A method and device for extracting features of urban carbon emission influencing factors and long-term prediction

Abstract

The application discloses a kind of urban carbon emission influence factor feature extraction and medium and long term prediction method and device, based on economic-energy-carbon emission balance system, selected urban carbon emission influence factor, and obtain influence factor historical data;To the data obtained, data preprocessing is carried out to obtain high-quality data set;Based on high-quality data set, the mechanism of action and contribution of the influence factors of urban carbon emission are effectively analyzed using the generalized Dittmar index decomposition method, and the key influence factors affecting urban carbon emission are obtained;Based on the obtained key influence factors, an integrated empirical mode decomposition-long short-term memory network method is used to predict a single variable;Based on the obtained influence factor prediction result, an XGBoost network optimized based on a particle swarm algorithm is used to construct a medium and long term prediction model for urban carbon emission.The present application considers factors such as energy consumption and economic growth, and considers the interaction and influence of multiple factors, thereby achieving accurate carbon emission prediction.

Description

Technical Field

[0001] This invention relates to the field of carbon emission prediction and analysis, and in particular to a method and apparatus for extracting the characteristics of factors influencing urban carbon emissions and for medium- and long-term prediction. Background Technology

[0002] Climate change has drawn widespread attention for its destructive impact on the ecological environment and its threat to human survival and development. A global effort to address climate change has become an inevitable choice. To achieve carbon emission reduction targets and improve energy efficiency and environmental protection, it is necessary to deeply explore and analyze the driving factors of carbon emissions, establish medium- and long-term carbon emission prediction models that can consider multiple influencing factors, and provide a scientific basis for formulating and implementing carbon emission reduction policies and improving energy efficiency and environmental protection.

[0003] Existing research indicates that carbon emissions are typically influenced by a combination of factors, including population, economy, technological progress, and industrial structure. Mainstream methods for studying the influencing factors of carbon emissions include the Log-Mean Dijkstra Index (LMDI) decomposition method and the STIRPAT model regression analysis method. Existing literature has used GDP, per capita energy consumption, urbanization rate, and CO2 emission intensity to decompose and predict China's carbon emission factors, and analyzed the contributions of different influencing factors through error decomposition. It has also used grey models and support vector regression models to predict the impact of energy consumption, industrial structure, and population growth, and proposed carbon reduction policies led by energy structure adjustment and technological progress. Furthermore, it has used time series models and grey correlation models to predict and analyze China's carbon emissions and their influencing factors, concluding that energy consumption, population size, economic growth rate, and energy structure have the most significant impacts on carbon emissions. Finally, it has used grey relational analysis and time series analysis to study the relationship between China's carbon emissions and factors such as energy structure, industrial structure, economic growth, and technological innovation, and predicted China's future carbon emission trends.

[0004] However, there are still some problems and challenges in the decomposition and prediction of carbon emission factors. First, existing models for decomposing and predicting carbon emission factors still have shortcomings, such as the influence of oversimplified factors, missing data, and insufficient sample size. Second, the factors affecting carbon emissions are very complex, requiring consideration of the interactions and influences of multiple factors; therefore, how to effectively extract and analyze information from each influencing factor still needs in-depth research. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for extracting the characteristics of factors affecting urban carbon emissions and predicting them in the medium and long term. This method comprehensively considers factors such as energy consumption and economic growth, as well as the interaction and influence of multiple factors, thereby achieving accurate carbon emission prediction.

[0006] This invention collects indicators of energy consumption, economic growth rate, electricity production, and carbon emissions, quantifies their impact contribution using the GDIM method, and selects highly correlated influencing factors. It uses the XGBoost algorithm with particle swarm optimization to predict future carbon emission trends, which can effectively identify and utilize the temporal characteristics of the data, thereby improving the estimation accuracy and reliability of the model.

[0007] The specific plan includes the following steps:

[0008] S1: Based on the "economy-energy-carbon emissions" balance system, select the factors influencing urban carbon emissions and obtain historical data on these factors;

[0009] S2: Perform data preprocessing on the dataset obtained in S1, including bad data identification, data cleaning, and data completion, to obtain a high-quality dataset;

[0010] S3: Based on the data provided in S2, the Generalized Dijkstra Index Decomposition Method (GDIM) is used to effectively analyze the mechanism and contribution of factors affecting urban carbon emissions, and to identify the key factors affecting urban carbon emissions.

[0011] S4: Based on S3, key influencing factors are obtained, and the EEMD-LSTM (Ensemble Empirical Mode Decomposition-Long Short-Term Memory Network) method is used to achieve univariate prediction, so as to accurately predict the factors in the medium and long term.

[0012] S5: Based on the prediction results of influencing factors obtained in S4, an XGBoost (PSO-XGBoost) network optimized by particle swarm optimization algorithm is used to construct a medium- and long-term prediction model for urban carbon emissions.

[0013] Furthermore, in S1, based on the "economy-energy-carbon emission" balance system, the factors influencing urban carbon emissions are selected, and historical data of these factors are obtained. The selected factors are as follows:

[0014] We establish the Kaya identity, which is mainly used for the decomposition of carbon emission influencing factors. Based on carbon emissions (C), total energy consumption (E), gross domestic product (GDP), and total population (P), we decompose the influencing factors of carbon emissions, expressed as:

[0015]

[0016]

[0017]

[0018] The influencing factors are defined as follows:

[0019]

[0020]

[0021] Furthermore, S2 performs data preprocessing on the dataset obtained from S1, including bad data identification, data cleaning, and data completion, to obtain a high-quality dataset.

[0022] The main task was to complete the missing monthly data on industrial product output and remove economic data with obvious anomalies.

[0023] Data cleaning was performed using LOF and GAKNNI methods;

[0024] Furthermore, S3 employs the Generalized Dichotomy Index Decomposition (GDIM) method to effectively analyze the mechanisms and contributions of factors influencing urban carbon emissions, identifying key factors affecting urban carbon emissions, specifically:

[0025] C = X3X4

[0026] X3X4-X5X6=0

[0027] X3X4-X1X2=0

[0028] X3-X1X8=0

[0029] X1 - X5X7 = 0

[0030] The contribution of factor X (X = X1, X2, X3, X4, X5, X6, X7, X8) to the change in carbon emissions is represented by the function C(X).

[0031] X1 is the region's real GDP, X2 is carbon emissions per unit of GDP (C / GDP), X3 is energy consumption (E), X4 is carbon emissions per unit of energy consumption (C / E), X5 is the resident population (P), X6 is carbon emissions per capita (C / P), X7 is real GDP per capita, and X8 is energy consumption per unit of GDP (E / GDP).

[0032] Construct the Jacobian matrix based on GDIM:

[0033]

[0034] According to the GDIM principle, the change in carbon emissions ΔC can be decomposed into the sum of the contributions of various factors:

[0035]

[0036] In the formula, L represents the time span; I represents the identity matrix; "+" represents the generalized inverse matrix; ▽C=(0 0 X4 X3 0 00 0) T Jacobi matrix If the column vectors are linearly independent, then

[0037]

[0038] The change in carbon emissions can be decomposed into the sum of eight effects: ΔX1, ΔX2, ΔX3, ΔX4, ΔX5, ΔX6, ΔX7, and ΔX8. ΔX1, ΔX3, and ΔX5 are absolute values, reflecting the impact of economic development level, energy consumption, and population size changes on carbon emissions from energy consumption, respectively. The remaining five effects are relative factors, representing, in order, the carbon emissions per unit of GDP, the carbon emissions per unit of energy consumed, per capita carbon emissions, per capita GDP, and the energy consumed per unit of GDP on the change in carbon emissions.

[0039] Furthermore, the EEMD-LSTM method used in S4 is specifically as follows:

[0040] This invention utilizes the EEMD method to obtain the overall trend characteristics, seasonal characteristics, and random fluctuation characteristics of various influencing factors. EEMD is used to decompose the data, and Long Short-Term Memory (LSTM) networks are built for each decomposed data point to perform predictions. The prediction results are then superimposed to obtain the final predicted value.

[0041] Using the EEMD method, different types of white noise need to be added to the initial time series to transform the electricity consumption time series data x k,t It is decomposed into a series of intrinsic mode functions (IMF) components.

[0042]

[0043] In the formula, imf k,i This represents the i-th intrinsic mode component obtained by decomposing the electricity consumption time-series data of the k-th user class using EEMD; r k,n It is the residual component after decomposing the original time series data.

[0044] EEMD decomposes time series data by connecting local maxima and minima with cubic spline curves to form upper and lower envelopes, and then taking the mean m. j-1 Then from the initial sequence x k,t Subtracting the average of these envelopes from the mean yields the remaining sequence r. k,i Therefore, when calculating each IMF process, we have:

[0045] r k,i =r k,i-1 -imf k,i

[0046] In the formula, when i = 1, r k,0=x k,t .

[0047] Iterate continuously on the remaining sequence r k,i Extract the (i+1)th IMF until the remaining sequence r k,n Since further decomposition is not possible, the mean of the IMF obtained by adding different amounts of white noise is finally calculated. Finally, the overall trend characteristics, seasonal characteristics, and random fluctuation characteristics of each influencing factor are obtained.

[0048] The specific training process of the LSTM model is mainly completed by three gate structures: the input gate, the forget gate, and the output gate. The input gate updates information by combining the current input content and the output of the previous time step, and selectively stores the updated part into the cell state c. t The forgetting gate determines the degree to which the cell state is forgotten or updated, identifying the forgotten information. The output gate calculates the output content at that moment based on the updated cell state, i.e., the user's response load adjustment. The specific process is shown in the attached diagram.

[0049] Furthermore, S5 employs an XGBoost (PSO-XGBoost) network optimized based on particle swarm optimization to construct a model for analyzing and predicting the influencing factors of urban carbon emissions in the medium and long term. The PSO-XGBoost algorithm is specifically described as follows:

[0050] The XGBoost algorithm improves its generalization ability by controlling the overall model complexity through a second-order Taylor expansion of the loss function and the introduction of a regularization term to achieve the global optimum. Furthermore, to prevent overfitting, the algorithm employs parallel processing of feature selection, resulting in faster execution and more interpretable results.

[0051] Suppose we have a dataset D = {(x i ,y i )∣x i ∈R m ,y i The set ∈R, i=1,2,…,n} consists of m features and a total of n samples, where Rm and R represent the m-dimensional real vector dataset and the set of real numbers, respectively.

[0052]

[0053] In the formula, fk is a regression tree, K is the total number of regression trees, and F is the regression tree space.

[0054] The objective function is expressed as:

[0055]

[0056] In the formula: l represents the loss function, which measures the error between the classification prediction value and the true value; For classification prediction; y i The true value; Ω(f) k ) is a regular term.

[0057] The XGBoost algorithm uses gradient boosting iterative computation. In each iteration, a new regression tree is added. Therefore, the result of the t-th iteration is:

[0058]

[0059] Combining the formulas, we can calculate the objective function expression for the t-th iteration as follows:

[0060]

[0061] Perform a second-order Taylor expansion and add a regularization term Ω(f) k To prevent overfitting.

[0062]

[0063] In the formula: γ represents the leaf tree penalty coefficient; T is the number of leaf nodes; ω is the leaf weight; λ is the weight penalty coefficient.

[0064] Particle Swarm Optimization (PSO) is a swarm intelligence algorithm based on mimicking the foraging behavior of bird flocks. PSO defines the solution to the optimization problem as searching for particles in a finite-dimensional space. Each particle consists of a position vector and a velocity vector. All particles cooperate to choose the best position, searching for a better position by combining their own optimal value with the optimal value of the swarm. Each particle calculates its fitness value using a fitness function to measure the quality of its position, while all particles in the swarm follow the currently best particle in the search for its position in the solution space.

[0065] The method for constructing a medium- and long-term prediction model for urban carbon emissions using an XGBoost network optimized by particle swarm optimization includes the following steps: model initialization; training the XGBoost model; calculating the fitness value of particles; calculating the individual optimal value and the group optimal value; iteratively updating the optimal solution; if the conditions are met, assigning the optimal parameters to the XGBoost model and retraining the XGBoost model with the optimal parameters; performing carbon emission prediction; if the conditions are not met, re-initializing the model.

[0066] Based on the principles of XGBoost and the theory of PSO algorithm, PSO is applied to the parameter optimization of XGBoost regression prediction, and a PSO-XGBoost medium- and long-term carbon emission regression prediction model is constructed. The process is as follows:

[0067] (1) Data cleaning, data repair, and standardization preprocessing are performed on carbon emission data and driving factor data obtained from statistical data:

[0068] (2) Perform GDIM analysis on the preprocessed carbon emission data, select important driving factors according to their contribution rates, and construct a dataset based on the analysis results:

[0069] (3) Divide the training set and the test set into an 8:2 ratio, set an appropriate fitness function and initialize the individual particle optimal value and the global optimal value, and optimize parameters such as learning rate and max_depth:

[0070] (4) The Dardai update particle velocity and position, and continuously update the individual optimal value and the global optimal value by calculating its fitness value until the generation termination condition is met:

[0071] (5) Select the optimal parameter values, construct the parameter-optimized XGBoost regression prediction model, and import the training set for model training and learning.

[0072] This invention also provides a device for extracting the characteristics of urban carbon emission influencing factors and predicting medium- and long-term trends, comprising:

[0073] The selection module, based on the economic-energy-carbon emission balance system, identifies the factors influencing urban carbon emissions and obtains historical data on these factors.

[0074] The data processing module is used for data preprocessing to obtain high-quality datasets;

[0075] The analysis module uses the generalized Dijkstra index decomposition method to effectively analyze the mechanism and contribution of factors affecting urban carbon emissions, and to identify the key factors affecting urban carbon emissions.

[0076] The prediction module uses an integrated empirical mode decomposition-long short-term memory network method to predict univariate variables; and employs an XGBoost network optimized based on particle swarm optimization to construct a medium- and long-term prediction model for urban carbon emissions.

[0077] The beneficial effects of the technical solution provided by this invention are:

[0078] (1) Compared with the prior art, the present invention takes into account factors such as energy consumption and economic growth, and considers the interaction and influence of multiple factors, thereby achieving accurate carbon emission prediction.

[0079] (2) Compared with the regression prediction methods commonly used in the prior art, this invention analyzes carbon emission trends and makes medium- and long-term predictions of carbon emissions using the PSO-improved XGBoost algorithm. The feasibility and accuracy of the proposed method are verified through actual research data.

[0080] Therefore, the method for extracting features of urban carbon emission influencing factors and predicting medium- and long-term emissions proposed in this invention provides a systematic approach to analyze important carbon emission drivers, achieve medium- and long-term carbon emission prediction, and formulate effective strategies to reduce these emissions. Attached Figure Description

[0081] Figure 1 This is a flowchart of the method of the present invention;

[0082] Figure 2 Here is a flowchart of the EEMD algorithm proposed in this invention;

[0083] Figure 3 This is a diagram illustrating the training and prediction process of an LSTM network.

[0084] Figure 4 Flowchart for PSO-XGBoost carbon emission prediction;

[0085] Figure 5 An example diagram is shown below, calculated using actual data from Tianjin City as an example, based on the method of this invention. Detailed Implementation

[0086] The present invention will be further described in detail below through specific embodiments. The following embodiments are merely descriptive and not limiting, and should not be used to limit the scope of protection of the present invention.

[0087] Example

[0088] A method for extracting characteristics of factors influencing urban carbon emissions and for medium- to long-term prediction, comprising the following steps:

[0089] 1: Analysis of carbon emission influencing factors based on GDMI;

[0090] To effectively analyze the mechanisms and contributions of various driving factors influencing carbon emissions, the exponential decomposition method was used to quantitatively measure the key factors affecting carbon emissions and analyze their degree of influence. The commonly used LMDI method only considers one absolute factor, and there are significant interdependencies among the factors, leading to insufficient accuracy. Therefore, the GDIM method was adopted to analyze the driving factors. The specific procedures are as follows:

[0091] 101: Establish the Kaya identity

[0092] We establish the Kaya identity, which is mainly used for the decomposition of carbon emission influencing factors. Based on carbon emissions (C), total energy consumption (E), gross domestic product (GDP), and total population (P), we decompose the influencing factors of carbon emissions, expressed as:

[0093]

[0094]

[0095]

[0096] 102: The influencing factors are defined as follows:

[0097] C Carbon emissions from energy consumption 10,000 tons X1 = GDP Regional Gross Domestic Product (Economic Development Level) 100 million yuan X2 = C / GDP Carbon emissions per unit of GDP produced (output carbon intensity) ton / 10,000 yuan X3 = E Energy consumption 10,000 tons of standard coal X4 = C / E Carbon emissions per unit of energy consumption (energy carbon intensity) tons / ton of standard coal X5 = P Resident population (population size) 10,000 people X6 = C / P per capita carbon emissions tons / person X7 = GDP / P Real GDP per capita Ten thousand yuan / person

[0098] 2: Data cleaning;

[0099] 201: Outlier Detection

[0100] The LOF algorithm is used to detect anomalies in the feature data to be tested. The outlier factor value of each data point in the dataset is calculated to determine the set of outliers, and it is then determined whether all the data to be tested has been detected.

[0101] Based on historical feature datasets, the GAKNNI algorithm is used to complete outlier and missing data, and it is determined whether all data to be completed is complete. The data completion process involves using GAKNNI to calculate the K nearest data points of the input sample to the training dataset in space, and then calculating the similarity or dissimilarity of the development trends between the data.

[0102] The method for using the LOF algorithm to detect anomalies in the feature data to be detected, and determining the set of outliers by calculating the outlier factor value of each data point in the dataset, is as follows:

[0103] First, analyze the average density of data points in the dataset. Based on the density distribution, determine the number of outliers M1 and the outlier set D1. Then, calculate the outlier factor to determine M2 (M2 = M1) outliers and the outlier set D2. Take the intersection of D1 and D2 as the final outlier group.

[0104] The method for determining outliers in a dataset based on density distribution is as follows:

[0105] The k-distance of a data object q is defined as the distance from the k-th point in the data set that is closest to the data object q to q, denoted as k-distance(q). Here, the distance refers to the Euclidean distance, i.e., the straight-line distance.

[0106] Calculate the data points in the dataset whose distance to the data object q is no greater than k.

[0107] Calculate p, where p and q are any two points in the dataset;

[0108] The local reachability density of q is calculated as the average local reachability density of q to all points in its neighborhood.

[0109] Calculate the outlier degree of the data;

[0110] Outliers in the data are identified and removed based on their degree of outlier.

[0111] 202: Data Completion

[0112] The method for completing outlier and missing data based on historical feature datasets and using the GAKNNI algorithm is as follows:

[0113] KNN is used to find the K data points that are spatially closest to the missing point in the training dataset. Then, based on the similarity weights of the K data points, the completion value of the missing point is the sum of the products of the K data points and their similarity weights.

[0114] The method for obtaining the padding value of missing points is as follows:

[0115] Establish time-series characteristic data of carbon emissions and their influencing factors, and obtain a sample matrix T without missing terms. train Missing item sample time series t lack ;

[0116] Calculate the sample matrix T without missing items train Time series of missing items t lack The Euclidean distance parameter d;

[0117] From T train Screening and t lack K nearest neighbor similar time series

[0118] Calculate t lack and Grey relational coefficients of K similar time series in the dataset;

[0119] calculate The similarity weight ω of K similarity times k ;

[0120] Calculate the completion value y for the missing points.

[0121] 3: Selection of key influencing factors

[0122] 301: Constructing the Jacobian Matrix

[0123] C = X3X4

[0124] X3X4-X5X6=0

[0125] X3X4-X1X2=0

[0126] X3-X1X8=0

[0127] X1 - X5X7 = 0

[0128] The contribution of factor X (X = X1, X2, X3, X4, X5, X6, X7, X8) to the change in carbon emissions is a function C(X). Based on the GDIM, the Jacobian matrix is ​​constructed as follows:

[0129]

[0130] 302: Contribution Breakdown

[0131] According to the GDIM principle, the change in carbon emissions ΔC can be decomposed into the sum of the contributions of various factors:

[0132]

[0133] In the formula, L represents the time span; I represents the identity matrix; "+" represents the generalized inverse matrix; ▽C=(0 0 X4 X3 0 00 0) T Jacobi matrix If the column vectors are linearly independent, then

[0134]

[0135] The change in carbon emissions can be decomposed into the sum of eight effects: ΔX1, ΔX2, ΔX3, ΔX4, ΔX5, ΔX6, ΔX7, and ΔX8. ΔX1, ΔX3, and ΔX5 are absolute values, reflecting the impact of economic development level, energy consumption, and population size changes on carbon emissions from energy consumption, respectively. The remaining five effects are relative factors, representing, in order, the carbon emissions per unit of GDP, the carbon emissions per unit of energy consumed, per capita carbon emissions, per capita GDP, and the energy consumed per unit of GDP on the change in carbon emissions.

[0136] 4: Prediction of Influencing Factors

[0137] 401: Temporal Feature Extraction

[0138] Add white noise

[0139] Using the EEMD method, different types of white noise need to be added to the initial time series to transform the electricity consumption time series data x k,t It is decomposed into a series of intrinsic mode functions (IMF) components.

[0140]

[0141] In the formula, imf k,i This represents the i-th intrinsic mode component obtained by decomposing the electricity consumption time-series data of the k-th user class using EEMD; r k,n It is the residual component after decomposing the original time series data.

[0142] Find the components

[0143] EEMD decomposes time series data by connecting local maxima and minima with cubic spline curves to form upper and lower envelopes, and then taking the mean m. j-1 Then from the initial sequence x k,t Subtracting the average of these envelopes from the mean yields the remaining sequence r. k,i Therefore, when calculating each IMF process, we have:

[0144] r k,i =r k,i-1 -imf k,i

[0145] In the formula, when i = 1, r k,0 =x k,t .

[0146] Iterate continuously on the remaining sequence r k,i Extract the (i+1)th IMF until the remaining sequence r k,n Since further decomposition is not possible, the mean of the IMF obtained by adding different white noises is finally calculated. This yields the regularity component, trend component, and random component.

[0147] 402: LSTM Network Prediction

[0148] The model training uses the LSTM algorithm for multivariate time series prediction.

[0149] F i =LSTM θ (P regular,m ,P trend,m ,P random,m )

[0150] In the formula, F i P represents the predicted value of the influencing factors. regular,m ,P trend,m ,P random,m These represent the regularity components, trend components, and random components of different influencing factors.

[0151] 5: Medium- and long-term carbon emission prediction model based on PSO-XGBOOST method

[0152] 501: PSO-XGBOOST Model Construction

[0153] The XGBoost algorithm improves its generalization ability by controlling the overall model complexity through a second-order Taylor expansion of the loss function and the introduction of a regularization term to achieve the global optimum. Furthermore, to prevent overfitting, the algorithm employs parallel processing of feature selection, resulting in faster execution and more interpretable results.

[0154] Suppose we have a dataset D = {(x i ,y i )∣x i ∈R m ,y i The set ∈R, i=1,2,…,n} consists of m features and a total of n samples, where Rm and R represent the m-dimensional real vector dataset and the set of real numbers, respectively.

[0155]

[0156] In the formula, fk is a regression tree, K is the total number of regression trees, and F is the regression tree space.

[0157] The objective function is expressed as:

[0158]

[0159] In the formula: l represents the loss function, which measures the error between the classification prediction value and the true value; For classification prediction; y i The true value; Ω(f) k ) is a regular term.

[0160] The XGBoost algorithm uses gradient boosting iterative computation. In each iteration, a new regression tree is added. Therefore, the result of the t-th iteration is:

[0161]

[0162] Combining the formulas, we can calculate the objective function expression for the t-th iteration as follows:

[0163]

[0164] Perform a second-order Taylor expansion and add a regularization term Ω(f) k To prevent overfitting.

[0165]

[0166] In the formula: γ represents the leaf tree penalty coefficient; T is the number of leaf nodes; ω is the leaf weight; λ is the weight penalty coefficient.

[0167] Based on the XGBoost principle and PSO algorithm theory, PSO is applied to the parameter optimization of XGBoost regression prediction, and a PSO-XGBoost long-term regression prediction model for carbon emissions is constructed.

[0168] 502: Carbon Emission Forecasting

[0169] (1) Data cleaning, data repair, and standardization preprocessing are performed on carbon emission data and driving factor data obtained from statistical data:

[0170] (2) Perform GDIM analysis on the preprocessed carbon emission data, select important driving factors according to their contribution rates, and construct a dataset based on the analysis results:

[0171] (3) Divide the training set and the test set into an 8:2 ratio, set an appropriate fitness function and initialize the individual particle optimal value and the global optimal value, and optimize parameters such as learning rate and max_depth:

[0172] (4) The Dardai update particle velocity and position, and continuously update the individual optimal value and the global optimal value by calculating its fitness value until the generation termination condition is met:

[0173] (5) Select the optimal parameter values, construct the parameter-optimized XGBoost regression prediction model, and import the training set for model training and learning.

[0174] The overall process is as follows Figure 1 As shown.

[0175] This invention further provides a device for extracting the characteristics of urban carbon emission influencing factors and predicting medium- and long-term trends, comprising:

[0176] The selection module, based on the economic-energy-carbon emission balance system, identifies the factors influencing urban carbon emissions and obtains historical data on these factors.

[0177] The data processing module is used for data preprocessing to obtain high-quality datasets;

[0178] The analysis module uses the generalized Dijkstra index decomposition method to effectively analyze the mechanism and contribution of factors affecting urban carbon emissions, and to identify the key factors affecting urban carbon emissions.

[0179] The prediction module uses an integrated empirical mode decomposition-long short-term memory network method to predict univariate variables; and employs an XGBoost network optimized based on particle swarm optimization to construct a medium- and long-term prediction model for urban carbon emissions.

[0180] Based on actual data from Tianjin, simulation analysis was conducted on the above embodiments, and the beneficial effects obtained by the present invention are as follows: Figure 5 As shown, the model proposed in this invention was used to make long-term predictions of carbon emissions in Tianjin. The error rate between the test set and the actual values ​​remained at approximately 3%, and the predictive ability remained good even when carbon emission data fluctuated drastically. It can be considered that the predicted values ​​obtained using the method described in this paper are sufficiently reasonable, practical, and informative.

[0181] In summary, the embodiments of the present invention provide a method for extracting features of factors influencing urban carbon emissions and for medium- to long-term prediction. The present invention makes full use of historical information to accurately and reliably predict the medium- to long-term changes in carbon emissions.

[0182] Unless otherwise specified, the model numbers of the various devices in this embodiment of the invention are not limited, and any device that can perform the above functions is acceptable.

[0183] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0184] 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 within the protection scope of the present invention.

Claims

1. A method for extracting the characteristics of factors influencing urban carbon emissions and for medium- to long-term prediction, characterized in that, The method includes the following steps: S1: Based on the economic-energy-carbon emission balance system, select the factors influencing urban carbon emissions and obtain historical data on these factors; S2: Perform data preprocessing on the data obtained from S1 to obtain a high-quality dataset; S3: Based on the high-quality dataset of S2, the generalized Dichroic index decomposition method is used to effectively analyze the mechanism and contribution of factors affecting urban carbon emissions, and to identify the key factors affecting urban carbon emissions. S4: Based on S3, key influencing factors are obtained, and the integrated empirical mode decomposition-long short-term memory network method is used to predict univariate variables. S5: Based on the influencing factor prediction results obtained in S4, an XGBoost network optimized by particle swarm optimization is used to construct a medium- and long-term prediction model for urban carbon emissions.

2. The method according to claim 1, characterized in that: The factors influencing carbon emissions in selected cities in S1 include: carbon emissions from energy consumption (C), regional real GDP, carbon emissions per unit of GDP (C / GDP), energy consumption (E), carbon emissions per unit of energy consumption (C / E), resident population (P), per capita carbon emissions (C / P), per capita real GDP, and energy consumption per unit of GDP (E / GDP).

3. The method according to claim 2, characterized in that: Data preprocessing in S2 includes: using the LOF algorithm to detect anomalies in the feature data to be detected, and using the GAKNNI algorithm to complete the abnormal and missing data.

4. The method according to claim 3, characterized in that: The generalized Dichotomy index decomposition method in S3 includes a Jacobian matrix construction step and a contribution decomposition step. The Jacobian matrix is: C = X3X4 X3X4-X5X6=0 X3X4-X1X2=0 X3-X1X8=0 X1 - X5X7 = 0 Among them, the contribution of factor X (X=X1,X2,X3,X4,X5,X6,X7,X8) to the change in carbon emissions is a function C(X); The contribution decomposition involves breaking down the change in carbon emissions ΔC into a summation of the contributions from various factors: In the formula, L represents the time span; I represents the identity matrix; and "+" represents the generalized inverse matrix. Jacobian matrix If the column vectors are linearly independent, then X1 is the region's actual GDP, X2 is the carbon emissions per unit of GDP C / GDP, X3 is energy consumption E, X4 is the carbon emissions per unit of energy consumption C / E, X5 is the resident population P, X6 is the carbon emissions per capita C / P, X7 is the actual GDP per capita, and X8 is the energy consumption per unit of GDP E / GDP. Changes in carbon emissions are decomposed into the sum of eight effects: ΔX1, ΔX2, ΔX3, ΔX4, ΔX5, ΔX6, ΔX7, and ΔX8. Among them, ΔX1, ΔX3, and ΔX5 are absolute quantities, reflecting the impact of economic development level, energy consumption, and population size changes on carbon emissions from energy consumption, respectively. The remaining five effects are relative factors, which, in order, represent carbon emissions per unit of GDP (C / GDP), carbon emissions per unit of energy consumption (C / E), carbon emissions per capita (C / P), real GDP per capita, and energy consumption per unit of GDP (E / GDP).

5. The method according to claim 4, characterized in that: The integrated empirical mode decomposition-long short-term memory network method used in S4 is as follows: The ensemble empirical mode decomposition method is used to obtain the overall trend characteristics, seasonal characteristics and random fluctuation characteristics of each influencing factor. The data is decomposed using ensemble empirical mode decomposition, and long short-term memory networks are built for prediction of the decomposed data. The prediction results are then superimposed to obtain the final prediction value.

6. The method according to claim 5, characterized in that: The integrated empirical mode decomposition method is as follows: different white noises are added to the initial time series, and the power consumption time series data x is processed. k,t It is decomposed into a series of intrinsic mode components. In the formula, imf k,i This represents the i-th intrinsic mode component obtained by applying integrated empirical mode decomposition to the electricity consumption time-series data of the k-th user class; r k,n It is the residual component after decomposing the original time series data; Decomposing time series data involves connecting local maxima and minima with cubic spline curves to form upper and lower envelopes, and then taking their mean m. j-1 Then from the initial sequence x k,t Subtracting the average of these envelopes from the mean yields the remaining sequence r. k,i Therefore, when calculating each IMF process, we have: r k,i =r k,i-1 -imf k,i In the formula, when i = 1, r k,0 =x k,t ; Iterate continuously on the remaining sequence r k,i Extract the (i+1)th IMF until the remaining sequence r k,n Since it cannot be further decomposed, the mean of the IMF obtained by adding different white noise is finally calculated, and the overall trend characteristics, seasonal characteristics and random fluctuation characteristics of each influencing factor are finally obtained.

7. The method according to claim 6, characterized in that: S5 employs an XGBoost network optimized using the particle swarm optimization algorithm to construct a medium- to long-term prediction model for urban carbon emissions. The method includes the following steps: Model initialization; Training the XGBoost model; Calculate the particle's fitness value; Calculate the individual optimal value and the group optimal value; Iteratively update the optimal solution; If the conditions are met, the optimal parameters are assigned to the XGBoost model, and the XGBoost model is retrained with the optimal parameters to predict carbon emissions; if the conditions are not met, the model is re-initialized.

8. A device for extracting the characteristics of factors influencing urban carbon emissions and for medium- to long-term prediction, characterized in that, include: The selection module, based on the economic-energy-carbon emission balance system, identifies the factors influencing urban carbon emissions and obtains historical data on these factors. The data processing module is used for data preprocessing to obtain high-quality datasets; The analysis module uses the generalized Dijkstra index decomposition method to effectively analyze the mechanism and contribution of factors affecting urban carbon emissions, and to identify the key factors affecting urban carbon emissions. The prediction module uses an integrated empirical mode decomposition-long short-term memory network method to predict univariate variables; and employs an XGBoost network optimized based on particle swarm optimization to construct a medium- and long-term prediction model for urban carbon emissions.

Images