+86 18988945661 contact@iflowpower.com +86 18988945661

Language

2022/04/08

Author ：Iflowpower – **Portable Power Station Supplier**

my country's energy storage network: Abstract: In order to solve the problem of difficulty in accuracy and efficiency in the efficiency of the available capacity of lithium-ion batteries, this paper proposes a feature processing and radial nerve network residual capacity estimation method. First, extract the feature amount associated with the remaining available capacity in the battery charging process data, and then use the local abnormal factor algorithm to accurately cleaning the abnormal point in the feature amount, improve the amount of effective information contained in the characteristics, and then use local linear embedding. The algorithm performs designing the resulting feature vector group, reducing data complexity, finally, introducing the estimation model of the remaining capacity in the radial nerve network.

The model is applied on different model batteries. The maximum average absolute error of the estimation result is 0.06.

The maximum root mean square error is 0.05, indicating that the model can effectively estimate the remaining available capacity of the lithium-ion battery and strong robust robust sex. Compared to the Elman neural network and the BP neural network algorithm, the method has faster estimation efficiency while ensuring high precision.

Key words: lithium-ion battery; feature treatment; radial nerve network; capacity estimated lithium-ion battery due to high energy density, long-circulating life, low cost, environmental pollution, etc. best choice. In practical applications, since the activity of lithium ions in the battery will gradually decline, the available capacity of the lithium ion battery will be lost during use, directly affecting its remaining service life (Remaininguseful, RUL).

Therefore, in order to efficiently utilize lithium ion batteries, it is effective to estimate the remaining available capacity of current lithium-ion batteries, which in turn diagnose its health status (StateOfhealth, SOH) is an important feature of BatteryManagementSystem, BMS. At present, the researchers have proposed a variety of estimation methods to obtain available capacity, and it is important to divide Bayesian method, related empirical methods and data drive algorithms, etc. Bayesian-based method includes Kalman filter and particle filtering and other related improvement algorithms.

The literature uses a dual-extension Kalman filtering algorithm to obtain model parameters and current battery status, but this method is poor in the late stage of battery cycle life. In response to this problem, the literature adopts the adaptive elsencing Kalman filtering algorithm. The multi-step prediction of SOH and RUL is realized by updating the noise covariance matrix, and the estimation accuracy in the battery life cycle is improved.

The literature uses no less particle filtering to estimate the RUL of the lithium-ion battery, and the actual estimation relative error is less than 5%. The above-based method is based on the closed loop filtering algorithm to estimate the key parameters of the battery health. It has strong robustness to modeling errors and measurement errors, but the modeling process is complex, and the actual estimation process is large.

In addition to using various filter algorithms, the method of the relevant empirical fit is also used in the state estimate of the lithium ion battery. The literature considers the use time, charge and discharge ratio and temperature of the lithium-ion battery, and establish a generalized battery life model to achieve effective estimates for the health status of the same type of battery. Document combined with lithium-ion battery recession, the growth and active material of the active substance is established by the loss of the active substance, and the minimum multiplier estimation model parameters are used, and the results show that the result of the method is less than 1%.

The method for fitting related experience is generally established by establishing aging models based on engineering related experience or electrochemical knowledge, high precision, but model is equipped with a large number of test data. Also, a particular model often only effectively estimates a particular type of battery or a particular aging battery condition, flexibility is poor. Today, with the rapid development of big data and artificial intelligence, data-driven methods are also widely used in the state estimate of the battery.

The literature proposes a SOH estimation method based on a priori neural network and a Markov chain. The results show that even under uncertain external conditions, the method can still effectively estimate the internal state of the battery. The literature uses the Gaussian ProcessRegression, GprocessRegression, GPR to obtain a mapping between battery capacity, temperature and tipping state (SOC), and realizes simultaneous estimation of multiple states of lithium-ion batteries.

The literature uses a long-term time memory neural network to further optimize the structure of the GPR model, resulting in an estimation model with high estimation accuracy. The above-based data-driven approach is a black box model, which has the advantages of flexible intelligence, but the black box model is more sensitive to the training data set, less data makes the model less effective training, more data sets It is possible to cause complicated redundancy of models, and may cause excessive extensions of the model. In order to solve the problem of difficult to obtain an efficient characteristic set, this paper performs the estimation of the remaining capacity of lithium-ion batteries based on the characteristic processing and radial basisFunction, RBF) neural network, and improves estimation efficiency while ensuring accuracy.

First, the amount of capacity characteristics extracted during the charging process is optimized, and the data sets of the data set are more concentrated, and the data set is selected, and the complexity of the data is further selected to improve the complexity of the data to improve the estimation. Efficiency, then use a concise and efficient radial nerve network to establish the remaining available capacity estimation model, and ultimately realize the rapid and accurate estimation of the remaining available capacity of lithium ion batteries. 1 Lithium Ion Battery Cycle Acting Experimenting After the recession of the dynamic lithium-ion battery, the available capacity with the change of the battery using the cycle of the battery.

This paper provides continuous constant current charge and discharge under the conditions of the two specifications under room temperature. Circular aging experiment, operational flow is shown in Figure 1, the battery-related information is shown in Table 1. Experiments were carried out using the BT-5HC-5V-100A battery test system developed by ARBIN.

Figure 1 Battery cycle aging Experimental Process Table 1 Test Battery Basic Parameter Capacity Decline Experiment After completion of the four sets of battery cell full life cycle test data, respectively, battery A-1, battery B-2, battery B-3, battery B-4, wherein the battery A-1 is a type A-model lithium-ion battery, and the other 3 groups are B model lithium ion batteries. Taking battery A-1 as an example, a cycle of capacity decay experimental voltage and current change As shown in Figure 2, each test cycle is composed of constant current charging, constant voltage charging, shelving, constant current discharge, shelving, etc. Among them, the A-model constant current charging process current is 2A, the B model is 1.

275A, the A model constant voltage charging process cutoff current is 80mA, the B model is 51mA, the A model constant current discharge process current is 8A, B model 5.1a. Figure 2 Cyclic testing process voltage and current change 2 discharge capacity estimation implementation method 2.

1 Capacity Decayment Characteristics 2.1.1 Features During the actual application process, considering that the charging process of the lithium ion battery is constant current constant pressure charging mode, The battery of different models has consistency, and the discharge process has different differences due to actual use, and it is difficult to extract stable feature.

Therefore, this paper extracts 7 sets of 7 groups of seven groups related to the remaining available capacity of lithium ion batteries in experimental data, and divided into two categories according to data sources. The first category is characterized by directly collecting data according to the sensor. The charging time CC-T, charging capacity CC-C, and constant voltage charging phase, charging time CV-T, charging capacity CV-C as characteristics.

Take battery A-1 as an example, the first type of feature extracted is shown in Figure 3. Figure 3 Features of the battery A-1 The second type of feature amount is the characteristics of further excavation of the sensor directly collects data. Among them, the peak value of the incrementalcapacity (IC) curve is effective to characterize the physical and chemical characteristics of the lithium ion battery, and the IC curve means differentiation of capacity and voltage, and the VT representative Time battery voltage.

In the constant current charging mode, the current is a constant value, and when taking a very small time interval picture, the formula (1) is established, thereby quickly plotting the IC curve. The drawing IC curve is shown in Figure 4, after filtering, visible curves present a significant peak characteristic. The maximum value of each loop IC curve is used as the second type of feature, called: MAX-IC.

Further, the ratio of charging time and corresponding cycle constant pressure phase is performed separately: CC-T / CV-T and constant current stage charging capacity ratio ratio: CC-C / CV- C as a second type of related characteristics. The second type of battery A-1 is shown in Figure 5. Figure 4IC Curve Filter Before After the Characteristics of Battery A-1 During the acquisition of experimental data, there is a case where the real battery characteristic data is doped with the error interference data, and the two categories shown in FIGS.

3 and 5. In the characteristics, there is a significant data ly, and the test error and the battery's own decay characteristics can cause the appearance of out of group. Therefore, this paper uses a local abnormal factor algorithm to eliminate the impact of test error on the rules of characteristics, retaining the characteristics reflected in the true physical and chemical characteristics of the lithium-ion battery recession, effectively improve the amount of information contained in the characteristics, and improves model learning efficiency.

The local abnormal factor algorithm (LOF) is to determine the degree of extent of the dispersion by calculating the localized district of each sample, is a density-based extensive point detection method. It is possible to prevent the problem of incorrectly cleaning effective data in accordance with the data distribution discrimination of the display value according to the set threshold accurate positioning. The average number of local delta points of the sampling point P neighborhood point in the characteristic amount and the number of partial delimited density of point P are represented as part of the localized dispensing factor, and in formula (2), NK (P) Indicates that the k distance from point P, point O is a certain point in the neighborhood; LRDK (P), LrDk (O) represents the local accessibility density of point P and point O, respectively.

The closer the partial dislocation factor, the point P and its neighborhood points are different, the higher the probability of P and neighborhoods, but in turn, the P point may be an abnormal point. Figure 6 shows the abnormality of the abnormality of the application of localized firing factor CC-T, the original data is 1 ~ 5, which is obviously visible, where points 1 and point 2 are due to The abnormality value caused by the test environment, and the change in the number of points 3 to 5 is the reflection of the battery's available capacity recovery in the characteristic amount of CC-T, cannot be smoothly processed, in the characteristics selected It should be reserved during the process. Figure 6 Schedule for abnormal lying group 2.

1.2 Correlation inspection is to verify whether the characteristic amount is closely correlated between the amount of available capacity, and the appropriate method is used to correlate the two. Considering the relationship between the characteristics and the available capacity is highly non-linear, it is necessary to select an analysis method that does not require a feature data distribution.

Spearmanrand correlation coefficient is also known as "grade difference method", with a value between -1 to +1, the absolute value is close to 1 indicates the correlation. Compared to other intensification correlation coefficients, the Spirman level correlation coefficient is not sensitive to data errors and extreme values, and only two variables are required to be paired, or transformed from continuous variable observation data. Get rating.

Regardless of the overall distribution of the two variables, the size of the sample capacity is, the difference between the Spirman level can be effectively identified. The Spirman relationship is represented as in the formula (3), and Ri and Si are the value grade of the observation value i; R * and S * are the average levels of variables x and y, respectively; Di = Ri-Si represents two columns The number of differentiated modes of the variable; N is the total number of observations. The Sprim grade correlation coefficient between the circulating discharge capacity of the four sets of batteries obtained by the experiment is shown in Table 2, and the correlation coefficients before and after the abnormality of abnormal isolation point are listed.

The results show that the correlation between the extracted characteristics and the corresponding remaining available capacity are around 0.9, which exhibits strong correlation. And after the LOF algorithm is processed, the correlation coefficient has improved, significant enhancement correlation between the characteristic vector and the corresponding cycle capacity.

Table 2 Features and available capacity Spearman grade correlation coefficient 2.1.3 Features During the actual application, handling large amounts of data will occupy a large number of car calculation modules, which affects the overall performance of BMS.

This paper uses a local linear embedding algorithm (LECALLINEAREMBEDDING, LLE) to reduce the design of the data set, reduce the storage space required for the system, speed up the calculation speed. LLLLLE is a rolling study, maintaining the linear relationship between the original high-dimensional spatial sample neighborhood in the low-dimensional space, can better maintain the geometrical structure and properties of the data, suitable for designing to depletion processing for highly nonlinear data. LLE algorithm for high-dimensional data low-dimensional reflection is divided into the following steps.

Step 1: Select the local neighbor point set of the sample point to the given data set x = {x1, x2, ., xn} ℛℛD × n, xiℛℛd × 1, i = 1, 2, ., N, calculate the European miles from the sample point, select K (K

, K (K

Constrained condition picture indicates each of the lines in the weight matrix Ω addition to 1. Step 3: By obtaining the value matrix ω to obtain the low-dimensional embedding of the sample set X to map all sample points to the low-dimensional space ℛD, the vector X of the D-dimensional X is reduced to D-D-D-Dimension, each high dimensional vector xi corresponds A low-dimensional vector yi, y = {Y1, Y2, ., YN} ℛℛD × n mapping conditions are minimizing map loss functions, see formula (5) to solve the Solution of Y.

Solution under certain constraints The characteristic vector of the matrix, the mapping results on the low-dimensional space can be obtained by the Lagrangian multiplier method. 2.2 Based on the computation completion characteristics of the RBF neural network, this paper uses the RBF neural network to carry out the construction of a lithium-ion battery capacity recession model.

The RBF neural network is a function of a function approximation, data mining and other application scenarios. The internet. Its structure is simple, the principle training network of local response, local approximation, is the advantage of faster convergence speed compared to algorithms such as global response BP neural network, and the activation function of hidden neurons of the RBF neural network is a non-linear function.

Make the network approximate any nonlinear function. Its structure consists of a multi-input multi-output feedforward neural network, such as the input layer, implies layer, and output layer. Figure 7RBF Neural Network Structure Model Data is passed from the input layer node to the hidden layer, and the hidden layer node is generally constructed of the radial basis function, where the Gaussian radial basis function is better, and the output node is nonlinear to the hidden layer The end result of linear operations for linear operations.

In the hidden layer node, the Gaussian radial function function is shown in formula (6), X is the S-dimensional input vector; Ci is the center of the i-th radial function function; Δi is the scope of the i-th neural perception, which determines The central width of the base function; M is the number of hidden layer neurons; || X-CI || Represents the European Distance between X and Ci. Y is the output vector of the network. The input layer realizes a nonlinear mapping from X → FI (X), and the output layer implements linear mapping from FI (X) to Y, ie, Wi (7), Wi is the i-th Gabos radial basis function to output layers Pass the weight of the data; W0 is deviation from the output layer.

Using the error reverse propagation algorithm to determine the internal parameters of the network, as shown in FIG. 8, first initialize the weight matrix and set the target error function and error target of network training, then the training set input network cycle update weight matrix, until Decide the error target, complete network training, the weight matrix at this time, and the parameters of each neuron radial basis function are structural parameters of the final network. Figure 8RBF Neural Network Training Process Lithium Ion Battery Remainable Capacity Estimation Process As shown in Figure 9, it is important to establish and online capacity estimation for offline estimation models.

In the offline phase, preprocessing the measured lithium ion battery cycle charge and discharge experimental data, as a training data set of RBF neural network. A series of characteristic vectors are extracted from the time, voltage, current, capacity, and IC curve of the charging stage, and eliminate the disturbance of the characteristic noise in the characteristic noise, and use Spil to eliminate the interference of the characteristic amount of the characteristics of the character. Correlation between the characteristics and capacity of the characteristics and capacity of the manan relations.

Reduce the amount of multidimensional characteristics by local linear embedding method, reduce the complexity of data, and improve the training efficiency of the model. Finally, the internal parameters in the RBF network are trained, and the capacity estimation model is determined. The online portion uses the same manner as the offline stage as a model input, and the well-trained estimation model outputs efficient and accurate capacity estimates.

Figure 9 Lithium ion battery available capacity estimation flowchart 3 Estimation results and discussion 3.1 Model estimates 3.1.

1 Effects of different training volumes on the results of the results in improving the learning speed and prediction accuracy of the network, normalize the characteristics of the desired characteristics After processing, the model output is the estimated discharge capacity as the input value of the RBF neural network. In the implementation of the data driver method, 60% of data is used for training models, 40% for test model [7]. In order to assess the effectiveness and robustness of the established model, this paper divides 50%, 60%, 70% of each group of data as a training set, and the rest as a test set.

After entering the network model, the training test results of the model under different training are shown in Fig. 10, which uses the mean square error (MSE), RMSE Error (RMSE), average absolute error (MAE). The effect of the model prediction section is analyzed, as shown in Table 3.

Figure 10 Battery A-1, B-2, B-3 and B-4 Training Test Results Table 3 Different battery estimation errors When 60% of data is used to train models, the estimation model is estimated to output the remaining available capacity of four groups of experimental batteries. Stable and effective estimation results, the prediction results of the test range are very close to the actual sampling data. The error value in Table 3, the maximum MAE appears in the prediction portion of the battery B-4, only 0.

0494, MSE is 0.0034, RMSE is 0.0581, indicating that the established RBF neural network estimation model is a, b two type power lithium ion Both the capacity after the decline can make more precise estimation.

As can be seen from Figure 10, when the amount of training data is increased to 70%, the prediction accuracy of the model has improved, the maximum MAE is 0.0381, the maximum MSE is 0.0022, the maximum RMSE is 0.

0472. When the amount of training data is reduced to 50%, the estimation accuracy of the model is only slightly reduced, in which the prediction error of the battery B-2 is large, the maximum MAE is 0.1067, the maximum MSE is 0.

0137, the maximum RMSE is 0.1171, the maximum RMSE is 0.1171, the prediction accuracy of other batteries The forecast results under 60% training data are different.

It indicates that the estimated model has good generalization ability and strong robustness, which can effectively estimate the availability of different types of batteries in the case of fluctuations. 3.1.

2 Different Estimation Methods Compare the advantages of the ELMAN neural network estimation method, the ELMAN neural network estimation method, the estimation method proposed by the Elman neural network estimation method, BP neural network estimation method, and the estimated model proposed in this paper, based on characteristic processing based on feature processing Application to estimate the capacity recession of the full life cycle of lithium ion batteries, capacity estimates, estimated errors, and time consumption are shown in Figures 11, Table 4 and Table 5. Figure 11 Battery B-3 predicts the result (a) of battery B-2 (a) and battery B-4 predicting battery B-3 results (b) Table 4 Estimation error and consumption timetable 5 of battery B-3 predictive battery B-2 Battery B-4 Predicting Battery B-3 Estimation Error and Consumption Time Figure 11 (a) Full life capacity prediction results of battery B-2, all models are from battery B-3. Figure 11 (b) shows three different models of the full life capacity prediction result of the battery B-3, all models of training are from the battery B-4.

It can be seen from the forecasting results that the estimated model proposed in this paper relative to the capacity prediction result of the battery B-2 is except 6%, and the rest is less than 6%, and the extent to which the extent to which the predicted extent is only 0.057. However, the relative error of the battery B-3 capacity prediction result is within 6%, showing high estimation accuracy.

Compared to the ELMAN neural network with a circulating structure, the Elman neural network has a total capacity of the battery B-2 and the battery B-3, and the maximum MAE is 0.0847, the maximum RMSE is 0.0874 The relative error in the predictive result of the BP neural network model is more than 6%, the maximum mae is 0.

0723, the maximum RMSE is 0.0869, and the maximum MAE of the RBF neural network model to the two battery capacity prediction results is 0.0612, the maximum RMSE is 0.

0570, The prediction accuracy and model stability are reflected in good practical use advantages. Further, as can be seen from Table 4, under the same calculation processing conditions, when the capacity of the battery B-2 is predicted, the RBF neural network consumes only 0.99 s, and the ELMAN neural network consumption is 3.

5 s, BP neural network performance. The worst, consumption is 8.5s.

According to Table 5, when the capacity of the battery B-3 is predicted, the RBF neural network consumes 0.9 s, and the ELMAN neural network consumption is 3.5 s.

The BP neural network is as high as 9.4S. Comparing three predictive models, it can be seen that the time consumed by the remaining available capacity estimation method is approximately ELMAN neural network and the BP neural network, respectively, showing a large efficiency advantage.

3.2 The influence of outbound point processing on estimation results In the process of capacity estimation, this paper uses the LOF algorithm to accurately cleaning the characteristic vector in the characteristic vector. In order to reflect different abnormal value processing methods, the effect of the model estimation result is compared, and the three processing methods of the prediction result is compared to the three processing methods of the prediction result, and the application of the HAMPEL filter smooth zoom point and the application LOF algorithm.

And shown in Figure 13. The predicted residual capacity is measured by battery B-3 experiment, and the battery A-1 and battery B-4 experiment data are training data sets. Figure 12 Different abnormal value processing mode Figure 13 Capacity estimation result error results show that the unprocessed feature is input to the RBF model after the LLE algorithm is reduced, and the capacity prediction result is poor, the maximum error is close to 1A · h And the result distribution is more dispersion, which means that the abnormal value is not processed, the neural network model cannot effectively learn the mapping relationship between the characteristics and capacity, and does not achieve the ideal forecast results.

After the smoothing of the Hampel filter, the error of the remaining available capacity estimated after the input model is less than 0.2A · h, but the estimation effect in the late battery cycle life is poor. The capacity estimation error corresponding to the characteristics of the LOF algorithm is maintained within 0.

2A · h. When estimating the available capacity of the recurring phenomenon of the capacity, the estimation result corresponding to the Hampel filter is compared with a large gap compared to the estimation result obtained after the LOF algorithm is processed, in the 113th experiment cycle, two The absolute error of the capacity is 0.0253A · H and 0.

1003A · H, respectively, 0.0158A · H and 0.1821A · H, respectively, in the 143th experimental cycle, respectively.

It can be seen that the characteristics of the LOF algorithm have included more detailed battery actual capacity decline information, enabling estimating models to better learn the relationship between lithium-ion battery capacity recession and characteristics, and output more accurate and reliable remaining available Capacity value. 4 Conclusion This paper proposes a method of efficiently estimating the remaining available capacity of lithium-ion batteries. 7 sets of 7 sets of residual capacity related characteristics are extracted from the data of the charging phase, and the LOF algorithm and the LLE algorithm are used to accurately discode the cleaning and low-dimensional mapping of the multidimensional feature quantity, and improve the value of effective information contained in the characteristic value.

Estimated the required calculation needs, enhance estimation efficiency. On this basis, a capacity estimation model is established using the RBF neural network. The results show that the estimation model can effectively estimate the remaining available capacity after the lithium-ion battery recession, which has great advantage over the estimation accuracy and estimation efficiency compared to the ELMAN neural network model and BP neural network model.

This method can achieve efficient and accurate estimation of available capacity after lithium ion battery recession. 聽聽 聽聽 聽聽 聽聽 聽聽 聽聽 聽聽 聽聽 聽聽 聽聽 聽聽 聽聽 聽聽 聽聽 聽聽 聽聽 聽聽 聽聽 聽聽 聽聽 聽聽 聽聽 聽聽 聽聽 聽聽 聽聽 聽聽 聽聽 聽聽 聽聽 聽聽 聽聽 聽聽 聽聽.(CHENZheng,LILeilei,SHUXing,etal.

EfficientremainingcapacityestimationmethodforLIBbasedonfeatureprocessingandtheRBFneuralnetwork[J].EnergyStorageScienceandTechnology,2021,10(01):261-270.)第一作者：陈峥（1982—），男，博士，教授，研究方向为动力Lithium-ion battery status estimation, E-mail: chen@kust.

edu.cn First author: Shenjiangwei, senior experiment, research direction is the status of dynamic lithium-ion battery, e-mail: shenjiangwei6@163.com.

.

CONTACT US

Just tell us your requirements, we can do more than you can imagine.

Send your inquiry

Copyright © 2024 iFlowpower - Guangong iFlowpower Technology Co., Ltd.

Chat with Us

Choose a different language

╳

English
العربية
Deutsch
Español
français
italiano
日本語
한국어
Português
русский
简体中文
繁體中文
Afrikaans
አማርኛ
Azərbaycan
Беларуская
български
বাংলা
Bosanski
Català
Sugbuanon
Corsu
čeština
Cymraeg
dansk
Ελληνικά
Esperanto
Eesti
Euskara
فارسی
Suomi
Frysk
Gaeilgenah
Gàidhlig
Galego
ગુજરાતી
Hausa
Ōlelo Hawaiʻi
हिन्दी
Hmong
Hrvatski
Kreyòl ayisyen
Magyar
հայերեն
bahasa Indonesia
Igbo
Íslenska
עִברִית
Basa Jawa
ქართველი
Қазақ Тілі
ខ្មែរ
ಕನ್ನಡ
Kurdî (Kurmancî)
Кыргызча
Latin
Lëtzebuergesch
ລາວ
lietuvių
latviešu valoda
Malagasy
Maori
Македонски
മലയാളം
Монгол
मराठी
Bahasa Melayu
Maltese
ဗမာ
नेपाली
Nederlands
norsk
Chicheŵa
ਪੰਜਾਬੀ
Polski
پښتو
Română
سنڌي
සිංහල
Slovenčina
Slovenščina
Faasamoa
Shona
Af Soomaali
Shqip
Српски
Sesotho
Sundanese
svenska
Kiswahili
தமிழ்
తెలుగు
Точики
ภาษาไทย
Pilipino
Türkçe
Українська
اردو
O'zbek
Tiếng Việt
Xhosa
יידיש
èdè Yorùbá
Zulu

Current language：English