A multidimensional dynamic carbon emission factor modeling method and system

Abstract

The present invention discloses a multi-dimensional dynamic carbon emission factor modeling method and system, which specifically relates to the field of carbon emission monitoring technology, and is used to solve the problem that existing methods cannot accurately characterize the impact of local abnormal emissions on global assessment in cross-scale data fusion; it is through collecting multi-dimensional monitoring data to construct a data set, screening key feature subsets based on statistical tests, establishing a correlation matrix to extract cross-dimensional main factors and generate hierarchical weight parameters, constructing a gradient mask feature importance network to optimize feature distribution, training multi-dimensional special sub-models and dynamically fusing output results, combining incremental learning and transfer learning to update model parameters and filter noise, and realizing dynamic assessment based on Monte Carlo simulation calibration uncertainty interval; through cross-dimensional dynamic feature fusion and hierarchical optimization mechanism, the sensitivity of carbon emission factor modeling and the reliability of assessment results are effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of carbon emission monitoring, and more specifically, to a multidimensional dynamic carbon emission factor modeling method and system. Background Art

[0002] When integrating data at different scales, existing carbon emission factor modeling methods usually adopt a unified aggregation strategy to process device-level monitoring data and regional-level energy data. Such methods standardize and integrate data at different levels based on preset rules, and fail to effectively distinguish the differential impacts of key local fluctuations and overall trends.

[0003] In the process of cross-scale data fusion, the existing methods fail to accurately characterize the contribution of local abnormal emission events to the global assessment due to the insufficient response of the standardized integration mechanism to the differences between micro-dynamic characteristics and macro-steady-state characteristics. This distortion in the transmission of influence between data levels will reduce the monitoring model's sensitivity to identifying key risk signals, forming a bottleneck in cross-scale data fusion. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide a multi-dimensional dynamic carbon emission factor modeling method and system to solve the problems raised in the above-mentioned background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A multi-dimensional dynamic carbon emission factor modeling method, including:

[0007] S1. Collect multi-dimensional monitoring data to construct a multi-dimensional dataset, and screen the key feature subsets that are significantly correlated with carbon emission factors based on statistical tests;

[0008] S2. Establish a correlation matrix between key feature subsets and carbon emission factors, extract cross-dimensional main factors through principal component analysis, quantify the nonlinear dependence intensity of each cross-dimensional main factor, and generate hierarchical weight parameters;

[0009] S3. Construct a gradient mask-based feature importance network based on hierarchical weight parameters, optimize the feature distribution of the multidimensional dataset through feature channel scaling, and generate an initial prediction model;

[0010] S4. Build dedicated sub-models for multiple preset dimensions, and use a dynamic weight allocation mechanism to fuse the output results of the dedicated sub-models to generate a multi-scenario fused carbon emission factor;

[0011] S5. Update the initial prediction model based on the incremental learning mechanism and generate historical verification data. Combined with transfer learning, transfer historical knowledge to the dedicated sub-model and filter input data noise.

[0012] S6. Calibrate the uncertainty interval of the multi-scenario fusion carbon emission factor based on historical verification data and output the dynamic optimization evaluation results.

[0013] In a preferred embodiment, multi-dimensional monitoring data is collected to construct a multi-dimensional dataset, and a subset of key features significantly correlated with carbon emission factors is screened based on statistical tests, including:

[0014] Collect monitoring indicators in fuel attribute dimensions, regional characteristics dimensions, time dynamic dimensions, process technology dimensions, equipment efficiency dimensions, and environmental parameter dimensions to construct a multidimensional data set;

[0015] After standardization preprocessing of the multidimensional data set, the chi-square test or mutual information method is used to calculate the correlation between the characteristics of each dimension and the carbon emission factor, and the features with correlation higher than the preset significance threshold are retained to form the key feature subset.

[0016] In a preferred embodiment, the fuel attribute dimension includes fuel type, calorific value and carbon content indicators, the regional characteristic dimension includes regional energy structure and policy intensity indicators, and the time dynamic dimension includes seasonal energy consumption fluctuations and equipment operation cycle indicators.

[0017] In a preferred embodiment, a correlation matrix between key feature subsets and carbon emission factors is established, cross-dimensional principal factors are extracted through principal component analysis, the nonlinear dependency strength of each cross-dimensional principal factor is quantified, and hierarchical weight parameters are generated, including:

[0018] Calculate the Spearman rank correlation coefficient between each dimension feature and the carbon emission factor in the key feature subset, and select features whose absolute value of the Spearman rank correlation coefficient is greater than the preset correlation threshold to generate an initial correlation matrix;

[0019] Perform eigenvalue decomposition on the initial correlation matrix and extract the principal components whose cumulative contribution rate exceeds the preset contribution threshold as the cross-dimensional principal factors;

[0020] The kernel density estimation method is used to fit the joint probability distribution of the cross-dimensional main factors and carbon emission factors, and the nonlinear dependence strength between the two is calculated based on the mutual information entropy formula;

[0021] The weight distribution ratio is calculated based on the entropy weight method combined with the nonlinear dependency intensity of each cross-dimensional main factor to generate the hierarchical weight parameters.

[0022] In a preferred embodiment, a gradient mask-based feature importance network is constructed based on hierarchical weight parameters, and the feature distribution of a multidimensional dataset is optimized by feature channel scaling to generate an initial prediction model, including:

[0023] The hierarchical weight parameters are used as the initial feature importance weights, and the weight gradients of each feature channel are dynamically updated through gradient backpropagation to generate dynamic feature importance masks;

[0024] Multiply the dynamic feature importance mask with the multidimensional dataset channel by channel to suppress the noise of low-importance features and enhance the key feature signals to generate an optimized multidimensional feature distribution;

[0025] Based on the optimized multidimensional feature distribution, the gradient boosting tree model architecture is selected, and the model hyperparameters are adjusted using cross-validation and Bayesian optimization to generate an initial prediction model.

[0026] In a preferred embodiment, dedicated sub-models are constructed for multiple preset dimensions, and the output results of the dedicated sub-models are integrated using a dynamic weight allocation mechanism to generate a multi-scenario integrated carbon emission factor, including:

[0027] Based on the optimized multi-dimensional feature distribution, the gradient boosting tree sub-models corresponding to the fuel attribute dimension, regional feature dimension, and process technology dimension are trained respectively;

[0028] Calculate the dynamic weight based on the historical prediction error of each gradient boosting tree sub-model, and sum the prediction results of the gradient boosting tree sub-model according to the dynamic weight;

[0029] The weighted summation results are normalized and the multi-scenario fusion carbon emission factor is output.

[0030] In a preferred embodiment, the initial prediction model is updated based on an incremental learning mechanism and historical verification data is generated. In combination with transfer learning, historical knowledge is transferred to a dedicated sub-model and input data noise is filtered, including:

[0031] The newly added monitoring data is input into the initial prediction model, and the model parameters are incrementally updated through the online sequence extreme learning machine to generate an updated initial prediction model and historical verification data;

[0032] Extract the weight distribution features of the updated initial prediction model and adapt them to the weight space of the gradient boosting tree sub-model through the feature mapping function;

[0033] The outliers in the newly added monitoring data are detected based on the isolation forest algorithm, and then input into the gradient boosting tree sub-model after removing the outliers.

[0034] In a preferred embodiment, the uncertainty interval of the multi-scenario fusion carbon emission factor is calibrated based on historical verification data, and the dynamic optimization assessment results are output, including:

[0035] Generate the probability distribution of carbon emission factors for multiple scenarios based on Monte Carlo simulation, and calculate the confidence interval with a preset threshold confidence level;

[0036] The upper and lower thresholds of the confidence interval are adjusted according to the preset rules to generate dynamic optimization evaluation results.

[0037] In a preferred embodiment, the execution steps of the Monte Carlo simulation include: randomly sampling from historical verification data to generate a simulation data set, iteratively calculating the distribution statistics of the multi-scenario fusion carbon emission factor; the preset rules include the statistical laws of historical data and external constraints.

[0038] In another aspect, the present invention provides a multi-dimensional dynamic carbon emission factor modeling system, comprising:

[0039] Multidimensional acquisition module: collects multidimensional monitoring data to construct a multidimensional dataset, and screens key feature subsets that are significantly correlated with carbon emission factors based on statistical tests;

[0040] Cross-dimensional analysis module: establishes a correlation matrix between key feature subsets and carbon emission factors, extracts cross-dimensional main factors through principal component analysis, quantifies the nonlinear dependency strength of each cross-dimensional main factor, and generates hierarchical weight parameters;

[0041] Weight Optimization Module: Constructs a gradient mask-based feature importance network based on hierarchical weight parameters, optimizes the feature distribution of multidimensional datasets through feature channel scaling, and generates an initial prediction model;

[0042] Dynamic Fusion Module: Builds dedicated sub-models for multiple preset dimensions, uses a dynamic weight allocation mechanism to fuse the output results of the dedicated sub-models, and generates a multi-scenario fused carbon emission factor;

[0043] Incremental migration module: Based on the incremental learning mechanism, it updates the initial prediction model and generates historical verification data. It also combines transfer learning to migrate historical knowledge to a dedicated sub-model and filter input data noise.

[0044] Dynamic calibration module: calibrates the uncertainty interval of the multi-scenario fusion carbon emission factor based on historical verification data, and outputs dynamic optimization evaluation results.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] 1. Through the multi-dimensional dynamic feature fusion and hierarchical weight optimization mechanism, the ability of carbon emission factor modeling to analyze the dynamic characteristics of cross-scale data is effectively improved. Based on the nonlinear dependency intensity quantification of key feature subsets and the gradient mask feature importance network, the contribution of local abnormal emission events to the global assessment is accurately identified, solving the problem of feature contribution distortion caused by linear aggregation in traditional methods. Through the coordinated fusion of dynamic weight allocation and multi-scenario sub-models, the model's adaptability to multi-dimensional differences such as region and process is enhanced, avoiding the risk of scenario bias under static weight allocation. Incremental learning and transfer learning are combined with historical verification data to continuously optimize model parameters to ensure the real-time response capability of the modeling process to the dynamic evolution of data. At the same time, noise filtering is used to improve the quality of input data and ensure model robustness.

[0047] 2. Through a dynamic uncertainty interval calibration mechanism, Monte Carlo simulations are integrated with external constraints to ensure the reliability and compliance of assessment results across multiple scenarios. The collaborative design of cross-dimensional principal factor extraction and feature channel scaling overcomes the technical bottleneck of traditional methods that hinder the compatibility of macro-steady-state features with micro-dynamic features, enabling the model to maintain high sensitivity and generalization performance even in complex data environments. By integrating closed-loop data flows and logical coupling to form a complete technical chain, this system systematically addresses issues such as local signal omissions, poor scenario adaptability, and inaccurate assessment results in carbon emissions monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is a flow chart of a multi-dimensional dynamic carbon emission factor modeling method of the present invention;

[0049] Figure 2 This is a structural schematic diagram of a multi-dimensional dynamic carbon emission factor modeling system of the present invention. DETAILED DESCRIPTION

[0050] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0051] Example 1: Figure 1 The present invention provides a multidimensional dynamic carbon emission factor modeling method, including:

[0052] S1. Collect multi-dimensional monitoring data to construct a multi-dimensional dataset, and screen the key feature subsets that are significantly correlated with carbon emission factors based on statistical tests;

[0053] S2. Establish a correlation matrix between key feature subsets and carbon emission factors, extract cross-dimensional main factors through principal component analysis, quantify the nonlinear dependence intensity of each cross-dimensional main factor, and generate hierarchical weight parameters;

[0054] S3. Construct a gradient mask-based feature importance network based on hierarchical weight parameters, optimize the feature distribution of the multidimensional dataset through feature channel scaling, and generate an initial prediction model;

[0055] S4. Build dedicated sub-models for multiple preset dimensions, and use a dynamic weight allocation mechanism to fuse the output results of the dedicated sub-models to generate a multi-scenario fused carbon emission factor;

[0056] S5. Update the initial prediction model based on the incremental learning mechanism and generate historical verification data. Combined with transfer learning, transfer historical knowledge to the dedicated sub-model and filter input data noise.

[0057] S6. Calibrate the uncertainty interval of the multi-scenario fusion carbon emission factor based on historical verification data and output the dynamic optimization evaluation results.

[0058] S1. Collect multi-dimensional monitoring data to construct a multi-dimensional dataset and select a subset of key features that are significantly correlated with carbon emission factors based on statistical tests. The specific implementation steps are as follows:

[0059] Fuel attribute monitoring indicators are collected, including fuel type, calorific value, and carbon content. Fuel type is obtained from the fuel classification list provided by the energy supplier. This list follows the fuel classification standards published by the International Energy Agency and includes three major categories: fossil fuels, biomass fuels, and synthetic fuels. Each category is further subdivided into specific fuel names, such as coal, which is divided into anthracite and lignite, and natural gas, which is divided into liquefied natural gas and pipeline natural gas. Calorific value is obtained through laboratory testing or standard calorific value tables provided by the fuel supplier. Laboratory testing uses oxygen bomb calorimetry, where a fuel sample is completely burned in an oxygen bomb calorimeter and the calorific value released is measured in megajoules per kilogram (MJ / kg). Standard calorific value tables reference the fuel calorific value database published by the International Organization for Standardization, matching the corresponding calorific value by fuel type index. Carbon content is measured using an elemental analyzer or references standard values ​​from the International Carbon Content Database. When measuring with an elemental analyzer, the fuel sample is burned and then analyzed for carbon dioxide emissions using gas chromatography. This is then converted to carbon emission intensity per unit of calorific value, in kilograms of carbon dioxide equivalent per megajoule (kg CO2e / MJ).

[0060] Monitoring indicators are collected based on regional characteristics, including regional energy structure and policy intensity indicators. Regional energy structure is obtained from annual energy consumption statistical reports published by local governments. The share of coal-fired power, hydropower, and wind power in these reports is calculated by dividing total energy consumption by the consumption of each energy type. The calculation formula is: the share of a particular energy source is equal to the consumption of that energy source divided by total energy consumption, and the result is expressed as a percentage. Policy intensity indicators are quantified using carbon emission regulatory policy documents published by environmental protection departments. Carbon tax rates are directly extracted from these documents, expressed in yuan per ton of carbon dioxide equivalent (yuan / tCO2e). The emission reduction target completion rate is calculated as the ratio of actual emission reductions to target emission reductions, and the result is expressed as a percentage.

[0061] Monitoring indicators for the time-dynamic dimension are collected, including seasonal energy consumption fluctuations and equipment operation cycle indicators. Seasonal energy consumption fluctuations are extracted from historical energy consumption records, and the standard deviation of each month's energy consumption is calculated on a quarterly basis. The standard deviation is calculated as the square root of the mean of the sum of the squared deviations of each month's energy consumption from the quarterly average energy consumption. This standard deviation represents the magnitude of energy consumption fluctuations, expressed in megawatt-hours (MWh). Equipment operation cycle indicators are collected through equipment sensors and include continuous operation time and downtime maintenance intervals. Continuous operation time is the cumulative number of hours (hours) between equipment startup and shutdown, and downtime maintenance intervals are the number of days (days) between two consecutive maintenance operations.

[0062] Monitoring indicators of process technology dimensions are collected, and process technology dimensions include production process parameters and technology iteration data. Production process parameters are obtained through the factory production management system, which records the reaction temperature, pressure and raw material ratio in real time. The reaction temperature unit is degrees Celsius (°C), the pressure unit is kilopascals (kPa), and the raw material ratio is the mass proportion of each raw material, expressed as a percentage (%). Technology iteration data is recorded in the company's R&D log, including the application time of the new process and the percentage of energy efficiency improvement. The percentage of energy efficiency improvement is calculated by dividing the difference in unit product energy consumption between the new and old processes by the energy consumption of the old process. The calculation formula is that the percentage of energy efficiency improvement is equal to (old process energy consumption - new process energy consumption) divided by the old process energy consumption multiplied by 100%.

[0063] Collect monitoring indicators related to equipment effectiveness, which includes operating efficiency and maintenance status parameters. Equipment operating efficiency is calculated by real-time monitoring of the equipment's energy consumption-to-output ratio. The calculation formula is: equipment operating efficiency equals the total mass of output (kg) divided by the equipment's total energy consumption (kWh), expressed in kilograms per kilowatt-hour (kg / kWh). Maintenance status parameters are obtained from equipment maintenance records, including the most recent maintenance date and fault code frequency. Fault code frequency is calculated as the number of times a specific fault code occurs per 1,000 hours of operation (times / 1000h).

[0064] Environmental parameters include temperature, humidity, and air pressure. Temperature and humidity are collected in real time via IoT sensors that meet industrial standards. Temperature measurement accuracy is ±0.5 degrees Celsius (°C), humidity measurement accuracy is ±3%, and data is collected once per minute. Air pressure is obtained via the weather station data interface, which updates hourly in kilopascals (kPa).

[0065] The monitoring indicators for all the above dimensions are aligned by timestamp to construct a multidimensional dataset. The specific alignment method is: using minute-level timestamps as the benchmark, linearly interpolate hourly or daily data (such as weather station air pressure) to fill in the minute level, ensuring that all dimensional data has the same time granularity. The dataset is stored as a two-dimensional table, with rows representing time series samples and columns representing the monitoring indicators for each dimension. Missing values ​​are handled using a forward filling method, which fills the current missing value with the valid value at the previous moment.

[0066] The multidimensional dataset was preprocessed for standardization using the Z-score standardization method. Z-score standardization was calculated for each monitoring indicator separately. The calculation formula is: the standardized value is equal to the original value minus the mean of the indicator, divided by the standard deviation of the indicator. After processing, each indicator has a mean of 0 and a standard deviation of 1. The mean and standard deviation are calculated based on the training set data, and the same parameters are used in the test set data to avoid data leakage.

[0067] Based on the standardized multidimensional data set, the chi-square test or mutual information method is used to calculate the correlation between each dimensional feature and the carbon emission factor. The chi-square test execution steps include: discretizing the continuous monitoring indicator using the equal-width binning method, and dynamically adjusting the number of bin intervals based on the data distribution. The specific rules are as follows: if the data distribution is approximately normal, the number of bins is 10; if the data distribution is skewed, the number of bins is 5; counting the actual frequency of the carbon emission factor in each bin interval and calculating the theoretical expected frequency, which is equal to the number of samples in the bin multiplied by the average distribution probability of the overall carbon emission factor; calculating the chi-square statistic, which is the sum of the square of the difference between the actual frequency and the expected frequency in each bin; consulting the chi-square distribution table based on the chi-square statistic and the degrees of freedom to obtain the P value, and retaining features with a P value less than 0.05; the 0.05 significance threshold is determined according to the general standard for statistical significance level.

[0068] The implementation steps of the mutual information method include: using the kernel density estimation method to fit the joint probability distribution of continuous monitoring indicators and carbon emission factors, selecting the Gaussian kernel as the kernel function, and automatically calculating the bandwidth parameter through the Silverman rule. The calculation formula is that the bandwidth is equal to 1.06 multiplied by the sample standard deviation multiplied by the negative fifth power of the sample size; calculating the marginal probability distribution of the monitoring indicators and carbon emission factors, and the marginal probability is obtained by integrating the joint probability distribution; calculating the mutual information value based on the entropy formula, and the expected value of the logarithm ratio of the joint probability to the marginal probability in the entropy formula is the mutual information value; sorting the mutual information values ​​of all features in descending order, and retaining the top 30% of the features as significantly correlated features; the 30% quantile threshold is determined through experimental verification, and the experimental results show that the model F1 score reaches above 0.85 under this threshold.

[0069] The results of the chi-square test and the mutual information method were combined, and features identified as significantly correlated by at least one method were retained to form a key feature subset. This key feature subset included fuel type, calorific value, regional energy structure, seasonal energy consumption fluctuations, equipment operating efficiency, and temperature and humidity indicators. The key feature subset was stored in a two-dimensional table, with rows representing time series samples and columns representing filtered feature indicators. This was used in the subsequent correlation matrix construction step.

[0070] S2. Establish a correlation matrix between key feature subsets and carbon emission factors, extract cross-dimensional principal factors through principal component analysis, quantify the nonlinear dependency strength of each cross-dimensional principal factor, and generate hierarchical weight parameters. The specific implementation steps are as follows:

[0071] Based on the key feature subset, the Spearman rank correlation coefficient of each dimensional feature and the carbon emission factor is calculated. The Spearman rank correlation coefficient is calculated by the following steps: sort the dimensional features and carbon emission factors in the key feature subset according to the numerical size and assign a rank. The rank is the ranking position of the data point in the sequence to which it belongs. For example, the data point with the smallest numerical value has a rank of 1, the second smallest has a rank of 2, and so on; calculate the sum of the squares of the rank differences of each pair of data points. The sum of the squares of the rank differences is the squares of the rank differences of the corresponding data points and then add them up; generate the correlation coefficient value according to the formula, which is 1 minus 6 times the sum of the squares of the rank differences divided by the sample size multiplied by the difference of the sample size squared minus 1. The correlation coefficient value ranges from -1 to 1. The larger the absolute value, the stronger the linear correlation. The default correlation threshold was set at 0.3, determined through cross-validation experiments. The specific experimental method involved trying thresholds of 0.2, 0.3, and 0.4 on the training set. The model was constructed by calculating the features selected at each threshold and evaluating the prediction accuracy on the test set. The results showed that a threshold of 0.3 reduced the root mean square error of the model by 15% compared to other thresholds, so 0.3 was selected as the optimal threshold. Features with absolute correlation coefficients greater than 0.3 were selected to generate an initial correlation matrix. The rows of the initial correlation matrix represent features in the key feature subset, the columns represent carbon emission factors, and the matrix elements are the corresponding Spearman rank correlation coefficient values. The matrix dimension is M × N, where M is the number of features in the key feature subset and N is the number of carbon emission factors.

[0072] Perform eigenvalue decomposition on the initial correlation matrix to calculate the matrix's eigenvalues ​​and corresponding eigenvectors. Eigenvalue decomposition is achieved by decomposing the matrix into the product of an eigenvector matrix and a diagonal eigenvalue matrix. Arrange the eigenvalues ​​in descending order, and calculate the cumulative contribution rate. The cumulative contribution rate is the percentage of the sum of the first N eigenvalues ​​to the total sum of all eigenvalues. For example, if the sum of the first three eigenvalues ​​is 80% of the total, and the sum of the first four eigenvalues ​​is 90% of the total, then the cumulative contribution rates are 80% and 90%, respectively. A preset contribution threshold is set at 85%, meaning that principal component extraction stops when the cumulative contribution rate reaches 85%. This threshold is set according to industry practice for principal component analysis to ensure that key information is retained while reducing data dimensionality. The principal components corresponding to the first N eigenvectors that meet the cumulative contribution rate threshold are used as cross-dimensional principal factors. Cross-dimensional principal factors represent the global coupling relationship between the characteristics of each dimension in the key feature subset and the carbon emission factor. For example, the first principal component can be interpreted as the combined influencing factor of fuel properties and regional energy structure.

[0073] The kernel density estimation method is used to fit the joint probability distribution of the cross-dimensional principal factors and the carbon emission factors. The kernel function of the kernel density estimation uses a Gaussian kernel. The Gaussian kernel function is a symmetric bell-shaped curve whose width is controlled by the bandwidth parameter. The bandwidth parameter is calculated using the Silverman rule. The calculation formula of the Silverman rule is that the bandwidth is equal to 1.06 multiplied by the sample standard deviation multiplied by the negative fifth power of the sample size. Based on the joint probability distribution and the marginal probability distribution, the nonlinear dependence strength of the cross-dimensional principal factors and the carbon emission factors is calculated using the mutual information entropy formula. The calculation steps of the mutual information entropy include: calculating the logarithm of the ratio of the product of the joint probability distribution and the marginal probability distribution, where the ratio is the joint probability distribution divided by the product of the marginal probability distribution; and finding the expected value of all possible values, that is, taking the weighted sum of the logarithms of the ratio under the joint probability distribution, with the weight being the corresponding joint probability value. The value range of nonlinear dependence intensity is 0 to positive infinity. The larger the value, the stronger the explanatory power of the cross-dimensional main factor on the carbon emission factor. For example, the nonlinear dependence intensity of a main factor is 1.2, which means that its explanatory power on the carbon emission factor is significantly higher than that of other main factors with a dependence intensity of 0.8.

[0074] Based on the nonlinear dependency strength of each cross-dimensional principal factor, the entropy weighting method is used to calculate the weight distribution ratio. The entropy weighting method includes the following steps: normalizing the nonlinear dependency strength to a probability distribution by dividing the dependency strength value of each principal factor by the sum of the dependency strength values ​​of all principal factors, so that the sum of the normalized values ​​is 1; calculating the information entropy value of each principal factor, which is the cumulative sum of the negative logarithm of the normalized value and its product; calculating the variance coefficient based on the information entropy value, which is equal to 1 minus the information entropy value. A larger variance coefficient indicates a higher weight for the principal factor; and finally, the weight is the ratio of the variance coefficient to the sum of the variance coefficients of all principal factors. The generated hierarchical weight parameters are stored in vector form. The vector elements represent the importance weight of each cross-dimensional principal factor in carbon emission factor modeling, and the weights sum to 1. For example, the weights of the three principal factors are 0.5, 0.3, and 0.2, respectively, indicating that the first principal factor has the highest contribution to the carbon emission factor.

[0075] Among them, the Spearman correlation coefficient threshold of 0.3 was determined through cross-validation. Experiments showed that the root mean square error of the model was reduced by 15% under this threshold; the cumulative contribution rate threshold of 85% was set according to the industry practice of principal component analysis to ensure that the main information was retained after dimensionality reduction; the kernel density estimation bandwidth parameter was automatically calculated by the Silverman rule to avoid manual intervention bias.

[0076] This step S2 constructs a correlation matrix using the Spearman correlation coefficient to capture the nonlinear correlation between features and carbon emission factors, breaking through the limitations of traditional linear correlation analysis; principal component analysis extracts cross-dimensional principal factors, solves the problem of multi-dimensional data redundancy, and retains key coupling information; kernel density estimation and mutual information entropy quantify the intensity of nonlinear dependencies and accurately identify the influence weights of cross-scale dynamic features; the entropy weight method combines information entropy to dynamically assign weights to avoid subjective pre-set bias. Compared with the data fusion methods based on mean aggregation or fixed weights in the existing technology, this step significantly improves the model's ability to analyze local abnormal signals and cross-dimensional interaction effects through nonlinear statistics and dynamic weight allocation, solves the problem of feature contribution distortion caused by linear assumptions in traditional methods, and enhances the sensitivity and scenario generalization of carbon emission factor modeling.

[0077] S3. Construct a gradient mask-based feature importance network based on the hierarchical weight parameters, optimize the feature distribution of the multidimensional dataset through feature channel scaling, and generate an initial prediction model. The specific implementation steps are as follows:

[0078] Using layer-wise weight parameters as initial feature importance weights, gradient backpropagation is used to dynamically update the weight gradients of each feature channel to generate a dynamic feature importance mask. The layer-wise weight parameters are derived from the calculation of the nonlinear dependency strength of the principal factors across dimensions and are stored as vectors, with each element representing the initial importance weight of each feature channel. Gradient backpropagation is implemented based on a neural network architecture. The neural network input is a multidimensional dataset, and the output is an estimate of the predicted carbon emission factor. The loss function uses the mean squared error (MSE) as the loss function. During backpropagation, the gradient of the loss function with respect to each feature channel weight is calculated. The gradient value reflects the contribution of the feature channel to the prediction error. The initial importance weights are dynamically updated to generate a dynamic feature importance mask. The dynamic feature importance mask is a weight matrix with the same dimension as the channels of the multidimensional dataset. The matrix elements range from 0 to 1, with larger values ​​indicating higher importance of the corresponding feature channel. For example, if the initial weight of a feature channel is 0.5, and after gradient update, the weight increases to 0.9, this indicates that the channel's contribution to the prediction result has significantly increased.

[0079] The dynamic feature importance mask is multiplied by the multidimensional dataset channel-by-channel to suppress noise from low-importance features and enhance the signal from key features, generating an optimized multidimensional feature distribution. This channel-by-channel multiplication is performed separately for each feature channel in the multidimensional dataset, multiplying the original data value of the feature channel by the weight value of the corresponding channel in the dynamic feature importance mask to achieve feature channel scaling. For example, if the original data value of a feature channel is 100 and the mask weight is 0.8, the scaled data value is 80; if the mask weight is 0.2, the scaled data value is 20. This operation suppresses data values ​​from low-weight feature channels while retaining or enhancing data values ​​from high-weight feature channels, forming an optimized multidimensional feature distribution. The optimized multidimensional feature distribution is stored in a two-dimensional table, with rows representing time series samples and columns representing scaled feature channel data values.

[0080] Based on the optimized multidimensional feature distribution, a gradient boosting tree model architecture was selected. Model hyperparameters were adjusted using cross-validation and Bayesian optimization to generate an initial prediction model. The tree depth of the gradient boosting tree model architecture was set between 3 and 8 layers, and the learning rate was set between 0.01 and 0.1. These hyperparameter ranges were determined based on experience and experimental validation. For example, a tree depth exceeding 8 layers is prone to overfitting, while a depth below 3 layers can result in underfitting. Cross-validation was performed using a 5-fold cross-validation algorithm. The optimized multidimensional feature distribution was divided into five subsets, with four subsets used as training sets and one as validation sets. The average model prediction accuracy was calculated as the evaluation metric. Bayesian optimization used Gaussian process regression to model the relationship between hyperparameters and model accuracy. The optimal hyperparameter combination was iteratively searched for with a set of 50 iterations. The optimization was terminated early when the accuracy improvement was less than 1% after 10 consecutive iterations. This termination criterion was determined through experimental validation and demonstrated to balance computational efficiency and optimization effectiveness. The initial prediction model, containing the trained gradient boosting tree structure and hyperparameter configuration, was stored in a serialized file format for subsequent incremental learning and transfer learning steps.

[0081] The depth range of the gradient boosting tree (3-8 layers) was determined through experimental verification. Experiments have shown that exceeding this range can easily lead to overfitting or underfitting. The number of Bayesian optimization iterations, 50, was determined through convergence testing to ensure a balance between search efficiency and accuracy. The early termination condition (improvement of <1% for 10 consecutive iterations) was set based on the statistical results of historical training data to avoid invalid calculations.

[0082] This step, S3, uses dynamic feature importance masks to adaptively scale feature channels, overcoming the limitations of traditional static weight assignments to accurately identify key features and suppress noise interference. Combining gradient boosting trees with Bayesian optimization, it dynamically adapts to data distribution characteristics, improving the model's ability to analyze cross-dimensional interactions. Compared to existing technologies that use fixed feature weights or a single model architecture, this step dynamically updates feature importance through gradient backpropagation. Combined with cross-validation and hyperparameter optimization, it effectively addresses the problem of insufficient model sensitivity caused by feature redundancy and noise, enhancing the accuracy of carbon emission factor predictions and the ability to respond to local anomalies.

[0083] S4. Build dedicated sub-models for multiple preset dimensions, use a dynamic weight allocation mechanism to fuse the output results of the dedicated sub-models, and generate a multi-scenario fused carbon emission factor. The specific implementation steps are as follows:

[0084] Based on the optimized multidimensional feature distribution, gradient boosting tree sub-models corresponding to the fuel attribute dimension, regional feature dimension, and process technology dimension are trained separately. The optimized multidimensional feature distribution is derived from the output of the feature channel scaling step and is stored in a two-dimensional table format, with rows representing time series samples and columns representing the scaled feature channel data values. For the fuel attribute dimension, feature channels related to fuel type, calorific value, and carbon content are selected from the optimized multidimensional feature distribution. Fuel type is identified by string matching, with columns containing the keywords "fuel type," "calorific value," and "carbon content" in the feature name. For example, columns with feature names such as "fuel type_coal" and "calorific value_MJ / kg" are selected. For the regional dimension, feature channels related to regional energy structure and policy intensity are selected by matching the "regional energy" and "policy intensity" fields in the feature name using regular expressions. For example, "regional energy_coal-to-power ratio" and "policy intensity_carbon tax rate" are selected. For the process technology dimension, feature channels related to production process parameters and technology iteration data are selected by extracting them based on the identifiers "reaction temperature," "raw material ratio," and "technology iteration" in the feature name. For example, "reaction temperature_°C" and "technology iteration_energy efficiency improvement percentage" are selected. The tree depth of each sub-model is set to 3 to 5 layers, and the learning rate is set to 0.05 to 0.1. The tree depth range is determined by cross-validation experiments. The experiment shows that when the tree depth exceeds 5 layers, the validation set error increases by 15%, indicating that the risk of overfitting increases significantly; the learning rate range is set according to the convergence of gradient boosting tree training. When the learning rate is lower than 0.05, the model convergence speed decreases by 30%, and when it is higher than 0.1, oscillation occurs.

[0085] Dynamic weights are calculated based on the historical prediction errors of each gradient boosting tree submodel. The prediction results of the gradient boosting tree submodels are weighted and summed according to the dynamic weights. The historical prediction errors are calculated using the root mean squared error (RMSE) of the submodels on the validation set. The validation set data is partitioned chronologically, with the most recent 30% of the time window used as the validation set to ensure that the evaluation results reflect the most up-to-date model state. The dynamic weight is calculated as the inverse of the submodel error divided by the sum of the inverses of all submodel errors. For example, if the fuel attribute submodel error is 10, the regional feature submodel error is 15, and the process technology submodel error is 20, then the fuel attribute weight is (1 / 10) / (1 / 10 + 1 / 15 + 1 / 20) = 0.5, the regional feature weight is 0.33, and the process technology weight is 0.17. If a submodel has a historical error of 0, its weight is forced to 1, and the weights of all other submodels are 0 to avoid division by zero errors. If all submodel errors are 0, the weights are evenly distributed. The weighted summation operation is to multiply the predicted value of each sub-model by its dynamic weight and then add them up to generate a preliminary fusion result. For example, the predicted value of the fuel attribute sub-model is 100, the predicted value of the regional characteristic sub-model is 110, and the predicted value of the process technology sub-model is 90. The fusion result is 100×0.5+110×0.33+90×0.17=101.6.

[0086] The weighted summation result is normalized to output the multi-scenario fusion carbon emission factor. The normalization process uses the maximum and minimum value normalization method to linearly map the preliminary fusion result to the range of 0 to 1. The normalization formula is that the fusion factor is equal to the preliminary fusion value minus the minimum value divided by the difference between the maximum and minimum values. The maximum and minimum values ​​are determined by historical data statistics, and the statistical time window is the past 12 months. For example, if the historical minimum fusion value is 50 and the maximum fusion value is 150, the current fusion value of 101.6 is normalized to (101.6-50) / (150-50)=0.516. If the maximum and minimum values ​​are equal, the normalization step is skipped and the original fusion value is output directly to avoid the error of the denominator being zero. The normalized fusion factor is stored in a standardized format with four decimal places of precision for subsequent uncertainty calibration steps.

[0087] The tree depth range (3-5 layers) was determined through cross-validation, and experiments have shown that exceeding this range leads to overfitting or underfitting. The learning rate range (0.05-0.1) was set based on training convergence, and its effectiveness is supported by experimental data. The validation set time window (30%) was determined through timeliness analysis to ensure that model evaluation reflects the latest data distribution.

[0088] S5. Update the initial prediction model based on the incremental learning mechanism and generate historical verification data. Combined with transfer learning, transfer historical knowledge to a dedicated sub-model and filter input data noise. The specific implementation steps are as follows:

[0089] The newly added monitoring data is fed into the initial prediction model, and the model parameters are incrementally updated using an online sequential extreme learning machine to generate an updated initial prediction model and historical validation data. The newly added monitoring data is derived from multi-dimensional monitoring indicators collected in real time. Its format is consistent with the training data, including characteristic channel data from fuel attribute dimensions, regional characteristics dimensions, and process technology dimensions. For example, the fuel attribute dimension includes fuel type, calorific value, and carbon content indicators, while the regional characteristics dimension includes regional energy structure and policy intensity indicators.

[0090] The incremental learning process of the online sequential extreme learning machine involves inputting new data into the model in chronological order, calculating the model output error, and updating the model weights via error backpropagation. The update step size is set to 0.001 to 0.01. The step size range was determined through experimental verification. The experimental method was to test the effect of different step sizes on the model convergence speed and stability on historical datasets. The results showed that when the step size was less than 0.001, the number of iterations required for model convergence increased by 50%. When the step size was greater than 0.01, the fluctuation of the validation set error exceeded 20%. Therefore, 0.001 to 0.01 was selected as a reasonable range. The historical validation data was stored as a two-dimensional table containing timestamps, input features, and model prediction results. It was used in the subsequent uncertainty calibration step. The storage format was Parquet files with the default compression rate.

[0091] The weight distribution features of the updated initial prediction model are extracted and adapted to the weight space of the gradient boosting sub-tree model using a feature mapping function. The weight distribution features are obtained by extracting the weight matrix of the fully connected layer of the initial prediction model. The rows of the matrix represent the input feature dimensions, and the columns represent the output nodes. For example, if the input feature dimension is 50 and the output node is 1, the weight matrix dimension is 50×1. The feature mapping function uses an attention mechanism. This mechanism is implemented by calculating the similarity between the weight of each tree node in the gradient boosting sub-tree model and the weight of the initial model. The similarity is calculated using the cosine similarity formula, which is the dot product of two vectors divided by the product of the vector modulus, and ranges from -1 to 1. The weights of tree nodes with a similarity greater than 0.8 are replaced with the corresponding weights of the initial model, while the weights of nodes with a similarity less than 0.8 are retained. For example, if the weight of the fuel attribute dimension in the initial model is 0.9, if the similarity of the corresponding feature of a tree node is 0.85, the weight of that node is updated to 0.9; if the similarity is 0.6, the original weight of 0.5 is maintained. The similarity threshold of 0.8 was determined through a transfer validity experiment. The experimental method was to compare the prediction accuracy of the sub-models under different thresholds. The results showed that when the threshold was 0.8, the root mean square error of the model was reduced by 8% compared with the threshold of 0.7.

[0092] The isolation forest algorithm detects outliers in newly added monitoring data, removes them, and then feeds them into the gradient boosting sub-model. The isolation forest algorithm partitions the data space by constructing multiple isolation trees. The number of isolation trees is set to 100, with a maximum depth of 8 layers per tree. The parameters are set according to the default configuration of the scikit-learn library. During anomaly detection, the path length of each data point is calculated to determine the degree of anomaly. The path length threshold is set to a value less than 1.5 times the average path length. This threshold is determined based on historical data distribution statistics. The statistical method calculates the path length distribution based on the past three months of data. For example, if the mean path length of normal historical data is 10 and the standard deviation is 2, the threshold is 10 + 2 × 1.5 = 13. After outliers are detected, they are marked and removed, and the remaining data is fed into the gradient boosting sub-model for prediction. If the proportion of outliers in a single batch of data exceeds 30%, a manual review mechanism is triggered, the automatic prediction process is suspended, and the administrator is notified via email.

[0093] Among them, the step size range (0.001-0.01) of the online sequence extreme learning machine was determined through convergence experiments. The experiments showed that a step size less than 0.001 resulted in a 50% decrease in convergence speed, and a step size greater than 0.01 caused model oscillation; the similarity threshold of 0.8 was determined through migration validity tests. Experiments showed that the model accuracy decreased by 12% below this threshold; the isolation forest path length threshold was set based on three months of historical data statistics to ensure that the threshold dynamically adapts to the data distribution.

[0094] S6. Calibrate the uncertainty interval of the multi-scenario integrated carbon emission factor based on historical verification data and output the dynamic optimization assessment results. The specific implementation steps are as follows:

[0095] Based on Monte Carlo simulation, a probability distribution of multi-scenario integrated carbon emission factors is generated, and a confidence interval with a preset threshold confidence level is calculated. The Monte Carlo simulation execution steps include: randomly sampling from historical verification data to generate a simulated data set. The historical verification data comes from the output of the incremental learning step and is stored as a two-dimensional table containing timestamps, input features, and model prediction results. The data format is Parquet file with the default compression rate. The random sampling method uses sampling with replacement, and the sampling volume is 80% of the total historical data. For example, if the historical data contains 1000 samples, 800 samples are sampled each time. The sampling ratio is determined through experimental verification, which shows that an 80% sampling volume can balance computational efficiency and result stability.

[0096] After the simulation dataset was generated, the distribution statistics of the multi-scenario fused carbon emission factors were iteratively calculated. The number of iterations was set to 10,000. The mean and standard deviation of the fused factors were calculated for each iteration to generate a probability density function. The probability density function was fitted using kernel density estimation, using a Gaussian kernel and the bandwidth parameter calculated using the Silverman rule. The confidence level was preset to a 95% threshold, and the upper and lower bounds of the calculated confidence interval were the 2.5% and 97.5% quantiles of the probability density function. For example, if the simulated dataset showed a fusion factor mean of 100 and a standard deviation of 10, the 95% confidence interval would be between 80 and 120.

[0097] The upper and lower confidence interval thresholds are adjusted based on preset rules to generate a dynamically optimized assessment result. These preset rules include historical data statistical patterns and external constraints. The historical data statistical patterns are determined by analyzing the fluctuation range of the fusion factor over the past year. This is done by calculating the maximum, minimum, and standard deviation of the historical data. For example, if the historical maximum is 130, the minimum is 70, and the standard deviation is 15, then the adjusted upper and lower confidence interval bounds must not exceed the maximum and minimum ranges, and the standard deviation fluctuation is limited to 1.5 times the historical standard deviation. External constraints are implemented by parsing the carbon emission intensity limit clauses in policy documents. This parsing method uses regular expressions to match keywords in the policy text (such as "carbon emission factor upper limit" and "allowable fluctuation range"), extract the numerical constraints, and convert them into thresholds. For example, if the policy requires that the carbon emission factor of a certain industry must not exceed 110, and the "upper limit of 110" is matched, the upper confidence interval threshold is adjusted from 120 to 110. The adjusted dynamic optimization evaluation results are stored in a standardized format, including the upper and lower bounds of the confidence interval, adjustment basis, and timestamp, for visualization and decision support. The storage format is JSON file, and the encoding method is UTF-8.

[0098] Among them, the number of Monte Carlo simulation iterations of 10,000 times was determined through convergence testing. The test method was to run 5,000, 10,000, and 20,000 iterations respectively, and calculate the fluctuation range of the confidence interval. The results showed that the fluctuation range was stable within 2% after 10,000 times; the confidence level of 95% was set according to general statistical standards and reliability requirements of industrial scenarios. Experiments showed that the 95% confidence level can cover more than 90% of the actual data distribution; the historical data fluctuation range (maximum value 130, minimum value 70) was obtained through descriptive statistical analysis, and the analysis window was a rolling 12-month data to ensure dynamic adaptation to data changes.

[0099] This embodiment overcomes the limitations of traditional linear aggregation methods by extracting cross-dimensional principal factors and quantifying nonlinear dependency strengths, addressing the problem of feature contribution distortion in multi-source heterogeneous data fusion. A dynamic feature importance network is constructed based on hierarchical weight parameters. Local anomaly signal enhancement is achieved through gradient masking and feature channel scaling, replacing fixed weights or manual experience settings, improving the model's sensitivity to microscopic dynamic features. Dedicated sub-models are constructed for multi-dimensional scenarios, and a dynamic weight allocation mechanism is designed to overcome the lack of adaptability of single models to cross-scenario data distribution differences. Through collaborative optimization of incremental learning and transfer learning, the model parameters achieve dual generalization capabilities in both time and scenario dimensions. Uncertainty interval calibration combines Monte Carlo simulation with dynamic rule feedback to dynamically integrate statistical inference results with external constraints, mitigating the risk of static confidence intervals becoming invalid due to policy changes or data distribution shifts. Data transfer and logical closure between each step form a complete technical chain. For example, a subset of key features drives the generation of cross-dimensional principal factors, hierarchical weight parameters guide the optimization of the feature importance network, and dynamic evaluation results inversely constrain model updates, forming an enhanced loop of "data screening-feature fusion-model iteration-result calibration." This embodiment systematically solves the problems of sensitivity attenuation, scene bias, and evaluation inaccuracy in multi-scale data fusion by coupling nonlinear statistics, dynamic weight migration, and cross-scenario generalization mechanisms.

[0100] Example 2: Figure 2 A structural schematic diagram of a multidimensional dynamic carbon emission factor modeling system of the present invention is provided. The multidimensional dynamic carbon emission factor modeling system includes:

[0101] Multidimensional acquisition module: collects multidimensional monitoring data to construct a multidimensional dataset, and screens key feature subsets that are significantly correlated with carbon emission factors based on statistical tests;

[0102] Cross-dimensional analysis module: establishes a correlation matrix between key feature subsets and carbon emission factors, extracts cross-dimensional main factors through principal component analysis, quantifies the nonlinear dependency strength of each cross-dimensional main factor, and generates hierarchical weight parameters;

[0103] Weight Optimization Module: Constructs a gradient mask-based feature importance network based on hierarchical weight parameters, optimizes the feature distribution of multidimensional datasets through feature channel scaling, and generates an initial prediction model;

[0104] Dynamic Fusion Module: Builds dedicated sub-models for multiple preset dimensions, uses a dynamic weight allocation mechanism to fuse the output results of the dedicated sub-models, and generates a multi-scenario fused carbon emission factor;

[0105] Incremental migration module: Based on the incremental learning mechanism, it updates the initial prediction model and generates historical verification data. It also combines transfer learning to migrate historical knowledge to a dedicated sub-model and filter input data noise.

[0106] Dynamic calibration module: calibrates the uncertainty interval of the multi-scenario fusion carbon emission factor based on historical verification data, and outputs dynamic optimization evaluation results.

[0107] The multi-dimensional acquisition module provides input for the cross-dimensional analysis module by collecting multi-dimensional monitoring data and screening key features; the cross-dimensional analysis module extracts cross-dimensional main factors and quantifies nonlinear dependencies, and the generated hierarchical weight parameters are input into the weight optimization module; the weight optimization module constructs a feature importance network based on the weight parameters, and the optimized feature distribution drives the dynamic fusion module to train the multi-dimensional sub-model and fuse the output; the multi-scenario carbon emission factors generated by the dynamic fusion module are input into the incremental migration module, which updates the model parameters and migrates historical knowledge through incremental learning, while filtering noise data to ensure input quality; the historical verification data output by the incremental migration module is input into the dynamic calibration module, which combines Monte Carlo simulation with dynamic rules to calibrate the uncertainty interval, and finally generates a dynamic optimization evaluation result.

[0108] Each module forms a collaborative mechanism through data flow and logical loops: the cross-dimensional analysis module addresses the distortion of feature contributions caused by linear aggregation of cross-scale data in traditional methods; the weight optimization module enhances the ability to identify local abnormal signals through dynamic feature scaling; the dynamic fusion module uses dynamic weight allocation to address model bias caused by differences in data distribution across multiple scenarios; the incremental migration module enables continuous model adaptation to dynamic data changes; and the dynamic calibration module enhances the reliability and compliance of assessment results through the fusion of statistics and rules. Through modular design and data transfer mechanisms, the overall solution systematically addresses the issues of insufficient sensitivity, poor scenario generalization, and inaccurate assessments in carbon emission factor modeling.

[0109] The calculations involved in the embodiments are all dimensionless numerical calculations, and the preset parameters and thresholds in the calculations are set by those skilled in the art according to actual conditions.

[0110] It should be noted that the present invention can be deployed on the device itself to implement embedded applications, and can also be run on a PC or other terminal with a user interface, thereby meeting various hardware environments and usage requirements.

[0111] The above embodiments can be implemented in whole or in part via software, hardware, firmware, or any other combination. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product comprises one or more computer instructions or computer programs. When loaded or executed on a computer, the processes or functions described in the embodiments of this application are fully or partially generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wireless or wired transmission (e.g., infrared, wireless, microwave, etc.); wired transmission methods include optical fiber, twisted pair, coaxial cable, etc.; wireless transmission methods include infrared, microwave, etc. The computer-readable storage medium can be any available medium accessible by a computer, or a data storage device such as a server or data center that contains a collection of one or more available media. The available medium can be magnetic media (e.g., floppy disk, hard disk, tape), optical media (e.g., DVD), or semiconductor media. The semiconductor medium may be a solid state drive.

[0112] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and modules described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0113] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the modules is only a logical function division. In actual implementation, there may be other division methods, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.

[0114] The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules, and may be located in one place or distributed across multiple network modules. Some or all of the modules may be selected to achieve the purpose of this embodiment according to actual needs.

[0115] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0116] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0117] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

[0118] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A multidimensional dynamic carbon emission factor modeling method, characterized in that: include: S1. Collect multi-dimensional monitoring data to construct a multi-dimensional dataset, and screen the key feature subsets that are significantly correlated with carbon emission factors based on statistical tests; S2. Establish a correlation matrix between key feature subsets and carbon emission factors, extract cross-dimensional main factors through principal component analysis, quantify the nonlinear dependence intensity of each cross-dimensional main factor, and generate hierarchical weight parameters; S3. Construct a gradient mask-based feature importance network based on hierarchical weight parameters, optimize the feature distribution of the multidimensional dataset through feature channel scaling, and generate an initial prediction model, including: The hierarchical weight parameters are used as the initial feature importance weights, and the weight gradients of each feature channel are dynamically updated through gradient backpropagation to generate dynamic feature importance masks; Multiply the dynamic feature importance mask with the multidimensional dataset channel by channel to suppress the noise of low-importance features and enhance the key feature signals to generate an optimized multidimensional feature distribution; Based on the optimized multi-dimensional feature distribution, the gradient boosting tree model architecture is selected, and the model hyperparameters are adjusted using cross-validation and Bayesian optimization to generate an initial prediction model. S4. Build dedicated sub-models for multiple preset dimensions, and use a dynamic weight allocation mechanism to fuse the output results of the dedicated sub-models to generate a multi-scenario fused carbon emission factor; S5. Update the initial prediction model based on the incremental learning mechanism and generate historical verification data. Combined with transfer learning, transfer historical knowledge to the dedicated sub-model and filter input data noise. S6. Calibrate the uncertainty interval of the multi-scenario fusion carbon emission factor based on historical verification data and output the dynamic optimization evaluation results.

2. A multidimensional dynamic carbon emission factor modeling method according to claim 1, characterized in that: Multi-dimensional monitoring data is collected to construct a multi-dimensional dataset, and a subset of key features that are significantly correlated with carbon emission factors is screened based on statistical tests, including: Collect monitoring indicators in fuel attribute dimensions, regional characteristics dimensions, time dynamic dimensions, process technology dimensions, equipment efficiency dimensions, and environmental parameter dimensions to construct a multidimensional data set; After standardization preprocessing of the multidimensional data set, the chi-square test or mutual information method is used to calculate the correlation between the characteristics of each dimension and the carbon emission factor, and the features with correlation higher than the preset significance threshold are retained to form a key feature subset.

3. A multidimensional dynamic carbon emission factor modeling method according to claim 2, characterized in that: The fuel attribute dimension includes fuel type, calorific value and carbon content indicators; the regional characteristic dimension includes regional energy structure and policy intensity indicators; and the time dynamic dimension includes seasonal energy consumption fluctuations and equipment operation cycle indicators.

4. A multidimensional dynamic carbon emission factor modeling method according to claim 1, characterized in that: Establish a correlation matrix between key feature subsets and carbon emission factors, extract cross-dimensional main factors through principal component analysis, quantify the nonlinear dependence intensity of each cross-dimensional main factor, and generate hierarchical weight parameters, including: Calculate the Spearman rank correlation coefficient between each dimension feature and the carbon emission factor in the key feature subset, and select features whose absolute value of the Spearman rank correlation coefficient is greater than the preset correlation threshold to generate an initial correlation matrix; Perform eigenvalue decomposition on the initial correlation matrix and extract the principal components whose cumulative contribution rate exceeds the preset contribution threshold as the cross-dimensional principal factors; The kernel density estimation method is used to fit the joint probability distribution of the cross-dimensional main factors and carbon emission factors, and the nonlinear dependence strength between the two is calculated based on the mutual information entropy formula; The weight distribution ratio is calculated based on the entropy weight method combined with the nonlinear dependency intensity of each cross-dimensional main factor to generate the hierarchical weight parameters.

5. The multidimensional dynamic carbon emission factor modeling method according to claim 1, characterized in that: Dedicated sub-models are constructed for multiple preset dimensions, and the output results of the dedicated sub-models are integrated using a dynamic weight allocation mechanism to generate multi-scenario integrated carbon emission factors, including: Based on the optimized multi-dimensional feature distribution, the gradient boosting tree sub-models corresponding to the fuel attribute dimension, regional feature dimension, and process technology dimension are trained respectively; Calculate the dynamic weight based on the historical prediction error of each gradient boosting tree sub-model, and sum the prediction results of the gradient boosting tree sub-model according to the dynamic weight; The weighted summation results are normalized and the multi-scenario fusion carbon emission factor is output.

6. A multidimensional dynamic carbon emission factor modeling method according to claim 1, characterized in that: Based on the incremental learning mechanism, the initial prediction model is updated and historical verification data is generated. In combination with transfer learning, historical knowledge is transferred to the dedicated sub-model and input data noise is filtered, including: The newly added monitoring data is input into the initial prediction model, and the model parameters are incrementally updated through the online sequence extreme learning machine to generate an updated initial prediction model and historical verification data; Extract the weight distribution features of the updated initial prediction model and adapt them to the weight space of the gradient boosting tree sub-model through the feature mapping function; The outliers in the newly added monitoring data are detected based on the isolation forest algorithm, and then input into the gradient boosting tree sub-model after removing the outliers.

7. The multidimensional dynamic carbon emission factor modeling method according to claim 1, characterized in that: The uncertainty interval of the multi-scenario integrated carbon emission factor is calibrated based on historical verification data, and dynamic optimization assessment results are output, including: Generate the probability distribution of carbon emission factors for multiple scenarios based on Monte Carlo simulation, and calculate the confidence interval with a preset threshold confidence level; The upper and lower thresholds of the confidence interval are adjusted according to the preset rules to generate dynamic optimization evaluation results.

8. A multi-dimensional dynamic carbon emission factor modeling method according to claim 7, characterized in that: The execution steps of the Monte Carlo simulation include: randomly sampling from historical verification data to generate a simulation data set, iteratively calculating the distribution statistics of the multi-scenario fusion carbon emission factors; the preset rules include the statistical laws of historical data and external constraints.

9. A multidimensional dynamic carbon emission factor modeling system, used to implement a multidimensional dynamic carbon emission factor modeling method according to any one of claims 1 to 8, characterized in that: include: Multidimensional acquisition module: collects multidimensional monitoring data to construct a multidimensional dataset, and screens key feature subsets that are significantly correlated with carbon emission factors based on statistical tests; Cross-dimensional analysis module: establishes a correlation matrix between key feature subsets and carbon emission factors, extracts cross-dimensional main factors through principal component analysis, quantifies the nonlinear dependency strength of each cross-dimensional main factor, and generates hierarchical weight parameters; Weight Optimization Module: Constructs a gradient mask-based feature importance network based on hierarchical weight parameters, optimizes the feature distribution of multidimensional datasets through feature channel scaling, and generates an initial prediction model; Dynamic Fusion Module: Builds dedicated sub-models for multiple preset dimensions, uses a dynamic weight allocation mechanism to fuse the output results of the dedicated sub-models, and generates a multi-scenario fused carbon emission factor; Incremental migration module: Based on the incremental learning mechanism, it updates the initial prediction model and generates historical verification data. It also combines transfer learning to migrate historical knowledge to a dedicated sub-model and filter input data noise. Dynamic calibration module: calibrates the uncertainty interval of the multi-scenario fusion carbon emission factor based on historical verification data, and outputs dynamic optimization evaluation results.

Images