-level policy network. Indicates the first The feature vectors for each interaction mode represent the business characteristics of that mode, such as the degree of automation, the granularity of information display, and the strength of risk control. These feature vectors are manually set by domain experts based on business needs or learned from historical data. For example, the feature vector corresponding to the guided learning mode... Efficiency optimization mode corresponds to Expert self-management mode corresponds to . and The dimensions are the same. When multiple patterns have similar matching degrees, the pattern with a lower degree of automation is selected first to ensure the security of business processing. At the same time, the system records the pattern selection history and optimizes the mapping relationship through statistical analysis to improve the accuracy of pattern selection.

[0064] Different interaction modes exhibit significant differences in the granularity of displaying auxiliary information and the degree of system automation, forming a hierarchical auxiliary system: In guided learning mode, the system provides the highest granularity of information display; the user interface will display the complete reasoning process of work order analysis, including the basis for problem location, detailed interpretation of relevant standard clauses, item-by-item explanation of risk points, and comparison of the advantages and disadvantages of each alternative solution; the automation level of the guided learning mode is the lowest, the system only provides information display, all decisions need to be manually confirmed by business personnel, and each operation step needs to be verified twice.

[0065] In efficiency optimization mode, the system uses medium-granularity information display; the user interface focuses on key decision points, displaying screened alternatives and corresponding core advantages and disadvantages comparisons, omitting the reasoning process; the automation level of the efficiency optimization mode is medium, the system can automatically fill in basic information and generate recommended solutions, but important decisions still need to be confirmed by business personnel.

[0066] In expert-driven mode, the system uses the smallest granularity of information display; the user interface only presents the final recommended solution and key indicators, omitting all intermediate processes and auxiliary explanations; the expert-driven mode has the highest level of automation, and the system supports one-click confirmation and batch processing functions, requiring only final review by business personnel.

[0067] Furthermore, the transitions between modes are smoothly achieved through animation effects, ensuring a consistent user experience. Manual switching between modes is also supported, balancing the intelligence of automated recommendations with the flexibility of human intervention.

[0068] S4: Based on different target interaction modes, display the corresponding set of reply assistance information on the user interface and receive the user's actual operation; the set of assistance information is generated based on the underlying reply action.

[0069] Furthermore, the guided learning mode's response support information set includes a complete problem analysis path, specifically presented in the form of: multi-dimensional breakdown of the work order problem, detailed interpretation of relevant regulatory clauses, comparative analysis of the advantages of each alternative solution, and reference to the handling process of similar historical cases. At the interaction level, the system requires business personnel to independently judge and manually confirm each key decision point, and to ensure correct understanding through phased verification.

[0070] When the system determines to enable the guided learning mode, it loads a template displaying the corresponding set of auxiliary response information. This template uses a step-by-step guided approach, breaking down the complex work order response task into several sequentially executed sub-tasks. At each sub-task stage, the system provides the necessary professional knowledge and decision support, awaiting the business personnel to complete the corresponding practical operations. These practical operations include, but are not limited to: selecting the most appropriate processing direction from multiple alternative solutions, adjusting the system-recommended wording according to the actual situation, and supplementing necessary on-site verification details. The system ensures that each operational step complies with business specifications through real-time compliance checks; only operations that pass verification can proceed to the next stage.

[0071] The response support information sets of the efficiency optimization mode and expert autonomy mode focus on core decision-making elements, mainly including optimized recommended solution summaries, key risk warning information, and necessary compliance verification results. At the interaction level, the system supports rapid selection and confirmation of solutions, provides batch processing entry points, and allows automated execution processes to be triggered by simple commands.

[0072] When the system determines to activate either the efficiency optimization mode or the expert self-management mode, it loads a template displaying the corresponding set of auxiliary response information. The template is results-oriented, directly presenting the optimal solution generated based on the underlying response actions. Business personnel's actual operations mainly focus on the review and confirmation of the solution, including: quickly browsing the core content of the system's recommended solutions, adjusting and optimizing specific details, and finally confirming the implementation of the solution. In this mode, the system provides quick operation functions, supporting one-click adoption of solutions and continuous processing of batch work orders, reducing manual intervention.

[0073] S5: Calculate the reward value based on the result of the actual operation, and update the hierarchical reinforcement learning model using the reward value.

[0074] In a preferred embodiment of this practice, step S5 employs a multidimensional composite reward function to calculate reward values ​​based on the degree of matching between the user's actual operation and the decision instruction. The multidimensional composite reward function includes an operation matching degree reward to assess the consistency between the user's actual operation and the system's recommended action, an audit result reward to reflect the final processing quality of the work order response, a risk coverage reward to assess the response content's ability to identify and respond to potential risks, a standard citation accuracy reward to assess the quality and appropriateness of the response content's citation of industry standards, and an efficiency reward to assess the time efficiency of work order processing.

[0075] The operation matching degree reward The calculation is expressed as:

[0076] in: The feature vector representing the user's actual operation; The feature vector representing the system's recommended action; The matching degree weight coefficient; the feature vector contains encoded information in three dimensions: operation type, operation object, and operation parameters.

[0077] The audit result reward The calculation is expressed as:

[0078] in: This indicates the work order approval rate. Indicates the benchmark pass rate. This indicates the rate of work order returns for modification. and This is a weighting coefficient. When a work order is approved on the first attempt, the system provides a positive reward; when a work order is returned for modification, the system provides a negative reward.

[0079] The risk coverage reward The calculation is expressed as:

[0080] in: This indicates the number of risk points identified in the response. This indicates the total number of risk points identified by the system. This indicates the quality score of the risk response measures. and These are the weighting coefficients.

[0081] The accuracy reward for the reference of the standard The calculation is expressed as:

[0082] in: The accuracy score of the cited clauses is indicated. This indicates the relevance score between the cited clauses and the work order issue. and These are the weighting coefficients.

[0083] The efficiency reward The calculation is expressed as:

[0084] in: Indicates the baseline processing time. Indicates the actual processing time. This indicates the on-time completion rate of emergency work orders. and These are the weighting coefficients.

[0085] The calculation of the reward value is expressed as follows:

[0086] in: This is the reward value.

[0087] Example 2 Accordingly, this embodiment provides an intelligent response system for power marketing inspection work orders, used to implement the method described in Embodiment 1, including a state vector construction module, a decision instruction generation module, an interaction pattern recognition module, an interaction execution module, and a model optimization module.

[0088] The state vector construction module is used to obtain the work order data of the current audit work order and the multi-dimensional capability indicators of the current business personnel, and construct a state vector based on the two.

[0089] The decision instruction generation module is used to input the state vector into a pre-trained hierarchical reinforcement learning model, and the hierarchical reinforcement learning model outputs decision instructions containing high-level policy instructions and low-level response actions.

[0090] The interaction pattern recognition module is used to determine the target interaction pattern from a set of preset interaction patterns based on the high-level strategy instructions; different interaction patterns correspond to different granularities and degrees of automation in the display of response auxiliary information.

[0091] The interaction execution module is used to display the corresponding set of reply assistance information on the user interface and receive the user's actual operation based on different target interaction modes; the set of assistance information is generated based on the underlying reply action.

[0092] The model optimization module is used to calculate the reward value based on the result of the actual operation, and to update the hierarchical reinforcement learning model using the reward value.

[0093] Example 3 This embodiment provides an electronic device, which includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the method described in Embodiment 1 of the present invention.

[0094] Example 4 This embodiment provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method described in Embodiment 1 of the present invention.

[0095] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, A and B simultaneously, or B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of singular or plural items. For example, at least one of a, b, and c can represent: a, b, c, a and b, a and c, b and c, or a and b and c, where a, b, and c can be single or multiple.

[0096] Those skilled in the art will recognize that the units and algorithm steps described in the embodiments disclosed herein can be implemented using electronic hardware, computer software, or a combination of electronic hardware and software. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0097] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0098] In the several embodiments provided in this application, any function, if implemented as a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, 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 described in the various embodiments of this application. 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.

[0099] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for intelligent reply of power marketing inspection work order, characterized in that, The method comprises the following steps: S1: obtaining work order data of a current inspection work order and multi-dimensional capability indexes of a current service personnel, and constructing a state vector based on the two; S2: inputting the state vector into a pre-trained hierarchical reinforcement learning model, and outputting decision instructions containing high-level strategy instructions and bottom-level reply actions from the hierarchical reinforcement learning model; S3: determining a target interaction mode from a plurality of preset interaction modes based on the high-level strategy instructions; different interaction modes correspond to different display granularities and automation degrees of reply auxiliary information; S4: displaying a corresponding set of reply auxiliary information on a user interface based on the different target interaction modes and receiving actual operations of a user; The set of auxiliary information is generated based on the bottom-level reply action; S5: calculating a reward value based on the results of the actual operations, and updating the hierarchical reinforcement learning model using the reward value.

2. The method as claimed in claim 1, wherein, The multi-dimensional capability indexes of the service personnel include risk sensitivity, rectification effectiveness, specification reference accuracy, and timeliness coefficient; The risk acuity The calculation is represented as: Wherein: represents the total number of risk points that the business staff actively identified in the historical work order; represents the total number of risk points that the system determines to exist in the historical work order; represents the one-time pass rate of the business staff's historical high-risk work order; represents the benchmark pass rate, i.e., the average pass rate of all business staff's high-risk work orders; and is a weight coefficient, which is set by the experience of field experts and calibrated by historical data regression analysis; The effectiveness of the rectification The calculation is expressed as: wherein: represents the recurrence rate of the same type of work order that the service personnel is responsible for; represents the average score of the operability of the proposed improvement measures, which is evaluated by the system according to the specificity, checkability and integrity of the measures, with a score range of 0 to 1; and is a weight coefficient; The specification reference accuracy The calculation is expressed as: wherein: represents the number of tickets in which the business agent correctly referenced the canonical clause; represents the total number of tickets handled by the business agent; represents the relevance score of the clause to the ticket issue, computed by the system based on a semantic similarity model, with a score ranging from 0 to 1; and are weight coefficients; The aging coefficient The calculation of the aging coefficient is represented as: wherein: represents the reference processing duration, i.e. the average duration of all service personnel to process the same type of work order; represents the average processing duration of the service personnel; represents the on-time completion rate of urgent work orders; and is a weight coefficient.

3. The method as claimed in claim 1, wherein, The state vector is constructed based on the work order data of the current inspection work order and the multi-dimensional capability indexes of the current service personnel as follows: Vectorizing the work order data: directly taking numerical work order data features as numerical vectors; converting category type work order data features into fixed length vectors through one-hot encoding; and extracting text type work order data features into semantic vectors through a pre-trained language model; Taking the multi-dimensional capability index values including risk sensitivity, rectification effectiveness, specification reference accuracy, and timeliness coefficient as numerical vectors; Concatenating the work order feature vector and the multi-dimensional capability index vector to form the state vector.

4. The method as claimed in claim 1, wherein, The hierarchical reinforcement learning model adopts a double network coupling architecture and is composed of a high-level strategy network and a bottom-level action network; The high-level strategy network is a fully connected neural network and includes an input layer, hidden layers, and an output layer; the dimension of the input layer is consistent with the dimension of the state vector; the high-level strategy network includes three hidden layers, and the number of neurons in each hidden layer is 512, 256, and 128, respectively; a ReLU activation function is used to introduce a nonlinear transformation; The output layer uses a Softmax activation function, and the output dimension is consistent with the number K of preset interaction modes; each output node corresponds to a probability distribution of a high-level strategy instruction; The bottom-level action network is a modular network set and includes K independent sub-networks, each of which corresponds to a high-level strategy instruction; each sub-network is a fully connected neural network and includes an input layer, hidden layers, and an output layer; the input layer receives the same state vector, and the hidden layer structure is dynamically adjusted according to the complexity of the corresponding high-level strategy instruction; the dimension of the output layer is consistent with the number of candidate bottom-level reply actions under the corresponding high-level strategy instruction, and a Softmax activation function is used to output the Q value of each action; The high-level strategy network and the bottom-level action network are connected through a strategy selector, which selects and activates the corresponding bottom-level action sub-network according to the probability distribution output by the high-level strategy network to form a complete decision path.

5. The method as claimed in claim 4, wherein, The generation process of the strategy instructions in step S2 is divided into two stages of strategy making and action selection: In the policy making stage, the state vector is input into the high-level policy network, and the policy probability distribution P is obtained through forward propagation calculation. The policy selector selects the current high-level policy instruction according to the probability distribution P using the ε-greedy strategy ; the policy selection formula is: wherein: To explore the rate parameter, gradually decays with the training process; In the action selection stage, based on the selected high-level policy instruction , the corresponding low-level action sub-network is activated, the state vector is input into the activated low-level action sub-network, the Q value estimation of all candidate reply actions is obtained, and the action selector selects a specific low-level reply action according to the Q value distribution by using the same ε-greedy strategy ; the final output decision instruction is a binary tuple ( , ), wherein represents a high-level policy instruction, defining the interaction mode that the system should adopt; represents a low-level reply action, specifying specific reply content or operation.

6. The method as claimed in claim 1, wherein, The interaction mode in the step S3 includes a guided learning mode, an efficiency optimization mode and an expert autonomous mode; In the guided learning mode, the system provides the highest granularity of information display; The user interface will display the complete reasoning process of the work order analysis, including problem positioning basis, detailed interpretation of relevant specification clauses, item-by-item explanation of risk points, and comparison of advantages and disadvantages of each alternative; the automation level of the guided learning mode is the lowest; In the efficiency optimization mode, the system uses medium-granularity information display; the user interface focuses on key decision points, displays filtered alternative solutions and corresponding core advantage and disadvantage comparisons, and omits the reasoning process; the automation level of the efficiency optimization mode is medium; In the expert autonomous mode, the system uses minimum-granularity information display; The user interface only presents the final recommended solution and key indicators, omitting all intermediate processes and auxiliary explanations; the automation level of the expert autonomous mode is the highest.

7. The intelligent response method for electricity marketing audit work orders according to claim 1, characterized in that, The multi-dimensional composite reward function is used in the step S5 to calculate a reward value based on the matching degree of the actual operation and the decision instruction; the multi-dimensional composite reward function includes operation matching degree reward, audit result reward, risk coverage degree reward, specification reference accuracy reward and efficiency reward; The operation matching degree reward The calculation is represented as: wherein: represents a feature vector of actual operation of the user; represents a feature vector of recommended action of the system; is a matching degree weight coefficient; the feature vector contains encoding information of three dimensions of operation type, operation object and operation parameter; The audit result reward The calculation is represented as: wherein: represents the audit pass rate of the work order, represents the benchmark pass rate, represents the modification return rate of the work order, and is a weight coefficient; The risk coverage reward The calculation is expressed as: wherein: represents the number of risk points identified in the reply, represents the total number of risk points determined by the system, represents the quality score of the risk countermeasures, and is a weighting factor; The norm reference accuracy reward The calculation is expressed as: wherein: represents an accuracy score of the cited clause, represents a relevance score of the cited clause to the ticket issue, and is a weight coefficient; The efficiency reward The calculation is represented as: wherein: represents a reference processing duration, represents an actual processing duration, represents an emergency work order on-time completion rate, and is a weight coefficient; The calculation of the reward value is represented as: wherein: is a reward value.

8. An intelligent reply system for power marketing inspection work order, characterized in that, The system is used to implement the method in any one of claims 1 to 7, and includes a state vector construction module, a decision instruction generation module, an interaction mode recognition module, an interaction execution module and a model optimization module; The state vector construction module is configured to obtain work order data of a current inspection work order and multi-dimensional capability indicators of a current business personnel, and construct a state vector based on the two; The decision instruction generation module is configured to input the state vector into a pre-trained hierarchical reinforcement learning model, and output a decision instruction including a high-level strategy instruction and a bottom-level reply action from the hierarchical reinforcement learning model; The interaction mode recognition module is configured to determine a target interaction mode from a plurality of preset interaction modes based on the high-level strategy instruction; different interaction modes correspond to different display granularities and automation levels of reply auxiliary information; The interaction execution module is configured to display a corresponding set of reply auxiliary information on a user interface based on different target interaction modes and receive actual operations of a user; The set of auxiliary information is generated based on the bottom-level reply action; The model optimization module is configured to calculate a reward value based on the result of the actual operation, and update the hierarchical reinforcement learning model using the reward value.

9. An electronic device, comprising: A memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the method in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the method in any one of claims 1 to 7.

Images