A method for identifying key influencing factors of carbon intensity based on random forest

Abstract

The present invention relates to the field of carbon emission reduction strategy calculation, and specifically to a method for identifying key influencing factors of carbon intensity based on random forests. Compared with the traditional carbon intensity influencing factor analysis method, the present invention has the characteristics of flexible data selection and no dimensionality restriction. It can more accurately identify the more important influencing factors among many potential influencing factors, make up for the defects of data dimension restriction of traditional analysis methods, provide scientific reference for the accurate implementation of carbon emission reduction plans, and also provide relative focus for regional carbon emission reduction policy formulation.

Description

Technical Field

[0001] The present invention relates to the field of carbon emission reduction strategy calculation, and in particular to a method for identifying key influencing factors of carbon intensity based on random forests. Background Art

[0002] The issue of climate change has received increasing attention worldwide and is seen as a challenge that the world needs to address together in the 21st century. There is a clear connection between human emission behavior and the impact on the natural climate system. Among greenhouse gases, excessive carbon dioxide emissions are the main cause of climate change. By controlling anthropogenic carbon dioxide emissions, there is hope to curb the trend of rising temperatures. Carbon intensity is usually used to measure the relationship between the national economic development of a country (region) and carbon emissions. It plays an important role in judging the relative emission reduction process in developing countries, that is, it represents the relative efficiency of the development process. Carbon intensity is the result of the combined effect of comprehensive factors. There are differences in resource endowments, pillar industries, and technological levels in various regions, and there are differences in key influencing factors of carbon intensity. Therefore, it is very necessary to identify the main factors affecting regional carbon intensity and to complete carbon intensity control work in a targeted manner.

[0003] At present, the research on the identification of carbon intensity influencing factors is mostly directed to a specific industry or type, such as the logarithmic mean Dirichlet decomposition method (LMDI), the scalable stochastic environmental impact assessment model (STIRPAT), the Kaya identity decomposition, the Laspeyres index decomposition model and other technical means. However, due to the data dimension limitations of these traditional algorithms, the perspective is often limited to a certain sub-field (such as energy structure, energy intensity, etc.), and lacks the control of sub-indicators within the scope of the overall social development perspective. Machine learning methods have no restrictions on data dimensions, that is, more types of parameters can be added to the identification of key influencing factors of carbon intensity, turning the perspective to the full range of the overall social development process. Summary of the invention

[0004] To solve the above problems, the present invention proposes a random forest-based method for identifying key influencing factors of carbon intensity. By collecting a data set of carbon intensity and potential influencing factors within a certain time range in a region and inputting them into the model, the identification and analysis of key influencing factors can be completed. The present invention can improve the efficiency of identifying targeted carbon intensity reduction.

[0005] The technical solution steps of the present invention are as follows:

[0006] A method for identifying key influencing factors of carbon intensity based on random forests, the specific steps are as follows:

[0007] 1) Data collection and classification: including carbon intensity data of the region to be identified for several years and data on potential carbon intensity influencing factors, and classifying carbon intensity influencing factors by category, including fossil energy output, power scale, high energy consumption industry scale, energy consumption, transportation industry development, residents' income and consumption, technological progress and land use;

[0008] 2) The carbon intensity of the region is taken as the dependent variable, the potential carbon intensity influencing factor data is taken as the independent variable, and the carbon intensity potential influencing factor data is normalized;

[0009] 3) Build a random forest model based on decision trees, that is, establish a model algorithm and verify it. The specific model building process is divided into the following steps:

[0010] S1, randomly divide the carbon intensity dataset into training set and test set in proportion, and input all potential influencing factor data into the RandomForestRegressor model;

[0011] S2, set the parameters of sklearn.RandomForestRegressor and train the data set. The parameters include: n_estimators (maximum number of iterations of decision tree), oob_score (out-of-bag sample evaluation), criterion (division feature evaluation criterion), max_features (maximum number of features), max_depth (maximum depth of decision tree), min_samples_split (minimum number of samples required for internal node re-division), min_samples_leaf (minimum number of samples for leaf nodes), min_weight_fraction_leaf (minimum sample weight for leaf nodes), max_leaf_nodes (maximum number of leaf nodes), min_impurity_split (minimum impurity for node partitioning), etc.

[0012] The mean square error is selected as the evaluation standard for feature division at the node. The construction process of each decision tree follows the principle of minimum mean square error of node features. The grid search (GridSearchCV) method is combined with manual parameter adjustment to determine the optimal parameters, and the RandomForestRegressor model construction can be completed.

[0013] S3, in order to verify the applicability of the training results in the overall data, the training model of the training set is used to test the model accuracy of the test set using the optimal parameters, and the mean square error (MSE) is used to characterize the model accuracy. The calculation formula is as follows:

[0014]

[0015] Where N is the number of samples, observed i is the true value of carbon intensity of the i-th sample, predicted i is the predicted value of carbon intensity of the ith sample. The smaller the MSE value, the higher the accuracy of the model.

[0016] 4) Result output: Based on the modeling results, output the importance of all potential influencing factors and rank them. By defining a certain number of key influencing factors, identify the key influencing factors of carbon intensity, and classify and analyze the key influencing factors according to the categories in step 1.

[0017] Beneficial effects of the present invention:

[0018] 1. The present invention overcomes the data dimension limitation of the traditional carbon intensity analysis model, provides the analysis capability to expand the scope of potential influencing factors related to carbon intensity, analyzes the sources of carbon intensity contribution from the overall perspective of social and economic development, and completes the model input as the entire data set.

[0019] 2. The present invention improves the identification method of key influencing factors of carbon intensity from the aspect of algorithm application, which can realize the identification of key influencing factors of regional carbon intensity and provide rich scientific basis for the realization of regional carbon intensity reduction goals.

[0020] 3. The present invention introduces machine learning algorithm modeling to achieve a classified discussion of key influencing factors of regional carbon intensity, and simply and intuitively shows the degree of influence of each category of influencing factors on carbon intensity. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is the overall flow chart of the present invention.

[0022] Figure 2 The potential influencing factors and classifications of the embodiments of the present invention.

[0023] Figure 3 It is the importance ratio of the key influencing factor categories in the embodiment of the present invention.

[0024] Figure 4 It is a flow chart of the model algorithm of the present invention. Specific implementation methods

[0025] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0026] Liaoning Province is selected as the example region of this embodiment, and the research time range is 2001-2019. The carbon intensity and potential influencing factor data of the region are integrated and processed. The overall process framework is as follows: Figure 1 As shown, the algorithm flow is as follows Figure 4 As shown, the specific steps are as follows:

[0027] 1) Data acquisition: In the carbon intensity data, carbon emissions are calculated using the estimation method in the 2006 IPCC National Greenhouse Gas Inventory Guidelines, and regional gross domestic product (GDP) is derived from the Liaoning Statistical Yearbook. The data on potential carbon intensity influencing factors are derived from the China Statistical Yearbook, China Energy Statistical Yearbook, Liaoning Statistical Yearbook, China High-tech Industry Statistical Yearbook, China Population and Employment Statistical Yearbook, Liaoning Provincial Bureau of Statistics Information Disclosure Database and International Energy Agency (IEA) Database, etc., and the numerical dimension is unified under each data category.

[0028] Taking into account the principle of avoiding duplication of information utility as much as possible and combining data availability, the potential carbon intensity influencing factors are divided into 8 categories: fossil energy output, power scale, high energy consumption industry scale, energy consumption, transportation industry development, residents' income and consumption, technological progress and land use, totaling 69 items, forming the total data set, totaling 1330 data. The data list and classification are shown in Table 1.

[0029] Table 1 Factors and categories affecting potential carbon intensity in Liaoning Province

[0030]

[0031] 2) Data normalization: Since all features of the sample have the same impact on the distance of the sample in the feature space, the scale module in sklearn.preprocessing is called to normalize the potential influencing factor data. Through the mean and variance of the data set, they are mapped to the same scale of [0,1] to generate a new data set.

[0032] 3) Carbon intensity data was used as the dependent variable, and the normalized potential influencing factor data was used as the independent variable to construct a random forest identification model based on decision trees.

[0033] The specific algorithm construction process is divided into the following steps:

[0034] S1, the data set generated in step 2 is divided into a training set and a test set in a ratio of 8:2 using random sampling with replacement (Bagging method).

[0035] S2, set the RandomForestRegressor parameters and train the data set: a, use grid search (GridSearchCV) to set the primary parameter n_estimators (the number of trees) to range (1, 20), the step size is 1, and the evaluation criterion parameter criterion is 'neg_mean_squared_error' (mean square error); b, max_depth (maximum depth of decision tree) parameter search range (1, 30), step size is 2, min_samples_split (minimum number of samples required for node re-division) parameter search is within [2, 3, 4], max_features (maximum number of features) parameter search is set to ["auto", "sqrt", "log2"]; c, manually adjust min_samples_leaf (minimum number of samples for leaf nodes) and min_weight_fraction_leaf (minimum sample weight for leaf nodes); d, manually adjust max_leaf_nodes (maximum number of leaf nodes) and min_impurity_split (minimum impurity for node division); e, output the best parameter to determine the model, the specific parameters are as follows:

[0036] n_estimators: 10,

[0037] criterion: 'neg_mean_squared_error'

[0038] oob_score: True

[0039] max_depth:13,

[0040] min_samples_split:3,

[0041] max_features:auto,

[0042] min_samples_leaf: 6,

[0043] min_weight_fraction_leaf: 0

[0044] max_leaf_nodes: None

[0045] min_impurity_split: 0

[0046] At this point, the model construction is completed. Through training, the feature importance value in the model corresponds to the importance of the impact of each potential influencing factor on carbon intensity in the data set.

[0047] S3, use the test set data to test the model accuracy. The calculation formula of mean square error (MSE) is as follows:

[0048]

[0049] Where N is the number of samples, observed i is the true value of carbon intensity of the i-th sample, Predicted i is the predicted value of carbon intensity of the ith sample.

[0050] The mean square error calculation result of the recognition model is 0.004139. The smaller the mean square error, the higher the accuracy of the model. The results show that the model has good prediction performance.

[0051] S4, output the identification results, arrange the feature importance from large to small, and obtain the influence of potential carbon intensity influencing factors on carbon intensity.

[0052] 4) After sorting out the identification results in step 3, it was found that nearly 1 / 2 of the indicators can reflect more than 90% of the importance of the entire data set. Considering the requirements of identification accuracy and simplicity, the number of identified key influencing factors of carbon intensity was determined to be 32. The identification results are shown in Table 2.

[0053] Table 2 Key influencing factors and categories of carbon intensity in Liaoning Province from 2001 to 2019

[0054]

[0055] From the perspective of individual key influencing factors of carbon intensity, during the development of Liaoning Province in the first 20 years of the 21st century, the factors that had the greatest impact on the overall carbon intensity were the energy intensity of the pulp and printing industry, natural gas production, crude oil production, power generation and electricity consumption, all of which accounted for more than 5% of the importance. The identified key influencing factors were classified and analyzed, and more than 5 indicators were selected in the categories of energy consumption, residents' income and consumption, scale of high-energy-consuming industries, and technological progress. From the perspective of category importance, 32 key influencing factors accounted for 90.39% of the overall carbon intensity importance, among which technological progress had the highest importance, at 24.81%. The categories with an importance share of more than 10% were residents' income and consumption, fossil energy production, and energy consumption, and the scale of high-energy-consuming industries also accounted for 9.02% of the importance. Figure 2 , Figure 3 shown.

[0056] The identification results can be understood as carbon emissions in the production process behind the top-ranked factors have a more significant impact on regional carbon intensity, and emission reduction policies formulated for them may achieve more prominent results. It can provide directional suggestions for the subsequent formulation of carbon emission reduction policies, and help policy adjustments to achieve targeted and efficient control of regional carbon intensity while ensuring the speed of economic development.

Claims

1. A method for identifying key influencing factors of carbon intensity based on random forests, characterized in that: The specific steps are as follows: 1) Data collection and classification: including carbon intensity data of the region to be identified for several years and data on potential carbon intensity influencing factors, and classifying carbon intensity influencing factors into categories, including fossil energy output, power scale, high energy consumption industry scale, energy consumption, transportation industry development, residents' income and consumption, technological progress and land use; 2) The carbon intensity of the region is taken as the dependent variable, the potential carbon intensity influencing factor data is taken as the independent variable, and the carbon intensity potential influencing factor data is normalized; 3) Construct a random forest model based on decision trees, that is, establish a model algorithm and verify it; the specific model construction process is divided into the following steps: S1, randomly divide the carbon intensity dataset into training set and test set in proportion, and input all potential influencing factor data into the RandomForestRegressor model; S2, set the parameters of sklearn.RandomForestRegressor and train the data set. The parameters include: the maximum number of iterations of the decision tree n_estimators, the out-of-bag sample evaluation oob_score, the partition feature evaluation criterion, the maximum number of features max_features, the maximum depth of the decision tree max_depth, the minimum number of samples required for internal node re-division min_samples_split, the minimum number of samples for leaf nodes min_samples_leaf, the minimum sample weight of leaf nodes min_weght_fraction_leaf, the maximum number of leaf nodes max_leaf_nodes and the minimum impurity of node partition min_impurity_split; The mean square error is selected as the evaluation standard for node division features. The construction process of each decision tree follows the principle of minimum mean square error of node features. The grid search method is combined with manual parameter adjustment to determine the optimal parameters, thus completing the construction of the RandomForestRegressor model. S3, in order to verify the applicability of the training results in the overall data, the training model of the training set is used to test the model accuracy of the test set using the optimal parameters, and the mean square error (MSE) is used to characterize the model accuracy. The calculation formula is as follows: Where N is the number of samples, observed i is the true value of carbon intensity of the i-th sample, predicted i is the predicted value of carbon intensity of the i-th sample; the smaller the MSE value, the higher the accuracy of the model; 4) Result output: Based on the modeling results, output the importance of all potential influencing factors and rank them. By defining a certain number of key influencing factors, identify the key influencing factors of carbon intensity, and classify and analyze the key influencing factors according to the categories in step 1.

Images