Cell capacity combination model prediction method

Abstract

The invention discloses a cell capacity combination model prediction method, which belongs to the technical field of battery management, and comprises the following steps: firstly, performing characteristic correlation analysis, then specifically adopting corresponding data of selected characteristics, and training the data. The method comprises the following steps: acquiring training data, processing the data by utilizing a wavelet transform algorithm, and performing prediction processing by respectively using LightGBM, improved DDPG and XGBoost so as to respectively acquire a prediction error A, a prediction error B and a prediction error C; according to the prediction error A, the prediction error B and the prediction error C, a prediction error quadratic sum reciprocal method is used for analysis to obtain the weight of a single model corresponding to each prediction error; weighting operation is carried out on all the models, so that a combined prediction model of WT + XGBoost + DDPG + LightGBM is formed; and finally, carrying out wavelet reconstruction and reverse normalization processing on the combined prediction model to obtain a prediction result. The technical problems that in the prior art, errors caused by a single prediction model are large, and the prediction model cannot be applied to production practice are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of battery management, and in particular to a method for predicting a battery cell capacity combination model. Background Art

[0002] Electric vehicle companies have generally launched new energy vehicle models that use lithium-ion power batteries as energy storage devices; moreover, in other power equipment fields, lithium-ion batteries are also the main type of power batteries. In the lithium battery industry, the capacity of the battery cell is an important indicator for measuring the performance of lithium batteries. The capacity decay of lithium batteries characterizes the length of battery life, so it is necessary to estimate the capacity of lithium batteries.

[0003] At present, many lithium battery capacity estimates are designed based on Kalman filters. For example, Chinese patent CN110320473A proposes a method for estimating the capacity of automotive lithium batteries based on Kalman filtering and fuzzy logic. The method aims to make predictions and corrections through Kalman filters. In this scheme, Kalman filtering believes that variables, that is, estimated voltage, estimated capacity, actual voltage, actual capacity and other variable parameters, have errors in the environment, and uses the normal distribution of estimated voltage and actual voltage to calculate a better predicted voltage; and make the predicted voltage more consistent with a normal distribution with a smaller variance (σ). At the same time, if Kalman filtering is used, the covariance between the optimal predicted voltage and the optimal predicted capacity must also be considered; because this is considered to be external noise in Kalman, which has an impact on the prediction.

[0004] Furthermore, in the production process of lithium-ion batteries, the capacity data in the battery capacity grading process is an important indicator for evaluating the quality of single lithium-ion batteries before leaving the factory. Lithium battery manufacturers will classify lithium-ion batteries according to different capacity levels and combine batteries of the same level into battery packs to obtain better stability and service life. However, the production time of the capacity grading process of lithium batteries is usually relatively long, which has a greater impact on production capacity. On the one hand, if the production time of the capacity grading process is too long, it will increase the time of equipment investment and cause unnecessary waste of manpower costs. On the other hand, if the production time of the capacity grading process is too long, it will greatly increase the production time of the battery cell, thereby compressing the production space, reducing production efficiency and affecting the shipment rate.

[0005] Among the related schemes disclosed in the prior art, there are studies on capacity prediction methods based on deep neural networks and on capacity prediction methods based on traditional machine learning. However, in the production practice of using a single prediction model to predict the capacity of lithium battery cells, there are generally technical problems such as slow convergence speed, large prediction error, and large interference from outliers. As a result, it is difficult to put a single prediction model into use in actual production. In other words, a single prediction model has the defect of poor practicality for the production process of battery cells. Summary of the invention

[0006] Based on this, it is necessary to provide a battery cell capacity combination model prediction method to solve the technical problem that a single prediction model in the existing technology has large errors and cannot be applied to production practice.

[0007] A method for predicting a battery cell capacity combination model comprises the following steps:

[0008] S1: Prepare training data of characteristic parameters that affect cell capacity in advance;

[0009] S2: Normalize and perform wavelet decomposition on the training data;

[0010] S3: Use LightGBM, improved DDPG and XGBoost to perform prediction processing on the processed data to obtain prediction error A, prediction error B and prediction error C respectively;

[0011] S4: According to the prediction error A, prediction error B and prediction error C, the prediction error square sum inverse method is used to analyze to obtain the weight of a single model corresponding to each prediction error; then each model is weighted to form a combined prediction model of WT+XGBoost+DDPG+LightGBM;

[0012] S5: Perform wavelet reconstruction and denormalization processing on the combined prediction model to obtain the prediction result.

[0013] Furthermore, in step S2, the training data is processed using a wavelet transform algorithm, and then the processed data is normalized, specifically:

[0014]

[0015] Where WT is the wavelet transform, s_{v} is the scaling variable, t_{v} is the translation variable, and g(t) is the signal that needs to be decomposed by wavelet transform.

[0016] Furthermore, in step S3, the reward function of DDPG adopts a reward coefficient variation reward, that is, the prediction error value of different models is used as the reward for each action, specifically:

[0017]

[0018] In the formula, R t is the reward function of DDPG, w i (i=1, 2, 3) are the corresponding error values ​​after processing the data using LightGBM, DDPG and XBGoost respectively.

[0019] Furthermore, in step S4, according to the prediction error values ​​of different single models, the prediction error sum of squares method is used to analyze them to obtain the weights of the relevant single models; then a weighted operation is performed on each relevant single model, and finally a combined prediction model of WT+XGBoost+DDPG+LightGBM is formed, which is specifically:

[0020] E x =σ1E 1x +σ2E 2x +σ3E 3x ,x=1,2,…,n

[0021]

[0022]

[0023]

[0024] Where n is the number of battery cells, σ1, σ2, and σ3 are the weight values ​​obtained after processing using LightGBM, DDPG, and XBGoost, respectively. 1t 、E 2t 、E 3t They are the predicted values ​​obtained by using LightGBM, DDPG, and XBGoost to predict the data.

[0025] Furthermore, in step S1, a back propagation algorithm is used to train the data.

[0026] Furthermore, in the back propagation algorithm, the discharge voltage and discharge current of the battery are used as input; and the remaining capacity of the battery is used as output.

[0027] Furthermore, the GA algorithm is used to optimize the structural parameters and network training parameters of the back propagation algorithm, which specifically includes the following steps:

[0028] Step S11: Initialize the population P, which includes the crossover scale, crossover probability and mutation probability. Real numbers are used for encoding, and the initial population is 20;

[0029] Step S12: Calculate the evaluation function of each individual and rank them;

[0030] Step S13: Perform crossover operation on individuals with a crossover probability of 0.6 to generate new individuals;

[0031] Step S14: using the mutation probability, mutate to generate new individuals;

[0032] Step S15: insert the new individual into the population P and calculate the evaluation function of the new individual;

[0033] Step S16: Calculate the square sum of the errors of the neural network. If it reaches a preset value, go to step S17; otherwise, go to step S13 to continue genetic operation.

[0034] Step S17: Using the optimized initial values ​​inherited by GA as initial weights and thresholds, the network is trained using the BP algorithm until the preset accuracy is met.

[0035] In summary, in a method for predicting a combined model of battery capacity of the present invention, one advantage is to improve DDPG; that is, a multi-strategy network is adopted, and the acquired training data is placed in experience replay pools of different priorities according to the abnormal large values, normal values, and abnormal small values ​​in the battery capacity data, and unnecessary search time is reduced through prior knowledge, thereby improving the convergence speed of the model. In addition, another advantage is the use of a multi-model combined prediction method; that is, through weight distribution, while processing nonlinear data prediction, it can solve the inherent limitations such as complex calculations, poor generalization ability, and overfitting, and can also improve the model prediction accuracy. In addition, another advantage is the method of using wavelet transform to decompose the data; because there will be some outliers in the original capacity data, the wavelet transform can decompose the time series signal into high and low frequency signals, which is helpful to make better predictions in terms of data outliers and, thereby improving the accuracy of battery capacity prediction. Moreover, the addition of wavelet transform can reduce the increase in model complexity caused by the combined model, so as to further optimize the model. Therefore, the method for predicting the combined model of battery cell capacity of the present invention solves the technical problem that the single prediction model in the prior art causes large errors and cannot be applied to production practice. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is a flow chart of a method for predicting a battery cell capacity combination model according to the present invention;

[0037] Figure 2 A schematic diagram of a method for predicting a battery cell capacity combination model according to the present invention;

[0038] Figure 3 A schematic diagram of a BP network model used in a method for predicting a battery cell capacity combination model of the present invention;

[0039] Figure 4 The present invention provides a GA optimized BP network flow chart of a method for predicting a battery cell capacity combination model. DETAILED DESCRIPTION

[0040] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention, so the present invention is not limited by the specific embodiments disclosed below.

[0041] Please also read Figure 1 to Figure 2 The present invention provides a method for predicting a battery cell capacity combination model, comprising the following steps:

[0042] S1: Prepare training data of characteristic parameters that affect cell capacity in advance;

[0043] S2: Normalize and perform wavelet decomposition on the training data;

[0044] S3: Use LightGBM, improved DDPG and XGBoost to perform prediction processing on the processed data to obtain prediction error A, prediction error B and prediction error C respectively;

[0045] S4: According to the prediction error A, prediction error B and prediction error C, the prediction error square sum inverse method is used to analyze to obtain the weight of a single model corresponding to each prediction error; then each model is weighted to form a combined prediction model of WT+XGBoost+DDPG+LightGBM;

[0046] S5: Perform wavelet reconstruction and denormalization processing on the combined prediction model to obtain the prediction result.

[0047] Specifically, assuming that we start from the perspective of practical application, we derive the prediction theory of the remaining battery capacity: at time t, the remaining battery capacity is recorded as C(t), the battery voltage is U(t), the current is I(t), the rated capacity of the battery is C0, the average internal resistance of the battery is R0, and the discharged capacity is C1(t), so we have:

[0048] C(t)=C0-C1(t)

[0049] Then the implicit function relationship of each parameter can be expressed as:

[0050] f2(U(t),I(t),R0,C0,t,C(t))=0

[0051] Although the rated capacity C0 and the average internal resistance R0 of batteries of different models are different, for a certain battery, their values ​​are constants, so that the above formula can be simplified to:

[0052] f2(U(t),I(t),t,C(t))=0

[0053] According to the battery discharge data, when the discharge time is known, the discharged capacity C1(t) can be calculated. However, usually, the battery discharge time is unknown, and the discharged capacity C1(t) is also unknown. Therefore, the relationship between the various parameters is:

[0054] F3(U(t),I(t),C(t))=0

[0055] Therefore, the remaining battery capacity can be expressed as:

[0056] C(t)=h(U(t),I(t))=0

[0057] It can be concluded from the above theoretical deduction that in the model prediction method of the battery cell capacity, the battery's discharge voltage U(t), discharge current I(t) and the remaining capacity C(t) at this moment can be used as characteristic parameters that affect the battery cell prediction capacity model.

[0058] Furthermore, if the back propagation algorithm, i.e., BP network, is used to train the data, the discharge voltage U(t) and discharge current I(t) should be used as input; the remaining capacity C(t) can be used as output. Figure 3 , neural network can realize highly nonlinear mapping from input to output, and any continuous function can be realized by a three-layer forward network. For example, a single hidden layer structure can be used. By changing the number of nodes, it is found that as the number of nodes increases, the accuracy of network training increases, but the speed and memory usage will increase. Considering practical applications, the number of hidden layer nodes is 13, the discharge current and discharge voltage are input, the remaining power is output, and the mean square error function is the performance function.

[0059] Specifically, the advantage of the BP algorithm is that it is accurate in optimization, but it also has some disadvantages, the most important of which is that it is easy to fall into local minima, converge slowly, and cause oscillation effects. In order to improve the convergence of network training and thus increase the rate of sample learning, the GA algorithm can be introduced to optimize the structural parameters and network training parameters of the BP network, so that the sample data can achieve better training results.

[0060] More specifically, the genetic algorithm, or Genetic Algorithm, is abbreviated as GA in English. This algorithm is a method inspired by the theory of evolution, species selection theory and genetic theory in biology. It solves the difficulties encountered by traditional artificial intelligence methods based on symbolic processing mechanisms in knowledge representation, information processing, solving combination explosion, etc. to a certain extent. Its self-organization, self-adaptation, self-learning and group evolution capabilities make it very suitable for large-scale complex optimization problems. Its biggest feature is implicit parallelism and global solution space search. The structural parameters of the BP network mainly include network weights and thresholds. GA is used to optimize the initial weights and thresholds of the network, making the BP network easier to train, and can also further improve the subsequent processing efficiency of the training data, and avoid the BP network from falling into the local optimum during data training.

[0061] Please continue reading Figure 4 , the process of using GA to optimize BP neural network is as follows:

[0062] Step S11: Initialize the population P, which includes the crossover scale, crossover probability, mutation probability, etc., and uses real numbers for encoding. The initial population is 20;

[0063] Step S12: Calculate each individual evaluation function and sort them;

[0064] Step S13: Perform crossover operation on individuals with a crossover probability of 0.6 to generate new individuals;

[0065] Step S14: using the mutation probability, mutate to generate new individuals;

[0066] Step S15: insert the new individual into the population P and calculate the evaluation function of the new individual;

[0067] Step S16: Calculate the square sum of the errors of the neural network. If it reaches a preset value, go to step S17; otherwise, go to step S13 to continue genetic operation.

[0068] Step S17: Using the optimized initial values ​​inherited by GA as initial weights and thresholds, the network is trained using the BP algorithm until the specified accuracy is reached.

[0069] Therefore, MATLAB can be used for programming simulation, and the weights and thresholds can be adjusted through GA to achieve the expected effect of the network. The collected data is used to test and verify the network after training, and the error analysis meets the requirements of the network model. According to the change of error, the number of hidden layer nodes is continuously selected, and it is finally determined that when the number of hidden layer nodes is 20, the number of iterations is 21, which reaches the set error, which can meet the convergence speed and accuracy.

[0070] Furthermore, in the above steps, several features with the greatest impact are selected through feature correlation analysis, such as voltage, current, battery temperature and cell capacity; then, the wavelet transform algorithm is used to process the data, so that the prediction on abnormal data is more accurate, and the accuracy of cell capacity prediction is improved. Finally, the data is normalized, specifically:

[0071]

[0072] Where WT is the wavelet transform, s_{v} is the scaling variable, t_{v} is the translation variable, and g(t) is the signal that needs to be decomposed by wavelet transform.

[0073] Furthermore, the data processed by normalization and wavelet transform algorithm are respectively predicted and processed using LightGBM, improved DDPG and XGBoost; among them, prediction error A can be obtained after processing the data using LightGBM, prediction error B can be obtained after processing the data using improved DDPG, and prediction error C can be obtained after processing the data using XGBoost. Among them, the DDPG reward function uses reward coefficient variation reward, which varies with other variables; and the prediction error values ​​of different models can be counted as the reward for each action, specifically:

[0074]

[0075] In the above formula, R t is the reward function of DDPG, w i (i=1, 2, 3) are the corresponding error values ​​after processing the data using LightGBM, DDPG and XBGoost respectively.

[0076] Furthermore, according to the prediction error values ​​of different single models, the inverse sum of squares of prediction errors can be used to analyze them to obtain the weights of relevant single models; then a weighted operation is performed on each relevant single model, and finally a combined prediction model of WT+XGBoost+DDPG+LightGBM is formed, as shown below:

[0077] E x =σ1E 1x +σ2E 2x +σ3E 3x ,x=1,2,…,n

[0078]

[0079]

[0080]

[0081] Where n is the number of battery cells, σ1, σ2, and σ3 are the weight values ​​obtained after processing using LightGBM, DDPG, and XBGoost, respectively. 1t 、E 2t 、E 3t They are the predicted values ​​obtained by using LightGBM, DDPG, and XBGoost to predict the data.

[0082] Specifically, by comparing 7 benchmark models, it is evaluated that the combined model prediction method of the battery capacity of the present invention can improve the accuracy of the battery capacity prediction. First, the combined model can achieve good prediction accuracy; although there are large fluctuations and shortcomings in the processing of certain abnormal values; however, by adding wavelet transform decomposition, the stability of the combined prediction method can be improved, so that the prediction accuracy can be further improved.

[0083] In summary, in a method for predicting a combined model of battery capacity of the present invention, one advantage is to improve DDPG; that is, a multi-strategy network is adopted, and the acquired training data is placed in experience replay pools of different priorities according to the abnormal large values, normal values, and abnormal small values ​​in the battery capacity data, and unnecessary search time is reduced through prior knowledge, thereby improving the convergence speed of the model. In addition, another advantage is the use of a multi-model combined prediction method; that is, through weight distribution, while processing nonlinear data prediction, it can solve the inherent limitations such as complex calculations, poor generalization ability, and overfitting, and can also improve the model prediction accuracy. In addition, another advantage is the method of using wavelet transform to decompose the data; because there will be some outliers in the original capacity data, the wavelet transform can decompose the time series signal into high and low frequency signals, which is helpful to make better predictions in terms of data outliers and, thereby improving the accuracy of battery capacity prediction. Moreover, the addition of wavelet transform can reduce the increase in model complexity caused by the combined model, so as to further optimize the model. Therefore, the method for predicting the combined model of battery cell capacity of the present invention solves the technical problem that the single prediction model in the prior art causes large errors and cannot be applied to production practice.

[0084] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0085] The above-mentioned embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A method for predicting a battery cell capacity combination model, characterized in that: It includes the following steps: S1: Prepare training data of characteristic parameters that affect cell capacity in advance; S2: Normalize and perform wavelet decomposition on the training data; S3: Use LightGBM, improved DDPG and XGBoost to perform prediction processing on the processed data to obtain prediction error A, prediction error B and prediction error C respectively; S4: According to the prediction error A, prediction error B and prediction error C, the prediction error square sum inverse method is used to analyze to obtain the weight of a single model corresponding to each prediction error; then each model is weighted to form a combined prediction model of WT+XGBoost+DDPG+LightGBM; S5: Perform wavelet reconstruction and denormalization processing on the combined prediction model to obtain the prediction result.

2. The method for predicting a battery cell capacity combination model according to claim 1, characterized in that: In step S2, the training data is processed using a wavelet transform algorithm, and then the processed data is normalized, specifically: Where WT is the wavelet transform, s_{v} is the scaling variable, t_{v} is the translation variable, and g(t) is the signal that needs to be decomposed by wavelet transform.

3. The method for predicting a battery cell capacity combination model according to claim 2, characterized in that: In step S3, the reward function of DDPG adopts reward coefficient variation reward, that is, the prediction error value of different models is used as the reward for each action, specifically: In the formula, R t is the reward function of DDPG, w i (i=1, 2, 3) are the corresponding error values ​​after processing the data using LightGBM, DDPG and XBGoost respectively.

4. The method for predicting a battery cell capacity combination model according to claim 3, characterized in that: In step S4, according to the prediction error values ​​of different single models, the prediction error sum of squares method is used to analyze them to obtain the weights of the relevant single models; then a weighted operation is performed on each relevant single model, and finally a combined prediction model of WT+XGBoost+DDPG+LightGBM is formed, which is specifically: E x =σ1E 1x +σ2E 2x +σ3E 3x ,x=1,2,…,n Where n is the number of battery cells, σ1, σ2, and σ3 are the weight values ​​obtained after processing using LightGBM, DDPG, and XBGoost, respectively. 1t 、E 2t 、E 3t They are the predicted values ​​obtained by using LightGBM, DDPG, and XBGoost to predict the data.

5. The method for predicting a battery cell capacity combination model according to claim 1, characterized in that: In step S1, the back propagation algorithm is used to train the data.

6. A method for predicting a battery cell capacity combination model according to claim 5, characterized in that: In the back propagation algorithm, the discharge voltage and discharge current of the battery are used as input; and the remaining capacity of the battery is used as output.

7. A method for predicting a battery cell capacity combination model according to claim 6, characterized in that: The GA algorithm is used to optimize the structural parameters and network training parameters of the back propagation algorithm, which specifically includes the following steps: Step S11: Initialize the population P, which includes the crossover scale, crossover probability and mutation probability. Real numbers are used for encoding, and the initial population is 20; Step S12: Calculate the evaluation function of each individual and rank them; Step S13: Perform crossover operation on individuals with a crossover probability of 0.6 to generate new individuals; Step S14: using the mutation probability, mutate to generate new individuals; Step S15: insert the new individual into the population P and calculate the evaluation function of the new individual; Step S16: Calculate the square sum of the errors of the neural network. If it reaches a preset value, go to step S17; otherwise, go to step S13 to continue genetic operation. Step S17: Using the optimized initial values ​​inherited by GA as initial weights and thresholds, the network is trained using the BP algorithm until the preset accuracy is met.

Images