Intelligent electricity bill interpretation and optimization system based on multi-source data fusion and rag

Abstract

The invention discloses an electric charge bill intelligent interpretation and optimization system based on multi-source data fusion and rag, and belongs to the technical field of energy consumption optimization. Comprising a multi-source data access and analysis module, an electricity price policy and charging rule knowledge base construction module, a multi-source data fusion and feature modeling module, a rate optimization inference module, a bill attribution analysis module, an intelligent explanation module, a front-end interaction display module and the like. Structured bill data, load time sequence features and unstructured electricity price policy document information are uniformly included in the same feature space, an electricity price policy knowledge base is used as a constraint source, and an explicit mapping relation from bottom layer data to upper layer rules is established. The rate optimization result strictly follows the current charging regulations in form and can be semantically attributed to specific terms, and a finer and more personalized rate and power utilization strategy scheme can be provided on the premise of ensuring compliance.

Description

Technical Field

[0001] This invention belongs to the field of energy optimization technology, specifically involving an intelligent interpretation and optimization system for electricity bills based on multi-source data fusion and rag. Background Technology

[0002] With the comprehensive deployment of smart meters, electricity consumption information collection systems, and marketing meter reading and billing systems in the distribution network, power companies can collect massive amounts of electricity consumption behavior data with minute-level or even second-level time resolution, and generate structured electricity bill data in the billing process. Meanwhile, my country's electricity pricing mechanism has undergone an evolutionary path over the past decade, from a "single electricity price" to a multi-dimensional coexistence of "time-of-use pricing—tiered pricing—power regulation fees—market-based pricing—comprehensive package pricing," resulting in significant heterogeneity and dynamism in billing rules across both time and space dimensions. However, existing electricity billing analysis and optimization still have the following shortcomings: 1. Fragmentation in data dimensions: Existing billing systems focus on structured electricity consumption, electricity charges, and archive data; load analysis systems focus on time-series load behavior; policy management systems focus on text rule extraction; and large model retrieval and enhancement generation systems focus on text question answering and document generation. There is still a lack of a unified data and knowledge representation framework that can model and reason about "bills-load-policy-optimization results" in the same semantic space.

[0003] 2. The singularity of decision-making objectives: Existing rate optimization and package recommendation methods often take "minimizing costs" or "maximizing benefits" as the sole objective, lacking comprehensive consideration of multiple objectives such as interpretability, compliance, and user cognitive load, and have not yet fully reflected the research trend of "explainable intelligent decision-making".

[0004] 3. Weak explanatory mechanism; Although existing methods can output a certain "optimal rate plan" or "optimization suggestion", most plans only provide numerical comparisons at the result level, without clearly connecting the complete causal chain of "electricity consumption behavior pattern - terms - cost changes - strategy suggestions" at the explanatory level, making it difficult for users to form stable behavioral changes and trust in the billing system.

[0005] 4. Lack of synergy between retrieval enhancement generation and optimization algorithms; existing retrieval enhancement generation applications mainly serve text question answering or document generation tasks, and are basically independent of the underlying constraint optimization model and simulation model. They have not yet formed a collaborative architecture that "uses optimization results as knowledge sources, clauses as constraints, and large models as interpretation interfaces". Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent interpretation and optimization system for electricity bills based on multi-source data fusion and rag, in order to solve the problems faced in the above-mentioned background technology.

[0007] The objective of this invention can be achieved through the following technical solutions: A smart electricity bill interpretation and optimization system based on multi-source data fusion and rag-and-drop functionality, comprising: The multi-source data access and parsing module is used to connect to multiple data interfaces, collect data information from multiple data interfaces, and perform unified parsing, standardization, and anomaly correction on the data information after preprocessing. The module for building a knowledge base of electricity pricing policies and billing rules is used to acquire electricity pricing-related information and, based on the analysis of this information, to build a knowledge base of electricity pricing policies and billing rules. The multi-source data fusion and feature modeling module is used to uniformly encode bill features, load time series features and clause features at the user granularity, and generate high-dimensional fusion feature vectors; The rate optimization inference module constructs a total cost objective function based on multiple rates and electricity consumption strategies, maps multiple features in the high-dimensional fused feature vector into constraints, and optimizes the total cost objective function based on the constraints. The billing attribution analysis module is used to analyze cost changes in both time series and billing structure dimensions, and to establish causal relationships between cost changes and load behavior and terms. The intelligent explanation module uses a knowledge base of electricity pricing policies and billing rules, as well as multi-source fusion features, as external knowledge sources. Through vector retrieval and large model generation mechanisms, it performs semantic understanding and generates answers to users' natural language questions. The front-end interactive display module is used to present the analysis results to users through a graphical and conversational interface.

[0008] Furthermore, the working method of the multi-source data access and parsing module is as follows: A billing cycle is set, and user data information is obtained within that cycle, including electricity consumption, electricity bill information, and electricity load information. The data is preprocessed, and an electricity consumption segmentation vector is constructed based on the processed electricity consumption information. Using a segmented vector of electrical charge A unified representation of electricity consumption across different segments is used; a cost vector is constructed based on the processed electricity bill information. Using cost vectors The system presents various user fees in a structured manner; based on the processed electricity load information, it constructs the user's electricity load value. Standardized electricity load values Perform load anomaly detection.

[0009] Furthermore, the working method of the knowledge base construction module for electricity pricing policies and billing rules is as follows: Electricity price-related information includes electricity price, electricity billing policy, and billing rule documents. After structured extraction of the electricity price, electricity billing policy, and billing rule documents, the clause text is obtained. semantic vectorization modeling of clause text The process is performed to obtain semantic vectors of the clauses. A knowledge base for electricity pricing policies and billing rules is constructed based on the obtained semantic vectors of each clause. The knowledge base stores the obtained electricity pricing-related information in the form of clause objects. The clause objects include at least: an applicable object field, a billing parameter field, an effective period or version field, and a conflict handling field. Incremental updates and version management are performed on the clause objects based on the version field.

[0010] Furthermore, the working method of the multi-source data fusion and feature modeling module is as follows: Statistical analysis is performed on billing fields in electricity billing information to calculate year-on-year and month-on-month electricity billing change rates. Simultaneously, user electricity billing information within the billing cycle is integrated to generate corresponding billing feature vectors. Load time-series characteristics include average load, maximum active load, load factor, and peak-to-valley ratio. Based on these characteristics, a corresponding load feature vector is generated. By using the semantic vectors of each clause in the electricity pricing policy and billing rules knowledge base, users are associated with the clauses, and association indicators are defined. Identify applicable clauses, construct a subset of applicable clauses, and build a policy feature vector based on the identified subset of applicable clauses. ;in , Terms and Conditions With users The applicable relationship between the terms indicates the quantity, when the terms Applicable to users The value is 1 if it is true, and 0 otherwise. bill feature vector Load characteristic vector and policy feature vectors By concatenating the features, a high-dimensional fused feature vector is obtained. .

[0011] Furthermore, the working method of the rate optimization inference module is as follows: Construct a total cost objective function with electricity consumption strategy as the independent variable and total cost as the dependent variable. , ,in, For electricity consumption strategy Next user During the billing cycle Total cost, For electricity consumption strategy Next user During the billing cycle The electricity bill portion, For electricity consumption strategy Next user During the billing cycle The basic electricity cost portion, For electricity consumption strategy Next user During the billing cycle The portion of the power-adjustable electricity fee, For users During the billing cycle The fixed cost portion within; Optimization includes single-objective function optimization and multi-objective function optimization. When it is single-objective function optimization, it is done by... optimization, The strategy variable vector is constructed for the electricity consumption strategy to be optimized. For the policy variable vector Next user In periodic sets The total cost objective function on the above, Represents the set of optimization cycles; When optimizing for a multi-objective function, through optimization, and Indicates the weighting coefficient. Let the total cost objective function for user u be , Let u be the risk objective function; The constraints are expressed as follows: , Indicates the first Constraint functions, This indicates the total number of constraints.

[0012] Furthermore, the bill attribution analysis module works as follows: The current electricity bill information is compared with the historical electricity bill information to identify cost changes. The cost changes are then broken down and combined with load time-series characteristics to identify electricity consumption behavior. Based on the mapping relationship between the electricity price policy and the billing rule knowledge base, a triplet set of cost change, behavior pattern and clause constraint is constructed. Cost changes include the amount of cost change. and cost change rate Cost Change Cost change rate In the formula, This represents the total cost for user u during billing period m. This represents the total cost for user u during the billing cycle m-1. Changes in costs Break it down into changes in electricity consumption. Electricity price change item ,in, , , The electricity price for user u at time t. For user u, the battery level at time t. For reference time The electricity price is below For reference time Electricity consumption.

[0013] Furthermore, the intelligent explanation module works as follows: The system acquires the user's natural language questions, which include the user's question field text and the user's trigger event field text. The user's natural language question field text is encoded to obtain a user text semantic vector. Based on the field text of each clause in the electricity pricing policy and billing rules knowledge base, multiple clause semantic vectors are constructed. The similarity score between the user text semantic vector and each clause semantic vector is calculated. Clauses corresponding to clause semantic vectors with similarity scores higher than a preset similarity threshold are selected as search results. Simultaneously, a large model prompt is constructed by combining the search results and structured data to generate response text containing clause references, numerical explanations, and behavioral suggestions.

[0014] Furthermore, the working method of the front-end interactive display module is as follows: The analysis results include breakdown of cost changes, rate comparisons, potential savings, and clause references. It also constructs a cost composition vector to show users a comparison of cost components.

[0015] Furthermore, when generating constraints, the rate optimization inference module performs consistency checks on the clause objects, maps only the clause objects that pass the consistency check to constraints, and triggers differential reconstruction of constraints when a version update is detected.

[0016] Furthermore, after generating the answer text, the intelligent explanation module performs a recalculation consistency check on the key numerical fields in the numerical description. If the consistency check fails, it triggers a secondary search and reconstructs the prompt words or reverts to a templated explanation output, while outputting the clause reference identifier and the numerical evidence identifier. If the consistency check passes, it is pushed to the front-end interactive display module.

[0017] The beneficial effects of this invention are: This invention integrates structured billing data, load time-series characteristics, and unstructured electricity pricing policy documents into the same feature space, and uses an electricity pricing policy knowledge base as a constraint source to establish an explicit mapping relationship between the underlying data and the upper-level rules. This ensures that the rate optimization results strictly follow the current billing regulations in form and can be attributed to specific clauses in semantics. It can provide more refined and personalized rate and electricity consumption strategy solutions while ensuring compliance. This invention introduces a billing attribution analysis module to decompose total cost changes into several explainable factors and establish a mapping between load behavior and terms. This allows for the generation of explanatory text containing a causal chain of "cost change - behavior pattern - terms constraint" during the interpretation phase. This structured interpretation method helps users build a stable understanding of the billing mechanism and also facilitates cross-validation by users with professional backgrounds. This invention utilizes a large model interpretation framework for retrieval-enhanced generation, injecting policy knowledge bases and optimization results as external knowledge into the generation process. This enables the system to maintain the traceability of clauses and the verifiability of values ​​when facing open-ended natural language questions. Compared with traditional template-based report generation, this invention has significant advantages in terms of interpretation flexibility, context adaptability, and interactivity, and can support multiple rounds of follow-up questions and in-depth consultation. This invention structures the clauses into clause objects that include applicable objects, pricing parameters, effective period / version information and priority. It also uniformly references the same version of the clauses in the three stages of constraint generation, bill attribution evidence construction and clause reference interpretation, so as to realize policy version consistency verification and traceability, and reduce the risk of cross-system rule drift and mixing of old and new policies. This invention performs consistency verification on key values ​​in the answer during the intelligent explanation output stage. If the verification fails, it triggers a secondary search or template rollback and outputs verifiable data fields, thereby reducing the risk of being misled by the illusion of a large model and increasing user trust.

[0018] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a system module block diagram of the present invention; Figure 2 This is a schematic diagram of the processing flow of the rate optimization inference module of the present invention; Figure 3 This is a schematic diagram of the processing flow of the bill attribution analysis module and the intelligent explanation generation module. Detailed Implementation

[0021] 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.

[0022] In one embodiment, a smart interpretation and optimization system for electricity bills based on multi-source data fusion and rag is disclosed, such as... Figure 1 As shown, the system mainly includes: The multi-source data access and parsing module is used to connect to multiple data interfaces, collect data information from multiple data interfaces, and perform unified parsing, standardization, and anomaly correction on the data information after preprocessing. The module for building a knowledge base of electricity pricing policies and billing rules is used to acquire electricity pricing-related information and, based on the analysis of this information, to build a knowledge base of electricity pricing policies and billing rules. The multi-source data fusion and feature modeling module is used to uniformly encode bill features, load time series features and clause features at the user granularity, and generate high-dimensional fusion feature vectors; The rate optimization inference module constructs a total cost objective function based on multiple rates and electricity consumption strategies, maps multiple features in the high-dimensional fused feature vector into constraints, and optimizes the total cost objective function based on the constraints. The billing attribution analysis module is used to analyze cost changes in both time series and billing structure dimensions, and to establish causal relationships between cost changes and load behavior and terms. The intelligent explanation module uses a knowledge base of electricity pricing policies and billing rules, as well as multi-source fusion features, as external knowledge sources. Through vector retrieval and large model generation mechanisms, it performs semantic understanding and generates answers to users' natural language questions. The front-end interactive display module is used to present the analysis results to users through a graphical and conversational interface.

[0023] Through the above scheme, this application constructs a bottom-up business chain. Specifically, the bottom layer uses a multi-source data access and parsing module to uniformly collect and standardize data information such as electricity consumption, electricity bills, electricity load, and user profile data, facilitating unified parsing and standardized processing of spatiotemporal data from different sources and with different granularities. The middle layer uses a knowledge base construction module for electricity pricing policies and billing rules, as well as a multi-source data fusion and feature modeling module, to map "bills-load-terms" into a unified computable representation. At the decision-making layer, a rate optimization inference module completes the simulation and solution of multiple rates and electricity consumption strategies under policy constraints, while a bill attribution analysis module decomposes cost changes into behavioral factors and term factors. Finally, at the service layer, an intelligent interpretation module collaborates with a front-end interactive display module to provide natural language bill interpretation, rate optimization suggestions, and visualized analysis results. Through the above layered and modular design, the organic coupling between data collection, knowledge modeling, optimization decision-making, and intelligent interpretation is achieved.

[0024] The multi-source data access and parsing module works as follows: A billing cycle is set, and user data information for that cycle is obtained, including electricity consumption, electricity bills, and electricity load. The data is preprocessed, and an electricity consumption segmentation vector is constructed based on the processed electricity consumption information. Using a segmented vector of electrical charge A unified representation of electricity consumption across different segments is used; a cost vector is constructed based on the processed electricity bill information. Using cost vectors The system presents various user fees in a structured manner; based on the processed electricity load information, it constructs the user's electricity load value. Standardized electricity load values Perform load anomaly detection.

[0025] In the above scheme, a billing cycle can be set independently, and user data information within the billing cycle can be obtained. The data information is first processed, for example, by setting hierarchical processing strategies for missing data, abnormal load data, and cross-cycle differences, including abnormal data correction, missing fragment imputation, and caliber alignment. The processed data is then used as one of the input features for subsequent multi-source data fusion and feature modeling modules. The data information includes the user's electricity consumption information, electricity bill information, and electricity load information. For example, in the user set, using... To represent a single user, use a set. Representing all users, in the time dimension, using To represent a time index, use a set. This represents a set of time indexes within a billing cycle. The multi-source data access and parsing module extracts the user's electricity consumption data, billing details, and necessary load curves for the current billing cycle from the metering master station, marketing system, and accounting system. It then performs time-series synchronization and cleaning, writing the data into the user-level data warehouse. The time-series electricity consumption measurements output by the smart meter or data acquisition system can be formalized as... , Indicates user At any moment The electricity consumption measurement values ​​are then used to construct a segmented electricity consumption vector based on the electricity consumption information. Using a segmented vector of electrical charge To uniformly express electricity consumption across different segments; for example, for users... and billing cycle A segmented vector of electrical charge can be defined: In the formula, For users During the billing cycle The power segment vector below, Indicates user During the billing cycle Inner Electricity consumption in segments (such as peak, off-peak, valley, or different tiers), Indicates the total number of segments. This approach, compared to existing technologies that often require customized fields and calculation logic for different peak and valley periods, time-sharing peak and valley periods, or stepped intervals in different regions, resulting in high system migration costs, utilizes segmented vectors... A unified representation of electricity consumption across different time zones allows for adaptation of billing discrepancies via vector dimensions rather than code branches, significantly reducing engineering complexity in environments with multiple policies and rates, and improving versatility for cross-regional deployment and expansion; furthermore, a cost vector is constructed based on electricity bill information. Using cost vectors Presenting various user fees in a structured manner. In the formula, For users During the billing cycle The cost vector below, Indicates user During the billing cycle The bill below Category fees (such as electricity charges, basic electricity charges, and additional charges). Indicates the total number of expense categories. This approach, compared to existing technologies where billing costs are often stored in loose key-value fields and the difficulty in maintaining consistency in itemization across detection, optimization, and display, structures costs into vectors. This allows for aligned calculation of itemized costs under the same caliber, direct participation in simulation and attribution, and seamless use in interpreting outputs. This enhances adaptability to complex billing structures and reduces misunderstandings and review costs caused by inconsistent calibers. Finally, based on electricity load information, user electricity load values ​​are constructed. Standardized electricity load values Perform load anomaly detection, such as analyzing raw time-series electrical measurements. After timestamp alignment and outlier processing, derived sequences such as daily curves, weekly curves, and billing cycle curves are generated. Anomaly detection can be represented by threshold rules. In the formula, For users At any moment The standardized electricity load value below; Indicates user Historical average load; Indicates user The historical standard deviation of electricity load, at this time when When the threshold is exceeded, it can be identified as an anomaly and corrected. This approach, compared to existing technologies that use fixed thresholds or empirical rules to handle spikes and missing load data, which are prone to false alarms and missed alarms across different users and seasons, and which can then affect cost simulation and optimization results, utilizes... and Constructing standardized indicators This enables anomaly detection and correction to be user-adaptive, thereby reducing the impact of data quality fluctuations on cost recalculation, strategy ranking, and cost-saving assessment, and improving the robustness and stability of recommendation results.

[0026] The working method of the knowledge base construction module for electricity pricing policies and billing rules is as follows: Electricity pricing-related information includes electricity prices, electricity pricing policies, and billing rules documents. After structured extraction of the electricity price, electricity pricing policy, and billing rules documents, the clause text is obtained. semantic vectorization modeling of clause text The process is performed to obtain semantic vectors of the clauses. A knowledge base for electricity pricing policies and billing rules is constructed based on the obtained semantic vectors of each clause. The knowledge base stores the obtained electricity pricing-related information in the form of clause objects. Each clause object includes at least the following fields: applicable object field, billing parameter field, effective period or version field, and conflict handling field. Incremental updates and version management are performed on the clause objects based on the version field.

[0027] In the above scheme, electricity price-related information, such as electricity price and electricity charge policies and billing rules documents issued by the National Development and Reform Commission, local governments, and power grid companies, is obtained and stored in the form of clause objects. These clause objects include at least: applicable object fields, pricing parameter fields, effective period or version fields, and conflict resolution fields. Incremental updates and version management are performed on the clause objects based on the version field. The electricity price, electricity charge policy, and billing rule document information is preprocessed, structured extracted, and semantically vectorized to obtain the clause text. semantic vectorization modeling of clause text The process is performed to obtain the semantic vector of the clause. For example, regarding a clause The corresponding text fragment is obtained as follows After structured extraction, it can be represented as In the formula, For the application of clause j, This is the set of pricing parameters for clause j (such as unit price, tiered threshold, etc.). This invention extracts the effective period or version information of clause j. Compared to existing technologies that store clauses as plain text or extract only pricing parameters while ignoring the applicable objects and effective period, leading to the risk of retrieval hits but no application or constraint construction lacking scope of application, this invention elevates the clause rules to structured, computable entities by extracting clauses as information triplets containing the applicable objects, pricing parameters, and effective period or version. This allows for simultaneous verification of the scope of application and invocation of corresponding parameters during retrieval, reference, and constraint construction optimization, improving policy adaptation accuracy and reducing compliance omissions and misuse risks. Furthermore, a semantic vector is obtained through an encoding function. ,in , Terms and Conditions semantic vectors, This refers to pre-trained language models or embedding models, which, compared to existing technologies that primarily rely on static rule tables, manual keyword retrieval, or are only geared towards clause-based questions and answers without being tied to the numerical scope of the bill, utilize the clause text. Encoded as semantic vectors This transforms policy clause knowledge from uncomputable natural language into a searchable, sortable, and reusable vector representation, allowing for backtracking to specific clause sources based on similarity during generative interpretation, thus significantly enhancing the accuracy and consistency of clause citations. Furthermore, a version-based dimension is introduced to achieve policy version management, such as clause text... The vector is obtained through the encoding function. And can be sorted by version form In the formula, Terms and Conditions In version The semantic vector below; This indicates the corresponding version of the clause text. In contrast to existing technologies where it is difficult to trace historical versions after knowledge base clauses are updated, and there is a tendency for interpretations to cite new clauses while calculations still use old parameters or vice versa, this solution explicitly introduces the version dimension in clause vector modeling. This enables the system to manage and trace policy changes in a versioned manner, referencing the version clauses that match the current billing cycle during the interpretation phase, and calling the same version parameters during the optimization phase. This significantly improves the consistency and auditability of calculation results and interpretation references in policy update scenarios.

[0028] The working method of the multi-source data fusion and feature modeling module is as follows: based on the billing fields in the electricity bill information, statistical analysis is performed to calculate the year-on-year change rate and month-on-month change rate of electricity bills. At the same time, the electricity bill information of users in the billing cycle is integrated to generate the corresponding bill feature vector. Load time-series characteristics include average load, maximum active load, load factor, and peak-to-valley ratio. Based on these characteristics, a corresponding load feature vector is generated. By using the semantic vectors of each clause in the electricity pricing policy and billing rules knowledge base, users are associated with the clauses, and association indicators are defined. Identify applicable clauses, construct a subset of applicable clauses, and build a policy feature vector based on the identified subset of applicable clauses. ;in , Terms and Conditions With users The applicable relationship between the terms indicates the quantity, when the terms Applicable to users The value is 1 if it is true and 0 otherwise; the bill feature vector is... Load characteristic vector and policy feature vectors By concatenating the features, a high-dimensional fused feature vector is obtained. .

[0029] In the above scheme, electricity bill information, load time-series characteristics, and clause characteristics are uniformly encoded at the user granularity to construct a high-dimensional feature representation that can be used for optimization and interpretation, such as for users. During the billing cycle load curve It allows defining load time-series characteristics such as average load, maximum active load, load factor, and peak-to-valley ratio, where average load is... The maximum load is The load factor is Peak-to-valley ratio is In the formula, Indicates user During the billing cycle Average load within, Indicates user During the billing cycle Maximum load within, Indicates the billing cycle Time index set, Indicates user During the billing cycle The load factor, Indicates user During the billing cycle Peak-to-valley ratio Indicates user During the billing cycle The minimum load within the range; it can also further define the electricity consumption and its proportion in three time periods: peak, flat, and valley, which are expressed as peak electricity consumption in order. Electricity consumption in flat sections Electricity consumption during off-peak hours , , , In the formula , , These represent the billing cycles. The peak, flat, and valley periods converge. Indicates period Total electricity consumption This represents the proportion of electricity consumption during peak hours; similar definitions apply to the proportions of electricity consumption during off-peak hours and valley hours, which will not be elaborated upon here. Based on these load time-series characteristics, corresponding load feature vectors are generated. Simultaneously, statistical analysis is performed based on the billing fields in the electricity bill information to calculate the year-on-year and month-on-month change rates of electricity costs. The calculation method is expressed as follows: as well as In the formula, Indicates user In the cycle The year-on-year change rate of electricity prices Indicates the month-on-month change rate. This represents the cost for the same period of the previous year; it also integrates the user's electricity bill information within the billing cycle, including numerical representations of fields such as time-of-use consumption, electricity cost per unit, basic electricity cost, and surcharges, to generate a corresponding bill feature vector. Then, by matching the electricity pricing policy with the terms in the billing rules knowledge base, the user is associated with the terms, and the associated indicator is defined. Identify applicable clauses, construct a subset of applicable clauses, and build a policy feature vector based on the identified subset of applicable clauses. For example, by matching the current user with specific terms in the knowledge base, an association indicator can be defined. In the formula Terms and Conditions With users The applicable relationship between the terms indicates the quantity, when the terms Applicable to users The value is 1 if the condition is met and 0 otherwise, thus obtaining a subset of applicable clauses and constructing a policy feature vector. Finally, the bill feature vector Load characteristic vector and policy feature vectors By concatenating the features, a high-dimensional fused feature vector is obtained. This provides computable input for optimizing inference and interpretation generation. Compared to existing technologies where multi-source data fusion is often used for single monitoring tasks such as electricity theft detection and anomaly alarms, or where rate optimization only utilizes load while the interpretation module only displays bills / terms, resulting in a disconnect between the two, this solution integrates bill fields at the user-period granularity. Load behavior characteristics Policy-side characteristics Unified coding is This allows simulation calculations, constraint optimization, and attribution interpretation to share the same feature base, thereby linking behavior, rules, and cost results in a unified model, improving the reproducibility of cost-saving assessments, and enabling interpretations to directly point to the specific behaviors and clauses that cause cost changes.

[0030] The working method of the rate optimization inference module is as follows: using electricity consumption strategy as the independent variable and total cost as the dependent variable, a total cost objective function is constructed. , ,in, For electricity consumption strategy Next user During the billing cycle Total cost, For electricity consumption strategy Next user During the billing cycle The electricity bill portion, For electricity consumption strategy Next user During the billing cycle The basic electricity cost portion, For electricity consumption strategy Next user During the billing cycle The portion of the power-adjustable electricity fee, For users During the billing cycle The fixed cost portion within; optimization includes single-objective function optimization and multi-objective function optimization. When it is a single-objective function optimization, it is achieved through... optimization, The strategy variable vector is constructed for the electricity consumption strategy to be optimized. For the policy variable vector Next user In periodic sets The total cost objective function on the above, Represents the set of optimization cycles; when optimizing a multi-objective function, it is done through... optimization, and Indicates the weighting coefficient. Let the total cost objective function for user u be , Let u be the risk objective function; the constraints are expressed as follows: , Indicates the first Constraint functions, This indicates the total number of constraints. When generating constraints, consistency checks are performed on the clause objects. Only clause objects that pass the consistency check are mapped as constraints, and differential reconstruction of constraints is triggered when a version update is detected.

[0031] In the above scheme, such as Figure 2 As shown, multi-source features Mapped to a set of decision variables, parameters, and constraints in a constrained optimization problem, this module simulates and solves for different tariff rates and electricity consumption strategies, outputting the optimal or near-optimal solution and its economic evaluation. In its implementation, this module is typically encapsulated as a backend microservice in the form of a tariff advisory service. On one hand, it calls historical or synthetic AMI load curve data to perform hourly energy cost calculations for candidate tariff schemes over the entire billing cycle. On the other hand, it constructs a total cost function by combining basic electricity charges, power adjustment charges, and other additional items. It also supports parallel simulations of multiple tariff schemes and user behavior adjustment schemes, generating a simulation result set containing cost results, savings, and scheme adoption statistics for each scheme. Specifically, it uses electricity consumption strategy as the independent variable and total cost as the dependent variable, under a single electricity consumption strategy... Below, user During the billing cycle The total cost can be written as ,in For electricity consumption strategy Next user During the billing cycle Total cost, For electricity consumption strategy Next user During the billing cycle The electricity bill portion, For electricity consumption strategy Next user During the billing cycle The basic electricity cost portion, For electricity consumption strategy Next user During the billing cycle The portion of the power-adjustable electricity fee, For users During the billing cycle The fixed cost portion; compared with the common practice in existing technologies that only give the total cost result or only perform local optimization around a single cost item, making it difficult to explain the source of changes in the sub-items, this solution will... Explicitly split into , , and fixed items and electricity consumption strategy Binding allows the system to recalculate and output item-by-item comparisons when simulating different tariff rates or electricity usage strategies. This provides auditable evidence of which cost changes, significantly enhancing the interpretability and reproducibility of the solution. Simultaneously, the tariff optimization inference module uses the minimum total cost within the billing cycle or forecast period, or the maximum user utility under cost constraints, as its basic objective function. It incorporates decision variables such as tariff structure selection, contract capacity adjustment, and adjustable load migration into a unified constrained optimization framework. Optimization includes both single-objective and multi-objective function optimization. When it is a single-objective function optimization, it uses... optimization, The strategy variable vector is constructed for the electricity consumption strategy to be optimized. For the electric strategy variable vector Next user In periodic sets The total cost objective function on the above, This approach represents a set of optimization periods. Compared to existing technologies that often rely on single-period experience-based recommendations or construct optimizations solely around a single objective like peak reduction, making it difficult to align with billing standards for period-by-period evaluation, this method formalizes the rate optimization objective as a set across periods. Minimizing the total cost enables the system to perform long-term simulations and comparisons of candidate strategies within a unified framework of historical and forecast periods, and outputs quantifiable and reproducible cost results, providing a consistent computational basis for subsequent interpretation and cost-saving proofs. When optimizing for multiple objective functions, a weighted summation form can be introduced, through... optimization, and Indicates the weighting coefficient. Let the total cost objective function be... The risk objective function, including weighting coefficients, can be set based on the expertise and experience of those in the field. Compared to existing technologies where rate optimization often focuses solely on minimizing electricity costs, neglecting user process constraints, comfort, and risk preferences, leading to implementation difficulties, this approach employs a multi-objective weighted sum, unifying cost and risk objectives into the evaluation function. This allows the system to explicitly balance acceptability and stability while ensuring energy savings, thereby improving the feasibility and adoption rate of the recommended strategy in real-world energy consumption scenarios. Finally, when generating constraints, consistency checks are performed on the clause objects. Only clause objects that pass the consistency check are mapped to constraints, and differential reconstruction of constraints is triggered when a version update is detected. The corresponding constraints can be expressed as: , Indicates the first Constraint functions, This represents the total number of constraints, including billing rules, contract capacity limits, and process load limits. Compared to existing technologies where constraints lack fixed thresholds or rule sets, making them difficult to dynamically adapt to user status and terms, and resulting in weak cross-scenario adaptability, this solution unifies the representation of constraints such as billing rules, contract capacity, and process load as follows: And introduce fusion features This allows constraints to be dynamically adjusted based on user profiles, load characteristics, and clause matching results, thereby improving the optimization solution's adaptability to diverse business scenarios, regional policy differences, and policy updates.

[0032] The bill attribution analysis module works as follows: It compares current electricity bill information with historical electricity bill information to identify cost changes. These cost changes are then broken down, and electricity consumption behavior is identified by combining load time-series characteristics. Based on the mapping relationship between electricity pricing policies and billing rules knowledge bases, a triplet set of cost change—behavioral pattern—term constraint is constructed. Here, cost change includes the amount of cost change. and cost change rate Cost Change Cost change rate In the formula, This represents the total cost for user u during billing period m. The total cost for user u during the billing cycle m-1; the change in cost. Break it down into changes in electricity consumption. Electricity price change item ,in, , , The electricity price for user u at time t. For user u, the battery level at time t. For reference time The electricity price is below For reference time Electricity consumption.

[0033] In the above scheme, such as Figure 3 As shown, when targeting explanatory objectives, the bill attribution analysis module compares the current bill with historical bills (such as the previous billing cycle or the same period last year) to identify bill items with significant cost changes, and further traces the underlying driving factors. Specifically, the bill attribution analysis module decomposes cost changes, allocating the total cost change to factors such as electricity consumption changes, electricity price changes, and policy adjustments. Simultaneously, it identifies significant changes in electricity consumption patterns (such as increased load during specific periods or seasonal air conditioning load increases) by combining load time-series characteristics. Based on this, through mapping the relationship with the knowledge base of electricity price policies and billing rules, it connects these behavioral changes with specific clauses, constructing a triplet set of "cost change—behavioral pattern—clause constraint," providing structured evidence for the retrieval enhancement generation model, enabling the subsequently generated explanatory text to achieve traceability and quantifiable verification of clauses. For example, for users... In two adjacent billing cycles and Cost variation can be defined. Cost change rate In the formula, This represents the total cost for user u during billing period m. This represents the total cost for user u during the billing cycle m-1. Compared to existing technologies that use fixed thresholds or rely solely on high / low amounts for coarse-grained alerts, making it difficult to align comparisons across user scales and fee rates, this solution uses... and This system provides absolute and relative measures of cost differences, enabling comparability across different electricity price levels and consumption scales. It also provides a unified numerical standard for subsequent attribution analysis, strategy selection, and proactive trigger threshold configuration, thereby improving the stability and consistency of anomaly identification and optimization opportunity discovery. To facilitate attribution analysis, the cost variation can be... Break it down into changes in electricity consumption. Electricity price change item ,in, , , The electricity price for user u at time t. For user u, the battery level at time t. For reference time The electricity price is below For reference time The method involves measuring electricity consumption; compared to existing technologies that often focus on cost increases / decreases or merely provide correlation explanations without distinguishing whether the cost difference is due to changes in user electricity consumption behavior or changes in electricity prices / rates, this solution... Explicitly broken down into changes in electricity consumption. Electricity price change item This enables the system to quantitatively distinguish the contributions of behavioral factors and policy / rate factors, and accordingly provide targeted load migration or rate switching recommendations, thereby enhancing the causal clarity and operability of attribution explanations; and simultaneously, for a given candidate rate scheme Compared with the benchmark scheme Period can be defined The amount saved is In the formula, Indicates the period The amount saved within, This represents the total electricity cost under the base rate plan. This represents the total electricity cost under the candidate tariff scheme. Compared to existing technologies where recommended schemes often only provide estimated savings without a unified and verifiable quantitative standard, this scheme explicitly defines the savings amount of candidate schemes relative to the benchmark scheme. This allows different candidate tariffs and behavioral strategies to be ranked and screened on the same scale, and also enables... As core evidence for explanation, it enhances the verifiability of optimization suggestions and user trust; the bill attribution analysis module analyzes the differences between the current bill and historical bills. When performing factor decomposition, identify the main cost drivers and correlate them with load behavior characteristics and the set of terms. Establish a mapping to form a triplet set of "cost changes - behavior patterns - terms and conditions"; when a user raises a question or the system detects a certain triggering condition (such as... Exceeding the threshold or When the value is greater than a certain expected value, an indicative quantity can be used. This indicates whether an explanation or alarm is triggered; in the formula, An indicator quantity that indicates whether an explanation and alarm are triggered, for example, if The value is 1 if the condition is met, and 0 otherwise. This indicates a preset threshold.

[0034] The intelligent explanation module works as follows: It acquires the user's natural language question, which includes the user's question field text and the user's trigger event field text. The user's natural language question field text is encoded to obtain a user text semantic vector. Based on the field text of each clause in the electricity pricing policy and billing rules knowledge base, multiple clause semantic vectors are constructed. The similarity score between the user text semantic vector and each clause semantic vector is calculated. Clauses corresponding to clause semantic vectors with similarity scores higher than a preset similarity threshold are selected as search results. Simultaneously, a large model prompt is constructed by combining the search results and structured data, generating an answer text containing clause citations, numerical explanations, and behavioral suggestions. After generating the answer text, the intelligent explanation module performs a recalculation consistency check on the key numerical fields in the numerical explanation. If the consistency check fails, a secondary search is triggered, and prompt words are reconstructed or the output reverts to a templated explanation, while also outputting clause citation identifiers and numerical evidence identifiers. When the consistency check passes, it is pushed to the front-end interactive display module.

[0035] In the above scheme, such as Figure 3 As shown, the intelligent interpretation module is a retrieval-enhanced generation-driven intelligent interpretation module. It first obtains the field text of the user's natural language question, including the user's question field text and the user's trigger event field text. Then, it encodes the field text of the user's natural language question to obtain a user text semantic vector. Based on the field text of each clause in the electricity pricing policy and billing rules knowledge base, it constructs multiple clause semantic vectors; for example, through... The user text semantic vector is obtained, where, Indicates user User text semantic vectors, The text field representing the user's natural language question. This represents the problem encoding function; in the knowledge base of electricity pricing policies and billing rules, based on the field text of each clause, a clause semantic vector is constructed for each clause. For each clause's semantic vector The matching score is calculated, specifically the similarity score between the user's text semantic vector and the semantic vectors of each clause. In the formula, The similarity score between the question vector and the clause vector is represented by a normalized weight, and clauses corresponding to the semantic vectors of clauses with similarity scores higher than a preset similarity threshold are selected as search results. Compared with existing technologies that rely on manual templates, fixed question classification, or keyword matching to trigger explanation logic, resulting in poor adaptability to colloquial and complex questions, this solution uses users' natural language questions. Encoded as semantic vectors This enables the system to understand intent based on semantic similarity rather than literal similarity and to directly match it with clause vectors, thereby improving its coverage and adaptability in complex policy Q&A, bill discrepancy follow-up questions, and multi-turn dialogue scenarios; it also provides and Matching score This allows for quantifiable ranking criteria in the search, enabling clause screening and noise reduction before generating answers, thus reducing the probability of irrelevant clauses appearing in the answer and improving the relevance and compliance consistency of cited clauses; it employs... To assign weights, in the formula, Terms and Conditions In users Weights under the current problem This represents the set of candidate clause indexes. Compared to existing techniques that often only provide a list of hit clauses when multiple clauses coexist without explaining the relative strength of each clause's impact, leading to a lack of focus in interpretation, this approach... Normalization to weights This enables the system to form a stable contribution allocation under multiple constraints or interpretations, and allows for... As the strength of citation follows the output of the response, this transforms the citation of specific clauses, the reasons for prioritizing them, and the strength of the citations into a verifiable chain of evidence, enhancing the traceability and auditability of the explanation; for example, when the aforementioned indicators... After an alarm occurs, the intelligent interpretation module encodes the problem, calls vector retrieval to select the most relevant clause fragment from the electricity pricing policy and billing rules knowledge base, and extracts the corresponding structured data from the bill attribution analysis module. and , , The system constructs a large model to provide prompts based on numerical values, generating response text containing clause references, numerical explanations, and behavioral suggestions. After generating the response text, it performs a recalculation-based consistency check on key numerical fields in the numerical explanations. If the consistency check fails, it triggers a secondary search and reconstructs the prompt words or reverts to a templated explanation output, while also outputting clause reference identifiers and numerical evidence identifiers. If the consistency check passes, it is pushed to the front-end interactive display module. In this way, during the intelligent explanation output stage, consistency verification is performed on key numerical values ​​in the response (including cost breakdowns, savings, and comparison conclusions). If the verification fails, a secondary search or templated reversion is triggered, and verifiable evidence fields are output, thereby reducing the risk of being misled by the illusion of a large model and increasing user trust.

[0036] The front-end interactive display module works as follows: the analysis results include information on cost changes, rate comparisons, potential savings, and clause references. At the same time, it constructs a cost composition ratio vector to display the cost composition comparison to the user.

[0037] In the above solution, the front-end interactive display module is mainly responsible for the interactive presentation with users. It uses graphical and conversational interfaces to display information such as bill breakdown, rate plan comparison, savings potential, and clause references. For example, it generates bill breakdown views, rate plan comparison views, savings potential views, and a natural language Q&A interface for users to view. This, combined with the visual output and the text explanations generated by the large model, helps users with higher professional skills to examine and verify the model's behavior. Simultaneously, for a given strategy... The system can construct proportional vectors with different cost compositions. To showcase, In the formula, Indicating in strategy The solution presents a vector representing the proportion of each cost component to the total electricity bill. This differs from existing technologies that merely display the total amount or list items in tables, making comparisons across different periods and tariff schemes difficult. By normalizing cost components into proportional vectors, the solution allows for intuitive comparisons across periods and schemes, and can be linked to attribution results. This transforms the abstract conclusions about cost changes into structured differences that users can understand, improving the consistency and efficiency of explanation. Furthermore, in terms of deployment, a cloud-edge-device collaborative architecture can be prioritized. The cloud hosts the electricity pricing policy knowledge base, large-scale model inference services, and the global optimization engine; edge nodes handle high-frequency load data preprocessing and local simulation; and the device focuses on interaction and lightweight display, achieving a balance between scalability, real-time performance, and resource utilization. This approach ensures both explanation quality and optimization accuracy while considering response time and computing costs in scenarios with large-scale concurrent user access.

[0038] It should be noted that the content of electricity bill information in this invention is not limited to traditional residential or industrial electricity bills, but can also extend to multi-energy flow integrated bills issued by integrated energy service providers, and even settlement bills in new business scenarios such as virtual power plants and shared energy storage; electricity price related information includes not only official electricity price and electricity fee policy documents issued by the government and power grid companies, but also regulatory documents with pricing effect or binding force, such as market transaction rules, ancillary service compensation methods, and demand response incentive rules; rate optimization includes not only selecting between different electricity price types, but also the joint optimization of multi-dimensional decision variables such as contract capacity, electricity consumption time structure, adjustable load response strategy, and market participation strategy.

[0039] For example, in urban residential users using a parallel mechanism of tiered and peak-valley electricity pricing, single-phase smart meters are configured on the user side, connected to the electricity information collection system via low-voltage centralized meter reading devices, with a collection granularity of 15 minutes. At the data access layer, the system periodically retrieves monthly electricity bill data from users (including time-of-use electricity consumption, electricity charges, basic electricity charges, tiered pricing information, power regulation charges (if applicable), and additional fees) and load curves at 15-minute resolution within the billing cycle through API interfaces with the marketing system and the main collection station. After performing consistency and outlier checks, the data is written to the user-level data warehouse. At the knowledge base construction layer, the system selects policy documents related to residential electricity pricing (such as the "Tiered Pricing for Residential Electricity Consumption"). Based on the "Implementation Rules" and "Implementation Measures for Residential Peak-Valley Time-of-Use Electricity Pricing," the system performs structural analysis and clause vectorization to construct a sub-knowledge base for residential scenarios, highlighting key aspects such as tiered intervals, peak-valley time period divisions, applicable user categories, and pricing formulas. At the feature modeling layer, the system calculates the electricity consumption and its proportion during peak, off-peak, and valley periods based on the user's load curve, and calculates indicators such as the maximum / minimum load ratio and load factor. At the billing dimension, it calculates the differences in total electricity bills and individual charges compared to the previous billing cycle and the same period last year. At the policy level, it determines the current electricity price type and applicable clause subset based on the user's profile. These features are then concatenated to form a residential user-cycle-level feature vector. At the rate optimization layer, the system constructs the following candidate... Strategies: Implement only tiered pricing, without peak-valley time-of-use pricing; superimpose peak-valley pricing on the existing tiered structure; under the premise of superimposed peak-valley pricing, assume a certain proportion of adjustable loads (such as washing machines, dishwashers, and electric water heaters) migrate from peak to valley periods; for each candidate strategy, the system reallocates time-series loads and recalculates electricity charges while keeping total electricity consumption constant, obtaining cost predictions and savings for the next billing cycle or year; at the attribution analysis layer, when a significant increase in current electricity charges compared to the previous period is detected, the system compares the bills and load characteristics of the two periods, identifies peak-valley electricity charges as the main incremental item, and infers a significant increase in air conditioning load during peak periods by combining weather and electricity consumption behavior data; simultaneously, it confirms the use of tiered pricing terms. The user has already crossed a certain threshold; from this, the system draws a combined driving conclusion of "increased air conditioning load + tiered pricing plus peak pricing"; in the interpretation generation layer, when a user asks through the App, "Why did my peak electricity bill increase so much this month?", the system inputs the question and feature vector into the retrieval module, retrieves clauses related to tiered pricing and peak-valley pricing from the knowledge base, and extracts numerical features such as "peak electricity consumption increased by 30% year-on-year" and "this month's electricity consumption exceeded the first tier by 20%", constructs prompts and sends them to the retrieval enhancement generation model; the answer generated by the model will clearly explain the driving mechanism in natural language and give a quantitative suggestion of "if the adjustable load is moved to the valley period, it is expected to reduce the electricity bill by X yuan per month".

[0040] For example, in 110kV large industrial users using a hybrid billing mechanism of "basic electricity fee + time-of-use pricing + market-based pricing," the user side is equipped with multi-functional electricity meters and an enterprise energy management system (EMS), connected to the electricity information collection system and the power market trading platform via dedicated lines or VPNs. At the data access layer, in addition to acquiring regular electricity consumption, electricity fees, and billing data, the system also needs to obtain the transaction volume and price of users participating in market-based transactions from the trading platform, thus completely reconstructing the cost structure of "basic electricity fee + electricity consumption fee + transaction fee + power regulation fee + other adjustment items." At the knowledge base construction layer, the system performs structured modeling of documents such as the "Implementation Rules for Electricity Prices and Fees for Large Industrial Users in a Certain Province," "Transmission and Distribution Prices and Line Loss Assessment Methods," and "Market-based Transaction Settlement Rules," paying particular attention to the basic electricity fee billing method selection rules, demand / capacity price calculation formulas, contract maximum demand deviation assessment, and transaction volume allocation mechanisms. At the feature modeling layer, the system combines load curves from many years with billing data... The system constructs annual load distribution, maximum demand variation trends, contract capacity utilization, and historical cost comparison features under different billing methods. Simultaneously, it generates optional basic electricity billing methods and contract capacity adjustment ranges based on current policy constraints. At the rate optimization layer, the system aims to minimize total annual costs, designing decision variables such as contract capacity, demand billing methods, and market transaction proportions to construct a mixed-integer optimization problem, solving for the optimal solution while satisfying production process and power factor constraints. At the interpretation and generation layer, when a user asks, "What risks will there be if we lower the maximum demand contract value?", the system retrieves relevant clauses (contract maximum demand deviation assessment, change rules, etc.) from the knowledge base and identifies historical moments exceeding the proposed new contract value based on historical load peak distribution. This allows the system to explain potential penalties and default risks in natural language and quantifies the cost-benefit trade-off between "reducing contract capacity—reducing basic electricity fees—potentially generating over-demand assessments."

[0041] In scenarios focusing on large-scale policy changes, such as adjustments to peak-valley divisions, changes to tiered thresholds, or the introduction of new market-based tariff rates, when a new electricity pricing policy document is detected, the knowledge base construction module automatically triggers a version comparison process. This process matches the old and new clauses, identifies changes to key fields such as billing formulas, tiered intervals, and peak-valley time periods, and generates a "clause-level change difference set." Subsequently, the system filters out affected user groups based on user profiles and electricity consumption characteristics (e.g., users at specific voltage levels or those subject to specific electricity pricing categories), and then optimizes the rates by comparing the old and new versions of the old rules. Under the "new rules," cost simulations were conducted to compare the cost differences over several billing cycles under the two rules. For users whose changes exceed a set threshold, the intelligent explanation module generates personalized notification text based on a retrieval-enhanced generation model, detailing the marginal impact of the policy change on their electricity bill composition, and providing "buffering strategies" in conjunction with potential electricity consumption behavior adjustment suggestions. For example: "Peak electricity prices have increased slightly, but off-peak discounts have been added; without adjusting electricity consumption behavior, the monthly electricity bill is expected to increase by approximately X yuan; if peak hours are appropriately shifted during non-critical production periods as suggested by the system, the impact can be controlled within Y yuan."

[0042] Without departing from the technical concept and principle of this invention, those skilled in the art can make various modifications or substitutions to the above embodiments, such as changing the specific optimization algorithm, replacing the large model implementation framework, adjusting the feature extraction method, or changing the knowledge base implementation technology, etc., and these modifications or substitutions should all be considered to fall within the protection scope of this invention.

[0043] The above description is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined in the claims, they should all fall within the protection scope of the present invention.

Claims

1. A smart electricity bill interpretation and optimization system based on multi-source data fusion and rag, characterized in that, The system includes: The multi-source data access and parsing module is used to connect to multiple data interfaces, collect data information from multiple data interfaces, and perform unified parsing, standardization, and anomaly correction on the data information after preprocessing. The knowledge base construction module for electricity pricing policies and billing rules is used to acquire electricity price-related information and, based on the analysis of this information, construct a knowledge base for electricity pricing policies and billing rules. The multi-source data fusion and feature modeling module is used to uniformly encode bill features, load time series features and clause features at the user granularity, and generate high-dimensional fusion feature vectors; The rate optimization inference module constructs a total cost objective function based on multiple rates and electricity consumption strategies, maps multiple features in the high-dimensional fused feature vector into constraints, and optimizes the total cost objective function based on the constraints. The billing attribution analysis module is used to analyze cost changes in both time series and billing structure dimensions, and to establish causal relationships between cost changes and load behavior and terms. The intelligent explanation module is used to perform semantic understanding and generate answers to users' natural language questions by using a knowledge base of electricity pricing policies and billing rules, as well as multi-source fusion features as external knowledge sources, and through vector retrieval and large model generation mechanisms. The front-end interactive display module is used to present the analysis results to users through a graphical and conversational interface.

2. The intelligent interpretation and optimization system for electricity bills based on multi-source data fusion and rag as described in claim 1, characterized in that, The working method of the multi-source data access and parsing module is as follows: A billing cycle is set, and user data information is obtained within that cycle, including electricity consumption, electricity bill information, and electricity load information. The data is preprocessed, and an electricity consumption segmentation vector is constructed based on the processed electricity consumption information. Using a segmented vector of electrical charge A unified representation of electricity consumption across different segments is used; a cost vector is constructed based on the processed electricity bill information. Using cost vectors The system presents various user fees in a structured manner; based on the processed electricity load information, it constructs the user's electricity load value. Standardized electricity load values Perform load anomaly detection.

3. The intelligent interpretation and optimization system for electricity bills based on multi-source data fusion and rag as described in claim 2, characterized in that, The working method of the knowledge base construction module for electricity pricing policies and billing rules is as follows: Electricity price-related information includes electricity price, electricity billing policy, and billing rule documents. After structured extraction of the electricity price, electricity billing policy, and billing rule documents, the clause text is obtained. semantic vectorization modeling of clause text The process is performed to obtain semantic vectors of the clauses, and a knowledge base of electricity pricing policies and billing rules is constructed based on the obtained semantic vectors of each clause. The electricity pricing policy and billing rules knowledge base stores the acquired electricity pricing-related information in the form of clause objects. The clause objects include: applicable object field, billing parameter field, effective period or version field, and conflict handling field. Incremental updates and version management are performed on the clause objects based on the version field.

4. The intelligent interpretation and optimization system for electricity bills based on multi-source data fusion and rag as described in claim 3, characterized in that, The working method of the multi-source data fusion and feature modeling module is as follows: Statistical analysis is performed on billing fields in electricity billing information to calculate year-on-year and month-on-month electricity billing change rates. Simultaneously, user electricity billing information within the billing cycle is integrated to generate corresponding billing feature vectors. Load time-series characteristics include average load, maximum active load, load factor, and peak-to-valley ratio. Based on these characteristics, a corresponding load feature vector is generated. By using the semantic vectors of each clause in the electricity pricing policy and billing rules knowledge base, users are associated with the clauses, and association indicators are defined. Identify applicable clauses, construct a subset of applicable clauses, and build a policy feature vector based on the identified subset of applicable clauses. ;in , Terms and Conditions With users The applicable relationship between the terms indicates the quantity, when the terms Applicable to users The value is 1 if it is true, and 0 otherwise. bill feature vector Load characteristic vector and policy feature vectors By concatenating the features, a high-dimensional fused feature vector is obtained. .

5. The intelligent interpretation and optimization system for electricity bills based on multi-source data fusion and rag as described in claim 4, characterized in that, The working method of the rate optimization inference module is as follows: Construct a total cost objective function with electricity consumption strategy as the independent variable and total cost as the dependent variable. , ,in, Electricity consumption strategy Next user During the billing cycle Total cost, For electricity consumption strategy Next user During the billing cycle The electricity bill portion, For electricity consumption strategy Next user During the billing cycle The basic electricity cost portion, For electricity consumption strategy Next user During the billing cycle The portion of the power-adjustable electricity fee, For users During the billing cycle The fixed cost portion within; Optimization includes single-objective function optimization and multi-objective function optimization. When it is single-objective function optimization, it is done by... optimization, The strategy variable vector is constructed for the electricity consumption strategy to be optimized. For the policy variable vector Next user In periodic sets The total cost objective function on the above, Represents the set of optimization cycles; When optimizing for a multi-objective function, through optimization, and Indicates the weighting coefficient. Let the total cost objective function for user u be , Let u be the risk objective function; The constraints are expressed as follows: , Indicates the first A constraint function, This indicates the total number of constraints.

6. The intelligent interpretation and optimization system for electricity bills based on multi-source data fusion and rag as described in claim 5, characterized in that, The working method of the bill attribution analysis module is as follows: The current electricity bill information is compared with the historical electricity bill information to identify cost changes. The cost changes are then broken down and combined with load time-series characteristics to identify electricity consumption behavior. Based on the mapping relationship between the electricity price policy and the billing rule knowledge base, a triplet set of cost change, behavior pattern and clause constraint is constructed. Cost changes include the amount of cost change. and cost change rate Cost Change Cost change rate In the formula, This represents the total cost for user u during billing period m. This represents the total cost for user u during the billing cycle m-1. Changes in costs Break it down into changes in electricity consumption. Electricity price change item ,in, , , The electricity price for user u at time t. For user u, the battery level at time t. For reference time The electricity price is below For reference time Electricity consumption.

7. The intelligent interpretation and optimization system for electricity bills based on multi-source data fusion and rag as described in claim 6, characterized in that, The working method of the intelligent explanation module is as follows: The system acquires the user's natural language questions, which include the user's question field text and the user's trigger event field text. The user's natural language question field text is encoded to obtain a user text semantic vector. Based on the field text of each clause in the electricity pricing policy and billing rules knowledge base, multiple clause semantic vectors are constructed. The similarity score between the user text semantic vector and each clause semantic vector is calculated. Clauses corresponding to clause semantic vectors with similarity scores higher than a preset similarity threshold are selected as search results. Simultaneously, a large model prompt is constructed by combining the search results and structured data to generate response text containing clause references, numerical explanations, and behavioral suggestions.

8. The intelligent interpretation and optimization system for electricity bills based on multi-source data fusion and rag as described in claim 7, characterized in that, The working method of the front-end interactive display module is as follows: The analysis results include breakdown of cost changes, rate comparisons, potential savings, and clause references. It also constructs a cost composition vector to show users a comparison of cost components.

9. The intelligent interpretation and optimization system for electricity bills based on multi-source data fusion and rag as described in claim 8, characterized in that, When generating constraints, the rate optimization inference module performs consistency checks on the clause objects, maps only the clause objects that pass the consistency check to constraints, and triggers differential reconstruction of constraints when a version update is detected.

10. The intelligent interpretation and optimization system for electricity bills based on multi-source data fusion and rag as described in claim 9, characterized in that, After generating the answer text, the intelligent explanation module performs a recalculation consistency check on the key numerical fields in the numerical description. If the consistency check fails, it triggers a secondary search and reconstructs the prompt words or reverts to a templated explanation output. At the same time, it outputs the clause reference identifier and the numerical evidence identifier. When the consistency check passes, it is pushed to the front-end interactive display module.

Images