CN112798961B

CN112798961B - Method for predicting remaining service life of power battery of electric automobile

Info

Publication number: CN112798961B
Application number: CN202110224012.2A
Authority: CN
Inventors: 肖迁; 穆云飞; 贾宏杰; 陆文标; 侯恺; 徐宪东; 余晓丹
Original assignee: Tianjin University
Current assignee: Tianjin University
Priority date: 2021-02-27
Filing date: 2021-02-27
Publication date: 2023-01-10
Anticipated expiration: 2041-02-27
Also published as: CN112798961A

Abstract

The invention discloses a method for predicting the residual service Life of a power battery of an electric automobile, which comprises the following steps of firstly, improving a Light Gradient Boosting Machine (LightGBM) through an Adaptive Robust Loss Function (ARLF), and improving the robustness of prediction of the residual service Life (RUL); then, processing the capacity attenuation data of the electric automobile in the whole life cycle under the running condition based on the multi-thread parallel histogram; secondly, formulating an RUL prediction model training mode based on a Leaf-wise growth mode limited by the bandwidth; thirdly, automatically adjusting and optimizing to obtain model optimal parameters based on meta-learning hyper-parameters and setting a performance evaluation function; and finally, the prediction of the residual service life is completed by utilizing the RUL prediction model for obtaining the optimal parameters.

Description

Method for predicting remaining service life of power battery of electric automobile

Technical Field

The invention relates to the field of electric vehicle battery fault prediction and health management, in particular to a method for predicting the remaining service life of a power battery of an electric vehicle.

Background

According to the energy storage industry research white paper 2020 issued by the energy storage industry alliance in Guancun, the cumulative installation scale of the lithium ion battery accounts for the largest proportion of the electrochemical energy storage, and is 88.8%. The lithium ion battery is widely applied to the fields of electric vehicles, energy storage systems, mobile wearing and the like by virtue of the advantages of high energy density, long service life, low self-discharge rate, wide working temperature range and the like.

As the number of times of charge and discharge cycles of a lithium ion battery increases, electrochemical reactions in the battery become more severe, and in addition to redox reactions occurring due to lithium ion deintercalation, there are many side reactions, such as decomposition of an electrolyte, dissolution of an active material, deposition of metallic lithium, etc., which cause capacity degradation of the battery. When the capacity of the battery is attenuated to 70-80% of the nominal capacity, the charging and discharging performance of the battery is seriously influenced, even the battery cannot be continuously used, and the battery is required to be replaced in time in order to avoid the occurrence of faults. As an important function of lithium ion battery fault Prediction and Health Management (PHM), remaining service Life (RUL) prediction can provide references for regular maintenance and safe and stable operation of a battery, reduce high maintenance cost, and reduce the occurrence probability of catastrophic consequences.

In order to achieve accurate RUL prediction, many researchers have conducted relevant studies. The model-based RUL prediction method generally utilizes prior knowledge to model a battery degradation mechanism, and identifies parameters in a model by means of a least square method or an observer through battery full-life cycle test data and an electrochemical alternating-current impedance spectrum. Although the RUL degradation model in the method can consider the influence of different temperature stresses and unobservable state quantities on the battery capacity degradation, has definite physical significance, and usually ignores the environmental conditions of the battery service, such as vibration stress and the like. The battery model parameters are different along with the change of a service environment, so that the model precision of fixed parameters is reduced along with the aging of a battery, and in order to improve the prediction precision, the model needs to be corrected according to different battery models and working states, so that huge parameter identification calculation amount and a complex physical model are brought, and the extended application of the battery model is difficult to realize. Meanwhile, the method belongs to an open-loop method, and robustness is difficult to consider.

Aiming at the problem that outliers exist in capacity data, the sensitivity of a common error form in machine learning to large errors is higher than that of small errors, namely, a model is biased to reduce the maximum error, the influence of the outliers on a prediction effect cannot be effectively reduced, and the common error form cannot be adaptively adjusted according to sample characteristics.

Disclosure of Invention

The invention aims to overcome the defects in the prior art, and provides a method for predicting the residual service life of a power battery of an electric automobile, which is described in detail in the following:

a method for predicting the residual service life of a power battery of an electric automobile comprises the following steps:

(1) Obtaining parameters related to the adaptive loss function: after acquiring the capacity data of the battery in the full life cycle under the running condition, acquiring a hyperparameter alpha and a coordination parameter c through a self-adaptive robust loss function, and configuring a loss function of a LightGBM;

(2) Processing data based on multi-threaded parallel histograms: for each dimension feature, dividing continuous floating point type data into discrete k ranges to obtain k "buckets" (bins), and constructing a histogram with a width of k, as shown in step 2 in fig. 1. Compared with the traditional Boosting mode, only k bins need to be traversed, and characteristic values do not need to be compared one by one, so that the calculation amount is greatly reduced, and the training speed is improved. Considering that the decision tree is a weak model, k determines the degree of regularization, thereby avoiding overfitting;

(3) Training a model based on a Leaf-wise growth mode with bandwidth limitation: as shown in step 3 in fig. 1, the growth mode according to Leaf growth (Leaf-wise) is that when each node is split, the Leaf node with the largest gain is selected for growth, so that the complexity of the model is limited, the training speed is increased, and overfitting is avoided;

(4) Obtaining model optimal parameters based on meta-learning hyper-parameter automatic optimization and setting a performance evaluation function: and constructing a model parameter space, a LightGBM model factory and a score acquirer according to the Hyperopt hyper-parameter optimization framework. The model factory is used for producing the required model, and the score obtainer is used for decoupling. And obtaining the optimal parameters of the model through multiple rounds of iteration. Setting an evaluation standard: using a global Root Mean Square Error (RMSE), a local RMSE, a global Mean Absolute Error (MAE), a local MAE, and an iteration number between the actual capacity and the predicted capacity as evaluation criteria;

(5) And completing the prediction of the residual service life: the remaining service life of the battery is the number of charge-discharge cycles before the available capacity of the battery is attenuated to 70-80% of the standard capacity (the service of the battery is finished), the full life cycle capacity data of the battery under the driving working condition is divided into a training set and a testing set, the training set is used for training a model and obtaining related parameters, and the testing set is used for verifying the effectiveness of the model;

the technical scheme provided by the invention has the beneficial effects that:

1) Compared with the traditional RUL prediction method, the RUL prediction method provided by the invention not only has higher prediction precision, but also has better adaptability, better meets the actual requirement, and is beneficial to improving the generalization capability of the RUL prediction model;

2) As the residual x (> | c |) increases, the influence of the RUL prediction method on the gradient is reduced. Therefore, in the process of gradient descent, the influence of outliers on the RUL prediction is effectively reduced;

3) Compared with the traditional loss function, the loss function used by the RUL prediction method provided by the invention can stably achieve the optimal effect under fewer iteration times, and has good robustness.

Drawings

FIG. 1 is a flow chart of the improved LightGBM based battery RUL prediction;

FIG. 2 is a system for testing the full life cycle capacity of a battery under driving conditions;

FIG. 3 is a power density spectrum of a vibration table;

FIG. 4 is a battery charge and discharge test curve under driving conditions;

FIG. 5 shows the capacity fading of a lithium ion battery under two working conditions;

FIG. 6 is a loss function;

FIG. 7 is the first derivative of the loss function;

FIG. 8 is a Loss value;

FIG. 9 shows the RUL prediction results;

FIG. 10 is a comparison of the effect of RUL prediction based on different loss functions.

Detailed Description

In order to reduce the influence of outliers on prediction accuracy and improve the generalization capability and robustness of a prediction model, the invention provides a method for predicting the residual service life of a power battery of an electric vehicle. In order to make the objects, technical solutions and advantages of the present invention more apparent, embodiments of the present invention are described in further detail below.

step 101: after acquiring the capacity data of the battery in the full life cycle under the driving condition, acquiring a super parameter alpha and a coordination parameter c through a self-adaptive robust loss function, and configuring a loss function of a LightGBM;

step 102: for each dimension feature, dividing continuous floating point type data into discrete k ranges to obtain k "buckets" (bins), and constructing a histogram with a width of k, as shown in step 2 in fig. 1. Compared with the traditional Boosting mode, only k bins need to be traversed, and characteristic values do not need to be compared one by one, so that the computation amount is greatly reduced, and the training speed is improved. Considering that the decision tree is a weak model, k determines the degree of regularization, thereby avoiding overfitting;

step 103: as shown in step 3 in FIG. 1, the Leaf-wise growth (Leaf-wise) mode is that the Leaf node with the maximum gain is selected to grow each time the node is split, the complexity of the model is limited, the training speed is increased, and overfitting is avoided

Step 104: and constructing a model parameter space, a LightGBM model factory and a score acquirer according to the Hyperopt hyper-parameter optimization framework. The model factory is used for producing the required model, and the fraction obtainer is used for decoupling. And obtaining the optimal parameters of the model through multiple iterations. Setting an evaluation standard: using a global Root Mean Square Error (RMSE), a local RMSE, a global Mean Absolute Error (MAE), a local MAE, and an iteration number between the actual capacity and the predicted capacity as evaluation criteria;

step 105: the remaining service life of the battery is the number of charge-discharge cycles before the available capacity of the battery is attenuated to 70-80% of the standard capacity (the service of the battery is finished), the full life cycle capacity data of the battery under the driving working condition is divided into a training set and a testing set, the training set is used for training a model and obtaining related parameters, and the testing set is used for verifying the effectiveness of the model.

The scheme in embodiment 1 is further described below with reference to specific calculation formulas, drawings and examples, and is described in detail in the following description:

example 2:

step 201: obtaining relevant parameters of an adaptive loss function;

wherein, the preliminary adaptive robust loss function form is as follows:

in the formula, c may be regarded as a tuning parameter for adjusting the scale of curve bending of the loss function 1 at x =0. The adaptive robust loss function mainly controls the robustness of the loss function through the hyper-parameter alpha, namely, when the alpha is different values, the adaptive robust loss function not only can represent other loss functions, but also can represent proper loss functions according to data characteristics. Although α is not defined at α =0 (Cauchy Loss), α =2 (L2 Loss), α = - ∞ (Welsch Loss), it can be approximated by taking the limit. Therefore, the final adaptive robust loss function can be expressed as a function set form as in equation (13).

The first derivative expression of the adaptive robust loss function is as follows:

and improving the original loss function of the LightGBM through an adaptive robust loss function.

LightGBM takes a decision tree as a base learner, which can be represented as:

in the formula, H _t Represents the ith learner and theta represents the collection space of all the learners.

LightGBM continuously improves the performance of the learner through multiple iterations, and the learner is used for acquiring the auxiliary input space X _s Mapping function to gradient space G. Assume that there is a training set of n data volumes { x } ₁ ,…,x _n In which x _i Is a space x _s The ith dimension is the vector of s. If the learner obtained in the previous iteration is H _t-1 (x) The loss function is L (y, H) _t-1 (x) The objective of the iteration is to find weak learner h) _t (x) The loss function of the current round is minimized as follows:

calculating the negative gradient of the loss function for obtaining an approximate value of the loss function of the current round, which can be expressed as:

the objective function is usually the squared difference, h _t (x) Can be approximately expressed as:

finally obtaining the strong learner of the iteration:

H _t (x)＝H _t-1 (x)+h _t (x) (8)

step 202: processing data based on the multi-threaded parallel histogram;

for each one-dimensional feature, dividing continuous floating-point data into discrete k ranges to obtain k "buckets" (bins), and constructing a histogram with a width of k, as shown in step 2 in fig. 1. Compared with the traditional Boosting mode, only k bins need to be traversed, and characteristic values do not need to be compared one by one, so that the calculation amount is greatly reduced, and the training speed is improved. Given that the decision tree is a weak model, k determines the degree of regularization, thus avoiding overfitting.

Step 203: training a model based on a Leaf-wise growth mode limited by the bandwidth;

as shown in step 3 in fig. 1, the growth mode according to Leaf growth (Leaf-wise) is that a Leaf node with the largest gain is selected to grow when each node is split, so that the complexity of the model is limited, the training speed is increased, and overfitting is avoided. In each gradient lifting iteration, the output value of the negative gradient of the current model loss function is { g } ₁ ,...,g _n In which g is _i Is x _i The corresponding negative gradient of the penalty function is at the value of the current model output. The basis learner segments at feature splitting points where the information gain is greatest, and the information gain is measured by the split back difference.

Let O be the data set in one fixed node of the base model. The variance gain of the feature at this node j at the segmentation point d is defined as:

in the formula, n _O Number of training set samples for a certain fixed leaf node, and n _O ＝∑I[x _i ∈O]，

The j-th characteristic median is less than or equal to the number of samples of d

Is the number of samples whose j-th median is greater than d

Traversing each split point of each feature to find

Calculating a corresponding maximum information gain

According to the point of splitting

The data is divided into left and right child nodes.

Step 204: automatically optimizing and obtaining model optimal parameters based on meta-learning hyper-parameters and setting a performance evaluation function;

the optimal parameters of the model are obtained based on meta-learning Hyperopt, and a performance evaluation function is set. And constructing a model parameter space, a LightGBM model factory and a score acquirer according to the Hyperopt hyper-parameter optimization framework. The model factory is used for producing the required model, and the fraction obtainer is used for decoupling. Through multiple iterations, the optimal parameters of the model are obtained, as shown in table 1.

TABLE 1 RUL prediction model hyper-parameters

Setting an evaluation standard: global Root Mean Square Error (RMSE), local RMSE, global Absolute Error (MAE), local MAE, and number of iterations between actual and predicted capacities were used as evaluation criteria. The global is the cycle number of the full life cycle, and the local is the cycle number of the capacity rebound position (namely 1300-1600 cycles). The number of iterations is the number of iteration rounds until the Loss value (Loss value) reaches a minimum.

Global

In the formula (I), the compound is shown in the specification,

is the ith cycle predicted capacity, y _true.i Is the ith actual capacity of the cycle and n is the number of cycles of the full life cycle.

Global System

In the formula (I), the compound is shown in the specification,

is the ith cycle predicted capacity, y _true.i Is the ith actual capacity of the cycle, and n is the number of cycles of the full life cycle.

Local part

In the formula (I), the compound is shown in the specification,

is the jth cycle predicted capacity, y _true.i Is the j-th cycle actual capacity, and m is the number of cycles at capacity rebound.

Local part

In the formula (I), the compound is shown in the specification,

is the j-th cycle prediction capacity, y _true.i Is the j-th cycle actual capacity, and m is the number of cycles at capacity rebound.

Step 205: completing the prediction of the residual service life;

in order to simulate the vibration stress and the charge-discharge stress of the electric automobile battery under the driving working condition, a full life cycle capacity testing system of the battery under the driving working condition is built, as shown in fig. 2. According to the urban driving condition, the power density spectrum of the vibration table is set as shown in figure 3. The most intuitive embodiment of the driving condition to the electric automobile is the driving speed of the automobile, which is specifically represented by uniform speed, acceleration, idling, deceleration and the like. The battery charge and discharge test curve under the set driving condition is shown in fig. 4.

And (3) carrying out RUL prediction research by using the full life cycle experimental sample of the electric vehicle battery under the running working condition as a data set. Taking 40% of the data set as a training set and the rest as a test set, and obtaining the hyperparameter alpha =0.809609 and the coordination parameter c =1.268496 through an adaptive robust loss function. According to equation (9), the Loss function corresponding to different values of α and its first derivative can be obtained, in addition to the conventional Loss function, by adding α =0.5 and α =1.5, which are intermediate values of the resulting α and L1 Loss and L2Loss losses, respectively, as shown in fig. 6 and 7.

As can be seen from fig. 6, the loss function is a smooth curve (differentiable) suitable for gradient-based algorithm optimization. In the interval of residual | x | > 0, the loss function is monotonically increasing. The hyper-parameter alpha is also monotonically increased, and the characteristic determines that the loss function has better robustness. The reason is that during the optimization the loss function can be gradually reduced starting from a larger value of alpha, so that the robust prediction avoids local minima. As can be seen from fig. 7, the adaptive robust loss function can perform a smooth transition of different loss functions.

As can be seen from fig. 7, as the value of α decreases, the amplitude saturation value of the first derivative of the loss function decreases, and its magnitude does not exceed 1/c. That is, as the residual x (> | c |) increases, its effect on the gradient decreases, and thus, the impact of outliers on the RUL prediction decreases as the gradient decreases.

As can be seen from fig. 8, the starting value of the Loss value of the adaptive Loss function is small and reaches a minimum value after 27 iterations, while the cases of the remaining Loss functions are: 44 iterations (Cauchy Loss), 32 iterations (a _0.5 Loss), 40 iterations (L1 Loss), 38 iterations (a _1.5 Loss), 49 iterations (L2 Loss), 40 (G-MC Loss), 36 iterations (Welsch Loss), and 77 iterations (LightGBM). Therefore, the adaptive loss function proves to have better robustness.

In order to verify the effectiveness of the improved RUL prediction method, the provided RUL prediction method is applied to the battery full-life-cycle capacity data under the driving condition, wherein the hyper-parameters of the LightGBM prediction model are set according to table 1, and the RUL prediction result is shown in fig. 9.

As can be seen from fig. 9, the RUL prediction method of the improved LightGBM can achieve a RUL prediction root mean square error value of 1.02% (the RUL prediction root mean square error value of the original LightGBM is 1.03%). Meanwhile, the prediction effect of the provided RUL prediction method is superior to that of the original RUL prediction method in local places.

As can be seen from fig. 10, the prediction effect of the RUL based on the AR Loss is better in both prediction accuracy and robustness, while the prediction effect of the RUL based on the Cauchy Loss is worse than others. The comparison result shows that compared with the traditional loss function, the ARLF has better self-adaptability, can generate different expression forms according to needs, and better meets the actual requirements, thereby being beneficial to improving the generalization capability of the model. Further illustrating the applicability of self ARLF to LightGBM in RUL prediction.

In conclusion, the method for predicting the remaining service life of the power battery of the electric automobile has the advantages that:

In the embodiment of the present invention, except for the specific description of the model of each device, the model of other devices is not limited, as long as the device can perform the above functions.

Those skilled in the art will appreciate that the drawings are only schematic illustrations of preferred embodiments, and the above-mentioned serial numbers of the embodiments of the present invention are only for description and do not represent the merits of the embodiments.

The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents, improvements and the like that fall within the spirit and principle of the present invention are intended to be included therein.

Claims

1. A method for predicting the residual service life of a power battery of an electric automobile is characterized by comprising the following steps:

(1) Obtaining relevant parameters of the adaptive loss function, specifically: after acquiring the capacity data of the battery in the full life cycle under the driving condition, acquiring a super parameter alpha and a coordination parameter c through a self-adaptive robust loss function, and configuring a loss function of a LightGBM;

(2) Processing data based on the multi-threaded parallel histogram;

(3) Training the model in a Leaf-wise growth (Leaf-wise) mode based on bandwidth limitation;

(4) Automatically optimizing and obtaining optimal parameters of the model based on meta-learning hyper-parameters;

(5) And finishing the prediction of the residual service life.

2. The method for predicting the remaining service life of the power battery of the electric vehicle according to claim 1, wherein the data processed based on the multi-thread parallel histogram specifically comprises:

for each dimension feature, dividing continuous floating point type data into discrete k ranges to obtain k buckets (bins), and constructing a histogram with the width of k.

3. The method for predicting the remaining service life of the power battery of the electric vehicle as recited in claim 1, wherein the Leaf-wise growth (Leaf-wise) training model based on the bandwidth limitation is specifically:

the Leaf-wise growth mode is that when each node is split, the Leaf node with the largest gain is selected to grow, the complexity of the model is limited, the training speed is increased, and overfitting is avoided.

4. The method for predicting the remaining service life of the power battery of the electric vehicle as recited in claim 1, wherein the obtaining of the optimal parameters of the model based on meta-learning hyper-parameter automatic tuning specifically comprises:

according to a Hyperopt hyper-parameter optimization framework, constructing a model parameter space, a LightGBM model factory and a score acquirer; the model factory is used for producing the required model, and the score acquirer is used for decoupling; and obtaining the optimal parameters of the model through multiple iterations.