BP neural network county area energy consumption carbon emission measuring and calculating method based on feature fusion

Abstract

The invention relates to a feature fusion-based county energy consumption carbon emission measuring and calculating method based on a BP neural network, and the method comprises the steps: analyzing the importance of seven types of features, namely, a regional total production value, a total fixed asset investment amount, the number of permanent residents at the end of a year, an urbanization rate, a green land area, a second industry specific gravity and a total night light value, in energy consumption carbon emission through a Gini coefficient; different weights are distributed to the features, and new features are obtained on the basis of weight fusion and serve as model input. The county energy consumption carbon emission measuring and calculating method based on the BP neural network based on feature fusion is established by adopting a modeling mode of the BP neural network, the current situation that a traditional energy consumption carbon emission measuring and calculating method is lack of data or low in measuring and calculating precision is changed, the modeling precision is greatly improved, county energy consumption carbon emission can be effectively measured and calculated, and the measuring and calculating efficiency is improved. And a scientific basis is provided for formulating a reasonable differentiated carbon emission reduction policy and developing green low-carbon economy.

Description

Technical Field

[0001] The present invention belongs to the technical field of carbon emission calculation, and relates to a method for calculating carbon emission from county energy consumption based on a BP neural network with feature fusion. Background Art

[0002] Climate warming is one of the biggest challenges facing mankind today. In recent years, extreme weather and sea level rise caused by the greenhouse effect have brought great harm. Therefore, it is necessary to analyze the spatial differences, dependencies and spatial heterogeneity of influencing factors of carbon emissions at the county level, and provide a scientific basis for formulating reasonable differentiated carbon emission reduction policies and developing a green and low-carbon economy. All of these are inseparable from the measurement of carbon emissions from county energy consumption. Accurate monitoring of carbon emissions over a long time series and a large range has always been one of the key contents of carbon emission-related research. Night light data is a more commonly used spatial resolution variable that can effectively reflect the activity intensity of human society. Compared with statistical data, night light data has the advantages of good economy, strong timeliness, large continuous time span and wide spatial coverage. More and more scholars have begun to use night light remote sensing data to spatially estimate carbon emissions. However, using only night light data cannot fully reflect the carbon emissions in the region, and the modeling accuracy is not high. Therefore, it is necessary to introduce other factors that have a greater impact on carbon emissions, improve the modeling method, and improve the accuracy of carbon emission calculation. Summary of the invention

[0003] The present invention provides a method for calculating carbon emissions from energy consumption in counties based on a BP neural network with feature fusion. The specific process is as follows:

[0004] Step a: Preprocess the NPP-VIIRS night light data of the study area;

[0005] Step b: Preprocessing the land use data of the study area;

[0006] Step c: Estimate carbon emissions from energy consumption in county-level administrative areas;

[0007] Step d: Collect seven indicators, including regional GDP, total fixed asset investment, permanent population at the end of the year, urbanization rate, green area, proportion of secondary industry and total value of night lights, to build an energy consumption carbon emission calculation model, analyze the impact of energy consumption carbon emissions in the administrative region based on the Gini coefficient, and perform feature fusion according to the impact;

[0008] Step e: Construct a BP neural network model to calculate the carbon emissions from energy consumption in each county.

[0009] Furthermore, step a preprocesses the NPP-VIIRS night light data of the study area, including the following steps:

[0010] (1) Using the NPP-VIIRS monthly data products, the VCM data from the monthly average radiation composite dataset from 2012 to 2017 were selected for processing;

[0011] (2) Perform projection transformation on the night light image, convert it into Lambert equal-area projection, and resample the NPP-VIIRS night light image to a spatial resolution of 500 meters;

[0012] (3) Values ​​less than 0.5 in night light images are considered noise, and their pixels are assigned a value of 0;

[0013] (4) Using the threshold method, the maximum brightness of the night light images of Beijing, Shanghai, and Guangzhou in the same year was selected as the threshold, and the 8-neighborhood method was used to eliminate the outliers of the night light images of the administrative area;

[0014] (5) Calculate the annual mean of NPP-VIIRS images using the following formula:

[0015]

[0016] Where: DN n Indicates the NPP-VIIRS mean night light data DN value in the nth year, DN i represents the NPP-VIIRS night light DN value in month i;

[0017] (6) Use the vector data of the study area to crop the night light images.

[0018] Furthermore, step b preprocesses the land use data, and has the following steps:

[0019] (1) Align the CLCD land use data with the coordinate system of the night light image, and use the vector data of the study area to clip the land use data of the study area;

[0020] (2) Each land use type was resampled and the current status of the nine types of land use in each district and county was calculated using ArcGIS zoning statistics. Woodlands and shrubs were classified as woodlands, and water bodies, wetlands, and ice were classified as water bodies.

[0021] Furthermore, step c estimates the carbon emissions from energy consumption in county-level administrative areas, specifically including the following steps:

[0022] (1) The direct carbon emission coefficient method was used to calculate the carbon emissions of cultivated land, forest land, grassland, water area and unused land. The carbon emission and absorption coefficients of different land types were 0.497, -0.623, -0.021, -0020 and -0.0005 (t·hm -2 ·a -1), the specific formula is as follows:

[0023] E x =ΣA i =ΣS i ×T i (2)

[0024] In the formula, E x is direct carbon emissions (t); A i is the carbon emission of the i-th land use type S i (t); is the area of ​​the i-th land use type (hm 2 );T i is the carbon emission coefficient of the i-th land use type (t·hm- 2 ·a- 1 );

[0025] (2) The net carbon emissions of each county and district provided by the China Carbon Accounting Database are used to calculate the carbon emissions from energy consumption in each county and district. The formula is as follows:

[0026] C E =C CEADs -E x (3)

[0027] In the formula, C E is the carbon emission of energy consumption (t); C CEADs Net carbon emissions (t) provided by China Carbon Accounting Database, E x is the direct carbon emission (t).

[0028] Furthermore, step d collects data such as regional GDP, total fixed asset investment, permanent population at the end of the year, urbanization rate, green area, proportion of secondary industry and total value of night lights, analyzes the impact of energy consumption and carbon emissions in the administrative region based on the Gini coefficient, and performs feature fusion according to the impact, which mainly includes the following steps:

[0029] (1) The socioeconomic indicators that affect carbon emissions from energy consumption were subjected to Pearson correlation analysis and VIF collinearity diagnosis, and weakly correlated and multicollinear indicators were removed. The energy consumption carbon emissions calculation model was constructed using seven indicators, namely, regional GDP, total fixed asset investment, year-end permanent population, urbanization rate, green area, proportion of secondary industry, and total value of night lights;

[0030] (2) Perform feature fusion based on the Gini coefficient. The Gini index is used to analyze the importance of the seven types of features in energy consumption and carbon emissions, and different weights are assigned to them. Based on the weight fusion, new features are obtained as model input. The calculation formula of the Gini index is:

[0031]

[0032] Among them, k means there are k categories, and P k represents the weight of the kth category. At the same time, the contribution of feature j is calculated as follows:

[0033]

[0034] Among them, VIM i represents the total Gini index change of feature j. N is the number of all features. Therefore, the feature importance is weighted and fused to obtain the input for the energy consumption carbon emission estimation model training, and its expression can be expressed as:

[0035]

[0036] Among them, HI i is the jth eigenvector. i Entered into the model as new input features.

[0037] Furthermore, step e constructs a BP neural network model to calculate the carbon emissions of energy consumption in each county, which specifically includes the following contents:

[0038] (1) Input data normalization: Before training, the input data needs to be normalized and mapped to the interval [0, 1]. This can eliminate the impact of different magnitudes of various factors on prediction accuracy and help speed up the learning process. After the network runs, the data is denormalized to obtain the predicted data and output it.

[0039] (2) Define the network topology structure When initializing the forward propagation process, it is necessary to determine the number of neurons n in the input layer, the number of layers and neurons l in the hidden layer, and the number of neurons m in the output layer. The number of neurons n in the input layer is the number of influencing factors, and the number of neurons m in the output layer is the number of predictive variables. If the number of neurons in the hidden layer is too small, the network structure will be insufficient and effective information cannot be extracted, which makes it difficult to meet the specific needs of network learning. If the number of neurons is too large, the training time of the neural network will increase and the amount of calculation will increase. Therefore, the number of neurons in the hidden layer can be selected by trial and error. There are three empirical formulas: l = log2n, a is a regulation constant between [1, 10]. This paper adopts the first empirical formula to determine the approximate range, and then selects the number of neurons that minimizes the error by calculating the error;

[0040] (3) Weight threshold initialization The weight threshold is initialized to a small random number. Since the weight not only needs to activate neurons but also needs to inhibit neurons, the connection weight and threshold of each neuron are assigned a random number in the interval [-1, 1];

[0041] (4) Calculate the output of the hidden layer according to the input sample x i , the connection weight w between the i-th neuron in the input layer and the j-th neuron in the hidden layer ij and the threshold a of the jth neuron in the hidden layer j The input S of each neuron in the hidden layer can be calculated j , and then apply the activation function to the input S j , we can get the output H of each neuron in the hidden layer j , the formula is as follows:

[0042]

[0043] H j =f(S j ) (8)

[0044] Where f(x) is the activation function sigmoid function of the hidden layer;

[0045] (5) Calculate the output of the output layer according to the output H of the hidden layer j , the connection weight w between the jth neuron in the hidden layer and the kth neuron in the output layer jk And the threshold b of the kth neuron in the output layer k The input P of each neuron in the output layer can be calculated k , and then apply the activation function to the input P k , the output of each neuron in the output layer can be calculated k , the formula is as follows:

[0046]

[0047] O k =f(P k ) (10)

[0048] (6) Calculate the error function, based on the actual output value of the BP neural network k and the expected output value Y k Calculate the error e k , the error function can be further calculated, that is, the mean square error E k , the weights and thresholds are repeatedly modified by the gradient descent method to minimize the error function. The formula is as follows:

[0049] e k =Y k -O k , k = 1, 2, ..., m (11)

[0050]

[0051] (7) When the training reaches the maximum number of iterations or reaches a certain learning rate, the training can be stopped. The predicted value is compared with the historical true value, and three evaluation indicators are selected to test the model accuracy:

[0052] (1) Coefficient of determination R 2 It is an important indicator used to measure the goodness of fit of a model, reflecting the influence of multiple independent variables on the dependent variable. 2 ∈[0, 1], and the closer it is to 1, the better the model fit is. 2 The calculation formula is as follows:

[0053]

[0054] Where: y i is the true value of the sample; is the sample prediction value; is the sample mean; n is the number of samples;

[0055] (2) Mean absolute percentage error is a relative error measure. It uses absolute values ​​to avoid the offset of positive and negative errors. It can well reflect the actual situation of the errors of each prediction result. MAPE∈[0,+∞]. A MAPE of 0% indicates a perfect model. A MAPE greater than 100% indicates a poor model. The smaller the MAPE value, the better the accuracy of the prediction model. The MAPE calculation formula is as follows:

[0056]

[0057] Where: y i is the true value of the sample; is the sample prediction value; is the sample mean; n is the number of samples;

[0058] (3) The root mean square error, also known as the standard error, can reflect the overall discreteness of a data set. RMSE∈[0,+∞]. When the predicted value is completely consistent with the historical true value, it is equal to 0, that is, a perfect model. The larger the error, the larger the value. The RMSE calculation formula is as follows:

[0059]

[0060] Where: y i is the true value of the sample; is the sample prediction value; is the sample mean; n is the number of samples.

[0061] Advantages of the present invention

[0062] Compared with the prior art, the present invention has the following technical effects:

[0063] 1. Compared with the traditional method of calculating carbon emissions only by using the DN value of night light data, the present invention adds feature fusion based on the Gini coefficient. According to the different importance of different data features for energy consumption carbon emissions, if different weights are assigned to each type of feature, the accuracy of county energy consumption carbon emissions calculation can be improved.

[0064] 2. Compared with the traditional method of single variable linear modeling, the present invention proposes a BP neural network modeling method based on feature fusion, which greatly improves the modeling accuracy and effectively measures the carbon emissions of county energy consumption, providing a scientific basis for formulating reasonable differentiated carbon emission reduction policies and developing a green and low-carbon economy. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] The description of the content of the present invention will become obvious and easy to understand in conjunction with the following drawings, in which:

[0066] Figure 1 It is a flowchart of a method for estimating carbon emissions from energy consumption in counties using a BP neural network based on feature fusion;

[0067] Figure 2 It is the result of preprocessing of NPP-VIIRS nighttime light images in the present invention;

[0068] Figure 3 It is the result of the importance of the carbon emission characteristics of energy consumption calculated by the present invention based on the Gini coefficient;

[0069] Figure 4 It is a schematic diagram of the training set test results of a BP neural network county energy consumption carbon emission estimation method based on feature fusion in the present invention. DETAILED DESCRIPTION

[0070] according to Figure 1 The steps shown are used to explain in detail a method for calculating carbon emissions from energy consumption in counties based on a BP neural network with feature fusion according to the present invention.

[0071] Step 1: Preprocess the NPP-VIIRS night light data of the study area. The specific steps are as follows:

[0072] (1) Using the NPP-VIIRS monthly data products, the VCM data from the monthly average radiation composite dataset from 2012 to 2017 were selected for processing;

[0073] (2) Perform projection transformation on the night light image and convert it into Lambert equal-area projection; resample the NPP-VIIRS night light image to a spatial resolution of 500 meters;

[0074] (3) Values ​​less than 0.5 in night light images are considered noise, and their pixels are assigned a value of 0;

[0075] (4) Using the threshold method, the maximum brightness of the night light images of Beijing, Shanghai, and Guangzhou in the same year was selected as the threshold, and the 8-neighborhood method was used to eliminate the outliers of the night light images of the administrative area;

[0076] (5) Calculate the annual mean of NPP-VIIRS images using the following formula:

[0077]

[0078] Where: DN n Indicates the NPP-VIIRS mean night light data DN value in the nth year, DN i represents the NPP-VIIRS night light DN value in month i;

[0079] (6) Use the vector data of the study area to crop the night light images.

[0080] Step 2: Preprocess the land use data of the study area. The specific steps are as follows:

[0081] (1) Align the CLCD land use data with the coordinate system of the night light image, and use the vector data of the study area to clip the land use data of the study area;

[0082] (2) Each land use type was resampled and the current status of the nine types of land use in each district and county was calculated using ArcGIS zoning statistics. Woodlands and shrubs were classified as woodlands, and water bodies, wetlands, and ice were classified as water bodies.

[0083] Step 3: Estimate the carbon emissions from energy consumption in county-level administrative areas. The specific steps are as follows:

[0084] (1) The direct carbon emission coefficient method was used to calculate the carbon emissions of cultivated land, forest land, grassland, water area and unused land. The carbon emission and absorption coefficients of different land types were 0.497, -0.623, -0.021, -0020 and -0.0005 (t·hm -2 ·a -1 ), the specific formula is as follows:

[0085] E x =ΣA i =ΣS i ×T i (2)

[0086] In the formula, E x is direct carbon emissions (t); A iis the carbon emission of the i-th land use type S i (t); is the area of ​​the i-th land use type (hm 2 );T i is the carbon emission coefficient of the i-th land use type (t·hm -2 ·a -1 );

[0087] (2) The net carbon emissions of each county and district provided by the China Carbon Accounting Database are used to calculate the carbon emissions from energy consumption in each county and district. The formula is as follows:

[0088] C E =C CEADs -E x (3)

[0089] In the formula, C E is the carbon emission of energy consumption (t); C CEADs Net carbon emissions (t) provided by China Carbon Accounting Database, E x is the direct carbon emission (t).

[0090] Step 4: Collect seven indicators, including regional GDP, total fixed asset investment, permanent population at the end of the year, urbanization rate, green area, proportion of secondary industry and total value of night lights, to build an energy consumption carbon emission calculation model, analyze the impact of energy consumption carbon emissions in the administrative region based on the Gini coefficient, and perform feature fusion according to the impact. The specific steps are as follows:

[0091] (1) The socioeconomic indicators that affect carbon emissions from energy consumption were subjected to Pearson correlation analysis and VIF collinearity diagnosis, and weakly correlated and multicollinear indicators were removed. A model for calculating carbon emissions from energy consumption was constructed using seven indicators, namely, regional GDP, total fixed asset investment, year-end permanent population, urbanization rate, green area, proportion of secondary industry, and total value of night lights.

[0092] (2) Perform feature fusion based on the Gini coefficient. The Gini index is used to analyze the importance of seven types of features in energy consumption and carbon emissions, assign different weights to them, and obtain new features based on weight fusion as model input. The calculation formula of the Gini index is:

[0093]

[0094] Among them, k means there are k categories, and P k represents the weight of the kth category. At the same time, the contribution of feature j is calculated as follows:

[0095]

[0096] Among them, VIMi represents the total Gini index change of feature j. N is the number of all features. Therefore, the feature importance is weighted and fused to obtain the input for the energy consumption carbon emission estimation model training, and its expression can be expressed as:

[0097]

[0098] Among them, HI i is the jth eigenvector. i Entered into the model as new input features.

[0099] Step 5: Construct a BP neural network model to calculate the carbon emissions of energy consumption in each county. The specific steps are as follows:

[0100] (1) Input data normalization: Before training, the input data needs to be normalized and mapped to the interval [0, 1]. This can eliminate the impact of different magnitudes of various factors on prediction accuracy and help speed up the learning process. After the network runs, the data is denormalized to obtain the predicted data and output it.

[0101] (2) Define the network topology structure When initializing the forward propagation process, it is necessary to determine the number of neurons n in the input layer, the number of layers and neurons l in the hidden layer, and the number of neurons m in the output layer. The number of neurons n in the input layer is the number of influencing factors, and the number of neurons m in the output layer is the number of predictive variables. If the number of neurons in the hidden layer is too small, the network structure will be insufficient and effective information cannot be extracted, which makes it difficult to meet the specific needs of network learning. If the number of neurons is too large, the training time of the neural network will increase and the amount of calculation will increase. Therefore, the number of neurons in the hidden layer can be selected by trial and error. There are three empirical formulas: l = log2n, a is a regulation constant between [1, 10]. This paper adopts the first empirical formula to determine the approximate range, and then selects the number of neurons that minimizes the error by calculating the error;

[0102] (3) Weight threshold initialization The weight threshold is initialized to a small random number. Since the weight not only needs to activate neurons but also needs to inhibit neurons, the connection weight and threshold of each neuron are assigned a random number in the interval [-1, 1];

[0103] (4) Calculate the output of the hidden layer according to the input sample x i , the connection weight w between the i-th neuron in the input layer and the j-th neuron in the hidden layer ij and the threshold a of the jth neuron in the hidden layer j The input S of each neuron in the hidden layer can be calculated j , and then apply the activation function to the input S j, we can get the output H of each neuron in the hidden layer j , the formula is as follows:

[0104]

[0105] H j =f(S j ) (8)

[0106] Where f(x) is the activation function sigmoid function of the hidden layer;

[0107] (5) Calculate the output of the output layer according to the output H of the hidden layer j , the connection weight w between the jth neuron in the hidden layer and the kth neuron in the output layer jk And the threshold b of the kth neuron in the output layer k The input P of each neuron in the output layer can be calculated k , and then apply the activation function to the input P k , the output of each neuron in the output layer can be calculated k , the formula is as follows:

[0108]

[0109] O k =f(P k ) (10)

[0110] (6) Calculate the error function, based on the actual output value of the BP neural network k and the expected output value Y k Calculate the error e k , the error function can be further calculated, that is, the mean square error E k , the weights and thresholds are repeatedly modified by the gradient descent method to minimize the error function. The formula is as follows:

[0111] e k =Y k -O k , k = 1, 2, ..., m (11)

[0112]

[0113] (7) When the training reaches the maximum number of iterations or reaches a certain learning rate, the training can be stopped. The predicted value is compared with the historical true value, and three evaluation indicators are selected to test the model accuracy:

[0114] (1) Coefficient of determination R 2 It is an important indicator used to measure the goodness of fit of a model, reflecting the influence of multiple independent variables on the dependent variable. 2∈[0, 1], and the closer it is to 1, the better the model fit is. 2 The calculation formula is as follows:

[0115]

[0116] Where: y i is the true value of the sample; is the sample prediction value; is the sample mean; n is the number of samples;

[0117] (2) Mean absolute percentage error is a relative error measure. It uses absolute values ​​to avoid the offset of positive and negative errors. It can well reflect the actual situation of the errors of each prediction result. MAPE∈「0,+∞), MAPE of 0% indicates a perfect model, MAPE greater than 100% indicates a poor model. The smaller the MAPE value, the better the accuracy of the prediction model. The calculation formula of MAPE is as follows:

[0118]

[0119] Where: y i is the true value of the sample; is the sample prediction value; is the sample mean; n is the number of samples;

[0120] (3) The root mean square error, also known as the standard error, can reflect the overall discreteness of a data set. RMSE∈[0,+∞). When the predicted value is completely consistent with the historical true value, it is equal to 0, that is, a perfect model. The larger the error, the larger the value. The RMSE calculation formula is as follows:

[0121]

[0122] Where: y i is the true value of the sample; is the sample prediction value; is the sample mean; n is the number of samples.

[0123] The present invention discloses a method for calculating carbon emissions from energy consumption in counties based on a BP neural network with feature fusion. The Gini coefficient is used to analyze the importance of seven types of features, namely, regional gross domestic product, total fixed asset investment, permanent population at the end of the year, urbanization rate, green area, proportion of secondary industry, and total value of nighttime lights, in carbon emissions from energy consumption. Different weights are assigned to them, and new features are obtained based on weight fusion as model input. A method for calculating carbon emissions from energy consumption in counties based on a BP neural network with feature fusion is established by adopting the modeling method of a BP neural network. This method changes the current situation of data missing or low calculation accuracy in traditional methods for calculating carbon emissions from energy consumption, greatly improves the modeling accuracy, and can effectively calculate carbon emissions from energy consumption in counties, providing a scientific basis for formulating reasonable differentiated carbon emission reduction policies and developing a green and low-carbon economy.

[0124] 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 principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for calculating carbon emissions from energy consumption in counties based on BP neural network with feature fusion, characterized in that: The specific steps include: Step 1: Preprocess the NPP-VIIRS night light data of the study area; Step 2: Preprocess the land use data of the study area; Step 3: Estimate carbon emissions from energy consumption in county-level administrative areas; Step 4: Collect seven indicators, including regional GDP, total fixed asset investment, permanent population at the end of the year, urbanization rate, green area, proportion of secondary industry and total value of night lights, to build an energy consumption carbon emission calculation model, analyze the impact of energy consumption carbon emissions in the administrative region based on the Gini coefficient, and perform feature fusion according to the impact; Step 5: Construct a BP neural network model to calculate the carbon emissions from energy consumption in each county.

2. A method for calculating carbon emissions from energy consumption in counties based on a BP neural network with feature fusion as claimed in claim 1, characterized in that: The step 1 includes the following sub-steps: Step 1.1, select the NPP-VIIRS monthly data product and the VCM data in the monthly average radiation composite dataset from 2012 to 2017 for processing; Step 1.2: Perform projection transformation on the night light image and convert it into Lambert equal-area projection; resample the NPP-VIIRS night light image to make the image spatial resolution 500 meters; Step 1.3: The values ​​less than 0.5 in the night light image are regarded as noise, and the pixel value is assigned to 0; Step 1.4, using the threshold method, select the maximum brightness of the night light images of Beijing, Shanghai, and Guangzhou in the same year as the threshold, and use the 8-neighborhood method to eliminate the abnormal values ​​of the night light images of the administrative area; Step 1.5, calculate the mean of the annual NPP-VIIRS image, the formula is as follows: （1） Where: Indicates Annual NPP-VIIRS mean night light data value, Indicates Moon NPP-VIIRS Night Lights value; Step 1.6: Use the vector data of the study area to crop the night light image.

3. According to the method for calculating carbon emissions from energy consumption in counties based on BP neural network with feature fusion as claimed in claim 1, step 2 comprises the following steps: Step 2.1, align the CLCD land use data with the coordinate system of the night light image, and use the vector data of the study area to clip the land use data of the study area; In step 2.2, each land use type was resampled and the current land use status of the nine types in each district and county was calculated using ArcGIS zoning statistics. Woodland and shrubs were classified as woodland, and water bodies, wetlands, and ice were classified as water bodies.

4. The method for calculating carbon emissions from energy consumption in counties based on BP neural network with feature fusion as claimed in claim 1 is characterized in that: Step 3 specifically includes the following sub-steps: Step 3.1, the direct carbon emission coefficient method is used to calculate the carbon emissions of cultivated land, forest land, grassland, water area and unused land. The carbon emission and absorption coefficients of different land types are 0.497, -0.623, -0.021, -0020 and -0.0005 respectively ( ), the specific formula is as follows: （2） In the formula, is direct carbon emissions (t); For the Carbon emissions from land use types (t); for the Area of ​​land use type ( ); For the Carbon emission coefficient of land use type ( ); Step 3.2, use the net carbon emissions of each county and district provided by the China Carbon Accounting Database to calculate the carbon emissions of energy consumption in each county and district. The formula is as follows: （3） In the formula, is the carbon emission from energy consumption (t); Net carbon emissions provided to China Carbon Accounting Database (t), is the direct carbon emission (t).

5. The method for calculating carbon emissions from energy consumption in counties based on BP neural network with feature fusion as claimed in claim 1 is characterized in that: The step 4 mainly includes the following sub-steps: Step 4.1: Perform Pearson correlation analysis and VIF collinearity diagnosis on the socioeconomic indicators that affect energy consumption carbon emissions, remove weakly correlated and multicollinear indicators, and use seven indicators, including regional GDP, total fixed asset investment, year-end permanent population, urbanization rate, green area, proportion of secondary industry, and total value of night lights, to build an energy consumption carbon emissions calculation model; Step 4.2, perform feature fusion based on the Gini coefficient. The Gini index is used to analyze the importance of the seven types of features in energy consumption and carbon emissions, and different weights are assigned to them. Based on the weight fusion, new features are obtained as model input. The calculation formula of the Gini index is: （4） in, Indicates that there is categories, and Indicates The weights of the categories, at the same time, the features The contribution of is calculated as follows: （5） in, Representation characteristics The total Gini index change is is the number of all features, therefore, the feature importance is weighted and fused to obtain the input for the energy consumption carbon emission estimation model training, which can be expressed as: （6） in For the feature vectors, Entered into the model as new input features.

6. The method for calculating carbon emissions from energy consumption in counties based on BP neural network with feature fusion as claimed in claim 1 is characterized in that: Step 5 includes the following sub-steps: Step 5.1, normalize the input data. Before training, the input data needs to be normalized and mapped to the interval [0, 1]. This can eliminate the impact of different magnitudes of various factors on the prediction accuracy and help speed up the learning process. After the network runs, the data is denormalized to obtain the predicted data and output it. Step 5.2, define the network topology structure. When initializing the forward propagation process, it is necessary to determine the number of neurons n in the input layer, the number of layers and neurons l in the hidden layer, and the number of neurons m in the output layer. Among them, the number of neurons n in the input layer is the number of influencing factors, and the number of neurons m in the output layer is the number of predictive variables. If the number of neurons in the hidden layer is too small, the network structure will be insufficient and effective information cannot be extracted, which makes it difficult to meet the specific needs of network learning; if the number of neurons is too large, the training time of the neural network will increase and the amount of calculation will increase. Therefore, the number of neurons in the hidden layer can be selected by trial and error. There are three empirical formulas: , , , is a regulation constant between [1, 10]. This paper adopts the first empirical formula to determine the approximate range, and then selects the number of neurons that minimizes the error by calculating the error. Step 5.3, weight threshold initialization. Initialize the weight threshold to a small random number. Since the weight not only needs to activate neurons but also needs to inhibit neurons, the connection weight and threshold of each neuron are assigned a random number in the interval [−1, 1]; Step 5.4, calculate the output of the hidden layer, according to the input sample , input layer neurons and hidden layers The connection weights between neurons And the hidden layer The threshold of a neuron The input of each neuron in the hidden layer can be calculated , and then apply the activation function to the input , we can get the output of each neuron in the hidden layer , the formula is as follows: （7） （8） In the formula, is the activation function sigmoid function of the hidden layer; Step 5.5, calculate the output of the output layer based on the output of the hidden layer , hidden layer neurons and the output layer The connection weights between neurons And the output layer The threshold of a neuron The input of each neuron in the output layer can be calculated , and then apply the activation function to the input , the output of each neuron in the output layer can be calculated , the formula is as follows: （9） （10） Step 5.6, calculate the error function, based on the actual output value of the BP neural network and expected output Calculated error , the error function can be further calculated, that is, the mean square error , the weights and thresholds are repeatedly modified by the gradient descent method to minimize the error function. The formula is as follows: （11） （12） 7. Step 5.7, when the training reaches the maximum number of iterations or reaches a certain learning rate, the training can be stopped, the predicted value is compared with the historical true value, and 3 evaluation indicators are selected to test the model accuracy: (1) Coefficient of determination It is an important indicator used to measure the goodness of fit of a model, reflecting the influence of multiple independent variables on the dependent variable. , and the closer it is to 1, the better the model fit is, and the coefficient of determination is The calculation formula is as follows: （13） Where: is the true value of the sample; is the sample prediction value; is the sample mean; n is the number of samples; (2) The mean absolute percentage error is a relative error measure that uses absolute values ​​to avoid the mutual cancellation of positive and negative errors. It can well reflect the actual situation of the errors of each prediction result. A MAPE of 0% indicates a perfect model, and a MAPE greater than 100% indicates a poor quality model. The smaller the MAPE value, the better the accuracy of the prediction model. The MAPE calculation formula is as follows: （14） Where: is the true value of the sample; is the sample prediction value; is the sample mean; n is the number of samples; (3) The root mean square error, also known as the standard error, can reflect the overall degree of dispersion of a data set. , when the predicted value is completely consistent with the historical true value, it is equal to 0, that is, a perfect model. The larger the error, the larger the value. The RMSE calculation formula is as follows: （15） Where: is the true value of the sample; is the sample prediction value; is the sample mean; n is the number of samples.

Images