A method for predicting the remaining useful life of a lithium-ion battery based on the FIG-ABC-SVR algorithm

Abstract

The present invention discloses a method for predicting the remaining useful life of a lithium-ion battery based on FIG-ABC-SVR. Predicting the remaining useful life of a lithium-ion battery is the core of battery health management. The specific steps of this method are as follows: First, use the fuzzy information granulation method to divide the time series of the lithium battery capacity degradation data into several subsequences as operation windows, and fuzzify each generated window to generate capacity fuzzy particles; then, use the ABC-SVR model to predict two parameters of the fuzzy particles. Finally, integrate the predicted fuzzy particles, use the interpolation method to complete the missing capacity data, obtain the complete capacity time series values, and further obtain the remaining useful life of the lithium battery by comparing with the capacity threshold. The method of the present invention solves the problem of poor accuracy of the long-term prediction effect of SVR. At the same time, fuzzy information granulation reduces the number of training samples of SVR and improves the calculation efficiency. The proposed method has a more accurate prediction effect and can be widely applied to the prediction of the remaining useful life of lithium batteries.

Description

Technical Field

[0001] The present invention relates to the technical field of testing the electrical condition of batteries, and particularly to a method for predicting the remaining useful life of a lithium battery. Technical Background

[0002] Lithium-ion batteries have the advantages of high working voltage, low self-discharge rate, high specific energy, no memory effect, environmental friendliness, long cycle life, flexible volume, etc., so they are widely used in the fields of electric vehicles, aerospace, transportation, and mobile electronic devices. However, as the battery is continuously used and the number of charge and discharge cycles increases, irreversible chemical reactions will occur inside the battery, specifically manifested as an increase in internal resistance, a decrease in the maximum available capacity, and a decrease in peak power. The degradation of lithium batteries will greatly reduce the performance of the equipment, reduce the reliability of the equipment, and even pose a safety hazard. Therefore, accurately predicting the remaining useful life (RUL) of lithium batteries is of great significance. The RUL of a lithium battery refers to the number of cycles that the lithium battery has to experience before it degrades to a certain specified failure threshold. The degradation state of a lithium battery can be evaluated using the capacity and internal resistance of the battery. Generally, the capacity is used to characterize the degradation state. It is usually considered that when the battery capacity decays to 70% to 80% of the initial capacity, the battery reaches its failure threshold.

[0003] In order to accurately predict the remaining useful life of lithium batteries, researchers have invested a lot of research work in this field. At present, the methods for predicting the remaining useful life of lithium batteries can be roughly divided into three categories: model-based methods, data-driven methods, and hybrid methods. The model-based method attempts to establish an accurate mathematical model through the internal reaction mechanism of the battery to describe the degradation phenomenon of the battery. Therefore, although the accuracy of the model-based method is relatively high, the model is too complex and the actual application efficiency is not high. The data-driven method extracts the hidden information between the capacity degradation through historical data, such as voltage, current, temperature, etc. This method does not need to understand the degradation mechanism of lithium batteries, avoiding complex mechanism modeling. In addition, this method has the advantages of high prediction accuracy, fast model training, and good universality, so it has been widely used. The fusion method uses a combination of model-based methods and data-driven methods to predict the RUL of lithium batteries, and has also received extensive attention in recent years.

[0004] In the data-driven method for predicting the remaining life of lithium batteries, support vector regression (SVR) is a typical algorithm. In order to obtain better prediction results, some scholars have proposed using artificial bee colony optimized support vector regression (ABC-SVR) to predict the RUL of lithium batteries. Although using the improved support vector regression to predict the RUL of lithium batteries will achieve better results, it still cannot change the problem of poor long-term prediction accuracy of support vector regression. Summary of the Invention

[0005] Aiming at the problem that the long-term prediction results in the existing technology are poor, leading to low accuracy of the remaining useful life prediction results, the present invention provides a method for predicting the remaining useful life of lithium batteries based on FIG-ABC-SVR, realizing a more accurate prediction of the RUL of lithium batteries.

[0006] The technical solution adopted by the present invention is as follows: including the following steps:

[0007] Step 1, collect the capacity data of the battery. Select the prediction starting point ST, divide the capacity data into a training set and a test set, and set the failure threshold Cap of the capacity EOL .

[0008] Step 2, use the fuzzy information granulation algorithm to divide the capacity data of the training set into several information granules, that is, fuzzy windows. Extract the maximum and minimum values of each window and construct the corresponding time series and where win is the number of windows.

[0009] Step 3, use the time series of the window maximum and minimum values obtained in the previous step to construct a new training set and Use the two training sets to train the SVR model respectively, and use the ABC algorithm to search for two parameters, the penalty coefficient c and the kernel function radius g, of the SVR model: the specific steps are as follows:

[0010] Step 3.1, initialize two parameters of the ABC algorithm: regard the penalty coefficient c and the kernel function radius g as the parameters x to be optimized in the ABC algorithm i (nectar source), initialize the bee colony size NP, the number of nectar sources FoodNumber, the maximum number of iterations M, the threshold Limit for the nectar source not being updated continuously, and the upper and lower bounds ub and lb of the parameter values, and the initial solution of x i is generated by Equation 1;

[0011] x i = lb + (ub - lb)·rand(0,1) Equation 1

[0012] Step 3.2, calculate the fitness value fit of each initial solution x i by Equation 2 i :

[0013]

[0014] where f i is the objective function value. Here, the root mean square error is defined as the objective function value.

[0015] Step 3.3, set the iteration count iter = 1, and in the leading bee search phase, find the optimal solution under the current population conditions. The specific steps are as follows:

[0016] Step (1), search for the nectar source x according to Equation 3 i ;

[0017] x’ ij = x ij + rand(-1,1)(x ij - x kj ) Equation 3

[0018] Step (2), calculate the fitness value of the nectar source after search and update according to Equation 3. The f in Equation 3 i is the objective function value, and use the greedy algorithm to select a better food source;

[0019] Step 3.4, in the follower bee search phase, calculate the probability that the nectar source x i is selected by the follower bees;

[0020]

[0021] Step 3.5, the follower bees search for and select a nectar source according to Equation 3, and calculate the corresponding fitness value by Equation 2;

[0022] Step 3.6, the follower bees use the greedy algorithm to select a better nectar source;

[0023] Step 3.7, determine whether there is a nectar source that has not been updated after Limit cycles. If so, the corresponding leading bee becomes a scout bee. And randomly generate a new food source according to Equation 5 to replace it;

[0024]

[0025] Step 3.8, determine whether iter reaches the maximum number of iterations. If it does, output the optimal solutions of parameters c and g; otherwise, iter = iter + 1, and return to Step 3.3 to repeat the execution;

[0026] Step 4, from Step 3, the predicted time series Up i and Low i can be obtained. Use the interpolation method to complete the missing sequence, and the complete predicted capacity value can be obtained. Determine whether the predicted capacity value reaches the capacity failure threshold Cap EOL . If it reaches the threshold, calculate the corresponding RUL result and calculate the corresponding evaluation index.

[0027] As an optimized design: The method of performing fuzzy information granulation in step 2 is to granulate the original capacity data using triangular fuzzy particles. The specific granulation method can be described as follows:

[0028]

[0029] Where x is the capacity sequence value included in each window, a is the minimum value of the capacity sequence in each window, b is the maximum value of the capacity sequence in each window, and m is the median of a and b.

[0030] As an optimized design: The interpolation method for filling in step 4 is to use a linear function to solve for the missing values in the middle, where the linear function can be expressed as:

[0031] y = βx Equation 7

[0032] Where y represents the information Up i or Low i of the window, x is the number of cycles, and β is the corresponding slope value when fitting with a linear function for each window.

[0033] The beneficial effects of the present invention are as follows:

[0034] 1. The present invention uses the method of fuzzy information granulation to perform fuzzy processing on the original capacity data, reducing the scale of the original capacity data set. By predicting the maximum and minimum value sequences of a small number of window information, the prediction accuracy of support vector regression is improved, effectively alleviating the problem of poor long-term prediction effect of support vector regression.

[0035] 2. The present invention uses the artificial bee colony algorithm to predict the maximum and minimum value sequences of window information respectively, constructing two independent regression prediction models. Compared with the traditional method of using a single model for regression prediction, the present invention has higher prediction accuracy. Brief Description of the Drawings

[0036] Figure 1 is the overall flowchart of the present invention.

[0037] Figure 2 is the capacity degradation diagram of the lithium-ion battery.

[0038] Figure 3 is the capacity prediction result diagram of Battery B5 using different algorithms.

[0039] Figure 4 is the capacity prediction result diagram of Battery B6 using different algorithms.

[0040] Figure 5 is the capacity prediction result diagram of Battery B7 using different algorithms. Detailed Embodiment

[0041] The specific embodiments of the present invention will be described in detail below. Using the lithium battery capacity degradation dataset provided by the NASA Excellence in Failure Prediction Research Center, three batteries, namely B5, B6, and B7, are selected as the implementation dataset in the present invention.

[0042] A method for predicting the remaining useful life of lithium-ion batteries using improved support vector regression with fuzzy information granulation (FIG-ABC-SVR) is as Figure 1 shown. The specific steps of this method are as follows:

[0043] Step 1: Extract the capacity degradation dataset from the original lithium battery dataset and perform normalization processing. Set the starting point ST for capacity prediction. Use the capacity data before the starting point ST as training data and the capacity data after ST as test data, and set the failure threshold Cap of the capacity EOL .

[0044] Step 2: Perform fuzzy granulation processing on the divided training dataset. Divide the training dataset into several windows, and extract the maximum value sequence {Up i} and the minimum value sequence {Low i} of the capacity data for each window. The specific steps are as follows:

[0045] Step 2.1: Initialize the size n of the window. According to the fuzzy information granulation criterion, the number of windows num for the fuzzy information granulation of the training set is num = N / n, where N is the number of samples contained in the training set.

[0046] Step 2.2: Perform triangular fuzzy granulation processing on the training data according to the set window information. The expression of triangular fuzzy granulation is as follows:

[0047]

[0048] where x is the capacity sequence value contained in each window of the training set, a is the minimum value of the capacity sequence in each window, b is the maximum value of the capacity sequence in each window, and m is the median of a and b. Extract the parameters a and b for each window, and respectively form the maximum value sequence {Up i} and the minimum value sequence {Low i} of the window information.

[0049] There are two major advantages in using triangular fuzzy granulation to process the original capacity degradation data. First, the original dataset contains many capacity regeneration samples. The capacity regeneration phenomenon will affect the model training effect, making it difficult for the regression algorithm to find the non-linear mapping relationship between capacity and cycle times. After triangular fuzzy granulation processing, the scale of the capacity dataset decreases significantly. Compared with the original capacity dataset, the number of regeneration samples contained in the data after triangular fuzzy granulation processing will be significantly reduced, and the model is more stable and efficient during training. Second, the long-term prediction effect of traditional regression algorithms is often poor. In battery life prediction, it is manifested as a large deviation in the predicted remaining useful life (RUL). After using triangular fuzzy information granulation, the data scale is reduced, and the medium- and long-term capacity prediction problem in long-term prediction is simplified to the problem of predicting the capacity in the early or middle stage, which will improve the capacity prediction accuracy, thereby improving the prediction accuracy of the remaining useful life (RUL).

[0050] Step 3, based on the window information sequences {Up i} and {Low i}, train the SVR model respectively, and use the artificial bee colony optimization algorithm (ABC) to select two parameters, the penalty coefficient c and the kernel function radius g, of the SVR model. Use the ABC-SVR model to predict the window information sequence values respectively. The specific steps are as follows:

[0051] Step 3.1, initialize the parameters: regard the two parameters c and g to be optimized as the nectar sources x i of the artificial bee colony algorithm, that is, x i = (c i , g i ). Initialize the swarm size NP, the number of nectar sources FoodNumber, the maximum number of iterations M, the threshold Limit for the nectar source not being updated continuously, and the upper and lower bounds ub and lb of the parameter values. And generate the initial solution of x i by Equation 2;

[0052] x i = lb + (ub - lb)·rand(0, 1) Equation 2

[0053] Step 3.2, calculate the fitness value fit i of each initial solution x i by Equation 3:

[0054]

[0055] where f i is the objective function value. Here, set the root mean square error mse as the objective function value.

[0056] Step 3.3, set the iteration count iter = 1, and in the leading bee search phase, find the optimal solution under the current population conditions. The specific steps are as follows:

[0057] Step (1), search for the nectar source x according to Equation 4 i ;

[0058] x’ ij = x ij + rand(-1, 1)(x ij - x kj ) Equation 4

[0059] Step (2), calculate the fitness value of the nectar source after search and update according to Equation 4. The f in Equation 4 i is the objective function value, and the greedy algorithm is used to select a better food source;

[0060] Step 3.4, in the follower bee search phase, calculate the probability that the nectar source x i is selected by the follower bees according to Equation 5;

[0061]

[0062] Step 3.5, the follower bees search for and select the nectar source according to Equation 4, and calculate the corresponding fitness value according to Equation 3;

[0063] Step 3.6, the follower bees use the greedy algorithm to select a better nectar source;

[0064] Step 3.7, determine whether a certain nectar source has not been updated after Limit times of iteration. If so, the corresponding leading bee becomes a scout bee. And a new food source is randomly generated according to Equation 6 to replace it;

[0065]

[0066] Step 3.8, determine whether iter reaches the maximum number of iterations. If so, output the optimal solutions of parameters c and g; otherwise, iter = iter + 1, and return to Step 3.3 to repeat; finally, output the optimal parameters c and g, and provide them to the SVR model respectively.

[0067] When using the traditional grid search method to search for the parameters c and g of the support vector regression model, the parameters are first roughly estimated, and a large range is defined for the parameters. Then, through the error grid diagram of the parameters, the parameters are finely selected. The parameters of the grid search method need to be finely selected manually. When the error coefficients of different parameters are the same, it is difficult to select parameters, and it may fall into a local optimal solution. Moreover, these two steps will reduce the model training efficiency, which is not conducive to the online application scenario of lithium batteries. When using the artificial bee colony optimization algorithm to search for the parameters c and g, different from the grid search method that generates several parameters at set intervals, the parameters c and g are randomly generated within the set range. In addition, the artificial bee colony optimization algorithm will record and compare the optimal parameters in each iteration process, and the finally selected output result is the global optimal solution. In the online prediction application of the remaining service life of lithium batteries, existing methods often cannot solve the balance problem between model efficiency and model complexity. Too low model efficiency is not conducive to rapid prediction, and too high model complexity often leads to the model falling into a local optimal solution. The artificial bee colony optimization algorithm ensures the model efficiency and effectively alleviates the occurrence of falling into a local optimal solution through mechanisms such as random generation of nectar sources and evolution of bee colonies, improving the efficiency and prediction accuracy of the online prediction of the remaining life of lithium batteries.

[0068] Step 4, based on the window information {Up i} and {Low i} of the training set for training, and more specifically, based on the set prediction starting point ST, predict the subsequent capacity window information. And use the linear interpolation method to complete the missing capacity data of the window. The equation of the linear interpolation method is y = βx. Where y is the dependent variable, x is the independent variable, and β is the slope of the linear equation. Then a complete capacity prediction sequence can be obtained. Determine whether the predicted capacity value reaches the set capacity threshold. When it reaches the threshold, calculate the RUL value. And calculate the error value, output the corresponding error, and evaluate the prediction performance of the model.

[0069] After using fuzzy information granulation to process the original data, the data scale will be reduced, and the capacity regeneration samples in the model will be reduced. However, there are still a small number of capacity regeneration samples in the processed samples. When training the window information after fuzzy information granulation, introducing the artificial bee colony optimization algorithm to optimize the parameters c and g of the support vector regression algorithm can effectively prevent the model from falling into a local optimal solution. Even if there are a small number of capacity regeneration samples in the training set, the model can still well capture the non-linear decreasing relationship between capacity and the number of cycles during training. In addition, fuzzy information granulation divides the original data into two groups of independent window information data {Up i} and {Low i}, Using two independent support vector regression models for training can reduce the training pressure of the model. Using the artificial bee colony optimization algorithm to optimize the c and g of the regression algorithm can improve the efficiency of a single model. Generally, it can improve the stability and computational efficiency of the model, making the prediction results more accurate and better meeting the requirements of online prediction of the remaining life of lithium batteries.

[0070] To accurately analyze the accuracy of the prediction results, the mean absolute error (MAE), root mean square error (RMSE), and remaining life absolute error (RE) are used to evaluate the model performance.

[0071]

[0072]

[0073] RE = |RP - RT|

[0074] In the above formula, y i is the true value of the capacity, is the predicted value of the capacity, and n is the number of samples in the test set. RP is the predicted remaining life, and RT is the true remaining life.

[0075] Next, taking batteries B5, B6, and B7 of NASA as examples, the prediction performance of the proposed model is verified.

[0076] The capacity degradation of the three batteries is as Figure 2 shown. The failure thresholds of batteries B5 and B6 are set to 70% of the initial capacity, that is, 1.4 Ah. Thus, the true RUL values are 124 (for battery B5) and 108 (for battery B6). Since the capacity of battery B7 did not degrade to 1.4 Ah at the last cycle, the failure threshold is set to 75% of the initial capacity, that is, 1.5 Ah, and the corresponding RUL value is 125 (for battery B7). The starting point of battery prediction is set near the 80th cycle.

[0077] The present invention compares the proposed fuzzy information granulation plus optimized algorithm support vector regression (FIG-ABC-SVR) with SVR and ABC-SVR algorithms. Figures 3 to 5 The different comparison results are given respectively, and Tables 1 to 3 give the error analysis. It can be found that the results of the method proposed by the present invention are more accurate. The mean absolute error and root mean square error of the capacity are the smallest, and the predicted RUL is also closer to the true value.

[0078] The first column in Tables 1 to 3 lists three algorithms. Among them, SVR is the support vector regression algorithm used alone, ABC-SVR is the support vector regression algorithm optimized by artificial bee colony, and FIG-ABC-SVR is the algorithm combining fuzzy information granulation and artificial bee colony optimization of support vector regression proposed in the present invention. The MAE, RMSE, RT, RP, and RE in the second to fifth columns respectively represent the mean absolute error, root mean square error, actual remaining useful life, predicted remaining useful life, and absolute error of the remaining useful life.

[0079] Table 1 Prediction Results of B5

[0080]

[0081] Table 2 Prediction Results of B6

[0082]

[0083] Table 3 Prediction Results of B7

[0084]

[0085] In view of the problem of predicting the remaining useful life of lithium-ion batteries, the present invention proposes a method combining fuzzy information granulation (FIG) and artificial bee colony optimization of support vector regression (ABC-SVR). The present invention uses the FIG method to divide the capacity degradation data into several windows, and then uses the ABC-SVR algorithm to perform regression predictions on the maximum and minimum values of each window respectively to obtain the information of the prediction window. Finally, the linear interpolation method is used to complete the missing values in the prediction window to obtain the complete capacity prediction value, and the remaining life of the battery is calculated according to the failure threshold. The present invention can provide relatively accurate RUL prediction results and provide reliable remaining useful life prediction information for lithium battery equipment.

Claims

1. A method for predicting the remaining useful life of a lithium-ion battery based on FIG-ABC-SVR, characterized in that The steps to implement this method are as follows: Step 1, collect the capacity data of the battery, select the prediction starting point ST, divide the capacity data into a training set and a test set, and set the failure threshold Cap of the capacity EOL ; Step 2: Use the fuzzy information granulation algorithm to divide the capacity data of the training set into several information granules, namely fuzzy windows, extract the maximum and minimum values of each window, and construct the corresponding time series and where win is the number of windows; Step 3: Construct a new training set using the window maximum and minimum time series obtained in the previous step and Train the SVR model using the two training sets respectively, and use the ABC algorithm to search for the two parameters of the penalty coefficient c and the kernel function radius g of the SVR model: Step 3.1, initialize two parameters of the ABC algorithm: Take the penalty coefficient c and the kernel function radius g as the parameters to be optimized in the ABC algorithm, i.e., the nectar source x i , initialize the swarm size NP, the number of nectar sources FoodNumber, the maximum number of iterations M, the threshold Limit for the nectar source not being updated continuously, and the upper and lower bounds ub and lb of the parameter values. The initial solution of x i is generated by Equation 1: x i = lb + (ub - lb)·rand(0,1) Equation 1 Step 3.2, calculate the fitness value fit of each initial solution x i according to Equation 2 i : where f i is the objective function value. Here, the root mean square error is defined as the objective function value; Step 3.3: Set the iteration count iter = 1, and in the leading bee search phase, find the optimal solution under the current population conditions. The specific steps are as follows: Step (1), search for the nectar source x according to Equation 3 i ; x i ' j = x ij + rand(-1, 1)(x ij - x kj ) Equation 3 Step (2), calculate the fitness value of the nectar source after search update according to Equation 3, where f i is the objective function value, and use the greedy algorithm to select a better food source; Step 3.4, in the follower bee search phase, calculate the nectar source \(x\) according to Equation 4 i The probability selected by the follower bees; Step 3.5: The follower bees search and select nectar sources according to Equation 3, and calculate the corresponding fitness values according to Equation 2; Step 3.6: The follower bees use the greedy algorithm to select better nectar sources; Step 3.7: Determine whether a certain nectar source has not been updated after Limit cycles. If so, the corresponding leading bee becomes a scout bee; and a new food source is randomly generated according to Equation 5 to replace it; Step 3.8: Determine whether iter reaches the maximum number of iterations. If it does, output the optimal solutions of parameters c and g; otherwise, iter = iter + 1, and return to Step 3.3 to repeat; Step 4: The predicted time series Up i and Low i can be obtained from Step 3. The missing sequences are complemented using the interpolation method to obtain the complete predicted capacity values. Determine whether the predicted capacity values reach the capacity failure threshold Cap EOL . If the threshold is reached, calculate the corresponding RUL results and calculate the corresponding evaluation metrics.

2. The method for predicting the remaining service life of a lithium-ion battery according to claim 1, wherein: The method of performing fuzzy information granulation in Step 2 is to granulate the original capacity data using triangular fuzzy particles. The granulation method is as follows: Among them, x is the capacity sequence value included in each window, a is the minimum value of the capacity sequence in each window, b is the maximum value of the capacity sequence in each window, and m is the median of a and b.

3. The method for predicting the remaining service life of a lithium-ion battery according to claim 1, wherein: The interpolation method for filling in Step 4 is to use a linear function to solve the missing values in the middle, where the linear function is expressed as: y = βx Equation 7 Among them, y represents the information Up of the window i or Low i , x is the number of cycles, and β is the slope value corresponding to each window when fitting with a linear function.

Images