Industrial enterprise carbon emission metering method and system based on multi-source data fusion

Abstract

The invention discloses an industrial enterprise carbon emission metering method and system based on multi-source data fusion, and belongs to the technical field of industrial carbon emission monitoring. The method comprises the following steps: S1, full-chain multi-source data acquisition and preprocessing; s2, carrying out hierarchical multi-source data fusion of digital twin driving; s3, constructing a dynamic factor adaptive carbon emission metering model; s4, dynamic feedback correction of reinforcement learning driving is carried out; and S5, multi-dimensional result output and decision support are carried out. According to the invention, a three-level fusion architecture driven by digital twinning and a dynamic feedback correction mechanism driven by reinforcement learning are combined, the accuracy of carbon emission measurement for industrial enterprises is improved through multi-source fusion and dynamic correction, and the adaptive ability of the system is enhanced through edge calculation and reinforcement learning. And clearer decision support is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of industrial carbon emission monitoring, and particularly relates to an industrial enterprise carbon emission metering method and system based on multi-source data fusion. BACKGROUND

[0002] With the continuous promotion of the "double carbon" strategy, as the core responsible subject of carbon emission, the accuracy, timeliness and traceability of the emission data of industrial enterprises have become the key basis for government regulation, carbon market trading and enterprise green transformation decision-making. Traditional carbon emission accounting generally uses emission factor method or mass balance method, relies on annual summary of manual account books, static emission coefficients and local sampling detection, and has some shortcomings: Only linear conversion based on energy consumption and corresponding emission factors cannot reflect nonlinear influences such as process parameter fluctuations, equipment aging and raw material composition changes, resulting in deviation of the accounting results from the true emission level; most enterprises form emission reports on an annual or quarterly basis, and the data lags for several months or even more than a year, which is difficult to support real-time carbon emission monitoring, early warning and process optimization; there is a lack of dynamic characterization of device-level, process-level and hourly-level emissions, which cannot accurately locate high-emission links and cannot provide targeted basis for deep emission reduction; the accounting of fugitive emissions (valve / flange micro-leakage, methane emission, wastewater treatment byproduct gas, etc.) and indirect emissions (purchased steam, green electricity proportion change) is rough, which is difficult to meet the needs of full-process and full-life-cycle management.

[0003] However, existing multi-source data fusion technologies mostly stay at the "data aggregation" level, and have not yet formed a complete metering system of "data cleaning-time and space alignment-trustworthy fusion-dynamic calibration-closed-loop optimization" for industrial complex processes. The main bottlenecks include: the time and space benchmarks of multi-source heterogeneous data are not unified, and the sampling frequency, dimension and precision differ greatly, direct fusion may easily cause noise amplification or signal cancellation; there is a lack of adaptive correction capability for dynamic production conditions (start / stop, load rise / fall, equipment switching), resulting in high false alarm rate of emission anomaly detection; existing algorithms mostly focus on a single industry or a single device, have poor reusability, and are difficult to extend to comprehensive industrial enterprises with long processes, many processes and complex emission sources.

[0004] Therefore, it is necessary to propose an industrial enterprise carbon emission metering method and system based on multi-source data fusion, which can improve the accuracy of carbon emission metering, enhance the real-time performance and adaptive ability of the system, and provide better decision support. SUMMARY

[0005] To address the aforementioned issues, the present invention aims to provide a method and system for measuring carbon emissions of industrial enterprises based on multi-source data fusion. By combining a three-level fusion architecture driven by digital twins with a dynamic feedback correction mechanism driven by reinforcement learning, the accuracy of carbon emission measurement for industrial enterprises is improved through multi-source fusion and dynamic correction. Edge computing and reinforcement learning enhance the system's adaptability and provide clearer decision support.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows: a method for measuring carbon emissions of industrial enterprises based on multi-source data fusion, comprising the following steps: S1. Full-chain multi-source data acquisition and preprocessing: Real-time acquisition of production equipment data, raw material or product data, environmental and emission data, supply chain and power grid data, and process and management data of industrial enterprises through terminal acquisition devices; cleaning, noise reduction, format conversion and structuring of the acquired data. S2. Digital Twin-Driven Layered Multi-Source Data Fusion: Adopting a three-tiered fusion architecture of "edge-workshop-enterprise," equipment-level data is integrated at the edge nodes using a lightweight time-series fusion model. Based on the digital twin's physical mirror of the workshop, data from multiple devices in the same process stage is aligned. Combined with a knowledge graph of "raw materials-process-emissions-supply chain," enterprise-level multi-dimensional data is linked to output a full-chain carbon emission dataset. This dataset includes basic identification fields: data ID, enterprise code, workshop number, timestamp, and emission source type; core measurement fields: carbon emissions at each stage, emission factors, and raw material or product association ID; association attribute fields: equipment operating parameter snapshot, raw material carbon content, and treatment method correction coefficient; and data traceability fields: data source identifier, data cleaning record, and model version number. S3. Construction of Dynamic Factor Adaptive Carbon Emission Measurement Model: Clarify the carbon emission measurement boundary of the entire chain of "raw material acquisition - production and processing - waste disposal - product storage and transportation", establish a measurement model with custom formulas for different links, and dynamically update the model factor weights and formula structure based on fused data; S4. Dynamic feedback correction driven by reinforcement learning: The deviation rate is calculated by comparing the model prediction value with the measured value at the edge. The reward function is to minimize the deviation rate and minimize the adjustment cost. The correction strategy is output through the reinforcement learning agent to dynamically adjust the model parameters and generate process optimization suggestions. S5. Multi-dimensional output and decision support: Generate enterprise-level and workshop-level carbon emission reports and carbon cost analysis data based on the revised model, and visualize the dynamic flow of carbon emissions through a digital twin dashboard.

[0007] The basic solution is based on the following principle: The core principle is to achieve accurate measurement, dynamic adaptation and carbon reduction support for industrial enterprises through the full-chain data integration, intelligent hierarchical fusion, dynamic model adaptation and closed-loop optimization decision-making.

[0008] First, let's discuss the principles behind building a comprehensive data foundation. Industrial carbon emissions exhibit significant characteristics of multi-stage coupling and multi-factor driving forces. Raw material carbon content, production process parameters, and supply chain transportation methods all directly impact emission results. Traditional partial data collection methods often lead to measurement errors due to the omission of key influencing factors. This solution constructs a comprehensive data foundation through the S1 step, covering the entire lifecycle from raw material input to production transformation, emission output, and supply chain extension. Specifically, it includes five core data categories: production equipment operation and energy consumption data, raw material / product traceability and specification data, environmental waste gas, wastewater, and solid waste monitoring data, supply chain carbon footprint and real-time power grid factor data, and process documents and management log data. This achieves comprehensive coverage of carbon emission influencing factors. Simultaneously, preprocessing operations such as cleaning and denoising to eliminate abnormal interference, format conversion to unify data standards, and structured processing to convert unstructured information transform heterogeneous multi-source data into high-quality standardized data, laying a solid foundation for subsequent fusion and modeling.

[0009] Secondly, there is the principle of layered fusion. Industrial enterprises face the dual challenges of dispersion and heterogeneity in multi-source data. Equipment data is high-frequency and real-time but has fine granularity, while supply chain data has a wide span but weak real-time performance. Unstructured data such as process documents differ significantly in format from structured data such as energy consumption, and direct fusion can easily lead to information distortion. The solution addresses this problem through a three-level fusion architecture of edge-workshop-enterprise in the S2 step: At the edge level, to meet the real-time requirements of equipment data, a lightweight time-series fusion model is used to quickly integrate single-equipment operation and energy consumption data, outputting equipment-level instantaneous carbon emission potential values ​​and ensuring low latency in data processing; At the workshop level, a physical mirror constructed by digital twins is used as an anchor point to accurately align data from multiple equipment in the same process step according to the production rhythm, achieving a one-to-one mapping between physical entities and data signals, and solving the problem of time-series misalignment of multi-equipment data; At the enterprise level, relying on the knowledge graph of raw materials-process-emissions-supply chain, workshop data is associated with cross-domain dimensional information to explore implicit coupling relationships such as raw material carbon content and reaction temperature, ultimately outputting a full-chain carbon emission dataset that connects material flow, energy flow, and carbon flow.

[0010] Furthermore, there is the principle of dynamic measurement modeling. Traditional carbon emission measurement models often rely on fixed industry average factors or single calculation formulas, making it difficult to adapt to dynamic scenarios such as fluctuations in raw material batches, adjustments in production processes, and changes in load conditions in industrial enterprises, which can easily lead to measurement errors. The solution constructs a dynamic factor adaptive model through the S3 steps: In terms of boundary definition, it clearly defines the entire chain from raw material acquisition to production and processing, waste disposal, and product storage and transportation, covering various carbon emission types such as implicit emissions, direct emissions, secondary emissions, and extended emissions, avoiding omissions in any link; in terms of formula design, it provides customizable link-specific formula templates to adapt to the process characteristics of multiple industries such as chemical and metallurgical, taking into account the differences in carbon emission mechanisms in different industrial links; in terms of parameter updates, based on the data results of hierarchical fusion output, it adjusts the model factor weights and formula structure in real time. For example, when the carbon content of raw materials increases, the weights of raw material link factors are automatically increased, and the emission calculation logic is switched synchronously when the process is optimized, achieving real-time adaptation of the model to actual operating conditions.

[0011] Furthermore, there is the principle of closed-loop optimization. Traditional feedback correction modes often employ threshold triggering plus manual inspection mechanisms, which have limitations such as response lag, high adjustment costs, and optimization of only the model without relating it to the process. The proposed solution achieves proactive closed-loop optimization through reinforcement learning-driven correction in four steps: real-time comparison of model predictions and edge-side measured values ​​to calculate the deviation rate; the reinforcement learning agent uses minimizing measurement deviation and minimizing adjustment costs as dual-objective reward functions, autonomously learning and outputting the optimal correction strategy; the correction action covers two dimensions, dynamically adjusting model parameters to improve measurement accuracy and generating targeted process optimization suggestions, realizing the linkage between measurement model optimization and actual carbon reduction actions; for extreme scenarios where deviations continuously exceed the threshold, digital twin simulation is triggered to verify the effectiveness of the correction scheme, avoiding production risks caused by blind adjustments and ensuring the safety and reliability of the optimization process.

[0012] Ultimately, it's about the principle of value transformation. The core value of carbon emission measurement lies in providing data support for corporate carbon management decisions, rather than simply outputting numerical results. The solution achieves the value transformation of measurement data through S5 steps: It generates layered results such as enterprise-level total carbon emission / intensity reports, workshop-level high-carbon process location reports, and carbon cost analysis data, tailored to the needs of different management levels within the enterprise. These are adapted to scenarios such as senior management compliance decisions, workshop process optimization, and financial cost control. Furthermore, it uses a digital twin dashboard to visualize the dynamic flow of carbon emissions, transforming abstract data into an intuitive physical spatial mapping. This helps managers quickly locate high-carbon processes, assess optimization effects, and achieve efficient transformation from measurement results to carbon reduction actions and management improvement.

[0013] In summary, this solution forms a complete closed-loop principle of "full-domain data collection (full coverage of elements) - hierarchical intelligent fusion (strong data correlation) - dynamic model adaptation (quasi-mapping of measurement) - reinforcement learning optimization (self-evolution of accuracy) - multi-dimensional decision output (real value realization)". Through deep collaboration between data, algorithms and physical entities, it solves the core pain points of inaccurate carbon emission measurement, poor adaptability and disconnect from carbon reduction in industrial enterprises, and achieves the integrated goal of accurate measurement and carbon reduction support.

[0014] Furthermore, in step S1: Production equipment data includes operating parameters and energy consumption data of machine tools and reactors, which are collected through the OPC UA protocol and broken down into equipment-level energy consumption using a load decomposition algorithm; Environmental and emission data include data related to exhaust gas, wastewater, and solid waste, which are collected through multi-gas sensors and noise reduction is performed using variational mode. Process and management data are structured through multi-engine OCR recognition and NLP technology parsing, and knowledge graph entity matching is used to associate raw material carbon footprint data.

[0015] Furthermore, the device-level edge fusion adopts the LSTM temporal fusion model. The input is the temporal stream of device operation parameters collected by the terminal layer, including temperature, pressure and energy consumption dimensions. The sampling frequency is 5-15s / time. The output is the device-level instantaneous carbon emission potential value and data credibility score. Workshop-level time-series fusion adopts the Transformer attention mechanism to highlight the data weight of key equipment and output the carbon emission sequence of the process. Enterprise-level integration employs graph neural networks to mine implicit relationships in data, including the coupling relationship between the carbon content of raw materials and reaction temperature.

[0016] Furthermore, in step S3: The implicit emissions from raw materials are modeled using a dynamic carbon footprint, and the emission factors are updated in real time with the Ecoinvent database and supplier data. Production process emissions are provided with customized core formulas for the chemical and metallurgical industries. The formula for the chemical industry is as follows: ; The waste disposal discharge incorporates a correction factor for the disposal method, calculated in conjunction with online solid waste weighing data.

[0017] Furthermore, in step S4: The formula for calculating the deviation rate is: The status of the merged data is updated every 15 minutes. When the deviation rate exceeds 15% twice consecutively, the digital twin simulation is triggered to screen the optimal correction scheme. The action space of a reinforcement learning agent includes model factor weight adjustment, process parameter suggestions, and raw material ratio optimization.

[0018] The beneficial effects of the above scheme are as follows: The industrial enterprise carbon emission measurement method based on multi-source data fusion effectively addresses the pain points of traditional measurement methods, such as fragmented data, rigid models, uncontrollable biases, and delayed decision-making, through a collaborative design of the entire chain of "collection-fusion-modeling-correction-decision". Step S1 utilizes multiple technologies to achieve comprehensive collection and standardized preprocessing of multi-source data, ensuring data integrity and accuracy and breaking down information blind spots. Step S2 combines a three-level fusion architecture with algorithms such as LSTM, Transformer, and digital twins to achieve low-latency data integration and implicit correlation mining, constructing a full-chain related dataset. Step S3's dynamic factor adaptive model adapts to changes in scenarios by updating factors in real time and using customized industry formulas, improving measurement accuracy. Step S4 uses reinforcement learning and digital twin simulation to achieve intelligent bias correction, balancing emission reduction effects and production economics. Step S5 empowers carbon reduction decisions through multi-dimensional reports and visualization outputs. The overall construction of a complete closed loop not only improves the comprehensiveness, accuracy, and real-time nature of measurement but also strengthens the guiding value of measurement for carbon reduction in production, providing systematic support for enterprises' low-carbon transformation.

[0019] An industrial enterprise carbon emission metering system based on multi-source data fusion includes: The terminal layer is used to sense and collect physical data related to corporate carbon emissions, and to complete the initial purification and standardization of the data. It includes gas sensors, smart meters, RFID readers and GPS terminals. The edge layer is used to provide near-field computing, storage and control capabilities near the physical data source, enabling localized real-time data processing and real-time decision feedback, including edge computing gateways and local servers; The cloud layer is used to receive and aggregate data from the edge layer for global modeling, in-depth analysis, and iterative optimization. It includes cloud servers, digital twin engines, and knowledge graph engines.

[0020] The beneficial effects of the above solution are as follows: 1. At the terminal layer, this layer achieves comprehensive and multi-dimensional perception of physical data related to carbon emissions of industrial enterprises through the combined application of diverse data collection devices such as gas sensors, smart meters, RFID readers, and GPS terminals. Gas sensors can accurately capture the concentration of greenhouse gases such as CO2 in exhaust gas; smart meters collect energy consumption data of production equipment in real time; RFID readers enable traceability management of raw materials and products; and GPS terminals track the transportation trajectory of the supply chain. These four types of devices work together to cover the core data dimensions of emission measurement, energy consumption, material flow, and supply chain extension, completely breaking the data blind spots caused by the reliance on a single data collection method in traditional metering systems and ensuring the comprehensiveness of data collection.

[0021] 2. For the edge layer, its near-field computing, storage, and control capabilities are precisely designed to meet the core requirements of industrial production for real-time performance and localization. Edge computing gateways and local servers can perform lightweight tasks such as device-level energy consumption decomposition, real-time deviation monitoring, and simple emission potential calculations near the data generation point, eliminating the need to transmit massive amounts of high-frequency data to the cloud for processing. This significantly reduces the computing load and data transmission bandwidth pressure on the cloud, while keeping data processing latency within the millisecond range acceptable for industrial scenarios. This enables the immediate output of device-level and process-level carbon emission data and rapid early warning of abnormal emissions. For example, when a reactor experiences a sudden increase in energy consumption leading to abnormal emissions, the edge layer can immediately capture and feed back to the on-site control system, buying time for production adjustments. This solves the response lag caused by data transmission and remote processing in traditional pure cloud architectures, enhancing the system's ability to adapt to dynamic changes in the production site in real time.

[0022] 3. At the cloud layer, its global modeling, in-depth analysis, and iterative optimization capabilities endow the system with core functions of full-chain insight and intelligent evolution. Supported by the powerful computing power of cloud servers, the digital twin engine can construct a full-dimensional mirror of the enterprise's physical space, integrating scattered data collected at the terminal layer with local data processed at the edge layer into a complete data chain of raw materials, equipment, processes, and emissions, enabling visualized traceability and full life-cycle measurement of carbon emissions. The knowledge graph engine can deeply mine the implicit correlations between raw material carbon content and reaction temperature, process parameters and emission intensity, providing interpretable logical support for the measurement model. Simultaneously, the cloud layer can conduct reinforcement learning model training and iterative optimization based on long-term accumulated data, distributing global optimization strategies (such as raw material ratio adjustments and cross-workshop process collaboration) to the edge and terminal layers, forming a closed loop of collection-processing-optimization-feedback. This allows the system's measurement accuracy to continuously improve over time, overcoming the limitations of traditional static measurement systems with fixed accuracy and difficulty in evolution.

[0023] Furthermore, it also includes a data access module, which supports OPC UA / Modbus protocol for interfacing with industrial equipment, API interface for interfacing with ERP and power grid platforms, and OCR / NLP for processing unstructured documents. This module is used to realize the automated, standardized access and governance of multi-source heterogeneous data.

[0024] The beneficial effects of the above solution are as follows: This data access module supports OPC UA / Modbus protocol to connect with industrial equipment, API interface to connect ERP and power grid platforms, and OCR / NLP to process unstructured documents, realizing automated and standardized access and governance of multi-source heterogeneous data; it effectively breaks down the barriers between equipment, systems and data forms, activates the value of implicit data such as process documents, reduces manual intervention costs and improves data quality, and provides high-quality input for real-time processing at the edge layer and intelligent modeling at the cloud layer to enhance the collaborative efficiency of each level of the system. At the same time, it significantly enhances the adaptability and scalability of the metering system to different industrial scenarios.

[0025] Furthermore, it also includes a digital twin fusion module, which is used to construct a digital mirror of the enterprise's physical space, map equipment operating status and carbon emission data in real time, and support the visualization adjustment and simulation verification of process parameters.

[0026] The beneficial effects of the above solution are as follows: By constructing a digital mirror of the enterprise's physical space, the scattered equipment operating status and carbon emission data are mapped and integrated in real time, effectively breaking down the fragmentation of carbon emission information caused by "data silos." This makes the correlation between equipment, processes, and emissions intuitive and traceable, significantly reducing the cost of interpreting and analyzing carbon emission data. Its supported process parameter visualization adjustment function allows enterprises to intuitively observe the dynamic impact of parameter changes on carbon emissions, while the simulation verification capability can pre-simulate the emission reduction effect of process optimization schemes in a virtual environment, avoiding the production risks and trial-and-error costs caused by blindly adjusting parameters in actual production. This provides a safe and efficient testing path for the accurate implementation of carbon reduction strategies. Simultaneously, as the core hub for data fusion and application, this module further strengthens the collaborative linkage between real-time data at the edge layer and the global model at the cloud layer. Instantaneous data from the edge layer can be quickly synchronized to the digital mirror for dynamic updates, and optimization strategies at the cloud layer can be deployed and executed after mirror simulation verification. This significantly improves the dynamic response capability and decision reliability of the entire metering system, providing key technical support for the accurate metering, dynamic monitoring, and efficient carbon reduction of enterprise carbon emissions.

[0027] Furthermore, it also includes a dynamic metering module, which provides a template library for the chemical, metallurgical, and equipment manufacturing industries, supports users to customize metering formulas and emission factors, and automatically generates full-chain carbon emission results.

[0028] The beneficial effects of the above solution are as follows: Through the collaboration of the industry template library, customization function, and automated generation, the industry adaptability, accuracy, and efficiency of carbon emission measurement have been significantly improved. The provided industry template libraries for chemical engineering, metallurgy, equipment manufacturing, etc. can directly adapt to the process characteristics of different industries, avoid the neglect of the special carbon emission mechanism of the industry by the general model, reduce the repeated development costs during enterprise deployment, and quickly meet the measurement requirements of multiple industries. The function that supports users to customize the formulas for measurement links and emission factors can accurately match the personalized processes of enterprises (such as specific reaction processes and raw material characteristics), break through the limitation that fixed models are difficult to adapt to the actual working conditions of enterprises, make the measurement logic more in line with production reality, and greatly improve the accuracy of the results. The ability to automatically generate the carbon emission results of the entire chain replaces the cumbersome mode of traditional manual calculation and summary by links, reduces human errors, improves measurement efficiency, and ensures the coherence and integrity of the data of the entire raw material-production-waste chain, avoiding measurement deviations caused by missing links. In addition, through flexible template calls and customized configurations, this module further enhances the dynamic adaptability of the entire measurement system to enterprise process adjustments and industry standard updates, providing core support for the long-term and accurate carbon emission management of enterprises.

[0029] Furthermore, it also includes an intelligent optimization module, which is used to display the correction strategies output by reinforcement learning and supports pushing process optimization suggestions to the production execution system at one click.

[0030] The beneficial effects of the above solution are as follows: 1. From the perspective of decision-making efficiency, the module intuitively presents the correction strategies generated by reinforcement learning (such as the model factor adjustment logic and the process parameter optimization direction), enabling managers to quickly understand the root causes of deviations and the optimization paths, avoiding the problems of high interpretation costs and decision-making lags caused by the abstraction of traditional optimization suggestions, and greatly improving the timeliness and accuracy of carbon emission reduction decisions.

[0031] 2. From the perspective of implementation and implementation, the one-click push function completely breaks the data barrier and execution fault between the carbon emission measurement system and the production execution system (MES). In the traditional solution, optimization suggestions need to be manually sorted and transmitted before adjusting production, which is prone to information distortion and response delays. However, this module realizes the standardized automatic distribution of suggestions, ensuring the rapid implementation of optimization strategies at the production end, especially adapting to scenarios that require immediate adjustment such as raw material fluctuations and equipment condition changes in industrial production, and winning a crucial time window for carbon emission reduction.

[0032] 3. From the perspective of system collaboration, this module takes over the precise data output of the dynamic metering module and the simulation verification results of the digital twin module, transforming the intelligent decision-making of reinforcement learning into executable production instructions, so that data acquisition-metering modeling-simulation optimization-production adjustment form a complete closed loop; at the same time, the execution feedback data from the production end can feed back into the reinforcement learning model iteration, further improving the adaptability of subsequent correction strategies, and realizing the continuous evolution of the system's metering accuracy and carbon reduction effect. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the carbon emission measurement method for industrial enterprises based on multi-source data fusion according to the present invention.

[0034] Figure 2 This is a schematic diagram of the modules of the industrial enterprise carbon emission metering system based on multi-source data fusion of the present invention.

[0035] Figure 3 This is a schematic diagram of the operation of the industrial enterprise carbon emission metering system based on multi-source data fusion according to the present invention. Detailed Implementation

[0036] The following detailed description illustrates the specific implementation method: Example 1 The basics are as follows: Figure 1 As shown: Taking the ethylene production project of a large petrochemical enterprise as a specific implementation scenario, the entire chain of carbon emission measurement is carried out based on the above-mentioned industrial enterprise carbon emission measurement method based on multi-source data fusion. The implementation process, mechanism of action, and effect of each step are as follows: I. Implementation and Role of Full-Chain Multi-Source Data Acquisition and Preprocessing In the ethylene production area of ​​this chemical enterprise, terminal data acquisition equipment completes full-area data coverage according to a preset plan: Regarding production equipment data, for core equipment such as reactors, distillation towers, and cracking furnaces, a standardized data transmission channel is established through the OPC UA protocol to collect parameters such as temperature (0-500℃), pressure (0-10MPa), stirring speed (0-300r / min), and total power consumption (0-1000kW) in real time; In order to achieve accurate breakdown of equipment-level energy consumption, a load decomposition algorithm is used to process the total power consumption of the cracking furnace area. Based on the equipment operating time, rated power, and load factor, a decomposition model is constructed. For example, the total hourly energy consumption of 1200kW is accurately broken down into 3 cracking furnaces (450kW, 420kW, and 330kW respectively), which solves the problem of the crude estimation of single equipment emissions in traditional total energy consumption estimation. For environmental and emission data acquisition, a multi-gas sensor array is deployed at the tail gas emission outlet of the pyrolysis furnace to simultaneously monitor concentration data such as CO2 (0-50% vol) and CH4 (0-1% vol), and uploads the raw monitoring values ​​every 5 seconds. Since the sensors are susceptible to interference from fluctuations in wind speed (0-15 m / s) and temperature (-10-40℃) in the plant area, a variational mode is introduced for noise reduction. By constructing a signal decomposition equation, recursive filtering is performed on 100 consecutive data points, which improves the signal-to-noise ratio of CO2 concentration data from 25dB to 40dB, significantly enhancing the reliability of the measured data.

[0037] The signal decomposition equation is as follows:

[0038] in, This is the noise term.

[0039] In process and management data processing, OCR technology is used to convert PDF-format ethylene cracking process specifications and equipment maintenance records into text data. Then, NLP technology is used to extract key entities (such as "cracking temperature 850℃" and "maintenance duration 4 hours") and logical relationships (such as "temperature ≥ 850℃ → residence time ≤ 0.5s"), forming a structured dataset. Furthermore, knowledge graph entity matching technology is used to associate the purchase batch ID of the feedstock ethylene with the feedstock carbon footprint data in the CLCD database, clarifying the implicit carbon emission baseline value (approximately 2.8 tCO2 / t ethylene) for each batch of ethylene from crude oil extraction to plant transportation, achieving cross-domain data connection between process parameters and feedstock carbon footprint. This step, through targeted acquisition protocols and processing technologies, constructs a high-quality data foundation covering "equipment-emissions-process-feedstock," providing fine-grained and highly reliable input for subsequent fusion and modeling, reducing the risk of measurement deviations from the source.

[0040] II. Implementation and Role of Digital Twin-Driven Hierarchical Multi-Source Data Fusion Based on the enterprise's physical space layout, a three-level integration system of "edge-workshop-enterprise" is built: Equipment-level edge integration deploys edge computing gateways near equipment such as pyrolysis furnaces and reactors, with built-in lightweight LSTM models. The input is the temperature, pressure, and energy consumption time-series data of the equipment within 1 minute (a total of 60 samples). The dynamic dependencies between parameters are captured through a gated loop unit, and the instantaneous carbon emission potential value of the equipment is output (such as the instantaneous potential value of 120kgCO2 / h for pyrolysis furnaces). The processing latency is controlled within 80ms, which meets the real-time monitoring needs of the production site and reduces the data transmission pressure on the cloud. Workshop-level time-series fusion takes the digital twin mirror of the ethylene synthesis workshop as its core, aligning the edge layer output data of 12 devices such as reactors, condensers, and separators according to the production cycle (one time unit per hour). The Transformer attention mechanism is introduced to build a fusion model. Through self-attention weight calculation (weight value range 0-1), the reactor (weight 0.42) and cracking furnace (weight 0.38) are automatically identified as key equipment, significantly reducing the interference of auxiliary equipment such as cooling pumps (weight 0.05), and outputting the carbon emission sequence of the entire process of "cracking-reaction-separation" (e.g., 1500 kg carbon emission per hour for the cracking process and 800 kg per hour for the reaction process), achieving accurate characterization of process-level emissions. Enterprise-level integration relies on a knowledge graph encompassing raw materials, processes, emissions, and the supply chain. It links workshop-level sequence data with multi-dimensional data such as raw material procurement data (ethylene carbon content 85%-87%), real-time grid factors (0.52-0.58 kg CO2 / kWh), and transportation data (transport distance 200-500 km). A graph neural network (GNN) is used to construct a node relationship model. The coupling coefficient between raw material carbon content and reaction temperature is calculated using an adjacency matrix (correlation coefficient 0.76), uncovering the implicit pattern that "for every 1% increase in ethylene carbon content, the reaction temperature needs to be reduced by 5°C to maintain stable emissions." The final output is a complete carbon emission dataset covering the entire chain from raw material entry to production and processing to product exit. This hierarchical integration strategy adapts to the characteristics of different data levels, achieving progressive data integration from "equipment instantaneous potential" to "workshop process sequence" to "enterprise-wide chain linkage." This ensures real-time performance while uncovering deeper data value, providing strong data support for the metrology model.

[0041] III. Implementation and Role of Dynamic Factor Adaptive Carbon Emission Measurement Model The measurement boundary is clearly defined as the entire process of "ethylene feedstock procurement - cracking reaction - waste catalyst disposal - product storage and transportation," and dynamic measurement models are established for each stage: the implicit emissions of feedstock are modeled using a dynamic carbon footprint model, with the core formula being:

[0042] in The baseline emission factor for ethylene (2.8 tCO2 / t) is from the Ecoinvent database. The batch variation coefficient provided to the supplier in real time (e.g., ΔX=0.03 for a batch due to a change in the origin of crude oil). Let i be the amount of raw material used in this batch (500t), i be the i-th batch of raw material, and n be the total number of batches of raw material in the statistical period. The total implicit carbon emissions are obtained by summing all batches. The implicit carbon emissions of this batch of raw material are calculated to be 2.8×(1+0.03)×500=1442 tCO2. This formula realizes the dynamic updating of emission factors with raw material batches, avoiding the time lag deviation of fixed factors.

[0043] Emissions from the production process are determined based on the characteristics of the chemical industry, using the following formula:

[0044] in This refers to the hourly energy consumption of equipment such as cracking furnaces and reactors (e.g., cracking furnace W=450kW). The real-time carbon emission factor of the power grid is 0.55 kg CO2 / kWh. The carbon conversion rate of the ethylene cracking reaction is 0.98. This refers to the raw material processing capacity (10t / h). Here, is the carbon content of the raw material (0.86), j is the specific energy-consuming equipment (such as No. 1 cracking furnace, No. 2 reactor, etc.), used to traverse all energy-consuming equipment in the production process; m is the total number of equipment involved in the energy consumption carbon emission calculation (such as the chemical enterprise's ethylene production involving cracking furnace, reactor, circulating pump, etc., totaling m=5 key energy-consuming equipment); k is the specific raw material batch or raw material type (such as No. 1 ethylene raw material, No. 2 catalyst, etc.), used to traverse all raw materials involved in the chemical reaction in the production process; p is the total number of raw materials involved in the reaction carbon emission calculation (such as in this scenario, ethylene raw material, catalyst, etc., totaling p=2 types of core raw materials).

[0045] The total emissions are obtained by adding up the carbon emissions from equipment energy consumption and raw material reaction. The calculated hourly carbon emissions of the production process are (450 × 0.55 × 10⁻⁶). -3 )+(0.98×10×0.86×3.67)=0.2475+30.12=30.3675 tCO2 / h (3.67 is the molar mass coefficient of carbon to CO2).

[0046] The correction factor for waste disposal and emission based on the disposal method is calculated using the following formula:

[0047] in, The weight of the spent catalyst (obtained via online weighing equipment, 20t) is given. The basic factor for landfill disposal is 0.15tCO2 / t. The correction factor for the disposal method is β=1.2 for incineration and β=0.3 for recycling. This project uses incineration for disposal. q represents the total number of waste types or the total number of waste disposals. l represents the l-th waste disposal behavior for the l-th type of waste. The waste emission is calculated to be 20×0.15×1.2=3.6 tCO2.

[0048] The dynamic updating of model factor weights and formula structure is achieved through weighting coefficients ω(t), as shown in the formula:

[0049] in, The attenuation coefficient is 0.7. To assess the reliability of the integrated data (calculated based on data integrity and timeliness, ranging from 0 to 1), when the reliability of the raw material data increases from 0.8 to 0.95, =0.7×0.3+0.3×0.95=0.495 (original weight 0.3), adjusted synchronously. The weighting of emissions in the calculation allows the model to adapt to changes in data quality. This step, through dynamic formulas and weighting mechanisms, enables the metrology model to adapt in real time to fluctuations in raw materials and changes in processes, significantly improving the accuracy of carbon emission calculations across the entire chain.

[0050] IV. Implementation and Role of Reinforcement Learning-Driven Dynamic Feedback Correction First, a real-time data comparison mechanism is constructed. Every 15 minutes, the measured CO2 value (C_meas) from the exhaust gas sensor is obtained and compared with the predicted value (C_pred) from the metrology model. The deviation rate is calculated using the following formula:

[0051] For example, if the measured CO2 / h at a certain moment is 30.5 tCO2 / h and the predicted value is 34.2 tCO2 / h, the calculated δ = 12.13%. The reinforcement learning agent designs its reward function with the objective of "minimizing the bias rate + minimizing the adjustment cost," and the formula is:

[0052] in, =0.8、 =0.2 is the weighting coefficient. To adjust costs (such as the energy consumption increase due to process parameter adjustments, and the procurement cost of raw material replacement).

[0053] The agent's action space is divided into three categories according to priority: model factor weight adjustment (ΔEF(t)= ×R(t), With an adjustment coefficient of 0.1, when δ = 12.13%, ΔEF = 0.1 × (0.8 × (1 - 0.1213) - 0.2 × 0.05) = 0.069, meaning the carbon conversion rate of the reaction is increased. From 0.98 to 0.9869; Recommended process parameters (ΔT(t) = ×(EF (t)- ), =-0.8, (For the optimal emission factor), the calculated ΔT = -0.8 × (0.55 - 0.52) = -2.4℃, therefore it is recommended to reduce the pyrolysis temperature from 850℃ to 847.6℃; Optimization of feedstock ratio (Δγ(t) = ×( (t)- ), =-0.5, =0.85), when the current carbon content is 0.86, Δγ = -0.5×(0.86-0.85) = -0.005, which suggests reducing the proportion of high-carbon raw materials by 0.5%. When the deviation rate exceeds 15% for two consecutive 15-minute cycles (e.g., δ1=15.8%, δ2=16.2%), a digital twin simulation is automatically triggered: simulating two schemes in the virtual mirror: "temperature reduced to 845℃ + raw material carbon content reduced to 0.85%" and "only adjusting model factors". The calculated R(t) for the first scheme is 0.8×(1-0.07)-0.2×0.1=0.694, which is significantly higher than the 0.58 for the second scheme. Therefore, this scheme is selected as the optimal correction strategy. After the correction strategy is implemented, the deviation rate between the model prediction and the measured value is reduced to 6.8%, while the adjustment cost is controlled within 1.2% of the original production budget. This mechanism, through active decision-making via reinforcement learning and simulation verification via digital twins, can achieve dual-dimensional optimization of model parameters and process execution. It not only solves the lag of traditional passive correction but also ensures the economy and safety of the adjustment.

[0054] V. Implementation and Role of Multi-Dimensional Output and Decision Support Based on the revised econometric model, three levels of output results are automatically generated: The enterprise-level report covers the total monthly carbon emissions across the entire chain (12,500 tCO2) and carbon emission intensity (3.2 tCO2 / t ethylene), and analyzes emission reduction potential against the industry benchmark (3.5 tCO2 / t ethylene); the workshop-level report identifies the cracking workshop as a high-carbon segment (accounting for 58%), with the No. 1 cracking furnace contributing 22% of emissions; carbon cost analysis data is calculated using the following formula:

[0055] in, The carbon price is 80 yuan / tCO2.

[0056] The monthly carbon cost is 1 million yuan, with 62% of that being hidden emissions from raw materials. These results are visualized through a digital twin dashboard: a top-down view shows a carbon emission heat map of the plant area, with red areas focusing on the area around the cracking furnace; a timeline view shows the daily emission trends of the three stages of "raw materials-production-waste" in a bar chart, marking the emission peaks caused by changes in raw material batches within 15 days; a comparison view dynamically displays the deviation between actual emissions and budgeted values, automatically highlighting a red warning when the deviation exceeds 5%. Enterprise managers can quickly locate the No. 1 cracking furnace as a key emission reduction point through the dashboard, click to view the intelligent optimization module's suggestion of "temperature reduced to 847℃ + replacement with low-carbon raw materials," and after confirmation, push it to the Production Execution System (MES) with one click, achieving seamless integration of measurement analysis, decision-making, and execution. This step transforms abstract measurement data into hierarchical and visualized decision-making basis, effectively supporting the enterprise's carbon quota compliance, optimization of high-carbon processes, and carbon cost control, driving the transformation of measurement value from "data statistics" to "carbon reduction empowerment."

[0057] Example 2 The difference from the above embodiments is that, as Figure 2 and Figure 3 As shown, the above-mentioned industrial enterprise carbon emission metering system was used to conduct a specific carbon emission metering experiment in the ethylene cracking production workshop of a large petrochemical enterprise to verify the actual performance of the present invention in a typical industrial scenario (taking ethylene production in a chemical enterprise as an example).

[0058] I. Experimental Objective The following capabilities of this invention will be evaluated: accuracy improvement: whether the carbon emission measurement deviation rate of the entire chain can be stably controlled within 8%; dynamic adaptability: whether the model can be automatically corrected in dynamic scenarios such as raw material fluctuations and process adjustments; carbon reduction value: whether the generated process optimization suggestions have a practical guiding role in emission reduction.

[0059] II. Experimental Environment and Scenarios Experimental system: Deploy the "cloud-edge-device" three-level system architecture described in this invention. The terminal layer includes multiple gas sensors, smart meters, RFID readers, and GPS terminals; the edge layer includes the EdgeX Foundry framework and LSTM model; and the cloud layer deploys a digital twin engine (Unity), a knowledge graph engine (Neo4j), and a reinforcement learning module (TensorFlow Agents).

[0060] Comparison object: The traditional accounting method currently used by enterprises based on IPCC static emission factors and monthly electricity settlement.

[0061] Data period: 30 consecutive days of production data.

[0062] Verification benchmark: The measured carbon emission data from a high-precision continuous emission monitoring system (CEMS) is used as the "true value".

[0063] III. Experimental Procedure Phase 1: System Deployment and Data Access (Days 1-3) Terminal layer deployment: Smart meters (sampling frequency 1Hz) are installed in key equipment such as cracking furnaces and reactors to collect temperature, pressure and energy consumption data via OPC UA protocol; RFID readers are deployed at the raw material inlet to track ethylene raw material batches and sources; a multi-gas sensor array (CO2, CH4) is deployed at the emission outlet with a sampling frequency of 0.2Hz; and the provincial power grid platform API is connected to obtain real-time carbon emission factors (updated every 15 minutes).

[0064] Edge layer deployment: An edge gateway was deployed in the workshop computer room, and a lightweight LSTM model (60 input dimensions, 32 hidden layer units) was configured to calculate the device-level carbon emission potential value in real time; the EdgeX Foundry framework was deployed to realize local data caching and protocol conversion.

[0065] Cloud-based deployment: Build a digital twin model of the ethylene production workshop (Unity engine) to map equipment layout and process flow; construct a knowledge graph (Neo4j) to integrate raw material, process, and emission data entities; and initialize a reinforcement learning agent.

[0066] Phase 2: Baseline Operation and Model Training (Days 4-10) The entire system is fully integrated, carrying out data collection, fusion, and model training; the digital twin engine completes historical data alignment and physical mapping; and the reinforcement learning agent conducts exploratory training (exploration rate 0.3, learning rate 0.001).

[0067] Phase 3: Steady-state operation and comparative experiment (days 11-25) The system enters steady-state operation, completing a "collection-fusion-metering-correction" cycle every 15 minutes; on the 16th day, a disturbance is introduced: 100 tons of alternative raw materials with a carbon content of 87% are mixed into the production line to simulate raw material fluctuations; the hourly metering values ​​of the system of this invention and the traditional method are recorded, and the deviation rate is calculated by comparing them with the actual CEMS values.

[0068] Phase 4: Optimization and Validation (Days 26-30) The intelligent optimization module outputs process suggestions: "Reduce the pyrolysis temperature by 2.4℃ + reduce the proportion of high-carbon feedstock by 0.5%"; after verification through digital twin simulation, it is pushed to the MES system for execution with one click; and the changes in carbon emission intensity and deviation rate are monitored after execution.

[0069] IV. Experimental Results As shown in the table below: Table 1. Comparison and statistics of measurement deviation rates (days 11-25)

[0070] Note: The maximum deviation of 9.5% occurred in the first cycle after the introduction of raw material disturbance, and the system corrected it to 7.1% within 15 minutes through reinforcement learning.

[0071] Table 2. Results of Process Optimization Implementation (Days 26-30)

[0072] V. Experimental Conclusions The system of this invention reduces the average deviation rate from 18.5% of traditional methods to 5.8%, with a maximum deviation rate not exceeding 9.5%, achieving the technical goal of "controlling the deviation rate within 8%". This demonstrates the effectiveness of digital twin-driven fusion and reinforcement learning dynamic correction.

[0073] In scenarios involving raw material disturbances, the deviation rate of traditional methods soars to 35%, while the system of this invention corrects the deviation rate to 7.1% within 15 minutes by dynamically updating model factors (such as adjusting the reaction carbon conversion rate from 0.98 to 0.9869) and process parameters (reducing the pyrolysis temperature by 2.4℃), demonstrating a strong adaptability to operating conditions.

[0074] After the system's output process optimization suggestions were implemented, the actual carbon emission intensity decreased by 2.2%, and the carbon cost decreased by approximately RMB 25,000 per month (calculated based on a carbon price of RMB 80 per tCO2). This demonstrates that the present invention not only improves metering accuracy but also directly empowers enterprises' carbon reduction practices.

[0075] The "cloud-edge-device" architecture enables the collaboration of data acquisition (device layer), real-time processing (edge ​​layer), and global optimization (cloud layer). The edge layer processing latency is controlled within 80ms, and the cloud reinforcement learning decision cycle is 15 minutes, which is superior to traditional pure cloud or stand-alone systems.

Claims

1. A method for measuring carbon emissions of industrial enterprises based on multi-source data fusion, characterized in that, Includes the following steps: S1. Full-chain multi-source data acquisition and preprocessing: Real-time acquisition of production equipment data, raw material or product data, environmental and emission data, supply chain and power grid data, and process and management data of industrial enterprises through terminal acquisition devices; cleaning, noise reduction, format conversion and structuring of the acquired data. S2. Digital Twin-Driven Layered Multi-Source Data Fusion: Adopting a three-level fusion architecture of "edge-workshop-enterprise", equipment-level data is integrated at the edge node through a lightweight time-series fusion model. Based on the physical mirror of the workshop of digital twin, data of multiple equipment in the same process link is aligned. Combined with the knowledge graph of "raw materials-process-emission-supply chain" to associate enterprise-level multi-dimensional data, the full-chain carbon emission dataset is output. The full-chain carbon emission dataset includes basic identification fields: data ID, enterprise code, workshop number, timestamp and emission source type. Core measurement fields: carbon emissions at each stage, emission factors, and raw material or product association ID; associated attribute fields: equipment operating parameter snapshots, raw material carbon content, and treatment method correction coefficients; and data traceability fields: data source identifier, data cleaning records, and model version number. S3. Construction of Dynamic Factor Adaptive Carbon Emission Measurement Model: Clarify the carbon emission measurement boundary of the entire chain of "raw material acquisition - production and processing - waste disposal - product storage and transportation", establish a measurement model with custom formulas for different links, and dynamically update the model factor weights and formula structure based on fused data; S4. Dynamic feedback correction driven by reinforcement learning: The deviation rate is calculated by comparing the model prediction value with the measured value at the edge. The reward function is to minimize the deviation rate and minimize the adjustment cost. The correction strategy is output through the reinforcement learning agent to dynamically adjust the model parameters and generate process optimization suggestions. S5. Multi-dimensional output and decision support: Generate enterprise-level and workshop-level carbon emission reports and carbon cost analysis data based on the revised model, and visualize the dynamic flow of carbon emissions through a digital twin dashboard.

2. The industrial enterprise carbon emission measurement method based on multi-source data fusion according to claim 1, characterized in that, In step S1: Production equipment data includes operating parameters and energy consumption data of machine tools and reactors, which are collected through the OPC UA protocol and broken down into equipment-level energy consumption using a load decomposition algorithm; Environmental and emission data include data related to exhaust gas, wastewater, and solid waste, which are collected through multi-gas sensors and noise reduction is performed using variational mode. Process and management data are structured through multi-engine OCR recognition and NLP technology parsing, and knowledge graph entity matching is used to associate raw material carbon footprint data.

3. The industrial enterprise carbon emission measurement method based on multi-source data fusion according to claim 1, characterized in that, In step S2: The device-level edge fusion adopts the LSTM temporal fusion model. The input is the temporal stream of device operation parameters collected by the terminal layer, including temperature, pressure and energy consumption dimensions. The sampling frequency is 5-15s / time. The output is the device-level instantaneous carbon emission potential value and data credibility score. Workshop-level time-series fusion adopts the Transformer attention mechanism to highlight the data weight of key equipment and output the carbon emission sequence of the process. Enterprise-level integration employs graph neural networks to mine implicit relationships in data, including the coupling relationship between the carbon content of raw materials and reaction temperature.

4. The industrial enterprise carbon emission measurement method based on multi-source data fusion according to claim 1, characterized in that, In step S3: The implicit emissions from raw materials are modeled using a dynamic carbon footprint, and the emission factors are updated in real time with the Ecoinvent database and supplier data. Production process emissions are provided with customized core formulas for the chemical and metallurgical industries. The formula for the chemical industry is as follows: ； The waste disposal discharge incorporates a correction factor for the disposal method, calculated in conjunction with online solid waste weighing data.

5. The industrial enterprise carbon emission measurement method based on multi-source data fusion according to claim 1, characterized in that, In step S4: The formula for calculating the deviation rate is: The status of the merged data is updated every 15 minutes. When the deviation rate exceeds 15% twice consecutively, the digital twin simulation is triggered to screen the optimal correction scheme. The action space of a reinforcement learning agent includes model factor weight adjustment, process parameter suggestions, and raw material ratio optimization.

6. An industrial enterprise carbon emission measurement system based on multi-source data fusion, used in the industrial enterprise carbon emission measurement method according to any one of claims 1-5, characterized in that, include: The terminal layer is used to sense and collect physical data related to corporate carbon emissions, and to complete the initial purification and standardization of the data. It includes gas sensors, smart meters, RFID readers and GPS terminals. The edge layer is used to provide near-field computing, storage and control capabilities near the physical data source, enabling localized real-time data processing and real-time decision feedback, including edge computing gateways and local servers; The cloud layer is used to receive and aggregate data from the edge layer for global modeling, in-depth analysis, and iterative optimization. It includes cloud servers, digital twin engines, and knowledge graph engines.

7. The industrial enterprise carbon emission metering system based on multi-source data fusion according to claim 6, characterized in that, It also includes a data access module, which supports OPC UA / Modbus protocol for interfacing with industrial equipment, API interface for interfacing with ERP and power grid platforms, and OCR / NLP for processing unstructured documents. This module is used to realize the automated, standardized access and governance of multi-source heterogeneous data.

8. The industrial enterprise carbon emission metering system based on multi-source data fusion according to claim 6, characterized in that, It also includes a digital twin fusion module, which is used to build a digital mirror of the enterprise's physical space, map equipment operating status and carbon emission data in real time, and support the visualization adjustment and simulation verification of process parameters.

9. The industrial enterprise carbon emission metering system based on multi-source data fusion according to claim 6, characterized in that, It also includes a dynamic metering module, which provides a template library for the chemical, metallurgical, and equipment manufacturing industries, supports users to customize metering formulas and emission factors, and automatically generates full-chain carbon emission results.

10. The industrial enterprise carbon emission metering system based on multi-source data fusion according to claim 6, characterized in that, It also includes an intelligent optimization module, which is used to display the correction strategy of the reinforcement learning output and supports pushing process optimization suggestions to the production execution system with one click.

Images