Electric power marketing intelligent agent real-time decision early warning method

Abstract

The application relates to the technical field of electric power marketing data processing, and discloses an electric power marketing intelligent agent real-time decision early warning method; the method first accesses electric power marketing real-time data streams and business rule source event streams, carries out source identification marking, semantic field alignment and business time alignment on marketing data, and constructs an event chain oriented to a business object; a rule version identification code is established for a business rule clause, and a rule change event is identified; an intelligent agent decision context is constructed according to the rule change event, the event chain and business situation information, candidate decision suggestions are generated; the candidate decision suggestions are subjected to compliance verification and backtracking correction, the early warning triggering threshold is adjusted in combination with the business situation and the rule change, conflict arbitration is carried out on the candidate decision suggestions that pass the verification, and a decision record is generated, so that the compliance of electric power marketing decision suggestions, the timeliness of risk early warning and the traceability of a decision process are improved.

Description

Technical Field

[0001] This invention relates to the field of power marketing data processing technology, specifically a real-time decision-making and early warning method for power marketing intelligent agents. Background Technology

[0002] With the continuous advancement of the construction of new power systems, the power marketing business faces an operating environment characterized by frequent adjustments to electricity pricing policies, continuous updates to audit rules, dynamic changes in credit evaluation, and the continuous integration of new businesses such as virtual power plants, green electricity trading, and demand response. This places higher demands on the real-time performance, accuracy, and compliance of marketing decisions and risk warnings. For example, the published invention patent application CN112862172B discloses a method, device, computer equipment, and storage medium for predicting power outage complaints on the State Grid 95598 hotline. It constructs a complaint prediction model by linking power outage data, emergency repair work order data, and complaint work order data to predict potential customer complaints after a power outage. The published invention patent application CN119557408B discloses a method and system for intelligent question answering of power audit knowledge based on big data AI. It realizes power audit knowledge question answering through knowledge graphs, semantic recognition models, and feedback recognition models. While the aforementioned technologies can provide some support in complaint prediction, knowledge-based question answering, or auxiliary analysis, their processing mainly relies on historical data models, existing knowledge bases, or fixed question-and-answer chains. This makes it difficult to adapt in a timely manner to dynamic updates to electricity marketing business rules as policy documents, auditing methods, credit rules, and regulatory requirements change. Furthermore, existing intelligent agents, when generating marketing decision suggestions and risk warning conclusions, typically rely on rule knowledge already solidified at the training time and use preset thresholds for warning triggering. When business rules are updated or rules across different business lines change cross-cuttingly, the system struggles to promptly detect rule version changes and accurately reflect the output suggestions. Verifying and retrospectively correcting the consistency between new rigid business rules can easily lead to issues such as compliance overreach, rule conflicts, and mismatches between the timing of early warnings and the actual business situation. This can result in increased audit misjudgment rates, delayed early warnings of electricity bill collection risks, a surge in customer complaints, and a decline in the compliance of marketing operations. Therefore, there is an urgent need to propose a real-time decision-making and early warning method for power marketing agents that can perceive changes in marketing business rules in real time, verify and retrospectively correct the compliance of decision-making suggestions, and adaptively adjust early warning thresholds based on the business situation. This would improve the compliance of decision-making suggestions, the timeliness of early warning conclusions, and the accuracy of risk identification. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a real-time decision-making and early warning method for intelligent agents in power marketing, which solves the problems mentioned in the background section.

[0004] To achieve the above objectives, the present invention provides the following technical solution: Real-time decision-making and early warning methods for intelligent agents in electricity marketing include: S1. Access the real-time data stream of electricity marketing and the source event stream of business rules, perform source identification annotation, semantic field alignment and business time alignment on the marketing data, and build an event chain oriented towards business objects; S2. Establish rule version identification codes for the accessed business rules and identify rule change events based on the semantics, scope of application, and strength of constraints of the clauses; S3. Based on the rule change event, retrieve the associated business rules, event chain and business status information, construct the agent decision context, and call the agent to generate candidate decision suggestions; S4. Perform compliance verification on candidate decision recommendations according to semantic consistency, rule boundaries, and business impact order, and backtrack and correct candidate decision recommendations that fail the compliance verification. S5. Adjust the warning trigger threshold based on business situation information and rule change events, conduct conflict arbitration on candidate decision suggestions that have passed compliance verification, and generate warning conclusions and decision records.

[0005] Preferably, S1 includes: Generate a source identifier for the accessed data, consisting of the data source type, sampling frequency, and semantic ontology version number; Field normalization is performed based on the business object thesaurus and field mapping table; Perform co-occurrence matching and pending confirmation processing on unmatched fields; Generate event chain nodes based on business objects according to the business event time correction results; Event chain association edges are generated based on the preceding and following relationships of business rules and the stable relationships of historical events.

[0006] Preferably, S2 includes: Extract the clause text, rule source, clause number, rule effective date, applicable objects, and binding expressions from the rule data to be versioned; Semantic summary identifiers are generated through key term extraction, term weight ranking, and fixed-length hashing. A rule version identifier code is generated from the rule source type, clause number, version timestamp, semantic summary identifier, and effective status identifier; The current rule clauses are compared with the previous effective version in terms of semantics, scope of application, and binding strength. Corresponding events are marked for newly added clauses, repealed clauses, and rule source unreachable status.

[0007] Preferably, the semantic summary identifier is generated through key term extraction, term weight ranking, and fixed-length hashing, including: When generating semantic summary identifiers, non-semantic information is cleaned up and sentences are segmented in the main text of the rule clauses; Extracting key business terms and constraint terms from a power marketing terminology glossary; Term weights are determined based on word frequency, inverse document frequency, and constraint term identifiers; Terms are selected according to their weights, and the selected terms, their clauses, clause numbers, and rule effective dates are concatenated into a summary input string. Perform fixed-length hashing on the digest input string.

[0008] Preferably, S3 includes: Parse the business object, business category, execution area, execution time period, and trigger source in the business request; Extract event chain segments by business type; Search for rule terms by business category, customer category, region, time period, and effective status, and divide them into hard constraint area, reference constraint area, and rule change prompt area; Write the rule difference field, business situation profile, historical similar decision summary and data credibility tag into the decision context; The confidence level for generating candidate decision recommendations is determined based on rule citation completeness, event chain matching degree, business situation matching degree, and similarity to similar historical decisions.

[0009] Preferably, the generation confidence level of candidate decision suggestions is determined based on rule citation completeness, event chain matching degree, business situation matching degree, and similarity to historical similar decisions, including: The rule reference completeness is obtained by matching the valid rules corresponding to the current business request with the referenced valid rules. The event chain matching degree is obtained based on the correspondence between key event nodes and referenced event chain nodes; The business situation matching degree is obtained based on the consistency between the candidate disposal level and the risk level of the business situation profile. The similarity to historical decisions is obtained based on the semantic similarity between the candidate decision action and the historical similar decision action. The generation confidence score is obtained by weighting the rule reference completeness, event chain matching degree, business situation matching degree, and similarity of historical similar decisions.

[0010] Preferably, S4 includes: Extract action phrases and constraint objects from candidate decision suggestions; The candidate decision-making recommendations are broken down into the operating subject, operating object, operating time period, operating scope, and the version on which they are based; Semantic confidence is generated based on the results of phrase matching according to hard rules; The rule confidence score is generated based on the comparison results between the action elements and the rule-allowed set. Impact confidence is generated based on downstream relationships in the event chain and similarity to historical risk decisions; The overall compliance confidence level is adjusted by combining data credibility, and candidate decision recommendations are classified into sets, partial correction processes, or backtracking requests carrying deviation information and regenerated according to confidence level ranges.

[0011] Preferably, the candidate decision recommendations are broken down into the operating subject, operating object, operating time period, operating scope, and supporting version, including: Extract the execution role, applicable objects, execution time, action range, and rule basis identifiers from the candidate decision recommendations; Compare the execution roles with the permission sets; compare the applicable objects with the scope of the rules; Compare the execution time with the rule's effective period and processing time limit; Compare the action range with the monetary boundary, time limit boundary, threshold boundary, and processing level boundary respectively; Compare the rule reference identifier with the latest valid rule version identifier code.

[0012] Preferably, S5 includes: Threshold adjustment factors are generated based on time period factors, seasonal factors, policy window period factors, and special event factors. Adjust the basic early warning threshold according to the type of business indicator and add rule drift compensation; The warning level is determined based on the adjusted warning trigger threshold; Iterative arbitration of support levels is used to address action conflicts. For timing conflicts, the event chain is rearranged according to the topological order of the event chain and observation waiting periods are inserted; The final decision content, warning level, rule version identifier, threshold adjustment factor, and arbitration trajectory are written into the decision record.

[0013] Preferably, the basic early warning threshold is adjusted according to the business indicator type and rule drift compensation is superimposed, including: When adjusting the basic early warning thresholds, identify the risk direction of business indicators; For business indicators where the risk increases with the increase in indicator value, the trigger threshold is lowered according to the threshold adjustment factor. For business indicators where the risk increases when the indicator value decreases, the trigger threshold is increased according to the threshold adjustment factor. The drift compensation magnitude is determined based on the change level and the number of changed fields of the rule change event, and the adjusted warning trigger threshold is superimposed and corrected.

[0014] Compared with existing technologies, this invention provides a real-time decision-making and early warning method for power marketing intelligent agents, which has the following beneficial effects: 1. This invention constructs an event chain oriented towards business objects by uniformly accessing, marking source identifiers, aligning semantic fields, and aligning business time with real-time data streams and business rule source event streams in power marketing. It also establishes rule version identifiers for business rule clauses and identifies rule change events based on clause semantics, scope of application, and constraint strength. This allows dispersed marketing business data, dynamically updated rule clauses, and real-time business situations to jointly enter the decision-making basis of the intelligent agent. Furthermore, it generates candidate decision suggestions by constructing a dynamic decision context and performs progressive compliance verification on these suggestions according to semantic consistency, rule boundaries, and business impact. Candidate decision suggestions that do not comply with rule constraints are partially corrected or back-generated. Simultaneously, it adjusts the warning trigger threshold based on business situations and rule changes, and performs conflict arbitration and record-keeping on verified candidate decision suggestions. This improves the compliance of power marketing decision suggestions, the timeliness of risk warnings, the accuracy of business risk identification, and the traceability of the decision-making process.

[0015] 2. This invention transforms business rule clauses into rule version identifiers and rule change events, and writes rule difference fields, business situation profiles, historical similar decision summaries, and data credibility markers into the intelligent agent's decision context. This creates a structured relationship between rule updates, business requests, and candidate decision suggestions. When electricity pricing policies, audit rules, credit evaluation details, or regulatory requirements change, the system can update, review, and regenerate the associated decision context based on the rule change events. The manual review results are then written back as historical similar decision records. This reduces the reliance on manually reconfiguring fixed rules and readjusting the intelligent agent model, thereby reducing the workload of rule maintenance and manual review. It also reduces redundant configurations and judgments caused by inconsistent rule definitions across multiple systems, improving the maintainability, scalability, and operational stability of the power marketing intelligent agent in scenarios with continuous access across multiple business lines. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the real-time decision-making and early warning method for an intelligent agent in power marketing according to the present invention; Figure 2 This is a schematic diagram illustrating the multi-source data access and business object event chain construction of the present invention; Figure 3 This is a schematic diagram illustrating the generation of rule version identifier codes and the identification of rule changes in this invention; Figure 4 This is a schematic diagram illustrating the construction of the dynamic decision-making context for the intelligent agent in this invention; Figure 5 This is a schematic diagram illustrating the compliance verification and backtracking correction of candidate decision recommendations in this invention; Figure 6 This is a schematic diagram illustrating the dynamic adjustment of the early warning threshold and conflict arbitration of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1: Figures 1-6 A real-time decision-making and early warning method for power marketing intelligent agents is presented, including: S1. Access the real-time data stream of electricity marketing and the source event stream of business rules, perform source identification annotation, semantic field alignment and business time alignment on the marketing data, and build an event chain oriented towards business objects; S2. Establish rule version identification codes for the accessed business rules and identify rule change events based on the semantics, scope of application, and strength of constraints of the clauses; S3. Based on the rule change event, retrieve the associated business rules, event chain and business status information, construct the agent decision context, and call the agent to generate candidate decision suggestions; S4. Perform compliance verification on candidate decision recommendations according to semantic consistency, rule boundaries, and business impact order, and backtrack and correct candidate decision recommendations that fail the compliance verification. S5. Adjust the warning trigger threshold based on business situation information and rule change events, conduct conflict arbitration on candidate decision suggestions that have passed compliance verification, and generate warning conclusions and decision records.

[0019] This method is applied to real-time decision-making and early warning scenarios in power marketing centers. The processing objects include at least power marketing business data, business rule sources, agent decision-making processes, compliance verification processes, and early warning handling processes. Input information originates from channels such as electricity consumption information collection, marketing business management, customer service, payment settlement, power trading, audit management, credit evaluation, and rule publication. By uniformly accessing, semantically organizing, aligning, and versioning the dispersed marketing business data and dynamically changing business rules, it transforms them into a unified event chain and rule view that can be invoked by agents. This allows candidate decision suggestions to be generated based on the latest business rules, business event chains, and real-time business status. Output results include at least candidate decision suggestions, compliance verification results, early warning conclusions, and decision records. This method executes in the following order: data and rule access, rule change identification, decision context construction, compliance verification backtracking, threshold adjustment, and conflict arbitration. Each processing stage is connected via event chains, rule version identifiers, rule change events, candidate decision suggestions, and compliance confidence levels as intermediate data, thus forming a closed-loop real-time decision-making and early warning process for power marketing business.

[0020] Specifically, such as Figure 2 As shown: When electricity marketing business data enters the decision-making process, the data streams from the marketing business system and the event streams from the rule release channels are aggregated, and a traceable data foundation is established according to the data source, field semantics, and business occurrence time. The accessed data streams come from the electricity information collection system, marketing business management system, customer service system, payment and settlement system, and electricity trading platform. The data content includes at least electricity consumption curves, customer files, contract information, work order information, payment records, and settlement data. For the business rule source event streams, the rule source, release time, clause number, rule effective time, and rule text summary are extracted, and the extraction results are transmitted to the rule change identification process as the rule data to be versioned. To avoid inconsistencies in field naming, sampling periods, and time bases across different data sources that could distort decision-making, a source identifier is assigned to each piece of accessed data. The source identifier consists of three parts: data source type, sampling frequency, and semantic ontology version number. The data source type can be one of the following: electricity information collection, marketing management, customer service, payment settlement, electricity trading, and rule publishing. The sampling frequency can be one of the following: second-level, minute-level, 15-minute-level, hour-level, day-level, or event-triggered level. The semantic ontology version number is used to mark the field interpretation version used when the data enters the unified semantic view. The source identifier is passed along with the data in subsequent processes and is used to locate the data source in the decision record. After completing the source identifier labeling, the fields from different data sources are semantically aligned. Semantic alignment is achieved using a power marketing business object thesaurus and a field mapping table. The business object thesaurus includes at least the following fields: customer number, meter number, transformer area number, line number, contract number, electricity price category, electricity consumption value, electricity bill amount, payment status, audit items, credit rating, work order type, and rule clause number. The field mapping table records the correspondence between fields from each source system and unified fields. For example, the request number and work order number in the customer service system are uniformly mapped to work order identifiers; the billing time and receipt time in the payment system are uniformly mapped to payment confirmation time; and the effective date and execution date in the rule source are uniformly mapped to rule effective time. Numerical fields are simultaneously unified in units during alignment: electricity consumption is uniformly converted to kilowatt-hours, amount to RMB yuan, and time to Beijing time millisecond-level timestamps. Enumerated fields are lexicalized during alignment: overdue, unpaid, and pending deductions are categorized as incomplete payment status, while paid, received, and settled are categorized as completed payment status. If a field cannot find a corresponding relationship in the field mapping table, the field is placed in the set of fields to be confirmed, and candidate matching is performed based on the co-occurrence relationship of fields of the same business object within the observation window. Preferably, the observation window is the most recent 30 days, which corresponds to the monthly billing and payment observation period commonly seen in electricity marketing business, and can cover the occurrence of fields of the same business object within a complete electricity billing cycle. In daily settlement, real-time settlement, or high-frequency trading scenarios, the observation window can be adjusted to 7 to 30 days. The field candidate matching similarity M is calculated according to... The following parameters are defined: M1 represents field name similarity, M2 represents field value type consistency, and M3 represents the co-occurrence ratio of fields within the same business object. Field name similarity is determined by the ratio of the number of shared terms after field name segmentation to the union of terms in the candidate field and the standard field, or by the similarity between the semantic vectors of field names. Field value type consistency is determined based on whether the data type, unit, numerical range, and enumeration set of the fields are consistent. The field co-occurrence ratio is determined by the proportion of the number of times the field appears simultaneously with a known standard field within the observation window to the total number of times the field appears. Field names directly reflect... The naming conventions of the Yingyuan system have a significant impact on the initial matching, so their weight is set to 0.4. Field value type and field co-occurrence ratio are used to correct the consistency of physical meaning and business scenario, respectively, and their weights are set to 0.3. When the candidate matching similarity is lower than 0.8, it means that there is a significant inconsistency in at least one of the field name, value type, or co-occurrence scenario. This field will not enter the agent's decision-making process and will enter the set of fields to be manually confirmed. 0.8 is used to achieve a balance between scenarios with similar field names but different business meanings and scenarios with different field names but the same business meaning, reducing subsequent decision-making bias caused by mismapping. After completing the semantic alignment of fields, the marketing data is aligned with business time. Business time represents the actual occurrence or effective time of a business event, which is different from the time the data enters the system. For example, the access time of customer payment data may be later than the actual receipt time, the release time of electricity price rule data may be earlier than the effective time of the rule, and the creation time of customer complaint work order may be later than the time of the customer's first call. To distinguish the time differences between data generation, data access, and business occurrence, three time markers are established for each piece of data: data generation time, data access time, and business event time. The data generation time comes from the original timestamp of the source system, the data access time is the timestamp generated when the data is received, and the business event time is obtained by parsing the business field. When there is a business occurrence time field in the data, the business occurrence time field is used as the business event time. When there is no business occurrence time field but there is a source system generation time, the source system generation time is used as the temporary business event time, and a time source marker is attached. Time synchronization is preferably achieved using BeiDou time synchronization or unified time synchronization via the dedicated power grid. Under normal network conditions, the time synchronization error should not exceed 100 milliseconds, and under weak network conditions, the time synchronization error should not exceed 500 milliseconds. Normal network conditions refer to a round-trip latency of no more than 200 milliseconds and a continuous packet loss rate of no more than 5%. Weak network conditions refer to a round-trip latency of more than 200 milliseconds or a continuous packet loss rate of more than 5%. 100 milliseconds is suitable for high-consistency event sequencing in dedicated power grid or dedicated line environments, while 500 milliseconds is used to accommodate weak network, retransmission, and cross-system forwarding scenarios. When the deployment environment does not support BeiDou time synchronization, NTP network time synchronization is used, and the time synchronization error is allowed to be relaxed to 1 second. A network time synchronization mark is added to this type of data. The above error range matches the event sequencing accuracy of work order generation, payment confirmation, rule activation, and settlement processing in power marketing business, which can meet the needs of business judgment while avoiding frequent corrections caused by short-term link jitter. If time drift occurs in three consecutive sampling periods from the same data source, and the drift value exceeds the tolerance range of the corresponding sampling frequency, the data generated by that data source is marked as time-series aberration data and placed in the correction queue. The time drift value is the difference between the data generation time and the unified time synchronization time of the same data source in the same sampling period, or the deviation between the source system time interval and the standard sampling interval of two adjacent sampling periods. Three consecutive sampling periods are used to exclude occasional deviations caused by single transmission jitter, retransmission, or short-term delays in the source system. The time tolerance for second-level data is set to 1 second, the time tolerance for minute-level data is set to 5 seconds, and the time tolerance for 15 minutes is set to... The time tolerance for hourly, daily, and hourly data is set to 30 seconds. Second-level data is mainly used for real-time events such as work orders, payments, and rule releases. Minute-level data is mainly used for business status updates. Data at 15 minutes and above is mainly used for power consumption curves, settlements, and periodic file updates. Therefore, the above tolerance range is matched with the sensitivity of the sequence of business events under different sampling granularities. When the corresponding tolerance range is exceeded, a secondary correction is performed. Data in the correction queue is preferentially corrected by using the source system's retransmission time, adjacent business event times, and the event chain order of the same business object. Data that still cannot be corrected does not directly enter the intelligent agent decision-making process. On a unified timeline, an event chain is constructed using business objects as the primary index. Business objects include customers, transformer substations, lines, contracts, meters, work orders, audit items, or rule clauses. Event chain nodes must record at least the following fields: event identifier, business object identifier, event type, business event time, data source, key fields, related vouchers, and upstream event pointers. Event chain edges are generated from explicit sequential relationships in business rules and stable sequential relationships in historical event sequences. Explicit sequential relationships in business rules include, for example, installation application precedes on-site inspection, payment confirmation precedes account reconciliation, and audit project initiation precedes audit handling. Stable sequential relationships in historical event sequences include, for example, a complaint work order for the same customer after an overdue payment reminder, or a centralized consultation work order for the same transformer substation after power outage restoration. For the sequential relationships formed in the historical event sequence, a stability score is used for screening. The stability score is the ratio between the number of times the target successor event occurs within a set observation window after the occurrence of the preceding event, and the total number of occurrences of the preceding event. The observation window for the target successor event is determined according to the type of business event: 7 days for customer complaint events, 30 days for electricity bill collection events, 90 days for audit events, and 45 days for business expansion and installation events. These windows correspond to the complaint response cycle, monthly electricity bill collection cycle, regular audit cycle, and standard installation processing cycle, respectively, so that the judgment of the event chain edge matches the actual business cycle of electricity marketing. Relationships with scores above 0.85 are considered strong correlation edges, relationships with scores between 0.6 and 0.85 are considered reference edges, and relationships with scores below 0.6 are not included in the event chain. In historical event replays, relationships between events with stability scores above 0.85 typically recur in different months and among different customer groups, and can be considered strong correlation edges for subsequent decision-making. Relationships with stability scores below 0.6 are more susceptible to occasional events, and statistical co-occurrence can easily be misjudged as business causality, therefore they are not included in the event chain. Relationships with scores between 0.6 and 0.85 have some business reference value, but still need further confirmation in subsequent decision generation and compliance verification in conjunction with rules and regulations. The above thresholds are calibrated through historical marketing event replays of the past 12 to 24 months. The 12-month period covers the complete annual billing, seasonal electricity consumption, and routine audit cycle, while the 24-month period further covers cross-year policy adjustments, changes in customer behavior, and differences in rule application. Therefore, this sample period matches the operational patterns of the electricity marketing business. After the event chain is constructed, the event chain oriented towards business objects is output to the intelligent agent decision context construction process, serving as the business basis for generating candidate decision suggestions; at the same time, the rule data to be versioned is output to the rule change identification process, serving as the data foundation for establishing rule version identification codes and identifying rule change events; through the above processing, real-time power marketing data forms a unified expression in the three dimensions of source, semantics, and time, providing a data foundation for rule change identification, decision context construction, and compliance verification.

[0021] Specifically, such as Figure 3 As shown: Upon receiving the rule data to be versioned, the system reads the rule clause text, rule source, clause number, publication time, rule effective time, applicable objects, applicable regions, and binding expressions, and performs version identification and change recognition on each rule clause. To avoid version confusion when electricity pricing policies, inspection methods, credit evaluation rules, contract templates, and regulatory notices are updated through multiple channels, a rule version identification code is established for each rule clause. The rule version identification code consists of rule source type, clause number, version timestamp, semantic summary identifier, and effective status identifier. The rule source type can be one of electricity pricing, inspection, credit, contract, or regulation. The clause number follows the rule publisher's number; if the original number is missing, a replacement number is generated according to the rule title, chapter level, and clause order. The version timestamp uses a combination of the rule publication time and the rule effective time. The semantic summary identifier is used to represent the core semantics of the rule clause text. The effective status identifier is used to distinguish between pending, effective, invalid, and repealed states. Through the above identification code structure, the same rule can be distinguished under different publication times, effective statuses, or clause content versions. The semantic summary identifier can be selected as 64 hexadecimal characters. 64 hexadecimal characters correspond to the commonly used representation length of a 256-bit summary, which can reduce the probability of summary conflicts when the scale of rule clauses is large, and at the same time facilitates storage and comparison in the rule index table. When generating the semantic summary identifier, non-semantic information such as headers and footers, serial number formats, spaces, line breaks, and layout symbols in the body of the rule clauses are first deleted. Then, the clauses are divided into sentences according to periods, semicolons, commas, and clause level symbols. Subsequently, key terms are extracted based on the electricity marketing terminology glossary. Key terms include at least the following items: electricity price tier, electricity consumption category, customer category, billing cycle, recovery rate, audit items, credit score, credit rating, installed capacity, default handling, and regulatory time limit. These items are used to characterize the business constraint objects and execution boundaries in the rule clauses. The extracted key terms are weighted and sorted, with the term weights W according to... The definition is as follows: F is the normalized term frequency value of the term in the current clause, I is the normalized inverse document frequency value of the term in the rule clause set, and C is the constraint word identifier value. When the term is a strong constraint word, C is 1; when the term is not a strong constraint word, C is 0. The normalized term frequency value is used to reflect the current constraint object of the clause and has a significant impact on the subject identification of the rule summary, so its weight is set to 0.4. The normalized inverse document frequency value is used to reflect the distinguishing ability of the term among different rule clauses, and the constraint word identifier value is used to reflect whether the clause involves rigid compliance boundaries. Both are used as distinguishing and binding correction factors, and their weights are set to 0.3 respectively. After completing the term weight sorting, the key terms with the highest weights, the clauses containing the key terms, the clause number, and the rule effective time are concatenated to form the summary input string. The summary input string is then subjected to fixed-length hashing to obtain the semantic summary identifier. This semantic summary identifier serves as the semantic distinguishing field in the rule version identifier code and is used to distinguish version differences that occur when the content or wording of the same clause changes. After establishing the rule version identifier, the changes of the current version of the rule clauses are identified by comparing them with the previous valid version. The previous valid version is the rule version that was in effect or expired but not repealed before the current version was published, under the same rule source type and the same clause number. Change identification is carried out from three dimensions: clause semantics, scope of application, and binding strength. In the semantic dimension of the clauses, a standardized text sequence of the current version clause and the previous valid version clause is used as input. A semantic vector is generated through a semantic encoding model in the power marketing field. The semantic vector dimension can be selected from 384 to 768 dimensions, where 384 dimensions are suitable for real-time scenarios with a large number of rules and requiring fast comparison, and 768 dimensions are suitable for fine-grained comparison scenarios with longer clauses and fine semantic differences. This dimension range matches the output dimension of commonly used Chinese semantic encoding models, achieving a balance between semantic expressiveness and processing latency. The semantic similarity between the two versions of the clauses is calculated using cosine similarity. When the original value of cosine similarity is in the range of -1 to 1, the normalized semantic similarity S is calculated according to... Determine that cos is the original value of cosine similarity; when the cosine similarity output by the semantic coding model is in the range of 0 to 1, the cosine similarity is directly used as the semantic similarity; the semantic difference D1 is 1 minus the normalized semantic similarity S; In terms of scope of application, the fields such as customer category, business line, applicable region, applicable time period, and business object in the rule clauses are analyzed, and the intersection-union ratio (IU) of the current version and the previous effective version on the corresponding sets of the above fields is calculated. 1 is subtracted from the IU to obtain the difference degree of the corresponding set. The scope of application difference degree D2 can be selected as the arithmetic mean of the difference degrees of each set, so that changes in customer category, business line, applicable region, applicable time period, and business object can all participate in the judgment of the rule's boundary of action. When both the current version and the previous effective version are empty on a certain scope of application set, the difference degree of that set is recorded as 0; when only one is empty, the difference degree of that set is recorded as 1. Through the above processing, the inability to calculate the difference degree due to empty sets can be avoided, and changes in the scope of application can reflect the actual changes in the rule's boundary of action. In terms of constraint strength, the binding expressions in the current version of the rules and the previous effective version of the rules are identified and divided into strong constraint words and flexible constraint words. Strong constraint words include at least words such as "must," "shall," "must not," "prohibited," "strictly prohibited," and "required" that indicate mandatory enforcement, prohibition of enforcement, or mandatory conditions. Flexible constraint words include at least words such as "may," "suggest," "encourage," and "appropriate" that indicate permission, recommendation, or preferential treatment. For each version of the rules, the number of occurrences of strong constraint words, flexible constraint words, and the total number of occurrences of binding constraint words are counted. The total number of occurrences of binding constraint words is the sum of the number of occurrences of strong constraint words and the number of occurrences of flexible constraint words. The percentage of strong constraint terms, Pstrong, is determined by the ratio of the number of occurrences of strong constraint terms to the total number of occurrences of constraint expressions. The percentage of flexible expression terms, Pflexible, is determined by the ratio of the number of occurrences of flexible expression terms to the total number of occurrences of constraint expressions. The percentage of strong constraint terms corresponding to the current version of the rule clause is denoted as Pstrongcurrent, and the percentage of strong constraint terms corresponding to the previous valid version of the rule clause is denoted as Pstrongprevious. The percentage of flexible expression terms corresponding to the current version of the rule clause is denoted as Pflexiblecurrent, and the percentage of flexible expression terms corresponding to the previous valid version of the rule clause is denoted as Pflexibleprevious. The constraint strength difference degree, D3, is determined according to the following formula: in, This indicates the change in the proportion of strongly binding terms in the two versions of the rules. This indicates the change in the proportion of flexible expressions in the two versions of the rules; strong constraint words directly affect whether business actions are allowed to be executed, whether they must be executed, and whether there are prohibited boundaries, and have a significant impact on compliance boundaries, so the weight is 0.7; flexible expressions mainly affect the recommendation direction, priority, and execution preference of candidate decision suggestions, and have a relatively weak impact on compliance boundaries, so the weight is 0.3; through the above weighting method, the change in the execution rigidity of the two versions of the rules is converted into the difference in constraint strength within the range of 0 to 1; When neither the current version of the rule clause nor the previous valid version of the rule clause recognizes strong constraint words or flexible expression words, it means that there are no constraint expressions that can be used for comparison between the two versions, and the constraint strength difference D3 is recorded as 0; when only one version of the rule clause recognizes a constraint expression word, and the other version of the rule clause does not recognize a constraint expression word, it means that the rule constraint attribute has been added or disappeared, and the constraint strength difference D3 is recorded as 1; when both versions recognize a constraint expression word, the constraint strength difference D3 is calculated according to the above formula, and this constraint strength difference degree is used as the constraint strength dimension input in the comprehensive change degree calculation. After obtaining the semantic difference degree D1, the scope difference degree D2, and the binding strength difference degree D3, the overall change degree D is calculated using the following formula: Changes in the semantics of the clauses directly affect the core content of the rules; changes in the scope of application affect the boundaries of the rules; and changes in the strength of constraints affect the rigidity of enforcement and the level of compliance risk. Therefore, the weights of the three are set to 0.5, 0.3, and 0.2, respectively. When the overall change degree D is below 0.15, it is judged as a text fine-tuning event, only the rule version identifier is updated, and the decision context reconstruction is not triggered; when the overall change degree D reaches 0.15 and is below 0.3, it is judged as a minor rule change event, triggering the update of the associated decision context; when the overall change degree D reaches 0.3 and is below 0.55, it is judged as a moderate rule change event, triggering the update of the associated decision context, and the associated decision tasks being processed are reviewed in parallel; when the overall change degree D reaches 0.55, it is judged as a severe rule change event, freezing the associated candidate decision suggestions that are being generated or have not yet passed compliance verification, and starting the backtracking generation process. The aforementioned thresholds were calibrated by replaying historical rule update records from the past 12 to 24 months. During calibration, the historical rule update records were categorized into four types based on manually confirmed update types: format fine-tuning, partial adjustment, parameter adjustment, and structural adjustment. The comprehensive change distribution corresponding to each type of record was calculated. 0.15 was used as the boundary between the format fine-tuning and partial adjustment categories, 0.3 as the boundary between the partial adjustment and parameter adjustment categories, and 0.55 as the boundary between the parameter adjustment and structural adjustment categories. The 12-month period covers annual electricity price adjustments, audit cycles, and credit ratings. The 24-month pricing cycle covers cross-year policy changes and regulatory rule adjustments, thus matching the update frequency of electricity marketing rules. In historical playback, records below 0.15 typically correspond to adjustments in punctuation, numbering, format, or wording; records between 0.15 and 0.3 typically correspond to adjustments in partial descriptions or partial scope of application; records between 0.3 and 0.55 typically correspond to adjustments in execution parameters such as amount, tier, time limit, or scoring weight; records reaching 0.55 typically correspond to new business terms, new audit items, new electricity price tiers, or overall changes in the applicable objects. If the current rule clause is a newly added clause and no previous valid version can be found, the overall change degree D is set to 1 and marked as a newly added rule event. Since the newly added clause does not have a previous valid version, it is impossible to compare the clause semantics, scope of application, and binding strength with the same version, so it is treated as a severe rule change. If the current rule clause is repealed, a repeal rule event is generated, and the candidate decision task referencing the rule version identifier is marked as pending review. If the rule source is unreachable for 5 consecutive reachability detection cycles, the most recent valid rule view is used as the temporary rule view. The reachability detection cycle can be selected as 60 seconds, and 5 cycles correspond to about 5 minutes, which is used to filter short-term network jitter and avoid using outdated rule views for a long time. During the period when the rule source is unreachable, the associated decision tasks generated based on the temporary rule view are marked as the temporary state of the rule view and enter the pending review branch in the subsequent compliance verification. After the rule source becomes reachable again, the rule data is retrieved again and compared with the temporary rule view. The rule version identifier establishment and rule change identification are re-executed for the affected rule clauses. After completing the rule change identification, the rule change event is output. The rule change event records at least the trigger time, rule version identifier code, change dimension, comprehensive change degree, change level, applicable business object and suggested response method, and serves as the input for the intelligent agent decision context construction process. Through the above processing, the business rule terms can form a traceable expression in terms of version identifier, semantic content, scope of application and constraint strength, enabling the intelligent agent to update the context, conduct parallel review or backtrack generation based on the latest effective rules.

[0022] Specifically, such as Figure 4As shown: Upon receiving a business request or rule change event, the system reads the aforementioned rule change event, the event chain oriented towards the business object, the current business request, and real-time business status information. Before generating each candidate decision suggestion, it reconstructs the dynamic decision context that the agent can call, so that the latest rule version, business event chain, and real-time business status jointly participate in the generation of candidate decision suggestions. The business request types include at least the following power marketing business data processing requests: electricity fee collection risk handling, marketing audit auxiliary judgment, customer complaint warning, credit rating adjustment, business expansion application review, and market settlement anomaly prompts. When constructing a dynamic decision context, the business object identifier, business category, business subcategory, customer group, execution area, execution time period, and trigger source in the business request are first parsed. The business object identifier is used to locate business objects such as customers, contracts, transformer areas, work orders, meters, or rule clauses. The business category and business subcategory are used to determine the rule retrieval scope and event chain interception scope. The customer group, execution area, and execution time period are used to determine the rule application boundaries. The trigger source is used to distinguish whether the business request is triggered by a risk threshold, manually initiated, triggered by a rule change event, or triggered by an external work order. For example, in the scenario of handling the risk of large customer arrears, the business object identifier can be selected as the customer number or contract number, the business category is electricity fee collection, the business subcategory is arrears risk handling, and the trigger source can be selected as the electricity fee collection rate being lower than the basic threshold, an abnormal scissors difference between the increase in electricity consumption and the decrease in payment amount, a downgrade of customer credit rating, or a rule change event. After parsing the business request, the event chain oriented towards the business object is retrieved based on the business category and business object identifier, and the event chain segment related to the current business request is extracted. This extraction window is used to determine the scope of historical events entering the dynamic decision context, which is different from the subsequent event observation window used for event chain edge stability scoring. The extraction window for the event chain segment is determined according to the business type: 30 to 90 days for electricity bill collection, 60 to 180 days for auditing, 7 to 30 days for customer complaint, and 45 days for business expansion and installation. The window period is 30 to 90 days; the above window periods are matched with the processing cycles of each business. Specifically, the 30- to 90-day window for electricity bill collection corresponds to 1 to 3 billing cycles, which can cover the process of arrears formation, collection, and review; the 60- to 180-day window for auditing corresponds to the common audit preparation, verification, and handling cycle, which can cover the continuous changes in abnormal electricity use; the 7- to 30-day window for customer complaints corresponds to the concentrated outbreak of complaints and the short-term handling cycle; and the 45- to 90-day window for business expansion and installation corresponds to the cycle of application acceptance, on-site inspection, solution response, and document correction. After retrieving event chain fragments, the latest rule view is retrieved. Rule retrieval uses business category, customer category, execution area, execution time period, and rule effective status as filtering conditions to retrieve the set of rule clauses related to the current business request into the dynamic decision context. During retrieval, the rule clauses are classified by constraint type. Rule clauses containing mandatory expressions such as "must," "must not," "prohibited," and "shall" are placed in the hard constraint area, while rule clauses containing flexible expressions such as "may," "encouraged," "suggested," and "appropriate" are placed in the reference constraint area. The hard constraint area is used to determine the rule boundaries that candidate decision suggestions must not cross, while the reference constraint area is used to provide preferred directions. If the rule change event is a moderate or severe rule change event, the changed clauses are written into the rule change prompt area, and the difference fields between the current version and the previous effective version are recorded. The difference fields include at least the electricity price level, inspection threshold, credit score weight, application time limit, and customer applicable scope, so that the agent can locate the specific clauses and difference fields of the rule change when generating candidate decision suggestions. Real-time business status information is used to express the operational status of the current decision-making scenario. Real-time business status information includes at least time-based factors, seasonal factors, policy window factors, special event factors, customer risk factors, and data credibility factors. Time-based factors distinguish between late night, early morning, day shift, evening peak, and night shift. Seasonal factors distinguish between normal periods, peak summer periods, peak winter periods, Spring Festival power supply guarantee periods, and major holiday power supply guarantee periods. Policy window factors indicate whether rules have just taken effect, are being implemented stably, or are awaiting updates. Special event factors indicate whether there are events such as large customer arrears, group complaints, sudden power outages, or regulatory inspections. Customer risk factors consist of indicators such as continuous arrears period, fluctuations in payment amounts, changes in credit rating, and work order growth rate. Data credibility factors are derived from missing data, time-series anomalies, and fields marked as unconfirmed. Each factor is normalized and then written into the business situation profile. For continuous indicators, minimum-maximum normalization is used, and the normalized value is the ratio of the current value minus the lower limit to the upper limit minus the lower limit. The normalized value is 0 when it is lower than 0 and 1 when it is higher than 1. For graded indicators, they are mapped to 0 to 1 according to the ratio of grade number to the total number of grades. For Boolean indicators, they are 1 when the event exists and 0 when it does not exist. Time period factors, seasonal factors, and policy window factors are converted to values ​​between 0 and 1 using an enumeration mapping method. When multiple special events exist at the same time, the maximum value or average value of each special event label is taken as the special event factor. The above normalization method can bring business states of different dimensions into the same numerical range, which is convenient for context assembly and candidate suggestion confidence evaluation. Customer risk value R according to The following parameters are defined: R1 is the normalized value for consecutive overdue payment periods, R2 is the normalized value for payment amount volatility, R3 is the normalized value for credit rating changes, and R4 is the normalized value for work order growth rate. The normalized value for consecutive overdue payment periods, R1, is determined by the ratio of the current number of consecutive overdue payment periods to the maximum number of observation periods. The maximum number of observation periods can be selected as 3 billing periods, which can cover the complete process of overdue payment formation, collection reminders, and review and processing. The normalized value for payment amount volatility, R2, is determined by the deviation ratio of the current billing period's payment amount from the average payment amount of the past 3 billing periods. The past 3 billing periods can reflect the customer's recent payment... The stability of the electricity bill collection is assessed. The normalized value R3 for credit rating changes is determined by the ratio of the number of credit rating declines to the total number of credit rating levels. The normalized value R4 for work order growth rate is determined by the growth rate of the current 7-day work order count relative to the previous 7-day work order count. The 7-day period can reflect short-term concentrated changes in customer demands or complaints. The above weights are set according to the degree of impact of each indicator on the risk of electricity bill collection. The continuous arrears period has the greatest impact, followed by fluctuations in payment amount. Credit rating changes and work order growth rate are used as auxiliary risk factors. The customer risk value is not directly used as the final warning result, but rather as contextual input when the agent generates candidate decision suggestions. After retrieving the rule clauses, event chain fragments, and business situation profile, a dynamic decision context is generated according to a fixed structure. The dynamic decision context includes at least the following fields: business request summary, basic information of the business object, event chain fragments, latest hard rules, latest reference rules, rule change prompts, business situation profile, historical similar decision summaries, and data credibility markers. Historical similar decision summaries are derived from archived decision records, prioritizing archived decision records with the same or similar business object type, business category, business subcategory, and risk level. When there are fewer than 10 candidate records, the scope is broadened to archived decision records under the same business category. Historical similar decision summaries can select 10 to 50 records with the most recent occurrence time and extract the decision actions, compliance verification results, and execution feedback. The range of 10 to 50 records is used to strike a balance between sample diversity, context length, and retrieval latency. Fewer than 10 records make it difficult to form a stable reference, while more than 50 records increase context redundancy and reduce real-time processing efficiency. After constructing the dynamic decision context, the intelligent agent is invoked to generate candidate decision suggestions. Each time, 3 to 5 candidate decision suggestions are generated. Three suggestions cover three types of handling paths: conservative, conventional, and proactive. Five suggestions increase alternative options while controlling the burden of subsequent compliance verification. The candidate decision suggestions output by the intelligent agent include at least the following fields: candidate suggestion identifier, business action type, operating subject, operating object, execution period, execution scope, list of rule version identifiers, list of referenced event chain nodes, and generation confidence level. The business action type can be selected as reminder, review, audit, postponement, supplementary materials, transfer to manual processing, or generate an early warning. The operating subject can be selected as marketing personnel, audit personnel, account managers, or an automated early warning process. The operating object can be selected as a customer, work order, contract, meter, distribution area, or rule clause. The execution period indicates the time range for the suggested execution. The execution scope indicates the action boundaries such as amount, time limit, number of notifications, and review scope. The confidence level G is generated with a value between 0 and 1, and is calculated according to... The following parameters are defined: G1 is rule reference completeness, G2 is event chain matching, G3 is business situation matching, and G4 is similarity to similar historical decisions. Rule reference completeness (G1) is determined by the ratio of the number of valid rules referenced in the candidate decision suggestion to the number of valid rules corresponding to the current business request. Event chain matching (G2) is determined by the ratio of the number of key nodes in the event chain referenced in the candidate decision suggestion to the number of key event nodes in the current dynamic decision context. Business situation matching (G3) is determined by the consistency between the handling level corresponding to the candidate decision suggestion and the risk level of the business situation profile. Similarity to similar historical decisions... Similarity G4 is determined based on the semantic similarity between candidate decision suggestions and historical similar decision actions; rule reference completeness has a significant impact on compliance foundation, with a weight of 0.35; event chain matching degree and business situation matching degree reflect scenario adaptability, with weights of 0.25 respectively; historical similar decision similarity serves as an empirical reference, with a weight of 0.15; if the rule change event is a severe rule change event, the candidate decision suggestion should reference the new version rule version identifier code; if not referenced, it will be marked as insufficient rule basis in compliance verification; after the candidate decision suggestion is generated, the candidate pool is output to the compliance verification and retrospective correction process.

[0023] Specifically, such as Figure 5As shown: Upon receiving the candidate pool, the system reads the candidate decision recommendations, their referenced rule version identifiers, referenced event chain nodes, business situation profiles, and data credibility markers, and performs compliance checks in the order of semantic consistency, rule boundaries, and business impact. Semantic consistency checks determine whether the content of the candidate decision recommendations contradicts mandatory rule clauses; rule boundary checks determine whether the action elements of the candidate decision recommendations fall within the permitted scope of the rules; and business impact checks determine whether the execution of the candidate decision recommendations may trigger secondary compliance risks in downstream business events. The above checks are performed in a progressive manner from expression to action to impact. Semantic consistency verification takes candidate decision suggestion text, business action type, and hard rule clause text as input. First, core action phrases and constraints are extracted from the candidate decision suggestions. Core action phrases include at least the expressions for actions such as collecting electricity fees, suspending collection, initiating an investigation, downgrading credit rating, requesting supplementary materials, and expediting processing. Then, prohibited phrases, mandatory phrases, and conditional phrases are extracted from the hard rule clauses. Prohibited phrases include at least the expressions such as "must not," "prohibited," "strictly prohibited," and "must not perform specific actions on specific customer categories." Mandatory phrases include at least the expressions such as "must," "shall," and "must be completed within the specified time limit." Conditional phrases include at least the expressions such as "can only be executed when specific conditions are met," "executed after approval," and "triggered after reaching a threshold." After extraction, candidate actions are compared with rule phrases using a thesaurus and semantic vector similarity. The thesaurus is used for deterministic matching of rule phrases, and semantic vector similarity is used to supplement the identification of similar actions with different expressions. The comprehensive similarity between candidate actions and rule phrases is determined by both synonym matching results and semantic vector similarity. When a synonym is matched, the synonym matching value is 1; when no synonym is matched, the synonym matching value is 0. The synonym matching value and semantic vector similarity are weighted equally to obtain the comprehensive similarity. Synonym matching results reflect the deterministic matching of rule phrases, while semantic vector similarity reflects the generalized matching of natural language expressions. The weighted fusion of these two values ​​balances the accuracy of rule expression with the diversity of business action expressions. When the comprehensive similarity between a candidate action and a prohibited phrase is higher than 0.85, semantic consistency is considered achieved. Semantic consistency checks are marked as high-risk when the candidate action is contrary to the requirements of the mandatory phrase; when the candidate action requires the fulfillment of preconditions but the candidate decision suggestion does not reference the corresponding condition field, the semantic consistency check is marked as condition missing; 0.85 can be determined by replaying manually annotated samples of historical candidate action phrases and rule-prohibited phrases, and is used to distinguish between synonymous or near-synonymous illegal actions and general related actions; below this threshold, it is easy to misjudge general business-related expressions as illegal actions, and above this threshold, it is easy to miss near-synonymous expressions such as collection, demand, and recovery. The semantic confidence C1 ranges from 0 to 1, and can be initially set to 1. When a prohibited phrase is hit at a high risk, 0.4 is deducted; when a mandatory phrase conflict exists, 0.3 is deducted; when a condition is missing, 0.2 is deducted. If the value is lower than 0 after deduction, it is counted as 0. The above deduction range is set according to the degree of direct impact of semantic risk on compliance boundaries. Prohibited phrase conflicts have the greatest impact, followed by mandatory phrase conflicts. Missing conditions can usually be corrected by supplementing condition fields, so the deduction range is relatively low. Rule boundary verification uses the structured action elements of candidate decision suggestions and the set of allowed rules as input. Each candidate decision suggestion is broken down into five elements: operating subject, operating object, operating time period, operating scope, and version based. The operating subject is used to determine whether the candidate decision suggestion is executed by a role with the corresponding authority; the operating object is used to determine whether the applicable object falls within the scope of the rule; the operating time period is used to determine whether it is within the rule's effective period or the business processing time limit; the operating scope is used to determine whether the amount, time limit, threshold, and processing level exceed the rule boundary; and the version based is used to determine whether the rule version referenced by the candidate decision suggestion is the latest valid version. For candidate decision suggestions related to electricity bill collection, the operating scope should at least include the number of reminders, the amount to be collected, and the grace period; for candidate decision suggestions related to auditing, the operating scope should at least include the deviation multiple of abnormal electricity use, the audit cycle, and the handling level; for candidate decision suggestions related to credit rating, the operating scope should at least include the credit score deduction and the rating adjustment difference; and for candidate decision suggestions related to business expansion and installation, the operating scope should at least include the processing time limit, the number of supplementary materials, and the on-site inspection arrangement. The rule boundary verification compares each of the above five action elements. If any action element exceeds the rule's allowed set, the candidate decision suggestion is marked as failing the rule boundary verification, and the deviation element, the deviation rule version identifier code, and the deviation magnitude are recorded. The rule confidence level C2 ranges from 0 to 1 and can be determined by the ratio of the number of action elements that pass the verification to 5. When all five action elements pass the verification, the rule confidence level C2 is 1. When there are action elements that fail the verification, the rule confidence level C2 decreases as the number of passing elements decreases. When the version used is not the latest valid version or the object of operation does not belong to the applicable object of the rule, the candidate decision suggestion enters the process of pending correction or retrospective regeneration. The five action elements correspond to the permission boundary, object boundary, time boundary, magnitude boundary, and version boundary, respectively, which can cover the main execution constraints in the power marketing compliance review. Business impact verification uses event chain fragments, action elements of candidate decision suggestions, and business situation profiles as inputs. First, the business actions corresponding to the candidate decision suggestions are mapped to event chain nodes oriented towards the business object. Then, short-term impact is extrapolated downstream along the event chain. This extrapolation range is used to identify the short-term business impact after the execution of candidate decision suggestions, and is different from the subsequent event observation window and dynamic decision context capture window during event chain construction. The extrapolation range is determined according to the business type: 7 days for customer complaints, 30 days for electricity bill collection, 90 days for audits, and 45 days for business expansion and installation applications. The 7-day range for customer complaints corresponds to a concentrated outbreak of complaints and a short-term handling period; the 30-day range for electricity bill collection corresponds to one billing cycle; the 90-day range for audits corresponds to a regular audit cycle; and the 45-day range for business expansion and installation applications corresponds to a common installation processing cycle. Therefore, the above ranges match the business feedback cycle after the execution of candidate decision suggestions. Business impact verification does not predict all business outcomes, but rather identifies the likelihood of triggering compliance risk markers. Compliance risk markers are categorized as low, medium, and high risk. Low risk indicates that a candidate decision suggestion may only cause minor process adjustments; medium risk indicates that a candidate decision suggestion may cause downstream process rescheduling or manual review; and high risk indicates that a candidate decision suggestion may trigger rule conflicts, duplicate processing, or group customer complaints. If a strong correlation exists in the downstream event chain, and the candidate action changes the triggering order or triggering condition of that strong correlation, a business impact risk marker is generated. Business impact verification also references historical records of similar decisions, comparing the semantic similarity of the current candidate decision suggestion with those that resulted in complaints, failed reviews, or rule conflicts after execution. When the similarity is higher than 0.8, the risk marker level is increased, with low risk upgraded to medium risk and medium risk upgraded to high risk. A similarity of 0.8 can be calibrated through playback of historical risk decision records to identify candidate decision suggestions with high semantic similarity to historical risk actions. Below this threshold, it is easy to introduce generally similar business actions; above this threshold, it is easy to miss actions with the same risk meaning but different expressions. The confidence level C3 ranges from 0 to 1, and can be initially set to 1. A deduction of 0.1 is applied when a low-risk marker is generated, 0.25 when a medium-risk marker is generated, and 0.45 when a high-risk marker is generated. If a candidate action changes the triggering order or triggering condition of a strongly correlated edge, another 0.2 is deducted. If the similarity between the candidate decision suggestion and historical risk decision records is higher than 0.8, another 0.15 is deducted; if the result is lower than 0 after deduction, it is counted as 0. The above deduction ranges increase according to the severity of the business impact. Changes in strongly correlated edges and similarity to historical risk actions are used to reflect causal chain disturbances and the probability of historical risk recurrence, respectively, and are therefore considered as additional deduction factors. After obtaining the semantic confidence C1, rule confidence C2, and impact confidence C3, the overall compliance confidence C is calculated using the following formula: Semantic consistency and rule boundaries directly determine whether candidate decision recommendations violate rigid rules, so their weights are both set to 0.4; business impact mainly reflects secondary compliance risks, so its weight is set to 0.2 but not by default. Data credibility is used to adjust the overall compliance confidence level; the data credibility Q value ranges from 0 to 1, and can be adjusted according to... The criteria are defined as follows: Q1 represents the proportion of missing data on which candidate decision recommendations are based; Q2 represents the proportion of time-series anomalies; and Q3 represents the proportion of fields requiring confirmation. Missing data has a significant impact on the completeness of candidate decision recommendations, therefore its weight is set to 0.4. Time-series anomalies and fields requiring confirmation affect the chronological order of events and the semantic accuracy of fields, respectively, and their weights are set to 0.3. If the data confidence level Q is below 0.9, the overall compliance confidence level C is multiplied by the data confidence decay coefficient. 0.9 is used to distinguish between scenarios with basically complete data and scenarios with obvious data anomalies; this threshold can be determined through historical data quality replay. When the data confidence level is below 0.9, it indicates that the data on which candidate decision recommendations are based contains anomalies that affect the automatic approval process. When the data credibility Q is below 0.9, the data credibility decay coefficient is further determined based on the missing proportion. When the missing proportion is below 10%, the decay coefficient is 0.9 to 1; when the missing proportion reaches 10% but is below 30%, the decay coefficient is 0.7 to 0.9; when the missing proportion reaches 30% but is not higher than 50%, the decay coefficient is 0.5 to 0.7; when the missing proportion is higher than 50%, the candidate decision suggestion will not enter the automatic approval process. Within each segment, the data credibility decay coefficient can be determined by linear interpolation according to the missing proportion. The higher the missing proportion, the lower the decay coefficient. The above proportion range matches the fault tolerance processing requirements of multi-source data in power marketing. Local missing data below 10% can usually be supplemented by event chains and historical data. Missing data between 10% and 30% will have a significant impact on the completeness of the candidate decision suggestion. Missing data between 30% and 50% requires a significant reduction in the automatic approval probability. When the missing proportion exceeds 50%, the reliability of automatic judgment is insufficient. When the corrected overall compliance confidence score is higher than 0.92, the candidate decision recommendation enters the pass set; when the overall compliance confidence score reaches 0.8 and is not higher than 0.92, the candidate decision recommendation enters the mild correction process; when the overall compliance confidence score is lower than 0.8, the candidate decision recommendation enters the retrospective regeneration process; 0.92 and 0.8 can be calibrated by replaying the compliance review records of historical decision recommendations; during calibration, historical candidate decision recommendations are divided into three categories according to the post-review results: pass, pass after partial correction, and fail, and the overall compliance confidence score distribution corresponding to the three categories of records is statistically analyzed; 0.92 is taken as the boundary value between the pass category and the pass category after partial correction, and 0.8 is taken as the boundary value between the pass category after partial correction and the fail category; among them, 0.92 corresponds to the automatic pass interval with a high review pass rate, and 0.8 corresponds to the boundary interval where compliance can be restored through partial correction. Candidate decision recommendations with a score lower than 0.8 usually have structural deviations such as missing rule basis, execution object exceeding the boundary, or action range exceeding the limit; The mild correction process is for candidate decision suggestions with small deviations and complete rule basis but unclear expression or boundary conditions. During correction, the entire candidate decision suggestion is not regenerated; instead, partial correction instructions are generated based on the deviation elements. These partial correction instructions include at least supplementing the execution period, narrowing the applicable customer scope, reducing the operational scope, adding pre-communication nodes, and supplementing the latest rule version identifier. After partial correction, three-layer verification of semantic consistency, rule boundaries, and business impact is re-executed. The backtracking and regeneration process is for candidate decision suggestions with structural deviations. During backtracking, the deviation rule version identifier, deviation dimension, deviation action phrase, and deviation event chain node are written into the backtracking request, and the backtracking request is sent back to the dynamic decision context construction process, ensuring that the regeneration carries the deviation rule, deviation dimension, and deviation action constraint. The number of backtracking regeneration attempts can be set to no more than three. Three attempts are sufficient to cover the initial deviation correction, secondary generation after supplementing rule constraints, and adjustment of alternative paths in the candidate pool. If a candidate decision suggestion with a comprehensive compliance confidence score higher than 0.92 cannot be obtained after more than three attempts, it usually indicates a structural conflict between the business request and the current rule boundary. Continuing to generate automatically will increase processing latency and uncertainty. In this case, the manual review process is initiated. The manual review result is written into the verification record of the candidate decision suggestion as the final verification result and serves as the data source for subsequent historical similar decision records and threshold replay calibration. After compliance verification and correction are completed, the set and its comprehensive compliance confidence score will be output to the threshold adjustment, conflict arbitration, and record generation processes.

[0024] Specifically, such as Figure 6As shown: Upon receiving the approved set, the system reads the candidate decision suggestions that have passed compliance verification, the comprehensive compliance confidence level, the business situation profile, rule change events, the event chain, and the basic early warning threshold. Based on the business situation profile and rule change events, it calculates the threshold adjustment factor to form the adjusted early warning trigger threshold. To avoid premature, false, delayed, or missed triggers caused by static thresholds near the rule change point, an early warning judgment is formed based on the adjusted early warning trigger threshold. When multiple candidate decision suggestions exist for the same business object, conflict arbitration is performed on the candidate decision suggestions, and a traceable early warning conclusion and decision record are generated. Threshold adjustments start with the basic early warning threshold. The basic early warning threshold is established according to business lines. Specifically, for electricity bill collection risk, the threshold can be set as follows: a customer or distribution area's electricity bill collection rate is below 95% for two consecutive billing cycles. Two consecutive billing cycles are used to exclude occasional fluctuations caused by payment delays in a single billing cycle, and 95% is used to identify risky entities with monthly collection rates significantly lower than normal. For abnormal electricity consumption in marketing audits, the threshold can be set as follows: customer electricity consumption deviates from the historical average for the same period by 3 to 5 standard deviations. 3 standard deviations are used for general abnormal warnings, and 5 standard deviations are used for strong abnormalities or high-confidence audit triggers. For a surge in customer complaints, the threshold can be set as follows: the number of complaint work orders in a single day or distribution area is more than 3 times the average of the previous 7 days, and the number of complaint work orders reaches 3 or more. The average of the previous 7 days is used to reflect short-term service conditions, and more than 3 complaints are used to avoid... To avoid false triggers due to an excessively low baseline, the risk of overdue business expansion and installation applications can be mitigated by setting the current processing time to 80% of the standard processing time limit. This 80% is reserved for correction, work assignment, and review before the standard processing time limit expires. Different regions or business units can use historical data from the past 12 to 24 months for localized calibration. The 12-month data can cover the complete annual billing, seasonal electricity consumption, and regular business cycle, while the 24-month data can further cover cross-year policy changes and differences in customer behavior. During localized calibration, historical risk events, manually confirmed early warning records, and post-event handling results are used as samples. The basic early warning threshold is reviewed step by step within the candidate range, and the threshold with both false trigger rate and missed trigger rate within the acceptable range and the early warning time meets the handling requirements is selected as the localized threshold. The threshold adjustment factor is derived by integrating the time-period factor, seasonal factor, policy window period factor, and special event factor from the business situation profile. A factor greater than 1 indicates increased early warning sensitivity, while a factor less than 1 indicates decreased early warning sensitivity. The time-period factor reflects the level of business activity: 0.9 for late night, 1 for early morning, 1.1 for day shift, 1.2 for evening peak, and 1 for night shift. The seasonal factor reflects power supply and settlement pressure: 1 for normal periods, 1.1 for peak summer and winter periods, 1.2 for the Spring Festival power supply guarantee period, and 1.2 for major holidays. The daily power supply guarantee period is set at 1.15; the policy window period factor is used to reflect the sensitivity to rule changes, with a value of 1.2 for the initial effective period, 1 for the stable implementation period, and 1.1 for the pending update period; the special event factor is used to reflect sudden risks, with a value of 1 when no special event occurs, and a value of 1.2 to 1.5 when there are large customer arrears, group complaints, sudden power outages, or regulatory inspections; the special event factor is determined according to the scope of the event's impact, with a value of 1.2 when affecting a single customer, 1.3 to 1.4 when affecting a distribution area or business line, and 1.5 when involving regulatory inspections, group complaints, or sudden power outages; The comprehensive threshold adjustment factor F is based on It is determined that F1 is the time-period factor, F2 is the seasonal factor, F3 is the policy window factor, and F4 is the special event factor. The policy window factor and the time-period factor directly affect the timing of rule changes and the sensitivity of early warning during peak business periods, so their weights are both set to 0.3. The seasonal factor and the special event factor serve as supplementary variables for power supply pressure and sudden risks, and their weights are both set to 0.2. The comprehensive threshold adjustment factor F usually falls between 0.9 and 1.35 under common factor combinations. To avoid drastic fluctuations in the threshold due to abnormal factor inputs, the allowable range of F is set to 0.5 to 1.5. When it is lower than 0.5, it is taken as 0.5, and when it is higher than 1.5, it is taken as 1.5. After the comprehensive threshold adjustment factor is generated, the direction of threshold tightening or loosening is determined according to the type of business indicator. For indicators with higher risk values, such as the number of complaint work orders, abnormal electricity consumption deviation multiple, and arrears risk index, the base threshold is divided by F, and the trigger threshold is lowered when F is greater than 1. For indicators with lower recovery rates, such as electricity bill recovery rate, the base threshold is multiplied by F, and a higher recovery rate is required when F is greater than 1. For moderate rule change events, drift compensation is superimposed during the 6-hour transition period; for severe rule change events, drift compensation is superimposed during the 24-hour transition period. Drift compensation; 6 hours corresponds to the common cycle of completing business synchronization within half a working day for parameter adjustment rules, and 24 hours corresponds to the full business day cycle required for confirmation of clause structure changes, new rules, or cross-line rules; the drift compensation range is determined between 5% and 15% according to the rule change level and the number of changed fields, of which 5% is used for minor parameter drift, 10% is used for multiple parameters changing simultaneously, and 15% is used when there are significant changes in the applicable objects, business levels, or clause structure; risk thresholds are further reduced according to this range, and recovery rate thresholds are further increased according to this range; The adjusted warning trigger threshold is used to generate warning conclusions. If the current business risk indicator reaches the adjusted warning trigger threshold, a warning conclusion is generated and the warning level is marked. Warning levels are divided according to the deviation multiple. For indicators with higher risk values, the deviation multiple is the ratio of the current indicator value to the adjusted threshold. For indicators with lower recovery rates, the deviation multiple is the ratio of the adjusted threshold to the current indicator value. When the deviation multiple is less than 1.5, it is marked as information level; when it reaches 1.5 and is less than 3, it is marked as warning level; when it reaches 3, it is marked as emergency level. 1.5 is used to identify risks that have clearly exceeded the trigger boundary but can still be handled by the business line. 3 is used to identify emergency situations where the risk level has been significantly amplified. Information level is used for daily monitoring, warning level is used for supervision by business supervisors, and emergency level is used for cross-line intervention. If the data source that triggers the warning has a data credibility downgrade mark, the warning conclusion is appended with a data credibility mark, and a review action is added to the handling suggestion to reduce the risk of false triggering caused by missing data, time sequence anomalies, or fields that need to be confirmed. After generating the early warning conclusion, conflict arbitration is conducted on the candidate decision suggestions that have passed the compliance verification. Conflict arbitration uses the same customer, the same transformer area, the same contract, the same audit matter, or the same work order as the aggregation object, and collects related candidate decision suggestions in the statistics window. The statistics window is used to merge multiple candidate decision suggestions generated by the same business object in a short period of time. The window length is set according to the urgency of the business: customer complaints change rapidly and require a quick response, so the customer complaint business can be set to 5 minutes; marketing audits need to avoid duplicate project initiation and duplicate processing, so the marketing audit business can be set to 15 minutes; electricity bill collection and business expansion application allow for a longer aggregation time to reduce duplicate reminders and duplicate work assignments, so the electricity bill collection business and business expansion application business can both be set to 30 minutes; conflict types are divided into action conflict and time sequence conflict. Action conflict refers to the opposite direction of action between candidate decision suggestions, such as the same customer simultaneously having immediate collection and deferred collection; time sequence conflict refers to the action direction is not opposite but the execution order is unreasonable, such as conducting audit penalties before conducting pre-conference communication. For action conflicts, the initial support level is determined by combining comprehensive compliance confidence and business risk level. The initial support level is the comprehensive compliance confidence of the candidate decision recommendation multiplied by the risk level coefficient, with the coefficient set to 1 for information level, 1.1 for warning level, and 1.2 for emergency level. The risk level coefficient increases by 0.1 to reflect the impact of the warning level on arbitration priority, while avoiding the warning level from completely covering the comprehensive compliance confidence. The initial support level is no higher than 1; if the calculated result is higher than 1, it is set to 1. Subsequently, the support level is iteratively updated, and each round is adjusted based on the performance of the candidate decision recommendation in the latest compliance feedback, event chain consistency feedback, and threshold matching feedback. The latest compliance feedback indicates whether the candidate decision recommendation still meets the compliance verification requirements under the current rule version; the event chain consistency feedback indicates whether the candidate decision recommendation conforms to the event chain topology order and strong correlation edge constraints; and the threshold matching feedback indicates whether the handling level corresponding to the candidate decision recommendation matches the current warning level and the adjusted threshold. If a candidate decision suggestion has a boundary compliance risk, the support is multiplied by 0.8; if a candidate decision suggestion conflicts with a strongly related node downstream of the event chain, the support is multiplied by 0.6; if the latest compliance feedback, event chain consistency feedback, and threshold matching feedback all pass, the support is multiplied by 1.1 but not exceeding 1; 0.8 is used to reflect local deviations from boundary compliance risks, 0.6 is used to reflect the impact of strongly related node conflicts on the business causal chain, and 1.1 is used to enhance candidate decision suggestions that pass all three types of feedback; when the support after the iterative update is lower than 0, it is taken as 0, and when it is higher than 1, it is taken as 1; When the difference between the highest and second-highest support values ​​exceeds 0.2, and the change in the highest support value is less than 0.05 for two consecutive rounds, the candidate decision suggestion with the highest support value is selected. 0.2 is used to confirm that the highest support value has a significant advantage over the second-highest support value, 0.05 is used to determine that the highest support value has stabilized, and two consecutive rounds are used to exclude single-round feedback fluctuations. If the convergence condition is not met after 5 rounds of iteration, a manual review process is triggered, and the candidate decision suggestion, comprehensive compliance confidence level, rule version identifier code, event chain evidence, and warning level are output to the business personnel for adjudication. The 5 rounds are used to control the real-time decision latency and avoid the conflict arbitration process occupying the business processing link for a long time. To address timing conflicts, the content of the candidate decision recommendations is not altered; instead, the execution order is rearranged according to the topological sequence of the event chain. If two adjacent actions require waiting for feedback from the previous action, an observation and waiting period is inserted. For electricity bill collection, the observation and waiting period can be selected as 7 days to cover the payment communication, payment confirmation, and reminder cycle. For auditing, the observation and waiting period can be selected as 14 days to cover the on-site verification, document confirmation, and preliminary handling cycle. For complaint handling, the observation and waiting period can be selected as 24 hours to meet the requirements for rapid complaint response. For business expansion and installation, the observation and waiting period can be selected as 3 days to cover the document correction or on-site investigation and coordination cycle. These waiting periods match the feedback cycles of each business, allowing for adjustments to the execution order without changing the content of the candidate decision recommendations. After conflict arbitration is completed, the final adopted decision recommendations, early warning conclusions, and arbitration process will be generated into a decision record. The decision record will include at least the following fields: business object identifier, final decision content, early warning level, comprehensive compliance confidence level, list of referenced rule version identifiers, business status profile, threshold adjustment factor, conflict arbitration trajectory, referenced event chain nodes, data credibility identifier, and generation time. To facilitate post-event auditing, a compliance watermark will be generated for the core fields. The compliance watermark is formed by fixed-length hashing of the business object identifier, final decision content, rule version identifier list, and generation time. The compliance watermark can be a fixed-length hash value of 64 hexadecimal characters. 64 hexadecimal characters correspond to the commonly used representation length of a 256-bit digest, which can reduce the probability of record collisions and facilitate storage and comparison. If a compliance watermark inconsistency is found during post-event review, a record integrity alarm will be triggered. After conflict arbitration and record generation are completed, the final adopted decision recommendations, early warning conclusions, and decision records will be output to the business execution process, and the rule version identifiers, event chain nodes, threshold adjustment factors, and arbitration trajectories in the decision record will be used for subsequent review and playback.

[0025] Example 2: Based on Example 1, the specific application process of the real-time decision-making and early warning method for power marketing intelligent agents is further explained: In a specific application scenario, a municipal power supply unit is in the monthly electricity billing period and has received new time-of-use pricing rules. Simultaneously, a large high-voltage customer within its jurisdiction is experiencing a combination of anomalies: increased electricity load, delayed payment schedules, and a surge in customer service work orders. This customer's electricity bill collection rate for the previous billing cycle was 94.6%, and the current billing cycle's collection rate up to day 20 is 93.8%. The electricity consumption information collection system shows that the customer's average electricity consumption over the past 7 days has increased by approximately 16% compared to the average of the past 3 similar working days. The payment settlement system shows that the customer's recent... The amount paid in a single transaction decreased by approximately 11% compared to the average of the past three billing cycles; the number of work orders recorded by the customer service system within the same distribution area related to electricity billing, electricity bill collection, and power outage risk consultation increased significantly compared to the average of the previous seven days; in this scenario, business requests are identified as composite business requests for handling electricity bill collection risks and providing early warnings for customer complaints. The business object identifier includes at least fields such as customer number, contract number, and distribution area number. Triggering sources include electricity bill collection rate falling below the basic threshold, an abnormal scissors difference between electricity consumption growth and payment amount decrease, and electricity price rule change events; During the data access process, the system accesses 15-minute-level electricity consumption curves generated by the electricity consumption information collection system, customer files and contract capacity generated by the marketing business management system, billing records generated by the payment and settlement system, work order records generated by the customer service system, and time-of-use pricing rules pushed by the rule publishing channel. Source identifiers are generated for each data source, and fields such as customer number, contract number, meter number, transformer area number, payment confirmation time, work order identifier, and rule effective time are mapped to a unified field system. The arrival time and billing time in the payment system are uniformly merged into the payment confirmation time; the execution date and effective date in the rule publishing channel are uniformly merged into the rule effective time. Subsequently, using the customer number and contract number as the main index, an event chain is constructed on a unified business timeline. The event chain sequentially records event nodes such as continuous increase in electricity consumption, delayed payment confirmation, payment reminders, increase in customer inquiry work orders, and the effective date of time-of-use pricing rules. Based on the historical event sequence, the relationship between overdue payment reminders and the increase in complaint work orders is marked as a reference edge, and the relationship between payment confirmation and account reconciliation is marked as a strong correlation edge. During the rule change identification process, the text, clause number, publication time, effective time, applicable area, customer category, and binding expression of the new time-of-use pricing rule are read, and a rule version identification code is established for the rule. For the new and old time-of-use pricing rules, the semantic difference, the difference in the scope of application, and the difference in the binding strength are calculated respectively. If the new rule adjusts the peak period from 18:00 to 21:00 to 17:00 to 22:00 and adds an applicable explanation for high-voltage large industrial customers, then both the applicable period set and the customer category set change. If the clause adds a mandatory expression that time-of-use pricing should be implemented according to the new period, then the binding strength increases accordingly. The comprehensive change degree is calculated based on the changes in clause semantics, scope of application, and binding strength. When the comprehensive change degree reaches 0.36, the rule change is identified as a moderate rule change event, and the changed period field, customer scope field, and pricing constraint field are written into the rule change event, so that the rule change event enters the subsequent decision context construction process. During the construction of the intelligent agent's decision context, the system retrieves the customer's electricity bill collection event chain fragments for the past 90 days, work order event chain fragments for the past 30 days, and the latest effective time-of-use pricing rules, electricity bill collection policies, and customer credit evaluation rules based on business requests. Rules containing mandatory expressions such as "should," "must," and "must" are written into the hard constraint area, while rules containing flexible expressions such as "may," "suggest," and "encourage" are written into the reference constraint area. Differences in time periods and applicable customer scope between the old and new time-of-use pricing rules are written into the rule change notification area. Subsequently, a business situation profile is generated based on information such as the current period being the second half of the electricity bill settlement cycle, the initial effective period of the policy, an increase in consultation work orders in the distribution area, and a low continuous collection rate for the customer. The customer risk value is calculated by integrating the continuous arrears period, payment amount volatility, credit rating changes, and work order growth rate, and serves as the context input for the intelligent agent to generate candidate decision suggestions. In the context of dynamic decision-making, the agent generates multiple candidate decision suggestions. The first candidate decision suggestion is for the account manager to contact the customer within 24 hours to verify the payment plan and explain the new time-of-use electricity price implementation period. The second candidate decision suggestion is to raise the customer's electricity bill collection risk level and initiate supervisor review within the 7-day observation period. The third candidate decision suggestion is to immediately initiate the marketing audit procedure and issue a collection notice simultaneously. The fourth candidate decision suggestion is to include the customer in the area's complaint risk observation list and merge and analyze related consultation work orders in the same area. Each candidate decision suggestion includes at least the following fields: candidate suggestion identifier, business action type, operating subject, operating object, execution period, execution scope, list of referenced rule version identifier codes, and list of referenced event chain nodes. During the compliance verification process, semantic consistency verification is first performed on each candidate decision suggestion. For the first candidate decision suggestion, its core action phrases are identified as contacting customers, explaining electricity pricing rules, and verifying payment plans. These actions do not conflict with the prohibited phrases in the hard rules. For the third candidate decision suggestion, it is identified that it includes actions such as immediately initiating an audit procedure and simultaneously issuing a collection notice. Since the event chain has not yet formed the pre-existing anomaly confirmation node required for audit initiation, and the electricity fee collection rules require prior communication reminders and verification confirmation, this candidate decision suggestion is marked as lacking conditions. Subsequently, rule boundary verification is performed on the candidate decision suggestions, breaking them down into five elements: operating subject, operating object, operating time period, operating scope, and version based. If the first candidate decision suggestion is executed by the account manager, the operating object is the corresponding customer, the execution time is within the rule's effective period, the operating scope is only communication and verification, and the latest time-of-use electricity pricing rule version is referenced, then the rule confidence is high. If the third candidate decision suggestion does not reference the latest time-of-use electricity pricing rule version or does not meet the pre-existing audit conditions, then the backtracking and regeneration process begins. During the business impact verification process, the first and second candidate decision recommendations are mapped to nodes in the event chain, such as customer communication, payment confirmation, and risk review, and the process is extrapolated downstream along the event chain. If the communication and explanation are executed first, and then a decision on whether to raise the risk level is made based on customer feedback within 7 days, this execution order is consistent with the event chain topology and will not change the triggering conditions of strongly related edges. If the audit and collection are executed directly, it may lead to an increase in complaint orders before customers understand the new time-of-use pricing rules. Therefore, this candidate decision recommendation is marked as having a higher business impact risk. Risk; After considering semantic confidence, rule confidence, and impact confidence, the overall compliance confidence of the first candidate decision recommendation is 0.94, and it enters the approval set; the overall compliance confidence of the second candidate decision recommendation is 0.87, and it enters the minor revision process. After supplementing the 7-day observation period, limiting the review trigger conditions, and supplementing the latest rule version identifier code, it is re-verified and passes; the third candidate decision recommendation has missing conditions and a high risk of business impact, and it enters the retrospective regeneration process. The regenerated recommendation is adjusted to initiate an investigation project after the customer fails to pay as promised and the review confirms the abnormality. During the adjustment of early warning thresholds and conflict arbitration, the basic threshold for electricity fee collection risk is used as the starting point. The threshold adjustment factor is calculated based on business situation information such as the initial policy implementation period, daytime business hours, the second half of the monthly settlement cycle, and the increase in consultation work orders in the distribution area. Since this rule change is a moderate rule change event, drift compensation is superimposed during the 6-hour transition period, temporarily tightening the electricity fee collection risk threshold. If the customer's current collection rate is lower than the adjusted early warning trigger threshold, and the deviation multiple reaches the warning level range, an electricity fee collection risk warning level early warning conclusion is generated. For candidate decision suggestions such as customer manager contact communication, supervisor review, and complaint risk observation existing simultaneously under the same customer, conflict arbitration is conducted within a 30-minute statistical window, using the customer number and contract number as the aggregation object. Since customer manager contact communication and supervisor review do not have conflicting action directions but have a sequential execution relationship, customer communication is placed as the first execution node and supervisor review as the second execution node according to the event chain topology, with a 7-day observation waiting period inserted between them. For the complaint risk observation suggestion, it is retained as a parallel action for continuous monitoring of work order changes in the same distribution area. After conflict arbitration is completed, a decision record is generated. The decision record includes at least the following fields: customer number, contract number, final decision content, warning level, comprehensive compliance confidence level, list of referenced rule version identifiers, business situation profile, threshold adjustment factor, conflict arbitration trajectory, referenced event chain nodes, data credibility identifier, and generation time. A 64-character compliance watermark is generated based on the business object identifier, final decision content, list of rule version identifiers, and generation time, and this compliance watermark is written into the decision record. When business personnel subsequently conduct customer communication, payment review, or complaint risk observation, they can use this decision record to trace the rule version, event chain evidence, threshold adjustment process, and arbitration process referenced by the candidate decision suggestion. When a subsequent review finds that the watermark in the decision record is inconsistent with the recalculated result, a record integrity alarm is triggered. When manual review confirms that the decision is effective, the review result is used as a historical similar decision record and enters the subsequent threshold replay calibration and agent context construction process. Through the above specific operation process, the scattered electricity consumption, payment, work order, customer files and rule release data are organized into a traceable event chain. The dynamically changing time-of-use electricity pricing rules are transformed into rule change events and enter the decision context of the intelligent agent. The candidate decision suggestions generated by the intelligent agent enter the early warning and arbitration process after being verified by semantic consistency, rule boundaries and business impact. This transforms the electricity marketing risk early warning from static threshold triggering to real-time decision early warning driven by business situation, rule changes and compliance evidence chain.

[0026] It should be noted that this invention can be deployed on the device itself to realize embedded applications, or it can run on a PC or other terminal with a user interface, thereby meeting various hardware environments and usage requirements.

[0027] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any other combination. When implemented in software, the above embodiments can be implemented in whole or in part by a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions of the embodiments of this application are implemented in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted wirelessly or wiredly from one website, computer, server, or data center to another website, computer, server, or data center. Wired methods include optical fiber, twisted pair, coaxial cable, etc. Wireless methods include infrared, microwave, etc. Available media include any available media that can be accessed by a computer or data storage devices such as servers and data centers that contain one or more sets of available media. Available media can be magnetic media (floppy disks, hard disks, magnetic tapes), optical media (DVDs), or semiconductor media. Semiconductor media can be solid-state drives.

[0028] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0029] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A real-time decision-making and early warning method for an intelligent agent in electricity marketing, characterized in that, include: S1. Access the real-time data stream of electricity marketing and the source event stream of business rules, perform source identification annotation, semantic field alignment and business time alignment on the marketing data, and build an event chain oriented towards business objects; S2. Establish rule version identification codes for the accessed business rules and identify rule change events based on the semantics, scope of application, and strength of constraints of the clauses; S3. Based on the rule change event, retrieve the associated business rules, event chain and business status information, construct the agent decision context, and call the agent to generate candidate decision suggestions; S4. Perform compliance verification on candidate decision recommendations according to semantic consistency, rule boundaries, and business impact order, and backtrack and correct candidate decision recommendations that fail the compliance verification. S5. Adjust the warning trigger threshold based on business situation information and rule change events, conduct conflict arbitration on candidate decision suggestions that have passed compliance verification, and generate warning conclusions and decision records.

2. The real-time decision-making and early warning method for power marketing intelligent agents according to claim 1, characterized in that, S1 includes: Generate a source identifier for the accessed data, consisting of the data source type, sampling frequency, and semantic ontology version number; Field normalization is performed based on the business object thesaurus and field mapping table; Perform co-occurrence matching and pending confirmation processing on unmatched fields; Generate event chain nodes based on business objects according to the business event time correction results; Generate event chain association edges based on the preceding and following relationships of business rules and the stable relationships of historical events.

3. The real-time decision-making and early warning method for power marketing intelligent agents according to claim 1, characterized in that, S2 includes: Extract the clause text, rule source, clause number, rule effective date, applicable objects, and binding expressions from the rule data to be versioned; Semantic summary identifiers are generated through key term extraction, term weight ranking, and fixed-length hashing. A rule version identifier code is generated from the rule source type, clause number, version timestamp, semantic summary identifier, and effective status identifier; The current rule clauses are compared with the previous effective version in terms of semantics, scope of application, and binding strength. Corresponding events are marked for newly added clauses, repealed clauses, and rule source unreachable status.

4. The real-time decision-making and early warning method for power marketing intelligent agents according to claim 3, characterized in that, Semantic summary identifiers are generated through key term extraction, term weight ranking, and fixed-length hashing, including: When generating semantic summary identifiers, non-semantic information is cleaned up and sentences are segmented in the main text of the rule clauses; Extracting key business terms and constraint terms from a power marketing terminology glossary; Term weights are determined based on word frequency, inverse document frequency, and constraint term identifiers; Terms are selected according to their weights, and the selected terms, their clauses, clause numbers, and rule effective dates are concatenated into a summary input string. Perform fixed-length hashing on the digest input string.

5. The real-time decision-making and early warning method for an intelligent agent in power marketing according to claim 1, characterized in that, S3 includes: Parse the business object, business category, execution area, execution time period, and trigger source in the business request; Extract event chain segments based on business type; Search for rule terms by business category, customer category, region, time period, and effective status, and divide them into hard constraint area, reference constraint area, and rule change prompt area; Write the rule difference field, business situation profile, historical similar decision summary and data credibility tag into the decision context; The confidence level for generating candidate decision recommendations is determined based on rule citation completeness, event chain matching degree, business situation matching degree, and similarity to similar historical decisions.

6. The real-time decision-making and early warning method for an intelligent agent in power marketing according to claim 5, characterized in that, The confidence level for generating candidate decision recommendations is determined based on rule citation completeness, event chain matching, business situation matching, and similarity to historical similar decisions. This includes: The rule reference completeness is obtained by matching the valid rules corresponding to the current business request with the referenced valid rules. The event chain matching degree is obtained based on the correspondence between key event nodes and referenced event chain nodes; The business situation matching degree is obtained based on the consistency between the candidate disposal level and the risk level of the business situation profile. The similarity to historical decisions is obtained based on the semantic similarity between the candidate decision action and the historical similar decision action. The generation confidence score is obtained by weighting the rule reference completeness, event chain matching degree, business situation matching degree, and similarity of historical similar decisions.

7. The real-time decision-making and early warning method for an intelligent agent in power marketing according to claim 1, characterized in that, S4 includes: Extract action phrases and constraint objects from candidate decision suggestions; The candidate decision-making recommendations are broken down into the operating subject, operating object, operating time period, operating scope, and the version on which they are based; Semantic confidence is generated based on the results of phrase matching according to hard rules; The rule confidence score is generated based on the comparison results between the action elements and the rule-allowed set. Impact confidence is generated based on downstream relationships in the event chain and similarity to historical risk decisions; The overall compliance confidence level is adjusted by combining data credibility, and candidate decision recommendations are classified into sets, partial adjustment processes, or backtracking requests with deviation information and regenerated according to confidence level ranges.

8. The real-time decision-making and early warning method for an intelligent agent in power marketing according to claim 7, characterized in that, The candidate decision-making recommendations are broken down into the operating subject, operating object, operating time period, operating scope, and supporting version, including: Extract the execution role, applicable objects, execution time, action range, and rule basis identifiers from the candidate decision recommendations; Compare the execution roles with the permission sets; compare the applicable objects with the scope of the rules; Compare the execution time with the rule's effective period and processing time limit; Compare the action range with the monetary boundary, time limit boundary, threshold boundary, and processing level boundary respectively; Compare the rule reference identifier with the latest valid rule version identifier code.

9. The real-time decision-making and early warning method for an intelligent agent in power marketing according to claim 1, characterized in that, S5 includes: Threshold adjustment factors are generated based on time period factors, seasonal factors, policy window period factors, and special event factors. Adjust the basic early warning threshold according to the type of business indicator and add rule drift compensation; The warning level is determined based on the adjusted warning trigger threshold; Iterative arbitration of support levels is used to address action conflicts. For timing conflicts, the event chain is rearranged according to the topological order of the event chain and observation waiting periods are inserted; The final decision content, warning level, rule version identifier, threshold adjustment factor, and arbitration trajectory are written into the decision record.

10. The real-time decision-making and early warning method for an intelligent agent in power marketing according to claim 9, characterized in that, Adjusting the basic early warning threshold according to the type of business indicator and overlaying rule drift compensation, including: When adjusting the basic early warning thresholds, identify the risk direction of business indicators; For business indicators where the risk increases with the increase in indicator value, the trigger threshold is lowered according to the threshold adjustment factor. For business indicators where the risk increases when the indicator value decreases, the trigger threshold is increased according to the threshold adjustment factor. The drift compensation magnitude is determined based on the change level and the number of changed fields of the rule change event, and the adjusted warning trigger threshold is superimposed and corrected.

Images