CN117081063A - A distributed charging load prediction method and system based on GCN-Crossformer model - Google Patents

Abstract

The invention discloses a distributed charging load prediction method and system based on the GCN-Crossformer model. First, historical charging load data and influencing factor data are obtained, and the data is cleaned; the contribution degree of each influencing factor to charging load prediction is calculated. Analyze and screen the importance of each influencing factor; divide the screened influencing factor data and historical charging load data into training sets and verification sets; build a GCN‑Crossformer charging load prediction model and train it through the training set and verification set , optimize the learning rate and number of attention heads of the trained GCN‑Crossformer charging load prediction model; predict the charging load by using the optimized GCN‑Crossformer charging load prediction model. A deep learning fusion model that combines graph convolutional network and Crossformer structure is adopted. This fusion model can make full use of the spatial relationship between charging stations and the time characteristics of the charging load to achieve more accurate distributed charging load prediction.

Description

A distributed charging load prediction method based on GCN-Crossformer model and system

Technical field

The present invention relates to charging load prediction, specifically to a distributed charging load prediction method and system based on the GCN-Crossformer model.

Background technique

In recent years, with the popularity of electric vehicles and the rapid development of charging facilities, distributed charging load prediction has become an important research field. Load forecasting is of vital significance to the reliable operation, maintenance and efficient management of smart grids. Accurate prediction of charging station load is an important measure to improve the safe operation and economic benefits of charging stations. For grid operators, accurate prediction of charging load can help them rationally dispatch power resources and ensure the stable operation of the grid.

At present, charging load prediction methods mainly include mathematical statistical models and neural network models. Common statistical models mainly include ARMA, SVM, multiple linear regression models, etc. However, these models are often ineffective when dealing with complex and nonlinear charging load data. They face the problem of low computational efficiency and limitations when dealing with large-scale charging networks. its promotion in practical applications. Neural network models such as RNN, CNN, and LSTM have powerful computing and learning capabilities. They can extract deep features of data and solve multi-variable, nonlinear complex modeling problems. In addition, deep learning methods combined with graph neural networks have made certain progress in charging load prediction, but there are still some difficulties and shortcomings. Many existing methods only focus on the extraction of spatial features or time features, ignoring the relationship between the two. interactions, which may lead to model performance degradation in the face of complex charging load changes.

Contents of the invention

Purpose of the invention: In view of the above shortcomings, the present invention provides a distributed charging load prediction based on the GCN-Crossformer model that improves the accuracy of distributed charging load prediction of electric vehicles, facilitates grid managers to perform energy scheduling and optimization, and improves energy utilization efficiency. Methods and systems.

Technical solution: In order to solve the above problems, the present invention adopts a distributed charging load prediction method based on the GCN-Crossformer model, which includes the following steps:

(1) Obtain historical charging load data and influencing factor data, and perform data cleaning;

(2) Calculate the contribution of each influencing factor to charging load prediction, analyze and screen the importance of each influencing factor; and divide the screened influencing factor data and historical charging load data into a training set and a verification set;

(3) Construct a GCN-Crossformer charging load prediction model, input the training set and verification set in step (2) into the GCN-Crossformer charging load prediction model for training, and extract spatial features from the data through the graph convolution network GCN. Capture the dependence of data on time series through the Crossformer module;

(4) Optimize the learning rate and number of attention heads of the trained GCN-Crossformer charging load prediction model;

(5) Predict the charging load through the optimized GCN-Crossformer charging load prediction model to obtain the final charging load prediction value.

Further, the influencing factors in step (1) include user travel characteristics, user residence, charging station locations, charging equipment distribution, number of electric vehicles, electricity prices, weather and holidays.

Further, in step (2), the XGBoost algorithm is used to calculate the contribution of each influencing factor to the charging load prediction; the structure of the XGBoost algorithm is as follows:

in, _is the label vector of the j-th sample _, indicating the charging load at this _time ; Mapping of nodes; F(x _j ) represents the set space of the basic learner; f _k ( _xi ) is the functional relationship between d(x _j ) and the weight w of the k-th tree, and K represents the total number of trees .

Further, in the step (3), the graph convolution network GCN is used to extract spatial features from the historical charging load data and influencing factor data; the operation formula of the graph convolution network GCN is:

Among them, σ(·) is the activation function; I is the identity matrix, A is the adjacency matrix;/> is a diagonal matrix; H ^l and W ^l are the output and parameter values of the l-th layer respectively.

Further, in step (4), the fitness distance balance strategy FDB and the random walk strategy are used to improve the dung beetle optimization algorithm DBO to obtain the IDBO algorithm; the IDBO algorithm is used to optimize the learning rate and sum of the GCN-Crossformer charging load prediction model. Number of attention heads.

Furthermore, the random walk strategy is used to improve the dung beetle position update in the dung beetle optimization algorithm DBO; the update formula of the dung beetle position is:

_{Among them , _} (t) and _di (t) are the minimum and maximum values of the i-th dimension variable in the t-th iteration respectively.

Further, the fitness distance balance strategy is used to find the candidate solution in the population that makes the greatest contribution to the search process, including the following steps:

(4.31) Calculate the fitness value _fi of the i-th candidate solution in the dung beetle algorithm;

(4.32) Calculate the Euclidean distance between the i-th candidate solution and the current optimal solution in the dung beetle algorithm

(4.33) Normalize the fitness value of the candidate solution and the Euclidean distance from the candidate solution to the optimal solution, and calculate the FDB score S _i of the i-th dimension candidate solution through weighted summation:

Among them, n is the total number of dimensions, ω∈(0,1) is the weight coefficient; norm(·) represents normalization;

(4.34) Based on the FDB score of the candidate solution, the roulette method is used to determine the final selected reference position. The improved position update formula is as follows:

in, is the initial individual randomly selected in the population, e is the scale coefficient, r is a uniformly distributed random number in [0,1], p is the number of dung beetle individuals, that is, the population size; num _p is the minimum distance threshold, X _fdb is FDB Reference individuals determined by the method.

The present invention also adopts a distributed charging load prediction system based on the GCN-Crossformer model, including:

The data acquisition module is used to obtain historical charging load data and influencing factor data, and perform data cleaning;

The data division module is used to calculate the contribution of each influencing factor to charging load prediction, analyze and screen the importance of each influencing factor; and divide the screened influencing factor data and historical charging load data into a training set and a verification set and test set;

The model building module is used to build the GCN-Crossformer charging load prediction model. The training set and verification set in step (2) are input into the GCN-Crossformer charging load prediction model for training, in which the data is spatially processed through the graph convolution network GCN. Feature extraction, capturing the dependence of data on time series through the Crossformer module;

Optimization module, used to optimize the learning rate and number of attention heads of the trained GCN-Crossformer charging load prediction model;

The prediction module is used to predict the charging load through the optimized GCN-Crossformer charging load prediction model to obtain the final charging load prediction value.

Beneficial effects: Compared with the existing technology, the significant advantage of this invention is that it adopts a deep learning fusion model that combines a graph convolution network and a Crossformer structure. Graph convolutional network is a neural network used to process graph-structured data. It can capture the spatial relationship between distributed charging stations; Crossformer can capture the dependence of data in time series. This fusion model can make full use of the spatial relationship between charging stations and the time characteristics of the charging load to achieve more accurate distributed charging load prediction.

The extreme gradient enhanced XGBoost algorithm is used to obtain the contribution of each influencing factor to the charging load prediction. By analyzing the importance of each influencing factor, the influencing factor data with high importance to the charging load is screened out; the GCN-Crossformer charging load can be improved Improve the precision and accuracy of prediction models and reduce unnecessary calculations and data processing.

Combining multiple improved strategies with the IDBO algorithm, the optimal dung beetle fitness value is perturbed by introducing a random walk strategy to improve the local exploration capability of the dung beetle optimization algorithm; the fitness distance balance strategy is used to find the population that makes the greatest contribution to the search process. The candidate solution enables the IDBO algorithm to better handle the differences between different charging needs in distributed charging load prediction and improve the accuracy and stability of the prediction results.

Description of the drawings

Figure 1 shows the overall flow chart of distributed charging load prediction according to the present invention.

Figure 2 shows a schematic flowchart of the XGBoost algorithm in the present invention.

Figure 3 shows a schematic diagram of the GCN model in the present invention.

Figure 4 shows a schematic diagram of the Crossformer model in the present invention.

Detailed ways

As shown in Figure 1, the distributed charging load prediction method based on the GCN-Crossformer model in this embodiment includes the following steps:

(1) Collect historical charging load data and influencing factor data, and perform data cleaning; influencing factors include user travel characteristics, user residence, charging station location, charging equipment distribution, number of electric vehicles, electricity prices, weather and holidays.

(2) Use the XGBoost algorithm to obtain the contribution of each influencing factor to the charging load prediction, analyze the importance of each influencing factor, and screen out the influencing factor data that is highly important to the charging load; combine the screened influencing factor data with The charging load data is divided into training set, verification set and test set;

The implementation process of XGBoost algorithm is as follows:

(2.1) Define the data set D = {(x _i ,y _i )}, and the model containing k trees can be expressed as:

in, _is the label vector of the j-th sample _, indicating the charging load at this _time ; Mapping of nodes; F(x _j ) represents the set space of the basic learner; f _k ( _xi ) is the functional relationship between d(x _j ) and the weight w of the k-th tree, and K represents the total number of trees .

(2.2) Determine the objective function of XGBoost training:

Among them, l() is the loss function and Ω() is the normal function. Ω(f _k ) represents the complexity of the generated regression tree, α is the regularization parameter to avoid overfitting, T is the number of leaf nodes, and β controls the leaf node weight w _j ;

(3) Build the GCN-Crossformer charging load prediction model through the graph convolution network GCN and Crossformer model, and input the training set and verification set obtained in step (2) into the GCN-Crossformer charging load prediction model for training;

The implementation process of constructing the GCN-Crossformer charging load prediction model is as follows:

(3.1) Extract spatial features from historical charging load data and influencing factor data through graph convolution network GCN;

The operation formula of graph convolution is:

In the formula, I is the identity matrix, A is the adjacency matrix;/> is a diagonal matrix; H ^l and W ^l are the output and parameter values of the l-th layer respectively, and σ(·) is the activation function;

Step 4.2: Capture the dependence of data on time series through Crossformer;

Represent the relative position of the embedding by adding a bias to the attention;

In the formula, are vectors in the sub-attention module respectively, /> is the constant normalization formula, and B is the RPB matrix.

(4) Use the fitness distance balance strategy FDB and the random walk strategy to improve the dung beetle optimization algorithm DBO to obtain the IDBO algorithm; use the IDBO algorithm to optimize the learning rate and the number of attention heads of the GCN-Crossformer charging load prediction model to improve The accuracy of model prediction; the implementation process is as follows:

(4.1) Initialize the relevant parameters of the IDBO algorithm, including population, dimension, maximum number of iterations, and upper and lower limits of the search space;

(4.2) Define the location update of the dung beetle;

Dung beetle rolling ball location update:

x _i (t+1)＝xi ₍ t)+α×k×x _i (t-1)+b×Ax,Δx＝|x _i (t)-X ^w | (8)

Among them, x _i (t+1) is the position of the i-th dung beetle at the t-th iteration, k∈(0,0.2) represents the deflection coefficient, b∈(0,1) is a natural coefficient; Δx is the light The degree of change in intensity, X ^w is the worst position within the current population. α=1 means that the natural environment does not affect the original direction, α=-1 means that it deviates from the original direction.

The location of the dancing dung beetle is updated to:

x _i (t+1)＝xi ₍ t)+tan(θ)|x _i (t)-x _i (t-1)| (9)

Among them, θ∈[0,π] is the deflection angle;

The location of dung beetle breeding is updated to:

B _i (t+1)＝X ^* +b ₁ ×(B _i (t)-Lb ^* )+b ₂ ×(B _i (t)-Ub ^* ) (10)

The location of dung beetles feeding has been updated to:

x _i (t+1)＝ _xi (t)+C ₁ ×( _xi (t)-Lb ^b )+C ₂ ×(x _i (t)-Ub ^b ) (11)

The location of Dung Beetle has been updated to:

x _i (t+1)＝X ^b +S×g×(|x _i (t)-X ^* |+|x _i (t)-X ^b |) (12)

Use random walk strategy to improve the location update of dung beetles:

X(t)＝[0,sum(2r(t ₁ )-1),…,sum(2r(t _n )-1)] (13)

Among them, X(t) is the number of steps; t is the maximum number of iterations; r(t) is a random function, defined as:

Among them, the random number rand is between [0,1]; the above formula is normalized. Since there is a boundary in the feasible region, the following formula is used to directly update the position of the dung beetle:

In the formula, a _i is the minimum value of the random walk of the i-th dimension variable; b _i is the maximum value of the random walk of the i-th dimension variable; c _i (t) and _di (t) are respectively the i-th dimension variable in the random walk. Minimum and maximum values for t iterations.

(4.3) Use the fitness distance balance strategy to find the candidate solution in the population that makes the greatest contribution to the search process; specifically, it includes the following steps:

(4.31) Calculate the fitness value f _i of the i-th dimension candidate solution in the dung beetle algorithm;

(4.32) Calculate the distance between the i-th dimension candidate solution and the current optimal solution Euclidean distance is used as the index of distance measurement;

(4.33) Perform normalization according to the fitness value of the solution and the distance to the optimal solution, and calculate the FDB score of the i-th dimension candidate solution through weighted summation:

where ω∈(0,1) is the weight coefficient;

(4.34) After obtaining the FDB score S _i , the roulette method is used to determine the final selected reference position. The improved position update formula is as follows:

in is a randomly selected individual in the population, e is the scale coefficient, r is a uniformly distributed random number in [0,1], p is the number of dung beetle individuals, that is, the population size; num _p is the minimum distance threshold, X _fdb is FDB Reference individuals determined by the method.

(4.4) Determine whether the maximum number of iterations has been reached. If so, output the optimal solution, otherwise return to step (4.2).

(5) Use the test set to predict the distributed charging load, verify the reliability of the final GCN-Crossformer charging load prediction model, predict the charging load through the optimized GCN-Crossformer charging load prediction model, and obtain the charging load prediction value.

Claims (10)

1. The distributed charge load prediction method based on the GCN-Crossformer model is characterized by comprising the following steps of:

(1) Acquiring historical charging load data and influence factor data, and cleaning the data;

(2) Calculating the contribution degree of each influence factor to the charge load prediction, and analyzing and screening the importance of each influence factor; dividing the screened influence factor data and the historical charging load data into a training set and a verification set;

(3) Constructing a GCN-cross former charge load prediction model, inputting the training set and the verification set in the step (2) into the GCN-cross former charge load prediction model for training, wherein the spatial feature extraction is carried out on data through a graph rolling network GCN, and the dependence of the data on time sequence is captured through a cross former module;

(4) Optimizing the learning rate and the attention head number of the trained GCN-Crossformer charge load prediction model;

(5) And predicting the charging load through the optimized GCN-Crossformer charging load prediction model to obtain a final charging load prediction value.

2. The distributed charge load prediction method according to claim 1, wherein the influencing factors in step (1) include user travel characteristics, user residence, charging station location, charging device distribution, electric car number, electricity price, weather, and holidays.

3. The method according to claim 1, wherein the contribution degree of each influencing factor to the charge load prediction is calculated by using XGBoost algorithm in the step (2);

the construction of the XGBoost algorithm is as follows:

wherein,a label vector which is the j-th sample and indicates the charging load at that time; x is x _j Representing influencing factors for the features of the j-th sample; f represents a function space formed by a tree; d (x) _j ) Representing characteristic x _j Mapping to leaf nodes; f (x) _j ) Representing a collection space of the base learner; f (f) _k (x _i ) Is d (x) _j ) And the weight w of the kth tree, wherein K represents the total number of the trees.

4. The method according to claim 1, wherein in the step (3), the spatial feature extraction is performed on the historical charging load data and the influence factor data through a graph rolling network GCN; the operational formula of the graph roll network GCN is:

wherein σ (·) is the activation function;i is an identity matrix, A is an adjacent matrix; />Is a diagonal matrix; h ^l And W is ^l The output of the first layer and the parameter value, respectively.

5. The method for predicting the distributed charging load according to claim 1, wherein in the step (4), an adaptability distance balance strategy FDB and a random walk strategy are adopted to improve a dung beetle optimization algorithm DBO, so as to obtain an IDBO algorithm; the learning rate and the attention head number of the GCN-Crossformer charge load prediction model are optimized by using an IDBO algorithm.

6. The method for predicting distributed charging load according to claim 5, wherein the position update of the dung beetles in the dung beetle optimization algorithm DBO is improved by using a random walk strategy; the updating formula of the dung beetle position is as follows:

wherein X is _i (t) represents the position of the ith dimension of the t-th iteration of the dung beetle, a _i Is the minimum value of the random walk of the variable in the ith dimension, b _i Maximum value of random walk for the ith dimension variable; c _i (t) and d _i (t) is the minimum and maximum values, respectively, of the ith dimension variable at the t-th iteration.

7. The method of claim 5, wherein the step of searching for a candidate solution in the population that contributes most to the search process using a fitness distance balancing strategy, comprises the steps of:

(4.31) calculating the fitness value f of the ith dimension candidate solution in the dung beetle algorithm _i ；

(4.32) calculating Euclidean distance between the ith dimension candidate solution and the current optimal solution in the dung beetle algorithm

(4.33) normalizing the fitness value of the candidate solution and the Euclidean distance from the candidate solution to the optimal solution, and calculating the FDB score S of the ith dimension candidate solution by means of weighted summation _i ：

Wherein n is the total number of dimensions, ω∈ (0, 1) is a weight coefficient; norm (·) represents normalization;

(4.34) determining the final selected reference position by a roulette method based on the FDB score of the candidate solution, the modified position update formula being as follows:

wherein,is the initial individual selected randomly in the population, e is the scale factor, r is [0,1]Wherein p is the individual number of the dung beetles, namely the population scale; num (num) _p Is the minimum distance threshold, X _fdb Is a reference individual determined by the FDB method.

8. A prediction system employing the distributed charge load prediction method according to any one of claims 1 to 7, characterized by comprising:

the data acquisition module is used for acquiring historical charging load data and influence factor data and cleaning the data;

the data dividing module is used for calculating the contribution degree of each influence factor to the charge load prediction and analyzing and screening the importance of each influence factor; dividing the screened influence factor data and the historical charging load data into a training set, a verification set and a test set;

the model construction module is used for constructing a GCN-Crossformer charging load prediction model, inputting the training set and the verification set in the step (2) into the GCN-Crossformer charging load prediction model for training, wherein the spatial feature extraction is carried out on data through a graph rolling network GCN, and the dependence of the data on time sequence is captured through the Crossformer module;

the optimization module is used for optimizing the learning rate and the attention head number of the trained GCN-Crossformer charge load prediction model;

the prediction module is used for predicting the charge load through the optimized GCN-Crossformer charge load prediction model to obtain a final charge load prediction value.

9. A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the processor implements the steps of the method according to any one of claims 1 to 7 when the computer program is executed by the processor.

10. A computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, implements the steps of the method of any of claims 1 to 7.