CN109815995B - Method for predicting remaining life of lithium battery under condition of missing observed value - Google Patents
Method for predicting remaining life of lithium battery under condition of missing observed value Download PDFInfo
- Publication number
- CN109815995B CN109815995B CN201910011104.5A CN201910011104A CN109815995B CN 109815995 B CN109815995 B CN 109815995B CN 201910011104 A CN201910011104 A CN 201910011104A CN 109815995 B CN109815995 B CN 109815995B
- Authority
- CN
- China
- Prior art keywords
- model
- complete
- elm
- observation
- training
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Abstract
The invention relates to a method for predicting the residual life of a lithium battery under the condition of missing observed values, which comprises the following steps: 1) designing a total framework of a lithium battery residual life prediction method under the condition of missing observation values; 2) designing a preprocessing module algorithm, including introducing an ELM algorithm of an extreme learning machine, an ELMSI (single-point interpolation algorithm) design and an ELMMI (multiple interpolation algorithm); 3) the prediction module algorithm design comprises a nuclear limit learning machine design and a multi-step advanced prediction algorithm design; 4) training an ELMMI multiple interpolation algorithm model; 5) and (3) carrying out model training on a multi-step advanced prediction method. The invention has the beneficial effects that: the invention provides a method for combining a multi-interpolation algorithm ELMMI with different multi-step advance prediction methods to construct a multi-step advance predictor based on the multi-interpolation algorithm, which can realize an accurate and stable estimation process.
Description
Technical Field
The invention relates to a method for predicting the residual service life of a lithium iron phosphate battery under the condition of a missing observation value, in particular to a method for filling the missing observation value based on a multiple interpolation technology and a method for predicting the residual service life of a single lithium battery based on multi-step advanced prediction.
Background
Lithium batteries are favored by the industry due to the advantages of high energy density, long service life, low self-discharge rate and the like, and currently occupy the main market of power batteries. During the use process, the lithium battery has the problems of rapid aging and energy exhaustion, which leads to the reduction of the overall performance of the equipment and unpredictable damage. Therefore, the life Prognosis and Health Management (PHM) of lithium batteries are increasingly gaining importance from the battery industry. The life prognosis of a lithium battery is defined as predicting the service life of a battery system according to the operating conditions and operating conditions of the system. Lithium battery remaining service life (RUL) prediction methods are mainly divided into two major categories, model-based prediction and data-driven prediction. The model-based prediction method is the most widely used RUL estimation method, but since all parameters of the prediction model need initialization and pre-adjustment, it is difficult to obtain accurate and perfect model parameters in practical applications. Meanwhile, the model parameters are difficult to update in real time under new working conditions. The above problem can be solved based on a data-driven prediction method that predicts future values using present and past observations in the battery degradation curve. The information characteristics reflected by the degradation curve, such as voltage values, capacity values, current values, and impedance values, can be extracted and used for accurate prediction of RUL. And the data driving algorithm calculates the degradation trend by learning the correlation among the information characteristics and accurately predicts the residual service life of the lithium battery.
A one-step look-ahead (OSP) method for estimating the short-term RUL of a battery uses a non-linear A-regression (NAR) structure to build a pre-decision model as shown in equation (1):
in the formula, C is the battery capacity in the discharge cycle, i is the cycle number, e is the estimation error of each cycle, n is the total number of cycles, and f is an approximate function generated by the training stage predictor and is a parameter matrix of the estimation function.
Disclosure of Invention
The invention aims to overcome the defects in the prior art and provide a method for predicting the residual life of a lithium battery under the condition of missing observed values.
The method for predicting the residual life of the lithium battery under the condition of the missing observation value comprises the following steps:
step 1: designing a total framework of a lithium battery residual life prediction method under the condition of missing observation values; calculating and predicting based on an ELM (extreme learning machine); the preprocessing module designs single-point interpolation and multiple interpolation based on an ELM algorithm, and the prediction module designs a multi-step advanced prediction method based on the traditional ELM, the nucleation ELM and the online sequential ELM;
step 2: designing an algorithm of a preprocessing module; designing single-point interpolation and multiple interpolation algorithms based on the traditional ELM algorithm;
step 2-1: introducing an ELM algorithm of an extreme learning machine; giving a group of m observation values, and randomly distributing input weight and hidden layer deviation by using ELM; ELM analyzes and adjusts output weight, and the algorithm formula is as follows:
where α is the output weight, hiFor non-linear feature mapping, wiTo connect the ith hidden node to the input node, biA threshold value of the ith hidden node; minimizing prediction error H alpha-F Y in ELM model training period2And the output weight rated value alpha is calculated according to the following formula:
wherein the content of the first and second substances,representing an output node xjThe error vector of (a) is calculated,for regularization parameters, F is a function FELMThe solution of (1); to solve the optimization problem, the following least-norm least-squares solution of the linear system is found:
Hα=F (4)
the least squares solution α ═ H is obtained from equation (4)TF;HTObtaining prior information by utilizing an orthogonal projection technology;
step 2-2: designing a single-point interpolation algorithm ELMSI: the observation data set is first divided into two distinct subsets, complete subset XcompleteAnd incomplete subset XuncomThen to the subset XuncomAny observation vector x in (2)nBy the use of XcompleteTraining an ELM model by all complete observations in (1); finally, a target feature vector X is definednAnd estimating X by using the trained ELM modelnThe missing observation feature of (a);
step 2-3: design of an ELMMI (multiple interpolation algorithm): in the initial state, ELMMI divides the data set into two subsets, complete subset XcompleteAnd incomplete subset Xuncom(ii) a By XuncomEach observation vector x ofnThe loop creates p estimation sets and satisfies p e [1,5 ]]Generating a complete estimation data set with a matrix size of m multiplied by n for each estimation set; by using XcompleteCircularly train the ELM model with xnAll incomplete observation features in (1) are target vectors, the rest observation features are input vectors, and an incomplete subset X is estimateduncomThe missing observation vector features in (1); computing recently input observations x using a similarity function λnAnd complete subset XcompleteA similarity between each of the observations; interpolation algorithm searching complete subset XcompleteSelecting the l nearest neighborsBuilding a data subset Xl(ii) a With XlUsing x as a training targetnTaking all corresponding incomplete observation features as input vectors, and training an ELM model; training the ELM model twice to obtain the final ELMMI model, and estimating XnAnd the complete observation is obtainedAdding complete subset XcompletePerforming the following steps;
and step 3: designing a prediction module algorithm; introducing one-step advanced prediction, and designing a multi-step advanced prediction algorithm as a component module of the lithium battery residual life prediction method; designing a nuclear limit learning machine based on the ELM;
step 3-1: designing a coring extreme learning machine; designing a nucleation extreme learning machine KELM based on the ELM, utilizing a kernel matrix of the ELM model, and meeting Mercer conditions, such as formula (5):
ΨKELM=HHT:fKELM(xi,xj)=h(xi)·h(xj) (5)
the output function of the KELM model is reconstructed as:
in the formula, Y and fKELM(. the) respectively represents the identification matrix and the kernel function, m observed values are selected, and the formula is as follows:
training a KELM model by utilizing various kernel functions, wherein the functions meet the Mercer condition; training a KELM model by utilizing a wavelet function WAV and a radial basis function RBF, and establishing formulas of the wavelet function and the RBF respectively as follows:
wherein tau, upsilon, zeta and xi are model parameters in the algorithm training process respectively;
step 3-2: designing a multi-step advanced prediction algorithm; multi-step advanced prediction MSP is classified into three major categories: iterative, DirRec and direct; designing a new model of multi-step advanced prediction based on a multiple interpolation technology and a DirRec method, wherein the model generates a new prediction model after each calculation step;
calculating a new predicted value in the training subset, and discarding the last observation value, namely keeping the same number of observation values of the training subset in each iterative calculation process; the new predicted value formula for estimating the L lithium battery capacities is as follows:
in the formula (f)lA prediction model for predicting step length of L order;
and 4, step 4: training an ELMMI multiple interpolation algorithm model;
step 4-1: introducing a set of missing observations in each cycle of a battery aging data setCreating incomplete dataset X'uncom(ii) a The missing observation values are randomly imported into different data set periods, and the import probability of the last period of the training sequence is highest;
step 4-2: selecting the voltage/current and SOC (state of charge) data of 40-50% of the front of the single battery to carry out ELMMI model training; vector X in complete subsetcompleteIs rewritten asThe model training formula is as follows:
in the formula, H is a hidden node matrix,is an output weight matrix; using a model function fELMMIPredicted X'uncomIncomplete vector ofxn∈XcompleteAnd introducing random missing observation values; selecting the first 40-50% sampling value to train the model, wherein the model training formula is as follows:
step 4-3: the complete observation vector is represented asThe training model selects a nonlinear model and a linear model, and compares the nonlinear model and the linear model with the real value of the target parameter;
and 5: training a multi-step advanced prediction method model; and for an incomplete data set containing less than 20% of missing observed values, performing data filling on the input subset by using a multiple interpolation algorithm.
Preferably, the method comprises the following steps: in step 2-3, the ELM-based multiple interpolation algorithm elmi further includes the following steps:
inputting: xm×nIs an incomplete data set, is a parameter set of an ELM model;
defining: c. CmisIs a missing feature of an observed value, cobsObservable features being an observation, dmisFor incomplete observation, dobsFor complete observations, X is the total observation dataset, XcompleteIs a complete subset of X, XuncomIs a non-complete subset of X, XestmationFor the estimated set of full observations,for some two observed values xnAnd xmThe similarity between them;
wherein GRC is a gray correlation coefficient and is expressed as GRC (x)nj,xmj)=0.5/(|xnj-xmj|+0.5);
1) Decomposition of X into XcompleteAnd Xuncom;
2) Mixing XuncomBy cmisThe data sets are represented and arranged from small to large;
3) normalization of XcompleteAnd Xuncom;
Within the p data sets, satisfying p ═ 1:1:5, the following steps are performed:
4) using the parameter XcompleteRewritten as xnTraining an ELM model;
in the formula, H is a hidden node matrix,to output the weight matrix, the weight matrix is output,is an observable in a complete subset;
5) using a model function fELMPrediction of XuncomIs incompletely observed value xn,xn∈Xuncom
6) calculating a weight coefficient gamma:
γ=(2β-1)/10
in the formula, beta is a parameter coefficient;
7) calculating a similarity equation:
the value M is taken out and circulated;
8) according to the parameter XcompleteWriting an equation lambda in descending order;
10) The ELM model was trained using the following formula:
n value cycle is finished;
13) set of return estimates Xestimation(a)=Xcomplete;
And finishing the p value cycle.
The invention has the beneficial effects that: the invention provides a method for combining a multi-interpolation algorithm ELMMI with different multi-step advance prediction methods to construct a multi-step advance predictor based on the multi-interpolation algorithm, which can realize an accurate and stable estimation process.
Drawings
FIG. 1 is an overall frame diagram of the present invention;
FIG. 2 is a graph of the results of a comparison of the linear regression model, the non-linear prediction model, and the true values of the target parameters, trained using the first 40% observations;
FIG. 3 is a graph of the results of RMSE values calculated by 5 OS predictors in conjunction with different step lead predictors.
Detailed Description
The present invention will be further described with reference to the following examples. The following examples are set forth merely to aid in the understanding of the invention. It should be noted that, for those skilled in the art, it is possible to make various improvements and modifications to the present invention without departing from the principle of the present invention, and those improvements and modifications also fall within the scope of the claims of the present invention.
The invention carries out pretreatment, estimation and prediction on the residual life of the lithium iron phosphate battery under the condition of missing observed values, innovatively provides a prediction method based on one-step advanced prediction and multi-step advanced prediction algorithms, and the overall framework comprises two modules: a preprocessing module and a prediction module. The preprocessing module includes two new techniques, namely an extreme learning machine-based single point interpolation technique (elsim) and a multiple interpolation technique (EMMI).
Step 1: and (3) designing the overall framework of the lithium battery residual life prediction method under the condition of missing observation values. As shown in fig. 1, the new method is based on dynamic Extreme Learning Machine (ELM) to make predictions and estimations. Since the hidden nodes of the ELM are randomly generated and the lost observation values can be classified, the algorithm can accurately predict the short-term and long-term life states of the lithium battery. The preprocessing module designs single-point interpolation and multiple interpolation based on an ELM algorithm, and the prediction module designs a multi-step advanced prediction method based on traditional ELM, nucleation ELM (KELM) and online sequential ELM (OS-ELM).
Step 2: and designing an algorithm of a preprocessing module. A preprocessing module in the lithium battery residual life prediction method mainly completes the task of filling missing observed values.
Step 2-1: an extreme learning machine ELM algorithm is introduced. Given a set of m observations, the ELM randomly assigns input weights and hidden layer biases, reducing computation time while maintaining overall prediction accuracy. ELM analyzes and adjusts output weight, and the algorithm formula is as follows:
where α is the output weight, hiFor non-linear feature mapping, wiTo connect the ith hidden node to the input node, biIs the threshold value of the ith hidden node. Minimizing prediction error H alpha-F Y in ELM model training period2And the output weight rated value alpha is calculated according to the following formula:
wherein the content of the first and second substances,representing an output node xjThe error vector of (a) is calculated,for the regularization parameters to balance the above optimization formula, F is a function FELMThe solution of (1). To solve the optimization problem, the minimum norm least squares solution of the following linear system can be found:
Hα=F (4)
the least squares solution α ═ H is obtained from equation (4)TF, the solution has good generalization ability, minimal error, and fast convergence. HTThe prior information is obtained using an orthogonal projection technique.
Step 2-2: and (4) designing a single-point interpolation algorithm. The invention designs a new single-point interpolation algorithm (ELMSI), firstly, an observation data set is divided into two different subsets, namely a complete subset XcompleteAnd incomplete subset XuncomThen to the subset XuncomAny observation vector x in (2)nBy the use of XcompleteAll perfect observations in (a) train the ELM model. Finally, a target feature vector X is definednAnd estimating X by using the trained ELM modelnThe missing observation feature of (1).
Step 2-3: and (4) designing a multiple interpolation algorithm. The invention designs a multi-interpolation algorithm (ELMMI) based on single-point interpolation, and the algorithm is a novel multi-filling technology. The relevant formula and calculation flow of the new algorithm are given below: (1) in the initial state, ELMMI divides the data set into two subsets, complete subset XcompleteAnd incomplete subset Xuncom(ii) a (2) By XuncomEach observation vector x ofnThe loop creates p estimation sets and satisfies p e [1,5 ]]Generating a complete estimation data set with a matrix size of m multiplied by n for each estimation set; (3) by using XcompleteCircularly train the ELM model with xnAll incomplete observation features in (1) are target vectors, the rest observation features are input vectors, and an incomplete subset X is estimateduncomThe missing observation vector features in (1); (4) computing recently input observations x using a similarity function λnAnd complete subset XcompleteThe similarity between each observation value in (the weighting coefficient of the similarity function is gamma, and gamma is 0.1-0.9); (5) interpolation algorithm searching complete subset XcompleteSelecting the l nearest neighborsBuilding a data subset Xl(the total number of nearest neighbors is not a fixed value, but a fraction (set to 10%) of the total number of observations); (6) with XlUsing x as a training targetnTaking all corresponding incomplete observation features as input vectors, and training an ELM model; (7) training the ELM model twice to obtain the final ELMMI model, and estimating XnAnd the complete observation is obtainedAdding complete subset XcompleteIn (1).
ELM-based multiple interpolation algorithm (elmi):
inputting:
Xm×nas an incomplete data set
Parameter set for ELM model
Defining:
cmismissing features being an observation
cobsObservable features being an observed value
dmisFor incomplete observation
dobsAs a complete observation
X is the total observation dataset
XcompleteComplete subset of X
XuncomIs a non-complete subset of X
XestmationSet of estimates as complete observations
Wherein GRC is a gray correlation coefficient and is expressed as GRC (x)nj,xmj)=0.5/(|xnj-xmj|+0.5)
1. Decomposition of X into XcompleteAnd Xuncom
2. Mixing XuncomBy cmisThe data sets are represented and arranged from small to large
3. Normalization of XcompleteAnd Xuncom
Within the p data sets, satisfying p ═ 1:1:5, the following steps are performed:
4. Using the parameter XcompleteRewritten as xnTraining the ELM model
In the formula, H is a hidden node matrix,to output the weight matrix, the weight matrix is output,is an observable in a complete subset
5. Using a model function fELMPrediction of XuncomIs incompletely observed value xn,xn∈Xuncom
6. Calculating weight coefficient gamma
γ=(2β-1)/10
In the formula, β is a parameter coefficient.
7. Equation of computational similarity
M value cycle ends
8. According to the parameter XcompleteWriting an equation lambda in descending order
10. ELM model training using the following equation
n value cycle ends
13. Set of return estimates Xestimation(a)=Xcomplete
And finishing the p value cycle.
And step 3: and (4) designing a prediction module algorithm. The invention introduces one-step advanced prediction and designs a multi-step advanced prediction algorithm as a component module of the lithium battery residual life prediction method. An extreme learning machine is designed based on ELM, and aims to complete the short-term and long-term residual life prediction tasks.
Step 3-1: and (5) verifying the extreme learning machine design. The invention designs a nucleation extreme learning machine KELM based on ELM, and the algorithm utilizes a kernel matrix of an ELM model and meets Mercer conditions, such as formula (5):
ΨKELM=HHT:fKELM(xi,xj)=h(xi)·h(xj) (5)
the output function of the KELM model is reconstructed as:
in the formula, Y and fKELM(. the) respectively represents the identification matrix and the kernel function, m observed values are selected, and the formula is as follows:
the KELM model is trained using a variety of kernel functions that must satisfy the Mercer condition. The present invention trains a KELM model using wavelet functions (WAV) and Radial Basis Functions (RBF). The formulas for establishing the wavelet function and the RBF function are respectively as follows:
wherein tau, upsilon, zeta and xi are model parameters in the algorithm training process respectively.
Step 3-2: and designing a multi-step advanced prediction algorithm. Multi-step advanced prediction (MSP) is suitable for estimating the long-term remaining life of a lithium battery, and the method is divided into three categories: iterative, DirRec and direct. The invention designs a new model of multi-step advanced prediction based on a multiple interpolation technology and a DirRec method, and the model can generate a new prediction model after each calculation step.
The main characteristics are as follows: and calculating a new predicted value in the training subset, and discarding the last observation value, namely keeping the same number of the observation values of the training subset in each iterative calculation process. The new predicted value formula for estimating the L lithium battery capacities is as follows:
in the formula (f)lFor the prediction model of L-order prediction step length, the estimation precision can be improved by carrying out cyclic training on the model by using the new prediction value.
And 4, step 4: and training an ELMMI multiple interpolation algorithm model.
Step 4-1: the invention introduces a group of missing observation values in each period of the battery aging data setArtificially creating incomplete dataset X'uncom. And the missing observation values are randomly imported into different data set periods, and the import probability of the last period of the training sequence is highest.
Step 4-2: and selecting the data of the voltage/current and SOC of the front 40 percent of the single battery to carry out ELMMI model training. Vector X in complete subsetcompleteIs rewritten asThe model training formula is as follows:
in the formula, H is a hidden node matrix,is the output weight matrix. Using a model function fELMMIPredicted X'uncomIncomplete vector ofxn∈XcompleteAnd a random missing observation is imported. Selecting the first 40% sampling value to train the model, wherein the model training formula is as follows:
step 4-3: the complete observation vector is represented asThe model training result is shown in fig. 2, and the training model selects a nonlinear model and a linear model and compares the nonlinear model and the linear model with the true value of the target parameter. FIG. 2 shows that even if the first 40% of observed values are used for training, the calculation result of the ELMMI nonlinear model provided by the invention is closest to the true value, and meanwhile, the nonlinear interpolation algorithm model used in the residual life prediction process of the lithium battery has more advantagesThe model can generate an effective predictive model with a limited number of observations.
And 5: and (3) carrying out model training on a multi-step advanced prediction method. For an incomplete data set containing 15% missing observations, a multi-interpolation algorithm is applied to fill data in an input subset, and a multi-step lead predictor is applied to calculate an RMSE value by respectively combining 5 different interpolation algorithms, as shown in Table 1. The interpolation algorithms include ELMSI, ELMMI, kNNI, LWLA, and MCMC, and the multi-step look-ahead predictor includes regression methods (RF), on-line sequential learning machine ELM (OS-ELM), radial basis function-nucleated ELM (KELM-RBF), wavelet function-nucleated ELM (KELM-WAV), and traditional ELM. The incomplete data set selected for this experiment contains three lag inputs, and model training is performed using the first 60% observations. Table 1 illustrates: (1)5 interpolation algorithms are combined with different predictors, and the predicted RMSE average values are equivalent; (2) RMSE of the KELM-RBF predictor is larger than that of other advanced predictors and reaches 18 percent, and the average value of the OS-ELM predictor is minimum and does not exceed 8.9 percent; (3) the ELMSI is combined with the OS-ELM, the RMSE average value is minimum, and the prediction effect is best.
TABLE 1 RMSE mean value (x 10) of multi-step look-ahead method combined with 5 interpolation algorithms3)
Interpolation algorithm | RF | OS-ELM | KELM-RBF | KELM-WAV | ELM |
ELMSI | 6.514 | 1.965 | 15.624 | 5.789 | 2.030 |
kNNI | 6.216 | 2.385 | 15.046 | 5.649 | 2.423 |
LWLA | 7.558 | 3.465 | 18.912 | 6.889 | 3.504 |
ELMMI | 8.726 | 7.587 | 12.815 | 7.982 | 7.630 |
MCMC | 9.795 | 8.883 | 12.705 | 9.158 | 8.628 |
Experimental result 1:
the results were calculated according to table 1, yielding fig. 3. FIG. 3 shows that ELMSI is combined with different multi-step advanced prediction methods, the prediction result is the same as the result of a single-point interpolation algorithm, and the KNNI interpolation algorithm is combined with an RF predictor and a KELM-RBF predictor to obtain the optimal prediction result. In summary, the invention proposes to combine the multi-interpolation algorithm elmi with different multi-step look ahead methods, construct the multi-step look ahead predictor based on the multi-interpolation algorithm, and can realize an accurate and stable estimation process.
TABLE 2 different MS predictors combine 5 interpolation algorithms to calculate the error value E of the end lifeRUL(%)
Interpolation algorithm | RF | OS-ELM | KELM-RBF | KELM-WAV | ELM |
ELMSI | -23 | -10.5 | -12.5 | -12.5 | -11 |
kNNI | -23 | -11 | -12.5 | -12.5 | -12.5 |
LWLA | -22 | -11 | -12.5 | -12.5 | -12.5 |
ELMMI | -24.4 | -20.4 | -20.4 | -20.5 | -21.0 |
MCMC | -46.2 | -50.4 | -61.2 | -46.4 | -44.2 |
Experimental results 2:
the invention designs a multi-step advanced prediction method applied to a prediction module in an integral frame, and an error value (E) of the end life is calculated by using a DirRec method in MSP (mixed Signal processor) in an experimentRUL) The results are shown in Table 2. When the multi-step look-ahead method is combined with the elmi interpolation algorithm, the calculation error of all multi-step look-ahead (MS) predictors is low. Table 2 illustrates: (1) the MCMC has the maximum overall error value which is more than 45 percent, and the error values of the ELMSI, the kNNI and the LWLA are equivalent and are not more than 23 percent; (2) the average error value of the predictor combining ELM and different interpolation technologies is minimum, the short-term and long-term remaining service life of the lithium battery can be well estimated, the establishment time of the predictor is short, and the estimation precision is high; (3) for the one-step look-ahead process, OS-ELM and ELMSI (ELMMI) are combined and comparedThe interpolation algorithm has better prediction performance; (4) obtaining the same error value of two coring functions of multi-step advanced prediction; (5) the performance of a prediction model established by the RF predictor is better, but the performance of the predictor is obviously reduced in one-step advanced prediction; (6) both the ELM predictor and the RF predictor require more calculation time and are not suitable for the online real-time working state of the lithium battery.
Effect of the Algorithm
The invention provides a method for predicting the residual life of a lithium iron phosphate battery under the condition of lacking an observed value, designs a single-point and multiple interpolation technology, introduces an advanced prediction algorithm, and provides a complete observation data set and a prediction module.
(1) The invention develops a predictor for battery current estimation based on an electro-dynamic learning machine (ELM) design, and performs simulation comparison with predictors such as a fuzzy neural Network (NFS), a data processing Grouping Method (GMDH) and a Random Forest (RF).
(2) The invention provides an ELM-based missing observation interpolation algorithm, namely single-point interpolation (ELMSI) and multiple interpolation (ELMMI).
(3) The invention constructs a multi-form prediction network based on ELM, and realizes short-term and long-term prediction calculation at the same time.
(4) The invention integrates the ELMSI and ELMMI interpolation algorithm and is used for the residual life prediction process of the lithium battery with a missing observation value.
(5) In the invention, random missing observation values are artificially introduced into a preprocessing module, and are classified for multiple times by using ELMMI, and simultaneously, the uncertainty and the confidence coefficient of a new algorithm are calculated.
FIG. 1 illustrates:
(1) the overall frame is divided into two modules: a left preprocessing module and a right prediction module;
(2) the total input of the prediction scheme is a lithium battery observation data set, and the output is a one-step advanced prediction/multi-step advanced prediction end life estimated value and RMSE;
(3) the preprocessing module comprises a single-point interpolation algorithm (ELMSI, kNNI and LWLA) and a multiple-interpolation algorithm (MCMC and ELMMI), and the observation data set is subjected to three processes of classification, normalization and interpolation to obtain a complete data set which is input into the prediction module;
(4) the prediction module comprises a one-step advanced prediction (NAR) and a multi-step advanced prediction (an iteration method, a DirDec method and a direct method), a complete data set obtained through an interpolation algorithm is used as an input value of the predictor, and the prediction method is used for estimating the residual life (RUL) and the RMSE average value of the lithium battery and comparing the values.
FIG. 2 illustrates:
(1) FIG. 2 is a graph showing the comparison result between the non-linear model, the linear model and the real value of the target parameter.
(2) FIG. 2 shows that: even if a small number of observations are used for training, the results obtained by the nonlinear prediction model are closer to the true values than the linear regression model.
(3) The results show that: the non-linear interpolation algorithm model is used in the lithium battery residual life prediction process to be more advantageous, and the model can generate an effective prediction model by using a limited number of observed values.
FIG. 3 illustrates:
(1) FIG. 3 shows the calculation of RMSE values using a multi-step look-ahead predictor in combination with 5 different interpolation algorithms.
(2) The incomplete data set selected in the experiment contains three lag inputs, the prediction model is trained by utilizing the first 60% of observed values, and the result shows that: the predicted result from the single point interpolation algorithm, when combined by OS-ELM and ELMSI, yields the lowest RMSE value.
(3) The different curves in fig. 3 reflect the RMSE average obtained by different interpolation algorithms.
(4) As can be seen from fig. 3, the elmi is combined with various predictors, the prediction result is the same as that of the single-point interpolation algorithm, and the KNNI interpolation algorithm is combined with the RF predictor and the KELM-RBF predictor to obtain the optimal prediction result.
(5) The invention provides a method for combining a multi-interpolation algorithm ELMMI with different multi-step advance prediction methods to construct a multi-step advance predictor based on the multi-interpolation algorithm, which can realize an accurate and stable estimation process.
Claims (2)
1. A method for predicting the residual life of a lithium battery under the condition of missing observed values is characterized by comprising the following steps:
step 1: designing a total framework of a lithium battery residual life prediction method under the condition of missing observation values; calculating and predicting based on an ELM (extreme learning machine); the preprocessing module designs single-point interpolation and multiple interpolation based on an ELM algorithm, and the prediction module designs a multi-step advanced prediction method based on the traditional ELM, the nucleation ELM and the online sequential ELM;
step 2: designing an algorithm of a preprocessing module; designing single-point interpolation and multiple interpolation algorithms based on the traditional ELM algorithm;
step 2-1: introducing an ELM algorithm of an extreme learning machine; giving a group of m observation values, and randomly distributing input weight and hidden layer deviation by using ELM; ELM analyzes and adjusts output weight, and the algorithm formula is as follows:
where α is the output weight, hiFor non-linear feature mapping, wiTo connect the ith hidden node to the input node, biA threshold value of the ith hidden node; minimizing prediction error H alpha-F Y in ELM model training period2And the output weight rated value alpha is calculated according to the following formula:
wherein the content of the first and second substances,representing an output node xjThe error vector of (a) is calculated,for regularization parameters, F is a function FELMThe solution of (1); to solve the optimization problem, the following least-norm least-squares solution of the linear system is found:
Hα=F (4)
the least squares solution is obtained from equation (4)α=HTF;HTObtaining prior information by utilizing an orthogonal projection technology;
step 2-2: designing a single-point interpolation algorithm ELMSI: the observation data set is first divided into two distinct subsets, complete subset XcompleteAnd incomplete subset XuncomThen to the subset XuncomAny observation vector x in (2)nBy the use of XcompleteTraining an ELM model by all complete observations in (1); finally, a target feature vector X is definednAnd estimating X by using the trained ELM modelnThe missing observation feature of (a);
step 2-3: design of an ELMMI (multiple interpolation algorithm): in the initial state, ELMMI divides the data set into two subsets, complete subset XcompleteAnd incomplete subset Xuncom(ii) a By XuncomEach observation vector x ofnThe loop creates p estimation sets and satisfies p e [1,5 ]]Generating a complete estimation data set with a matrix size of m multiplied by n for each estimation set; by using XcompleteCircularly train the ELM model with xnAll incomplete observation features in (1) are target vectors, the rest observation features are input vectors, and an incomplete subset X is estimateduncomThe missing observation vector features in (1); computing recently input observations x using a similarity function λnAnd complete subset XcompleteA similarity between each of the observations; interpolation algorithm searching complete subset XcompleteSelecting the l nearest neighborsBuilding a data subset Xl(ii) a With XlUsing x as a training targetnTaking all corresponding incomplete observation features as input vectors, and training an ELM model; training the ELM model twice to obtain the final ELMMI model, and estimating XnAnd the complete observation is obtainedAdding complete subset XcompletePerforming the following steps;
and step 3: designing a prediction module algorithm; introducing one-step advanced prediction, and designing a multi-step advanced prediction algorithm as a component module of the lithium battery residual life prediction method; designing a nuclear limit learning machine based on the ELM;
step 3-1: designing a coring extreme learning machine; designing a nucleation extreme learning machine KELM based on the ELM, utilizing a kernel matrix of the ELM model, and meeting Mercer conditions, such as formula (5):
ΨKELM=HHT:fKELM(xi,xj)=h(xi)·h(xj) (5)
the output function of the KELM model is reconstructed as:
in the formula, Y and fKELM(. the) respectively represents the identification matrix and the kernel function, m observed values are selected, and the formula is as follows:
training a KELM model by utilizing various kernel functions, wherein the functions meet the Mercer condition; training a KELM model by utilizing a wavelet function WAV and a radial basis function RBF, and establishing formulas of the wavelet function and the RBF respectively as follows:
wherein tau, upsilon, zeta and xi are model parameters in the algorithm training process respectively;
step 3-2: designing a multi-step advanced prediction algorithm; multi-step advanced prediction MSP is classified into three major categories: iterative, DirRec and direct; designing a new model of multi-step advanced prediction based on a multiple interpolation technology and a DirRec method, wherein the model generates a new prediction model after each calculation step;
calculating a new predicted value in the training subset, and discarding the last observation value, namely keeping the same number of observation values of the training subset in each iterative calculation process; the new predicted value formula for estimating the L lithium battery capacities is as follows:
in the formula (f)lA prediction model for predicting step length of L order;
and 4, step 4: training an ELMMI multiple interpolation algorithm model;
step 4-1: introducing a set of missing observations in each cycle of a battery aging data setCreating incomplete data set Xu'ncom(ii) a The missing observation values are randomly imported into different data set periods, and the import probability of the last period of the training sequence is highest;
step 4-2: selecting the voltage/current and SOC (state of charge) data of 40-50% of the front of the single battery to carry out ELMMI model training; vector X in complete subsetcompleteIs rewritten asThe model training formula is as follows:
in the formula, H is a hidden node matrix,is an output weight matrix; using a model function fELMMIPredicted X'uncomIncomplete vector ofxn∈XcompleteAnd introducing random missing observation values; selecting the first 40-50% sampling value to train the model, wherein the model training formula is as follows:
step 4-3: the complete observation vector is represented asThe training model selects a nonlinear model and a linear model, and compares the nonlinear model and the linear model with the real value of the target parameter;
and 5: training a multi-step advanced prediction method model; and for an incomplete data set containing less than 20% of missing observed values, performing data filling on the input subset by using a multiple interpolation algorithm.
2. The method for predicting the remaining life of the lithium battery under the condition of missing observed values as claimed in claim 1, wherein in the step 2-3, the ELM-based multiple interpolation algorithm ELMMI further comprises the following steps:
inputting: xm×nIs an incomplete data set, is a parameter set of an ELM model;
defining: c. CmisIs a missing feature of an observed value, cobsObservable features being an observation, dmisFor incomplete observation, dobsFor complete observations, X is the total observation dataset, XcompleteIs a complete subset of X, XuncomIs a non-complete subset of X, XestmationFor the estimated set of full observations,for some two observed values xnAnd xmThe similarity between them;
wherein GRC is a gray correlation coefficient and is expressed as GRC (x)nj,xmj)=0.5/(|xnj-xmj||+0.5);
1) Decomposition of X into XcompleteAnd Xuncom;
2) Mixing XuncomBy cmisThe data sets are represented and arranged from small to large;
3) normalization of XcompleteAnd Xuncom;
Within the p data sets, satisfying p ═ 1:1:5, the following steps are performed:
4) using the parameter XcompleteRewritten as xnTraining an ELM model;
in the formula, H is a hidden node matrix,to output the weight matrix, the weight matrix is output,is an observable in a complete subset;
5) using a model function fELMPrediction of XuncomIs incompletely observed value xn,xn∈Xuncom
6) calculating a weight coefficient gamma:
γ=(2β-1)/10
in the formula, beta is a parameter coefficient;
7) calculating a similarity equation:
the value M is taken out and circulated;
8) according to the parameter XcompleteWriting an equation lambda in descending order;
10) The ELM model was trained using the following formula:
n value cycle is finished;
13) set of return estimates Xestimation(a)=Xcomplete;
And finishing the p value cycle.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910011104.5A CN109815995B (en) | 2019-01-07 | 2019-01-07 | Method for predicting remaining life of lithium battery under condition of missing observed value |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910011104.5A CN109815995B (en) | 2019-01-07 | 2019-01-07 | Method for predicting remaining life of lithium battery under condition of missing observed value |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109815995A CN109815995A (en) | 2019-05-28 |
CN109815995B true CN109815995B (en) | 2020-10-27 |
Family
ID=66603946
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910011104.5A Active CN109815995B (en) | 2019-01-07 | 2019-01-07 | Method for predicting remaining life of lithium battery under condition of missing observed value |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109815995B (en) |
Families Citing this family (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111366848A (en) * | 2019-12-31 | 2020-07-03 | 安徽师范大学 | Battery health state prediction method based on PSO-ELM algorithm |
CN111833990A (en) * | 2020-07-17 | 2020-10-27 | 电子科技大学 | Method for filling missing items of psychological assessment scale |
CN112731183B (en) * | 2020-12-21 | 2023-04-21 | 首都师范大学 | Improved ELM-based lithium ion battery life prediction method |
CN113203953B (en) * | 2021-04-02 | 2022-03-25 | 中国人民解放军92578部队 | Lithium battery residual service life prediction method based on improved extreme learning machine |
CN113567863B (en) * | 2021-06-11 | 2022-04-01 | 北京航空航天大学 | Abnormal degraded lithium battery capacity prediction method based on quantum assimilation and data filling |
CN113361692B (en) * | 2021-06-28 | 2023-05-23 | 福建师范大学 | Lithium battery remaining life combined prediction method |
CN113850016B (en) * | 2021-08-16 | 2024-04-05 | 国网江苏省电力有限公司技能培训中心 | Method for predicting residual life of storage battery of simulated transformer substation in intermittent working mode |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2014125340A1 (en) * | 2013-02-15 | 2014-08-21 | Pharmaday S.R.L. | Smoke liquid for atomizers and/or vaporizers |
CN105183994A (en) * | 2015-09-10 | 2015-12-23 | 广西大学 | Method and device for predicting powder battery SOC on basis of improved I-ELM |
CN106897794A (en) * | 2017-01-12 | 2017-06-27 | 长沙理工大学 | A kind of wind speed forecasting method based on complete overall experience mode decomposition and extreme learning machine |
CN107590537A (en) * | 2016-07-08 | 2018-01-16 | 香港理工大学 | For constructing the granulation Forecasting Methodology in probabilistic forecasting section |
-
2019
- 2019-01-07 CN CN201910011104.5A patent/CN109815995B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2014125340A1 (en) * | 2013-02-15 | 2014-08-21 | Pharmaday S.R.L. | Smoke liquid for atomizers and/or vaporizers |
CN105183994A (en) * | 2015-09-10 | 2015-12-23 | 广西大学 | Method and device for predicting powder battery SOC on basis of improved I-ELM |
CN107590537A (en) * | 2016-07-08 | 2018-01-16 | 香港理工大学 | For constructing the granulation Forecasting Methodology in probabilistic forecasting section |
CN106897794A (en) * | 2017-01-12 | 2017-06-27 | 长沙理工大学 | A kind of wind speed forecasting method based on complete overall experience mode decomposition and extreme learning machine |
Non-Patent Citations (2)
Title |
---|
Extreme learning machine for missing data using multiple imputations;DušanSovilj 等;《Neurocomputing》;20161122;第174卷;第220-231页 * |
锂电池剩余寿命的ELM间接预测方法;姜媛媛 等;《电子测量与仪器学报》;20160229;第30卷(第2期);第179-185页 * |
Also Published As
Publication number | Publication date |
---|---|
CN109815995A (en) | 2019-05-28 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109815995B (en) | Method for predicting remaining life of lithium battery under condition of missing observed value | |
Ren et al. | A data-driven auto-CNN-LSTM prediction model for lithium-ion battery remaining useful life | |
Yang et al. | Cars: Continuous evolution for efficient neural architecture search | |
CN106600059B (en) | Intelligent power grid short-term load prediction method based on improved RBF neural network | |
CN112036084B (en) | Similar product life migration screening method and system | |
US20190034784A1 (en) | Fixed-point training method for deep neural networks based on dynamic fixed-point conversion scheme | |
CN108181591B (en) | Battery SOC value prediction method based on improved BP neural network | |
Li et al. | A comparative study of battery state-of-health estimation based on empirical mode decomposition and neural network | |
CN111832825B (en) | Wind power prediction method and system integrating long-term memory network and extreme learning machine | |
CN110007235A (en) | A kind of accumulator of electric car SOC on-line prediction method | |
Tang et al. | Skfac: Training neural networks with faster kronecker-factored approximate curvature | |
CN111063398A (en) | Molecular discovery method based on graph Bayesian optimization | |
CN115374853A (en) | Asynchronous federal learning method and system based on T-Step polymerization algorithm | |
CN113935489A (en) | Variational quantum model TFQ-VQA based on quantum neural network and two-stage optimization method thereof | |
Nazari et al. | Multi-level binarized lstm in eeg classification for wearable devices | |
CN112307667A (en) | Method and device for estimating state of charge of storage battery, electronic equipment and storage medium | |
CN114580747A (en) | Abnormal data prediction method and system based on data correlation and fuzzy system | |
CN111832817A (en) | Small world echo state network time sequence prediction method based on MCP penalty function | |
CN116579371A (en) | Double-layer optimization heterogeneous proxy model assisted multi-objective evolutionary optimization computing method | |
CN110471768A (en) | A kind of load predicting method based on fastPCA-ARIMA | |
Li et al. | Resource usage prediction based on BiLSTM-GRU combination model | |
CN109558898B (en) | Multi-choice learning method with high confidence based on deep neural network | |
CN115528750B (en) | Power grid safety and stability oriented data model hybrid drive unit combination method | |
CN117406100A (en) | Lithium ion battery remaining life prediction method and system | |
CN112731183A (en) | Lithium ion battery life prediction method based on improved ELM |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |