Electric carbon calculation collaborative data quality control method and system based on anomaly identification

Abstract

The application relates to the technical field of data quality control of a data center computing power carbon footprint database, and discloses an electric carbon computing collaborative data quality control method and system based on abnormality identification. The method accesses multi-source electric carbon computing physical entity data of a server power, a power meter power, a carbon factor version and scheduling records of an object in a time slice, calculates energy consumption balance deviation or carbon factor time sequence deviation, and aggregates the deviations into a comprehensive abnormality score. According to an abnormality level, the carbon footprint database publishing state is migrated to isolation or freezing, affected tasks are located along a blood relationship diagram, and a recalculation queue is generated. According to a homologous abnormality correlation coefficient, candidate sources are removed and weighted to generate candidate correction values that do not cover original data. After physical consistency rechecking, the candidate correction values are published and trigger recalculation, which helps to reduce the risk of abnormal physical data entering the formal carbon footprint version.

Description

Technical Field

[0001] This application relates to the field of data quality control technology for data center task-level computing power carbon footprint databases, and in particular to a collaborative data quality control technology for carbon footprint databases based on anomaly identification and linking the anomaly identification results with the migration of the carbon footprint database release status. Background Technology

[0002] As large data centers, intelligent computing centers, and cloud computing platforms provide computing power services to users, the need to calculate their computing power carbon footprint at the task level is becoming increasingly prominent. For example, in computing power service scenarios for artificial intelligence model training and inference tasks, it is necessary to aggregate the electrical energy consumed and the corresponding carbon emissions of a single training task during its operation to that task; in multi-tenant cloud computing scenarios, it is necessary to allocate data center electricity consumption, power distribution losses, and cooling energy consumption according to tenant tasks, and calculate the carbon footprint of each task accordingly; in scenarios where carbon emission reports are issued to external customers, it is also necessary to ensure that the calculated task carbon footprint results are traceable, verifiable, and publishable. In the above scenarios, the calculation of task-level computing power carbon footprint often relies on multiple sources of data, including task logs, server power collected through server baseboard management controllers or intelligent platform management interfaces, electricity meter power measured by power distribution units, cooling power from cooling allocation units, carbon factor versions, and scheduling records characterizing task scheduling lineage. Moreover, the three types of data—electricity, carbon, and computing—are intercoupled and mutually verified.

[0003] However, anomalies in any of the aforementioned multi-source data sources—such as missing task logs, abnormal resource utilization, sudden spikes in server power, unbalanced electricity meter readings, decoupling of cooling power and load, or incorrect carbon factor versions—can contaminate the task carbon footprint dataset along the accounting chain. Existing general anomaly detection methods in such scenarios mostly only provide anomaly alerts, failing to link the anomaly identification results with release control actions such as isolation, freezing, candidate correction, rollback, and task recalculation in the carbon footprint database. They also struggle to traceably identify the specific tasks affected by the anomaly and to automatically generate corresponding candidate correction and recalculation actions, allowing contaminated data to potentially still enter the official carbon footprint release version. Therefore, a new technical solution is urgently needed that can block multi-source anomaly data in the release chain of the task-level computing power carbon footprint database, traceably locate affected tasks, and generate correction and recalculation actions. Summary of the Invention

[0004] The purpose of this application is to provide a collaborative data quality control method and system for carbon footprint analysis based on anomaly identification. This method addresses the technical problem that when anomalies occur in multi-source carbon footprint analysis data in existing task-level computing power carbon footprint databases, the anomaly identification results often only result in alarms. It is difficult to link these anomalies with release control actions such as isolation, freezing, candidate correction, rollback, and task recalculation in the carbon footprint database. This makes it difficult to trace and locate the affected tasks and to automatically generate corresponding candidate correction and recalculation actions, thus allowing contaminated data to still be included in the official carbon footprint release version.

[0005] This application provides a collaborative data quality control method for carbon footprint analysis based on anomaly identification, applied to an electronic device including a processor, memory, and multi-source data acquisition interfaces, for controlling the release status of a data center task-level computing power carbon footprint database. The method includes: S1. Through the multi-source data acquisition interface, access the multi-source electrical carbon calculation physical entity data of the object within the time slice. The multi-source electrical carbon calculation physical entity data includes at least the server power collected by the server baseboard management controller or intelligent platform management interface, the meter power measured by the power distribution unit, the cooling power from the cooling allocation unit, the carbon factor version, and the scheduling record characterizing the data lineage. S2. Based on the multi-source carbon computing physical entity data mentioned in S1, calculate the various types of carbon computing anomaly indicators of the object within the time slice and summarize them into a comprehensive anomaly score. The various types of carbon computing anomaly indicators include at least the energy consumption balance deviation characterizing energy conservation or the carbon factor time series deviation characterizing time series jump. S3. Based on the anomaly level of the comprehensive anomaly score mentioned in S2, control the release status of the object in the release status table of the task-level computing power carbon footprint database, and when the comprehensive anomaly score meets the high anomaly level condition, migrate the release status to the isolation status or the frozen status. S4. For the objects that are in the isolated or frozen state in S3, propagate along the lineage graph of the task-level computing power carbon footprint database to locate the affected task set, calculate the electricity consumption ratio of each affected task affected by the anomaly, and when the electricity consumption ratio exceeds the task threshold, update the carbon footprint result of the corresponding task in the affected task table to the state that needs to be reviewed and generate a recalculation queue. S5. For the object in the isolated or frozen state, extract candidate sources to generate candidate correction values ​​that do not cover the original data, and calculate the homology abnormal correlation coefficient between each candidate source and the object in the lineage diagram in S4 during generation. Based on this, candidate sources that are directly homologous are eliminated and candidate sources that are multi-hop homologous are downgraded in weight. S6. According to the anomaly type of the object, input the candidate correction value in S5 into the corresponding physical consistency verification gate for closed-loop verification. When the candidate correction value meets the physical consistency verification threshold, migrate the release state to the new release state and trigger the corresponding task in the recalculation queue in S4 to call the candidate correction value to recalculate the carbon footprint.

[0006] In a preferred embodiment, the energy consumption balance deviation described in S2 Time deviation of carbon factor The calculation formulas are as follows: in, This indicates the power measured by the power distribution unit within the time slice. This indicates the total power consumption of servers within the coverage area of ​​this electricity meter. Indicates unassigned power; This indicates the carbon factor version value of the current time slice. This represents the median value of the carbon factor version within adjacent windows centered on the current time slice. This represents the median absolute deviation of the carbon factor version values ​​within the adjacent windows.

[0007] In a preferred embodiment, the comprehensive anomaly score described in S2 The calculation formula is: in, This indicates a deviation in the completeness of the task log. This indicates the energy consumption balance deviation. Indicates cooling coupling deviation. This indicates the time-series deviation of the carbon factor. This indicates a risk bias related to missing bloodlines. to These are the weights of the corresponding abnormal indicators; The rule for determining the anomaly level described in S3 is: when When the threshold is greater than or equal to the high anomaly threshold, the system migrates to an isolated or frozen state; when... When the value is less than the high anomaly threshold and greater than or equal to the medium anomaly threshold, the state transitions to a candidate correction state; when... When the value is less than the medium anomaly threshold, the status is transitioned to a marked release state; wherein, the high anomaly threshold used for formal disclosure tasks is stricter than the high anomaly threshold used for internal tasks.

[0008] In a preferred embodiment, the node types of the lineage graph described in S4 include tasks, servers, rack groups, electricity meters, cooling distribution units, carbon factor versions, and scheduling versions, and the edge types include at least task execution relationships, electricity meter coverage relationships, carbon factor call relationships, scheduling allocation relationships, and cooling distribution relationships. The electricity consumption percentage is for tasks. Task Influence The calculation formula is as follows: in, Indicates task In time slice Electricity consumption Represents an exception object In time slice The aforementioned comprehensive anomaly score, Indicates the abnormal threshold. For indicator functions, when The value is 1 if the condition is met, and 0 otherwise. The propagation range of the anomaly along the lineage graph is determined based on the anomaly type of the object. Specifically, the propagation range of the meter balance anomaly is the server pool covered by the same meter, and the propagation range of the carbon factor timing anomaly is the task in the same region and time slice.

[0009] In a preferred embodiment, the candidate sources mentioned in S5 include adjacent time period data, data from the same server pool, data of the same task type, deduction data from the superior meter, and data from the previous trusted version; Wherein, direct homology represents the candidate source. Sharing the same abnormal meter, abnormal carbon factor, abnormal cooling unit, or abnormal scheduling version with the aforementioned object, and ; The comprehensive weight is calculated based on the aforementioned homologous anomaly correlation coefficient. And generate the candidate correction value. : in, Candidate source type The baseline credibility, For time proximity, The similarity of operating conditions is determined based on server pool consistency and task type consistency; the candidate correction value Candidate fields are written into the isolated data table and stored alongside the object's original observation data, the previous trusted version data, and the source of the correction. An audit snapshot is generated after a new release, which includes the cause of the anomaly, the affected tasks, and the difference in carbon footprint before and after the correction.

[0010] In a preferred embodiment, the step S6 of inputting the corresponding physical consistency verification gate according to the anomaly type of the object includes: When the anomaly type of the object belongs to the energy consumption or meter channel, the candidate correction value is input into the energy consumption channel verification gate to calculate the energy consumption correction verification value. : When the anomaly type of the object belongs to the carbon factor channel, the candidate correction value is input into the carbon factor channel verification gate to calculate the carbon factor correction verification value. : in, This is the energy consumption re-verification threshold, and it is stricter than the high energy consumption anomaly threshold corresponding to the isolation state. This is the carbon factor re-verification threshold, and it is stricter than the high anomaly threshold for carbon factors corresponding to a frozen state.

[0011] In a preferred embodiment, for cases where the physical consistency verification gate is not passed, iterative and degenerate processing is further included: when the candidate correction value fails the corresponding physical consistency verification gate, the current homogeneous abnormal correlation coefficient is removed. After identifying the largest single candidate source, the process returns to recalculate the candidate correction value and performs closed-loop verification again, with the number of iterations not exceeding the total number of extracted candidate sources. When the iteration reaches the total number and still fails, or when the sum of the combined weights of all candidate sources is reached... When the automatic generation of the candidate correction value is stopped, the object is kept in an isolated or frozen state, and the object's data is rolled back to the previous trusted version and a manual review is triggered; when the calculation of the carbon factor correction verification value encounters... When necessary, expand the range of the adjacent windows or switch to historical adaptive thresholds.

[0012] This application provides a collaborative data quality control system for carbon footprint analysis based on anomaly identification, applied to an electronic device including a processor, memory, and multi-source data acquisition interfaces, for controlling the release status of a data center task-level computing power carbon footprint database. The system includes: The multi-source access module is configured to access the multi-source electrical carbon calculation physical entity data of the object within the time slice through the multi-source data acquisition interface. The multi-source electrical carbon calculation physical entity data includes at least the server power collected by the server baseboard management controller or the intelligent platform management interface, the electricity meter power measured by the power distribution unit, the cooling power from the cooling allocation unit, the carbon factor version, and the scheduling record characterizing the data lineage. The anomaly index module is configured to calculate multiple types of anomaly indexes of the object within the time slice based on the multi-source electric carbon computing physical entity data and summarize them into a comprehensive anomaly score. The multiple types of electric carbon computing anomaly indexes include at least energy consumption balance deviation characterizing energy conservation or carbon factor time-series deviation characterizing time-series jumps. The state transition module is configured to control the release status of the object in the release status table of the task-level computing power carbon footprint database according to the anomaly level to which the comprehensive anomaly score belongs, and to transition the release status to an isolated state or a frozen state when the comprehensive anomaly score meets the high anomaly level condition. The impact set module is configured to, for the object in the isolated state or frozen state, propagate along the lineage graph of the task-level computing power carbon footprint database to locate the affected task set, calculate the electricity consumption ratio of each affected task affected by the anomaly, and when the electricity consumption ratio exceeds the task threshold, update the carbon footprint result of the corresponding task in the affected task table to the state that needs to be reviewed and generate a recalculation queue. The candidate correction module is configured to extract candidate sources for the object in the isolated or frozen state to generate candidate correction values ​​that do not cover the original data, and calculate the homology abnormal correlation coefficient between each candidate source and the object in the lineage diagram during generation, thereby eliminating directly homologous candidate sources and downgrading the weight of multi-hop homologous candidate sources. The re-verification release module is configured to input the candidate correction value into the corresponding physical consistency re-verification gate for closed-loop re-verification according to the anomaly type of the object. When the candidate correction value meets the physical consistency re-verification threshold, the release state is migrated to the new release state, and the corresponding task in the recalculation queue is triggered to call the candidate correction value to recalculate the carbon footprint.

[0013] This application provides an electronic device, including: a processor; and a memory storing an executable computer program thereon, wherein when the computer program is executed by the processor, it implements the steps of the above-described anomaly-based collaborative data quality control method for carbon dioxide calculation.

[0014] This application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method for collaborative data quality control based on anomaly identification.

[0015] By integrating multi-source electricity carbon computing physical entity data into the task-level computing power carbon footprint database under the same time slice, server power, electricity meter power, cooling power, carbon factor version, and scheduling records can be calculated, compared, and traced under a unified object and a unified time dimension. This allows subsequent anomaly identification to no longer rely on isolated data values ​​but to be based on the technical fields of the actual power supply and distribution, cooling allocation, and task execution links of the data center. Since the integrated data all have clear physical or engineering sources, this process helps reduce the risk of misusing abstract statistical anomalies for carbon footprint release control and provides a verifiable data foundation for subsequent database state migration.

[0016] In the anomaly indicator calculation stage, energy consumption balance deviation is constrained by the balance relationship between the metered power of the power distribution unit, the total power of the server, and the unassigned power. It identifies sudden meter jumps, server power omissions, or energy consumption anomalous assignments from the perspective of physical consistency between the power distribution side and the load side. Carbon factor time series deviation identifies sudden jumps in factor versions that may affect the task's carbon footprint conversion by using the relative deviation between the current time slice's carbon factor version value and the statistical benchmark of adjacent windows. When task log completeness deviation, cooling coupling deviation, and lineage loss risk deviation are all incorporated into the comprehensive anomaly score, the impact of different anomaly sources on the credibility of the same object's release can be expressed under a unified scoring framework, thus providing a more stable basis for judging object state transitions.

[0017] Based on the anomaly level assigned to the comprehensive anomaly score, the object status in the release status table is controlled, ensuring that anomaly identification results directly impact the release process rather than merely generating alarms: high-anomaly-level objects are moved to an isolated or frozen state, helping to block abnormal physical data or anomaly factor versions before they enter the official release; medium-anomaly-level objects are moved to a candidate correction state, preserving correction space while preventing the direct release of suspicious data; low-anomaly-level objects are released with tags, maintaining data continuity while retaining quality indicators. This establishes a clear technical connection between anomaly detection, database object status, and task carbon footprint recalculation.

[0018] Once an object enters an isolated or frozen state, it propagates along the lineage graph of the task-level computing power carbon footprint database to locate the affected task set. This allows the abnormal event to be transmitted from its location to the carbon footprint result of the task that actually calls the data. Since the lineage graph simultaneously includes tasks, servers, rack groups, electricity meters, cooling allocation units, carbon factor versions, and scheduling versions, and establishes associations based on edge types such as task execution, electricity meter coverage, carbon factor invocation, scheduling allocation, and cooling allocation, different abnormal types can have their propagation range limited along the corresponding physical or engineering links. Combined with the task impact degree, the electricity consumption percentage of each task affected by the abnormality is calculated, which helps to limit the status requiring review and recalculation queue to the actually affected tasks. This reduces the possibility of irrelevant tasks being mishandled and also lowers the risk of affected tasks being missed.

[0019] In the generation of candidate correction values, the correlation coefficient between candidate sources (such as adjacent time periods, the same server pool, the same task type, deductions from the superior meter, and the previous trusted version) and the abnormal object in the lineage graph is calculated. This allows for reverse constraints on candidate sources that appear similar but actually share the same abnormal source—eliminating those with direct similarity and reducing the weight of those with multiple similarity across different hops. This helps reduce the risk of abnormal meters, abnormal carbon factors, abnormal cooling units, or abnormal scheduling versions re-entering the release chain through the candidate correction process. The candidate correction values ​​are written to the candidate field of the isolated data table and stored alongside the original observation data, the previous trusted version, and the correction source, which also helps maintain the traceability of the original data without disrupting the data lineage. Inputting the candidate correction values ​​into the corresponding physical consistency verification gate according to the abnormality type allows for closed-loop verification using the energy conservation channel or carbon factor time series channel before the correction values ​​enter the new release state. Furthermore, the verification threshold is stricter than the initial high-abnormality threshold, helping to avoid situations where the correction values ​​are only statistically usable but still unreleaseable in terms of physical consistency. For boundary conditions such as failed re-verification, zero sum of candidate weights, and zero median absolute deviation, the iterative elimination, stopping automatic correction, maintaining isolation or freezing, rolling back to the previous trusted version, and switching historical adaptive thresholds are implemented to help avoid state oscillations or infinite loops caused by repeated elimination and recalculation. These effects can be achieved in various implementations, including methods, systems, electronic devices, and computer-readable storage media. It should be noted that these effects are optimizations of the data quality control process under corresponding constraints, not a complete elimination of any anomalies; the specific effects may vary under different data center environments and parameter calibrations.

[0020] The specification of this application contains numerous technical features distributed across various technical solutions. Listing all possible combinations of these technical features (i.e., technical solutions) would make the specification excessively lengthy. To avoid this problem, the various technical features disclosed in the above-described invention, the various technical features disclosed in the following embodiments and examples, and the various technical features disclosed in the accompanying drawings can be freely combined to form various new technical solutions (all of which should be considered as described in this specification), unless such a combination of technical features is technically infeasible. For example, one example discloses feature A+B+C, and another example discloses feature A+B+D+E. Features C and D are equivalent technical means that serve the same function, and technically only one needs to be used; they cannot be used simultaneously. Feature E can technically be combined with feature C. Therefore, the solution A+B+C+D should not be considered as described because it is technically infeasible, while the solution A+B+C+E should be considered as described. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating an embodiment of the collaborative data quality control method for carbon dioxide based on anomaly identification according to this application.

[0022] Figure 2 This is a schematic diagram of the structure of an anomaly-identification-based collaborative data quality control system for carbon dioxide calculation according to an embodiment of this application. Detailed Implementation

[0023] In the following description, many technical details are presented to help the reader better understand this application. However, those skilled in the art will understand that the technical solutions claimed in this application can be implemented even without these technical details and various variations and modifications based on the following embodiments.

[0024] The following is combined Figure 1 This application describes the overall technical concept.

[0025] In summary, this application does not simply attach anomaly detection results in the conventional sense to the data center carbon footprint database, but rather... Figure 1 In the processing flow shown, the object In time slice The server power signal is obtained from the multi-source data acquisition interface. Electricity meter power signal Unattended power signal Cooling power signal, carbon factor version signal And the scheduling bloodline signal is first converted into an energy consumption balance deviation signal with electrochemical physical meaning. Carbon factor time series deviation signal Cooling coupling deviation signal Task log integrity deviation signal Risk signals related to missing bloodline These signals are then aggregated into a comprehensive anomaly scoring signal. This score directly drives the migration of objects in the release status table between isolated, frozen, candidate correction, new release, and rollback states. This process unifies the physical metering chain, carbon factor call chain, and task scheduling chain where the anomaly occurs into the same database status control chain.

[0026] Furthermore, for objects that have entered an isolated or frozen state, this application does not directly replace the original values ​​with data from adjacent or similar operating conditions. Instead, it determines the set of affected tasks along the task execution relationships, meter coverage relationships, carbon factor call relationships, scheduling allocation relationships, and cooling distribution relationships in the lineage diagram, and forms a task impact signal accordingly. With the recalculation queue; in generating candidate correction value signals At that time, the correlation coefficient of homologous anomalies was used. For candidate sources Apply reverse constraints to prevent candidate sources that directly share the same abnormal meter, abnormal carbon factor, abnormal cooling unit, or abnormal scheduling version as the abnormal object from participating. The weighted generation method differs from the conventional approach of generating candidate values ​​solely based on temporal proximity or similar operating conditions. Therefore, candidate correction, lineage propagation, and state transition are not independent data processing actions, but rather form a closed loop nested around the release control objective of "whether anomalous physical entity data can enter the official version of the task-level computing power carbon footprint database": state transition defines the objects that need to be addressed, lineage propagation defines the anomalous boundary of these objects, and homologous elimination defines the credible sources that can be used for correction; and Furthermore, the physical consistency verification gates in the energy consumption or carbon factor channels must be entered according to the anomaly type. Only after the verification result meets the corresponding threshold constraints will the carbon footprint recalculation of the corresponding task in the new release version and recalculation queue be triggered. This continuous coupling from physical entity signals, database status signals, lineage influence signals to candidate correction release signals distinguishes this application from schemes that only generate alarms or only perform general data cleaning. This forms an internally closely related technical closed loop in the scenario of releasing the carbon footprint of data center task-level computing power, and helps to reduce the risk of abnormal physical data entering the official carbon footprint version.

[0027] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application are described clearly and completely below in conjunction with specific embodiments. Obviously, the described embodiments are only a part of the embodiments of this application, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application. It should be noted that, without departing from the main technical line of this application—"multi-source carbon calculation of physical entity data access, anomaly indicator calculation, database release state migration, lineage influence propagation, candidate correction, and physical consistency re-verification"—those skilled in the art can make adaptive adjustments to specific thresholds, time slice lengths, field names, data table names, or module deployment methods.

[0028] It should be noted that the object processed in this embodiment is not abstract data detached from physical quantities, but rather physical entity data of electrical carbon calculation that has definite physical and engineering meaning generated during the operation of the data center. Electrical carbon calculation refers to a collaborative system of three types of technical data coupled together during the data center task-level computing power carbon footprint accounting process: electricity (electricity consumption and distribution metering), carbon (carbon factors and carbon footprint results), and computing (computing power tasks and their resource consumption). The object refers to the smallest data unit included in quality control, such as the observation set of a certain electricity meter coverage area, server pool, rack group, cooling distribution unit, carbon factor version, or task carbon footprint result within a certain time slice. The time slice refers to a discrete time unit for collecting and aligning the above physical entity data, which can be exemplarily taken as fifteen minutes, but this application is not limited to this specific value.

[0029] This embodiment provides a collaborative data quality control method for electricity carbon footprint based on anomaly identification. This method is applied to an electronic device including a processor, memory, and multi-source data acquisition interfaces to control the release status of a data center task-level computing power carbon footprint database. The task-level computing power carbon footprint database is not only used to store report results, but also to store database objects and their release status related to task execution, server power, electricity metering, cooling allocation, carbon factor version, and scheduling lineage. Specifically, the database includes, exemplarily, a raw observation table, an anomaly index table, an anomaly event table, an isolation data table, a release status table, and an impact task table: the raw observation table stores the object, time slice, observation value, source, and quality marker; the anomaly index table stores the object's index type, deviation value, threshold, and contribution; the anomaly event table stores anomaly events, anomaly objects, anomaly levels, main causes, triggering rules, and event status; the isolation data table stores the original value, candidate corrected value, previous confidence value, and review status; the release status table stores the object version and its release, freeze, isolation, and rollback status; and the impact task table stores anomaly events, affected tasks, carbon footprint impact, and recalculation status. The following description will first focus on method embodiments, and then provide corresponding system embodiments, electronic device embodiments, and computer-readable storage medium embodiments.

[0030] I. Method Implementation Examples like Figure 1 As shown, the method in this embodiment includes steps 100 to 700.

[0031] Step 100: Connect to multi-source carbon computing physical entity data In step 100, the multi-source electrical carbon computing physical entity data of the object within the time slice is accessed through the multi-source data acquisition interface. This multi-source electrical carbon computing physical entity data includes at least: server power collected by the Server Baseboard Management Controller (BMC) or Intelligent Platform Management Interface (IPMI), electricity meter power measured by the Distribution Unit (PDU), cooling power from the cooling allocation unit, carbon factor version, and scheduling records characterizing data lineage. Furthermore, the scheduling records can characterize which server or server pool the task is executed on, which electricity meter covers that server or server pool, which carbon factor version the task calls, and how the cooling power is allocated to the corresponding server pool, thus forming a data lineage foundation connecting electricity, carbon, and computing. Therefore, the data processed by this method has a clear physical or engineering origin, rather than abstract data detached from the data center operating object.

[0032] Specifically, step 100 may further include step 110. In step 110, the electronic device writes the server power, meter power, cooling power, carbon factor version, and scheduling record within the same time slice into the original observation table, and records the source identifier, collection time, object identifier, and quality tag for each observation data. For example, server power data can be associated with a server identifier or server pool identifier, meter power data can be associated with a distribution unit identifier, cooling power can be associated with a cooling allocation unit identifier, carbon factor version can be associated with a region, time slice, and version number, and scheduling record can be associated with a task identifier, executing server, scheduling version, and task start and end time. It should be noted that the above fields are not used to form management rules, but are used for subsequent calculation of abnormal indicators, construction of lineage graphs, and control of database release status. For example, each collection result is aligned according to a unified time slice, so that the electricity, carbon, and computation data within the same time slice are comparable in the time dimension. Furthermore, when key observation data within a certain time slice cannot be aligned, or is missing, causing the corresponding abnormal indicator to be mathematically uncalculate (e.g., the meter power as the denominator is zero or missing), the electronic device does not generate an abnormal score for that indicator. Instead, it records a quality mark for the object and writes it into the pending review record. It should be noted that this processing only applies to degradation cases where the indicator itself cannot be calculated. For missing or under-collected cases that can be normally characterized by task log completeness deviation or energy consumption balance deviation, the subsequent abnormal scoring and release status migration process will still be followed. Step 100 corresponds to... Figure 1 The multi-source data access link in the process.

[0033] Step 200: Calculate multiple types of abnormal indicators for carbon dioxide and summarize them into a comprehensive abnormality score. In step 200, based on the multi-source carbon computing physical entity data accessed in step 100, multiple types of carbon computing anomaly indicators for the object within the time slice are calculated, and these are summarized into a comprehensive anomaly score. The calculation results are written into the anomaly indicator table. The multiple types of carbon computing anomaly indicators include at least energy consumption balance deviation characterizing energy conservation or carbon factor time-series deviation characterizing time-series jumps, and may also include task log integrity deviation, cooling coupling deviation, and lineage loss risk deviation. Step 200 corresponds to... Figure 1 The abnormal indicator calculation and comprehensive abnormal score generation process.

[0034] Specifically, in step 210, the energy consumption balance deviation is calculated. The physical essence of this deviation is a test of whether the natural law of energy conservation holds true between the distribution side and the load side: under normal circumstances, the actual power measured by the meter should be basically balanced with the total power of the servers within its coverage area and the unassigned power. Once a sudden metering spike, missed measurement, or incorrect assignment occurs, this balance is disrupted. Its calculation method can be expressed as: in, Indicates the energy consumption balance deviation; This indicates the power measured by the power distribution unit within the time slice; This represents the total server power collected by the server baseboard management controller or intelligent platform management interface within the coverage area of ​​this power distribution unit; This indicates unassigned power that could not be clearly attributed to a specific server or task. When counted as electricity consumption, , and The electricity consumption should be converted to the amount of electricity consumed within the same time slice before being substituted into the proportional deviation formula of the same form to maintain dimensional consistency. The technical implication of this formula is that it identifies anomalies by analyzing the energy balance between meter readings and server-side data collection, rather than simply making statistical judgments based on a single value. The larger the value, the more unbalanced the distribution side and the load side are, and the more likely there are anomalies in the energy consumption data for that time slice.

[0035] Calculate the carbon factor time series deviation in step 220 This bias is used to identify sudden jumps in the carbon factor version over time, i.e., situations where the carbon factor value shows a significant jump at a certain time slice when no corresponding changes are shown in the publicly available grid data and adjacent time periods. Its calculation method can be expressed as: in, Indicates the time-series deviation of the carbon factor; This indicates the carbon factor version value for the current time slice; Indicates the current time slice The set of carbon factor versions within adjacent windows formed by taking a number of time slices centered on the center, forward, and backward; This represents the median of the carbon factor version values ​​within the adjacent window; This represents the median absolute deviation of the carbon factor version values ​​within the adjacent window. Using the median and median absolute deviation instead of the mean and standard deviation is to reduce the impact of individual extreme values ​​on the reference benchmark, making abrupt jumps more robust. This indicator is further used to control whether the task carbon footprint results calling this carbon factor version can enter the formal release state.

[0036] It should be noted that, in the first round of calculation of the time series deviation of the carbon factor, when When the value is below the preset minimum positive number, the electronic device expands the range of the adjacent window, or switches to a historical adaptive threshold or absolute deviation threshold based on historical reliable samples for judgment. Its processing mechanism is consistent with the degradation processing in step 730 for the re-verification stage. If an effective judgment benchmark still cannot be formed, the abnormal score of the indicator is not generated, the object is written into the record to be reviewed and the original release status is maintained.

[0037] In step 230, the remaining anomaly indicators for the electrical carbon calculation are calculated. This includes the task log completeness deviation. Based on the task start time, end time, task identifier, execution server identifier, resource utilization records, and whether there are any missing, duplicate, time-inverted, or uncorrelated conditions among scheduling versions, a normalized deviation value can be generated, which is exemplified as follows: in, Indicates the number of missing fields. Indicates the number of duplicate records. This indicates the number of records where the task start and end times are reversed or the time slices are not consecutive. This indicates the number of records that could not be associated with the execution server, scheduling version, or carbon factor version. This indicates the number of records that the task log should have within the corresponding time slice.

[0038] For cooling coupling deviation The correspondence between cooling power and server power can be determined based on historical reliable time slices, and the cooling power of the current time slice can be compared with the reference cooling power under this correspondence, as exemplified as follows: in, This indicates the cooling power of the current time slice. This represents the reference cooling power determined by historical reliable samples or operation and maintenance rules. and For target Calibrated parameters This is a preset minimum positive number to avoid the denominator being zero.

[0039] For the risk bias of missing bloodline Deviation values ​​can be generated based on whether there is a traceable correlation between tasks, servers, electricity meters, cooling allocation units, carbon factor versions, and scheduling versions, as exemplified as follows: in, This indicates the number of valid edges that have been successfully established for the current object in the lineage graph. This indicates the number of edges that should be established according to the object type and exception type; if If the electronic device does not automatically publish a quality conclusion for the object, it will maintain the original publication status and generate a review record. It should be noted that all the above deviations are calculated from specific technical fields and are related to the physical or engineering links upon which the data center task-level computing power carbon footprint calculation relies, rather than being arbitrary values.

[0040] In step 240, the electronic device summarizes the various electrical carbon calculation anomaly indicators into a comprehensive anomaly score. Its calculation method can be expressed as: in, Representation Object In time slice The overall abnormality score; to These are the weights corresponding to task log completeness deviation, energy consumption balance deviation, cooling coupling deviation, carbon factor timing deviation, and lineage loss risk deviation, respectively. For example, to The weights are non-negative and can satisfy the normalization condition. Each weight can be determined based on object type, anomaly type, historical reliable samples, or operational rules, and different configurations can be adopted across different object types. Furthermore, it is correlated with the degree of impact of the corresponding anomaly on the reliability of the carbon footprint results. This ensures that the comprehensive anomaly score is not a purely mathematical weighted formula with arbitrary values, but rather directly corresponds to the data quality technical issues to be addressed. For example, for meter coverage objects, energy consumption balance deviation can have a higher contribution; for carbon factor version objects, carbon factor time-series deviation can have a higher contribution; and for task carbon footprint result objects, the contribution of lineage loss risk deviation and task log completeness deviation can be increased.

[0041] It should be noted that the specific judgment logic of the above-mentioned abnormal indicators can be implemented by various interpretable mechanisms and can be selected and developed as follows: Optionally, energy consumption balance deviation and carbon factor time series deviation can be directly judged according to the rules and thresholds of Formula 1 and Formula 2; Optionally, the interquartile range method can be used to define the normal interval with the upper and lower quartiles of historical samples; Optionally, isolated forest can be used to score the anomalies of multidimensional features; Optionally, a time series prediction model can be used, with the difference between the predicted value and the measured value as the deviation; Optionally, a graph model can be used to propagate the anomaly reliability on the pedigree map. This embodiment preferentially uses interpretable mechanisms such as rules, thresholds, median absolute deviation, and state machines as the main implementation method, and machine learning models are only used as optional implementations, so that the solution can be clearly understood and implemented by those skilled in the art.

[0042] Step 300: Control the database release status based on the anomaly level. In step 300, based on the anomaly level of the comprehensive anomaly score described in step 240, the release status of the object in the release status table is controlled. This step differs from general anomaly detection in that the anomaly identification result is not merely a warning, but is used to control the release status migration of the underlying database object. Specifically, when When the value is greater than or equal to the high anomaly threshold, the electronic device will transition the object's published state to an isolated or frozen state; when When the threshold is less than the high anomaly threshold and greater than or equal to the medium anomaly threshold, the state transitions to a candidate correction state; when... When the anomaly level is below the medium anomaly threshold, the data is moved to a tagged publication status. Therefore, the higher the anomaly level, the more restricted the object's publication status, and the stronger the barrier to entry of anomalous data into the official carbon footprint dataset. Conversely, higher non-anomaly levels require stricter manual review. Step 300 corresponds to... Figure 1 The release status migration process.

[0043] Furthermore, objects in the candidate correction state do not directly enter the official release version; their candidate values ​​can be temporarily stored in the candidate field. When the comprehensive anomaly score of the object rises to the high anomaly level, or when it is confirmed after review that correction is required, it will then enter the candidate correction process corresponding to the isolation state or the frozen state, i.e., execute subsequent steps 500 to 600.

[0044] Furthermore, the methods for determining the high anomaly threshold, medium anomaly threshold, and threshold used for release verification may include: static thresholds, used for hard constraints, such as power not being negative; historical adaptive thresholds, determined by the absolute deviation between the median and the median in a sliding window; and business-level thresholds based on task release requirements, determined according to the task's report level. It should be noted that stricter anomaly thresholds can be used for data objects that require formal release versions, while relatively lenient thresholds can be used for intermediate data objects used only for internal analysis. That is, the high anomaly threshold used for formal disclosure tasks is stricter than the high anomaly threshold used for internal tasks, which can be expressed as follows: The technical reason is that once the carbon footprint results of a formally disclosed task are published, they are difficult to retract. Using stricter thresholds can block potential abnormal data at an earlier stage, thereby exercising stronger control over the publication process, rather than simply converting the level of abnormality into the intensity of manual approval.

[0045] In step 310, the electronic device can classify the object state into several states: raw, candidate, published, quarantined, corrected, frozen, and rolledback. Specifically, the raw state represents the data that has been initially accessed but not yet published; the candidate state represents the data that has entered the candidate correction process; the published state represents the data that has generated a formally released version; the quarantined state represents the data that has been isolated due to energy balance anomalies, server power anomalies, or other physical quantity anomalies; the frozen state represents the data that has been frozen due to carbon factor version anomalies, cooling distribution anomalies, or call chain anomalies; the corrected state represents the state where candidate corrected values ​​have been generated and passed the corresponding verification but a new release version has not yet been formed; and the rolledback state represents the state that reverts to the previous trusted version when re-verification or review fails. Through this state transition, the anomaly identification results no longer remain at the alarm level but directly affect the release chain of the task-level computing power carbon footprint database.

[0046] Step 400: Locate the affected task set along the lineage graph and generate a recalculation queue. In step 400, for objects that were isolated or frozen in step 300, the electronic device propagates along the lineage graph of the task-level computing power carbon footprint database to locate the affected task set and writes the relevant results into the affected task table. Step 400 corresponds to... Figure 1 The bloodline transmission and recalculation queue generation process.

[0047] Specifically, the node types of the lineage graph include tasks, servers, rack groups, electricity meters, cooling distribution units, carbon factor versions, and scheduling versions; the edge types of the lineage graph include at least task execution relationships, electricity meter coverage relationships, carbon factor call relationships, scheduling allocation relationships, and cooling distribution relationships. In step 410, starting from the abnormal object, reachability propagation is performed along the above edge types to obtain a set of objects that are related to the abnormal object. For ease of description, abnormal events can be referred to as abnormal events. The corresponding scope of influence is expressed as follows: in, Indicates an abnormal event The set of objects affected; This represents the set of nodes that can be reached from an abnormal object along a specified edge type in the lineage graph; Indicates an abnormal event The corresponding exception object; Indicates the coverage relationship of the electricity meters; Indicates the task execution relationship; This indicates the invocation of the carbon factor relation; Indicates the scheduling and allocation relationship; This indicates the cooling distribution relationship. It should be noted that the propagation range of an anomaly along the lineage graph is determined by the anomaly type of the object. For example, the propagation range of a meter balance anomaly can be limited to the server pool covered by the same meter and its executed tasks, and the propagation range of a carbon factor timing anomaly can be limited to tasks calling that carbon factor version in the same region and time slice.

[0048] In step 420, the electronic device further calculates the percentage of power consumption affected by the anomaly for each affected task, i.e., the percentage of power consumption affected by the anomaly for each task. Task Influence Its calculation method can be expressed as: in, Indicates task The percentage of electricity consumption affected by abnormal time slots; Indicates task In time slice Electricity consumption; Represents an exception object In time slice The overall abnormality score; The threshold for abnormality is indicated; For indicator functions, when The value is 1 if the time is abnormal, and 0 otherwise. The physical meaning of this formula is: the higher the proportion of a task's electricity consumption in an abnormal time slice to its total electricity consumption, the greater the likelihood that the task's carbon footprint results will be contaminated. It should be noted that when the task... Total electricity consumption At that time, the electronic device does not calculate the task impact of the task, but marks the task as inapplicable or transfers it to review, in order to avoid the denominator being zero.

[0049] In step 430, when the electricity consumption ratio exceeds the task threshold, the electronic device updates the carbon footprint result of the corresponding task in the affected task table to a status requiring review and generates a recalculation queue so that the carbon footprint of the task can be recalculated after obtaining a reliable correction value. Furthermore, when the same task is hit by multiple anomalous events, the anomalous event number, task number, and object version can be used as deduplication criteria, and the task status can be determined according to the highest anomalous level, thereby avoiding duplicate recalculation records or repeated freezing for the same object version. This fine-grained lineage tracing, where the propagation range is determined by the annomalous type, automatically defines pollution boundaries, avoiding both freezing or recalculating irrelevant tasks and reducing the risk of missing affected tasks.

[0050] Step 500: Extract candidate sources and generate candidate correction values ​​that do not overwrite the original data. In step 500, for objects in an isolated or frozen state, the electronic device extracts candidate sources to generate candidate correction values ​​that do not overwrite the original data. The key to this step is that the candidate correction values ​​do not overwrite the original data, and during generation, the homologous anomaly correlation coefficient between each candidate source and the anomalous object is calculated. Based on this, directly homologous candidate sources are eliminated outright, and multi-hop homologous candidate sources are graded and downweighted. Step 500 corresponds to... Figure 1 The candidate correction value generation stage in the process.

[0051] Specifically, in step 510, candidate sources are extracted. These candidate sources include data from adjacent time periods, data from the same server pool, data from the same task type, deduction data from the superior meter, and data from the previous trusted version. Each candidate source retains its source type, source time slice, source version, and its associated task, meter, server pool, or carbon factor version. It is important to emphasize that the candidate correction value does not overwrite the original outlier value; instead, it is written to the candidate field of the isolated data table and stored alongside the original observation data, the previous trusted version data, the correction source, and the version record.

[0052] In step 520, for each candidate source Calculate the homology correlation coefficient between it and the anomalous object in the lineage diagram. Its calculation method can be expressed as: in, Indicates the first The correlation coefficient of the same source anomaly among candidate sources; direct same source means that the candidate source and the anomaly object share the same abnormal meter, abnormal carbon factor, abnormal cooling unit or abnormal scheduling version. This represents the one-hop homology correlation coefficient; Let f(x) represent the multi-hop homology association coefficient, and satisfy the following conditions: In this application, the term "multi-hop homology" is used to summarize candidate sources by bloodline. Figure 1 Indirect same-origin cases where an exception object is connected by one or more indirect relationships (one hop or more), where one hop same-origin is used... For multi-hop homogeneous systems with more than one hop, the following approach is adopted: Both belong to the category of indirect homologous sources that should be graded and weighted down. The design motivation for this grading is that the closer the candidate source is to the anomalous object, the greater the possibility that it itself has been contaminated by the same anomalous source, and therefore it should be assigned a higher correlation coefficient to suppress it further. When the candidate source is considered a directly homologous candidate source, the electronic device will reject it outright and exclude it from the weighted calculation of the candidate correction value; when When the candidate source is classified and weighted according to its degree of homology, the weighting is reduced; when In this case, the candidate source can participate in the calculation normally according to source credibility, time proximity, and operating condition similarity. In this way, the system avoids introducing candidate sources that share the same anomaly source as the anomaly object into the correction value again.

[0053] In step 530, the electronic device calculates the comprehensive weight of each candidate source based on the homologous anomaly correlation coefficient. Its calculation method can be expressed as: in, Indicates the first The overall weight of each candidate source; Indicates the candidate source type The baseline credibility, for example, is that the baseline credibility of the deduction by the superior meter is higher than that of the same server pool, the same server pool is higher than that of the adjacent time period, and the adjacent time period is higher than the previous trusted version. Indicating temporal proximity, the closer the time slice of the candidate source is to the time slice of the anomaly, the better. The larger; The similarity of operating conditions can be determined based on server pool consistency and task type consistency, and may be further combined with load range consistency if necessary. This represents a homology repressor. Therefore, direct homology candidate sources... And make If the value is zero, it does not participate in the weighting; one-hop or multi-hop same-source candidate source factors This reduces its contribution. As can be seen from Formula 7, the overall weight is determined by the specific physical basis, rather than being arbitrarily set.

[0054] In step 540, the electronic device generates candidate correction values ​​based on the comprehensive weights. Its calculation method can be expressed as: in, Indicates the candidate correction value; Indicates the first Candidate values ​​corresponding to each candidate source. The generated candidate correction values ​​are written to the candidate field of the isolated data table and stored alongside the original observation data of the object, the previous trusted version data, and the correction source, thus ensuring that the original data is not overwritten and the correction process is traceable. It should be noted that if the sum of the comprehensive weights of all candidate sources... In this case, the electronic device does not generate an automatic correction value, but instead keeps the object in an isolated or frozen state and generates a review task to avoid forcibly issuing a correction value when all candidate sources are contaminated by the same source anomaly.

[0055] Step 600: Perform physical consistency re-verification and control release according to anomaly type for each channel. In step 600, the electronic device, according to the anomaly type of the object, inputs the candidate correction value described in step 540 into the corresponding physical consistency verification gate for closed-loop verification. This verification gate ensures that the candidate correction value must pass physical consistency verification before release, thus differing from the prior art practice of only isolating or correcting abnormal data without physical verification. Step 600 corresponds to... Figure 1 The physical consistency verification and new release generation process.

[0056] Specifically, when the anomaly type of the object belongs to the energy consumption or meter channel, in step 610, the candidate correction value is input into the energy consumption channel verification gate to calculate the energy consumption correction verification value. It satisfies: in, This represents the energy consumption balance deviation after replacing the original meter power with the candidate correction value; This represents the total server power within the coverage area of ​​the corresponding electricity meter. Indicates unassigned power; This represents the energy consumption re-verification threshold, which is stricter than the high energy consumption anomaly threshold corresponding to the isolation state. In other words, for a candidate correction value to enter the new release state, it should meet more stringent energy consumption balance requirements than the initial isolation judgment.

[0057] When the anomaly type of the object belongs to the carbon factor channel, in step 620, the candidate correction value is input into the carbon factor channel verification gate to calculate the carbon factor correction verification value. It satisfies: in, This represents the time series deviation of the carbon factor after replacing the original outlier factor with a candidate correction factor; In this channel, the candidate correction value is represented as a candidate correction factor; This represents the carbon factor re-verification threshold, which is stricter than the high anomaly threshold corresponding to the state of freezing. A stricter re-verification threshold means that candidate corrected values ​​must not only be better than the original anomaly value that triggered isolation or freezing, but also converge to a range closer to physical equilibrium to be considered as passing the re-verification. Furthermore, when the carbon footprint result of the same task is simultaneously affected by both meter anomalies and carbon factor anomalies, the energy consumption channel re-verification gate and the carbon factor channel re-verification gate can be set in parallel. A new release version is generated only after the candidate corrected value simultaneously meets both re-verification thresholds.

[0058] In step 630, after the candidate correction value passes the physical consistency verification of the corresponding channel, the electronic device writes the candidate correction value, candidate source version, candidate weight, same-source candidate removal record, verification result, and optional review record into the audit snapshot. It then sequentially migrates the object from an isolated or frozen state to a correction state and generates a new release version in the newly released state. Simultaneously, the electronic device triggers the recalculation queue generated in step 430, causing the corresponding task to call the candidate correction value for carbon footprint recalculation. The system retains the version before recalculation, the version after recalculation, the exception event number, the correction source, and the recalculation status in the affected task table. The audit snapshot exemplarily includes the exception cause, the affected task, and the carbon footprint difference before and after correction. Therefore, candidate correction is not a replacement of the original data, but rather forms a traceable versioned correction chain.

[0059] Step 700: Iteration and Degradation Processing In step 700, for cases where the physical consistency verification gate is not passed, the electronic device performs iterative and degradation processing to help the correction process converge, become traceable, and avoid falling into a vicious cycle. Step 700 corresponds to... Figure 1 The iterative and degradation processing steps in the process.

[0060] Specifically, in step 710, when a candidate correction value fails to pass the corresponding physical consistency verification gate, the electronic device removes the current homogeneous anomaly correlation coefficient. For the largest single candidate source, return to step 530 to recalculate the candidate correction value and re-input the corresponding physical consistency verification gate, with the number of iterations not exceeding the total number of extracted candidate sources. In each round, only one source with the highest correlation coefficient is removed, making the correction process gradually traceable and avoiding information loss caused by removing too many sources at once. When multiple candidate sources have the same highest correlation coefficient, they can be selected to be removed in descending order of their residual contribution or comprehensive weight. In step 720, if the re-verification fails after the total number of candidate sources has been reached, or if the sum of the comprehensive weights of all candidate sources is insufficient, the source will be removed. When this occurs, the electronic device stops automatically generating candidate correction values, maintains the object's isolated or frozen state, reverts the object data to the previous trusted version, and triggers manual review. In step 730, for the carbon factor channel, when calculating the carbon factor correction verification value encounters... In the event of degradation, expand the range of adjacent windows or switch to historical adaptive thresholds; if an effective threshold still cannot be formed, remain frozen and do not publish automatically.

[0061] Optionally, for degradation scenarios where the meter power is zero or the candidate correction value is zero (e.g., the denominator is zero), the system will also revert to the previous trusted version without forcibly releasing the new version. Therefore, this embodiment provides clear handling methods for both critical scenarios (e.g., the correction value is exactly at the revalidation threshold boundary) and extreme scenarios (e.g., the bottleneck source has been eliminated), helping to avoid repeated oscillations between released, frozen, corrected, and rolled-back states. Optionally, to prevent repeated state oscillations, version locks, state locks, or one-time migration records can be set.

[0062] II. Exemplary Examples To facilitate understanding of the synergy between the above steps and their specific effects, exemplary calculation examples are provided below. It should be noted that the following values ​​are merely anonymized examples for algorithm illustration, used to demonstrate the calculation process of each formula and the connection between each step. The threshold values ​​involved are all exemplary and do not constitute a limitation of this application.

[0063] In an exemplary energy consumption channel calculation, during a certain time slice, the power meter reading recorded by the power distribution unit suddenly jumps to 900 kW, while the total server power collected by the baseboard management controller within the meter's coverage area is only 620 kW, leaving approximately 20 kW of unassigned power. According to Formula 1, the energy balance deviation is approximately... In this example, if the deviation reaches a high anomaly level, the electronic device will migrate the energy consumption data of the meter object or its coverage area within that time slice to an isolated state, and record the abnormal object, abnormal time slice, anomaly level, and triggering rule in the anomaly event table. Its physical balance benchmark is... kW. Subsequently, the electronic device locates the affected tasks along the meter coverage relationship and task execution relationship in the lineage diagram, and calculates the electricity consumption ratio of each task affected by the anomaly according to Formula 5. For tasks whose electricity consumption ratio exceeds the task threshold, their carbon footprint results are updated to the status of needing review and written into the recalculation queue.

[0064] In the aforementioned energy consumption anomaly example, candidate sources extracted by the electronic equipment include, for instance, the previous reliable version (approximately 640 kW, from a different source than the anomaly), adjacent time period data (approximately 642 kW), deduction data from the upstream meter (approximately 645 kW), and data from the same cabinet group during the same time period (approximately 870 kW). Among these, the data from the same cabinet group during the same time period is determined to be directly from the same source because it shares the same anomaly feeder as the anomaly, and its correlation coefficient is set to 1. For example, if this directly from the same source candidate is not eliminated and approximate equal weighting is applied, the resulting candidate correction value will be significantly increased to approximately 699 kW. Substituting this into Formula 9, the energy consumption correction verification value is approximately... The percentage exceeds the exemplary energy consumption re-verification threshold and fails; however, according to the rules of Formulas 6 and 7, after eliminating the direct source of the same energy consumption, the candidate correction value converges to approximately 642 kW. Substituting this into Formula 9, the energy consumption correction re-verification value is approximately... If the value falls within the exemplary energy consumption re-verification threshold range and passes the re-verification, the system generates a new release version accordingly, and the affected task is recalculated at approximately 642 kW. This comparison illustrates that homologous candidate sources significantly pull candidate correction values ​​away from the physical equilibrium point, preventing them from passing physical re-verification. In contrast, the homologous hierarchical elimination in this embodiment causes candidate correction values ​​to converge to near the physical equilibrium benchmark. The homologous anomaly correlation coefficient is not simply used to increase or decrease statistical weight, but rather to block homologous anomaly candidate sources from entering the correction release chain.

[0065] In another exemplary carbon factor channel calculation, the carbon factor version value of a certain region suddenly increases from the adjacent level to 1.35 in the current time slice, with a median of 0.50 in the adjacent windows and an absolute deviation of 0.16 from the median in the adjacent windows. According to Formula 2, the carbon factor time series deviation is approximately... The carbon factor version deviates significantly from the adjacent window baseline, therefore it is migrated to a frozen state. The electronic device searches for the task calling this carbon factor version along the carbon factor call relationships in the lineage graph; candidate correction factors can be extracted from adjacent time period factors, trusted factors in the same region, and the previous trusted version, and candidate sources directly related to the abnormal version are eliminated. After the candidate correction factor passes through the carbon factor channel verification gate shown in Formula 10, the electronic device generates a new release version and triggers the corresponding task in the recalculation queue to call the candidate correction factor for carbon footprint recalculation. For example, if a task's electricity consumption in the abnormal time slice is 30 kWh and its total electricity consumption is 80 kWh, then according to Formula 5, the proportion of its electricity consumption affected by the abnormality is approximately... The electronic device can use this percentage to determine whether to set the task result to a status requiring review and include it in the recalculation queue. It should be noted that in the above example, the deviations such as energy consumption balance deviation, energy consumption correction verification value, and electricity consumption ratio are expressed as percentages. When summarizing them into a comprehensive anomaly score or comparing them with the threshold, they are all included in the corresponding dimensionless values ​​according to the same conversion relationship. The carbon factor time series deviation is calculated as a multiple of the absolute deviation of the median of adjacent windows, which is itself a dimensionless quantity. The scale differences between the indicators are absorbed by the weight calibration process, so the deviations are comparable in the weighting of Formula 3.

[0066] As can be seen from the above example, steps 100 to 700 are not isolated from each other, but form a closed loop consisting of multi-source access, anomaly index calculation, comprehensive anomaly scoring, database state migration, lineage influence propagation, candidate correction generation, and physical verification and release across channels. The natural law verification represented by energy balance deviation and carbon factor temporal deviation is mapped to the release status action of the database object through comprehensive anomaly scoring; the release status action then acts back along the lineage graph to recalculate the carbon footprint of the affected tasks; and the candidate correction value converges and is released under the dual constraints of source elimination and physical verification. There are clear technical connections between the inputs, processing, and outputs of each step, thereby enabling control over the carbon footprint calculation and release process.

[0067] III. System Implementation Examples Based on the same inventive concept as the above-described method embodiments, this embodiment also provides an anomaly identification-based collaborative data quality control system for carbon footprint analysis. This system is applied to an electronic device including a processor, memory, and multi-source data acquisition interfaces, and is used to control the release status of the data center's task-level computing power carbon footprint database. For example... Figure 2 As shown, the system includes a multi-source access module, an anomaly indicator module, a state transition module, an impact set module, a candidate correction module, and a re-verification and release module. The processing performed by each module is consistent with the corresponding steps in the aforementioned method embodiments, so only a brief description is given.

[0068] The multi-source access module is configured to access the multi-source carbon calculation physical entity data of the object within the time slice through the multi-source data acquisition interface. This multi-source carbon calculation physical entity data includes at least the server power collected by the server baseboard management controller or intelligent platform management interface, the meter power measured by the power distribution unit, the cooling power from the cooling allocation unit, the carbon factor version, and scheduling records representing data lineage. The anomaly indicator module is configured to calculate multiple types of carbon calculation anomaly indicators of the object within the time slice based on the multi-source carbon calculation physical entity data and summarize them into a comprehensive anomaly score. These multiple types of carbon calculation anomaly indicators include at least energy consumption balance deviation representing energy conservation or carbon factor timing deviation representing timing jumps. The state transition module is configured to control the release status of the object in the release status table according to the anomaly level to which the comprehensive anomaly score belongs, and to transition the release status to an isolated state or a frozen state when the comprehensive anomaly score meets the high anomaly level condition. The impact set module is configured to propagate along the lineage graph for objects in isolated or frozen states to locate the affected task set, calculate the electricity consumption percentage of each affected task affected by the anomaly, and update the carbon footprint result of the corresponding task in the impact task table to a status requiring review when the electricity consumption percentage exceeds a task threshold, and generate a recalculation queue. The candidate correction module is configured to extract candidate sources for objects in isolated or frozen states to generate candidate correction values ​​that do not overwrite the original data, and calculate the homogeneity anomaly correlation coefficient between each candidate source and the object in the lineage graph during generation, thereby eliminating directly homogeneous candidate sources and downgrading the weight of multi-hop homogeneous candidate sources. The re-verification and release module is configured to input the candidate correction values ​​into the corresponding physical consistency re-verification gate for closed-loop re-verification according to the anomaly type of the object. When the candidate correction value meets the physical consistency re-verification threshold, the release state is migrated to a new release state, and the corresponding task in the recalculation queue is triggered to call the candidate correction value for carbon footprint recalculation. It should be noted that the above module division is only an example. In actual implementation, several modules can be merged or further split as needed, as long as the whole can perform the steps of the aforementioned method embodiment.

[0069] IV. Examples of Electronic Devices and Examples of Computer-Readable Storage Media This embodiment also provides an electronic device, including a processor and a memory. The memory stores an executable computer program. When the computer program is executed by the processor, it implements the steps of the anomaly-based collaborative data quality control method for carbon dioxide computation based on anomaly identification in the aforementioned method embodiments. This embodiment also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the steps of the anomaly-based collaborative data quality control method for carbon dioxide computation based on anomaly identification in the aforementioned method embodiments. Those skilled in the art will understand that the computer-readable storage medium can be any tangible medium capable of storing computer program code.

[0070] In any of the foregoing embodiments, optionally, when determining the propagation range of an anomaly along the lineage graph, the electronic device may also configure anomaly type-edge type propagation gating relationships for different anomaly types. Let an anomaly event... The exception type is The edge type in the kinship graph is Electronic devices according to the propagation gating function Determine if the corresponding edge type allows traversal: when Propagation is allowed to continue along that edge type when Propagation stops at the specified edge type. For example, meter balance anomalies propagate primarily along the meter coverage relationship and task execution relationship; carbon factor timing anomalies propagate primarily along the carbon factor invocation relationship and task time slice relationship; and cooling coupling anomalies propagate primarily along the cooling distribution relationship and server pool relationship. Furthermore, each propagation can be recorded to form a propagation path identifier. Used to indicate abnormal events To the affected tasks The propagation path follows a sequence of node versions and edge types. When the propagation path encounters a rolled-back version, a previous trusted version, or an edge type inconsistent with the current anomaly type, the electronic device terminates the propagation of that path. Therefore, the scope of anomaly impact is not indiscriminately spread according to the lineage graph, but is limited by the combined constraints of anomaly type, edge type, and object version, ensuring that the location of affected tasks, elimination of candidate sources, and generation of recalculation queues all follow a consistent propagation basis.

[0071] In any of the foregoing embodiments, optionally, when generating the recalculation queue and performing release state migration, the electronic device can also set versioned idempotent control for abnormal events, object versions, and task versions. The state migration idempotent key can be generated jointly by the abnormal event identifier, object identifier, object version identifier, target state, and time slice identifier, to limit the same abnormal event to generate only one valid state migration record for the same object version; the recalculation queue can be formed by the abnormal event identifier, task identifier, object version identifier, and candidate correction version identifier to form a deduplication key, to limit the same task to retain only one record to be recalculated for the same abnormal object version. Further, once an object has migrated from an isolated or frozen state to a newly released state, a rollback state, or a previous trusted version state, unless a new data collection source version or a new carbon factor version enters the database, the electronic device will no longer repeatedly trigger freezing, rollback, or recalculation based on the same abnormal event. As a result, a one-way migration constraint with version as the boundary is formed between the release status table, the affected task table and the isolation data table. This ensures that the exception handling maintains good traceability in scenarios where multiple events concurrently hit the same task or the same object, and helps to reduce the risk of repeated freezing, repeated recalculation or state oscillation.

[0072] In any of the foregoing embodiments, optionally, when the carbon footprint result of the same object or the same task falls into both the energy consumption channel and the carbon factor channel, in addition to requiring the candidate correction values ​​to meet the corresponding re-verification thresholds in each channel, the electronic device may, exemplarily, also apply a cross-channel joint re-verification constraint: the carbon footprint result of the object in that time slice is recalculated jointly using the candidate correction values ​​of the energy consumption channel and the candidate correction values ​​of the carbon factor channel, and compared with the carbon footprint reference value derived from adjacent reliable time slices. The relative deviation between them is used as the joint verification residual. Only when Not exceeding the joint verification threshold When the time-space migration is to a newly released state, it can be represented as: in, Indicates the candidate correction value of the energy consumption channel. With carbon factor channel candidate correction value The carbon footprint results obtained through joint recalculation This represents the carbon footprint reference value derived from adjacent reliable time slices. The value can be determined by historical reliable samples or according to operation and maintenance rules. It should be noted that the passing of the two channel re-verifications does not necessarily mean that the carbon footprint result recalculated using the candidate correction values ​​of both channels is within the physically consistent range; when... Exceed When this happens, maintain the object's isolation or frozen state and roll back according to iteration and degradation processing. When referencing the carbon footprint... ( When the value is a preset minimum positive number, the electronic device does not calculate the joint verification residual, but maintains the original verification state and generates a verification record to avoid the denominator being zero. Therefore, in the case of abnormal concurrency in dual channels, this helps to further reduce the risk of the corrected carbon footprint result deviating from physical consistency.

[0073] V. Technical Effects The above embodiments have the following technical effects. First, by establishing a hard linkage between energy consumption balance deviation and carbon factor time series deviation, which have natural law implications, and the release status of the underlying objects in the task-level computing power carbon footprint database, such as isolation, freezing, candidate correction, and new release, the results of anomaly identification no longer merely remain as alarms but directly affect the database object status and release chain, helping to reduce the risk of anomalous physical data entering the official carbon footprint version. Second, by determining the propagation range of anomalies along the heterogeneous lineage graph based on the anomaly type and automatically defining pollution boundaries and generating recalculation queues in conjunction with task impact, it helps to traceably locate affected tasks and trigger corresponding carbon footprint recalculations when anomalies occur, reducing the false freezing of irrelevant tasks and the omission of affected tasks. Third, a candidate reverse elimination mechanism with a correlation coefficient of common anomalies is introduced. Candidate sources that directly share the same anomaly source are eliminated with a single vote, and candidates with one or more hops of common source are graded and downweighted. This differs from the conventional practice of generating candidate values ​​based solely on temporal proximity or similar operating conditions. This helps to converge candidate correction values ​​to near the physical equilibrium benchmark, reducing the risk of secondary contamination from the source. Fourth, a physical consistency verification release gate is set up according to anomaly type. Candidate correction values ​​must first pass through a verification channel of energy consumption conservation or factor time series that is stricter than the initial screening threshold before they can be released. This helps to avoid secondary contamination caused by the release of correction values ​​that still do not meet the physical consistency requirements. Fifth, degradation and rollback rules are set up, such as iterative elimination not exceeding the total number of candidate sources, the sum of comprehensive weights being zero, and the absolute deviation of the median being zero. Candidate correction values ​​are not overwritten with the original observation data but are coexisted with the previous reliable version and correction source, and an audit snapshot is generated. This helps to make the correction, release, rollback, and recalculation processes have better traceability and stability. It should be noted that the above-mentioned technical effects are optimizations to the data quality control process under corresponding constraints, rather than the complete elimination of any anomalies; the specific effects may vary under different data center environments, acquisition conditions, and parameter calibrations.

[0074] The embodiments of this application are further explained below by way of example.

[0075] Optional, , satisfy The weighting can be determined based on lineage distance and edge type, with one-hop homologous candidate sources having a higher weighting than multi-hop homologous candidate sources; for example... A value between 0.5 and 0.8 is acceptable. The value can be between 0.2 and 0.5, and the specific value is determined by regression calibration of historical reliable samples or by the measured impact of operation and maintenance rules on the correction error in combination with the lineage path length.

[0076] Optionally, the high and medium anomaly thresholds can be determined based on the distribution of anomaly indicators in historical trusted time slices or configured according to operation and maintenance rules; the physical consistency re-verification threshold adopts a stricter configuration than the high anomaly threshold. For example, if the high energy consumption anomaly threshold is set to approximately 20%, then... It can be taken as approximately 5%. It can be set as a multiple of the absolute deviation of the median in adjacent windows; for task carbon footprint results that need to be officially released, a more stringent threshold than that for internal analysis data is used.

[0077] Optional, The consistency can be determined by a weighted average of three consistency factors: server pool consistency, task type consistency, and load range consistency. When the candidate source and the anomaly object belong to the same server pool and have the same task type... Take the higher value, and take the lower value only when the load ranges are close but the server pools or task types are different. The weights of each item are configured by historical trusted samples or operation and maintenance experience. When only partial consistency information can be obtained, at least one of them should be used to determine the weights.

[0078] Optionally, the time slice length is exemplarily set to fifteen minutes, with adjacent windows... For example, the time slices are taken as 4 to 12 time slices before and after, depending on the carbon factor update cycle and the frequency of grid data release.

[0079] Optional, The historical correction hit rate is sorted by source type, and for example, the deduction of the upper-level electricity meter is greater than that of the same server pool, the same server pool is greater than that of the adjacent time period, and the adjacent time period is greater than that of the previous trusted version. to The weight of anomalies is determined by normalizing their measured contribution to the deviation of carbon footprint results, with anomalies contributing more weighted higher. Optionally, when an anomaly indicator is not applicable to a specific object type, the weight of that anomaly indicator is set to zero. This is equivalent to pruning the set of indicators participating in the weighting according to the object type, without needing to define a separate indicator system for different object types. For example, for carbon factor version objects, the weights of energy balance deviation and cooling coupling deviation can be set to zero.

[0080] Optionally, the main implementation prioritizes rule-based and threshold-based judgment, that is, directly judging anomalies based on Formula 1, Formula 2, static boundary, and absolute deviation of sliding window median; when the data dimension is high or there are complex correlations, the interquartile range method or isolated forest can be used as a supplement, and the parameters such as the quantile interval or the number of trees are determined by cross-validation of historical samples.

[0081] Optionally, server power data should include at least the server identifier, server pool identifier, time slice identifier, power value, and data collection source; power distribution unit metering data should include at least the meter identifier, coverage cabinet group identifier, time slice identifier, and meter power value; carbon factor version data should include at least the area identifier, time slice identifier, version number, and carbon factor value; and scheduling records should include at least the task identifier, executing server identifier, scheduling version, and task start and end times.

[0082] Optionally, the release status table may include at least the object identifier, object version, old status, new status, status transition time, trigger exception event number, candidate correction value version, and rollback version; when an object is transitioned from an isolated or frozen state to a newly released state, the candidate source, candidate weight, same-source candidate removal record, re-verification result, and recalculation queue number are recorded simultaneously for subsequent traceability.

[0083] Optionally, when multiple abnormal events hit the same task, the task identifier, abnormal event number, and carbon footprint result version are used as deduplication keys to determine the task status according to the highest abnormality level, and multiple abnormal event numbers are retained in the affected task table to avoid repeated freezing or recalculation of the same version; if necessary, the risk score can be weighted and merged according to the contribution of each abnormal event to the electricity consumption ratio of the task.

[0084] Optionally, the carbon factor version can be associated with the time slice of a region, power grid zone, or the power supply area where the data center is located. When calculating the carbon footprint of a task, the corresponding carbon factor version is called according to the task execution time slice. When a carbon factor version is frozen and a candidate correction version is generated, the task that calls the carbon factor version recalculates the carbon footprint result of the corresponding time slice according to the new candidate correction factor.

[0085] It should be noted that in this application, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one" does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. In this application, if it refers to performing an action according to an element, it means performing the action at least according to that element, including two cases: performing the action only according to that element, and performing the action according to that element and other elements. Expressions such as "multiple," "repeatedly," and "various" include two, two times, two kinds, and more than two, more than two times, and more than two kinds.

[0086] All documents mentioned in this application are considered to be incorporated in their entirety into the disclosure of this application so that they can serve as a basis for modifications if necessary. Furthermore, it should be understood that after reading the foregoing disclosure of this application, those skilled in the art can make various alterations or modifications to this application, and these equivalent forms also fall within the scope of protection claimed in this application.

Claims

1. A method for collaborative data quality control based on anomaly identification in carbon computing, characterized in that, An electronic device comprising a processor, memory, and multi-source data acquisition interfaces, used to control the release status of a data center task-level computing power carbon footprint database, the method comprising: S1. Through the multi-source data acquisition interface, access the multi-source electrical carbon calculation physical entity data of the object within the time slice. The multi-source electrical carbon calculation physical entity data includes at least the server power collected by the server baseboard management controller or intelligent platform management interface, the meter power measured by the power distribution unit, the cooling power from the cooling allocation unit, the carbon factor version, and the scheduling record characterizing the data lineage. S2. Based on the multi-source carbon computing physical entity data mentioned in S1, calculate the various types of carbon computing anomaly indicators of the object within the time slice and summarize them into a comprehensive anomaly score. The various types of carbon computing anomaly indicators include at least the energy consumption balance deviation characterizing energy conservation or the carbon factor time series deviation characterizing time series jump. S3. Based on the anomaly level of the comprehensive anomaly score mentioned in S2, control the release status of the object in the release status table of the task-level computing power carbon footprint database, and when the comprehensive anomaly score meets the high anomaly level condition, migrate the release status to the isolation status or the frozen status. S4. For the objects that are in the isolated or frozen state in S3, propagate along the lineage graph of the task-level computing power carbon footprint database to locate the affected task set, calculate the electricity consumption ratio of each affected task affected by the anomaly, and when the electricity consumption ratio exceeds the task threshold, update the carbon footprint result of the corresponding task in the affected task table to the state that needs to be reviewed and generate a recalculation queue. S5. For the object in the isolated or frozen state, extract candidate sources to generate candidate correction values ​​that do not cover the original data, and calculate the homology abnormal correlation coefficient between each candidate source and the object in the lineage diagram in S4 during generation. Based on this, candidate sources that are directly homologous are eliminated and candidate sources that are multi-hop homologous are downgraded in weight. S6. According to the anomaly type of the object, input the candidate correction value in S5 into the corresponding physical consistency verification gate for closed-loop verification. When the candidate correction value meets the physical consistency verification threshold, migrate the release state to the new release state and trigger the corresponding task in the recalculation queue in S4 to call the candidate correction value to recalculate the carbon footprint.

2. The method according to claim 1, characterized in that, The energy consumption balance deviation mentioned in S2 Time deviation of carbon factor The calculation formulas are as follows: in, This indicates the power measured by the power distribution unit within the time slice. This indicates the total power consumption of servers within the coverage area of ​​this electricity meter. Indicates unassigned power; This indicates the carbon factor version value of the current time slice. This represents the median value of the carbon factor version within adjacent windows centered on the current time slice. This represents the median absolute deviation of the carbon factor version values ​​within the adjacent windows.

3. The method according to claim 1, characterized in that, The comprehensive anomaly score described in S2 The calculation formula is: in, This indicates a deviation in the completeness of the task log. This indicates the energy consumption balance deviation. Indicates cooling coupling deviation. This indicates the time-series deviation of the carbon factor. This indicates a risk bias related to missing bloodlines. to These are the weights of the corresponding abnormal indicators; The rule for determining the anomaly level described in S3 is: when When the threshold is greater than or equal to the high anomaly threshold, the system migrates to an isolated or frozen state; when... When the value is less than the high anomaly threshold and greater than or equal to the medium anomaly threshold, the state transitions to a candidate correction state; when... When the value is less than the medium anomaly threshold, the status is transitioned to a marked release state; wherein, the high anomaly threshold used for formal disclosure tasks is stricter than the high anomaly threshold used for internal tasks.

4. The method according to claim 1, characterized in that, The node types of the lineage graph described in S4 include tasks, servers, rack groups, electricity meters, cooling distribution units, carbon factor versions, and scheduling versions. The edge types include at least task execution relationships, electricity meter coverage relationships, carbon factor call relationships, scheduling allocation relationships, and cooling distribution relationships. The electricity consumption percentage is for tasks. Task Influence The calculation formula is as follows: in, Indicates task In time slice Electricity consumption Represents an exception object In time slice The aforementioned comprehensive anomaly score, Indicates the abnormal threshold, For indicator functions, when The value is 1 if the condition is met, and 0 otherwise. The propagation range of the anomaly along the lineage graph is determined based on the anomaly type of the object. Specifically, the propagation range of the meter balance anomaly is the server pool covered by the same meter, and the propagation range of the carbon factor timing anomaly is the task in the same region and time slice.

5. The method according to claim 1, characterized in that, The candidate sources mentioned in S5 include adjacent time period data, data from the same server pool, data of the same task type, deduction data from the superior meter, and data from the previous trusted version; Wherein, direct homology represents the candidate source. Sharing the same abnormal meter, abnormal carbon factor, abnormal cooling unit, or abnormal scheduling version with the aforementioned object, and ; The comprehensive weight is calculated based on the aforementioned homologous anomaly correlation coefficient. And generate the candidate correction value. : in, Candidate source type The baseline credibility, For time proximity, The similarity of operating conditions is determined based on server pool consistency and task type consistency; the candidate correction value Candidate fields are written into the isolated data table and stored alongside the object's original observation data, the previous trusted version data, and the source of the correction. An audit snapshot is generated after a new release, which includes the cause of the anomaly, the affected tasks, and the difference in carbon footprint before and after the correction.

6. The method according to claim 1, characterized in that, The step S6, which involves inputting the corresponding physical consistency verification gate based on the anomaly type of the object, includes: When the anomaly type of the object belongs to the energy consumption or meter channel, the candidate correction value is input into the energy consumption channel verification gate to calculate the energy consumption correction verification value. : When the anomaly type of the object belongs to the carbon factor channel, the candidate correction value is input into the carbon factor channel verification gate to calculate the carbon factor correction verification value. : in, This is the energy consumption re-verification threshold, and it is stricter than the high energy consumption anomaly threshold corresponding to the isolation state. This is the carbon factor re-verification threshold, and it is stricter than the high anomaly threshold for carbon factors corresponding to a frozen state.

7. The method according to claim 6, characterized in that, For cases where the physical consistency verification gate is not passed, iterative and degenerate processing is also included: when the candidate correction value fails the corresponding physical consistency verification gate, the current homogeneous anomaly correlation coefficient is removed. After identifying the largest single candidate source, the process returns to recalculate the candidate correction value and performs closed-loop verification again, with the number of iterations not exceeding the total number of extracted candidate sources. ; If the iteration reaches the total number and still fails, or if the sum of the combined weights of all candidate sources is reached... When the automatic generation of the candidate correction value is stopped, the object is kept in an isolated or frozen state, and the object's data is rolled back to the previous trusted version and a manual review is triggered; when the calculation of the carbon factor correction verification value encounters... When necessary, expand the range of the adjacent windows or switch to historical adaptive thresholds.

8. A collaborative data quality control system for carbon dioxide based on anomaly identification, characterized in that, An electronic device comprising a processor, memory, and multi-source data acquisition interfaces, used to control the release status of a data center task-level computing power carbon footprint database, the system comprising: The multi-source access module is configured to access the multi-source electrical carbon calculation physical entity data of the object within the time slice through the multi-source data acquisition interface. The multi-source electrical carbon calculation physical entity data includes at least the server power collected by the server baseboard management controller or the intelligent platform management interface, the electricity meter power measured by the power distribution unit, the cooling power from the cooling allocation unit, the carbon factor version, and the scheduling record characterizing the data lineage. The anomaly index module is configured to calculate multiple types of anomaly indexes of the object within the time slice based on the multi-source electric carbon computing physical entity data and summarize them into a comprehensive anomaly score. The multiple types of electric carbon computing anomaly indexes include at least energy consumption balance deviation characterizing energy conservation or carbon factor time-series deviation characterizing time-series jumps. The state transition module is configured to control the release status of the object in the release status table of the task-level computing power carbon footprint database according to the anomaly level to which the comprehensive anomaly score belongs, and to transition the release status to an isolated state or a frozen state when the comprehensive anomaly score meets the high anomaly level condition. The impact set module is configured to, for the object in the isolated state or frozen state, propagate along the lineage graph of the task-level computing power carbon footprint database to locate the affected task set, calculate the electricity consumption ratio of each affected task affected by the anomaly, and when the electricity consumption ratio exceeds the task threshold, update the carbon footprint result of the corresponding task in the affected task table to the state that needs to be reviewed and generate a recalculation queue. The candidate correction module is configured to extract candidate sources for the object in the isolated or frozen state to generate candidate correction values ​​that do not cover the original data, and calculate the homology abnormal correlation coefficient between each candidate source and the object in the lineage diagram during generation, thereby eliminating directly homologous candidate sources and downgrading the weight of multi-hop homologous candidate sources. The re-verification release module is configured to input the candidate correction value into the corresponding physical consistency re-verification gate for closed-loop re-verification according to the anomaly type of the object. When the candidate correction value meets the physical consistency re-verification threshold, the release state is migrated to the new release state, and the corresponding task in the recalculation queue is triggered to call the candidate correction value to recalculate the carbon footprint.

9. An electronic device, characterized in that, include: processor; And a memory having an executable computer program stored thereon, which, when executed by the processor, implements the steps of the anomaly identification-based collaborative data quality control method for carbon dioxide calculation as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the electro-carbon computing collaborative data quality control method based on anomaly identification as described in any one of claims 1 to 7.

Images