A method for evaluating and predicting factors affecting electric vehicle charging load based on combined weights

Abstract

The present invention discloses a method for evaluating and predicting factors affecting electric vehicle charging load based on combined weights, comprising: obtaining historical charging load data in advance, preprocessing the load data, and dividing the preprocessed charging load data; constructing an index system of charging load prediction factors based on different scenarios and calculating the combined weights of the indexes; constructing a GATv2-Linformer hybrid prediction model; obtaining an IKOA based on the Kepler optimization algorithm (KOA) by fusing multiple improvement strategies, and optimizing the GATv2-Linformer prediction model parameters using the IKOA; constructing an XGBoost error correction model; obtaining an initial prediction value and an error prediction value using the improved GATv2-Linformer prediction model and the XGBoost error correction model, respectively; and correcting the generated prediction value using the prediction error value to obtain a more accurate electric vehicle charging load prediction result. The present invention expands the search scale of the Kepler optimization algorithm, improves the algorithm's ability to escape local optimality, and obtains a more accurate prediction output.

Description

Technical Field

[0001] The present invention relates to the field of new energy vehicles, and in particular to a method for evaluating and predicting factors affecting electric vehicle charging load based on combined weights. Background Art

[0002] Electric vehicle charging load forecasting is a rapidly emerging field in the context of the rapid development of electric vehicles. With increasing environmental awareness and the demand for sustainable energy utilization, electric vehicles have become a leading alternative to traditional fuel vehicles. However, the energy consumption during EV charging and its impact on the power system have become urgent challenges that need to be addressed. Charging load forecasting is crucial for grid operation planning and the planning and construction of charging infrastructure. Summary of the Invention

[0003] Purpose of the invention: The purpose of the present invention is to provide a method for evaluating and predicting factors affecting electric vehicle charging load based on combined weights.

[0004] Technical solution: The present invention comprises the following steps:

[0005] (1) Obtain the historical charging load data of electric vehicles in three charging scenarios: urban residential areas, commercial areas, and industrial parks, pre-process the load data, obtain the annual load curve, and divide the pre-processed charging load data;

[0006] (2) Construct an index system based on the factors affecting charging load prediction in different scenarios, calculate the index combination weights through expert experience and improved entropy weight method, and use the characteristic values ​​of the factors with higher weights as input and the charging load as output;

[0007] (3) Introducing the graph attention mechanism of GATv2 into the Linformer attention layer to build a GATv2-Linformer hybrid prediction model;

[0008] (4) Based on the Kepler optimization algorithm KOA, multiple improvement strategies are integrated to obtain IKOA, and IKOA is used to optimize the parameters of the GATv2-Linformer prediction model;

[0009] (5) Build the XGBoost error correction model;

[0010] (6) The improved GATv2-Linformer prediction model and XGBoost error correction model are used to obtain the initial prediction value and error prediction value respectively. The prediction error value is used to correct the generated prediction value to obtain a more accurate electric vehicle charging load prediction result.

[0011] Furthermore, the step (1) divides the obtained historical data into three types, namely, working day charging load data, non-working day charging load data and holiday charging load data, and divides the three types of charging load data into training set 1, training set 2 and test set respectively.

[0012] Furthermore, the step (2) includes:

[0013] (2.1) Establish an impact index system for multiple scenarios. Considering the different factors affecting charging load in different scenarios, a combined weight index based on expert experience and improved entropy weight method is constructed to analyze the weights of each influencing factor in different scenarios.

[0014] (2.2) Weight indicators based on expert experience are weighted by comprehensive research data and experience;

[0015] (2.3) Weight index based on improved entropy weight method Based on the original entropy weight method, the information entropy principle is introduced, and the information entropy calculation formula is:

[0016]

[0017] Where p(j) is the probability of event j occurring, and N is the total number of samples.

[0018] Calculate the weight index of each influencing factor using the following formula:

[0019]

[0020] in is the mean value of information entropy not equal to 1, H j is the information entropy value corresponding to the jth parameter;

[0021] The entropy weight is calculated by the weight index:

[0022]

[0023] where the exponent a is defined as 8, The information entropy weight is calculated using the entropy value for the traditional information entropy weight.

[0024] (2.4) Based on the multi-factor impact of charging load prediction, the combined weight index is constructed as follows:

[0025]

[0026] where α j is the subjective weight.

[0027] Furthermore, the step (3) GATv2 modifies the order of operations based on GAT, which is expressed as:

[0028]

[0029] Among them, || represents the splicing operation, w gat represents the shared weight matrix, h i，t Represents the characteristics of the node, a T is the attention kernel function, L is the activation function, α ij，t is the weight coefficient at time t;

[0030] When calculating the weight coefficient, GATv2 first performs feature concatenation and then multiplies it by the weight matrix, using a multi-head attention mechanism:

[0031]

[0032] Among them, W k To learn the parameters, is the feature vector obtained by summing the single-head attention coefficients, α ij，k is the learned weight coefficient, and σ is an arbitrary activation function.

[0033] Furthermore, the Linformer attention layer in step (3) adds two k×n matrices E and F to the attention calculation, so that the n-order attention is reduced to a fixed k-dimensional matrix:

[0034]

[0035] Reduce the number of parameters, set E = F, and define the dimension of k as 128.

[0036] Furthermore, the step (4) of optimizing the GATv2-Linformer prediction model parameters using IKOA includes:

[0037] (4.1) Use the good point set strategy initialization instead of the original random initialization:

[0038]

[0039]

[0040] x1(k)=(upper j -lower j ){P n (k)}+lower j

[0041] Its deviation satisfies:

[0042] φ(n)=C(r,δ)n -1+δ

[0043] Among them, P n(k) is a set of good points, r is a good point, n is the number of points, C(r, δ) is a constant that is only related to r and δ, δ is an arbitrarily small positive number, p is the smallest prime number that satisfies (p-3) / 2>s, s represents the dimension, upper represents the upper bound, and lower represents the lower bound;

[0044] (4.2) Using the elitist strategy in position update:

[0045]

[0046] Introducing Cauchy mutation into the position update strategy increases population diversity, improves the algorithm's global search capability, and expands the search space:

[0047] x newbest =x best +x best ×Cauchy(0,1)

[0048] Calculate the fitness value of the latest planet position, update the global optimal solution in the iterative process, take the position of the planet corresponding to the optimal fitness value as the output, and update the learning parameters of the GATv2-Linformer prediction model and the number of encoding and decoding layers of the Linformer.

[0049] Furthermore, the step (5) uses the XGBoost algorithm to build an error prediction model. XGBoost gradually improves the prediction performance of the model by connecting multiple decision trees in series:

[0050]

[0051] Among them, x i represents the i-th sample, f k (x) is the calculation result of the k-th tree model, n is the number of decision trees, and XGBoost performs some processing in the loss function and objective function. The objective function is as follows:

[0052]

[0053] Among them, l() is the loss function; Ω() is the regularization term, representing the complexity of the tree; y i is the true value of the observation, is the predicted value, K is the total number of trees in the algorithm; g i 、h i It is known in the algorithm process; γ and λ represent parameters; T represents the number of leaf nodes, ω j Represents the weight of the j-th leaf node.

[0054] Furthermore, the step (5) includes:

[0055] (5.1) Prepare the training set 2 dataset and define the parameters of the XGBoost error prediction model, including the maximum number of iterations, tree depth, number of trees, and learning rate;

[0056] (5.2) Use the training set to train the XGBoost model. During the training process, the model will gradually learn how to fit the data and optimize the prediction effect in each round of iteration to reach the maximum number of iterations and output the optimal prediction value.

[0057] Furthermore, the maximum number of iterations is 1000, the number of trees is set to 150, the depth of the tree is set to 3, and the learning rate is 0.1.

[0058] Furthermore, the step (6) includes:

[0059] (6.1) Optimize the parameters of the GATv2-Linformer prediction model using the data from training set 1 and the improved Kepler optimization algorithm (IKOA);

[0060] (6.2) Use the GATv2-Linformer prediction model to predict the charging load prediction value of training set 2, and subtract it from the true value of training set 2 to obtain an error sequence. Use this error sequence to train the XGBoost error prediction model;

[0061] (6.3) Based on the GATv2-Linformer prediction model and the XGBoost error prediction model, the initial prediction value and error correction value of the test set data are obtained, and the final charging load prediction value is obtained by adding them together.

[0062] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0063] A GATv2-Linformer model was constructed, and the graph attention mechanism was introduced into the encoder and decoder of the Linformer. The graph attention mechanism is used to capture the correlation of input features without prior knowledge, and the multi-head attention mechanism is used to capture temporal dependencies. The planet positions are initialized with the good point set strategy, and the eccentricity is initialized with the sobol sequence, so that the planets are evenly distributed in the initial solution space, making the algorithm more comprehensive. At the same time, the Cauchy mutation is used to perturb the planets, thereby expanding the search scale of the Kepler optimization algorithm and improving the algorithm's ability to escape local optimality. The XGBoost model is used to predict the error sequence, and the initial prediction value is corrected with the obtained error correction value to obtain a more accurate prediction output. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 It is a structural schematic diagram of the present invention;

[0065] Figure 2 Flowchart for improving Kepler optimization algorithm;

[0066] Figure 3 Flowchart of optimizing the GATv2-Linformer model for IKOA;

[0067] Figure 4 It is the GATv2-Linformer model block. DETAILED DESCRIPTION

[0068] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0069] like Figure 1 As shown, the present invention provides a method for evaluating and predicting factors affecting electric vehicle charging load based on combined weights, comprising the following steps:

[0070] (1) Obtain the historical charging load data of electric vehicles in three charging scenarios: urban residential areas, commercial areas, and industrial parks in advance, pre-process the load data to obtain the annual load curve, and divide the pre-processed charging load data; obtain the historical charging load data of the three scenarios at the same time every day in urban residential areas, commercial areas, and industrial parks, and clean the obtained historical charging load of electric vehicles to prevent data errors caused by equipment instability, system failure, etc.; draw the annual load curve according to the chronological order of the historical charging load data.

[0071] Since the date type has a huge impact on the charging load and is not conducive to collation and prediction, the present invention divides the obtained historical data into three types, namely, working day charging load data, non-working day charging load data and holiday charging load data, and divides the three types of charging load data into training set 1, training set 2 and test set respectively, and divides the charging load data into training set 1, training set 2 and test set in a ratio of 6:2:2.

[0072] (2) Construct an index system based on the factors affecting charging load prediction in different scenarios, calculate the index combination weights through subjective weights and improved entropy weight method, and use the characteristic values ​​of the factors with higher weights as input and the charging load as output;

[0073] (2.1) Establish an influencing index system under multiple scenarios. In view of the different factors affecting charging load in different scenarios, a combined weight index based on subjective weight and objective weight is constructed to analyze the weight of each influencing factor in different scenarios. In view of the different situations of urban residential areas, commercial areas and industrial parks, five factors that have a greater impact on the charging load are selected. There are more factors that affect the prediction of electric vehicle charging load. The present invention gives priority to factors that have a greater impact on the prediction of charging load in the studied area, namely temperature, precipitation, charging mode, number of charging piles and number of electric vehicles. The characteristics of each influencing factor are used as input values, and the charging load is used as output value. Obtain the data of each influencing factor. Since the scale and order of magnitude of each type of data are different, the data of each influencing factor are normalized. Enable the learning model to converge when the gradient model descends, and at the same time improve the prediction accuracy.

[0074] (2.2) Subjective weights are determined by expert experience. The subjective weights are weighted based on comprehensive research data and experience. The weights of influencing factors are shown in the following table:

[0075]

[0076]

[0077] The numbers 1.0-2.2 represent the degree of influence of each influencing factor on the charging load prediction.

[0078] (2.3) The objective weight is calculated by the improved entropy weight method. On the basis of the original entropy weight method, the information entropy principle is introduced. The information entropy calculation formula is:

[0079]

[0080] Where p(j) is the probability of event j occurring, and N is the total number of samples.

[0081] Calculate the weight index of each influencing factor using the following formula:

[0082]

[0083] in is the mean value of information entropy not equal to 1, H j is the information entropy value corresponding to the jth parameter;

[0084] The entropy weight is calculated by the weight index:

[0085]

[0086] The index a is subjectively defined as 8, The information entropy weight is calculated using the entropy value for the traditional information entropy weight.

[0087] (2.4) Based on the multi-factor impact of charging load prediction, the combined weight index is constructed as follows:

[0088]

[0089] where α j is the subjective weight.

[0090] (3) Construct a GATv2-Linformer hybrid prediction model; introduce the GATv2 graph attention mechanism into the Linformer attention layer to construct a GATv2-Linformer hybrid prediction model, such as Figure 4 shown.

[0091] (3.1) Compared to the original GAT, GATv2 is able to compute dynamic attention, which means that for each query node, the attention function is not monotonic with respect to the neighbor scores, but can vary depending on the query node. This dynamic feature enables the model to better adapt to the representation and learning of different types of node features. GATv2 modifies the order of operations based on GAT.

[0092]

[0093] Among them, || represents the splicing operation, w gat represents the shared weight matrix, h i，t Represents the characteristics of the node, a T is the attention kernel function, k is the activation function, α ij，t is the weight coefficient at time t.

[0094] Compared with GAT, GATv2 first performs feature concatenation when calculating weight coefficients, and then multiplies them with the weight matrix. To enhance learning ability, a multi-head attention mechanism is adopted:

[0095]

[0096] Among them, W k is the learning parameter, α ij，k is the learned weight coefficient, σ is any activation function, It is the feature vector obtained by summing the single-head attention coefficients.

[0097] (3.2) Linformer is a variant of transformer. It uses linear attention mechanism instead of traditional self-attention mechanism based on classic transformer. The main idea of ​​linear attention mechanism is to project the original input sequence into a low-dimensional subspace, calculate the relevance score in the subspace, and then map the score back to the original space. In this way, the linear attention mechanism can preserve the relevance information between sequences while reducing the computational complexity. The time complexity of transformer self-attention mechanism is O(n 2 ), the calculation method is:

[0098]

[0099] Among them, Q, K, and V all come from the input features themselves, and their dimensions are n*d, W Q 、W K 、W V is the learned matrix, QW Q ∈R n ×d , KW K ∈R n×d 、VW V ∈R n×d , where n is the sequence length and d is the hidden dimension, then the calculation of softmax() here will be O(n 2 ), this complexity is not friendly to long sequence prediction.

[0100] Linformer adds two k×n matrices E and F to the attention calculation, reducing the n-order attention dimension to a fixed k-dimensional matrix:

[0101]

[0102] Furthermore, we can reduce the number of parameters, that is, let E = F, and subjectively define the dimension of k as 128.

[0103] (4) Based on the Kepler optimization algorithm KOA, multiple improvement strategies are integrated to obtain IKOA, and IKOA is used to optimize the parameters of the GATv2-Linformer prediction model; Figure 3 As shown in the figure, the objective function of the Kepler optimization algorithm is set as the learning parameters of GATv2 in the GATv2-Linformer model, the number of encoding and decoding layers of Linformer, and these three parameters are used as the input of the algorithm, with upper and lower limits set; the fitness value is the training set error index;

[0104] (4.1) Figure 2As shown in the figure, the initialization of the original Kepler optimization algorithm is random initialization, so the population distribution is random, which easily causes KOA to fall into local optimality. The present invention uses the good point set strategy initialization instead of the original random initialization:

[0105]

[0106] Its deviation satisfies:

[0107]

[0108] Among them, P n (k) is the set of good points, r is the good point, n is the number of points, C(r, δ) is a constant that is only related to r and δ, δ is an arbitrarily small positive number, p is the smallest prime number that satisfies (p-3) / 2>s, s represents the dimension, upper represents the upper bound, and lower represents the lower bound.

[0109] (4.2) Initialize the orbital eccentricity e and orbital period T. Use the Sobol sequence to initialize the orbital eccentricity e of the algorithm:

[0110]

[0111] Among them S i is the i-th random number generated by the sobol sequence, S i ∈[0, 1], r is a random number that conforms to the normal distribution.

[0112] (4.3) The speed of celestial bodies is affected by the gravitational pull of the Sun. When a planet approaches the Sun, the gravitational pull increases, and the planet increases its speed to prevent being pulled toward the Sun. Conversely, when it moves away from the Sun, the gravitational pull decreases, and the planet slows down. The relationship is as follows:

[0113]

[0114]

[0115]

[0116] M=(r3×(1-r4)+r4) (15)

[0117]

[0118]

[0119]

[0120]

[0121]

[0122]

[0123] Among them, V i (t) is the speed required for i to reach the new position; X i is the planet position, X a 、X b represents a solution randomly selected from the population, represents a solution randomly selected from the population; R i-norm (t) is R i Normalization, R i is the Euclidean distance between the planet and the sun; F is a flag to change the search direction; μ(t) is a constant that decreases with time to control the search accuracy; a i Represents celestial body m i The semi-major axis of the elliptical orbit; M s 、m i represents the mass of the sun and planets; L represents the step size; r3 and r4 are randomly generated values ​​ranging from 0 to 1; r5 is a vector containing random values ​​between 0 and 1; is a vector; is the vector form of M, where M represents mass;

[0124] (4.4) Update planetary positions:

[0125]

[0126] in, Represents universal gravitation, and r1 is a randomly generated value between 0 and 1. Its purpose is to provide more variations for the gravitation value during the optimization process.

[0127] (4.5) Update the distance between the planet and the sun:

[0128]

[0129] Among them, h is a parameter that changes with time and determines whether the algorithm focuses on optimizing the development operator or the exploration operator.

[0130] (4.6) The algorithm uses an elite strategy in position update:

[0131]

[0132] Introducing Cauchy mutation into the position update strategy increases population diversity, improves the algorithm's global search capability, and expands the search space:

[0133] x newbest =x best +x best ×Cauchy(0,1) (25)

[0134] Calculate the fitness value of the latest planet position, update the global optimal solution in the iterative process, take the position of the planet corresponding to the optimal fitness value as the output, and update the learning parameters of the GATv2-Linformer prediction model and the number of encoding and decoding layers of the Linformer.

[0135] (5) Build the XGBoost error correction model;

[0136] (5.1) Use the XGBoost algorithm to build an error prediction model. The core idea of ​​XGBoost is to gradually improve the prediction performance of the model by connecting multiple decision trees in series:

[0137]

[0138] Among them, x i represents the i-th sample, f k (x) is the calculation result of the k-th tree model, and n is the number of decision trees. Compared with GBDT, XGBoost has made some changes in the loss function and objective function. The objective function is as follows:

[0139]

[0140] Where l() is the loss function; Ω() is the regularization term, representing the complexity of the tree; y i is the true value of the observation, is the predicted value, K is the total number of trees in the algorithm; g i 、h i It is known in the algorithm process; γ and λ represent parameters; T represents the number of leaf nodes, ω j Represents the weight of the j-th leaf node.

[0141] (5.2) Build the XGBoost error prediction model:

[0142] Prepare the training set 2 dataset and define the parameters of the XGBoost error prediction model, including the maximum number of iterations, tree depth, number of trees, and learning rate. Use the training set to train the XGBoost model. During training, the model gradually learns how to fit the data and optimizes the prediction results in each iteration until the maximum number of iterations is reached and the optimal prediction value is output. For the model parameters, the present invention sets the maximum number of iterations to 1000, the number of trees to 150, the tree depth to 3, and the learning rate to 0.1.

[0143] (6) The improved GATv2-Linformer prediction model and XGBoost error correction model are used to obtain the initial prediction value and error prediction value respectively. The prediction error value is used to correct the generated prediction value to obtain a more accurate electric vehicle charging load prediction result.

[0144] (6.1) Optimize the parameters of the GATv2-Linformer prediction model using the data from training set 1 and the improved Kepler optimization algorithm (IKOA);

[0145] (6.2) Use the GATv2-Linformer prediction model to predict the charging load prediction value of training set 2, and subtract it from the true value of training set 2 to obtain an error sequence. Use this error sequence to train the XGBoost error prediction model;

[0146] (6.3) Based on the GATv2-Linformer prediction model and the XGBoost error prediction model, the initial prediction value and error correction value of the test set data are obtained, and the final charging load prediction value is obtained by adding them together.

Claims

1. A method for evaluating and predicting factors affecting electric vehicle charging load based on combined weights, characterized in that: The steps include: (1) Obtain the historical charging load data of electric vehicles in three charging scenarios: urban residential areas, commercial areas, and industrial parks, pre-process the load data, obtain the annual load curve, and divide the pre-processed charging load data; (2) Construct an index system based on the factors affecting charging load prediction in different scenarios, calculate the index combination weights through expert experience and improved entropy weight method, and use the characteristic values ​​of the factors with higher weights as input and the charging load as output; (3) Introducing the graph attention mechanism of GATv2 into the Linformer attention layer to build a GATv2-Linformer hybrid prediction model; (4) Based on the Kepler optimization algorithm KOA, multiple improvement strategies are integrated to obtain IKOA, and IKOA is used to optimize the parameters of the GATv2-Linformer prediction model; (5) Build the XGBoost error correction model; (6) The improved GATv2-Linformer prediction model and XGBoost error correction model are used to obtain the initial prediction value and error prediction value respectively, and the prediction error value is used to correct the generated prediction value to obtain a more accurate electric vehicle charging load prediction result; The step (4) of optimizing the GATv2-Linformer prediction model parameters using IKOA includes: (4.1) Use the good point set strategy initialization instead of the original random initialization: Its deviation satisfies: Among them, P n (k) is a set of good points, r is a good point, n is the number of points, C(r, δ) is a constant that is only related to r and δ, δ is an arbitrarily small positive number, p is the smallest prime number that satisfies (p-3) / 2>s, s represents the dimension, upper represents the upper bound, and lower represents the lower bound; (4.2) Using the elitist strategy in position update: Introducing Cauchy mutation into the position update strategy increases population diversity, improves the algorithm's global search capability, and expands the search space: Calculate the fitness value of the latest planet position, update the global optimal solution in the iterative process, take the position of the planet corresponding to the optimal fitness value as the output, and update the learning parameters of the GATv2-Linformer prediction model and the number of encoding and decoding layers of the Linformer.

2. The method for evaluating and predicting factors affecting electric vehicle charging load based on combined weights according to claim 1, characterized in that: The step (1) divides the obtained historical data into three types, namely working day charging load data, non-working day charging load data and holiday charging load data, and divides the three types of charging load data into training set 1, training set 2 and test set respectively.

3. The method for evaluating and predicting factors affecting electric vehicle charging load based on combined weights according to claim 1, characterized in that: The step (2) comprises: (2.1) Establish an impact index system for multiple scenarios. Considering the different factors affecting charging load in different scenarios, a combined weight index based on expert experience and improved entropy weight method is constructed to analyze the weights of each influencing factor in different scenarios. (2.2) Weight indicators based on expert experience are weighted by comprehensive research data and experience; (2.3) Weight index based on improved entropy weight method Based on the original entropy weight method, the information entropy principle is introduced, and the information entropy calculation formula is: Where p(j) is the probability of event j occurring, and N is the total number of samples; Calculate the weight index of each influencing factor using the following formula: in is the mean value of information entropy not equal to 1, H j is the information entropy value corresponding to the jth parameter; The entropy weight is calculated by the weight index: where the exponent a is defined as 8, The information entropy weight calculated using entropy value is used for traditional information entropy weight; (2.4) Based on the multi-factor impact of charging load prediction, the combined weight index is constructed as follows: where α j is the subjective weight.

4. The method for evaluating and predicting factors affecting electric vehicle charging load based on combined weights according to claim 1, characterized in that: The step (3) GATv2 modifies the order of operations based on GAT, which is expressed as: Among them, || represents the splicing operation, w gat represents the shared weight matrix, h i，t Represents the characteristics of the node, a T is the attention kernel function, L is the activation function, α ij,t is the weight coefficient at time t; When calculating the weight coefficient, GATv2 first performs feature concatenation and then multiplies it by the weight matrix, using a multi-head attention mechanism: Among them, W k To learn the parameters, is the feature vector obtained by summing the single-head attention coefficients, α ij,k is the learned weight coefficient, and σ is an arbitrary activation function.

5. The method for evaluating and predicting factors affecting electric vehicle charging load based on combined weights according to claim 1, characterized in that: In step (3), the Linformer attention layer adds two k×n matrices E and F to the attention calculation, so that the n-order attention is reduced to a fixed k-dimensional matrix: Reduce the number of parameters, set E = F, and define the dimension of k as 128.

6. The method for evaluating and predicting factors affecting electric vehicle charging load based on combined weights according to claim 1, characterized in that: Step (5) uses the XGBoost algorithm to build an error prediction model. XGBoost gradually improves the prediction performance of the model by connecting multiple decision trees in series: Among them, x i represents the i-th sample, f k (x) is the calculation result of the k-th tree model, n is the number of decision trees, and XGBoost performs some processing in the loss function and objective function. The objective function is as follows: Among them, l( ) is the loss function; Ω( ) is the regularization term, representing the complexity of the tree; y i is the true value of the observation, is the predicted value, K is the total number of trees in the algorithm; g i 、h i It is known in the algorithm process; γ and λ represent parameters; T represents the number of leaf nodes, ω j Represents the weight of the j-th leaf node.

7. The method for evaluating and predicting factors affecting electric vehicle charging load based on combined weights according to claim 1, characterized in that: The step (5) comprises: (5.1) Prepare the training set 2 dataset and define the parameters of the XGBoost error prediction model, including the maximum number of iterations, tree depth, number of trees, and learning rate; (5.2) Use the training set to train the XGBoost model. During the training process, the model will gradually learn how to fit the data and optimize the prediction effect in each round of iteration to reach the maximum number of iterations and output the optimal prediction value.

8. The method for evaluating and predicting factors affecting electric vehicle charging load based on combined weights according to claim 7, characterized in that: The maximum number of iterations is 1000, the number of trees is set to 150, the depth of the tree is set to 3, and the learning rate is 0.

1.

9. The method for evaluating and predicting factors affecting electric vehicle charging load based on combined weights according to claim 1, characterized in that: The step (6) comprises: (6.1) Optimize the parameters of the GATv2-Linformer prediction model using the data from training set 1 and the improved Kepler optimization algorithm (IKOA); (6.2) Use the GATv2-Linformer prediction model to predict the charging load prediction value of training set 2, and subtract it from the true value of training set 2 to obtain an error sequence. Use this error sequence to train the XGBoost error prediction model; (6.3) Based on the GATv2-Linformer prediction model and the XGBoost error prediction model, the initial prediction value and error correction value of the test set data are obtained, and the final charging load prediction value is obtained by adding them together.

Images