Carbon emission accounting method based on machine learning algorithm fusion
By introducing a method of fusion of feature engineering and multi-type regression algorithms in the carbon emission accounting method, the problems of variable screening and algorithm limitations in the existing technology are solved, the flexibility and stability of the carbon accounting model are realized, and the efficient accounting ability of carbon emissions is improved.
Patent Information
- Application Number
- CN202411659842.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-20
- Publication Date
- 2025-03-21
AI Technical Summary
Existing carbon emission accounting methods cannot automatically filter effective variables, and a single machine learning algorithm cannot cover multiple application scenarios, resulting in limited application scope and poor stability of the model.
Carbon emission accounting method based on machine learning algorithm fusion is adopted to construct a method of dynamically screening effective variables through feature engineering, and algorithm fusion calculation is carried out in combination with multi-type regression algorithms (such as Lasso, ARDL, LightGBM) to enhance the application scope and stability of the model.
The flexibility and scope of application of the carbon accounting model are achieved, the stability and accuracy of carbon accounting results are ensured, and the efficient accounting ability of carbon emissions is improved.
Smart Images
Figure CN119671023A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of carbon accounting technology, and in particular to a carbon emission accounting method based on the fusion of machine learning algorithms. Background Art
[0002] At present, various units of State Grid Corporation have carried out a large number of pilot practices, inherited and expanded the IPCC carbon accounting system, introduced machine learning, big data analysis and other theories, and gradually formed a carbon emission accounting method based on machine learning algorithm models. It mainly determines the input variables in advance based on data characteristics and expert experience, and performs carbon accounting based on fixed input variables of regions, industries and enterprises. This method has standardized carbon accounting logic, but it also has the following shortcomings:
[0003] (1) Unable to automatically screen effective variables: Traditional model carbon accounting methods use fixed variable data such as historical electricity, GDP and energy activities, and industrial output to construct regression analysis models. However, in actual applications, not all variables can be obtained or are beneficial to the model, resulting in the inability to directly apply the model or poor carbon emission accounting results.
[0004] (2) A single algorithm cannot cover multiple types of application scenarios: The selection of a specific machine learning method has certain limitations. Due to the different production processes and data distribution of different application entities, the degree of deviation between the model carbon accounting results and the actual value varies, affecting the model stability assessment. Summary of the invention
[0005] The present invention provides a carbon emission accounting method based on the fusion of machine learning algorithms. The method introduces a feature engineering construction method in variable screening to achieve dynamic screening of effective variables as the measurement subject changes; in algorithm optimization, multiple types of machine learning methods are added to achieve comprehensive measurement of carbon emissions by various algorithms through algorithm fusion, thereby increasing the application scope of the carbon accounting model and ensuring the stability of the model measurement deviation.
[0006] In order to solve the above technical problems, the present invention provides a carbon emission accounting method based on machine learning algorithm fusion, including:
[0007] S1. Build model-based feature engineering;
[0008] S2. Build a model based on multi-type regression algorithm;
[0009] S3. Algorithm fusion to calculate carbon emissions.
[0010] Preferably, S1 comprises the following steps:
[0011] S1.1 Multicollinearity analysis among independent variables based on Lasso model;
[0012] The Lasso regression model expression is:
[0013] y=Xβ (1)
[0014] The objective function that Lasso needs to solve is:
[0015] arg min||y-Xβ|| 2 +λ||β||1 (2)
[0016] In the formula, λ represents the degree of variable compression. The larger its value, the more redundant variables are identified and the more variables are eliminated. This method uses cross-validation to select the optimal λ value.
[0017] S1.2 Analyze carbon emission influencing factors based on Pearson correlation;
[0018] Calculate the correlation coefficient ρ between variable X and carbon emissions Y X,Y , the calculation formula is as follows:
[0019]
[0020] Where: μ X , μ Y , σ X , σ Y are the mean and standard deviation of X and Y respectively;
[0021] Pearson correlation coefficient ρ X,Y The value range of: -1≤ρ X,Y ≤1, indicating different degrees of linear correlation; |ρ X,Y ∣≥0.7 indicates strong correlation, 0.3≤∣ρ X,Y ∣≤0.7 indicates moderate correlation, ∣ρ X,Y |≤0.3 indicates weak correlation; the threshold used in this method is 0.3;
[0022] S1.3 Feature importance assessment based on Gini impurity;
[0023] Assuming that there are N categories in the data set, the probability that a data sample is exactly the i-th category is p(i), then the Gini impurity calculation formula of variable V is:
[0024]
[0025] In the scenario of conditional variable M, assuming that M has C classifications or discretization intervals, the conditional Gini impurity calculation formula of V is:
[0026]
[0027] From this, the Gini gain value, which measures the importance of the feature, can be calculated:
[0028] Gini 增益 (V,M)=Gini(V)-Gini(V|M) (6)
[0029] Then, we screen out the independent variables whose importance is greater than 0.1.
[0030] Preferably, S2 comprises the following steps:
[0031] S2.1LR algorithm to build carbon accounting model:
[0032] y t,e =α0+βx t,e +u t,e (7)
[0033] Where: β is the variable regression coefficient;
[0034] y t,e : The dependent variable of the model, representing the carbon emissions of the carbon accounting subject e in the current period;
[0035] α0: represents the basic influencing factor on y t,e Contribution
[0036] βx t,e : The independent variable of the model, indicating the influence factor of carbon accounting subject e on y t,e Contribution
[0037] u t,e : represents the random deviation of carbon emissions calculated by carbon accounting subject e;
[0038] S2.2ARDL algorithm to build carbon accounting model:
[0039]
[0040] In the formula, β, γ are variable regression coefficients, p is the lag order of the dependent variable, q is the lag order of the explanatory variable, s is the number of supplementary variables, and the explanatory variables and supplementary variables constitute the independent variables of the entire model;
[0041] y t,e : The dependent variable of the model, representing the carbon emissions of the carbon accounting subject e in the current period;
[0042] α0: represents the basic influencing factor on y t,e Contribution
[0043] represents the carbon emissions of carbon accounting subject e in historical period p to y t,e Contribution
[0044] represents the carbon accounting subject e historical period q and the current explanatory variable y t,e Contribution
[0045] Represents the carbon accounting subject e's s supplementary variables to y t,e Contribution
[0046] u t,e : represents the random deviation of carbon emissions calculated by carbon accounting subject e;
[0047] S2.3LightGBM algorithm builds carbon accounting model:
[0048]
[0049] In the formula, M is the number of base learner decision trees, β m is the weight of the prediction result of the mth decision tree, α m is the parameter of the mth decision tree;
[0050] y t,e : The dependent variable of the model, representing the carbon emissions of the carbon accounting subject e in the current period;
[0051] x t,e : The independent variable of the model, which represents all the independent variable input data of the carbon accounting subject e in the current period;
[0052] h(x t,e ; α m ): indicates the prediction result of the mth decision tree; when m=1, h(x t,e ; α1) is the predicted value of the current carbon emissions. When m≥2, h(x t,e ; α m ) is the residual prediction value of the current carbon emissions, and the sum of the prediction results of M decision trees is the current carbon emissions accounting result.
[0053] Preferably, in S3, the carbon accounting algorithms are integrated by weighted average method, and the weights are calculated according to the prediction deviation rate of each algorithm:
[0054]
[0055] In the formula, A i is the actual value of carbon emissions of the i-th sample data, F i is the predicted value of carbon emissions of the i-th sample data, and N is the number of data samples;
[0056]
[0057]
[0058]
[0059]
[0060] In the formula, MAPE LR ,MAPE ARDL ,MAPE LightGBM are the carbon emission accounting deviation rates of the three algorithms, from which the weight value of the carbon accounting results of each algorithm is calculated LR ,ω ARDL ,ω LightGBM , thereby calculating the carbon emissions result C calculated by the comprehensive model E :
[0061] C E =ω LR ×C LR +ω ARDL ×C ARDL +ω LightGBM ×C LightGBM (15)
[0062] In the formula, C LR , C ARDL , C LightGBM The carbon emissions are calculated by three algorithms respectively.
[0063] Compared with the related art, the present invention has the following beneficial effects:
[0064] The present invention provides a carbon emission accounting method based on the fusion of machine learning algorithms
[0065] 1. This method takes into account the ease and effectiveness of obtaining relevant data for carbon accounting application entities, constructs a feature engineering process, gets rid of the shortcomings of inputting fixed feature variables, and makes the application of carbon accounting model methods more flexible.
[0066] 2. It can automatically screen important carbon emission influencing factors for different carbon accounting application entities, and is compatible with multi-scenario data of regions, industries, and enterprises. The carbon accounting model method has a wider scope of application.
[0067] 3. We have built multiple types of machine learning algorithm models, including traditional regression, time series regression, and tree regression, to comprehensively capture the linear, time series, and nonlinear characteristics of carbon accounting data, and through algorithm fusion methods, make the carbon accounting results more accurate. BRIEF DESCRIPTION OF THE DRAWINGS
[0068] Figure 1 Constructing a flow chart for the feature engineering of the present invention;
[0069] Figure 2 A process diagram for constructing a carbon accounting model based on the fusion of machine learning algorithms of the present invention;
[0070] Figure 3 This is a flow chart of the carbon emission accounting model application method of the present invention. DETAILED DESCRIPTION
[0071] The present invention will be further described below in conjunction with the accompanying drawings and implementation modes.
[0072] The present invention divides the carbon accounting method based on machine learning algorithm into three steps: model-based feature engineering construction, model construction based on multi-type regression algorithm, and algorithm fusion calculation of carbon emissions.
[0073] (1) Model-based feature engineering construction
[0074] Due to the differences between different carbon accounting entities, in order to accurately capture the important influencing factors of carbon emissions of each accounting entity, algorithm analysis is carried out from the aspects of multicollinearity of independent variables, variable correlation, and feature importance, thereby completing the construction of feature engineering.
[0075] 1) Multicollinearity analysis between independent variables based on Lasso model
[0076] In a regression analysis model, if there is serious multicollinearity between independent variables, they will affect each other and it will be impossible to obtain the true correlation between each independent variable and the dependent variable. Lasso has a strong advantage in screening variables with high explanatory power and solving multicollinearity problems. This model adds a penalty term to the model estimation, which can compress the regression coefficients of some redundant variables to zero and then remove them from the model to achieve the purpose of variable screening.
[0077] The Lasso regression model expression is:
[0078] y=Xβ (1)
[0079] The objective function that Lasso needs to solve is:
[0080] arg min||y-Xβ|| 2 +λ||β||1 (2)
[0081] In the formula, λ represents the degree of variable compression. The larger its value, the more redundant variables are identified and the more variables are eliminated. This method uses cross-validation to select the optimal λ value.
[0082] 2) Analysis of carbon emission influencing factors based on Pearson correlation
[0083] Pearson correlation can measure the linear correlation between characteristic variables and carbon emissions. To calculate the correlation coefficient ρ between variable X and carbon emissions Y X,Y , the calculation formula is as follows:
[0084]
[0085] Where: μ X , μ Y , σ X , σ Y are the mean and standard deviation of X and Y respectively.
[0086] Pearson correlation coefficient ρ X,Y The value range of: -1≤ρ X,Y ≤1, indicating different degrees of linear correlation. X,Y ∣≥0.7 indicates strong correlation, 0.3≤∣ρ X,Y ∣≤0.7 indicates moderate correlation, ∣ρ X,Y |≤0.3 indicates weak correlation. The threshold used in this method is 0.3.
[0087] 3) Feature importance evaluation based on Gini impurity
[0088] Gini impurity can measure the amount of information contained in the independent variable relative to the dependent variable. The smaller its value, the more feature information the independent variable has; the larger the Gini gain, the greater the impact of the feature value change on the dependent variable, that is, the more important the feature is. Assuming that there are N categories in the data set (if it is a numerical variable, it is discretized to generate N interval categories), the probability that a data sample is exactly the i-th category is p(i), then the Gini impurity calculation formula of variable V is:
[0089]
[0090] In the scenario of conditional variable M, assuming that M has C categories (or discretization intervals), the conditional Gini impurity calculation formula of V is:
[0091]
[0092] From this, the Gini gain value, which measures the importance of the feature, can be calculated:
[0093] Gini 增益 (V,M)=Gini(V)-Gini(V|M) (6)
[0094] This method uses the random forest algorithm with Gini impurity to evaluate the importance of independent variable features and screen out independent variables with an importance greater than 0.1.
[0095] (2) Model construction based on multi-type regression algorithm
[0096] According to the analysis of the application scenarios of carbon accounting methods, it can be known that the construction of carbon accounting models can adopt both time series prediction algorithms and regression prediction algorithms. Combined with the characteristics of the algorithm itself, this method uses three algorithms: traditional regression-linear regression model (LR), time series regression-autoregressive distributed lag model (ARDL), and tree regression-gradient boosting decision tree model (LightGBM) to meet the needs of carbon accounting scenarios.
[0097] 1) LR algorithm to build carbon accounting model
[0098] LR predicts future trends based on the linear relationship of existing data. The LR carbon accounting model assumes that there is a strong linear relationship between the carbon emissions of the measuring subject and related variables such as output, output value, and energy. The algorithm construction expression is as follows:
[0099] y t,e =α0+βx t,e +u t,e (7)
[0100] Where: β is the variable regression coefficient;
[0101] y t,e : The dependent variable of the model, representing the carbon emissions of the carbon accounting subject e in the current period;
[0102] α0: represents the basic influencing factor on y t,e Contribution
[0103] βx t,e : The independent variable of the model, indicating the influence factor of carbon accounting subject e on y t,e Contribution
[0104] u t,e : Represents the random deviation of carbon emissions calculated by carbon accounting entity e.
[0105] 2) ARDL algorithm to build carbon accounting model
[0106] ARDL is used to establish long-term relationships between variables. It can comprehensively consider long-term and short-term effects to better describe the relationship between variables. Its core idea is that the future value of the dependent variable in the time series depends on its own lagged value and the independent variable and its lagged value. The algorithm construction expression is as follows:
[0107]
[0108] In the formula, β, γ are variable regression coefficients, p is the lag order of the dependent variable, q is the lag order of the explanatory variable, s is the number of supplementary variables, and the explanatory variables and supplementary variables constitute the independent variables of the entire model;
[0109] yt,e : The dependent variable of the model, representing the carbon emissions of the carbon accounting subject e in the current period;
[0110] α0: represents the basic influencing factor on y t,e Contribution
[0111] represents the carbon emissions of carbon accounting subject e in historical period p to y t,e Contribution
[0112] represents the carbon accounting subject e historical period q and the current explanatory variable y t,e Contribution
[0113] Represents the carbon accounting subject e's s supplementary variables to y t,e Contribution
[0114] u t,e : Represents the random deviation of carbon emissions calculated by carbon accounting entity e.
[0115] 3) LightGBM algorithm builds carbon accounting model
[0116] LightGBM is a nonlinear regression algorithm based on the gradient boosting decision tree algorithm, which can capture both the linear and nonlinear relationships between independent variables and carbon emissions. The algorithm construction expression is as follows:
[0117]
[0118] In the formula, M is the number of base learner decision trees, β m is the weight of the prediction result of the mth decision tree, α m is the parameter of the mth decision tree;
[0119] y t,e : The dependent variable of the model, representing the carbon emissions of the carbon accounting subject e in the current period;
[0120] x t,e : The independent variable of the model, which represents all the independent variable input data of the carbon accounting subject e in the current period;
[0121] h(x t,e ; α m ): indicates the prediction result of the mth decision tree; when m=1, h(x t,e ; α1) is the predicted value of the current carbon emissions. When m≥2, h(x t,e ; α m ) is the residual prediction value of the current carbon emissions, and the sum of the prediction results of M decision trees is the current carbon emissions accounting result.
[0122] The historical carbon emissions of the carbon accounting subject e and its important influencing factors are respectively input into the LR model, ARDL model and LightGBM model for training. The regression coefficients of each model are learned and substituted into the algorithm expression for carbon accounting calculations.
[0123] (3) Algorithm fusion to calculate carbon emissions
[0124] Machine learning algorithm fusion is a method to improve model prediction performance by combining the prediction results of multiple models in the hope of achieving better overall performance than a single model. This method uses the weighted average method to achieve the fusion of the above three carbon accounting algorithms, and its weight is calculated based on the prediction deviation rate (MAPE, mean absolute percentage error) of each algorithm:
[0125]
[0126] In the formula, A i is the actual value of carbon emissions of the i-th sample data, F i is the predicted value of carbon emissions of the i-th sample data, and N is the number of data samples.
[0127]
[0128]
[0129]
[0130]
[0131] In the formula, MAPE LR ,MAPE ARDL ,MAPE LightGBM are the carbon emission accounting deviation rates of the three algorithms, from which the weight value of the carbon accounting results of each algorithm is calculated LR ,ω ARDL ,ω LightGBM , so that the carbon emission result C calculated by the comprehensive model can be calculated E :
[0132] C E =ω LR ×C LR +ω ARDL ×C ARDL +ω LightGBM ×C LightGBM (15)
[0133] In the formula, C LR , C ARDL , C LightGBM The carbon emissions are calculated by three algorithms respectively.
[0134] In order to more clearly illustrate the specific implementation mode of the present invention, an embodiment is provided below:
[0135] The present invention provides a carbon emission accounting method based on the fusion of machine learning algorithms. In the process of building a carbon accounting model, a feature engineering construction method is introduced to realize the dynamic screening of characteristic variables of the carbon accounting subject, so as to improve the flexibility of the application of the carbon accounting method; a method of fusing multiple types of machine learning algorithms is adopted to carry out comprehensive calculation of carbon emissions, thereby ensuring the stability of the carbon accounting model, and realizing efficient calculation of carbon emissions while improving the accuracy of carbon accounting.
[0136] like Figure 1 As shown, the present invention constructs feature engineering for the input carbon accounting main data, including preprocessing processes such as variable expansion, anomaly detection, and missing value filling, as well as analysis processes such as variable collinearity, variable correlation, and feature importance. The variables thus screened out are used as input variables of the final carbon accounting model, ensuring the validity of the variables.
[0137] like Figure 2 As shown, the carbon accounting method of the present invention focuses on electricity and energy consumption, and the correlation between energy consumption and carbon emission data. First, an "electricity-based energy" model is constructed based on a variety of machine learning algorithms, and then energy-to-carbon conversion is performed based on the energy consumption values calculated by the algorithm and the carbon emission factors disclosed by the corresponding regions / industries, thereby realizing carbon emission accounting.
[0138] like Figure 3 As shown, the present invention formulates an application method of the carbon accounting model based on the known characteristics of the electricity-carbon data of the carbon accounting subject such as the region / industry / key enterprise, or the carbon accounting requirements. One method is to perform carbon emission accounting based on the self-training model, and the other method is to perform carbon emission accounting based on the pre-trained model of the same type of carbon accounting subject. If the given sample data volume is sufficient to meet the model training requirements, then the carbon accounting subject's own data can be directly used to perform autonomous model training and complete carbon emission accounting; if the training requirements cannot be met or no training is required, the pre-trained model of the same type that has been trained in advance is directly applied to perform carbon emission accounting to solve the problem of insufficient data of the carbon accounting subject.
[0139] The above are only embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.
Claims
1. A carbon emission accounting method based on machine learning algorithm fusion, characterized in that: include: S1. Build model-based feature engineering; S2. Build a model based on multi-type regression algorithm; S3. Algorithm fusion to calculate carbon emissions.
2. The carbon emission accounting method based on machine learning algorithm fusion according to claim 1 is characterized in that: The S1 comprises the following steps: S1.1 Multicollinearity analysis among independent variables based on Lasso model; The Lasso regression model expression is: y=Xβ (1) The objective function that Lasso needs to solve is: arg min||y-Xβ|| 2 +λ||β||1 (2) In the formula, λ represents the degree of variable compression. The larger its value, the more redundant variables are identified and the more variables are eliminated. This method uses cross-validation to select the optimal λ value. S1.2 Analyze carbon emission influencing factors based on Pearson correlation; Calculate the correlation coefficient ρ between variable X and carbon emissions Y X,Y , the calculation formula is as follows: Where: μ X , μ Y , σ X , σ Y are the mean and standard deviation of X and Y respectively; Pearson correlation coefficient ρ X,Y The value range of: -1≤ρ X,Y ≤1, indicating different degrees of linear correlation; |ρ X,Y ∣≥0.7 indicates strong correlation, 0.3≤∣ρ X,Y ∣≤0.7 indicates moderate correlation, ∣ρ X,Y |≤0.3 indicates weak correlation; the threshold used in this method is 0.3; S1.3 Feature importance assessment based on Gini impurity; Assuming that there are N categories in the data set, the probability that a data sample is exactly the i-th category is p(i), then the Gini impurity calculation formula of variable V is: In the scenario of conditional variable M, assuming that M has C classifications or discretization intervals, the conditional Gini impurity calculation formula of V is: From this, the Gini gain value, which measures the importance of the feature, can be calculated: Gini 增益 (V,M)=Gini(V)-Gini(V|M) (6) Then, we screen out the independent variables whose importance is greater than 0.
1.
3. The carbon emission accounting method based on machine learning algorithm fusion according to claim 1 is characterized in that: The S2 comprises the following steps: S2.1LR algorithm to build carbon accounting model: y t,e =α0+βx t,e +u t,e (7) Where: β is the variable regression coefficient; y t,e : The dependent variable of the model, representing the carbon emissions of the carbon accounting subject e in the current period; α0: represents the basic influencing factor on y t,e Contribution βx t,e : The independent variable of the model, indicating the influence factor of carbon accounting subject e on y t,e Contribution u t,e : represents the random deviation of carbon emissions calculated by carbon accounting subject e; S2.2ARDL algorithm to build carbon accounting model: In the formula, β, γ are variable regression coefficients, p is the lag order of the dependent variable, q is the lag order of the explanatory variable, s is the number of supplementary variables, and the explanatory variables and supplementary variables constitute the independent variables of the entire model; y t,e : The dependent variable of the model, representing the carbon emissions of the carbon accounting subject e in the current period; α0: represents the basic influencing factor on y t,e Contribution represents the carbon emissions of carbon accounting subject e in historical period p to y t,e Contribution represents the carbon accounting subject e historical period q and the current explanatory variable y t,e Contribution Represents the carbon accounting subject e's s supplementary variables to y t,e Contribution u t,e : represents the random deviation of carbon emissions calculated by carbon accounting subject e; S2.3LightGBM algorithm builds carbon accounting model: In the formula, M is the number of base learner decision trees, β m is the weight of the prediction result of the mth decision tree, α m is the parameter of the mth decision tree; y t,e : The dependent variable of the model, representing the carbon emissions of the carbon accounting subject e in the current period; x t,e : The independent variable of the model, which represents all the independent variable input data of the carbon accounting subject e in the current period; h(x t,e ; α m ): indicates the prediction result of the mth decision tree; when m=1, h(x t,e ; α1) is the predicted value of the current carbon emissions. When m≥2, h(x t,e ; α m ) is the residual prediction value of the current carbon emissions, and the sum of the prediction results of M decision trees is the current carbon emissions accounting result.
4. The carbon emission accounting method based on machine learning algorithm fusion according to claim 1 is characterized in that: In S3, the carbon accounting algorithms are integrated by weighted average method, and the weights are calculated according to the prediction deviation rate of each algorithm: In the formula, A i is the actual value of carbon emissions of the i-th sample data, F i is the predicted value of carbon emissions of the i-th sample data, and N is the number of data samples; In the formula, MAPE LR ,MAPE AEDL ,MAPE LightGBM are the carbon emission accounting deviation rates of the three algorithms, from which the weight value of the carbon accounting results of each algorithm is calculated LR ,ω ARDL ,ω LightGBM , thereby calculating the carbon emissions result C calculated by the comprehensive model E : C E =ω LR ×C LR +oh ARDL ×C ARDL +oh LightGBM ×C LightGBM (15) In the formula, C LR , C ARDL , C LightGBM The carbon emissions are calculated by three algorithms respectively.