ISMA-SVR regional carbon emission prediction method based on multi-layer influence factor screening

Abstract

The invention discloses an ISMA-SVR regional carbon emission prediction method based on multi-layer influence factor screening, and the method comprises the steps: firstly accounting the carbon emission of a target region, and selecting potential influence factors related to the carbon emission of the target region for preprocessing; secondly, screening the pre-processed potential influence factors based on grey correlation analysis, and finally screening by adopting an extended STIRPAT model in combination with a Lasso regression algorithm to obtain a key feature set; and then combining a random inertia weight and a Levy flight strategy with a mucus algorithm, applying the combination to hyper-parameter optimization of a support vector regression model, and constructing a regional carbon emission prediction model. And finally, taking the key feature set obtained by final screening as input, and substituting the key feature set into the model for regional carbon emission prediction to obtain a regional carbon emission prediction result. The method is suitable for small sample scene prediction, overcomes the limitation of a single machine learning prediction model, and improves the precision of carbon emission prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of carbon emission prediction, in particular to an ISMA-SVR regional carbon emission prediction method based on multi-layer influence factor screening. BACKGROUND

[0002] Global warming has become a major global challenge to human sustainable development. The series of environmental problems caused by global warming, such as extreme weather and ecological degradation, have attracted widespread attention and high attention from the international community. As the core carrier of economic activities, regions are the main source of carbon emissions. Therefore, the low-carbon transformation and emission reduction control at the regional level can accurately predict the trend of regional carbon emissions, which can provide core decision-making basis for identifying emission reduction pressure, developing scientific emission reduction strategies, and promoting low-carbon development.

[0003] However, the existing carbon emission prediction technology still has many technical defects to be solved, which makes it difficult to meet the actual application requirements in terms of prediction accuracy and stability. Specifically, the following aspects are involved:

[0004] 1. The existing carbon emission influence factor analysis technology is mostly limited to a single dimension, lacking a systematic multi-dimensional screening system. On the one hand, it is easy to introduce redundant features with weak correlation with carbon emissions, and on the other hand, it may miss key dimensions of economic, energy, and industrial influence factors, resulting in incomplete coverage of influence factors. The above problems directly cause the lack of effectiveness of the input features of the prediction model, and further limit the accuracy of the prediction results.

[0005] 2. The carbon emission prediction data generally has the inherent characteristic of "small sample size", but the existing prediction technology mostly relies on traditional prediction models or single prediction algorithms, without targeted adaptation and optimization for this characteristic. It is difficult to fully tap the potential laws in small sample data, resulting in a significant reduction in prediction accuracy and reliability of the model in small sample scenarios, and unable to meet the needs of fine control of regional carbon emissions.

[0006] 3. Support Vector Regression (SVR) is widely used in the field of carbon emission prediction due to its strong generalization ability and adaptation to small sample data. However, its prediction accuracy is highly dependent on the selection of model hyperparameters: penalty parameter C, insensitive loss function parameter ε, and kernel function parameter γ. In existing technology, the hyperparameters of SVR mostly rely on manual experience setting, which is prone to cause overfitting or underfitting problems of the model, resulting in large prediction errors. Even if intelligent algorithms (such as Genetic Algorithm (GA) and traditional Slime Mould Algorithm (SMA)) are used for hyperparameter optimization, there are inherent defects: GA has low search efficiency and long iteration time; traditional SMA is prone to local optimization in the later iteration period, and cannot achieve global optimization of hyperparameters, ultimately leading to the prediction effect of SVR model still unable to meet the actual needs of high-precision control of carbon emissions.

[0007] In summary, the existing carbon emission prediction model has obvious deficiencies in the comprehensiveness of the influencing factor screening, the adaptability of the small sample scene and the effectiveness of the model hyperparameter optimization, which makes it difficult to improve the prediction accuracy. Therefore, it is urgent to develop a prediction model integrating multi-level influencing factor screening mechanism and improved optimization algorithm to solve the defects of the existing technology and improve the accuracy and stability of carbon emission prediction, providing reliable support for regional low-carbon development decision-making. SUMMARY

[0008] To solve the problems in the prior art, the present application provides an ISMA-SVR regional carbon emission prediction method based on multi-level influencing factor screening. The method realizes the accurate extraction of key influencing factors through a multi-level feature screening mechanism of "multi-dimensional preliminary selection-gray correlation analysis screening-STIRPAT combined with Lasso regression final screening"; at the same time, an improved slime optimization algorithm ISMA is proposed by combining random inertia weight and Levy flight strategy, which optimally configures the global parameters of the SVR model, and finally constructs a high-precision and high-stability regional carbon emission prediction model, providing reliable technical support for regional low-carbon development decision-making.

[0009] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0010] The ISMA-SVR regional carbon emission prediction method based on multi-level influencing factor screening provided by the present application comprises the following steps:

[0011] Step 1: calculate the carbon emission of the target region, construct the carbon emission basic dataset, and select the potential influencing factors related to the carbon emission of the target region for preprocessing;

[0012] Step 2: screen the preprocessed potential influencing factors based on gray correlation analysis to obtain an influencing factor subset;

[0013] Step 3: use the extended STIRPAT model combined with Lasso regression algorithm to perform final screening on the influencing factor subset to obtain a key feature set;

[0014] Step 4: combine random inertia weight and Levy flight strategy with traditional slime algorithm SMA to construct an improved slime optimization algorithm ISMA;

[0015] Step 5: apply the constructed ISMA algorithm to the hyperparameter optimization of support vector regression SVR model to construct an ISMA-SVR regional carbon emission prediction model;

[0016] Step 6: input the key feature set obtained by final screening into the ISMA-SVR model for regional carbon emission prediction, and obtain the regional carbon emission prediction result.

[0017] Furthermore, in step 1, the carbon emissions of the target area are calculated based on the carbon emission factor method, covering major carbon emission sources such as energy consumption, industrial production, and transportation. The carbon emission factor method, due to its ease of operation, economy, and wide applicability, has become the most widely used carbon emission accounting method globally. The formula for the carbon emission factor method is as follows:

[0018]

[0019] Where m represents the total energy, c represents the energy category, CE represents the total carbon emissions of the target area, and AD c EF represents the activity level data for the c-th energy source. c This represents the carbon emission factor of the c-th energy source.

[0020] The carbon emission factors are all derived from official documents such as the "Guidelines for the Compilation of Provincial Greenhouse Gas Inventories" and the "2006 IPCC Guidelines for National Greenhouse Gas Inventories".

[0021] Furthermore, in step 1, q potential influencing factors related to carbon emissions are selected from three dimensions: population size, economic level, and technological level.

[0022] Preprocessing of the original data of the selected potential influencing factors: Given that the proportion of missing data is less than 5%, and that linear interpolation has the characteristics of high computational efficiency and can maintain the local trend and continuity of data, linear interpolation is used to complete the missing values ​​to obtain the dataset of influencing factors, providing a data foundation for subsequent feature selection.

[0023] Furthermore, in step 2, the grey relational analysis method is used to calculate the grey relational degree between the selected potential influencing factors and regional carbon emissions. After setting a correlation degree threshold, redundant potential influencing factors with grey relational degrees below the threshold are eliminated to obtain a subset of influencing factors with strong correlation to carbon emissions. This method is chosen for feature screening because it requires less sample size, does not require strict data distribution assumptions, and has stronger interpretability and stability.

[0024] The gray relational degree calculation process is as follows:

[0025] The comparison sequence X consists of q potential influencing factors. i =(x i (1),x i (2),...,x i (n)), i=1,2,...,q;Regional carbon emissions are used as a reference sequence X0=(x0(1),x0(2),...,x0(n)); where n is the length of each sequence.

[0026] The reference and comparison sequences are dimensionless using a mean-based method. The formula for the mean-based method is as follows:

[0027]

[0028] Where k is the sequence index, k = 1, 2, ..., n, x i '(k) represents the dimensionless value of the i-th sequence at time k, x i (k) represents the original data of the i-th sequence at time k.

[0029] Calculate the correlation coefficient ξ i (k), the formula is as follows:

[0030]

[0031] Where, ξ i (k) represents the correlation coefficient between the i-th comparison sequence and the reference sequence at time k, and x0'(k) and x i '(k) represents the reference sequence and the comparison sequence at time k, respectively, min i min k |x0'(k)-x i '(k)| represents the minimum difference between two levels, max i max k |x0'(k)-x i '(k)| represents the maximum difference between the two levels, and ρ represents the resolution coefficient.

[0032] The final grey relational degree r is obtained by averaging all correlation coefficients for each comparison sequence. i .

[0033] Furthermore, in step 3, an extended STIRPAT model, incorporating newly added influencing factors, is used to characterize the relationship between potential influencing factors and regional carbon emissions. Due to its flexible and extensible structure, the extended STIRPAT model can systematically incorporate multi-dimensional influencing factors such as socioeconomic, technological, and policy factors, thereby effectively quantifying the differentiated impact of different variables on carbon emissions. The extended STIRPAT model formula is as follows:

[0034]

[0035] Where a represents a constant term, β1, β2...β q Let represent the parameter to be estimated, e represent the error term, I represent the regional carbon emissions, and P, A...U represent q potential influencing factors related to carbon emissions.

[0036] Taking the natural logarithm of both sides of the above equation, we can transform it into a linear form that can be solved by regression:

[0037] lnI=lna+β1lnP+β2lnA+...+β q lnU+lne

[0038] Where lnI is the dependent variable, lnP, lnA...lnU are the independent variables of the model, lna is the constant term, and lne is the error term.

[0039] Compared to ridge regression, Lasso regression, when dealing with multicollinearity, not only stabilizes parameter estimates but also compresses the coefficients of unimportant or redundant variables to zero through sparsity features. This allows the model to effectively identify and eliminate highly correlated or weakly contributing factors while retaining key explanatory variables, thereby improving the interpretability of variable selection and model efficiency. The objective function of Lasso regression is:

[0040]

[0041] in, This represents the coefficient estimation results obtained through the Lasso regression method, where n is the sample size and y is the coefficient value. i Let X be the dependent variable value of the i-th sample. i Let be the vector of independent variables for the i-th sample, and β be the vector of coefficients representing all β values ​​to be estimated. j The set, β j Let λ represent the coefficient corresponding to the j-th independent variable, and λ be the regularization coefficient.

[0042] Furthermore, in step 4, a stochastic inertial weight and a Lévy flight strategy are introduced to improve the traditional SMA algorithm and reconstruct the slime mold position update formula. The stochastic inertial weight dynamically adjusts the search step size of the traditional slime mold algorithm, breaking the limitation of the imbalance between exploration and development capabilities under fixed weights. The Lévy flight strategy, by introducing random walk characteristics, improves the slime mold position update mechanism and effectively compensates for the defect of the traditional algorithm being prone to getting trapped in local optima. Experiments are conducted using multiple benchmark test functions (such as Sphere, Rastrigin, Ackley, etc.) to compare and evaluate the optimization performance of ISMA and traditional SMA from two dimensions: convergence speed and optimization accuracy.

[0043] The optimization capability of the slime mold algorithm is based on a mathematical model of slime mold propagation waves, which can be divided into three stages: approaching food, surrounding food, and acquiring food. The approaching and surrounding food stages are distinguished by the parameters rand and r. The original slime mold position update formula is:

[0044]

[0045] p=tanh|S(h)-DF|, (h=1,2,...,N)

[0046] Where rand < z represents the approaching food stage, and r < p and r ≥ p represent the surrounding food stage. t represents the current iteration number, and X(t) represents the position of the slime mold at time t. b (t) represents the optimal position of the slime mold at time t, X A (t) and X B (t) represents two randomly selected positions from the slime mold at time t, N represents the total number of individuals in the slime mold population, r and rand are random values ​​between [0,1], ub and lb represent the upper and lower bounds of the search range, z is the probability that the isolated slime mold individual searches for other food sources, and vb represents the parameter between [-g,g]. max(t) represents the maximum number of iterations, vc represents the decreasing variable between [0,1], W is the weight of the slime mold, p is the condition parameter controlling the update of the slime mold position, S(h) is the individual fitness value of the h-th slime mold, and DF is the optimal fitness value among all iterations.

[0047] In the food encirclement phase, random inertial weights are introduced to dynamically adjust the search step size of the SMA algorithm, balancing its global exploration and local exploitation capabilities; at the same time, the Levy flight strategy is incorporated to enhance the SMA algorithm's ability to escape local optima.

[0048] The formula for random inertia weights is:

[0049] w = w min +(w max -w min )·rand()+σ·randn()

[0050] Where w is the dynamic weight, w min and w max Let represent the upper and lower limits of the weights, respectively. `rand()` represents a uniformly distributed random number within the interval (0,1), `randn()` represents a random number distributed according to a standard normal distribution, and σ represents the standard deviation. The position update formula for the slime mold is then:

[0051]

[0052] When r ≥ p, a Lévy flight strategy is introduced. Lévy flight adopts a random walk pattern, and its step size follows a Lévy distribution, as shown in the formula:

[0053]

[0054] in, This represents the position of the i-th individual at the t-th iteration, and α represents the step size control parameter. Let Levy(λ) represent point-to-point multiplication, and let Levy(λ) represent the random step size. The final formula for updating the position of the slime mold population is as follows:

[0055]

[0056] Furthermore, in step 5, the ISMA algorithm is used to perform a global search on the hyperparameters (penalty parameter C, insensitive loss function parameter ε, kernel function parameter γ) of the SVR model to obtain the optimal hyperparameter combination; the convergence iteration curves of ISMA and traditional SMA in the hyperparameter optimization process are plotted, and the convergence speed and optimization effect of the two are compared and analyzed to verify the superiority of ISMA.

[0057] Furthermore, in step 6, the ISMA-SVR regional carbon emission prediction model is compared with existing mainstream prediction models such as the traditional SVR model, SMA-SVR model, and GA-SVR model. The comparison is conducted using mean squared error (MSE), mean absolute error (MAE), mean absolute percentage error (MAPE), root mean square error (RMSE), and coefficient of determination (R²). 2 The indicators are used to quantitatively evaluate the prediction accuracy and stability of each model.

[0058] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0059] To address the problems of existing carbon emission influencing factor screening being limited to a single dimension and lacking a systematic multi-dimensional screening system, which easily introduces redundant features or omits key factors, this invention proposes a multi-level feature screening strategy of "preliminary multi-dimensional selection - grey relational analysis screening - final screening by extending the STIRPAT model and combining Lasso regression", thereby improving the comprehensiveness and reliability of influencing factor selection.

[0060] To address the limitations of existing carbon emission prediction technologies, which rely heavily on traditional models or single algorithms and fail to adapt to the inherent characteristics of "small sample" carbon emission datasets, resulting in insufficient prediction accuracy, this invention constructs an improved SVR prediction model suitable for small sample scenarios, thereby overcoming the limitations of single machine learning prediction models.

[0061] To address the limitations of existing SVR models, where hyperparameters rely heavily on manual experience and intelligent optimization algorithms lack the ability to achieve globally optimal hyperparameter configuration, this invention improves the traditional slime mold algorithm by introducing stochastic inertial weights and the Lévy flight strategy. The improved ISMA algorithm iteratively optimizes the hyperparameters of the SVR model, achieving globally optimal hyperparameter configuration, thereby improving the accuracy of the carbon emission prediction model and preventing the model from getting trapped in local optima. Attached Figure Description

[0062] Figure 1 This is the overall flowchart of the present invention;

[0063] Figure 2Flowchart of the multi-level influencing factor screening system framework;

[0064] Figure 3 A comparison chart of ISMA and SMA iterations across four benchmark functions;

[0065] Figure 4 Flowchart for ISMA-SVR model parameter optimization;

[0066] Figure 5 A comparison chart of the iterations of the ISMA-SVR model and the SMA-SVR model;

[0067] Figure 6 A comparison chart of the predicted values ​​of the ISMA-SVR model with those of other models;

[0068] Figure 7 This is a comparison chart of the residuals between the predicted and actual values ​​of the ISMA-SVR model and other models. Detailed Implementation

[0069] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.

[0070] The following is a detailed description of the embodiments in conjunction with the accompanying drawings:

[0071] The ISMA-SVR regional carbon emission prediction method proposed in this invention, based on multi-layer influencing factor screening, has the following overall process: Figure 1 As shown, it includes the following steps:

[0072] Step 1: Calculate the carbon emissions of the target area, construct a basic carbon emissions dataset, and select potential influencing factors related to carbon emissions in the target area for preprocessing;

[0073] Step 2: Based on grey relational analysis, screen the preprocessed potential influencing factors to obtain a subset of influencing factors;

[0074] Step 3: Use the extended STIRPAT model combined with the Lasso regression algorithm to perform final screening of the subset of influencing factors and obtain the key feature set;

[0075] Step 4: Combine the stochastic inertial weights and the Levy flight strategy with the traditional slime mold algorithm SMA to construct the improved slime mold optimization algorithm ISMA;

[0076] Step 5: Apply the constructed ISMA algorithm to the hyperparameter optimization of the support vector regression (SVR) model to construct the ISMA-SVR regional carbon emission prediction model;

[0077] Step 6: Using the carbon emission basic dataset as the target output and the final selected key feature set as the input, substitute them into the ISMA-SVR model to predict regional carbon emissions and obtain the regional carbon emission prediction results.

[0078] Furthermore, the specific implementation process of step 1 is as follows: From the three dimensions of population size, economic level and technological level, the full names and abbreviations of the nine potential influencing factors are as follows: Urbanization level (UL), Population size (P), Gross Regional Product (GP), Total Fixed Asset Investment (FI), Total Total Consumption (TC), R&D input (RD), Energy structure (ES), Energy intensity (EI), and Industrial structure (IS).

[0079] Wherein, energy intensity = total energy consumption / regional GDP; energy structure = consumption of a certain type of energy / total energy consumption; industrial structure = (value of tertiary industry / regional GDP) * 100%. The remaining data are from official sources such as the China Statistical Yearbook, the China Energy Statistical Yearbook, and the Jiangsu Statistical Yearbook.

[0080] For missing data on influencing factors, linear interpolation is used to fill in the gaps, and the formula is as follows:

[0081]

[0082] Where x is the position of the target independent variable to be interpolated, y is the interpolation result to be calculated, and x0, y0, x1, and y1 are the independent variable values ​​and function values ​​of the first and second points, respectively.

[0083] Furthermore, the specific implementation process of step 2 is as follows: using the carbon emission data of Jiangsu Province from 2000 to 2021 as a reference sequence, and using nine potential influencing factors as a comparison sequence, the grey correlation degree between each potential factor and regional carbon emissions is shown in Table 1 below:

[0084] Table 1

[0085]

[0086]

[0087] As shown in the table above, the grey relational degree of all potential influencing factors exceeds the correlation threshold of 0.6, indicating a significant correlation between the initially selected influencing factors and regional carbon emissions. Therefore, after screening using grey relational analysis, nine influencing factors were identified as a subset of influencing factors.

[0088] Furthermore, the specific implementation process of step 3 is as follows: By extending the STIRPAT model, the following equation is used to conduct a quantitative analysis of the influencing factors of regional carbon emissions. This equation incorporates a subset of influencing factors obtained through grey relational analysis, and the formula is as follows:

[0089]

[0090] Linearizing the model equations using a logarithmic transformation yields the following simplified expression:

[0091] lnI=lna+β1lnUL+β2lnP+β3lnGP+β4lnFI+β5lnTC

[0092] +β6lnRD+β7lnES+β8lnEI+β9lnIS+lne

[0093] Where lnI is the dependent variable, lnUL, lnP, lnGP, lnFI, lnTC... are independent variables, lne is the error term, and a, β1, β2, β3, β4... represent all parameters to be estimated.

[0094] After clarifying the mathematical relationship between carbon emissions and various influencing factors, a multiple linear regression analysis was conducted on the variables involved in the model. The regression results are shown in Table 2 below:

[0095] Table 2

[0096]

[0097] In the table, Sig represents the significance level of the variable, and VIF represents the variance inflation factor.

[0098] The results in the table show that the variance inflation factor of most variables exceeds 10, indicating significant multicollinearity among the currently selected influencing factors. Lasso regression is introduced to address this issue. After optimization of the Lasso regression objective function, key factors with a significant impact on carbon emissions are finally selected, and their corresponding regression coefficients are shown in Table 3 below. For the complete framework and process of the multi-level influencing factor screening system, please refer to [link to relevant documentation]. Figure 2 The flowchart is shown.

[0099] Table 3

[0100]

[0101] As can be seen from the six influencing factors finally selected in the table, population size, fixed asset investment, and urbanization level have been eliminated. The remaining six influencing factors become the key feature set of carbon emissions in Jiangsu Province. The final extended STIRPAT model expression is as follows:

[0102] lnI=0.698lnGP+0.114lnTC+0.462lnRD+0.0577lnES+0.0208lnEI-0.217lnIS

[0103] Furthermore, the specific implementation process of step 4 is as follows: Four benchmark test functions are selected: Sphere, Rastrigin, Ackley, and Griewank, comprehensively covering different optimization scenarios. A systematic comparative analysis is conducted on the original SMA algorithm and the improved ISMA algorithm under various test environments.

[0104] The algorithm parameters for the simulation experiment are set as follows: population size 30, dimension 30, maximum number of iterations 100. Figure 3 The convergence characteristics of the ISMA algorithm and the original SMA algorithm are compared, with the horizontal axis representing the number of iterations and the vertical axis representing the fitness value of the benchmark function.

[0105] As can be seen from the convergence curves of the four benchmark functions, the ISMA algorithm exhibits a more significant advantage over the original SMA algorithm in both convergence speed and solution accuracy. Regarding convergence speed, the fitness value of ISMA decreases rapidly in the later stages of iteration, while the fitness value of SMA often remains stable for a long time, making effective convergence difficult. In terms of solution accuracy, ISMA can ultimately achieve an extremely low fitness value close to the theoretical optimum, while SMA maintains a consistently high fitness level, making it difficult to obtain an exact solution. This core advantage enables ISMA to accurately and efficiently optimize the hyperparameters of SVR.

[0106] Furthermore, the specific implementation process of step 5 is as follows: A radial basis function kernel is selected for the SVR model; its hyperparameters are iteratively optimized using the ISMA algorithm; the following are the specific steps of ISMA-SVR hyperparameter optimization, and its core process can be found in [reference needed]. Figure 4 :

[0107] (1) Data splitting: Jiangsu Province carbon emission data (as output variable) and corresponding key influencing factor data (as input variable) are divided into two subsets, with the training set accounting for 70% and the test set accounting for 30%.

[0108] (2) Parameter configuration: The population size of slime mold is set to 30, the maximum number of iterations is 200, and the search ranges of hyperparameters C, γ and ε are [0.001,1000], [0.001,1000] and [0.0001,0.1], respectively.

[0109] (3) Population initialization: Population initialization was completed with 30 slime mold locations. The fitness of each slime mold location was evaluated, and the optimal location of the current population and its corresponding fitness were determined.

[0110] (4) Position update: Combine random inertial weights and Levy flight strategy to optimize the slime mold position update formula. Compare the fitness of the newly generated slime mold position with the original position. If the new position performs better, update the optimal position of the population.

[0111] (5) Fitness evaluation: The fitness function is evaluated on the optimal position obtained in step (4) to determine whether it meets the preset convergence criterion. If it meets the criterion, the search process is terminated and the SVR hyperparameters corresponding to the global optimal position are output; if it does not meet the criterion, the process returns to step (4) to continue the iterative search.

[0112] like Figure 5 As shown, the ISMA-SVR model proposed in this invention exhibits a rapid and continuous decrease in fitness value during the iteration process, outperforming the SMA-SVR model in both convergence speed and final optimization accuracy. This performance verifies the excellent ability of the ISMA algorithm to effectively identify hyperparameters close to the optimal value.

[0113] Furthermore, the specific implementation process of step 6 is as follows: Model prediction and validation are carried out based on the 22-year time series dataset of Jiangsu Province from 2000 to 2021. The dataset is divided into training and test sets in a 7:3 ratio. To comprehensively compare and evaluate the prediction accuracy, the ISMA-SVR model is compared with six other benchmark models, specifically including: Backpropagation Neural Network (BPNN), basic SVR model, Genetic Algorithm-Optimized SVR (GA-SVR), Particle Swarm Optimized SVR (PSO-SVR), Marine Predator Algorithm-Optimized SVR (MPA-SVR), and Primitive Slime Mold Algorithm-Optimized SVR (SMA-SVR). The error evaluation indicators and prediction value comparison results of each model are shown in Table 4 below. Figure 6 , Figure 7 As shown.

[0114] Table 4

[0115]

[0116] Experimental results show that the BPNN model performs the worst in terms of fitting accuracy and error control, with R... 2 The value was only 0.8954, while the MAPE was as high as 1.291%, significantly lower than that of all SVR series models. The SVR model performed well in baseline predictions, with R... 2 The value is 0.9402, but its error index still has room for improvement. The GA-SVR model optimized by the genetic algorithm will have R... 2 The R² value increased to 0.9472, while the MAPE decreased to 0.799%, indicating a performance improvement. The PSO-SVR model optimized using the particle swarm optimization algorithm achieved an R² value of 0.9470. 2With a value and MAPE of 0.797%, the performance is comparable to GA-SVR. The MPA-SVR model, optimized by the marine predator algorithm, will achieve R... 2 The R² value was increased to 0.9714, and the MAPE decreased to 0.585%, showing a slight advantage over PSO-SVR and GA-SVR in terms of fitting accuracy and error control. However, the performance of MPA-SVR still lags behind models optimized based on slime mold-inspired algorithms. The SMA-SVR optimized using the original slime mold algorithm achieved an R² value of 0.9821. 2 The MAPE value decreased to 0.504%, showing a significant improvement in error control and fitting ability. In contrast, ISMA-SVR performed best on the test set, with R... 2 The MAPE was 0.9877, and the MAPE was as low as 0.428%. All other evaluation indicators were also better than other comparative prediction models.

[0117] Specifically, compared to SMA-SVR, ISMA-SVR reduced MAPE by 15.08%, R 2 It increased by 0.56 percentage points; compared with GA-SVR, its MAPE decreased by 46.43%, R 2 It increased by 4.05 percentage points; compared to the baseline SVR model, ISMA-SVR significantly reduced MAPE by 60.4%, R 2 This represents an improvement of 4.75 percentage points. This clearly demonstrates that optimizing the hyperparameters of the SVR model using the ISMA algorithm can significantly improve the model's prediction accuracy, validating the effectiveness of the improved algorithm.

[0118] This invention employs the aforementioned technology and systematically analyzes the influencing factors of regional carbon emissions through three steps: (1) preliminary selection of potential influencing factors; (2) screening through grey relational analysis; and (3) application of an extended STIRPAT model combined with Lasso regression to finally determine key factors, thereby enhancing the effectiveness of the input features of the prediction model. Simultaneously, by introducing stochastic inertial weights and the Lévy flight strategy to improve the traditional SMA algorithm, an ISMA algorithm is constructed, which then globally optimizes the hyperparameters of the SVR model, establishing an ISMA-SVR model with better prediction performance. Finally, through comparative verification of multiple models, the prediction accuracy of the model proposed in this invention is significantly better than existing comparative models, fully demonstrating the superiority of the technical solution.

Claims

1. A regional carbon emission prediction method based on multi-layered influencing factor screening for ISMA-SVR, characterized in that, Includes the following steps: Step 1: Calculate the carbon emissions of the target area, construct a basic carbon emissions dataset, and select potential influencing factors related to carbon emissions in the target area for preprocessing; Step 2: Based on grey relational analysis, screen the preprocessed potential influencing factors to obtain a subset of influencing factors; Step 3: Use the extended STIRPAT model combined with the Lasso regression algorithm to perform final screening of the subset of influencing factors and obtain the key feature set; Step 4: Combine the stochastic inertial weights and the Levy flight strategy with the traditional slime mold algorithm SMA to construct the improved slime mold optimization algorithm ISMA; Step 5: Apply the constructed ISMA algorithm to the hyperparameter optimization of the support vector regression (SVR) model to construct the ISMA-SVR regional carbon emission prediction model; Step 6: Using the carbon emission basic dataset as the target output and the final selected key feature set as the input, substitute them into the ISMA-SVR model to predict regional carbon emissions and obtain the regional carbon emission prediction results.

2. The ISMA-SVR regional carbon emission prediction method based on multi-layer influencing factor screening according to claim 1, characterized in that, The specific implementation of the preprocessing of potential influencing factors related to carbon emissions in the target area is as follows: q potential influencing factors related to carbon emissions are selected from three dimensions: population size, economic level, and technological level; the original data of the selected potential influencing factors are preprocessed: missing values ​​are filled in using linear interpolation to obtain the influencing factor dataset, which provides a data foundation for subsequent feature selection.

3. The ISMA-SVR regional carbon emission prediction method based on multi-layer influencing factor screening according to claim 2, characterized in that, Step 2 is implemented as follows: Grey relational analysis was used to calculate the grey relational degree between the selected potential influencing factors and the regional carbon emissions. After setting a correlation degree threshold, redundant potential influencing factors with grey relational degrees below the threshold were removed to obtain a subset of influencing factors related to carbon emissions.

4. The ISMA-SVR regional carbon emission prediction method based on multi-layer influencing factor screening according to claim 3, characterized in that, The gray relational degree is calculated in detail as follows: The comparison sequence X consists of q potential influencing factors. i =(x i (1),x i (2),...,x i (n)), i=1,2,...,q:Regional carbon emissions are used as a reference sequence X0=(x0(1),x0(2),...,x0(n)); where n is the length of each sequence; The reference and comparison sequences are dimensionless by means of the average method. This is applied to the original data x of the i-th sequence at time k. i (k) yields the corresponding dimensionless value x. i '(k); Calculate the correlation coefficient ξ i (k), the formula is as follows: Where, ξ i (k) represents the correlation coefficient between the i-th comparison sequence and the reference sequence at time k, and x0'(k) and x i '(k) represents the reference sequence and the comparison sequence at time k, respectively, min i min k |x0'(k)-x i '(k)| represents the minimum difference between two levels, max i max k |x0'(k)-x i '(k)| represents the maximum difference between the two levels, and ρ represents the resolution coefficient; The final grey relational degree r is obtained by averaging all correlation coefficients for each comparison sequence. i .

5. The ISMA-SVR regional carbon emission prediction method based on multi-layer influencing factor screening according to claim 4, characterized in that, Step 3 is implemented as follows: The extended STIRPAT model, which introduces new influencing factors, characterizes the relationship between potential influencing factors and regional carbon emissions. The extended STIRPAT model formula is as follows: Where a represents a constant term, β1, β2...β q Let represent the parameter to be estimated, e represent the error term, I represent the regional carbon emissions, and P, A...U represent q potential influencing factors related to carbon emissions, respectively. Taking the natural logarithm of both sides of the above equation, we can transform it into a linear form that can be solved by regression: lnI=lna+β1lnP+β2lnA+...+β q lnU+lne Where lnI is the dependent variable, lnP, lnA...lnU are the independent variables of the model, lna is the constant term, and lne is the error term; Lasso regression's L1 regularization property is used to sparsify the parameters to be estimated, eliminating minor influencing factors with parameters of 0 and retaining factors that have a significant impact on regional carbon emissions as the final key feature set. The Lasso regression objective function is: in, This represents the coefficient estimation results obtained through the Lasso regression method, where n is the sample size and y is the coefficient value. i Let X be the dependent variable value of the i-th sample. i Let be the vector of independent variables for the i-th sample, and β be a vector of coefficients representing all β values ​​to be estimated. j The set, β j Let λ represent the coefficient corresponding to the j-th independent variable, and λ be the regularization coefficient.

6. The ISMA-SVR regional carbon emission prediction method based on multi-layer influencing factor screening according to claim 5, characterized in that, Step 4 is implemented as follows: The optimization capability of the slime mold algorithm is based on the mathematical model of slime mold propagation waves, which is divided into three stages: approaching food, surrounding food, and acquiring food; the approaching food and surrounding food stages are distinguished by the two parameters rand and r. rand<z represents the approaching food stage, and r<p and r≥p represent the surrounding food stage. t represents the current iteration number, X(t) represents the position of the slime mold at time t, X b (t) represents the optimal position of the slime mold at time t, X A (t) and X B (t) represents two randomly selected positions from the slime mold at time t, N represents the total number of individuals in the slime mold population, r and rand are random values ​​between [0,1], ub and lb represent the upper and lower bounds of the search range, z is the probability that the isolated slime mold individual searches for other food sources, and vb represents the parameter between [-g,g]. max(t) represents the maximum number of iterations, vc represents the decreasing variable between [0,1], W is the weight of the slime mold, p is the condition parameter controlling the update of the slime mold position, S(h) is the individual fitness value of the h-th slime mold, and DF is the optimal fitness value among all iterations. In the food encirclement phase, a random inertial weight is introduced to dynamically adjust the search step size of the SMA algorithm, balancing global exploration and local exploitation capabilities. Simultaneously, the Levy flight strategy is incorporated to enhance the SMA algorithm's ability to escape local optima. The formula for the random inertial weight is: w=w min +(in max -In min )·rand()+σ·randn() Where w is the dynamic weight, w min and w max Let represent the upper and lower limits of the weights, respectively; rand() represents a uniformly distributed random number in the interval (0,1); randn() represents a random number distributed according to a standard normal distribution; and σ represents the standard deviation. The formula for updating the position of the slime mold is as follows: p=tanh|S(h)-DF|, (h=1,2,...,N) When r ≥ p, a Lévy flight strategy is introduced, with the step size following a Lévy distribution, as shown in the formula: in, This represents the position of the i-th individual at the t-th iteration, and α represents the step size control parameter. Let Levy(λ) represent point-to-point multiplication, and let Levy(λ) represent the random step size. The final formula for updating the position of the slime mold population is as follows:

Images