Electric vehicle ownership prediction method and system

Abstract

The invention provides an electric vehicle ownership prediction method and system, and belongs to the technical field of electric vehicle ownership prediction. The method comprises the following steps: acquiring historical time sequence data of a multi-source input variable related to the inventory of the electric vehicle, and preprocessing and standardizing the historical time sequence data; decomposing the data into a trend term, a season term and a residual term by adopting a time sequence decomposition algorithm; screening out an optimal feature subset for each component by using an information measurement algorithm; constructing a combined kernel function formed by linearly combining a plurality of basic kernel functions; optimizing and solving an optimal kernel weight combination by adopting an improved particle swarm optimization algorithm; and training a relevance vector machine prediction model by using the optimal kernel weight combination and the optimal feature subset, and integrating prediction results of all components to obtain a final electric vehicle inventory prediction value. According to the method, multiple influence factors are comprehensively considered, feature selection and kernel function combination are optimized, and the accuracy and stability of electric vehicle inventory prediction are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of electric vehicle ownership prediction technology, and in particular to a method and system for predicting electric vehicle ownership. Background Technology

[0002] The widespread adoption of electric vehicles not only helps reduce greenhouse gas emissions but also lowers dependence on traditional fossil fuels, promoting the optimization and upgrading of the energy structure. Therefore, accurately predicting the number of electric vehicles in use is of significant practical importance for formulating relevant policies, planning the construction of charging infrastructure, and guiding industrial development.

[0003] However, predicting the number of electric vehicles (EVs) faces numerous challenges. First, the market penetration of EVs is influenced by a combination of factors, including economic development levels, residents' income levels, advancements in battery technology, the availability of charging infrastructure, and policy support. These factors intertwine, resulting in complex, non-linear characteristics in changes in EV ownership. Second, the rapid development and uncertainty of the EV market increase the difficulty of prediction. Fluctuations in market demand, the emergence of new technologies, and policy adjustments can all lead to unexpected changes in EV ownership. Furthermore, traditional prediction methods often struggle to effectively handle the fusion of multi-source data and complex time-series characteristics, resulting in low prediction accuracy and failing to meet the needs of practical applications.

[0004] While existing technologies and methods exist for predicting electric vehicle (EV) ownership, most have limitations. For example, some methods consider only a few influencing factors, failing to fully reflect the complexity of EV ownership changes; others rely on a single prediction model, lacking optimized combinations of different features and kernel functions, resulting in insufficient accuracy and stability of prediction results. Therefore, how to comprehensively consider multiple influencing factors and construct an EV ownership prediction model that can effectively handle complex nonlinear relationships and has high prediction accuracy is a pressing technical problem that needs to be solved. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for predicting the number of electric vehicles. By comprehensively considering multiple input variables and employing techniques such as time series decomposition, feature optimization selection, combined kernel function construction, and intelligent optimization algorithms, the accuracy and stability of the prediction are effectively improved.

[0006] To achieve the above objectives, the present invention provides a method for predicting the number of electric vehicles in use, comprising the following steps:

[0007] Step S1: Obtain historical time series data of multi-source input variables related to electric vehicle ownership, and preprocess and standardize the historical time series data to obtain standardized data;

[0008] Step S2: Use a time series decomposition algorithm to decompose the electric vehicle ownership data into three components: trend, seasonality, and residual.

[0009] Step S3: Based on the three components and standardized data, use the information measurement algorithm to select the corresponding optimal feature subset for each component;

[0010] Step S4: Construct a combined kernel function that is a linear combination of multiple basic kernel functions;

[0011] Step S5: Using an intelligent optimization algorithm, the performance of the prediction model constructed based on the optimal feature subset and combined kernel function is used as the fitness value to optimize and solve for the optimal kernel weight combination.

[0012] Step S6: Substitute the optimal kernel weight combination into the combined kernel function, and use the optimal feature subset to train the correlation vector machine prediction model corresponding to each component respectively;

[0013] Step S7: Integrate the prediction results of each component to obtain the final predicted value of electric vehicle ownership.

[0014] Preferably, in step S1, the multi-source input variables include: the annual growth rate of GDP, the annual growth rate of per capita disposable income, the annual increase in the unit energy density of power lithium batteries, the annual growth rate of the number of public charging piles, and the weighted sum of the number of national and provincial annual new energy vehicle promotion policies.

[0015] Preferably, in step S1, preprocessing and standardization refer to: filling missing values ​​with linear interpolation and standardizing the complete data with Z-Score to finally obtain standardized data.

[0016] Preferably, step S2 specifically includes:

[0017] Using the STL decomposition method, the electric vehicle ownership data is decomposed into three additive components, expressed as follows:

[0018] Y v (τ)=T τ +S τ +R τ ;

[0019] Among them, Y v (τ) represents the number of electric vehicles at time τ, where T τ S represents the trend term at time τ. τ R represents the seasonal term at time τ with a period of 12 months. τThis represents the residual term at time τ.

[0020] Preferably, in step S3, the information measurement algorithm is the maximum information coefficient, specifically:

[0021] For each factorization term D∈{T t ,S t ,R t}, calculate the maximum information coefficient MIC(X) for each input variable in the standardized data. j D), the calculation formula is:

[0022] MIC(X j ,D)=max G I(X j ,D)|G;

[0023] Where G represents the variable (X) j The scatter plot of X(D) is divided into grids, where I(.) represents the mutual information value at grid G, and X(D) represents the mutual information value at grid G. j This represents the j-th input variable in the standardized data;

[0024] For each decomposition term D, select one that makes MIC(X) j The top K largest input variables (D) constitute the optimal feature subset F. D .

[0025] Preferably, in step S4, the combined kernel function consists of a linear kernel, a Gaussian kernel, a polynomial kernel, a periodic kernel, and an exponential sine square kernel, and its expression is:

[0026] K combined (x i ,x j )=β1K lin (x i ,x j )+β2K rbf (x i ,x j )+β3K poly (x i ,x j )

[0027] +β4K peri (x i ,x j )+β5K ess (x i ,x j );

[0028] Among them, K lin (x i ,x j ) represents a linear kernel function, K rbf (xi ,x j ) represents the Gaussian kernel function, K poly (x i ,x j ) represents the polynomial kernel function, K peri (x i ,x j ) represents a periodic kernel function, K ess (x i ,x j () represents the exponential sine square kernel function, where β1, β2, β3, β4, and β5 are kernel weight parameters, satisfying β1 + β2 + β3 + β4 + β5 = 1 and all weights are non-negative. i x j Both represent the feature vectors of the input samples.

[0029] The five kernel function combinations selected in this invention (the calculation formulas for the five kernel functions are existing technologies and will not be elaborated here) are fundamentally intended to construct a prediction model with strong expressive power and wide adaptability, so as to accurately capture the various complex patterns contained in the time series of electric vehicle ownership.

[0030] Linear kernel functions are used to capture the global long-term linear growth trend of stock ownership resulting from the development of macroeconomic factors (such as economy, technology, and policy), which forms the main framework for forecasting. Gaussian kernel functions (also known as radial basis function kernels) have strong local fitting capabilities and can flexibly characterize smooth nonlinear changes and short-term fluctuations in data, enabling models to adapt to bends and turns in growth trends.

[0031] The polynomial kernel function further extends the model's ability to capture nonlinear relationships. It can characterize more complex interactions between variables, such as the pattern of acceleration or deceleration of the growth rate itself, thus better fitting the acceleration or saturation phenomena that may occur in the actual development process.

[0032] To address the significant cyclical fluctuations in electric vehicle ownership (such as annual or quarterly cycles), a periodic kernel function is introduced. This function can accurately learn and predict these recurring seasonal patterns. Meanwhile, the exponential sine square kernel function, as a supplement to the periodic kernel, offers smoother and more flexible characteristics, enabling it to capture quasi-periodic patterns with slight variations in cycle length or amplitude, better reflecting the reality that seasonal effects are not entirely consistent year after year.

[0033] By linearly combining these five distinct kernel functions and automatically learning their weights using optimization algorithms, the resulting combined kernel function becomes a comprehensive covariance structure. This enables the predictive model to simultaneously characterize trends, seasonality, nonlinearity, and local volatility in the data, and to autonomously determine which pattern dominates. Consequently, it exhibits excellent fitting ability and predictive accuracy even when faced with complex and variable small sample data.

[0034] Preferably, in step S5, the intelligent optimization algorithm is an improved particle swarm optimization algorithm. The improved particle swarm optimization algorithm adopts a two-factor inertia weight dynamic adjustment strategy based on iteration progress and population diversity, specifically:

[0035] Step S51: Initialize the particle swarm. The position vector of each particle represents a set of kernel weights [β1,β2,β3,β4,β5].

[0036] Step S52: For each particle, construct a combined kernel function using the weights it represents, and train a related vector machine model based on this kernel function and the optimal feature subset.

[0037] Step S53: Calculate the fitness value of the relevant vector machine model on the validation set. The fitness function is:

[0038] Fitness=λ1(1 / MAPE)+λ2(1 / RMSE);

[0039] Where MAPE represents the mean absolute percentage error, RMSE represents the root mean square error, and λ1 and λ2 both represent weighting coefficients.

[0040] Step S54: By iteratively updating the particle's velocity and position, the dynamic adjustment formula for the inertia weight ω is:

[0041]

[0042] Where, ω max and ω min α and β represent the maximum and minimum values ​​of the inertia weight, respectively; t represents the current iteration number; T represents the maximum iteration number; D' represents the population diversity index; and α, β, and γ all represent adjustment parameters.

[0043] Preferably, in step S6, the relevant vector machine model is:

[0044]

[0045] Where y(x) represents the predicted value of the decomposition term D, N represents the number of training samples, and ω i W0 represents the model weights, W0 represents the bias term, and K represents the model weights. combined (x,xi ,β opt ) represents the combined kernel function whose weights are determined by optimization in step S5, and x represents the feature vector of the sample to be predicted; x i Let β represent the feature vector of the i-th sample in the training set. opt This represents the optimal kernel weight combination obtained from step S5.

[0046] The present invention also provides an electric vehicle ownership prediction system, comprising:

[0047] Data acquisition module: used to acquire historical time series data of multi-source input variables related to electric vehicle ownership;

[0048] Data preprocessing module: used to preprocess and standardize the acquired historical time series data;

[0049] Time series decomposition module: The STL decomposition method is used to decompose the electric vehicle ownership data into three components: trend, seasonality, and residual.

[0050] Feature selection module: Based on three components and standardized data, the maximum information coefficient algorithm is used to select the corresponding optimal feature subset for each component;

[0051] Kernel function construction module: Constructs combined kernel functions that are linear combinations of linear kernels, Gaussian kernels, polynomial kernels, periodic kernels, and exponential sine square kernels;

[0052] Model optimization module: An improved particle swarm optimization algorithm is used to optimize and solve for the optimal combination of kernel weights by using the performance of the prediction model built based on the optimal feature subset and combined kernel function as the fitness value.

[0053] Model training module: Substitute the optimal kernel weight combination into the combined kernel function, and use the optimal feature subset to train the correlation vector machine prediction model corresponding to each component respectively;

[0054] Results integration module: Integrates the prediction results of each component to obtain the predicted value of electric vehicle ownership.

[0055] Therefore, the present invention employs the above-described method and system for predicting the number of electric vehicles, and the beneficial technical effects are as follows:

[0056] (1) By comprehensively considering multiple input variables, including economic, technological, infrastructure and policy factors, this invention fully covers various factors affecting the change in electric vehicle ownership, solving the problem of incomplete prediction caused by considering only one factor in the prior art, thereby significantly improving the accuracy and reliability of the prediction.

[0057] (2) To address the problem that existing technologies struggle to effectively handle multi-source data fusion and complex time series features, this invention employs a time series decomposition algorithm to decompose data into trend, seasonal, and residual terms, and constructs a combined kernel function composed of linear combinations of various basic kernel functions. This decomposition and combination method effectively handles complex nonlinear relationships in the data, improves the model's expressive power and adaptability, and thus enhances prediction accuracy.

[0058] (3) This invention employs an improved particle swarm optimization algorithm. By using a dynamic adjustment strategy based on a two-factor inertial weight based on iteration progress and population diversity, it effectively avoids the problem of traditional optimization algorithms easily getting trapped in local optima, thereby improving global search capability and convergence speed. This improvement significantly enhances the training efficiency of the model and the stability of the prediction results, ensuring that the performance of the prediction model reaches its optimal level. Attached Figure Description

[0059] Figure 1 This is a flowchart of a method for predicting the number of electric vehicles according to the present invention;

[0060] Figure 2 Flowchart of the improved particle swarm optimization algorithm;

[0061] Figure 3 This is a flowchart of an electric vehicle ownership prediction system according to the present invention. Detailed Implementation

[0062] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0063] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0064] Example 1

[0065] like Figure 1 As shown, a method for predicting the number of electric vehicles includes the following steps:

[0066] Step S1: Obtain historical time series data of multi-source input variables related to electric vehicle ownership (in this embodiment, monthly historical data of a certain region from January 2015 to December 2023 are used), and preprocess and standardize the historical time series data to obtain standardized data.

[0067] The multi-source input variables include: annual growth rate of GDP (%), annual growth rate of per capita disposable income (%), annual increase in unit energy density of power lithium batteries (Wh / kg), annual growth rate of the number of public charging piles (%), and the weighted sum of the number of national and provincial annual new energy vehicle promotion policies (items).

[0068] The weighted sum of the number of national and provincial annual new energy vehicle promotion policies is calculated as follows:

[0069] A = B + 0.6C;

[0070] Where A represents the weighted sum of the number of policies, B represents the number of national-level policies, and C represents the number of provincial-level policies.

[0071] Preprocessing and standardization refer to: using linear interpolation to fill in missing values ​​for individual months, and using Z-Score standardization on the complete data to make its mean 0 and standard deviation 1, thus obtaining standardized data.

[0072] Step S2: Use a time series decomposition algorithm to decompose the electric vehicle ownership data into three components: trend, seasonality, and residual.

[0073] Using the STL (Seasonal and Trend decomposition using Loess) method, the electric vehicle ownership data is decomposed into three additive components, expressed as follows:

[0074] Y v (τ)=T τ +S τ +R τ ;

[0075] Among them, Y v (τ) represents the number of electric vehicles at time τ, where T τ S represents the trend term at time τ. τ R represents the seasonal term at time τ with a period of 12 months. τ This represents the residual term at time τ.

[0076] Step S3: Based on the three components and standardized data, use the information measurement algorithm to select the corresponding optimal feature subset for each component.

[0077] The information measurement algorithm is the maximum information coefficient, specifically:

[0078] For each factorization term D∈{T t ,S t ,R t}, calculate the maximum information coefficient MIC(X) for each input variable in the standardized data. j D), the calculation formula is:

[0079] MIC(X j ,D)=max G I(X j ,D)|G;

[0080] Where G represents the variable (X) j The scatter plot of X(D) is divided into grids, where I(.) represents the mutual information value at grid G, and X(D) represents the mutual information value at grid G. j This represents the j-th input variable in the standardized data;

[0081] For each decomposition term D, select one that makes MIC(X) j The top K largest input variables (D) constitute the optimal feature subset F. D .

[0082] In this embodiment, the top three variables with the highest MIC values ​​are selected as the optimal feature subset. This selection is based on a trade-off made in multiple experiments: selecting too few features (e.g., the first two) may result in the loss of crucial information; selecting too many features (e.g., the first four) may introduce noise, increase model complexity, and potentially lead to overfitting. The first three features achieve a good balance between information preservation and model simplicity.

[0083] The optimal feature subset of the trend term: annual GDP growth rate, annual increase in unit energy density of power lithium batteries, and the weighted sum of the number of national and provincial annual new energy vehicle promotion policies (reflecting the core driving forces of long-term development: economy, technology and policy).

[0084] The optimal subset of seasonal features: annual growth rate of per capita disposable income and annual growth rate of the number of public charging stations (reflecting the cyclical impact of consumption capacity and infrastructure convenience).

[0085] The optimal feature subset of the residual term: the annual growth rate of the number of public charging piles and the annual increase in the unit energy density of power lithium batteries (reflecting short-term fluctuations that are difficult to explain by trends and seasonality, and may be related to breaking news of infrastructure upgrades and technological breakthroughs).

[0086] Step S4: Construct a combined kernel function that is a linear combination of multiple basic kernel functions.

[0087] The combined kernel function consists of a linear kernel, a Gaussian kernel, a polynomial kernel, a periodic kernel, and an exponential squared kernel, and its expression is:

[0088] K combined (x i ,x j )=β1K lin (x i ,x j )+β2K rbf (x i ,x j )+β3K poly (x i ,x j )

[0089] +β4Kperi (x i ,x j )+β5K ess (x i ,x j );

[0090] Among them, K lin (x i ,x j ) represents a linear kernel function, K rbf (x i ,x j ) represents the Gaussian kernel function, K poly (x i ,x j ) represents the polynomial kernel function, K peri (x i ,x j ) represents a periodic kernel function, K ess (x i ,x j () represents the exponential sine square kernel function, where β1, β2, β3, β4, and β5 are kernel weight parameters, satisfying β1 + β2 + β3 + β4 + β5 = 1 and all weights are non-negative. i x j Both represent the feature vectors of the input samples.

[0091] The initial parameters of each basic kernel function are set as follows: Gaussian kernel bandwidth is 1.0, polynomial kernel degree is 3 and constant term is 1, periodic kernel and exponential sine square kernel period is 12 and length scale is 1.0.

[0092] Step S5, as follows Figure 2 As shown, an intelligent optimization algorithm is used to optimize the optimal combination of kernel weights by using the performance of the prediction model constructed based on the optimal feature subset and combined kernel function as the fitness value.

[0093] The intelligent optimization algorithm is an improved particle swarm optimization algorithm. The improved particle swarm optimization algorithm adopts a two-factor inertia weight dynamic adjustment strategy based on iteration progress and population diversity (setting the population size to 50 and the maximum number of iterations to 100), specifically:

[0094] Step S51: Initialize the particle swarm. The position vector of each particle represents a set of kernel weights [β1,β2,β3,β4,β5].

[0095] Step S52: For each particle, construct a combined kernel function using the weights it represents, and train a related vector machine model based on this kernel function and the optimal feature subset.

[0096] Step S53: Calculate the fitness value of the relevant vector machine model on the validation set. The fitness function is:

[0097] Fitness=λ1(1 / MAPE)+λ2(1 / RMSE);

[0098] Where MAPE represents the mean absolute percentage error, RMSE represents the root mean square error, both calculated based on the prediction results of the validation set, and λ1 and λ2 represent weighting coefficients.

[0099] Step S54: By iteratively updating the particle's velocity and position, the dynamic adjustment formula for the inertia weight ω is:

[0100]

[0101] Where, ω max and ω min α and β represent the maximum and minimum values ​​of the inertia weight, respectively; t represents the current iteration number; T represents the maximum iteration number; D' represents the population diversity index (calculated by the standard deviation of particle positions); and α, β, and γ all represent adjustment parameters.

[0102]

[0103] Where d represents the particle dimension, which is equal to the number of kernel weights in the combined kernel function, i.e., 5, corresponding to the weights of the linear kernel, Gaussian kernel, polynomial kernel, periodic kernel, and exponential squared kernel, respectively. σ n Let x represent the standard deviation of all particles at their nth position, M represent the population size, and x represent the standard deviation of all particles at their nth position. mn This represents the position value of the m-th particle in the n-th dimension (i.e., the n-th kernel weight).

[0104] This formula allows the inertia weight to maintain a large value in the early stage of the iteration due to the high diversity of the population, thereby enhancing the global exploration capability. In the later stage of the iteration, it adaptively decreases as the diversity decreases, thereby improving the local search accuracy. Ultimately, the optimal combination of kernel weights that maximizes Fitness is found.

[0105] Step S6: Substitute the optimal kernel weight combination into the combined kernel function, and use the optimal feature subset to train the correlation vector machine prediction model corresponding to each component.

[0106] The relevant vector machine model is:

[0107]

[0108] Where y(x) represents the predicted value of the decomposition term D, N represents the number of training samples, and ω i W0 represents the model weights, W0 represents the bias term, and K represents the model weights. combined (x,x i ,β opt) represents the combined kernel function whose weights are determined by optimization in step S5, and x represents the feature vector of the sample to be predicted; x i Let β represent the feature vector of the i-th sample in the training set. opt This represents the optimal kernel weight combination obtained from step S5.

[0109] Step S7: Integrate the prediction results of each component to obtain the final predicted value of electric vehicle ownership.

[0110] To verify the effectiveness of the method of this invention, data from January 2015 to December 2021 was used as the training set, and data from January 2022 to December 2023 was used as the test set for prediction. LSTM, SVM (Gaussian kernel), and ARIMA models were selected as comparison methods. All comparison models were trained using the same feature subset as the trend term of this invention to ensure fairness. The prediction accuracy results are shown in Table 1.

[0111] Table 1 Prediction Results

[0112] Predictive Model MAPE (%) RMSE (in ten thousand vehicles) This invention 5.2 0.87 LSTM 8.7 1.45 SVM 12.4 2.01 ARIMA 10.1 1.68

[0113] As shown in Table 1, LSTM models typically require a large amount of data for sufficient training. In this embodiment, they are prone to overfitting on medium-sized time series data, resulting in inferior generalization ability compared to the method of this invention. By employing a decomposition-component prediction-ensemble strategy, combined with an optimized combined kernel function, the method of this invention can better capture complex trends, seasonality, and residual patterns in time series, achieving performance far exceeding that of SVM using a single kernel function and the traditional time series model ARIMA.

[0114] Example 2

[0115] like Figure 3 As shown, an electric vehicle ownership prediction system includes:

[0116] Data acquisition module: used to acquire historical time series data of multi-source input variables related to electric vehicle ownership;

[0117] Data preprocessing module: used to preprocess and standardize the acquired historical time series data;

[0118] Time series decomposition module: The STL decomposition method is used to decompose the electric vehicle ownership data into three components: trend, seasonality, and residual.

[0119] Feature selection module: Based on three components and standardized data, the maximum information coefficient algorithm is used to select the corresponding optimal feature subset for each component;

[0120] Kernel function construction module: Constructs combined kernel functions that are linear combinations of linear kernels, Gaussian kernels, polynomial kernels, periodic kernels, and exponential sine square kernels;

[0121] Model optimization module: An improved particle swarm optimization algorithm is used to optimize and solve for the optimal combination of kernel weights by using the performance of the prediction model built based on the optimal feature subset and combined kernel function as the fitness value.

[0122] Model training module: Substitute the optimal kernel weight combination into the combined kernel function, and use the optimal feature subset to train the correlation vector machine prediction model corresponding to each component respectively;

[0123] Results integration module: Integrates the prediction results of each component to obtain the predicted value of electric vehicle ownership.

[0124] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.

[0125] Therefore, the present invention adopts the above-mentioned method and system for predicting the number of electric vehicles. By comprehensively considering multiple input variables and employing techniques such as time series decomposition, feature optimization selection, combined kernel function construction, and intelligent optimization algorithms, the accuracy and stability of the prediction are effectively improved.

[0126] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method of predicting the number of electric vehicles in use, characterized by, The method comprises the following steps: Step S1, obtaining historical time series data of multiple source input variables related to the electric vehicle population, and preprocessing and standardizing the historical time series data to obtain standardized data; Step S2, decomposing the electric vehicle population data into three components of trend, season and residual by using a time series decomposition algorithm; Step S3, based on the three components and the standardized data, using an information measurement algorithm to select the corresponding optimal feature subset for each component; Step S4, constructing a combined kernel function composed of multiple basic kernel functions; Step S5, using an intelligent optimization algorithm, taking the performance of the prediction model constructed based on the optimal feature subset and the combined kernel function as the fitness value, and optimizing and solving the optimal kernel weight combination; Step S6, substituting the optimal kernel weight combination into the combined kernel function, and training the related vector machine prediction model corresponding to each component by using the optimal feature subset; Step S7, integrating the prediction results of each component to obtain the final electric vehicle population prediction value.

2. The method of claim 1, wherein In step S1, the multiple source input variables include: annual growth rate of gross domestic product, annual growth rate of per capita disposable income, annual increment of unit energy density of power lithium battery, annual growth rate of public charging pile number, and weighted sum of the number of national and provincial annual new energy vehicle promotion policies.

3. The method of claim 1, wherein In step S1, preprocessing and standardization refers to: using linear interpolation method to fill in missing values, and using Z-Score standardization for complete data, finally obtaining standardized data.

4. The method of claim 1, wherein Step S2 is specifically: Using the STL decomposition method, the electric vehicle population data is decomposed into three additive components, and the expression is: Y v (τ) = T τ + S τ + R τ ; where Y v (τ) denotes the data of the number of electric vehicles at time τ, T τ denotes the trend term at time τ, S τ denotes the seasonal term at time τ with a period of 12 months, R τ denotes the residual term at time τ.

5. The method of claim 1, wherein In step S3, the information measurement algorithm is the maximum information coefficient, and the specific expression is: For each decomposition item D ∈ {T t ,S t ,R t}, the maximum information coefficient MIC(X j ,D) with each input variable in the standardized data is calculated respectively, and the calculation formula is: MIC(X j ,D) = max G I(X j ,D) | G; where G represents a grid divided on a scatter plot of variables (X j ,D), I(.) represents a mutual information value under the grid G, X j represents the jth input variable in the normalized data; Select the top K input variables that maximize MIC(X j , D) for each decomposition term D to form the optimal feature subset F D .

6. The method of claim 1, wherein In step S4, the combined kernel function is composed of linear kernel, Gaussian kernel, polynomial kernel, periodic kernel and exponential sine square kernel, and the expression is: wherein K lin (x i ,x j ) represents a linear kernel function, K rbf (x i ,x j ) represents a Gaussian kernel function, K poly (x i ,x j ) represents a polynomial kernel function, K peri (x i ,x j ) represents a periodic kernel function, K ess (x i ,x j ) represents an exponential sine square kernel function, β1, β2, β3, β4, β5 all represent kernel weight parameters, satisfying β1+β2+β3+β4+β5=1 and each weight being non-negative, x i , x j all represent feature vectors of input samples.

7. The method of claim 1, wherein In step S5, the intelligent optimization algorithm is an improved particle swarm optimization algorithm, and the improved particle swarm optimization algorithm adopts a double-factor inertia weight dynamic adjustment strategy based on iteration progress and population diversity, and the specific expression is: Step S51, initializing the particle swarm, and the position vector of each particle represents a set of kernel weights [β1, β2, β3, β4, β5]; Step S52, for each particle, using the weight represented thereby to construct a combined kernel function, and training a related vector machine model based on the kernel function and the optimal feature subset; Step S53, calculating the fitness value Fitness of the related vector machine model on the validation set, and the fitness function is: Fitness=λ1(1 / MAPE)+λ2(1 / RMSE); Wherein, MAPE represents the mean absolute percentage error, RMSE represents the root mean square error, and λ1, λ2 represent weight coefficients; Step S54, updating the speed and position of the particle by iteration, and the dynamic adjustment formula of the inertia weight ω is: where ω max and ω min represent the maximum and minimum values of the inertia weight, respectively, t represents the current iteration number, T represents the maximum iteration number, D' represents the population diversity index, and α, β, γ each represent a tuning parameter.

8. The method of claim 6, wherein In step S6, the related vector machine model is: wherein y(x) represents the predicted value of the decomposition term D, N represents the number of training samples, ω i represents the model weight, W0 represents the bias term, K combined (x, x i , β opt ) represents the combined kernel function determined by the optimization in step S5, x represents the feature vector of the sample to be predicted; x i represents the feature vector of the i-th sample in the training set, β opt represents the optimal kernel weight combination obtained by the optimization in step S5.

9. An electric vehicle population prediction system characterized by comprising: It comprises: A data acquisition module for acquiring historical time series data of multiple source input variables related to the electric vehicle population; A data preprocessing module for preprocessing and standardizing the acquired historical time series data; Time series decomposition module: the STL decomposition method is used to decompose the electric vehicle ownership data into three components: trend, seasonality and residual; Feature selection module: based on the three components and standardized data, the maximum information coefficient algorithm is used to select the optimal feature subset for each component; Kernel function construction module: a combination kernel function is constructed by linear kernel, Gaussian kernel, polynomial kernel, periodic kernel and exponential sine square kernel linear combination; Model optimization module: the improved particle swarm optimization algorithm is used to optimize the optimal kernel weight combination by taking the performance of the prediction model based on the optimal feature subset and the combination kernel function as the fitness value; Model training module: the optimal kernel weight combination is substituted into the combination kernel function, and the optimal feature subset is used to train the related vector machine prediction model corresponding to each component; Result integration module: integrate the prediction results of each component to get the electric vehicle ownership prediction value.

Images