1. A lithium battery SOH estimation method based on secondary fusion is characterized by comprising the following steps:

collecting an SOH data set of a lithium battery, wherein the SOH data set comprises SOH data of a lithium battery in a key life period;

dividing the SOH data set into M equally-spaced data sets according to the SOH size of each piece of data of the SOH data set;

respectively inputting each segment of data set into N models for training, and screening out all models with prediction precision meeting preset precision;

initializing training parameters of all models meeting preset precision corresponding to each data set, and performing Stacking fusion through a preset meta-learner to obtain a Stacking model corresponding to each data set;

taking the output of the Stacking model corresponding to each data set and each model with the prediction precision meeting the preset precision as the input of a layer of BP neural network for secondary fusion training to obtain a secondary fusion model corresponding to each data set;

inputting the obtained lithium battery data of the SOH to be predicted into an LSTM neural network to preliminarily predict the size of the SOH of the lithium battery;

and determining a data set segment into which the lithium battery data of the SOH to be predicted fall according to the size of the SOH of the lithium battery preliminarily predicted by the LSTM neural network, and selecting a secondary fusion model corresponding to the data set segment into which the lithium battery data of the SOH to be predicted fall to perform final SOH prediction.

2. The lithium battery SOH estimation method based on secondary fusion of claim 1, wherein after the step of collecting the SOH data set of the lithium battery, the SOH data set comprising SOH data of a critical life period of the lithium battery, and before the step of dividing the SOH data set into M-segment data sets with equal intervals according to the SOH size of each piece of data of the SOH data set, the method further comprises:

and judging whether the sample data of the SOH data set is enough, if not, expanding the SOH data set by adopting a generative countermeasure network.

3. The lithium battery SOH estimation method based on secondary fusion of claim 1, wherein the N models comprise: a support vector machine regression model, a neural network model, a decision tree model, a limit tree model, a K-nearest neighbor model, and a linear model.

4. The lithium battery SOH estimation method based on secondary fusion of claim 1, wherein the SOH prediction result expression of the secondary fusion model is as follows:

SOC_prediction＝W₁×SOC₁+W₂×SOC₂+…+W_n×SOC_n+W_n+1×SOC_n+1

Therein, SOC₁As a predictor of the Stacking model, SOC₂、…、SOC_n、SOC_n+1Predicted values, W, of the model satisfying the preset accuracy, respectively₁、W₂、W_n、W_n+1And respectively training the weights of the BP neural network.

5. A lithium battery SOH estimation system based on secondary fusion is characterized by comprising:

the data acquisition module is used for acquiring an SOH data set of the lithium battery, wherein the SOH data set comprises SOH data of the lithium battery in a key life period;

the data segmentation module is used for dividing the SOH data set into M equally-spaced data sets according to the SOH size of each piece of data of the SOH data set;

the training module is used for inputting each section of data set into N models respectively for training, and screening out all models with prediction precision meeting preset precision;

the fusion module is used for initializing training parameters of all models meeting preset precision corresponding to each section of data set and performing Stacking fusion through a preset meta-learner to obtain a Stacking model corresponding to each section of data set;

the secondary fusion module is used for performing secondary fusion training by taking the Stacking model corresponding to each section of data set and the output of each model with the prediction precision meeting the preset precision as the input of a layer of BP neural network to obtain a secondary fusion model corresponding to each section of data set;

the data input module is used for inputting the acquired lithium battery data of the SOH to be predicted into the LSTM neural network to preliminarily predict the size of the SOH of the lithium battery;

and the prediction output module is used for determining a data set segment into which the lithium battery data of the SOH to be predicted falls according to the size of the SOH of the lithium battery preliminarily predicted by the LSTM neural network, and selecting a secondary fusion model corresponding to the data set segment into which the lithium battery data of the SOH to be predicted falls to perform final SOH prediction.

6. The lithium battery SOH estimation system based on secondary fusion of claim 5, further comprising:

and the data expansion module is used for judging whether the sample data of the SOH data set is enough or not, and if not, expanding the SOH data set by adopting a generative countermeasure network.

7. The lithium battery SOH estimation system based on secondary fusion of claim 5, wherein the N models comprise: a support vector machine regression model, a neural network model, a decision tree model, a limit tree model, a K-nearest neighbor model, and a linear model.

8. The lithium battery SOH estimation system based on secondary fusion of claim 5, wherein the SOH prediction result expression of the secondary fusion model is as follows:

SOC_prediction＝W₁×SOC₁+W₂×SOC₂+…+W_n×SOC_n+W_n+1×SOC_n+1

Therein, SOC₁As a predictor of the Stacking model, SOC₂、…、SOC_n、SOC_n+1Predicted values, W, of the model satisfying the preset accuracy, respectively₁、W₂、W_n、W_n+1And respectively training the weights of the BP neural network.

Background

The lithium battery has the advantages of high energy density, long service life, reliable cleanness, good safety performance and the like, and is widely applied to electric automobiles, electronic products and energy storage equipment. However, in the long-term use process Of the lithium battery, excessive charging and discharging and local high temperature are inevitably generated, which leads to the condition that the Health State Of the lithium battery is reduced too fast, and the lithium battery aged to a certain degree not only greatly reduces the capacity, but also easily causes major safety accidents, so that the Health State (SOH) Of the lithium battery needs to be estimated.

The direct measurement method is simple in principle and can be directly applied to SOH estimation of each lithium battery, but the direct measurement method belongs to an open-loop method, is poor in robustness, has high requirements on a measuring instrument in order to achieve high accuracy, is harsh in measuring environment, is only suitable for laboratory environment, and is not suitable for popularization and application. The model-based method comprises an electrochemical model method and an equivalent circuit model method, for the electrochemical model, the parameter change of the chemical model is large due to the complex chemical reaction in the lithium battery, and the SOH estimation of the battery is difficult to carry out accurately, the equivalent circuit model can be divided into a Rint model, a Thevenin circuit, a first-order RC circuit, a second-order RC circuit and the like, generally speaking, although the state estimation of the battery is easy to carry out by a simple circuit, the precision is low, the complex circuit model has more parameters, and the parameters can change along with the aging of the battery, so that the calculated amount is large, and the expansion application is difficult. The data driving method does not need to clarify the internal change mechanism of the battery, realizes the estimation of the SOH of the lithium battery only through the data measured in the battery aging process and a machine learning method, and is an estimation method integrating practicability and accuracy. The conventional data-driven method uses a fusion model to perform SOH regression of a lithium battery, as shown in fig. 1, firstly, data sets are respectively input into models 1 to N for training, wherein the models 1 to N may adopt a conventional machine learning method, such as a support vector machine, a neural network, a decision tree, a limit tree, a K neighbor model, a linear model, and the like, in the training process, the models 1 to N are fused in a Stacking or bagging manner, and the fused model is used for estimating the SOH of the lithium battery. However, the existing SOH estimation method based on the fusion model only performs simple fusion on the sub-models, so that the used models cannot be sufficiently fitted with complex functional relationships, and the SOH estimation accuracy of the whole interval still cannot be guaranteed. Therefore, the method for estimating the SOH of the lithium battery based on the secondary fusion is provided, and is used for solving the technical problems that the conventional method for estimating the SOH of the lithium battery cannot be used for fully fitting a complex functional relation and cannot ensure high-precision estimation of the SOH of the whole interval.

Disclosure of Invention

The invention provides a lithium battery SOH estimation method and system based on secondary fusion, which are used for solving the technical problems that the conventional lithium battery SOH estimation method cannot fully fit a complex functional relation and cannot ensure high-precision estimation of the whole-interval SOH.

In view of this, the first aspect of the present invention provides a lithium battery SOH estimation method based on secondary fusion, including:

collecting an SOH data set of a lithium battery, wherein the SOH data set comprises SOH data of a lithium battery in a key life period;

dividing the SOH data set into M equally-spaced data sets according to the SOH size of each piece of data of the SOH data set;

respectively inputting each segment of data set into N models for training, and screening out all models with prediction precision meeting preset precision;

initializing training parameters of all models meeting preset precision corresponding to each data set, and performing Stacking fusion through a preset meta-learner to obtain a Stacking model corresponding to each data set;

taking the output of the Stacking model corresponding to each data set and each model with the prediction precision meeting the preset precision as the input of a layer of BP neural network for secondary fusion training to obtain a secondary fusion model corresponding to each data set;

inputting the obtained lithium battery data of the SOH to be predicted into an LSTM neural network to preliminarily predict the size of the SOH of the lithium battery;

and determining a data set segment into which the lithium battery data of the SOH to be predicted fall according to the size of the SOH of the lithium battery preliminarily predicted by the LSTM neural network, and selecting a secondary fusion model corresponding to the data set segment into which the lithium battery data of the SOH to be predicted fall to perform final SOH prediction.

Optionally, after the step of collecting a SOH data set of the lithium battery, the SOH data set including SOH data of a critical life period of the lithium battery, before the step of dividing the SOH data set into M equally spaced data sets according to the SOH size of each piece of data of the SOH data set, the method further includes:

and judging whether the sample data of the SOH data set is enough, if not, expanding the SOH data set by adopting a generative countermeasure network.

Optionally, the N models include: a support vector machine regression model, a neural network model, a decision tree model, a limit tree model, a K-nearest neighbor model, and a linear model.

Optionally, the SOH prediction result expression of the secondary fusion model is:

SOC_prediction＝W₁×SOC₁+W₂×SOC₂+…+W_n×SOC_n+W_n+1×SOC_n+1

Therein, SOC₁As a predictor of the Stacking model, SOC₂、…、SOC_n、SOC_n+1Predicted values, W, of the model satisfying the preset accuracy, respectively₁、W₂、W_n、W_n+1And respectively training the weights of the BP neural network.

The invention provides a lithium battery SOH estimation system based on secondary fusion, which comprises:

the data acquisition module is used for acquiring an SOH data set of the lithium battery, wherein the SOH data set comprises SOH data of the lithium battery in a key life period;

the data segmentation module is used for dividing the SOH data set into M equally-spaced data sets according to the SOH size of each piece of data of the SOH data set;

the training module is used for inputting each section of data set into N models respectively for training, and screening out all models with prediction precision meeting preset precision;

the fusion module is used for initializing training parameters of all models meeting preset precision corresponding to each section of data set and performing Stacking fusion through a preset meta-learner to obtain a Stacking model corresponding to each section of data set;

the secondary fusion module is used for performing secondary fusion training by taking the Stacking model corresponding to each section of data set and the output of each model with the prediction precision meeting the preset precision as the input of a layer of BP neural network to obtain a secondary fusion model corresponding to each section of data set;

the data input module is used for inputting the acquired lithium battery data of the SOH to be predicted into the LSTM neural network to preliminarily predict the size of the SOH of the lithium battery;

and the prediction output module is used for determining a data set segment into which the lithium battery data of the SOH to be predicted falls according to the size of the SOH of the lithium battery preliminarily predicted by the LSTM neural network, and selecting a secondary fusion model corresponding to the data set segment into which the lithium battery data of the SOH to be predicted falls to perform final SOH prediction.

Optionally, the method further comprises:

and the data expansion module is used for judging whether the sample data of the SOH data set is enough or not, and if not, expanding the SOH data set by adopting a generative countermeasure network.

Optionally, the N models include: a support vector machine regression model, a neural network model, a decision tree model, a limit tree model, a K-nearest neighbor model, and a linear model.

Optionally, the SOH prediction result expression of the secondary fusion model is:

SOC_prediction＝W₁×SOC₁+W₂×SOC₂+…+W_n×SOC_n+W_n+1×SOC_n+1

Therein, SOC₁As a predictor of the Stacking model, SOC₂、…、SOC_n、SOC_n+1Predicted values, W, of the model satisfying the preset accuracy, respectively₁、W₂、W_n、W_n+1And respectively training the weights of the BP neural network.

According to the technical scheme, the embodiment of the invention has the following advantages:

the invention provides a lithium battery SOH estimation method based on secondary fusion, which comprises the following steps: collecting an SOH data set of a lithium battery, wherein the SOH data set comprises SOH data of a lithium battery in a key life period; dividing the SOH data set into M equally-spaced data sets according to the SOH size of each piece of data of the SOH data set; respectively inputting each segment of data set into N models for training, and screening out all models with prediction precision meeting preset precision; initializing training parameters of all models meeting preset precision corresponding to each data set, and performing Stacking fusion through a preset meta-learner to obtain a Stacking model corresponding to each data set; taking the output of the Stacking model corresponding to each data set and each model with the prediction precision meeting the preset precision as the input of a layer of BP neural network for secondary fusion training to obtain a secondary fusion model corresponding to each data set; inputting the obtained lithium battery data of the SOH to be predicted into an LSTM neural network to preliminarily predict the size of the SOH of the lithium battery; and determining a data set segment into which the lithium battery data of the SOH to be predicted fall according to the size of the SOH of the lithium battery preliminarily predicted by the LSTM neural network, and selecting a secondary fusion model corresponding to the data set segment into which the lithium battery data of the SOH to be predicted fall to perform final SOH prediction. Because a single model is only sensitive to a specific functional relationship, the model has a good effect on fitting the functional relationship with small change degree, so that the data set is segmented according to the size of the SOH, the input and output functional relationships of the corresponding models of each segment are stable, and the fitting accuracy of the models is improved; according to the characteristic that a plurality of submodels exist in the fusion regression model, specific weights are given to different submodels through the BP neural network, the functional relation of each section of data set can be fitted in a targeted manner, the advantages of each submodel are exerted to the maximum extent to fit the SOH functional relation of different intervals, and therefore the accuracy and the stability of SOH estimation are further improved compared with those of a common fusion model. The technical problems that the existing lithium battery SOH estimation method cannot fully fit a complex functional relation and cannot ensure the high-precision estimation of the whole-interval SOH are solved.

Furthermore, the invention adopts a Generative Adaptive Network (GAN) to expand the data set, and the GAN can still expand data conforming to the rule of the original data set without using a complex Markov chain, thereby training the machine learning model more fully.

Drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings needed to be used in the description of the embodiments are briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments described in the present invention, and it is obvious for those skilled in the art to obtain other drawings according to these drawings.

FIG. 1 is a schematic diagram of a conventional data-driven method for estimating SOH of a lithium battery using a fusion model;

fig. 2 is a schematic flow chart of a lithium battery SOH estimation method based on secondary fusion provided in an embodiment of the present invention;

FIG. 3 is a flow chart of data set processing provided in an embodiment of the present invention;

fig. 4 is a schematic processing flow diagram of a secondary fusion model for the ith segment of data set provided in the embodiment of the present invention;

FIG. 5 is a schematic structural diagram of a single-layer neural network for secondary fusion provided in an embodiment of the present invention;

FIG. 6 is a flow chart illustrating the process of predicting the SOH of a lithium battery using a secondary fusion model according to an embodiment of the present invention;

fig. 7 is a schematic structural diagram of a lithium battery SOH estimation system based on secondary fusion provided in an embodiment of the present invention.

Detailed Description

In order to make the technical solutions of the present invention better understood, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

For easy understanding, referring to fig. 1, the present invention provides an embodiment of a method for estimating SOH of a lithium battery based on secondary fusion, including:

step 101, collecting an SOH data set of a lithium battery, wherein the SOH data set comprises SOH data of a lithium battery in a critical life period.

It should be noted that, before model training, an SOH data set of a lithium battery needs to be acquired, and sample data of the SOH data set may be represented as:

SOH＝f(U_t1,U_t2,…,U_tn,I_t1,I_t2,…,I_tn,T_t1,T_t2,…,T_tn,C_t1,C_t2,…C_tn)

wherein, U_ti，I_ti，T_ti，C_tiAnd respectively represent the voltage, the current, the temperature and the current available electric quantity which are respectively measured from the lithium battery at the moment ti.

The SOH of the lithium battery reflects the aging degree of the lithium battery, so that the SOH data of the lithium battery in each life cycle, i.e., the critical life time period, needs to be collected, and generally, 10% of the SOH data can be used as a critical node, and the critical life time period is (100% -91%, 90% -81%, 80% -71%, 70% -61%, 60% -51%, 50% -41%, 40% -31%, 30% -21%, 20% -11%, 10% -1%).

In one embodiment, it is further determined whether the sample data of the SOH data set is sufficient, and if not, the SOH data set is extended by using a generative countermeasure network, as shown in fig. 3. Because the collection of lithium battery data needs to pass through a lengthy charging and discharging process, and meanwhile, different physical quantities need to be accurately measured during the collection, a large amount of manpower, material resources and financial resources can be consumed, and the collection of large-scale lithium battery data is very difficult, so that the expansion of public or self-testing data by utilizing a machine learning technology is very important. Data set expansion generally expands the data volume in the data set through various technical means of machine learning, and makes the newly added data rule similar to the original data rule. The invention adopts a generating type confrontation network (GAN) to expand a data set, and the principle is that a generator continuously generates a newly added data set with a rule similar to the original data rule, the newly added data set is put into a discriminator to be discriminated, when the discriminator judges that the newly added data is similar to the original data rule, the newly added data set is formally generated, and when the discriminator judges that the newly added data is greatly different from the original data rule, the generator is retrained until the newly added data set meeting the requirement is generated, and parameters such as iteration times, network structures, batch training amount and the like are continuously adjusted during the training period, so that the satisfied newly added data set is obtained. And compared with data set expansion methods such as a Boltzmann machine and GSNs (generalized likelihood factors), the GAN only uses back propagation and does not need a complicated Markov chain. Compared with data set expansion methods such as white noise and self-encoder, the method can generate data which is more real and accords with the rule of the original data set.

Step 102, dividing the SOH data set into M data sets with equal intervals according to the SOH size of each piece of data of the SOH data set.

In the data expression of SOH, the function f varies with the degree of aging of the battery, and it can be considered that when the degree of aging of the battery does not vary greatly, the functional relation f also does not vary greatly, and the important index reflecting the degree of aging of the battery is the SOH size. The relatively stable f ensures that the fitting of the model to the functional relation is more stable and the fitting effect is more accurate. It is noted that, with the use of lithium battery, the SOH basically decreases gradually with the increase of the battery cycle number, the invention segments the data set into M segments according to the size range of the SOH, each data of the first segment of data set corresponds to the SOH at X₀-X₁In between, each data in the second segment data set corresponds to a SOH of X₁-X₂In between, each data set of the M-th data set corresponds to SOH at X_M-1-X_MIn which X₀＞X₁＞…＞X_M-1＞X_MSo that the SOH of each segment is f (U)_t1，U_t2，...，U_tn，It₁，It₂，...，I_tn，T_t1，T_t2，...，T_tn，C_t1，C_t2，...，C_tn) The functional relationship is close, and the corresponding cycle times of each piece of data are also close.

And 103, inputting each segment of data set into N models respectively for training, and screening out all models with prediction precision meeting preset precision.

And 104, initializing training parameters of all models meeting preset precision corresponding to each data set, and performing Stacking fusion through a preset meta-learner to obtain a Stacking model corresponding to each data set.

It should be noted that, as shown in fig. 4, each segment of data set is trained by using N models, all models whose prediction accuracy meets the preset accuracy are screened out, and only submodels meeting the accuracy requirements are subjected to Stacking fusion. The number of N models used varies, and the types of models may be various machine learning models such as support vector regression, neural networks, decision trees, limit trees, K-nearest neighbor models, linear models, ridge regression, and the like.

And 105, performing secondary fusion training by taking the Stacking model corresponding to each data set and the output of each model with the prediction precision meeting the preset precision as the input of a layer of BP neural network to obtain a secondary fusion model corresponding to each data set.

Because f does not vary much for each dataset, it is not certain how f is a function of itself. In fact, each sub-model is sensitive to a specific functional relationship, so that the fitting of the functional relationship f is realized by adopting a quadratic fusion model. Because the functional relation f of the data set can be fitted to the maximum extent through different weight combinations of different types of sub models. And performing secondary fusion on the submodel meeting the precision requirement and the Stacking model together, and automatically allocating proper weight coefficients to the submodel meeting the precision requirement and the Stacking model subjected to primary fusion by utilizing the process of single-layer neural network BP training during the secondary fusion. Therefore, the advantages of each sub-model can be exerted to the maximum extent to fit the SOH of different intervals, and the accuracy and stability of the SOH estimation are further improved compared with those of a common fusion model. The structure of the single-layer neural network is shown in fig. 5. Training the neural network of one layer to distribute trained weight to the Stacking model and the submodel with the accuracy RMSE meeting the specified threshold, multiplying the output of the submodel with the prediction accuracy RMSE meeting the specified threshold and the fused Stacking by the corresponding distributed weight, and adding to obtain the final SOH predicted value of the battery, wherein the calculation process is as follows:

SOC_prediction＝W₁×SOC₁+W₂×SOC₂+…+W_n×SOC_n+W_n+1×SOC_n+1

Therein, SOC₁As a predictor of the Stacking model, SOC₂、…、SOC_n、SOC_n+1Predicted values, W, of the model satisfying the preset accuracy, respectively₁、W₂、W_n、W_n+1Respectively, the weights of the BP neural network.

Step 106, inputting the obtained lithium battery data of the SOH to be predicted into an LSTM neural network to preliminarily predict the size of the SOH of the lithium battery;

and 107, determining a data set segment into which the lithium battery data of the SOH to be predicted fall according to the size of the lithium battery SOH preliminarily predicted by the LSTM neural network, and selecting a secondary fusion model corresponding to the data set segment into which the lithium battery data of the SOH to be predicted fall to perform final SOH prediction.

The secondary fusion model is used for predicting the SOH of the lithium battery, the flow is as shown in figure 6, data are collected from the battery to be subjected to SOH estimation, the obtained lithium battery data of the SOH to be predicted are input into an LSTM neural network to predict the size of the SOH of the lithium battery, the size of the SOH can be rapidly and preliminarily predicted by the pre-trained LSTM neural network (the value at the moment is a rough predicted value), then which data set the data correspond to is judged according to the SOH predicted by the LSTM neural network, the collected data are subjected to accurate SOH prediction estimation by the secondary fusion model corresponding to the data set, and an SOH prediction result is output.

According to the method, the data sets are segmented according to the size of the SOH, so that the input-output function relation of the model corresponding to each segment of data set is stable, and the model fitting accuracy is improved; according to the characteristic that a plurality of submodels exist in the fusion regression model, specific weights are given to different submodels through the BP neural network, the functional relation of each section of data set can be fitted in a targeted manner, the advantages of each submodel are exerted to the maximum extent to fit the SOH functional relation of different intervals, and therefore the accuracy and the stability of SOH estimation are further improved compared with those of a common fusion model. The technical problems that the existing lithium battery SOH estimation method cannot fully fit a complex functional relation and cannot ensure the high-precision estimation of the whole-interval SOH are solved.

For easy understanding, please refer to fig. 7, an embodiment of the present invention provides a system for estimating SOH of a lithium battery based on secondary fusion, including:

the data acquisition module is used for acquiring an SOH data set of the lithium battery, wherein the SOH data set comprises SOH data of the lithium battery in a key life period;

the data segmentation module is used for dividing the SOH data set into M equally-spaced data sets according to the SOH size of each piece of data of the SOH data set;

the training module is used for inputting each section of data set into N models respectively for training, and screening out all models with prediction precision meeting preset precision;

the fusion module is used for initializing training parameters of all models meeting preset precision corresponding to each section of data set and performing Stacking fusion through a preset meta-learner to obtain a Stacking model corresponding to each section of data set;

the secondary fusion module is used for performing secondary fusion training by taking the Stacking model corresponding to each section of data set and the output of each model with the prediction precision meeting the preset precision as the input of a layer of BP neural network to obtain a secondary fusion model corresponding to each section of data set;

the data input module is used for inputting the acquired lithium battery data of the SOH to be predicted into the LSTM neural network to preliminarily predict the size of the SOH of the lithium battery;

and the prediction output module is used for determining a data set segment into which the lithium battery data of the SOH to be predicted falls according to the size of the SOH of the lithium battery preliminarily predicted by the LSTM neural network, and selecting a secondary fusion model corresponding to the data set segment into which the lithium battery data of the SOH to be predicted falls to perform final SOH prediction.

Further comprising:

and the data expansion module is used for judging whether the sample data of the SOH data set is enough or not, and if not, expanding the SOH data set by adopting a generative countermeasure network.

The N models include: a support vector machine regression model, a neural network model, a decision tree model, a limit tree model, a K-nearest neighbor model, and a linear model.

The SOH prediction result expression of the secondary fusion model is as follows:

SOC_prediction＝W₁×SOC₁+W₂×SOC₂+…+W_n×SOC_n+W_n+1×SOC_n+1

Therein, SOC₁As a predictor of the Stacking model, SOC₂、…、SOC_n、SOC_n+1Predicted values, W, of the model satisfying the preset accuracy, respectively₁、W₂、W_n、W_n+1Respectively, the weights of the BP neural network.

It should be noted that, before model training, an SOH data set of a lithium battery needs to be acquired, and sample data of the SOH data set may be represented as:

SOH＝f(U_t1，U_t2，…，U_tn，I_t1，I_t2，…，I_tn，T_t1，T_t2，…，T_tn，C_t1，C_t2，…C_tn)

wherein the content of the first and second substances,U_ti，I_ti，T_ti，C_tiand respectively represent the voltage, the current, the temperature and the current available electric quantity which are respectively measured from the lithium battery at the moment ti.

The SOH of the lithium battery reflects the aging degree of the lithium battery, so that the SOH data of the lithium battery in each life cycle, i.e., the critical life time period, needs to be collected, and generally, 10% of the SOH data can be used as a critical node, and the critical life time period is (100% -91%, 90% -81%, 80% -71%, 70% -61%, 60% -51%, 50% -41%, 40% -31%, 30% -21%, 20% -11%, 10% -1%).

In one embodiment, it is further determined whether the sample data of the SOH data set is sufficient, and if not, the SOH data set is extended by using a generative countermeasure network. Because the collection of lithium battery data needs to pass through a lengthy charging and discharging process, and meanwhile, different physical quantities need to be accurately measured during the collection, a large amount of manpower, material resources and financial resources can be consumed, and the collection of large-scale lithium battery data is very difficult, so that the expansion of public or self-testing data by utilizing a machine learning technology is very important. Data set expansion generally expands the data volume in the data set through various technical means of machine learning, and makes the newly added data rule similar to the original data rule. The invention adopts a generating type confrontation network (GAN) to expand a data set, and the principle is that a generator continuously generates a newly added data set with a rule similar to the original data rule, the newly added data set is put into a discriminator to be discriminated, when the discriminator judges that the newly added data is similar to the original data rule, the newly added data set is formally generated, and when the discriminator judges that the newly added data is greatly different from the original data rule, the generator is retrained until the newly added data set meeting the requirement is generated, and parameters such as iteration times, network structures, batch training amount and the like are continuously adjusted during the training period, so that the satisfied newly added data set is obtained. And compared with data set expansion methods such as a Boltzmann machine and GSNs (generalized likelihood factors), the GAN only uses back propagation and does not need a complicated Markov chain. Compared with data set expansion methods such as white noise and self-encoder, the method can generate data which is more real and accords with the rule of the original data set.

In the data expression of SOH, the function f varies with the degree of aging of the battery, and it can be considered that when the degree of aging of the battery does not vary greatly, the functional relation f also does not vary greatly, and the important index reflecting the degree of aging of the battery is the SOH size. The relatively stable f ensures that the fitting of the model to the functional relation is more stable and the fitting effect is more accurate. It is noted that, with the use of lithium battery, the SOH basically decreases gradually with the increase of the battery cycle number, the invention segments the data set into M segments according to the size range of the SOH, each data of the first segment of data set corresponds to the SOH at X₀-X₁In between, each data in the second segment data set corresponds to a SOH of X₁-X₂In between, each data set of the M-th data set corresponds to SOH at X_M-1-X_MIn which X₀＞X₁＞…＞X_M-1＞X_MSo that the SOH of each segment is f (U)_t1，U_t2，...，U_tn，I_t1，I_t2，...，I_tn，T_t1，T_t2，...，T_tn，C_t1，C_t2，...，C_tn) The functional relationship is close, and the corresponding cycle times of each piece of data are also close.

And (3) training each section of data set by using N models, screening out all models with prediction precision meeting preset precision, and only performing Stacking fusion on the sub-models meeting the precision requirement. The number of N models used varies, and the types of models may be various machine learning models such as support vector regression, neural networks, decision trees, limit trees, K-nearest neighbor models, linear models, ridge regression, and the like.

Because f does not vary much for each dataset, it is not certain how f is a function of itself. In fact, each sub-model is sensitive to a specific functional relationship, so that the fitting of the functional relationship f is realized by adopting a quadratic fusion model. Because the functional relation f of the data set can be fitted to the maximum extent through different weight combinations of different types of sub models. And performing secondary fusion on the submodel meeting the precision requirement and the Stacking model together, and automatically allocating proper weight coefficients to the submodel meeting the precision requirement and the Stacking model subjected to primary fusion by utilizing the process of single-layer neural network BP training during the secondary fusion. Therefore, the advantages of each sub-model can be exerted to the maximum extent to fit the SOH of different intervals, and the accuracy and stability of the SOH estimation are further improved compared with those of a common fusion model. The structure of the single-layer neural network is shown in fig. 4. Training the neural network of one layer to distribute trained weight to the Stacking model and the submodel with the accuracy RMSE meeting the specified threshold, multiplying the output of the submodel with the prediction accuracy RMSE meeting the specified threshold and the fused Stacking by the corresponding distributed weight, and adding to obtain the final SOH predicted value of the battery, wherein the calculation process is as follows:

SOC_prediction＝W₁×SOC₁+W₂×SOC₂+…+W_n×SOC_n+W_n+1×SOC_n+1

Therein, SOC₁As a predictor of the Stacking model, SOC₂、…、SOC_n、SOC_n+1Predicted values, W, of the model satisfying the preset accuracy, respectively₁、W₂、W_n、W_n+1Respectively, the weights of the BP neural network.

The secondary fusion model is used for predicting the SOH of the lithium battery, the flow is as shown in figure 5, data are collected from the battery to be subjected to SOH estimation, the obtained lithium battery data of the SOH to be predicted are input into an LSTM neural network to predict the size of the SOH of the lithium battery, the size of the SOH can be rapidly and preliminarily predicted by the pre-trained LSTM neural network (the value at the moment is a rough predicted value), then which data set the data correspond to is judged according to the SOH predicted by the LSTM neural network, the collected data are subjected to accurate SOH prediction estimation by the secondary fusion model corresponding to the data set, and an SOH prediction result is output.

According to the method, the data sets are segmented according to the size of the SOH, so that the input-output function relation of the model corresponding to each segment of data set is stable, and the model fitting accuracy is improved; according to the characteristic that a plurality of submodels exist in the fusion regression model, specific weights are given to different submodels through the BP neural network, the functional relation of each section of data set can be fitted in a targeted manner, the advantages of each submodel are exerted to the maximum extent to fit the functional relation of different intervals, and therefore the accuracy and the stability of SOH estimation are further improved compared with those of a common fusion model. The technical problems that the existing lithium battery SOH estimation method cannot fully fit a complex functional relation and cannot ensure the high-precision estimation of the whole-interval SOH are solved.

The above-mentioned embodiments are only used for illustrating the technical solutions of the present invention, and not for limiting the same; although the present invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some technical features may be equivalently replaced; and such modifications or substitutions do not depart from the spirit and scope of the corresponding technical solutions of the embodiments of the present invention.