Zero-carbon park-oriented AI pollution and carbon reduction evaluation method and system and medium

Abstract

The invention discloses a zero-carbon park-oriented AI pollution reduction and carbon reduction evaluation method and system and a medium, and relates to the technical field of carbon reduction management. The method comprises the following steps: executing multi-source data acquisition of a park, and establishing a multi-source data set; establishing a fusion data set; predicting park carbon footprint and pollutant distribution through a multi-mode machine learning model, and establishing a prediction result; executing causal analysis and anti-factual reasoning of a prediction result by utilizing a pollution backtracking channel, and after identifying a key pollution source and a hidden collaborative emission reduction point, generating an intervention priority map according to a fusion data set in combination with causal relationship strength and a potential collaborative emission reduction effect; and constructing a multi-scale park digital twinborn model, and executing simulation evaluation of a pollution reduction and carbon reduction strategy. The technical problem that in the prior art, park pollution reduction and carbon reduction evaluation depends on static statistics and lacks dynamic data support, and consequently the reliability of the evaluation result is insufficient is solved, and the technical effect of improving the reliability of the evaluation result through multi-source data fusion and causal reasoning is achieved.

Description

Technical Field

[0001] This invention relates to the field of carbon reduction management technology, specifically to an AI-based method, system, and medium for assessing pollution and carbon reduction in zero-carbon industrial parks. Background Technology

[0002] With the continuous advancement of zero-carbon industrial park construction, refined assessment of carbon emissions and pollutants has gradually become a crucial foundation for achieving green transformation. However, current assessment methods largely rely on historical data and static statistical models, often providing only phased total analysis and failing to comprehensively reflect the dynamic evolution of emissions and energy consumption. In practical applications, this static data-driven assessment method easily overlooks the impact of temporal fluctuations, changes in production load, and technological differences, leading to a lack of sufficient credibility in the assessment results. Especially when facing complex industrial park energy systems and diverse emission sources, single statistical methods are insufficient to reveal key pollution sources and potential synergistic emission reduction opportunities, and the assessment conclusions often present uncertainty in guiding intervention strategies. These shortcomings result in industrial parks lacking a scientifically reliable basis for formulating and implementing pollution reduction and carbon reduction measures, directly hindering the improvement of emission reduction effectiveness and the refinement of management decisions. Summary of the Invention

[0003] This application provides an AI-based pollution reduction and carbon reduction assessment method, system, and medium for zero-carbon industrial parks, which solves the technical problem that existing technologies rely on static statistics and lack dynamic data support for pollution reduction and carbon reduction assessments in industrial parks, resulting in insufficient reliability of assessment results.

[0004] The first aspect of this application provides an AI-based pollution reduction and carbon reduction assessment method for zero-carbon industrial parks, the method comprising: Multi-source data collection is performed on the industrial park to establish a multi-source dataset, which includes data on carbon emissions, energy consumption, production processes, and equipment status. The dataset dynamically generates confidence level indicators based on the reliability of the data sources. After time-series alignment, the confidence level indicators are used for cross-modal consistency constraints and proactive anomaly repair analysis. Adaptive data fusion is then performed on the multi-source dataset to establish a fused dataset. Using the fused dataset as input, a multi-modal machine learning model is used to predict the park's carbon footprint and pollutant distribution, generating prediction results. Causal analysis and counterfactual reasoning are performed on the prediction results using a pollution backtracking channel to identify key pollution sources and hidden synergistic emission reduction points. Based on the fused dataset, combined with causal relationship strength and potential synergistic emission reduction effects, an intervention priority map is generated. A multi-scale digital twin model of the industrial park is constructed using the fused dataset and the intervention priority map, and simulation evaluation of pollution reduction and carbon reduction strategies is performed.

[0005] A second aspect of this application provides an AI-based pollution and carbon reduction assessment system for zero-carbon industrial parks, the system comprising: Data Acquisition Module: Performs multi-source data acquisition for the park, establishing a multi-source dataset including data on carbon emissions, energy consumption, production processes, and equipment status. The dataset dynamically generates confidence level indicators based on the reliability of the data sources. Data Fusion Module: After time-series alignment of the multi-source dataset, it uses the confidence level indicators for cross-modal consistency constraints and proactive anomaly repair analysis, adaptively fusing the multi-source dataset to establish a fused dataset. Prediction Module: Using the fused dataset as input, it predicts the park's carbon footprint and pollutant distribution through a multimodal machine learning model, establishing prediction results. Map Generation Module: Utilizes pollution backtracking channels to perform causal analysis and counterfactual reasoning on the prediction results, identifying key pollution sources and hidden synergistic emission reduction points. Based on the fused dataset, it generates an intervention priority map combining causal relationship strength and potential synergistic emission reduction effects. Evaluation Module: Constructs a multi-scale digital twin model of the park using the fused dataset and the intervention priority map, and performs simulation evaluation of pollution reduction and carbon reduction strategies.

[0006] A third aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements an AI-based pollution reduction and carbon reduction assessment method for zero-carbon industrial parks provided in this application.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages: First, multi-source data collection was performed on the park to establish a multi-source dataset, including data on carbon emissions, energy consumption, production processes, and equipment status. Confidence markers were dynamically generated for each dataset based on the reliability of the data sources. Next, the multi-source dataset was time-series aligned, and cross-modal consistency constraints and proactive anomaly remediation analysis were performed using the confidence markers. Adaptive data fusion was then performed on the multi-source dataset to establish a fused dataset. Further, using the fused dataset as input, a multi-modal machine learning model was used to predict the park's carbon footprint and pollutant distribution, establishing prediction results. Then, causal analysis and counterfactual reasoning were performed on the prediction results using a pollution backtracking channel. After identifying key pollution sources and hidden synergistic emission reduction points, an intervention priority map was generated based on the fused dataset, combining causal relationship strength and potential synergistic emission reduction effects. Finally, a multi-scale digital twin model of the park was constructed using the fused dataset and the intervention priority map, and a simulation evaluation of pollution reduction and carbon reduction strategies was performed. This approach solves the technical problem in existing technologies where pollution reduction and carbon reduction assessments rely on static statistics and lack dynamic data support, leading to insufficient reliability of assessment results. It achieves the technical effect of improving the reliability of assessment results through multi-source data fusion and causal reasoning. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 A schematic diagram of an AI-based pollution reduction and carbon reduction assessment method for zero-carbon industrial parks, provided as an embodiment of this application; Figure 2 This is a schematic diagram of the structure of an AI-based pollution reduction and carbon reduction assessment system for zero-carbon industrial parks, provided as an embodiment of this application.

[0010] Figure labeling: Data acquisition module 11, data fusion module 12, prediction module 13, map generation module 14, evaluation module 15. Detailed Implementation

[0011] This application provides an AI-based pollution reduction and carbon reduction assessment method, system, and medium for zero-carbon parks, which solves the technical problem that existing park pollution reduction and carbon reduction assessments rely on static statistics and lack dynamic data support, resulting in insufficient reliability of assessment results.

[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and 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.

[0013] It should be noted that the terms "comprising" and "having" are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to these processes, methods, products, or devices.

[0014] Example 1, as Figure 1 As shown, this application provides an AI-based pollution and carbon reduction assessment method for zero-carbon industrial parks, wherein the method includes: Perform multi-source data collection in the park and establish a multi-source dataset. The multi-source dataset includes data on carbon emissions, energy consumption, production processes, and equipment status in the park. The multi-source dataset dynamically generates confidence level labels based on the reliability of the data sources.

[0015] In this embodiment, multi-source operational data is collected within the target industrial park area. This multi-source data includes carbon emission monitoring data, energy consumption data, production process data, and equipment status data. Specifically, carbon emission monitoring data is acquired in real-time by sensors deployed at flue gas emission outlets, wastewater discharge points, and environmental monitoring stations, covering the concentrations of major pollutants such as carbon dioxide, nitrogen oxides, and particulate matter. Energy consumption data is collected from electricity meters, water meters, gas meters, and the energy management system, including the consumption and load curves of energy sources such as electricity, steam, and natural gas. Production process data originates from the process control system and production execution system, reflecting the operating parameters, output, and emission factors of key process stages. Equipment status data is acquired through the equipment operation monitoring system, including equipment start-up and shutdown status, operating efficiency, fault records, and maintenance logs.

[0016] After establishing a multi-source dataset, a data source reliability assessment is performed, specifically including: pre-setting a set of reliability evaluation indicators for each type of data source. These indicators include sampling frequency stability (sampling period variance ≤ first threshold), historical missing rate (number of missing records / total number of records ≤ second threshold), outlier rate (number of outliers / total number of records ≤ third threshold), and cross-source consistency (correlation coefficient with adjacent monitoring points ≥ fourth threshold). For each data source, a single score is calculated according to the set of reliability evaluation indicators, and an overall reliability score R is generated based on a weighted summation. The calculation formula is as follows: ,in, This represents the score value of the i-th evaluation indicator. This represents the weight of the indicator, and the sum of these weights equals 1. The reliability score R is normalized and mapped to the interval [0, 1], serving as the confidence level identifier. If R ≥ 0.8, the confidence level identifier is recorded as high (≥ 0.8); if 0.5 ≤ R < 0.8, the confidence level identifier is recorded as medium (0.5~0.8); if R < 0.5, the confidence level identifier is recorded as low (< 0.5). When storing multi-source datasets, each data record is accompanied by its confidence level identifier.

[0017] After aligning the multi-source datasets by time series, the confidence index is used to perform cross-modal consistency constraints and proactive anomaly repair analysis, and adaptive data fusion of the multi-source datasets is performed to establish a fused dataset.

[0018] In this embodiment, the multi-source datasets are aligned according to a unified time base. Since the sampling frequencies of data such as carbon emissions, energy consumption, production processes, and equipment status differ—for example, carbon emission data may be collected at the second level, energy consumption data at the minute level, process records at the hour level, and equipment status data at event-triggered levels—a unified time granularity is set. Data with different sampling frequencies are mapped to the same time series using interpolation, downsampling, or segmented aggregation, thereby ensuring that various types of data are comparable within the same time window.

[0019] After time series alignment, confidence levels are used to constrain consistency across different modalities. If a data type exhibits significantly low correlation with other modalities within the same time window, and its confidence level is below a preset threshold, the data is considered an anomaly. In this case, proactive anomaly repair is performed: firstly, a reasonable range for the anomaly is inferred based on the causal relationship with high-confidence data; secondly, trend prediction is performed using historical data to obtain alternative values. The results of causal inference and historical prediction are weighted and integrated to obtain the repaired value, which replaces the original anomaly, ensuring temporal and logical consistency across multiple data sources.

[0020] After the repair is completed, the aligned multi-source data is weighted and fused according to the confidence level. Specifically, within the same time window, data with higher confidence levels have a larger weight in the fusion result, while data with lower confidence levels have a correspondingly smaller weight.

[0021] Furthermore, the analysis of cross-modal consistency constraints and proactive anomaly repair using the aforementioned confidence level identifier includes: Data integrity and consistency analysis are performed on the multi-source dataset. If an anomaly active repair mechanism is triggered, active data repair is performed through a dual-channel approach of causal relationship inference and time series prediction to establish active repair results.

[0022] First, a data integrity analysis is performed on the multi-source dataset to detect missing data, sampling gaps, or incomplete records. Simultaneously, a data consistency analysis is performed to compare the numerical trends and correlations of different modalities within the same time window, such as comparing the synchronicity of carbon emission monitoring data and energy consumption data during load increases. When the detection results show that a certain modality's data is missing, abruptly changes, or is significantly inconsistent with other modalities, and its confidence level is below a preset threshold, an active anomaly repair mechanism is triggered. After triggering, the abnormal data is repaired through a dual-channel approach: causal relationship inference and time-series prediction. The causal relationship inference channel infers the reasonable value range of the abnormal modality within the current time window based on historical high-confidence data and causal dependencies between modalities. The time-series prediction channel uses the modality's own historical time-series data to predict the current value using an autoregressive model or recurrent neural network. The causal relationship inference results and time-series prediction results are then fused according to confidence levels to generate an active repair result, which is used to replace outliers or fill in missing values, thereby improving the completeness and reliability of the dataset while maintaining cross-modal consistency.

[0023] Using the fused dataset as input data, a multimodal machine learning model is used to predict the carbon footprint and pollutant distribution of the park, and prediction results are established.

[0024] In this embodiment, a fused dataset is input into a multimodal machine learning model to predict the carbon footprint and pollutant distribution of the industrial park. Specifically, the fused dataset includes multi-dimensional features such as carbon emission monitoring, energy consumption, production processes, and equipment status. First, a preprocessing layer extracts and normalizes the input data to eliminate differences in units and numerical ranges between different modalities. Then, a multimodal feature interaction mechanism, such as a joint attention layer or a mutual information constraint layer, is introduced into the multimodal machine learning model to model the correlations between different modalities, thereby capturing the potential nonlinear coupling relationship between carbon emissions and pollutant diffusion.

[0025] In the prediction phase, a multimodal machine learning model establishes two branches: a time-series carbon footprint prediction branch and a spatial distribution pollutant prediction branch. The time-series carbon footprint prediction branch uses residual recursive networks or long short-term memory networks to mine long-term dynamic trends and short-term fluctuations. The spatial distribution pollutant prediction branch, based on the spatial topology of the industrial park, models the diffusion relationships between different emission points, equipment, and pipelines using graph convolutional networks. Finally, the two types of prediction results are subjected to multi-objective consistency regularization to ensure the overall logical consistency between time-series and spatial predictions, outputting complete prediction results.

[0026] Furthermore, using the fused dataset as input data, a multimodal machine learning model is used to predict the park's carbon footprint and pollutant distribution, establishing prediction results, including: Feature extraction of the fused dataset is performed using the preprocessing layer of the multimodal machine learning model, including extracting feature subsets of carbon emissions, energy consumption, production processes, and equipment status. A multi-scale window slicing mechanism is introduced in the time series dimension to establish cross-scale input features. A joint attention layer is constructed in the multimodal machine learning model, and the joint attention layer is used to learn mutual information constraints on feature subsets of different modalities to capture the potential nonlinear coupling relationship between carbon emissions and pollutant diffusion. The deep features output by the joint attention layer are used to establish a carbon footprint time series prediction branch and a pollutant spatial distribution prediction branch. The carbon footprint time series prediction branch is constructed by extracting long-term dynamic trends through a residual recursive network, and the pollutant spatial distribution prediction branch is constructed by modeling the spatial topology of the park through a graph convolutional network. The carbon footprint time-series prediction branch and the pollutant spatial distribution prediction branch are subjected to multi-objective consistency regularization to establish the prediction results.

[0027] Feature extraction is performed on the fused dataset in the preprocessing layer of the multimodal machine learning model. Specifically, subsets of carbon emission features, energy consumption features, production process features, and equipment status features are extracted from the fused dataset, and normalization and feature denoising are performed on each subset to ensure the comparability of data from different sources on a numerical scale. Simultaneously, a multi-scale window slicing mechanism is introduced in the time series dimension to simultaneously preserve both long-term trend features and short-term instantaneous change features, establishing cross-scale input features to provide multi-level information support for subsequent predictions.

[0028] In the feature interaction stage of the multimodal machine learning model, a joint attention layer is constructed to perform mutual information constraint learning on the feature subsets of different modalities. By constraining the information entropy difference and mutual information gain between different modalities during training, the model can capture the potential nonlinear coupling relationship between carbon emission levels and pollutant diffusion trajectories, thereby avoiding information loss caused by single-modal prediction.

[0029] Furthermore, by utilizing the deep features output by the joint attention layer, a carbon footprint temporal prediction branch and a pollutant spatial distribution prediction branch are established, respectively. The carbon footprint temporal prediction branch is constructed based on a residual recursive network, which extracts long-term dynamic trends while maintaining short-term fluctuation characteristics through the recursive unit and residual connection mechanism. The pollutant spatial distribution prediction branch utilizes the physical layout and pipeline structure information of the park, and models spatial topological relationships based on graph convolutional networks to model and predict different emission points and diffusion paths.

[0030] The outputs of the carbon footprint time-series prediction branch and the pollutant spatial distribution prediction branch are subjected to multi-objective consistency regularization. Specifically, firstly, the accuracy of the outputs of the carbon footprint time-series prediction branch and the pollutant spatial distribution prediction branch is evaluated. The accuracy evaluation can use error comparison with actual monitoring values, trend fit, or spatial coverage indicators. When the deviation of the prediction result of either branch exceeds a preset allowable range, a consistency constraint mechanism is triggered. Under the consistency constraint mechanism, the trend of carbon footprint prediction and the trend of pollutant spatial diffusion prediction within the same time window are compared. If the directions of the two are inconsistent or the difference in the magnitude of change exceeds a threshold, it is determined to be an inconsistent prediction, and parameter correction is performed. The parameter correction operation includes: increasing the number of training iterations of the abnormal prediction branch, reducing its learning rate, or introducing high-confidence features from another branch as auxiliary input to gradually reduce the difference between the prediction results of the two branches. After completing the above correction, the outputs of the two branches are updated synchronously to ensure that the carbon footprint time-series prediction results and the pollutant spatial distribution prediction results are consistent in logical trends and change patterns, thereby establishing coordinated prediction results.

[0031] By utilizing pollution retrospective channels to perform causal analysis and counterfactual reasoning on the prediction results, key pollution sources and hidden synergistic emission reduction points are identified. Based on the fused dataset, an intervention priority map is generated by combining the strength of causal relationships and potential synergistic emission reduction effects.

[0032] In this embodiment, a pollution backtracking channel is used to perform causal analysis and counterfactual reasoning on the prediction results. Specifically, the carbon footprint time-series distribution and pollutant spatial diffusion prediction results obtained through a multimodal machine learning model are input into the pollution backtracking channel. Based on the time-series mapping relationship, the trajectory of carbon emission concentration changes is compared with energy consumption fluctuations, process status, and equipment operating parameters to form causal candidate chains. Subsequently, a causal constraint network is introduced into the pollution backtracking channel to perform causal path analysis on local abnormal emission points in the prediction results, identifying the process links that directly cause abnormal emissions, as well as influencing factors that indirectly cause abnormal emissions through energy-driven processes or reduced equipment efficiency.

[0033] Building upon causal path analysis, counterfactual reasoning is employed. This involves constructing several counterfactual scenarios by virtually replacing key variables or perturbing input parameters. For example, keeping other conditions constant, assuming an increase in equipment efficiency or a decrease in a process parameter, carbon emissions and pollutant distribution are recalculated, yielding counterfactual predictions. The difference between the counterfactual predictions and the original predictions is quantified; the difference value represents the causal contribution of that variable to emissions. By comparing the cumulative effects of pollutant diffusion under different counterfactual scenarios, the synergistic emission reduction potential among multiple variables can be identified, thus uncovering hidden synergistic emission reduction points.

[0034] Based on global information from the fused dataset, and combining the contribution intensity and potential synergistic emission reduction effects of each causal variable, an intervention priority map is generated. This intervention priority map ranks different intervention measures using quantitative indicators (such as causal contribution scores and synergistic emission reduction effect indices), providing the park with a scientific basis for pollution reduction and carbon reduction intervention decisions.

[0035] Furthermore, utilizing pollution tracing channels to perform causal analysis and counterfactual reasoning on predicted results includes: The carbon footprint distribution and pollutant diffusion trajectory from the predicted results are input into the pollution backtracking channel. Carbon emission concentrations are time-series mapped to equipment operating conditions, process steps, and energy consumption. A causal constraint network is introduced into the pollution backtracking channel. This network is used to analyze the causal paths of local abnormal emission points in the predicted results, identifying direct process steps and indirect energy driving factors, and establishing causal path analysis results. Based on the causal path analysis results, key variables are virtually replaced and parameters are perturbed. Counterfactual reasoning is performed to establish counterfactual prediction scenarios. The difference between the counterfactual prediction scenarios and the predicted results is calculated, and the difference is quantified to determine the causal contribution of pollution sources, identifying key pollution sources. Finally, pollution synergy analysis is performed on different counterfactual prediction scenarios to identify hidden synergistic emission reduction points.

[0036] The carbon footprint temporal distribution and pollutant spatial diffusion trajectory from the prediction results are input into the pollution backtracking channel. Carbon emission concentrations are mapped to equipment operating conditions, process parameters, and energy consumption curves on a unified time axis, thus constructing a temporal correlation between emissions and operating conditions. Subsequently, a causal constraint network is introduced into the pollution backtracking channel to analyze the causal paths of local anomalous emission points appearing in the prediction results. This causal constraint network includes an input layer, a causal structure learning layer, a constraint optimization layer, and an output layer: the input layer receives multimodal feature vectors of carbon emissions, energy consumption, process parameters, and equipment status; the causal structure learning layer calculates the conditional correlation coefficients, mutual information, and time lag parameters between variables based on directed acyclic graphs or causal graph neural networks to construct candidate causal links; the constraint optimization layer uses confidence indicators as weighting factors, assigning higher weights to high-confidence data in the causal links, while performing regularization on the causal structure to avoid spurious causality; the output layer generates the causal path analysis results, distinguishing between direct process parameters and indirect energy-driven factors.

[0037] Building upon causal path analysis, counterfactual reasoning is further applied to key variables: multiple counterfactual prediction scenarios are generated by virtually replacing a variable (such as equipment load or process parameters) or perturbing its value. The differences between the counterfactual prediction scenarios and the original prediction results are compared, and these differences are used as a quantitative indicator of causal contribution to measure the variable's impact on overall emissions, thereby identifying key pollution sources. Finally, multiple counterfactual scenarios are combined for analysis. If the emission reduction effect after simultaneously perturbing multiple variables is greater than the arithmetic sum of the emission reduction effects of perturbing each variable individually, a synergistic effect is identified, thus revealing hidden synergistic emission reduction points.

[0038] Using the fused dataset and intervention priority map, a multi-scale digital twin model of the park was constructed, and a simulation evaluation of pollution reduction and carbon reduction strategies was performed.

[0039] In this embodiment, sub-models at different levels of the industrial park are established based on the fused dataset, including physical layer sub-models (such as plant equipment layout, pipeline structure, and energy flow paths), process layer sub-models (such as the operating logic and energy consumption characteristics of major production processes), and management layer sub-models (such as emission standards, energy consumption indicators, and scheduling strategies). During the construction process, intervention priority maps are embedded into the multi-scale sub-models, and the intensity of pollution source impact and potential synergistic emission reduction effects between different sub-models are transmitted and quantified through a causal constraint coupling mechanism. This ensures that the digital twin not only has physical consistency but also dynamically reflects the emission evolution process driven by causality.

[0040] After the digital twin model is established, simulations and evaluations of pollution reduction and carbon reduction strategies are conducted. By setting different combinations of intervention measures (such as process parameter optimization, equipment energy efficiency improvement, and energy structure adjustment), virtual experiments are carried out in the digital twin environment to calculate the changing trends of carbon emissions and pollutant diffusion in real time. By comparing with a baseline scenario, the emission reduction effects, energy consumption changes, and synergistic benefits of each strategy are evaluated. Simultaneously, to enhance the practicality of the simulation, a dynamic scenario generation mechanism can be introduced into the twin model to generate various simulation scenarios based on energy price fluctuations, load changes, and meteorological conditions. This tests the adaptability of different strategies under complex external conditions, thereby providing park managers with a scientific and verifiable basis for pollution reduction and carbon reduction decision-making.

[0041] Furthermore, a multi-scale digital twin model of the park is constructed using the fused dataset and intervention priority map, including: Based on the fused dataset, a multi-scale sub-model of the park's physical layer, process layer, and management layer is established. The intervention priority map is embedded into the multi-scale sub-model. Through the causal constraint coupling mechanism, the influence intensity of pollution sources and the synergistic emission reduction effect are transmitted between the sub-models, generating a multi-scale digital twin model of the park.

[0042] A multi-scale sub-model of the park was established based on the fused dataset. The multi-scale sub-model includes a physical layer sub-model, a process layer sub-model, and a management layer sub-model. The physical layer sub-model is built based on the park's equipment layout, energy pipeline structure, and spatial topology, and is used to simulate energy transmission paths and emission diffusion patterns. The process layer sub-model is built based on the production process flow and operational parameters, and is used to reflect the energy consumption levels and emission characteristics of different process stages. The management layer sub-model is built based on energy consumption standards, emission policies, and scheduling rules, and is used to simulate the park's carbon emissions and pollutant control under different management strategies.

[0043] Based on the aforementioned multi-scale sub-models, intervention priority maps generated by causal analysis and counterfactual reasoning are embedded into each sub-model as policy-driven constraints. By introducing a causal constraint coupling mechanism between sub-models, the impact intensity of key pollution sources and synergistic emission reduction effects are transmitted across layers. For example, changes in equipment energy efficiency at the physical layer can be transmitted to the process layer, affecting the energy consumption and emissions of a certain production link; emission reduction measures at the process layer can then be fed back to the management layer for optimizing the overall scheduling strategy.

[0044] Based on the above cross-layer transmission and coupling modeling, a digital twin model of the park with multi-dimensional correlation characteristics is generated. This model can dynamically reflect the evolution of carbon emissions and pollutant distribution in a virtual environment and support the scenario-based simulation and effect evaluation of different pollution reduction and carbon reduction strategies.

[0045] Furthermore, the simulation assessment of implementing pollution reduction and carbon reduction strategies includes: A dynamic scenario generator is introduced into the multi-scale digital twin model of the park. The dynamic scenario generator is used to establish a scenario set based on energy price fluctuations, load changes, and meteorological conditions. Virtual intervention analysis of different pollution reduction and carbon reduction strategies is carried out under the scenario set, and simulation evaluation is performed.

[0046] A dynamic scenario generator is introduced into a multi-scale digital twin model of a park. This generator produces different sets of operational scenarios based on changes in external and internal conditions. Specifically, the generator establishes a multi-dimensional scenario set by setting parameters for energy price fluctuations (such as the upward and downward fluctuation ranges of electricity, natural gas, and steam prices), load change parameters (such as energy load adjustments caused by increases or decreases in production demand), and meteorological conditions (such as the impact of temperature, humidity, and wind speed on energy consumption and emissions). Each scenario represents the park's operational status under specific external conditions, thus providing a more realistic and comprehensive virtual environment for strategy evaluation. Under these scenario sets, the multi-scale digital twin model of the park is invoked, and different pollution reduction and carbon reduction strategies are loaded, such as process parameter optimization, equipment energy efficiency improvement, energy structure adjustment, and load shifting, to perform virtual intervention analysis. By comparing the changes in total carbon emissions, pollutant diffusion and distribution, energy utilization efficiency, and synergistic emission reduction effects before and after the intervention, different strategies are simulated, evaluated, and ranked. Finally, the emission reduction effects and adaptability of each strategy under different scenarios are output, providing quantitative decision-making basis for park managers.

[0047] Furthermore, the simulation assessment of implementing pollution reduction and carbon reduction strategies also includes: After anchoring the simulation evaluation results with key data, they are mapped to the blockchain; the anchored data is then encrypted and stored based on the mapping results.

[0048] After completing the virtual intervention analysis for different strategies, the simulation evaluation results are anchored to key data in the park's operation. This key data includes carbon emission monitoring values, energy consumption statistics, equipment operating status, and process parameters. By establishing this anchoring relationship, the simulation results are mapped one-to-one with actual observation data in both time and indicator dimensions, thus ensuring the traceability and reliability of the simulation results.

[0049] After data anchoring is completed, the anchored data is mapped to the blockchain system. As a distributed ledger, the blockchain ensures that the simulated evaluation results and their associated data are immutable and traceable. Specifically, a hash algorithm is used to digest the anchored data, generating a unique identifier, and this identifier, along with its corresponding evaluation result, is written to the blockchain nodes. Subsequently, the anchored data is encrypted and stored according to the mapping results, ensuring data security and privacy during transmission and storage, while providing a reliable basis for subsequent external audits, regulatory verifications, and cross-departmental sharing within the park.

[0050] Furthermore, the simulation assessment of implementing pollution reduction and carbon reduction strategies also includes: Set the data update cycle for the multi-source dataset; adaptively perform multi-source dataset updates during the data update cycle and record the data differences before and after the update; if the data differences meet a preset difference threshold, use the simulation evaluation result as the benchmark result, perform incremental prediction based on the data differences, and establish a prediction evaluation result.

[0051] A data update cycle is set for the multi-source dataset. This update cycle can be set according to the park's monitoring frequency, data acquisition capabilities, and management needs, for example, it can be on a minute, hour, or day basis. When the data update cycle is reached, an update operation for the multi-source dataset is adaptively triggered. The new monitoring data is compared with the data from the previous cycle, and the differences between the data before and after the update are recorded. These differences include numerical deviation, temporal fluctuation amplitude, and the number of outliers.

[0052] After recording the data discrepancies, the discrepancy values ​​are compared with a preset discrepancy threshold. If the discrepancy value does not exceed the threshold, it indicates that the overall operation of the park is stable, and the original simulation evaluation results can be maintained. If the discrepancy value exceeds the threshold, it indicates that the operation status has changed significantly, and the simulation results need to be corrected. In this case, the simulation evaluation results of the previous cycle are used as the baseline results, and incremental predictions are performed based on the newly added data discrepancy information. That is, only the variables and sub-models affected by the discrepancies are updated and calculated, rather than retraining and predicting the entire model. In this way, predictive evaluation results are generated quickly, reducing computational costs while ensuring the timeliness of the evaluation results.

[0053] In summary, the embodiments of this application have at least the following technical effects: First, multi-source data collection was performed on the park to establish a multi-source dataset, including data on carbon emissions, energy consumption, production processes, and equipment status. Confidence markers were dynamically generated for each dataset based on the reliability of the data sources. Next, the multi-source dataset was time-series aligned, and cross-modal consistency constraints and proactive anomaly remediation analysis were performed using the confidence markers. Adaptive data fusion was then performed on the multi-source dataset to establish a fused dataset. Further, using the fused dataset as input, a multi-modal machine learning model was used to predict the park's carbon footprint and pollutant distribution, establishing prediction results. Then, causal analysis and counterfactual reasoning were performed on the prediction results using a pollution backtracking channel. After identifying key pollution sources and hidden synergistic emission reduction points, an intervention priority map was generated based on the fused dataset, combining causal relationship strength and potential synergistic emission reduction effects. Finally, a multi-scale digital twin model of the park was constructed using the fused dataset and the intervention priority map, and a simulation evaluation of pollution reduction and carbon reduction strategies was performed. This approach solves the technical problem in existing technologies where pollution reduction and carbon reduction assessments rely on static statistics and lack dynamic data support, leading to insufficient reliability of assessment results. It achieves the technical effect of improving the reliability of assessment results through multi-source data fusion and causal reasoning.

[0054] Example 2, based on the same inventive concept as the AI-based pollution reduction and carbon reduction assessment method for zero-carbon parks described in the previous examples, such as... Figure 2 As shown, this application provides an AI-based pollution and carbon reduction assessment system for zero-carbon industrial parks, wherein the system includes: Data Acquisition Module 11: Performs multi-source data acquisition for the park, establishing a multi-source dataset including data on carbon emissions, energy consumption, production processes, and equipment status. The multi-source dataset dynamically generates confidence level indicators based on the reliability of the data sources. Data Fusion Module 12: After time-series alignment of the multi-source dataset, it uses the confidence level indicators to perform cross-modal consistency constraints and proactive anomaly repair analysis, adaptively fusing the multi-source dataset to establish a fused dataset. Prediction Module 13: Using the fused dataset as input data, it predicts the park's carbon footprint and pollutant distribution through a multimodal machine learning model, establishing prediction results. Map Generation Module 14: Utilizes pollution backtracking channels to perform causal analysis and counterfactual reasoning on the prediction results, identifying key pollution sources and hidden synergistic emission reduction points. Based on the fused dataset, it generates an intervention priority map combining causal relationship strength and potential synergistic emission reduction effects. Evaluation Module 15: Constructs a multi-scale digital twin model of the park using the fused dataset and the intervention priority map, and performs simulation evaluation of pollution reduction and carbon reduction strategies.

[0055] Furthermore, the map generation module 14 is used to perform the following method: The carbon footprint distribution and pollutant diffusion trajectory from the predicted results are input into the pollution backtracking channel. Carbon emission concentrations are time-series mapped to equipment operating conditions, process steps, and energy consumption. A causal constraint network is introduced into the pollution backtracking channel. This network is used to analyze the causal paths of local abnormal emission points in the predicted results, identifying direct process steps and indirect energy driving factors, and establishing causal path analysis results. Based on the causal path analysis results, key variables are virtually replaced and parameters are perturbed. Counterfactual reasoning is performed to establish counterfactual prediction scenarios. The difference between the counterfactual prediction scenarios and the predicted results is calculated, and the difference is quantified to determine the causal contribution of pollution sources, identifying key pollution sources. Finally, pollution synergy analysis is performed on different counterfactual prediction scenarios to identify hidden synergistic emission reduction points.

[0056] Furthermore, the prediction module 13 is used to perform the following method: Feature extraction of the fused dataset is performed using the preprocessing layer of the multimodal machine learning model, including extracting feature subsets of carbon emissions, energy consumption, production processes, and equipment status. A multi-scale window slicing mechanism is introduced in the time series dimension to establish cross-scale input features. A joint attention layer is constructed in the multimodal machine learning model, and mutual information constraint learning is performed on feature subsets of different modalities to capture the potential nonlinear coupling relationship between carbon emissions and pollutant diffusion. The deep features output by the joint attention layer are used to establish a carbon footprint time-series prediction branch and a pollutant spatial distribution prediction branch. The carbon footprint time-series prediction branch is constructed by extracting long-term dynamic trends through a residual recursive network, and the pollutant spatial distribution prediction branch is constructed by modeling the spatial topology of the park through a graph convolutional network. The carbon footprint time-series prediction branch and the pollutant spatial distribution prediction branch are subjected to multi-objective consistency regularization to establish the prediction results.

[0057] Furthermore, the evaluation module 15 is used to perform the following method: Based on the fused dataset, a multi-scale sub-model of the park's physical layer, process layer, and management layer is established. The intervention priority map is embedded into the multi-scale sub-model. Through the causal constraint coupling mechanism, the influence intensity of pollution sources and the synergistic emission reduction effect are transmitted between the sub-models, generating a multi-scale digital twin model of the park.

[0058] Furthermore, the evaluation module 15 is used to perform the following method: A dynamic scenario generator is introduced into the multi-scale digital twin model of the park. The dynamic scenario generator is used to establish a scenario set based on energy price fluctuations, load changes, and meteorological conditions. Virtual intervention analysis of different pollution reduction and carbon reduction strategies is carried out under the scenario set, and simulation evaluation is performed.

[0059] Furthermore, the data fusion module 12 is used to perform the following methods: Data integrity and consistency analysis are performed on the multi-source dataset. If an anomaly active repair mechanism is triggered, active data repair is performed through a dual-channel approach of causal relationship inference and time series prediction to establish active repair results.

[0060] Furthermore, the evaluation module 15 is used to perform the following method: After anchoring the simulation evaluation results with key data, they are mapped to the blockchain; the anchored data is then encrypted and stored based on the mapping results.

[0061] Furthermore, the evaluation module 15 is used to perform the following method: Set the data update cycle for the multi-source dataset; adaptively perform multi-source dataset updates during the data update cycle and record the data differences before and after the update; if the data differences meet a preset difference threshold, use the simulation evaluation result as the benchmark result, perform incremental prediction based on the data differences, and establish a prediction evaluation result.

[0062] Example 3: Based on the same inventive concept as the AI-based pollution reduction and carbon reduction assessment method for zero-carbon parks in the previous examples, this example provides a computer-readable storage medium for storing software programs, computer-executable programs, and modules, such as the program instructions / modules corresponding to the AI-based pollution reduction and carbon reduction assessment method for zero-carbon parks in this application. The processor executes the software programs, instructions, and modules stored in the memory to perform various functional applications and data processing of the computer device, thereby realizing the aforementioned AI-based pollution reduction and carbon reduction assessment method for zero-carbon parks.

[0063] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0064] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0065] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. An AI-based pollution reduction and carbon reduction assessment method for zero-carbon industrial parks, characterized in that, The method includes: Perform multi-source data collection in the park and establish a multi-source dataset. The multi-source dataset includes data on carbon emissions, energy consumption, production processes, and equipment status in the park. The multi-source dataset dynamically generates confidence level labels based on the reliability of the data sources. After aligning the multi-source datasets by time series, the confidence index is used to perform cross-modal consistency constraints and proactive anomaly repair analysis, and adaptive data fusion of the multi-source datasets is performed to establish a fused dataset. Using the fused dataset as input data, a multimodal machine learning model is used to predict the carbon footprint and pollutant distribution of the park, and prediction results are established. By utilizing pollution retrospective channels to perform causal analysis and counterfactual reasoning on the prediction results, key pollution sources and hidden synergistic emission reduction points are identified. Based on the fused dataset, an intervention priority map is generated by combining the strength of causal relationships and potential synergistic emission reduction effects. Using the fused dataset and intervention priority map, a multi-scale digital twin model of the park was constructed, and a simulation evaluation of pollution reduction and carbon reduction strategies was performed.

2. The AI-based pollution reduction and carbon reduction assessment method for zero-carbon industrial parks as described in claim 1, characterized in that, Using pollution tracing channels to perform causal analysis and counterfactual reasoning on predicted results, including: The carbon footprint distribution and pollutant diffusion trajectory in the prediction results are input into the pollution backtracking channel to perform a time-series mapping of carbon emission concentration with equipment operating conditions, process steps and energy consumption. A causal constraint network is introduced into the pollution tracing channel. The causal constraint network is used to analyze the causal path of local abnormal emission points in the prediction results, identify direct process links and indirect energy driving factors, and establish the causal path analysis results. Based on the causal path analysis results, key variables are virtually replaced and parameters are perturbed to perform counterfactual reasoning and establish counterfactual prediction scenarios. Calculate the difference between the counterfactual prediction scenario and the prediction result, and quantify the causal contribution of the pollution source by the difference to identify key pollution sources; Perform pollution synergy analysis under different counterfactual prediction scenarios to identify hidden synergistic emission reduction points.

3. The AI-based pollution reduction and carbon reduction assessment method for zero-carbon industrial parks as described in claim 1, characterized in that, Using the fused dataset as input data, a multimodal machine learning model is used to predict the park's carbon footprint and pollutant distribution, establishing prediction results, including: The preprocessing layer of the multimodal machine learning model is used to perform feature extraction of the fused dataset, including extracting feature subsets of carbon emissions, energy consumption, production processes and equipment status, and introducing a multi-scale window slicing mechanism in the time series dimension to establish cross-scale input features; A joint attention layer is constructed in the multimodal machine learning model, and the joint attention layer is used to perform mutual information constraint learning on feature subsets of different modalities to capture the potential nonlinear coupling relationship between carbon emissions and pollutant diffusion. A carbon footprint temporal prediction branch and a pollutant spatial distribution prediction branch are established using the deep features output by the joint attention layer. The carbon footprint temporal prediction branch is constructed by extracting long-term dynamic trends through a residual recursive network, and the pollutant spatial distribution prediction branch is constructed by modeling the spatial topology of the park through a graph convolutional network. The carbon footprint time-series prediction branch and the pollutant spatial distribution prediction branch are subjected to multi-objective consistency regularization to establish the prediction results.

4. The AI-based pollution reduction and carbon reduction assessment method for zero-carbon industrial parks as described in claim 1, characterized in that, A multi-scale digital twin model of the park is constructed using the fused dataset and intervention priority map, including: Based on the fused dataset, a multi-scale sub-model of the park's physical layer, process layer, and management layer is established. The intervention priority map is embedded into the multi-scale sub-model. Through the causal constraint coupling mechanism, the influence intensity of pollution sources and the synergistic emission reduction effect are transmitted between the sub-models, generating a multi-scale digital twin model of the park.

5. The AI-based pollution reduction and carbon reduction assessment method for zero-carbon industrial parks as described in claim 4, characterized in that, Simulation assessments of implementing pollution reduction and carbon reduction strategies include: A dynamic scenario generator is introduced into the multi-scale digital twin model of the park, and the dynamic scenario generator is used to establish a scenario set based on energy price fluctuations, load changes, and meteorological conditions. Virtual intervention analysis of different pollution reduction and carbon reduction strategies was conducted under the aforementioned scenario set, and simulation evaluation was performed.

6. The AI-based pollution reduction and carbon reduction assessment method for zero-carbon industrial parks as described in claim 1, characterized in that, The analysis of cross-modal consistency constraints and proactive anomaly repair using the confidence level identifier includes: Data integrity and consistency analysis are performed on the multi-source dataset. If an anomaly active repair mechanism is triggered, active data repair is performed through a dual-channel approach of causal relationship inference and time series prediction to establish active repair results.

7. The AI-based pollution reduction and carbon reduction assessment method for zero-carbon industrial parks as described in claim 1, characterized in that, The simulation assessment of implementing pollution reduction and carbon reduction strategies also includes: After anchoring the simulation evaluation results with key data, they are mapped to the blockchain; The anchored data is encrypted and stored based on the mapping results.

8. The AI-based pollution reduction and carbon reduction assessment method for zero-carbon industrial parks as described in claim 1, characterized in that, The simulation assessment of implementing pollution reduction and carbon reduction strategies also includes: Set the data update cycle for multi-source datasets; The multi-source dataset is adaptively updated during the data update cycle, and the data differences before and after the update are recorded. If the data difference meets the preset difference threshold, the simulation evaluation result is used as the benchmark result, and incremental prediction is performed based on the data difference to establish the prediction evaluation result.

9. An AI-based pollution reduction and carbon reduction assessment system for zero-carbon industrial parks, characterized in that, The system is used to implement the AI-based pollution reduction and carbon reduction assessment method for zero-carbon industrial parks as described in any one of claims 1-8, the system comprising: Data acquisition module: Performs multi-source data acquisition for the park and establishes a multi-source dataset. The multi-source dataset includes data on carbon emissions, energy consumption, production processes, and equipment status of the park. The multi-source dataset dynamically generates confidence level labels based on the reliability of the data sources. Data fusion module: After aligning the multi-source datasets by time series, it uses the confidence index to perform cross-modal consistency constraints and proactive anomaly repair analysis, adaptively fuses the multi-source datasets, and establishes a fused dataset; Prediction module: Using the fused dataset as input data, it predicts the carbon footprint and pollutant distribution of the park through a multimodal machine learning model and establishes prediction results; Map generation module: After performing causal analysis and counterfactual reasoning on the prediction results using the pollution backtracking channel, identifying key pollution sources and hidden synergistic emission reduction points, it generates an intervention priority map based on the fused dataset, combining the strength of causal relationships and potential synergistic emission reduction effects. Evaluation module: Utilize the fused dataset and intervention priority map to construct a multi-scale digital twin model of the park, and perform simulation evaluation of pollution reduction and carbon reduction strategies.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements an AI-based pollution reduction and carbon reduction assessment method for zero-carbon industrial parks as described in any one of claims 1-8.

Images