Carbon emission data intelligent analysis method

Abstract

The invention discloses a carbon emission data intelligent analysis method, particularly relates to the technical field of carbon emission data intelligent analysis, and is used for solving the problem of how to improve the data analysis efficiency in bill line projects with free Chinese texts, missing attributes and mixed dimensions on the premise of not transforming an existing business system. The row-level explainable emission factor unique mapping is realized, and the problems of region, year, version and caliber are locked. Bill lines are connected into a chain, namely, unified receiving segmentation, Chinese element normalization and dimension alignment, version positioning and locking according to an account period and a delivery place, unique factors based on attributes and time-space judgment, original text solidification, conversion and other evidences, and hierarchical recheck and regular recharge are triggered by an uncertainty threshold. Therefore, line-level unique mapping, auditing, reproducible and traceable performance, remarkable voltage drop mismatching, evidence completeness and recalculation consistency close to full score, recheck focusing on high-risk items, and parallel check of a place method and a market method are realized in a steady state, so that calibers are naturally paired.

Description

Technical Field

[0001] This invention relates to the field of intelligent carbon emission data analysis technology, specifically to intelligent carbon emission data analysis methods. Background Technology

[0002] Currently, enterprise carbon accounting largely relies on a pipeline of "exporting from business systems + general OCR / extraction + caliber rule engine + report aggregation". Data such as purchase invoices, inventory lists, and electricity meter readings are typically extracted into tabular fields first, and then emission factors are selected and evidence is supplemented during the reporting or auditing stage. Emission factor libraries generally provide tabular data by region and year, with a few systems supporting basic retrieval and alternative factor selection. In electricity consumption scenarios, the location method and market method calculations are mostly performed in parallel at the aggregation level, with the two calibers rarely maintained simultaneously at the row level. Evidence materials (original invoice fragments, extraction instructions, conversion basis, factor sources) are mostly archived afterward as attachments or notes, and version changes and caliber switching usually lack fine-grained traceability. Overall, existing solutions are more focused on reporting and compliance presentation, and do not adequately support the "auditability and reproducibility of row-level data processing".

[0003] In genuine invoices, descriptions of goods are primarily in free Chinese text, with numerous aliases and abbreviations, inconsistent supplier coding standards, and frequent omissions or ambiguities in key attributes such as material, specifications, and measurement standards. Units include cross-level conversions such as "box, roll, kilogram, ton," and density, specific gravity, and temperature / pressure conditions are inconsistent. Scanned versions are of varying quality and their positioning is unstable. When these line items are stably mapped to a "unique emission factor (including region, year, version, and caliber label)," existing processes are prone to problems such as difficulty in choosing among multiple candidates, insufficient basis for version locking, and opaque downgrade / alternative paths. In case of disputes, it is often difficult to review at the line level "why this factor was chosen and what the conversion basis is." Regarding dual-caliber electricity consumption, existing systems mostly perform parallel accounting at the aggregation end, lacking parallel and consistency checks on the same line of records, resulting in insufficient comparability and reconciliation between entities and between accounting periods. Furthermore, row-level evidence is usually not "born and coexist" with the results. Extraction traces, conversion trajectories, factor sources and version paths lack structured traces. Manual review relies heavily on experience-based allocation and lacks a diversion mechanism based on uncertainty. Monthly batch review is costly and difficult to reproduce.

[0004] Based on the above situation, a key issue remains for project implementation: without modifying existing business systems, how to develop a feasible processing method at the row level that allows each bill line, after unifying Chinese elements and aligning dimensions, to be stably and interpretably mapped to a unique emission factor, while simultaneously solidifying an auditable chain of evidence (including original text location, extraction traces, unit conversion paths, factor sources, and version locking / downgrade paths), and implementing row-level parallelism and consistency verification of the location method and market method in electricity consumption scenarios, as well as triggering tiered review and rule-based updates based on uncertainty indicators. Existing publicly available solutions generally lack: row-level version locking and downgrade path recording, evidence chain objectification and storage in the same domain as the results, step-by-step trajectory and conditional definitions for unit conversion, row-level dual-caliber parallel consistency governance, and uncertainty-driven diversion and reproducible parameter thresholds for batch monthly settlements. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent carbon emission data analysis method. By linking the receiving data segmentation → element normalization and dimensional alignment → spatiotemporal positioning and version locking → unique factor determination → evidence chain solidification and uncertainty diversion into a row-level method chain, each row is stably mapped to a unique emission factor, and the results are traceable, reproducible, and auditable, thereby solving the problems mentioned in the background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an intelligent analysis method for carbon emission data, comprising: S1. Receive business data and segment it by line, mark the billing period, subject and delivery location, retain the original text fragments and location index, and form a line-level unit to be processed. S2. Extract elements from Chinese cargo descriptions, unify names and attributes based on alias ledgers, unify units of measurement, and generate unit conversion trajectories. S3. Load the emission factor index containing regional, annual, version and caliber labels, parse the spatiotemporal information based on the payment period and delivery location, screen out candidate factors and complete version locking; S4. Calculate the matching score based on attribute fit, spatiotemporal consistency and version priority, output a unique emission factor, and generate a review task and record the judgment path if uniqueness is not achieved. S5. Calculate row-level emissions using activity data and a unique factor, and simultaneously generate location-based and market-based results in parallel on the same row, perform dimension and caliber consistency checks and set markers; S6. Generate an evidence chain object for each line, record text fragments, extraction traces, unit conversion trajectories, emission factor sources and version locking paths, construct uncertain quantification indicators, trigger a review when the threshold is exceeded, and write the results back to the alias ledger and rule snapshot.

[0007] In a preferred embodiment, S1 includes: A unified channel is established on the receiving side to stably access invoices from the financial system and energy consumption ledger, receiving them according to fixed and incremental rhythms. A one-to-one mapping is established between ticket numbers and subject codes, with unified time and unit, and duplicates are removed based on source priority and arrival order; Missing payment terms are made up according to the invoice month. Delivery location is mapped in the order of receiving address, contract address, and invoice address. If mapping fails, a temporary processing mark is set. The text is segmented line by line according to the format, and original fragments and location indexes are generated. The billing period, subject, and delivery location are added. The fragments are packaged into line-level unprocessed units containing batch number, line number, and content summary, and delivered to the message queue with an idempotent identifier composed of batch number and line number.

[0008] In a preferred embodiment, S2 includes: The elements of the Chinese product description are extracted and normalized, and the brand, material, specifications, shape, model and unit of measurement are grouped into standard names and attribute groups using an alias ledger; When the normalization is difficult to determine, the context fields are called sequentially based on adjacent rows of the same invoice, the same supplier in the same batch, and the contract material strip. Units of measurement are unified to the standard units. Unit conversion trajectories are generated step by step according to the conversion table. The source and scope are recorded and the trajectory number is written back to the line. For cases where the model and unit are inconsistent, the material and shape are inconsistent, the specifications are missing and the context is insufficient, a gray mark is set and a checklist is generated; The alias ledger and conversion table have versions and effective ranges, and their history is unified and not rolled back; When multiple materials are listed side-by-side, they are split into sub-items by a separator while maintaining their order and position index; The normalization results are stored in the normalization region for factor selection and version locking.

[0009] In a preferred embodiment, S3 includes: When loading the emission factor index, the index includes the region identifier, year, version number, caliber label, source identifier, and source verification fingerprint; Based on the payment term mapping year and the delivery location mapping region and zone, first determine the region and year, then filter candidate factors and lock the latest valid version that has not been withdrawn; When an annual entry is missing, only the most recent year is downgraded and the locked path is recorded. When a region entry is missing, the system will attempt to match it item by item according to the order of adjacent partitions in the filing and record the locking level. If no match is found, the approved alternative will be used and the responsibility information will be written. Write the candidate list along with the version lock path into the factor reference area, and backfill the list number and path summary into the corresponding row-level record.

[0010] In a preferred embodiment, S4 includes: After version locking is completed, attribute anchor synonym merging, unit benchmark verification and spatiotemporal pre-check are performed on each record to eliminate candidates that are inconsistent with the payment period and delivery location. The remaining candidates are matched based on a combination of attribute fit, spatiotemporal consistency, and version priority according to preset weights to obtain a matching score. When the difference between the highest and second-highest matching scores reaches a preset threshold, a unique emission factor is determined, and a decision path containing participating fields, scores, weights, reasons for removal, and version information is generated. If the threshold is not reached, a review task is generated and the candidate list, weights, thresholds and rule versions are frozen, and timestamps are added to ensure consistency when rerunning with the same version. The unique emission factor number and determination path are written into the factor decision area and backfilled to the row-level record.

[0011] In a preferred embodiment, S5 includes: At the row level, based on activity volume and unique emission factor, both location-based and market-based results are generated on the same row and labeled with caliber labels. The market coverage is allocated according to the contract ledger in the order of site mapping, time period mapping, and calendar equalization. Uncovered residual amounts are accounted for using the same factor caliber as the location method, and the allocation path and caliber label are recorded in this line; Cross-dimensional superposition is prohibited within the industry; the unit conversion trajectory that has been saved shall be used as the standard for unit uniformity.

[0012] In a preferred embodiment, a consistency check is performed on the location-based approach and the market-based approach results for the same entity and the same payment period; In case of inconsistency, the attribution shall be determined according to the priority order: insufficient contract coverage, insufficient number of vouchers, inconsistent regional mapping, and abnormal measurement conversion. A pending status shall be generated in the bank and the attribution, allocation path and evidence fragment shall be written. Green electricity certificates are verified for uniqueness by number and can only be used by a single entity. All results and tags are stored in the database with batch number and row number as idempotent keys and can be used for subsequent evidence chain generation.

[0013] In a preferred embodiment, S6 includes: Immediately after the row-level results are generated, an evidence chain object is created for that row. The object includes the original text fragment and location index, feature extraction traces, unit conversion trajectory, emission factor source and version locking path, and the signature and timestamp of the person in charge. Generate a content summary and write it to an increment-only log; any changes are appended to the new version while retaining the original version. Set viewing and modification permissions and only allow supplementary entries; The chain of evidence and the results are persisted in the same domain and partitioned by batch and date, allowing for playback and location; Batch number and line number are used as idempotent identifiers for transmission and access.

[0014] In a preferred embodiment, uncertain quantitative indicators are constructed based on the evidence chain object, including activity volume, emission factor, unit conversion, and caliber selection, and thresholds and priorities are set. When any uncertain quantification exceeds its threshold, a review task is automatically generated and the main cause is highlighted. The current candidates and parameters are frozen. The review conclusion is written back to the alias ledger and rule snapshot and the version, coverage and effective time are recorded. It is then automatically adopted in subsequent similar rows. Review tasks are processed according to grade and time limits and can be upgraded.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. By integrating bill line items onto the same method chain—first, uniformly receiving and segmenting; then, unifying and aligning Chinese elements with dimensions; then, performing spatiotemporal positioning and locking the version based on payment period and delivery location; then, determining a unique factor based on attribute fit and spatiotemporal consistency; and finally, solidifying the original text fragments, location indexes, extraction traces, unit conversion trajectories, and factor source paths into evidence chain objects, and driving hierarchical review and rule reinjection with uncertainty thresholds—this addresses line-level pain points such as high noise in free Chinese text, missing attributes, inconsistent definitions, and difficulty in reviewing factor selection. It achieves the goal of "each line being stably mapped to a unique emission factor that is auditable, reproducible, and traceable": significantly reduced line-level mismatches (target less than one percent), evidence chain field completeness and result reproducibility approaching 100%, human-machine collaboration focusing review on high-uncertainty stages, significantly reducing review time (target no less than half), parallel implementation of the electricity usage scenario location method and market method at the line level with consistency verification, and natural alignment of report definitions.

[0016] 2. By embedding engineering governance measures such as idempotency identification, at least-once-arrival semantics, backpressure and elastic scaling, hot caching and shard loading, version locking and orderly degradation, whitelisting and stop-loss measures, sampling and acceptance criteria, and rule snapshots and effective intervals into the method chain, operational risks such as unstable latency in large-scale processing, version drift, tamperable evidence, and deviations in cross-currency and cross-boundary criteria are resolved, achieving the goal of "stable operation at scale + long-term compliance and auditability"; end-to-end latency is controllable within a predetermined window (reception to encapsulation not exceeding ten minutes, batch criterion verification not exceeding fifteen minutes), and online parallel capabilities meet monthly settlement and peak demand requirements. In scenarios involving sharding and scaling up of throughput from thousands to hundreds of thousands per minute, the evidence chain is enhanced with enhanced traceability and long-term retention to ensure readily available audit replays. Version withdrawals, annual omissions, and regional omissions all have clearly defined classification and rollback paths. In terms of consistency, the differences between the electricity location method and the market method are addressed by subject-based tiered thresholds, with a target pass rate of no less than 98% and a dimensional alignment accuracy rate of no less than 99.5%. Overall, compliance and audit requirements are transformed into an executable "evidence object + threshold + backfeed closed loop," which not only stabilizes on-site indicators but also provides a replicable foundation for future expansion to more categories and subjects. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the intelligent carbon emission data analysis method of the present invention. Detailed Implementation

[0018] 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.

[0019] Example: Figure 1 A flowchart illustrating the intelligent carbon emission data analysis method of the present invention is provided. The intelligent carbon emission data analysis method includes: S1. Receive business data and segment it by line, mark the billing period, subject and delivery location, retain the original text fragments and location index, and form a line-level unit to be processed. S2. Extract elements from Chinese cargo descriptions, unify names and attributes based on alias ledgers, unify units of measurement, and generate unit conversion trajectories. S3. Load the emission factor index containing regional, annual, version and caliber labels, parse the spatiotemporal information based on the payment period and delivery location, screen out candidate factors and complete version locking; S4. Calculate the matching score based on attribute fit, spatiotemporal consistency and version priority, output a unique emission factor, and generate a review task and record the judgment path if uniqueness is not achieved. S5. Calculate row-level emissions using activity data and a unique factor, and simultaneously generate location-based and market-based results in parallel on the same row, perform dimension and caliber consistency checks and set markers; S6. Generate an evidence chain object for each line, record text fragments, extraction traces, unit conversion trajectories, emission factor sources and version locking paths, construct uncertain quantification indicators, trigger a review when the threshold is exceeded, and write the results back to the alias ledger and rule snapshot.

[0020] The technical connections and implementation logic of the six steps are as follows: First, in S1, the invoice is cut into rows, and each row is labeled with the payment period, subject, and delivery location. The original text fragments and layout positions are also stored and packaged into a "row-level pending unit," which is then placed into a queue using the batch number and row number. In S2, the row is retrieved based on this identifier. The Chinese description of goods is restored to the standard name and attribute group based on the alias ledger. At the same time, the unit is unified according to the benchmark dimension, and a traceable conversion trajectory is generated. These two results are written back to the same row. With the payment period and delivery location, S3 performs spatiotemporal positioning in the emission factor index, filters out candidates from the same region and year, and records version locking and any downgrade attempts as a "locked path." In S4, the attribute group, conversion trajectory, and candidate list of the row are read. The system scores and selects one of the following criteria: attribute fit, spatiotemporal consistency, and version priority. This generates a unique emission factor and a replayable "judgment path." If the difference is not significant, a review task is generated and the parameters for that step are frozen. In step S5, after aligning the unique factor with the activity level, two results—the location method and the market method—are generated in parallel within the row. A difference threshold check is performed on the same entity and the same accounting period; if the threshold is crossed, it is marked for investigation. Finally, step S6 merges the original text fragments, extraction traces, unit conversion trajectories, factor sources, and version locking / judgment paths accumulated from the first five steps into an "evidence chain object." Simultaneously, four uncertain quantitative indicators—activity level, factor, conversion, and caliber—are calculated. If the threshold is exceeded, a review is pushed, and the review conclusion is written back to the alias ledger and rule snapshot. The entire process is driven by the idempotent key of "batch number + row number," with the status and identifier of the previous step serving as the entry point for the next. Each step only adds traces without overwriting history, ensuring unique row-level mapping and comparability between the two calibers, while also allowing evidence and rules to continuously converge in a closed loop.

[0021] S1. Receive business data and segment it by line, marking the payment period, subject, and delivery location, retaining the original text fragments and location indexes, forming line-level units to be processed. The specific implementation is as follows: A unified data access channel is established on the receiving side. Without altering the existing business system's rhythm, invoice-related information is stably accessed from the financial system and energy consumption ledger, employing a parallel access rhythm of daily batches and hourly increments. Each invoice specifies the required fields: invoice number, invoice date, payment period, entity name and unified code, delivery location, goods description, quantity and unit, amount, and tax. All units adhere to the enterprise's standard system (kilograms for mass, cubic meters for volume, kilowatt-hours for electricity, and pieces for quantity). The data source is limited to the financial system and energy consumption ledger. Invoice dates are allowed a maximum difference of one day, and quantity and amount are allowed an input error within one percent; errors exceeding this range are flagged. The format and value requirements for key fields are clearly defined: the invoice number consists only of uppercase letters and numbers, with a maximum length of thirty digits; the unified code for the entity is based on the unified social credit code, with a length of eighteen digits; the payment period uses the Gregorian calendar year-month format; the delivery location uses a standard place name and includes the administrative division code; and the amount and tax are retained to the cent.

[0022] After receiving a batch of documents, the system first maps the invoice numbers to the entity codes, unifying the time to Beijing time. Units are merged according to equivalence. Duplicate records are retained only once, prioritizing the source over the arrival time (the financial system has a higher priority than the energy consumption ledger; in cross-source conflicts, the higher priority and newer time take precedence). If the payment period is missing, it is supplemented once according to the month of the invoice date. If the delivery location is missing, it is mapped once in the order of "receiving address → contract address → invoice address". If the mapping still fails, the process is temporarily suspended and "delivery location missing" is written. To avoid overload or starvation, the receiving process adopts two parallel rules: "triggering at the top of the hour" and "triggering when 100 lines have accumulated". The first to arrive is processed first. A single batch does not exceed 100,000 lines, and the threshold is adaptively adjusted monthly based on the percentile statistics of the business distribution over the past three months. If the arrival rate drops below the historical lower limit within a ten-minute observation window, it automatically switches to low-frequency mode (receiving a batch every thirty minutes). After two consecutive observation windows, it recovers to above the historical median and then returns to normal frequency.

[0023] Each time a batch is processed, a batch list (including batch number, source tag, number of rows, and time range) is generated. At the same time, the original content is completely saved as an "original image". The original image is stored in UTF-8 character set and line by line in JSON format. The length of a single line does not exceed 10,000 characters. If it exceeds this, it is split and a mapping table is saved. The mapping table is kept for at least 90 days. If it is an electronic invoice with an unstructured attachment, it is archived as is and a batch number is added. If it is a scanned document, the layout is located and the characters are recognized. Only when the layout stability is not less than 95% and the character recognition accuracy is not less than 97% is it included in the process. Otherwise, the entire order is temporarily suspended and manually entered.

[0024] Then, the document is segmented line by line according to the invoice format, retaining the description text fragment of each line and its position index in the original format. The position index is represented by a coordinate system of "page number count + character offset". Cross-page merging will generate a mapping table simultaneously. After segmentation, the payment period, subject, and delivery location are immediately added to the line entries: if the payment period is missing, it is filled in with the invoice month, and if the contract specifies the delivery month, the contract shall prevail; the subject is identified by a unified code, and aliases are merged first; if the three sources of delivery location are inconsistent, the receiving address shall prevail, and the conflict shall be recorded in the notes of the batch list. After completing these steps, the row number, batch number, payment period, subject, delivery location, goods description text, quantity and unit, amount, text fragment and location index are encapsulated into a row-level processing unit and written to the row-level storage area. At the same time, "system source + batch number + row number + content hash" is delivered to the message queue as an idempotent identifier, and the downstream uses this identifier to pull the message. The message channel uses this idempotent identifier on both the receiving and delivering sides to prevent duplicate consumption. Duplicate events are recorded as deduplication logs, and the deduplication window is fixed at seven days.

[0025] The entire process, from receiving the data to completing the encapsulation, normally takes no more than ten minutes. At least ten batches are processed in parallel at the same time. The queue backpressure threshold is set to more than twenty batches to be processed, an average waiting time of more than fifteen minutes, or memory utilization of more than 80%. If any of these conditions are met, the queue is expanded by adding two execution instances each time. If two consecutive observation windows are below the threshold, the queue is recycled. If any step fails, it is retried three times at intervals of one, three, and five minutes. Batches that still fail are sent directly to the manual channel and automatic re-triggering is prohibited until manual confirmation and recovery.

[0026] Time is uniformly recorded using Beijing time. Cross-border invoices will retain both the original time zone and the conversion timestamp. Daylight saving time will be implemented according to local rules on the conversion date. Amounts are recorded by default including tax, with both the tax rate and currency written in. When conversion is involved, the exchange rate will be the closing value of the company's financial master database on that day, with the version number locked. The "one percent error" judgment is based on the relative error calculated using the original precision. Lines exceeding the threshold are marked "precision questionable" and added to the manual verification list. Page numbers and character offsets in the location index must be replayable and locating. If playback fails, a mark will be immediately set, and the error must be corrected before public disclosure.

[0027] All abnormal situations throughout the process are uniformly assigned a cause code. The cause codes are fixed into six categories: missing subject, missing delivery location, duplicate number, date out of bounds, questionable accuracy, and unstable format, with reserved extension space. Each record is recorded with the first discovery time and the person in charge. An alarm will be triggered immediately if the percentage of a single batch that is temporarily suspended exceeds 5%. At the same time, if any indicator of the percentage of temporarily suspended processing, back pressure, or latency exceeds the threshold for two consecutive observation windows, it will be escalated to the duty manager through both enterprise instant messaging and email channels until the processing loop is closed.

[0028] All batch lists, reasons for delayed processing, alarms, and manual processing records must be retained for at least two years. Searching by batch number, line number, source system, reason code, handler, and timestamp is supported. Access follows the minimum necessary permissions, and access is audited. To ensure auditability, at least 10,000 lines are sampled monthly using stratified random sampling (stratified by source system, invoice type, and amount range). The sampled data is evaluated for coverage, field completeness, deduplication accuracy, and end-to-end latency compliance. A rectification list is generated within ten working days of the sampling results and fed back into the rule base. The coverage target is no less than 99%, the field completeness target no less than 95%, the deduplication accuracy target no less than 99.5%, and the end-to-end latency compliance target no less than 95%.

[0029] In practical applications, consider the following scenario: One day, the system receives a 40-line steel invoice along with the monthly electricity bill. The system merges these two into a single batch and enters the receiving process. Following the deduplication, merging, and completion rules, the data is rendered as an original image. The system then segments the data line by line, recording text fragments and character positions. After expanding the payment period, subject, and delivery location, it is packaged into over 40 line-level units and delivered downstream. One line lacking a delivery location is supplemented according to the contract ledger and marked "Completed Once." Two other lines are temporarily suspended due to scan quality falling below the threshold, awaiting manual entry. The entire batch processing time is within ten minutes, with one elastic expansion triggered when the queue reaches a high threshold. After the batch is completed, it participates in monthly spot checks, and all indicators are above the red line.

[0030] Up to this point, the three main lines of fields, time, and region have been aligned from the moment of receipt, and the original evidence and location information have been completely preserved. If subsequent name unification, factor filtering, or uniqueness determination requires backtracking, the entire process, from the batch list to the original mirror and then to the row-level location index, can be traced back along the identifiers. This allows for quick location and correction of problems within a small scope, ensuring timeliness while also meeting the requirements of traceability, reproducibility, and verifiability.

[0031] S2. Extract elements from the Chinese cargo description, unify names and attributes based on the alias ledger, standardize units of measurement, and generate unit conversion trajectories. The specific implementation is as follows: In the process of extracting elements from Chinese product descriptions and completing the normalization of names and attributes, this step aims to stably restore the colloquial, abbreviated, and even typo-containing descriptions in first-line bills into comparable and verifiable standard items, facilitating subsequent factor screening, version locking, and evidence chain generation. The scope of application covers three types of entries: materials, equipment, and services. The text is received on-site at the same rhythm as the upstream bank-level records, and six types of elements are fixedly collected: brand, material, specification, shape, model, and measurement unit. The unit should always fall within the enterprise's benchmark system (kilogram for mass, cubic meter for volume, kilowatt-hour for electricity, and piece for quantity). The sources mainly include financial document texts, supplier product catalogs, and enterprise's common directories, and are cross-checked with historical purchase records. Single-line keyword fields are allowed to be missing once, but consecutive missing is not allowed.

[0032] Perform normalization before going live: unify the coding, clean non-printable characters and extra spaces, handle the mixed use of full-width and half-width characters, make equivalent replacements for common oral expressions (for example, regard "φ two" as "two millimeters in diameter" and "stainless steel 316" as "316 stainless steel"), and align the dimension with the unit word list. The unit conversion table fixedly records the conversion starting point and ending point, reference source, conversion accuracy and tolerance, applicable product categories. The accuracy is generally reserved to three decimal places, and the tolerance is defaulted to one percent. Prefer to adopt the supplier's specification sheet and contract agreement. For the general conversion without evidence, confirm it through sampling weighing and reconciliation record. After one confirmation, write it into the conversion table and indicate the sample source.

[0033] The normalization action is carried out according to the "alias ledger": The alias ledger is used to merge colloquial Chinese product descriptions into a collection of entries of standard names and attribute groups, including alias, standard name, material / specification key points, common units, version number, effective period, and source record. The ledger is maintained weekly. The fields fixedly include alias, standard name, material key points, specification key points, shape category, common units, measurement benchmark, common misspellings, scope of application, source record, version number, reviewer, and effective time. Any new addition or modification is only used for on-site parsing after the effective time. The in-stock lines before the effective time are not recalculated to avoid historical result drift. When parsing on-site, the ledger is preferentially hit; if the hit is insufficient, use the upper and lower text fields to联动补强判断 (the specific translation of this term needs to be adjusted according to the actual situation) in the order of "adjacent lines in the same bill → same batch and same supplier → contract material item" to make a judgment. If the clues at any level are not sufficient to form a stable conclusion, stop immediately and do not make cross-level guesses.

[0034] To ensure pacing, a one-second limit is set for single-line parsing. An interpretable threshold is set to measure normalization reliability, using a percentage scale ranging from 60% to 85%. The initial value is taken from the quantiles of stable samples over the past quarter, and is automatically fine-tuned using rolling samples on the first working day of each month. If a supplier's writing style changes abruptly within the same batch, a local adjustment is performed on the judgment threshold for that batch without altering the global threshold. This adjustment only takes effect within that batch, and reverts to the monthly baseline after the batch ends. If three consecutive lines fail to reach the threshold, a manual assistance and rule convergence task is automatically triggered, generating a list of supplementary terms and example sentences.

[0035] Unit conversions require step-by-step recording of the path and basis, such as "box → roll → kilogram," with each step clearly stating the source, precision, and tolerance. These paths, along with standard names and attribute groups, are stored in a unified area, and the trajectory number and attribute group are written back to the row-level record for direct reference later, eliminating the need for repeated reasoning. Throughput boundaries are clearly defined: no less than 3,000 lines per minute per machine; automatic expansion occurs when backlog exceeds the threshold; single-line parsing failures can be retried twice, and those still failing to meet the standard are added to the to-do list.

[0036] When obvious conflicts occur, they are uniformly grayed out. Graying out is divided into three categories: contradiction between model and unit, contradiction between material and shape, and missing specifications with insufficient context. For each category, the trigger words, the location of occurrence, and suggested supplementary items are recorded. When the proportion of grayed-out items in a single batch exceeds 5%, the system automatically generates a rule adjustment work order, which includes a detailed list of new aliases, corrected rules, and supplementary conversion items. After review, a new version is formed. Issues found in monthly spot checks are included in the monthly convergence task. The generated new version takes effect on the first day of the following month and the effective time is written in. On-site switching is based on this.

[0037] The normalization results are stored in the normalization area using key-value pairs. Fields include the standard name, attribute group, and unit conversion trajectory number. Quick retrieval by batch number and row number is supported. The unit conversion trajectory unifies activity volume into a step-by-step conversion path to the baseline unit, recording the source / target unit, basis, precision and tolerance, confirmer, and time. Downstream users can retrieve the standard item and trajectory number for a row using only the batch number and row number, avoiding duplicate parsing. Service items use lightweight normalization, extracting three fixed items: service type, billing unit, and billing quantity, aligned with the unit thesaurus. When multiple units are listed in the same row, the measurement baseline is used as the primary unit, and other units are converted step-by-step according to the conversion table order, with the complete path recorded. If the path is broken at an intermediate node, the conversion stops and is marked as pending confirmation.

[0038] To avoid discrepancies in terminology, when the same statement appears in multiple available records in an alias ledger, the supplier-specific entry takes precedence over the industry-general entry, and the industry-general entry takes precedence over the enterprise-general entry. The entry with the most recent effective date takes precedence. If priority and effective date are the same, the source credibility is used to determine the order of credibility: contract appendix, supplier specifications, and historical reconciliation. When cross-dimensional conversions involving length and mass, volume and mass, etc., the material density and specific gravity library is used. Density values ​​are subdivided by material and specification, and the source and temperature conditions are noted. When density values ​​are missing, they are downgraded in the order of "historical weighing of the same material and specification → interpolation of adjacent specifications of the same material → publicly available industry reference values," and the downgrade level is marked in the conversion path.

[0039] When encountering multiple parallel materials (separated by commas, semicolons, Chinese commas, or hyphens), first break them down into sub-items according to the separator, then perform normalization and conversion separately, retaining the original sequence number and position index, and adding a merge mark at the end of the line for traceability. All numerical notation must be standardized to the company standard: Chinese numerals, full-width numerals, numerals with thousands separators, and multipliers must be uniformly converted to Arabic numerals and base units. If obvious errors are found (e.g., excessive decimal places, logical inconsistencies between quantity and unit), immediately suspend processing and record correction suggestions, then rewrite after manual confirmation.

[0040] The alias ledger and unit conversion table support obsolescence markers and effective ranges. Obsolescence does not affect historical records that have been normalized before taking effect. Any addition, modification, or obsolescence generates a change entry and retains the reviewer, time, and effective scope, with a minimum retention period of five years. It also clearly defines operational hard boundaries: the density and specific gravity database has a default temperature of 25 degrees Celsius, a default error limit of 5%, and an update frequency of once a month. New values ​​only apply to newly added rows after taking effect. The alias ledger and unit conversion table are published at fixed times: Tuesday at midnight and the first day of each month at midnight. Both have obsolescence markers and effective ranges, and historically normalized records are not rolled back. The priority of separators for splitting multiple materials is, in order: comma, semicolon, Chinese comma, hyphen. Service items are fixedly enumerated as installation, maintenance, testing, and transportation. Billing units are based on times, man-hours, sets, and kilometers. Supplier-defined units are converted according to the conversion table path and recorded. This approach ensures that the on-site personnel clearly know which information to use, the standard to read it, how to clean and normalize it, when to submit it for review, where to store the normalized records, and how to use it directly downstream. It also clearly defines the boundaries of threshold range, update rhythm, version activation, gray dictionary, and multi-unit and cross-dimensional boundaries, allowing peers to implement it independently and reproduce it stably.

[0041] S3. Load the emission factor index containing regional, year, version, and caliber labels; parse the spatiotemporal information based on payment terms and delivery locations; filter candidate factors and complete version locking; the specific implementation is as follows: When loading emission factor indexes containing regional, annual, version, and caliber labels, the usable scope is first compressed to the correct spatiotemporal coordinates, and then the specific version is completely locked to facilitate clean implementation of subsequent row-level unique determination. The index is maintained as a set of "regional identifier, year, version number, caliber label, source identifier, and source verification fingerprint." The source is simultaneously connected to the enterprise factor database and the authoritative release database, and is normally synchronized annually or quarterly. To ensure clear evolution, old and new versions are retained in parallel and their inheritance relationship is clearly stated. Regional names are unified to the national administrative region standard, and then the regional identifier is mapped according to the power grid zoning table. Common aliases are first merged into the same entry. Emission factors refer to the coefficients that convert activity quantities into greenhouse gas emissions, and have attributes such as dimensions, caliber labels, regional identifiers, and years. In this application, emission factors are measured in "emissions in benchmark units such as mass / energy / electricity," and caliber labels are limited to location method, market method, supplier measured, industry average, and expenditure method; each factor is accompanied by a source identifier, version number, and source verification fingerprint. The location-based approach refers to the method of calculating energy emissions based on the geographical area of ​​energy supply and its corresponding year's regional average emission factor, without considering contract reductions or green certificate deductions. In this application, the location-based approach results are consistent with the geographic mapping and year-based locking. The market-based approach refers to the method of calculating emissions based on the factors corresponding to the remaining electricity volume after deducting activity volume based on contract terms and market-based instruments such as green electricity / certificates; contract coverage and document uniqueness must be recorded at the row level. Candidate factors are the set of factors with the same coordinates selected from the factor index after mapping the payment period to the year and the delivery location to the geographic / regional area.

[0042] Before deployment, a full deduplication and source identifier integrity check are performed. During runtime, a dual mechanism of event triggering and scheduled inspection is employed: if the external library is updated, the hot cache is immediately invalidated and rebuilt; otherwise, the hot cache remains valid for two hours, automatically refreshing upon expiration, which can be shortened to one hour during peak periods. The cache is managed using sixteen fixed shards. Origin retrieval failures follow a rhythm of one second, two seconds, and four seconds, with a maximum of three failures. Further failures are recorded as a failure, and concurrent requests at the same coordinates are merged for rate limiting. Each row-level record relies solely on the previously marked payment period and delivery location as anchor points: the payment period is first mapped to the Gregorian calendar year, and the delivery location is located using standard place names and administrative division codes, then mapped to the power grid partition. Based on these two coordinates, the region and year are determined first, and then candidate entries are filtered from the index.

[0043] If multiple versions exist under the same coordinates, the latest valid version that has not been withdrawn will be selected by default. If an annual entry is missing, it is only allowed to advance one year and only once, and the time, reason, and basis of the "annual downgrade" will be recorded in the locked path. If a regional entry is missing, adjacent partitions will be tried item by item according to the partition adjacency table filed by the enterprise (the execution version number is based on TAB-NEIGH-202509-V.1.0). The process will stop on the first successful match and must not cross more than two levels of adjacency. Only when no match is found can the filed alternative caliber be used, and the approval invoice number and the responsible person code will be written into the locked path. Calibration labels are limited to five categories: location method, market method, supplier actual measurement, industry average, and expenditure method. The location method and market method are only used for energy purchases and scenarios where settlement can be based on contracts. Supplier actual measurement must have a verifiable third-party report or a manufacturer report that has been traced through measurement. The industry average and expenditure method are only used as downgrade calibers when the aforementioned evidence is unavailable.

[0044] The source verification fingerprint is generated by concatenating the source organization name, version number, official release date, acquisition channel identifier, and the standardized byte sequence of the first paragraph of the original document to create a short code (only visible characters and line breaks are retained, and headers, footers, and extra whitespace are removed). This code is stored with the entry for easy spot checks. The latest valid version is determined based on the "latest release date that has not been withdrawn"; if an errata is issued, the release date after the errata is corrected is used as the benchmark; if a new version is released and then a withdrawal statement is issued, the system immediately reverts to the previous valid version, and the time and basis for the "withdrawal and rollback" are indicated in the locked path. Both regional mapping and power grid partitioning indicate the version number used; if the delivery location undergoes boundary adjustments during the billing period, the version for the current billing period is used and an explanation is recorded; cross-border and Hong Kong, Macao, and Taiwan deliveries are mapped to available overseas or regional partitions according to the enterprise equivalent caliber table, and the basis and responsibility for "cross-domain mapping" are indicated in the locked path.

[0045] The on-site execution is based on clearly defined failure criteria: a single screening takes more than two seconds, the candidate list is empty, regional and annual conflicts cannot be resolved, the source verification fingerprint fails, or the candidate version has a withdrawal mark and there is no previous valid version. Any of these situations is considered a failure; if the same record fails twice, it will be transferred to manual confirmation. To balance throughput and latency, index loading uses local hot caching and sharded fetching, with a single loading time limited to within 30 seconds and a concurrent screening capacity of no less than 3,000 rows per minute. The factor reference area provides a candidate list and locking path in key-value format. The minimum set of fields includes candidate number set, regional identifier, year, version number, caliber label, source identifier, source verification fingerprint, locking level, path text, and generation time. The interface uses an internal network authenticated channel, with an expected response time of 200 milliseconds, a 95th percentile of no more than one second, a timeout threshold of two seconds, automatic retries three times, and idempotency. The idempotency key is a combination of batch number, row number, and second-level timestamp. Repeated submissions are considered successful only on the first successful write and a "pre-existing" mark is returned.

[0046] Upon completion of loading, a candidate list is generated and written into the factor reference area along with the progressively detailed version locking path. The list number and path summary are then filled back into the corresponding row-level record, which is used by the next step to make a unique determination. The version locking path is a complete record of the steps to determine the latest valid version from the candidate factor set or to downgrade and replace it in a predetermined order, including the locking level, time, and basis. On the quality control side, cases such as unclear region, missing version pages, and source verification failures are all recorded as "downgrade reason + timestamp + responsible side" and transferred to the review channel. If the proportion of downgraded entries in the same batch exceeds one-tenth, the system automatically initiates a governance task, prioritizing checks on whether the region mapping table is outdated, whether the partition table has been changed, and whether the external database has been missed in synchronization. Sampling is conducted using a stratified equidistant method, with levels divided by region and year, each level having no fewer than fifty entries and the full sample having no fewer than one thousand entries. If a level fails to meet the standards, the sampling ratio of that level is immediately doubled, and a re-inspection is completed within seven days. The re-inspection results are used for fine-tuning the threshold and caching strategy for the following month. Logs and evidence retention periods are consistent with billing records and are no less than five years, covering locked paths, downgrade reasons, approval records, and source verification fingerprints.

[0047] For enterprises with higher timeliness requirements, an alternative approach can be adopted: maintain a local mirror within the data warehouse, perform incremental synchronization every Friday at 02:00 (Beijing time), allow a delay of 24 hours, and conduct mirror health checks every four hours. If a lag is detected, immediately switch back to the remote repository and record the reason to prevent systemic deviations caused by mirror lag. By clearly defining the above boundary, sequence, threshold, and logging requirements, loading and locking not only constrain the candidate range to the correct coordinates, but also allow for a quick review of the reasons for each step of the selection in case of disputes. Interface performance, caching strategies, failure mitigation, and sampling loop all have specific values ​​that can be executed against the table.

[0048] S4. Calculate the matching score based on attribute fit, spatiotemporal consistency, and version priority, and output a unique emission factor. If uniqueness is not achieved, generate a review task and record the judgment path. The specific implementation is as follows: In the process of determining the unique emission factor, the goal is to converge the candidates from the previous step to a definite value, while also fully documenting the process of the determination for easy review during subsequent accounting and auditing. The information relied upon includes: the attribute groups (material, specifications, shape, brand, measurement standard, etc.) and unit conversion trajectory (the source and accuracy of each step are recorded), the candidate list formed by loading the factor index (including region, year, version, caliber labels and source identifiers), and the version locking path (if a region or year downgrade occurs, the complete downgrade sequence will be recorded).

[0049] Before proceeding with the judgment, let's clarify the terminology: An attribute anchor is a field marked as strongly constrained in the company directory, containing at least material and model number, and optionally brand and shape. The minimum required set of attribute anchors is material and model number; if both are missing, the judgment proceeds directly to the next review stage without further boundary tightening attempts. Spatiotemporal consistency means that the year mapped to the payment period and the region, zone, and candidate record mapped to the delivery location are completely consistent; fuzzy matching is not accepted. Version priority follows a two-dimensional sorting based on the issuing organization and the release date. The order of issuing organization levels is national, industry, local, and company-built; within the same level, the later the release date, the higher the priority. Simultaneously, the measurement standards are defined: the company's standard units are fixed as kilograms, cubic meters, kilowatt-hours, meters, square meters, and pieces; non-standard units cannot directly participate in uniqueness judgment and must first register a complete conversion trajectory before scoring.

[0050] The preparatory steps are as follows: attribute synonym merging, unit benchmark verification, and spatiotemporal pre-check. Any candidates whose payment terms and delivery locations map to a region or year inconsistent with the initial candidate list are eliminated, and the reasons are recorded. Subsequent scoring revolves around three factors: attribute fit, spatiotemporal consistency, and version priority. To ensure reproducibility, a clear default weighting is provided: attribute factors account for 50%, spatiotemporal factors for 40%, and version factors for 10%. Before deployment, historical acceptance batches are used for calibration, followed by minor monthly adjustments based on actual operational data. Each adjustment cannot exceed 10%. Any changes to weighting, thresholds, or the whitelist require double confirmation and take effect uniformly at the top of the hour, while a change summary is automatically generated and written into the decision path.

[0051] The threshold uses a difference method, comparing the difference between the highest and second-highest scores. Uniqueness is only recognized when the difference reaches a predetermined threshold. The default threshold band is set between the 10th and 20th percentiles of the distribution over the past quarter, with an initial value of 10%, adjusted monthly on a rolling basis, with a single adjustment not exceeding 3%. The sample source for weights and thresholds is the acceptance batches from the most recent three months, covering three major categories: materials, equipment, and services. Samples are drawn proportionally by subject stratification, with no fewer than 200 rows per subject. To facilitate on-site understanding and spot checks, confidence levels are divided into three tiers: high-level confidence levels are defined as a difference of not less than 0.15 and a key field completeness of not less than 95%; medium-level confidence levels are defined as a difference between 0.08 and 0.15 or a completeness between 85% and 95%; and low-level confidence levels are defined as a difference below 0.08 or a completeness below 85%.

[0052] The decision-making order when scores are tied is also explained: If the overall scores are tied, first check if the attribute anchors are completely aligned, then check the spatiotemporal alignment. If they are still tied, choose the version with the higher issuing agency level and later release date. If the enterprise whitelist already covers the corresponding suppliers and categories, and the aforementioned rules still cannot create a significant difference, then the whitelist takes priority, and the item number and applicable period are recorded in the track. If the overall scores have already created a significant difference, but the whitelists point to different factors, then the score result shall prevail, and the differences in the whitelists shall be recorded for audit interpretation.

[0053] To balance stability and timeliness, each line is limited to two seconds. If the threshold is not reached on the first calculation, a small-scale tightening of the boundary is allowed, such as slightly adjusting the similarity threshold of the synonym mapping towards the robust side by a small step, with the step size not exceeding two percent and only allowed once. If the threshold is still not reached, it is considered that the uniqueness is temporarily not met and a review task is generated. At the same time, the candidate list, weight values, threshold parameters, rule version number and timestamp are frozen to ensure that the same conclusion can be obtained when rerunning with the same version at any point in time.

[0054] Each judgment generates three pieces of information: a unique emission factor number and its confidence level, a complete judgment path, and necessary prompts. The candidate list retains the regional code, year, version number, caliber label, issuing agency number, and source verification summary. The minimum set of fields for the judgment path includes a list of participating fields, scores, weights, differences, eliminated candidates and reasons, and, if downgrading occurs, a list of downgrading order and time points, rule version number, and source verification summary. All of the above information is uniformly written into the factor decision area, and the number and confidence level are backfilled in the current row record. If a review task is generated, a card with the original text fragment location and suggested focus points is pushed to the collaborative channel. The review time limit is four hours on weekdays, and automatic escalation occurs after the deadline. When the same item enters review for the second time within thirty days, a review is triggered, and the adaptive adjustment of that category is frozen to prevent the indicator from being skewed by short-term samples. The connection with upstream and downstream processes is completed through an ordered queue: the previous stage submits candidate lists and attribute groups by batch and row number; after this stage completes the uniqueness verification, the number and judgment path are submitted to the downstream for row-level numerical and caliber consistency verification.

[0055] To ensure throughput and stability, high-water level expansion and queuing protection are implemented. Under normal circumstances, the throughput for downstream judgments is no less than 2,000 rows per minute. Queue waiting times exceeding 30 seconds trigger expansion. A single row timeout is recorded as a failure, with a maximum of two attempts, retrying at intervals of one second and two seconds respectively. If the collaborative platform is unavailable for more than 15 minutes, the public disclosure of low-confidence results is suspended; only the pending queue is accumulated, and centralized processing resumes once the platform is restored. Three common unexpected events also have established handling paths: When factors with the same name are duplicated across regions and have inconsistent sources, the newer version is used as a temporary value and a mandatory review is performed; if all candidates are eliminated after spatiotemporal pre-check and dimensional verification, it indicates a possible mismatch in payment terms or delivery location upstream, and processing of that row is immediately suspended, with a governance request sent upstream. The conclusion for that row is not released until governance is completed; if the unit conversion trajectory is found to be inconsistent with the candidate dimensionality, a correction is requested from upstream. If the issue remains unresolved after the waiting time, it also enters the review process, and the trajectory nodes requiring attention are highlighted in the judgment path.

[0056] To demonstrate that the mechanism is operational and verifiable on-site, at least 2,000 rows are randomly sampled monthly, with sampling proportionally based on subject and category. The uniqueness success rate is calculated as the percentage of rows forming unique factors, with a target of no less than 95%. The first-time pass rate for review is calculated as the percentage of rows whose initial review conclusion matches the provisional judgment, with a target of no less than 90%. Decisions with high confidence levels are subject to further on-site sampling, with a sampling rate of no less than 10% and a target pass rate of no less than 98%. All judgment paths support replay at any time, using a four-tuple of batch number, row number, rule version number, and timestamp as the idempotent key. An immediate alarm is triggered when the replay result differs from the original record by more than 0.02%. Judgment paths are retained in hot storage for three years and in cold backup for seven years. Upon expiration, they are uniformly de-identified and destroyed, with the destruction action linked to the batch number and generating a receipt for inclusion in the annual compliance audit.

[0057] To illustrate the logic on-site, let's take a practical example: Under the steel category, using default weights, if two candidates in a certain row, with the same region and year, differ only in material matching, and the difference in their overall scores reaches 0.16, the higher-scoring item is directly determined as the sole emission factor. Another row in the same batch, due to incomplete material fields, has a difference of only 0.06 and is automatically sent to the collaborative platform. The card is highlighted with "Missing Material" and "Incomplete Specifications" as points of concern. The reviewer can then complete the missing information and trigger the judgment again. If there are still cases where the candidates are tied and cannot be differentiated, and if the enterprise's whitelist covers the supplier and category, the whitelist rules are used to prioritize determining the sole emission factor when candidates are tied or the matching difference does not reach the preset threshold. The item number and applicable period are written into the track. If the whitelist and the score conclusions contradict each other, the score is still used and the difference is recorded for easy explanation. Through the above guidelines and record keeping, this step clearly explains "why it is selected, how to review it after selection, what to do if it is not selected, what to focus on manually, and how to set and modify parameters." Boundaries, thresholds, samples, permissions, timing, and disclosure gates are all implemented on an executable scale.

[0058] S5. Calculate row-level emissions using activity data and a unique factor. Simultaneously generate location-based and market-based results in parallel on the same row, perform dimensional and caliber consistency checks, and set markers. The specific implementation is as follows: Having already identified a unique emission factor and locked the version and caliber labels for each row record in the previous step, while retaining unit conversion traces, payment terms, and entity information, this step counts row by row within the payment term window on a daily batch basis. The goal is to parallelize and cross-check the results of the location-based method and the market-based method within the same row, ensuring comparability of subsequent statistical calibers and smooth reconciliation. The caliber of activity volume uniformly adopts the enterprise's benchmark unit: materials are mainly in kilograms, cubic meters, or pieces; electricity is mainly in kilowatt-hours; gas is in cubic meters; and liquid fuels are in liters or kilograms. The conversion of volume and mass is based on 20 degrees Celsius and one standard atmosphere. If the temperature or pressure stated on the invoice is inconsistent with the aforementioned benchmark, the invoice conditions shall prevail, and the source of the deviation and the parameter origin shall be recorded in the evidence collection.

[0059] In electricity usage scenarios, readings shall be based on the active energy meter installed at the main metering point, with a metering level not lower than Level II as stipulated by national standards, and a verification cycle not exceeding twenty-four months. If dual meters or multi-rate meters exist, the main meter used for financial settlement shall prevail, while retaining the secondary meter as supporting evidence. For non-electrical energy metering such as material weighing and gas flow, the national verification procedures corresponding to the company's metering ledger shall be followed, and the equipment level and verification cycle shall not be lower than the minimum internal control requirements. The latest verification certificate number shall be retained on-site and linked in the records.

[0060] When encountering returns or red-ink invoices, a negative quantity is generally considered an offset against the original record: if the offsetting bank falls in the previous accounting period and is no more than three months away from the current period, an inter-period relationship is established and a negative adjustment is made in the original accounting period; red-ink invoices exceeding three months are recalculated according to the current accounting period, and the reason and original invoice number are noted in the record. When the unit of the line record is inconsistent with the unit required by the factor, it is converted step by step according to the saved unit conversion trajectory; if the conversion is missing key parameters (such as density, moisture content, or packing factor), it is first supplemented according to the supplier specifications or the company's commonly used tables. If it still cannot be determined, it is marked "conversion missing parameter", and the line will not participate in the consistency verification and subject-level summary for the time being.

[0061] To prevent subjective adjustments to the scope, direct cross-dimensional aggregation is prohibited for the same entity within the same billing period. The system only performs the conversion within the line before proceeding to the entity-level verification. Market-related contract information (covered product type, covered electricity volume or percentage, start and end dates, settlement entity, traceable number, and signatory) is retrieved once from the contract ledger by billing period. When the same entity has multiple valid contracts, priority is given to the contract with the earliest start time that is still valid, and the rest are allocated in descending order of coverage volume. For contracts that only cover a portion of the days, priority is given to site mapping, followed by time period mapping, and finally equal allocation by calendar, with the allocation path written into the evidence set. Site mapping is considered a successful match if any one of the three—power supply account number, metering point number, and contract site code—is consistent, with the granularity defaulting to hourly aggregation.

[0062] The deduction of green electricity vouchers and renewable electricity is subject to strict uniqueness and validity management: the same voucher number cannot be used repeatedly by multiple entities within the same accounting period. If duplicates are detected, the first record will be confirmed according to the first-come, first-served principle, and the remaining entries will be transferred to pending investigation with their source indicated; the voucher number must meet the length and check digit rules, and any verification failure will be regarded as not being covered and included in pending investigation. The region and year used for the location method are based on the previously locked results. If missing, they will be mapped according to the power supply unit on the enterprise's electricity bill; the residual electricity formed after the market method deduction will continue to use the location method factor of the same region and year.

[0063] The system generates two parallel results, one from the location approach and one from the market approach, and performs consistency checks on the same entity and the same payment period. The difference threshold is set in a tiered manner: 20% by default for regular entities, 15% for large energy-consuming entities, and 10% to 25% for entities in special industries, which is fixed after annual review. Large energy-consuming entities are defined as those with annual electricity consumption of 5 million kWh or more, and the threshold is fixed by annual review. Minor adjustments can be made quarterly, with each adjustment not exceeding ±5%. Once the threshold is exceeded, the system is marked as pending investigation, and the difference is attributed to the following in priority: insufficient contract coverage, insufficient number of certificates, inconsistent regional mapping, and abnormal metering conversion. If multiple factors are identified simultaneously, the primary attribution is determined in that order, and secondary attributions and supporting evidence are listed in the notes.

[0064] To meet the requirements of monthly closing and large-scale batch processing, the end-to-end time limit for batch verification is controlled within 15 minutes, with a parallel scale of no less than 200,000 rows. When the queue level approaches the high threshold, it is divided into multiple shards by subject and billing period for parallel execution. Within each shard, idempotent writing is performed using the batch number and row number as unique identifiers. When a duplicate row arrives again, only the status bit and timestamp are refreshed, and it is not counted repeatedly. For row-level data entry failures, two retries are performed with gradually increasing intervals. If the third attempt still fails, the row is directly placed under investigation and subsequent operations related to that row are frozen. When the data entry failure rate of the same subject and the same billing period exceeds 0.5%, an alarm is triggered and a governance task is generated.

[0065] After the calculation is completed, each row will generate three core records and synchronize them: First, the row-level value saved according to the enterprise's benchmark unit, internally retained to the fifth decimal place, and displayed uniformly to the third decimal place; when the factor release precision is less than three decimal places, the factor precision is used as the upper limit and the source precision is marked in the record; Second, the caliber label, using a fixed code set, with the location method recorded as A and the market method recorded as B; when both are used in parallel, they must appear on the same row and cannot be split; Third, the consistency status, divided into three types: normal, pending investigation, and deferred processing. The three records are uniformly written into the row-level value area, and the status bit is refreshed on the row record; all "pending investigation" are aggregated into a list by subject and payment period and pushed to the on-site inspection channel. The list is accompanied by the primary attribution, secondary attribution, and necessary context fragments to reduce back-and-forth communication.

[0066] Regarding time standardization, all time fields uniformly use Beijing time; if the source system is located in another time zone, the time zone conversion is completed on the receiving side, and the original time zone is retained as a footnote field. To ensure institutionalized operation, row-level records are retained for a minimum of five years; monthly reconciliation should be completed within the fifth working day after the end of the accounting period, and quarterly reconciliation should be completed and archived within the tenth working day after the end of the quarter; contracts and vouchers involving the market approach that are expired, duplicated, or have failed number verification are considered as not covered and included in the pending review; after subsequent supplementation, they are re-entered according to the original accounting period, and correction traces are retained. Process effectiveness is based on two hard indicators as permanent benchmarks: the pass rate for consistency verification between the location method and the market method is no less than 98%, and the accuracy rate for dimensional alignment is no less than 99.5%; sampling adopts a proportional strategy by subject and accounting period, with a confidence level of no less than 95%, and a sample size of no less than 10,000 rows per month; details that fail to pass must be manually reviewed and rewritten within five working days.

[0067] A typical electricity consumption scenario in operation is as follows: In April, a manufacturing base recorded a total electricity consumption of 100,000 kWh at its metering point. The green electricity contract took effect on the 15th and covered 30,000 kWh. The system calculated the total amount based on the factor of the corresponding year in the East China region using the location method, and deducted the amount based on the number of days in effect and the covered amount using the market method. The remaining part was still calculated according to the location method. The difference between the two results fell within the stratification threshold and was marked as normal. In the same batch, a leased site recorded zero electricity consumption under the market method because the contract was not registered. The difference between the location method and the market method exceeded the threshold. The system listed it as pending investigation and gave the primary attribution prompt of "insufficient contract coverage".

[0068] The material scenarios also have clear boundaries: a batch of lubricating oil is measured in liters and the invoice specifies 15 degrees Celsius; the system converts it to kilograms based on the conversion factor at that temperature. Another item in the same batch is "boxes," but the catalog lacks a packing factor for this category, so it is marked as "conversion parameter missing" and entered into the pending investigation, not included in the main summary, until the supplier provides the packing factor before being added back. If the business side is resource-constrained or has higher timeliness requirements, a conservative approach can be adopted: first, quickly generate the location method and complete the dimensional verification; after the contract ledger is asynchronously backfilled, then generate the market method in batches and backfill the same line. At the same time, add a "subsequent cause and allocation path" explanation to the evidence set to ensure clear audit traceability and consistency. The entire process is seamlessly connected with the preceding unique factor determination, the subsequent evidence chain, and the uncertainty assessment, not only firmly anchoring the common dimensions and contract boundaries of enterprises, but also realistically describing thresholds, stratification, and fallback paths, ensuring that personnel in this field can reliably reproduce it under general system conditions.

[0069] S6. Generate an evidence chain object for each line, recording text fragments, extraction traces, unit conversion trajectories, emission factor sources, and version locking paths. Construct uncertain quantification indicators, triggering a review when the threshold is exceeded, and writing the results back to the alias ledger and rule snapshot. The specific implementation is as follows: After the row-level results are generated, the system immediately establishes a chain of evidence for each row that can be directly compared. The goal is to fix where the information in this row was obtained, how it was identified, how the unit conversion was completed, which version of the emission factor was used, and when and by whom it was confirmed, so as to achieve the traceability requirements of audit level. At the same time, it makes the uncertainty explicit, triggers review when it goes out of bounds, and sinks the review conclusion into the enterprise rule base for subsequent automatic adoption. The evidence chain elements include: the original text fragment and its location in the invoice format (marked by page number, inline offset, column offset, and character range; character encoding is uniformly the enterprise's common code; time is uniformly Beijing time, allowing clock drift of no more than one minute, otherwise a time synchronization event will be recorded); the terms hit during element extraction, the corresponding directory source, the identified field and time, and the current rule version number; the complete trajectory of unit conversion (source unit, target unit, conversion basis, accuracy and tolerance, confirmer and time for each step); the source identifier, region and year, version number and caliber label of the emission factor; if a substitution occurs, the downgrade order and reason are recorded as is; and the signature and timestamp of the person handling the transaction.

[0070] The evidence chain object refers to a traceable record that coexists with a specific row of results. It includes original text fragments and location indexes, feature extraction traces, unit conversion trajectories, factor sources and version locking paths, operator signatures and timestamps, and an incremental log summary. After generation, the content summary is calculated on the spot and written to the incremental log. Any changes are appended as a new version, while the original version is retained. The log records the batch number, row number, summary value, operator, and timestamp, supporting cross-database verification and spot checks. Access control identifiers are set, with viewing and modification belonging to different roles. Permissions are configured to be minimal, and modifications are only allowed to be added to the record; the original record cannot be overwritten. To ensure long-term verifiability of evidence, the evidence chain and row-level results are stored in the same domain, managed by batch and date partitions, and single-line writes are limited to within one second; frequently used fields reside in online hot storage, while original long texts and bitmap snapshots are placed in cold backups for a retention period of no less than ten years; message delivery adopts at least-once semantics, consumers use "batch number + row number" as an idempotent key for deduplication, and the queue automatically expands and throttles when it reaches a high watermark and remains so for fifteen minutes, clearing in-transit tasks before restoring the standard rhythm.

[0071] Before evidence is released, a boundary check is performed: the original text fragment should be able to be replayed and located in the original format, the position index should not exceed the boundary, and the source identifier must be able to be retrieved to the corresponding entry in the authoritative database; for formats derived from scanned documents, the recognition quality score is also checked, and if it is below 80%, it is considered unstable and a row position map snapshot must be saved simultaneously; if it is below 70%, it will directly enter the manual channel. If the proportion of evidence fields missing, source verification failure, or signature abnormalities in the same batch exceeds 1%, the system will automatically generate a compliance verification list; if it exceeds 3%, public disclosure will be immediately stopped, only the lines that have passed the review will be released, and the remaining ones will be processed according to the emergency verification process. To avoid the loss of accuracy due to cross-currency amounts, if the bill contains multiple currencies, it will be uniformly converted according to the enterprise settlement exchange rate table of the day, and the source and time of exchange rate will be written into the evidence object; suppliers and bill formats with repeated problems will be included in the quality control list, and entities that are included in the list three times in a row will be blacklisted and trigger on-site verification.

[0072] To address uncertainty, the system sets interpretable indicators and assigns priorities based on four aspects: activity volume, emission factors, unit conversion, and caliber selection. A difference exceeding 3% in the measurement chain is considered an overreach on the activity volume side; a deviation of more than one year between the payment period and version, or an untraceable factor source, is considered an overreach on the factor side; more than two cascaded unit conversions, or a tolerance exceeding 0.5% in any step, are considered an overreach on the conversion side; a difference exceeding 10% between the location method and the market method for the same entity and payment period, or exceeding the contractually agreed upper limit, is considered an overreach on the caliber side. Upon any overreach, the system immediately generates a review task, highlighting the main causes and suggested verification paths on the task sheet. Tasks are categorized into high, medium, and general levels: high-level tasks are processed within four hours, medium-level tasks within twenty-four hours, and general tasks within three working days. Tasks exceeding these limits are automatically escalated and the responsible party is notified. Thresholds are adjusted using a daily rolling adjustment mechanism. When the rolling result fluctuates by more than 2% of the original value within three consecutive days, it automatically reverts to the previous week's stable period and freezes for seven days, only to be unfrozen after review by the rules administrator.

[0073] Once the review conclusion is confirmed, it is written back to the alias ledger and rule snapshot, synchronously recording the version number, coverage, and effective time. This ensures that subsequent similar lines directly adopt the improved conversion method, avoiding duplicate alerts. For example, in a steel transaction, the conversion between "box → kilogram → ton" caused the conversion-side indicator to exceed the limit. The system automatically generates a review task, highlights the "unit conversion," and provides a related playback link. After the reviewer verifies according to the prompts, they adopt the more rigorous conversion method in the enterprise's local standard, sign to confirm, and re-inject the rule. From the next batch onwards, the same path will no longer exceed the limit. The end-to-end concurrency capability is configured to be no less than 5,000 lines per minute, with cold backup asynchronous deployment. During peak business periods, online writes are prioritized. If a single line write fails to meet the time limit, it is allowed to retries three times with increasing intervals. If it still fails, it is transferred to a manual checklist, and the original evidence is retained to avoid link interruption. Sensitive fields such as taxpayer identification numbers are stored using face masks, which can support spot checks while reducing the risk of leakage.

[0074] To measure the degree of implementation and on-site performance, monthly spot checks with sufficient coverage are conducted: at least one thousand lines are sampled, covering at least three entities, three product categories, and three payment terms. The checks are conducted by two independent reviewers with consistency as the standard. The completeness and reproducibility of evidence elements are statistically analyzed. At the same time, the average working hours for manual review, concurrent latency, and peak stability are continuously monitored. If the indicators fail to meet the standards, a rule optimization work order is triggered.

[0075] Through the above arrangements, evidence and results coexist, uncertainty is no longer "hidden," reviews have clear priorities and time limits, and improvements made after reviews are remembered by the system for a long time and automatically applied. Combined with tamper-proof incremental logs, long-term retention and hierarchical storage, arrival and idempotency constraints, compliance stop-loss line and blacklist mechanism, the whole chain achieves a verifiable balance between time, resources, compliance and auditability. On-site personnel can also quickly locate problems and close the loop with a piece of evidence and a task link, and directly reuse the confirmed statements in the next similar situation, reducing disputes and rework.

[0076] All calculations involved in the embodiments are dimensionless numerical calculations, and the preset parameters and thresholds in the calculations are set by those skilled in the art according to the actual situation.

[0077] 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.

[0078] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as 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, all or part of the processes or functions described in the embodiments of this application are generated. 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 from one website, computer, server, or data center to another website, computer, server, or data center via wireless or wired transmission; wired transmission methods include optical fiber, twisted pair, coaxial cable, etc.; wireless transmission includes infrared, microwave, etc. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center containing one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0079] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0080] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0081] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0082] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0083] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0084] 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.

[0085] 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 method for intelligent analysis of carbon emission data, characterized in that, include: S1. Receive business data and segment it by line, mark the billing period, subject and delivery location, retain the original text fragments and location index, and form a line-level unit to be processed. S2. Extract elements from Chinese cargo descriptions, unify names and attributes based on alias ledgers, unify units of measurement, and generate unit conversion trajectories. S3. Load the emission factor index containing regional, annual, version and caliber labels, parse the spatiotemporal information based on the payment period and delivery location, screen out candidate factors and complete version locking; S4. Calculate the matching score based on attribute fit, spatiotemporal consistency and version priority, output a unique emission factor, and generate a review task and record the judgment path if uniqueness is not achieved. S5. Calculate row-level emissions using activity data and a unique factor, and simultaneously generate location-based and market-based results in parallel on the same row, perform dimension and caliber consistency checks and set markers; S6. Generate an evidence chain object for each line, record text fragments, extraction traces, unit conversion trajectories, emission factor sources and version locking paths, construct uncertain quantification indicators, trigger a review when the threshold is exceeded, and write the results back to the alias ledger and rule snapshot.

2. The intelligent carbon emission data analysis method according to claim 1, characterized in that, S1 includes: A unified channel is established on the receiving side to stably access invoices from the financial system and energy consumption ledger, receiving them according to fixed and incremental rhythms. A one-to-one mapping is established between ticket numbers and subject codes, with unified time and unit, and duplicates are removed based on source priority and arrival order; Missing payment terms are made up according to the invoice month. Delivery location is mapped in the order of receiving address, contract address, and invoice address. If mapping fails, a temporary processing mark is set. The text is segmented line by line according to the format, and original fragments and location indexes are generated. The billing period, subject, and delivery location are added. The fragments are packaged into line-level unprocessed units containing batch number, line number, and content summary, and delivered to the message queue with an idempotent identifier composed of batch number and line number.

3. The intelligent analysis method for carbon emission data according to claim 1, characterized in that, S2 include: The elements of the Chinese product description are extracted and normalized, and the brand, material, specifications, shape, model and unit of measurement are grouped into standard names and attribute groups using an alias ledger; When the normalization is difficult to determine, the context fields are called sequentially based on adjacent rows of the same invoice, the same supplier in the same batch, and the contract material strip. Units of measurement are unified to the standard units. Unit conversion trajectories are generated step by step according to the conversion table. The source and scope are recorded and the trajectory number is written back to the line. For cases where the model and unit are inconsistent, the material and shape are inconsistent, the specifications are missing and the context is insufficient, a gray mark is set and a checklist is generated; The alias ledger and conversion table have versions and effective ranges, and their history is unified and not rolled back; When multiple materials are listed side-by-side, they are split into sub-items by a separator while maintaining their order and position index; The normalization results are stored in the normalization region for factor selection and version locking.

4. The intelligent analysis method for carbon emission data according to claim 1, characterized in that, S3 include: When loading the emission factor index, the index includes the region identifier, year, version number, caliber label, source identifier, and source verification fingerprint; Based on the payment term mapping year and the delivery location mapping region and zone, first determine the region and year, then filter candidate factors and lock the latest valid version that has not been withdrawn; When an annual entry is missing, only the most recent year is downgraded and the locked path is recorded. When a region entry is missing, the system will attempt to match it item by item according to the order of adjacent partitions in the filing and record the locking level. If no match is found, the approved alternative will be used and the responsibility information will be written. Write the candidate list along with the version lock path into the factor reference area, and backfill the list number and path summary into the corresponding row-level record.

5. The intelligent analysis method for carbon emission data according to claim 1, characterized in that, S4 include: After version locking is completed, attribute anchor synonym merging, unit benchmark verification and spatiotemporal pre-check are performed on each record to eliminate candidates that are inconsistent with the payment period and delivery location. The remaining candidates are matched based on a combination of attribute fit, spatiotemporal consistency, and version priority according to preset weights to obtain a matching score. When the difference between the highest and second-highest matching scores reaches a preset threshold, a unique emission factor is determined, and a decision path containing participating fields, scores, weights, reasons for removal, and version information is generated. If the threshold is not reached, a review task is generated and the candidate list, weights, thresholds and rule versions are frozen, and timestamps are added to ensure consistency when rerunning with the same version. The unique emission factor number and determination path are written into the factor decision area and backfilled to the row-level record.

6. The intelligent analysis method for carbon emission data according to claim 1, characterized in that, S5 include: At the row level, based on activity volume and unique emission factor, both location-based and market-based results are generated on the same row and labeled with caliber labels. The market coverage is allocated according to the contract ledger in the order of site mapping, time period mapping, and calendar equalization. Uncovered residual amounts are accounted for using the same factor caliber as the location method, and the allocation path and caliber label are recorded in this line; Cross-dimensional superposition is prohibited within the industry; the unit conversion trajectory that has been saved shall be used as the standard for unit uniformity.

7. The intelligent analysis method for carbon emission data according to claim 6, characterized in that: For the same entity and the same payment period, a consistency check should be performed between the location-based approach and the market-based approach. In case of inconsistency, the attribution shall be determined according to the priority order: insufficient contract coverage, insufficient number of vouchers, inconsistent regional mapping, and abnormal measurement conversion. A pending status shall be generated in the bank and the attribution, allocation path and evidence fragment shall be written. Green electricity certificates are verified for uniqueness by number and can only be used by a single entity. All results and tags are stored in the database with batch number and row number as idempotent keys and can be used for subsequent evidence chain generation.

8. The intelligent analysis method for carbon emission data according to claim 1, characterized in that, S6 include: Immediately after the row-level results are generated, an evidence chain object is created for that row. The object includes the original text fragment and location index, feature extraction traces, unit conversion trajectory, emission factor source and version locking path, and the signature and timestamp of the person in charge. Generate a content summary and write it to an increment-only log; any changes are appended to the new version while retaining the original version. Set viewing and modification permissions and only allow supplementary entries; The chain of evidence and the results are persisted in the same domain and partitioned by batch and date, allowing for playback and location; Batch number and line number are used as idempotent identifiers for transmission and access.

9. The intelligent analysis method for carbon emission data according to claim 8, characterized in that: Based on the evidence chain object, uncertain quantitative indicators are constructed for activity volume, emission factor, unit conversion, and caliber selection, and thresholds and priorities are set; When any uncertain quantification exceeds its threshold, a review task is automatically generated and the main cause is highlighted. The current candidates and parameters are frozen. The review conclusion is written back to the alias ledger and rule snapshot and the version, coverage and effective time are recorded. It is then automatically adopted in subsequent similar rows. Review tasks are processed according to grade and time limits and can be upgraded.

Images