Multi-model fusion battery state of charge prediction method and system based on GA-PSO optimization

By using genetic algorithm and particle swarm mixing algorithm to optimize multiple models in the SOC estimation of lithium-ion batteries, the error, real-time and computational complexity of SOC estimation methods in the prior art is solved, and the accuracy of SOC SOCs in lithium-ion batteries is achieved online Forecasting and efficient calculations.

CN115219906BActive Publication Date: 2024-11-05WUHAN UNIV OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210742084.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-06-27
Publication Date
2024-11-05
Estimated Expiration
2042-06-27

AI Technical Summary

Technical Problem

The existing lithium-ion battery state of charge (SOC) estimation methods have problems such as accumulation of errors, poor real-time performance, high computational complexity and high model accuracy requirements. Especially when processing large amounts of data, neural networks are prone to gradients during backpropagation. Disappearance and gradient explosion problems.

Method used

Using a multi-model fusion method based on genetic algorithm and particle swarm hybrid algorithm optimization, GA-PSO-LSTM, GA-PSO-SVM and GA-PSO-GRU models are constructed, and the hyperparameters of LSTM neural network are optimized through genetic algorithm and particle swarm algorithm. Combining various algorithms with other models to predict SOC to improve the accuracy and efficiency of prediction.

Benefits of technology

Accurate online estimation of lithium-ion battery SOC is realized, prediction efficiency and accuracy are improved, problems of low optimization efficiency of a single algorithm and local optimal solutions are avoided, and the overall performance of the model is improved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115219906B_ABST
    Figure CN115219906B_ABST
Patent Text Reader

Abstract

The present invention discloses a multi-model fusion battery state of charge prediction method and system based on GA-PSO optimization, including: collecting external parameters of lithium batteries to obtain a data set sorted by time; data cleaning and normalization processing, and dividing into training set, verification set and test set; constructing three base models GA-PSO-LSTM, GA-PSO-SVM, A-PSO-GRU, and integrating them into a strong learning model; training the strong learning model to obtain the characteristics of a new training set and a new test set; using the real SOC value as output for the second round of training, using the new test set for prediction, and finally weighted averaging the single prediction value to obtain the final SOC prediction value. The present invention solves the problem of low efficiency of a single genetic algorithm in the later stage and a single particle swarm algorithm falling into a local optimal solution, improves the efficiency and effect of optimization, has the advantages of long-term data memory, and improves the accuracy of prediction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of battery prediction, and specifically relates to a battery state of charge estimation method and system based on optimizing LSTM neural network by hybrid algorithm of genetic algorithm and particle swarm algorithm. Background Art

[0002] As environmental pollution and energy shortage become more serious, new energy electric vehicles are gradually replacing traditional fuel vehicles. Lithium-ion batteries have been widely used as power components of electric vehicles due to their high energy density, long service life, and good environmental performance. In order to avoid the phenomenon of energy waste caused by overcharging and discharging of lithium-ion batteries, it is very important to accurately predict the state of charge (SOC) of the battery. However, SOC, as an internal characteristic of the battery, cannot be directly measured and can only be estimated through some parameters that can be actually measured, such as voltage, temperature, and current.

[0003] At present, there are two main methods for estimating the state of charge of lithium-ion batteries: model-based methods and data-driven methods. Model-based methods mainly include ampere-hour integration method, open circuit voltage method, electrochemical impedance method and Kalman filter method. The ampere-hour integration method estimates the SOC online by integrating the sampling value of the current sensor. It does not have the ability to correct the initial error, and the error accumulation may cause a serious deviation between the estimated value and the actual value; the open circuit voltage method estimates the SOC through the numerical relationship between the open circuit voltage and SOC, and the battery needs to be shelved for a long time, and the real-time performance is poor; the electrochemical impedance method estimates the SOC by measuring the internal impedance of the battery, which is greatly affected by the external environment and difficult to measure online; the Kalman filter method requires the establishment of a complex equivalent circuit model, and the model accuracy requirements are high. Based on the limitations of the traditional model method, the data-driven method has been widely valued because it does not need to consider the complex physical and chemical reactions inside the battery, and only needs to use the historical data of the battery to predict the SOC.

[0004] Data-driven methods are the current mainstream methods, mainly including neural networks, support vector machines and other methods. Support vector machines have good results in processing small and medium-sized data, but when the amount of data is too large, the model calculation becomes complicated, resulting in slower calculation speed and reduced model accuracy. Neural networks have commendable effects when processing large sample data sets. Among them, RNN has a strong processing ability for time series data, so it is often used as one of the SOC estimation methods. However, RNN often has the problem of gradient disappearance and gradient explosion during the back propagation process. LSTM can solve this shortcoming well, but the LSTM network structure is changeable and it is difficult to find the optimal hyperparameters. Summary of the invention

[0005] In order to solve the shortcomings of the above-mentioned prior art methods for battery SOC estimation and prediction, the present invention proposes a multi-model fusion battery state of charge prediction method and system based on GA-PSO optimization, which improves the prediction efficiency and accuracy and can realize accurate online estimation of battery SOC.

[0006] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0007] A multi-model fusion battery state of charge prediction method based on GA-PSO optimization, characterized by comprising the following steps:

[0008] S1: Collect the external parameters of the lithium battery and calculate the actual battery SOC value; finally obtain a dataset of lithium battery current, voltage, temperature, and SOC value original features with labels sorted by time;

[0009] S2: Data cleaning: Data analysis is performed on the missing values, duplicate values, and outliers in the data set. Interpolation is performed on a small number of missing values, duplicate values ​​are removed, and outliers are deleted. Then, Max-min normalization is performed on the cleaned data, and the normalized sample data is divided into training set, validation set, and test set.

[0010] S3: Construct multiple base models and integrate them into a strong learning model: Construct a GA-PSO-LSTM model, and use a genetic algorithm and a particle swarm hybrid algorithm to optimize and determine the optimal hyperparameters of the LSTM neural network; construct a GA-PSO-SVM model, and optimize and determine the optimal penalty coefficient and Gaussian radial basis width coefficient of the SVM; construct a GA-PSO-GRU model, and optimize and determine the number of hidden layer nodes and learning step size of the GRU;

[0011] S4: Establish new features: Integrate the GA-PSO-LSTM model, GA-PSO-SVM model, and GA-PSO-GRU model containing the optimal hyperparameters obtained in step S3, and use four of the five working conditions as training sets each time, and the data of the other working condition as the validation set, to obtain the predicted values ​​of each model validation set, and use the predicted values ​​of each model validation set as the features of the new training set. At the same time, another working condition is used as the predicted value of the test set, and the weighted average of the predicted values ​​of the test set is used as the features of the new test set;

[0012] S5: SOC prediction: Take the predicted value of each model validation set obtained in step S4 as input, and the true SOC value as output. Select GA-PSO-LSTM, GA-PSO-SVM, and GA-PSO-GRU model training. After the model is trained, take the features of the new test set in step S4 as input to obtain the SOC prediction value of each model of the three models. Finally, take the weighted average of the SOC prediction values ​​of the individual models to obtain the final SOC prediction value, and evaluate the model error.

[0013] Furthermore, in step S1, when the electric vehicle is running, during the battery discharge process, the external parameters of the lithium battery are collected in real time at intervals of 0.1s: current (i), voltage (v), temperature (T), discharge ampere-hour (Ah), and rated capacity (Ah) of the electric vehicle; the actual battery SOC value is calculated by the following formula using the external parameters:

[0014]

[0015] Among them, the SOC value represents the current power of the lithium battery, Q res Indicates the battery's discharge ampere-hour, Q rated Indicates the rated capacity of the battery.

[0016] Furthermore, the step S1 finally obtains the lithium battery data set Df with the original features of each external parameter sorted by time and labeled:

[0017] Df={(Df i1 ,Df v1 , Df t1 , Df s1 ), (Df i2 , Df v2 , Df t2 , Df s2 ),…,(Df in , Df vn , Df tn , Df sn )};

[0018] Where n represents the number of samples in the collected data set, Df ik , Df vk , Df tk , Df sk They respectively represent the lithium battery current, voltage, temperature, and SOC value data of the kth sample, where k is between 1 and n.

[0019] Furthermore, in step S2, the normalization processing formula for normalizing the deviation of the current, voltage, temperature and true SOC value of the processed battery data is as follows:

[0020]

[0021] Among them, Df * Represents the normalized lithium battery current, voltage, temperature, and SOC value data, Df i Indicates the lithium battery current, voltage, temperature, and SOC value data before normalization, Df i min represents the minimum sample value of lithium battery current, voltage, temperature, and SOC value, Df i max represents the maximum sample value of lithium battery current, voltage, temperature, and SOC value.

[0022] Furthermore, in step S2, the training set, validation set and test set are divided according to the ratio: the normalized data Df * 64% of the total set is used as the training set, Df * 16% of the total set is used as the validation set, Df * The last 20% is used as the test set.

[0023] Furthermore, in step S2, Df * Contains SOC data for multiple working conditions. The SOC values ​​of the training set, validation set, and test set all decrease from 1 to 0.

[0024] Furthermore, in step S3, the various hyperparameters of the LSTM neural network include: the number of LSTM layers, the number of hidden layer Dense layers, and the number of nodes of each layer of neurons; the genetic algorithm and particle swarm hybrid algorithm optimizes the LSTM neural network model in the parameter search space, taking the minimum SOC prediction error as the objective function, and performs optimal combination of various hyperparameters. The parameter search space is specifically as follows: the lower limit of the number of LSTM and Dense layers is 1, and the upper limit is 3; the lower limit of the number of neuron nodes in each layer is 32, and the upper limit is 256.

[0025] Furthermore, in step S3, constructing the GA-PSO-LSTM model includes the following steps:

[0026] Step S31, randomly initialize the population;

[0027] Step S32, setting the fitness function to the inverse of the mean square error of the LSTM neural network, and using the fitness function to calculate the fitness of the initial population;

[0028] Step S33, performing particle swarm evolution operation on all individuals of the population, updating the speed and position of the particles, and generating offspring population 1; at the same time, performing inheritance, crossover, and mutation operations on all individuals of the population to obtain offspring population 2;

[0029] Step S34, sorting the offspring population 1 and the offspring population 2 according to the fitness value from large to small, and selecting the one with the highest fitness as the offspring population 3;

[0030] Step S35, setting the termination condition to the number of iterations of the population. If the termination condition is met, the individual with the best fitness is output, and the optimization process ends. At this time, the LSTM network is the global optimal combination; otherwise, proceed to step S32.

[0031] Furthermore, in step S3, the method for constructing the GA-PSO-SVM model and the GA-PSO-GRU model is the same as that of the GA-PSO-LSTM model. The hyperparameters for optimizing the GA-PSO-SVM model are the penalty coefficient C and the Gaussian radial basis width coefficient σ. The search space of the penalty coefficient C is (0, 1000], and the search space of the Gaussian radial basis width coefficient σ is (0.01, 100]; the hyperparameters for optimizing the GA-PSO-GRU model are the number of hidden layer neurons and the learning step size; the number of hidden layer neurons ranges from (32 to 256), and the learning step size ranges from (10 to 200).

[0032] Furthermore, in step S4, the operating condition simulates the vehicle operating conditions in the area, including at least acceleration climbing, deceleration, and constant speed; each operating condition constitutes a battery data set, and the SOC of each battery data set is a data set that decreases from 1 to 0. All the operating condition data sets are spliced ​​together to form the entire battery data set.

[0033] Furthermore, in step S4, the entire battery data set contains data of 6 working conditions: data1, data2, data3, data4, data5, data6, among which data6 remains unchanged as a test set, and the remaining 5 are used as training sets and validation sets. The training set and validation set change according to the number of training times: in the first training, data1 is used as a validation set, and the remaining 4 are used as training sets. In the second training, data2 is used as a validation set, and the remaining 4 are used as training sets. And so on, the training is repeated 5 times. In the fifth training, data5 is used as a validation set, and the remaining 4 are used as training sets.

[0034] Furthermore, the training process of the GA-PSO-LSTM model, the GA-PSO-SVM model, and the GA-PSO-GRU model in step S4 is as follows:

[0035] Each training takes the current, voltage, and temperature included in the training set as input, and the true SOC value as output. Each training will obtain three trained models: GA-PSO-LSTM model, GA-PSO-SVM model, and GA-PSO-GRU model. For the trained models obtained after each training, the current, voltage, and temperature in the validation set and test set corresponding to each training are used as inputs and input into the corresponding trained models to obtain the output validation set prediction values ​​and test set prediction values.

[0036] Finally, the validation set prediction values ​​obtained by the GA-PSO-LSTM model after five trainings are concatenated in time series as a feature A1, the test set prediction values ​​obtained by the GA-PSO-SVM model after five trainings are concatenated in time series as feature A2, and the test set prediction values ​​obtained by the GA-PSO-GRU model after five trainings are concatenated in time series as feature A3; each data in A1 is added and divided by 5 to obtain the test set feature B1; each data in A2 is added and divided by 5 to obtain the test set feature B2; each data in A3 is added and divided by 5 to obtain the test set feature B3; the obtained test set features B1, B2, and B3 are weighted averaged to obtain the SOC prediction value B0 of the test set.

[0037] Furthermore, in step S5, the final SOC prediction value is the weighted average of the prediction results of the LSTM, SVM, and GRU models; the mean square error (MSE), root mean square error (RMSE), and R 2 The coefficient is used as the evaluation index of the model. When the MSE and RMSE values ​​are smaller, the accuracy of the model is higher; when R 2 The closer the coefficient is to 1, the more accurate the model's predictions are:

[0038] Mean Square Error:

[0039] Root Mean Square Error:

[0040] R 2 coefficient:

[0041] Where n is the number of data sets, y i is the real SOC value, is the SOC value predicted by the model, is the average SOC of all samples.

[0042] Furthermore, in step S5, A1, A2, and A3 are used as new feature inputs for the second round of training, and the actual SOC value is used as the output. The models with the optimal hyperparameters obtained in step S3 are used for training respectively to obtain three models GA-PSO-LSTM, GA-PSO-SVM, and GA-PSO-GRU; finally, B0 is input into the trained model to obtain the predicted values ​​predict1, predict2, and predict3 of the three models GA-PSO-LSTM, GA-PSO-SVM, and GA-PSO-GRU, and the obtained predicted values ​​are weighted averaged to obtain the final model predicted value of SOC.

[0043] The root mean square error (RMSE) is used as the loss function for model training. The bathsize of the training set data is set to 32, epochs to 50, and the optimization algorithm is selected as Adam for gradient calculation. The network is iteratively updated to continuously update the weights of the model to minimize the prediction error RMSE. The RMSE calculation formula is as follows:

[0044]

[0045] Where n is the number of data sets, y i is the real SOC value, is the SOC value predicted by the model.

[0046] Furthermore, in step S5, the process of predicting the test set using the trained GA-PSO-LSTM model is as follows: the predicted value B0 of the first round of training of the test set in step S4 is input as a new feature into the LSTM neural network for forward calculation and output h(t):

[0047] Forget gate: f t =σ(W f x t +U f h t-1 +b f );

[0048] Input gate: i t =σ(W i x t +U i h t-1 +b i );

[0049] Unit status: g t =tanh(W g x t +U g h t-1 +b g );

[0050] Output gate: o t =σ(W o x t +U o h t-1 +b o );

[0051] Current status: c t =f t *c t-1 +i t *g t ;

[0052] Output: h t =ot tanh(c t )

[0053] Among them, f t Indicates the information to be discarded at the tth time node of the battery data, i t Indicates the information to be updated at the tth time node of the battery data, g t represents the cell state of the battery data at the tth time node, o t Indicates the information to be output at the tth time node of the battery data, c t represents the current state of the battery data at the tth time node, h t represents the output value of the battery data at the tth time node; W i , W f , W o , W g are the weight matrices of the input gate, forget gate, output gate, and cell state respectively; U i , U f , U o , U g is also the corresponding weight matrix, b i , b f , b o , b g is the corresponding bias term; h t-1 is the hidden layer state at time t-1; x t is the input vector at time t; tanh is the hyperbolic tangent activation function, and σ is the sigmoid function;

[0054] The predicted value B0 of the first round of training of the test set is used as a new feature to input the LSTM neural network for forward calculation, and the predicted value predict1 of the SOC of the LSTM model is output.

[0055] Furthermore, in step S5, the method of training the GA-PSO-SVM model is specifically as follows:

[0056] The new features A1, A2, A3 (x1, x2, ... x n ) as input, the actual SOC value (y1,y2,…y n ) as the output, and seek the regression function fi(·) such that:

[0057] y=f(x)=wφ(x)+b;

[0058]

[0059] Use the training data to find the appropriate vectors w and b so that the regression error of the data for the hyperplane is minimized:

[0060]

[0061] sty-f(x)≤ε+ξ

[0062] yf(x)≤ε+ξ *

[0063] ξ,ξ * ≥0;

[0064]

[0065] Where ξ and ξ* are the relaxation variables of the two boundaries respectively, C is the penalty coefficient, which is used to control the weight between the objective function "optimization" and "minimum deviation", K is the kernel function, and RBF radial basis function is preferably used as the kernel function; j is the eigenvector, σ is the Gaussian radial basis width coefficient;

[0066] The predicted value B0 of the first round of training of the test set in step S4 is input as a new feature into the SVM model containing the optimal parameters w and b for forward calculation, and the predicted value predict2 of the SOC of the SVM model is output.

[0067] Furthermore, in step S5, the GRU model uses the update gate and the reset gate as the core, uses a function to transform the matrix of the current input sequence and the memory variable at the previous moment, and updates it, and inputs the updated value into the update gate to compare and determine the importance of the current state and the state at the previous moment; in the reset gate, determine whether the previous moment needs to be written into the candidate set;

[0068] The specific method of training the GRU model is: the new feature A1(x1, x2,…x n ), A2, A3 as input, the actual SOC value (y1, y2, ... y n ) as the output, and the trained GRU model is obtained; after the model is trained, B0 in the test set is used as the input, and h is calculated through forward propagation t That is the predicted SOC value predict3, and the calculation method is as follows:

[0069]

[0070] Among them, h t-1 refers to the state variable at the previous moment, h t refers to the state variable at that moment, r t refers to the state variable of the update gate, z t refers to the state variable of the reset gate, refers to the current candidate set state, y t refers to the current state, Wo , W z , W r is the output vector and x t ,h t-1 The weights, candidate sets, reset gates, and update gates that make up the matrix multiplication.

[0071] φ is the tanh activation function, σ is the sigmoid function, and the expression is:

[0072]

[0073]

[0074] The present invention also provides a prediction system, characterized in that a program for implementing the above method is stored in the system.

[0075] Compared with the prior art, the present invention introduces a genetic algorithm and a particle swarm fusion algorithm to optimize the LSTM network results, and the beneficial effects are:

[0076] 1. A hybrid algorithm of genetic algorithm and particle swarm algorithm is used to optimize the structure of LSTM neural network, which solves the problem of low efficiency of single genetic algorithm in the later stage and single particle swarm algorithm falling into local optimal solution; the hybrid algorithm has both the good global search ability of genetic algorithm and the fast search ability of particle swarm algorithm, which improves the efficiency and effect of LSTM algorithm optimization.

[0077] 2. Genetic algorithm, particle swarm algorithm and LSTM neural network model are used to predict the battery state of charge. The model has the advantage of long-term data memory, which improves the accuracy of prediction.

[0078] 3. The base learner that integrates multiple algorithms such as LSTM, SVM, and GRU forms a strong learner, which effectively improves the accuracy of prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0080] Figure 1 It is a schematic diagram of the LSTM network of the present invention.

[0081] Figure 2 It is a schematic diagram of the GRU network of the present invention.

[0082] Figure 3 This is a chromosome explanation diagram of the present invention.

[0083] Figure 4 This is an explanatory diagram of the characteristic engineering of the fusion model of the present invention.

[0084] Figure 5 It is the overall flow chart of the present invention.

[0085] Figure 6 These are some verification results of the present invention on the Panasonic battery dataset. DETAILED DESCRIPTION

[0086] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0087] like Figure 1-6 A battery state of charge prediction method and system based on a multi-model fusion (optimizing LSTM network structure based on genetic algorithm and particle swarm fusion algorithm) optimized by GA-PSO according to the present invention. Figure 5 A method or network flow chart comprising the following steps:

[0088] S1: Data collection. During the battery discharge process, relevant sensors are used to collect characteristic parameters of the lithium battery for estimating the battery SOC (collecting data of 6 working conditions), such as current (i), voltage (v), temperature (T), discharge ampere-hour (Ah), and rated capacity (Ah) of the electric vehicle.

[0089] S2: Data cleaning. Perform data analysis on missing values, duplicate values, and outliers in the collected data. Interpolation is used for a small number of missing values, duplicate values ​​are removed, and outliers are deleted. Then, the cleaned data is normalized using Max-min, and the normalized sample data is divided into training set, validation set, and test set.

[0090] S3: Build multiple base models and integrate them into a strong learning model. Build a GA-PSO-LSTM model, use a genetic algorithm and a particle swarm hybrid algorithm to optimize the optimal hyperparameters of the LSTM neural network; build a GA-PSO-SVM model, optimize the optimal penalty coefficient and Gaussian radial basis width coefficient of the SVM; build a GA-PSO-GRU model, and optimize the number of hidden layer nodes and learning step size of the GRU.

[0091] S4: Feature engineering, establish new features. Integrate the LSTM model, SVM model, and GRU model containing the optimal hyperparameters obtained in step S3, keep one working condition data as the test set, use four of the five working conditions as the training set each time, and use the data of the other working condition as the validation set, and obtain the predicted values ​​val_pre_lstm, val_pre_svm, and val_pre_gru of each model in the validation set. Use the predicted values ​​of the validation sets of the three models, LSTM model, SVM model, and GRU model, as the features of the new training set. At the same time, the predicted values ​​of the test set test_pre_lstm, test_pre_svm, and test_pre_gru are also obtained, and the weighted average of the results of test_pre_lstm, test_pre_svm, and test_pre_gru is used as the features of the new test set.

[0092] S5: SOC prediction. Take the predicted value features val_pre_lstm, val_pre_svm, val_pre_gru obtained in step S4 as input, take the true label SOC value as output, select GA-PSO-LSTM, GA-PSO-SVM, GA-PSO-GRU model training, and after the model is trained, take test_pre_lstm, test_pre_svm, test_pre_gru in step S4 as input to obtain the predicted SOC value of the three models. Finally, take the weighted average of the SOC predicted values ​​of the individual models to obtain the final predicted SOC value, and evaluate the model error.

[0093] Furthermore, the step S1 is specifically as follows: during the operation of the electric vehicle, the external parameters of the lithium battery including current (i), voltage (v), temperature (T), discharge ampere-hour (Ah), and rated capacity (Ah) of the electric vehicle are collected in real time at intervals of 0.1s each time. Since the SOC value cannot be directly measured, the actual SOC value is calculated using the following formula:

[0094]

[0095] Among them, SOC represents the current power of the lithium battery, Q res Indicates the battery's discharge ampere-hour, Q rated Indicates the rated capacity of the battery.

[0096] Finally, a dataset Df with three original features of lithium battery current, voltage, temperature, and SOC value sorted by time and labeled is obtained. Df = {(Df i1 ,Df v1 , Df t1 , Df s1 ), (Df i2 , Dfv2 , Df t2 , Df s2 ),…,(Df in , Df vn , Df tn , Df sn )}, where n represents the number of samples in the collected data set, Df ik , Df vk , Df tk , Df sk They respectively represent the lithium battery current, voltage, temperature, and SOC value data of the kth sample.

[0097] Furthermore, in step S2, the current, voltage, temperature and real SOC value of the processed battery data are normalized by deviation standardization, and the calculation formula is as follows:

[0098]

[0099] Among them, Df * Represents the normalized lithium battery current, voltage, temperature, and SOC value data, Df i Indicates the lithium battery current, voltage, temperature, and SOC value data before normalization, Df i min represents the minimum sample value of lithium battery current, voltage, temperature, and SOC value, Df i max represents the maximum sample value of lithium battery current, voltage, temperature, and SOC value.

[0100] Furthermore, the method of dividing the training set, the validation set and the test set according to the ratio is as follows: the normalized data Df * 64% of the total is used as the training set, Df * 16% of the total set is used as the validation set, Df * The last 20% is used as the test set.

[0101] Among them, Df * The SOC data contains multiple working conditions, and the SOC values ​​of the training set, validation set, and test set are all in the process of decreasing from 1 to 0. The SOC data contains multiple working conditions, and the SOC values ​​of the training set, validation set, and test set are all in the process of decreasing from 1 to 0.

[0102] Furthermore, in step S3, the hyperparameters that need to be optimized in the LSTM neural network include: the number of LSTM layers, the number of hidden layer Dense layers, and the number of nodes of each layer of neurons. The genetic algorithm and particle swarm hybrid algorithm optimizes the LSTM neural network model in the parameter search space, taking the minimum SOC prediction error as the objective function, and performs the optimal combination of various hyperparameters. The parameter search space is specifically: the lower limit of the number of LSTM and Dense layers is 1, and the upper limit is 3; the lower limit of the number of neuron nodes in each layer is 32, and the upper limit is 256.

[0103] Furthermore, in step S3, constructing the composite GA-PSO-LSTM model includes the following steps:

[0104] Step S31, randomly initialize the population.

[0105] Step S32, setting the fitness function to the inverse of the mean square error of the LSTM neural network, and using the fitness function to calculate the fitness of the initial population.

[0106] Step S33, perform particle swarm evolution operation on all individuals of the population, update the speed and position of the particles, and generate offspring population 1. At the same time, perform inheritance, crossover, and mutation operations on all individuals of the population to obtain offspring population 2.

[0107] Step S34, sorting offspring population 1 and offspring population 2 according to fitness values ​​from large to small, and selecting the one with high fitness as offspring population 3.

[0108] Step S35, setting the termination condition to the number of iterations of the population. If the termination condition is met, the individual with the best fitness is output, and the optimization process ends. At this time, the LSTM network is the global optimal combination; otherwise, proceed to step S32.

[0109] Furthermore, in step S3, a composite GA-PSO-SVM model is constructed. The GA-PSO-GRU model method is the same as the GA-PSO-LSTM model. The hyperparameters for optimizing the SVM model are the penalty coefficient C and the Gaussian radial basis width coefficient σ. The search space of parameter C is (0,1000], and the search space of σ is (0.01,100]). The hyperparameters for optimizing the GRU model are the number of hidden layer neurons and the learning step size. The number of hidden layer neurons ranges from (32 to 256), and the learning step size ranges from (10 to 200).

[0110] Furthermore, in step S4, the working condition of the present invention simulates the operating conditions of cities or suburbs in the region (from another perspective, passenger cars or buses or other special vehicles), including acceleration and climbing, deceleration, constant speed, etc. Collecting battery consumption data sets of different working conditions for training can be extended to the battery prediction value of any time and any situation, which has the value of practical use.

[0111] Each working condition can be understood as a battery data set, and the SOC of each battery data set is a data set that decreases from 1 to 0. Simply put, each working condition is a small data set, and all working condition data sets are spliced ​​together to form the entire battery data set.

[0112] The entire data set contains data from 6 working conditions (data1, data2, data3, data4, data5, data6), of which the test set remains unchanged (the test set is data6), and the other 5 sets are used as training sets and test sets. The training set and validation set are changing (the first training uses data1 as the validation set and the other 4 sets as the training set. The second training uses data2 as the validation set and the other 4 sets as the training set. ... Train 5 times like this. The fifth training uses data5 as the validation set and the other 4 sets as the training set.)

[0113] Furthermore, the process of obtaining the predicted value of each model in the validation set in step S4 is as follows:

[0114] The entire data set contains data from 6 working conditions (data1, data2, data3, data4, data5, data6), among which the test set remains unchanged (the test set is data6), and the other 5 are used as training sets and test sets. The training set and validation set are changing (the first training uses data1 as the validation set and the other 4 as the training set, the second training uses data2 as the validation set and the other 4 as the training set, ... and trains 5 times like this. The fifth training uses data5 as the validation set and the other 4 as the training set). Each training will obtain three models: LSTM, SVM, and GRU.

[0115] For example, in the first training, data2, data3, data4, and data5 are used as training sets, and the current, voltage, and temperature of the training sets are used as inputs, and the true SOC value is used as output. Three models model_LSTM1, model_SVM1, and model_GRU1 are obtained through training. The current, voltage, and temperature of the data1 validation set are input as inputs into the three models model_LSTM1, model_SVM1, and model_GRU1 to obtain val_pre_lstm1, val_pre_svm1, and val_pre_gru1. The current, voltage, and temperature of the test set data6 are input as inputs into the three models model_LSTM1, model_SVM1, and model_GRU1 to obtain test_pre_lstm1, test_pre_svm1, and test_pre_gru1.

[0116] The second training uses data1, data3, data4, and data5 as training sets, and the current, voltage, and temperature of the training sets as inputs, and the true SOC value as output. Three models are trained, model_LSTM2, model_SVM2, and model_GRU2. The current, voltage, and temperature of the data2 validation set are input as inputs into the three models model_LSTM2, model_SVM2, and model_GRU2 to obtain val_pre_lstm2, val_pre_svm2, and val_pre_gru2. The current, voltage, and temperature of the test set data6 are input as inputs into the three models model_LSTM2, model_SVM2, and model_GRU2 to obtain test_pre_lstm2, test_pre_svm2, and test_pre_gru2.

[0117] The second training, the third training, the fourth training, and the fifth training use data2, data3, data4, and data5 as verification sets, and the remaining four of the five data sets as training sets. Each training can obtain three models model_LSTMi, model_SVMi, and model_GRUi, and then use the model to obtain the verification set results val_pre_lstmi, val_pre_svmi, val_pre_grui (i=2,3,4,5) and the test set prediction results test_pre_lstmi, test_pre_svmi, test_pre_grui.

[0118] Finally, val_pre_lstm1, val_pre_lstm2, val_pre_lstm3, val_pre_lstm4, val_pre_lstm5 are spliced ​​together in time series as a feature A1. Similarly, val_pre_svmj and val_pre_gruj are spliced ​​together in time series as two features A2 and A3. Each data corresponding to test_pre_lstm1, val_pre_lstm2, val_pre_lstm3, val_pre_lstm4, val_pre_lstm5 is added and divided by 5 to obtain the feature B1 of the test set. Similarly, each data corresponding to test_pre_svmj and test_pre_gruj is added and divided by 5 to obtain the features B2 and B3 of the test set. (j=1,2,3,4,5)

[0119] Furthermore, the model training process of the present invention in step S4 is as follows:

[0120] Take A1, A2, A3 obtained in step S4 as input x and the actual SOC value as output y to obtain three models GA-PSO-LSTM, GA-PSO-SVM, and GA-PSO-GRU. After obtaining the model, input the test set features B1, B2, and B3 obtained in step S4 into the corresponding GA-PSO-LSTM, GA-PSO-SVM, and GA-PSO-GRU models to obtain three predicted values ​​predict1, predict2, and predict3. The predicted values ​​of the three models are weighted averaged to obtain the final SOC predicted value. The final predicted SOC = (predict1+predict2+predict3) / 3.

[0121] Furthermore, in step S4, the current, voltage, and temperature in the training set are used as inputs, and the true SOC value is used as output to input into the LSTM neural network model, SVM model, and GRU neural network model containing the optimal hyperparameters. The training set has a total of 5 working condition data, and each time 4 working condition data are used as training sets, and the remaining working condition data are used as validation sets. Each validation set obtains a predicted SOC value, and the predicted values ​​val_pre_lstm of the 5 validation sets of the LSTM model are sequentially spliced ​​into A1, the predicted values ​​val_pre_svm of the 5 validation sets of the SVM model are sequentially spliced ​​into A2, and the predicted values ​​val_pre_gru of the 5 validation sets of the GRU model are sequentially spliced ​​into A3. The weighted average of the predicted values ​​test_pre_lstm, test_pre_svm, and test_pre_gru of the test set of the LSTM model, the SVM model, and the GRU model after 5 trainings is B1, B2, and B3. The obtained B1, B2, and B3 are weighted averaged to obtain the SOC prediction value B0 of the test set. A1, A2, and A3 are new features of the training set, and B0 is a new feature of the test set.

[0122] Furthermore, in step S5, A1, A2, and A3 are used as new features for the second round of training, and the real SOC is used as a label. The models with the optimal hyperparameters obtained in step S3 are used for training respectively. Finally, B0 is input into the trained model to obtain the prediction values ​​predict1, predict2, and predict3 of the three models of GA-PSO-LSTM, GA-PSO-SVM, and GA-PSO-GRU. The obtained prediction values ​​are weighted averaged to obtain the final model prediction value of SOC. Use the root mean square error (RMSE) as the loss function for model training. Set the training set data bathsize = 32, epochs = 50, select Adam for gradient calculation, iteratively update the network and continuously update the model weights to minimize the prediction error RMSE. The RMSE calculation formula is as follows:

[0123]

[0124] Where n is the number of data sets, y i is the real SOC value, is the SOC value predicted by the model.

[0125] Furthermore, in step S5, the use of the trained LSTM model to predict the test set is specifically as follows: the new feature B0 obtained by the first round of training of the test set in step S4 is input into the model, and the LSTM outputs h(t) through forward calculation:

[0126] Forget gate: f t =σ(W f x t +U f h t-1 +b f );

[0127] Input gate: i t =σ(W i x t +U i h t-1 +b i );

[0128] Unit status: g t =tanh(W g x t +U g h t-1 +b g );

[0129] Output gate: o t =σ(W o x t +U o h t-1 +b o );

[0130] Current status: c t =f t *c t-1 +i t *g t ;

[0131] Output: h t =o t tanh(ct )

[0132] Among them, f t Indicates the information to be discarded at the tth time node of the battery data, i t Indicates the information to be updated at the tth time node of the battery data, g t represents the cell state of the battery data at the tth time node, o t Indicates the information to be output at the tth time node of the battery data, c t represents the current state of the battery data at the tth time node, h t Represents the output value of the battery data at the tth time node. W i , W f , W o , W g are the weight matrices of the input gate, forget gate, output gate, and cell state respectively; U i , U f , U o , U g is also the corresponding weight matrix, b i , b f , b o , b g is the corresponding bias term; h t-1 is the hidden layer state at time t-1; x t is the input vector at time t; tanh is the hyperbolic tangent activation function, and σ is the sigmoid function.

[0133] The predicted value B0 of the first round of training of the test set is used as a new feature to input the LSTM neural network for forward calculation, and the predicted value predict1 of the SOC of the LSTM model is output.

[0134] Furthermore, in step S5, the method of training the SVM model is specifically as follows: the new features A1, A2, A3 (x1, x2, ... x n ) as input, the actual SOC value (y1,y2,…y n ) as output. Seek the regression function fi(·) such that:

[0135] y=f(x)=wφ(x)+b;

[0136]

[0137] Use the training data to find the appropriate vectors w and b so that the regression error of the data for the hyperplane is minimized:

[0138]

[0139] sty-f(x)≤ε+ξ

[0140] yf(x)≤ε+ξ *

[0141] ξ,ξ * ≥0;

[0142]

[0143] In the formula, ξ and ξ* are the relaxation variables of the two boundaries, C is the penalty coefficient, which is used to control the weight between the objective function "optimization" and "minimum deviation", and K is the kernel function. This paper adopts the RBF (radial basis function) kernel function. j is the eigenvector, and σ is the Gaussian radial basis width coefficient.

[0144] The predicted value B0 of the first round of training of the test set is input as a new feature into the SVM model containing the optimal parameters w and b for forward calculation, and the predicted value predict2 of the SOC of the SVM model is output.

[0145] Furthermore, in step S5, the GRU model uses the update gate and reset gate as the core, uses a function to transform the matrix of the current input sequence and the memory variable of the previous moment, and updates it, and inputs the updated value into the update gate to compare and determine the importance of the current state and the state of the previous moment. In the reset gate, determine whether the previous moment needs to be written into the candidate set. The specific method of training the GRU model is: the new feature A1 (x1, x2, ... x n ), A2, A3 as input, the actual SOC value (y1, y2, ... y n ) as output. After the model is trained, B0 in the test set is used as input and h is calculated through forward propagation t That is the predicted SOC value predict3, and the calculation method is as follows:

[0146]

[0147] Among them, h t-1 refers to the state variable at the previous moment, h t refers to the state variable at that moment, r t refers to the state variable of the update gate, z t refers to the state variable of the reset gate, refers to the current candidate set state, y t refers to the current state, W o , W z , W r is the output vector and x t ,h t-1The weights, candidate sets, reset gates, and update gates that make up the matrix multiplication.

[0148] φ is the tanh activation function, σ is the sigmoid function, and the expression is:

[0149]

[0150]

[0151] Furthermore, in step S5, the final SOC prediction value is the weighted average of the prediction results of the LSTM, SVM, and GRU models.

[0152] The mean square error (MSE), root mean square error (RMSE) and R 2 The coefficients are used as evaluation indicators of the model. When the MSE and RMSE values ​​are smaller, the accuracy of the model is higher; when R 2 The closer the coefficient is to 1, the more accurate the model's predictions are.

[0153] Mean Square Error:

[0154] Root Mean Square Error:

[0155] R 2 coefficient:

[0156] Where n is the number of data sets, y i is the real SOC value, is the SOC value predicted by the model, is the average SOC of all samples.

[0157] Specific steps S3 implementation (simplified):

[0158] Parameters of the genetic algorithm: Set the population to 20 individuals, the number of iterations to 100, and the crossover probability p c =0.7, mutation probability p m =0.1. The encoding is binary. The genes in an individual include at least 4 chromosomes, which represent the number of LSTM layers, the number of Dense layers, the number of LSTM nodes in the first layer, (the number of LSTM nodes in the second layer...), and the number of nodes in the Dense layer. (Chromosome interpretation diagram as shown in Figure 3 shown).

[0159] For each prediction result, the inverse of the root mean square error is used as the fitness value. In the selection stage, the roulette selection method is selected, and in the crossover stage, the single point crossover method is selected.

[0160] Parameters of the particle swarm algorithm: Set the population to 20 individuals, the number of iterations to 100, the learning factor c1=c2=1, and the velocity V boundary and position X boundary of each particle to [-256,256;-3,3]. The inertia factor w is set to a dynamically changing factor ranging from 0.4 to 0.9, and the calculation formula is:

[0161]

[0162] Among them, w max represents the maximum value of w, w min represents the minimum value of w, run represents the current number of iterations, run max Indicates the number of iterations at which termination occurs.

[0163] In each iteration, the particle will update its speed and position according to the individual value and group value. The formula for updating speed and position is:

[0164]

[0165]

[0166] Where w is the inertia factor; i = 1, 2…n; n is the number of particle populations; k is the current number of iterations; V i is the velocity of the particle; X i is the position of the particle; P i is the individual extreme value; P g is the individual extreme value; c1, c2 are learning factors; r1 and r2 are random numbers between [0,1].

[0167] During the iteration process, if the updated speed and position exceed the given boundary value, the boundary value is used to replace the speed and position that exceed the range, and the fitness value is recalculated.

[0168] In each iteration of the genetic algorithm, a new offspring 1 is generated by selecting a crossover mutation operation, and the fitness value of offspring 1 is calculated; in each iteration of the particle swarm algorithm, a new offspring 2 is also generated by updating the speed and position, and the fitness values ​​of offspring 1 and offspring 2 are sorted, and the one with the largest fitness value is selected as offspring 3; after continuous iterative updates, when the termination condition is met, the optimal individual output is the optimal hyperparameter of the network.

[0169] Specific step S4 implementation (simplified):

[0170] Collect 5 sets of working condition data sets as training sets and 1 new data set as test set. Among the 5 training sets, use 4 of them as real training sets and the remaining 1 as validation set. The first model training uses current, voltage and temperature as input and the real SOC value as output. The LSTM, SVR and GRU models with optimal parameters in step S3 are trained 5 times. Each time the validation set is replaced with 1 working condition data, the other 4 working conditions are used as training sets to obtain the SOC values ​​val_pre_lstm and val_p predicted by the validation set. re_svm, val_pre_gru, and get the predicted SOC value of the test set test_pre_lstm, test_pre_svm, test_pre_gru at the same time. The predicted values ​​of 5 validation sets of LSTM, SVM, and GRU models are sequentially spliced ​​into A1, A2, and A3. The predicted values ​​of a single test set of 5 trainings of LSTM, SVM, and GRU models are weighted averaged to get B1, B2, and B3. Then, the weighted average of B1, B2, and B3 is calculated to get the SOC predicted value B0 of the test set. B0 is the average value of the corresponding values ​​of B1, B2, and B3. A1, A2, and A3 are new features of the training set, and B0 is a new feature of the test set (the explanation diagram of the new features generated by the fusion model is shown in the figure). Figure 4 shown).

[0171] Figure 6 This is a comparison chart of the predicted effect and the actual detection effect obtained by collecting data on Panasonic batteries based on the method and system. The actual effect is the blue thin line and the predicted effect is the red thick line.

[0172] It should be understood that those skilled in the art can make improvements or changes based on the above description, but the specific implementation of the present invention is not limited to this example. All alternative solutions with similar effects proposed according to the principles and ideas of the present invention should be considered as the protection scope of the present invention.

Claims

1. A multi-model fusion battery state of charge prediction method based on GA-PSO optimization, characterized in that The steps include: S1: Collect the external parameters of the lithium battery: including current, voltage, temperature, discharge ampere-hour, and rated capacity of the electric vehicle, and calculate the actual battery SOC value from the external parameters to obtain a time-ordered data set including the original features of current, voltage, temperature, and SOC value plus labels; S2: Data cleaning: Data analysis is performed on the missing values, duplicate values, and outliers in the data set. Interpolation is performed on a small number of missing values, duplicate values ​​are removed, and outliers are deleted. Then, Max-min normalization is performed on the cleaned data, and the normalized sample data is divided into training set, validation set, and test set. S3: Construct multiple base models and integrate them into a strong learning model: Construct a GA-PSO-LSTM model, and use a genetic algorithm and a particle swarm hybrid algorithm to optimize and determine the optimal hyperparameters of the LSTM neural network; construct a GA-PSO-SVM model, and optimize and determine the optimal penalty coefficient and Gaussian radial basis width coefficient of the SVM; construct a GA-PSO-GRU model, and optimize and determine the number of hidden layer nodes and learning step size of the GRU; S4: Establish new features: Integrate the GA-PSO-LSTM model, GA-PSO-SVM model, and GA-PSO-GRU model containing the optimal hyperparameters obtained in step S3, keep one of the working conditions as the test set, and use four of the remaining five working conditions as training sets each time, and the data of the other working condition as the verification set, input the predicted values ​​of each model verification set in turn, and use the predicted values ​​of each model verification set as the features of the new training set; also obtain the predicted values ​​of the test set, and take the weighted average of the predicted values ​​of the test set as the features of the new test set; S5: SOC prediction: Take the features of the new training set obtained in step S4 as input, take the true SOC value as output, select GA-PSO-LSTM, GA-PSO-SVM, GA-PSO-GRU model training, and after the model is trained, take the features of the new test set in step S4 as input to obtain the SOC prediction values ​​of the three models. Finally, take the weighted average of the SOC prediction values ​​of the individual models to obtain the final SOC prediction value, and evaluate the model error. In step S3, constructing the GA-PSO-LSTM model includes the following steps: Step S31, randomly initialize the population; Step S32, setting the fitness function to the inverse of the mean square error of the LSTM neural network, and using the fitness function to calculate the fitness of the initial population; Step S33, performing particle swarm evolution operation on all individuals of the population, updating the speed and position of the particles, and generating offspring population 1; at the same time, performing inheritance, crossover, and mutation operations on all individuals of the population to obtain offspring population 2; Step S34, sorting the offspring population 1 and the offspring population 2 according to the fitness value from large to small, and selecting the one with the highest fitness as the offspring population 3; Step S35, setting the termination condition to the number of iterations of the population. If the termination condition is met, the individual with the best fitness is output, and the optimization process ends. At this time, the LSTM network is the global optimal combination; otherwise, proceed to step S32; The method of constructing GA-PSO-SVM model and GA-PSO-GRU model is the same as that of GA-PSO-LSTM model; The training process of the GA-PSO-LSTM model, the GA-PSO-SVM model, and the GA-PSO-GRU model in step S4 is as follows: Each training takes the current, voltage, and temperature included in the training set as input, and the true SOC value as output. Each training will obtain three trained models: GA-PSO-LSTM model, GA-PSO-SVM model, and GA-PSO-GRU model. For the trained models obtained after each training, the current, voltage, and temperature in the validation set and test set corresponding to each training are used as inputs and input into the corresponding trained models to obtain the output validation set prediction value and test set prediction value. Finally, the validation set prediction values ​​obtained by the GA-PSO-LSTM model after five trainings are spliced ​​together in time series as the first feature A1, the test set prediction values ​​obtained by the GA-PSO-SVM model after five trainings are spliced ​​together in time series as the second feature A2, and the test set prediction values ​​obtained by the GA-PSO-GRU model after five trainings are spliced ​​together in time series as the third feature A3; each data in the first feature A1 is added and divided by 5 to obtain the first test set feature B1; each data in the second feature A2 is added and divided by 5 to obtain the second test set feature B2; each data in the third feature A3 is added and divided by 5 to obtain the third test set feature B3; the weighted average of the three test set features B1, B2, and B3 is obtained to obtain the SOC prediction value B0 of the test set; In step S4, the entire battery data set contains data of 6 working conditions: data1, data2, data3, data4, data5, data6, among which data6 remains unchanged as a test set, and the other 5 are used as training sets and validation sets. The training set and validation set change according to the number of training times: in the first training, data1 is used as a validation set, and the other 4 are used as training sets. In the second training, data2 is used as a validation set, and the other 4 are used as training sets, and so on. The training is repeated 5 times. In the fifth training, data5 is used as a validation set, and the other 4 are used as training sets.

2. The multi-model fusion battery state of charge prediction method based on GA-PSO optimization according to claim 1 is characterized in that In step S1, when the electric vehicle is running, during the battery discharge process, the external parameters of the lithium battery are collected in real time at intervals of 0.1s: current, voltage, temperature, discharge ampere-hour, and rated capacity of the electric vehicle; the actual battery SOC value is calculated by the following formula using the external parameters: Among them, the SOC value represents the current power of the lithium battery, Q res Indicates the battery's discharge ampere-hour, Q rated Indicates the rated capacity of the battery.

3. The multi-model fusion battery state of charge prediction method based on GA-PSO optimization according to claim 1 is characterized in that The step S1 obtains a lithium battery data set Df with original features of each external parameter sorted by time and labeled: Df={(Df i1 ,Df v1 ,Df t1 ,Df s1 ),(Df i2 ,Df v2 ,Df t2 ,Df s2 ),…,(Df in ,Df vn ,Df tn ,Df sn )}; Where n represents the number of samples in the collected data set, Df ik , Df vk , Df tk , Df sk They respectively represent the lithium battery current, voltage, temperature, and SOC value data of the kth sample, where k is between 1 and n.

4. The multi-model fusion battery state of charge prediction method based on GA-PSO optimization according to claim 1 is characterized in that In step S2, the normalization formula for normalizing the deviation of the current, voltage, temperature and actual SOC value of the processed battery data is as follows: Among them, Df * Represents the normalized lithium battery current, voltage, temperature, and SOC value data, Df i Indicates the lithium battery current, voltage, temperature, and SOC value data before normalization, Df i min represents the minimum sample value of lithium battery current, voltage, temperature, and SOC value, Df i max represents the maximum value of the samples of lithium battery current, voltage, temperature, and SOC value; In step S2, the training set, validation set, and test set are divided according to the proportion: the normalized data Df * 64% of the total set is used as the training set, Df * 16% of the total set is used as the validation set, Df * The last 20% is used as the test set.

5. The multi-model fusion battery state of charge prediction method based on GA-PSO optimization according to claim 1 is characterized in that In step S2, Df * Contains SOC data for multiple working conditions. The SOC values ​​of the training set, validation set, and test set all decrease from 1 to 0.

6. The multi-model fusion battery state of charge prediction method based on GA-PSO optimization according to claim 1 is characterized in that In step S5, the final SOC prediction value is the weighted average of the prediction results of the LSTM, SVM, and GRU models; the mean square error (MSE), root mean square error (RMSE), and R 2 The coefficient is used as the evaluation index of the model. When the MSE and RMSE values ​​are smaller, the accuracy of the model is higher; when R 2 The closer the coefficient is to 1, the more accurate the model's predictions are: Mean Square Error: Root Mean Square Error: R 2 coefficient: Where n is the number of data sets, y i is the real SOC value, is the SOC value predicted by the model, and y is the average SOC of all samples.

7. A multi-model fusion battery state of charge prediction system based on GA-PSO optimization, characterized by A program for implementing the method according to any one of claims 1 to 6 is stored.