Lithium ion battery health state evaluation method applied to energy storage

Abstract

The invention discloses a state of health assessment method for an energy storage lithium ion battery. The method comprises the following steps: step 1, constructing a one-dimensional CNN model based on residual connection; step 2, parameter optimization is carried out on the one-dimensional CNN model through a Hyperband hyper-parameter optimization method, and an optimized one-dimensional CNN model is obtained; step 3, training the optimized one-dimensional CNN model; and 4, obtaining the SOH curve of the lithium battery to be evaluated. According to the method, a CNN model is trained by using part of data with a fixed length in a constant current (CC) charging stage, and SOH is estimated on line by using a data segment with a random variable length. According to the method, the situation that SOH estimation fails due to the fact that the charging starting point and the charging ending point are uncertain due to random charging and discharging behaviors of a user is greatly avoided; and an enhanced CNN framework is established for SOH estimation.

Description

Technical Field

[0001] The present invention belongs to the field of batteries, and in particular relates to a method for evaluating the health status of lithium-ion batteries used for energy storage. Background Art

[0002] In recent years, data-driven methods have been widely used for battery SOH estimation. These methods rely on data and do not require accurate modeling of the complex electrochemical mechanisms inside the battery. Their effectiveness mainly depends on the selection of features and algorithms. Commonly used machine learning algorithms include kernel-based methods such as support vector machines, Gaussian process regression, and neural network algorithms such as convolutional neural networks (CNNs), long short-term memory (LSTM) networks, Elman, improved radial basis function neural networks (IRBFNNs), etc. Selecting appropriate health features as model inputs is crucial for accurately evaluating SOH. Due to the large uncertainty in the battery discharge process, health features are usually extracted from the charging process. According to the extraction of health features, current research methods can be divided into two categories: manual feature extraction and automatic feature extraction using neural networks. Manual feature extraction generally extracts various features that are highly correlated with SOH from voltage, current, temperature, and incremental capacity (IC) curves.

[0003] However, most current studies are picky about raw data. For manual feature extraction, the above methods usually require a large amount of measurement data within a running cycle to generate statistical or temporal features, such as average voltage and average current. Other methods require complex data processing. For example, the integrated circuit curve needs to be eliminated by advanced filtering methods to eliminate the influence of noise, which undoubtedly increases the computational burden of data processing. Moreover, the IC curve requires a low charging current rate throughout the constant current charging stage, which is not suitable for fast charging scenarios, which limits the online application of such technologies. For automatic feature extraction using neural networks, although complex feature engineering is avoided, most of these studies still require a specific voltage range and a fixed starting point or end point to ensure that the curves of different periods have the same reference point. However, in actual applications, there is uncertainty in the user's charging and discharging behavior, resulting in the charging starting point and end point being not fixed, which makes most of the above methods difficult to apply in practice.

[0004] Therefore, in order to solve the above problems, it is necessary to develop a SOH evaluation method that can reduce the requirements for raw data. Therefore, the main purpose of this invention is to introduce a new method for estimating the SOH of lithium-ion batteries, which can accurately estimate the health status of the battery using random charging segment data extracted from the constant current charging process. Summary of the invention

[0005] In order to solve the above problems, the present invention discloses a health status assessment method for lithium-ion batteries used in energy storage.

[0006] The present invention is achieved through the following technical solutions:

[0007] A method for evaluating the health status of a lithium-ion battery for energy storage comprises the following steps:

[0008] Step 1: Construct a one-dimensional CNN model based on residual connection;

[0009] Step 2: Optimize the parameters of the one-dimensional CNN model by using the Hyperband hyperparameter optimization method to obtain an optimized one-dimensional CNN model;

[0010] Step 3: Train the optimized one-dimensional CNN model:

[0011] 3.1 Using a sliding window, the charging voltage curve of the CC charging stage in each cycle is truncated into multiple data segments with a charging capacity increment of ΔQ and a step size of ΔQ / 10; performing data preprocessing on the multiple data segments to obtain a training set;

[0012] 3.2 Use the training set to train the optimized one-dimensional CNN model to obtain the one-dimensional CNN model after preliminary training;

[0013] 3.3 Use the data segments of random length in the actual environment to re-learn the trained one-dimensional CNN model to obtain the trained one-dimensional CNN model;

[0014] Step 4: Get the SOH curve of the lithium battery to be evaluated:

[0015] Randomly cut out the original data segment with a charging capacity increment not less than ΔQ from the voltage curve of the CC charging stage of the lithium battery to be evaluated, and use a sliding window with a step size of ΔQ / 100 to intercept multiple data segments with a charging capacity increment of ΔQ, and perform preprocessing to obtain the original data set, and remove the abnormal data in the original data set to obtain the predicted data set; input the predicted data set into the trained one-dimensional CNN model to obtain the SOH curve of the lithium battery to be evaluated;

[0016] Where, ΔQ = k × C0 (2)

[0017] Among them, k∈(0,1) represents the segment length coefficient, C0 represents the nominal capacity;

[0018] When intercepting the charging voltage curve, since the capacity value cannot be measured directly, the ampere-hour integration method is used for calculation:

[0019]

[0020] I(t) represents the charging current, t represents the charging time, t0 and t end are the start and end times of the selected voltage curve.

[0021] For further improvement, the pre-processing method in step 3.1 is as follows:

[0022] 3.1.1 Divide the data segment with a charging capacity increment of ΔQ into 10 voltage curves with equal charging capacity increments, take the voltage values ​​of the right endpoints of the segments on the 10 voltage curves, and use these 10 voltage values ​​as the input data of the first column; the calculation method of the charging capacity increment Δq is as follows:

[0023] 3.1.2 The differential voltage data calculated by formula (4) is used as the input data of the second column of the input matrix;

[0024] dv n =v n -v n-1 (4)

[0025] v n and v n-1 Represent the sampling voltage at the nth moment and the (n-1)th moment respectively; dv n is the nth element in the DV sequence;

[0026] 3.1.3: The charging capacity from the start point of the data segment to the current moment forms a capacity sequence as the input data of the third column;

[0027] 3.1.4: The input data of the training set obtained is as follows:

[0028]

[0029] As a further improvement, the one-dimensional CNN model in step 1 includes 22 one-dimensional convolutional layers, 7 addition layers, a flattening layer, a discard layer, a first fully connected layer, a second fully connected layer, a flattening layer, and a fully connected layer in sequence along the data processing direction;

[0030] After each one-dimensional convolutional layer, an activation layer is set. An activation layer uses the "tanh" activation function to perform nonlinear transformation on the output, which can be expressed as:

[0031]

[0032] f(x) represents the output, x represents the input, and e is the natural logarithm.

[0033] Further improvement, in step 3.1, when the segmentation coefficient k in formula (2) is set to 0.3, the constant current charging voltage data of 30% of the rated capacity length is taken, and after cleaning the data, the voltage data is extracted every 0.3% of the rated capacity using the linear interpolation method, and the voltage sequence obtained by linear interpolation is used as the input of the next data preprocessing process;

[0034] The differential voltage is normalized using formula (9), and for the capacity series, formula (10) is used for normalization.

[0035]

[0036]

[0037]

[0038] v max and v min are the upper and lower cut-off voltages of the battery, v is the open circuit voltage of the battery, q max and q min They represent the maximum and minimum values ​​of the capacity sequence in the segmented data, q is the capacity of the battery; v n is the normalized battery voltage at time n, q n is the normalized battery capacity at time n.

[0039] As a further improvement, in step 3.2, the loss function loss is minimized to obtain the trained one-dimensional CNN model:

[0040]

[0041] Where N is the number of samples in the training set. is the estimated SOH value of the i-th sample, SOH i,real is the actual value of SOH.

[0042] The present invention has the following beneficial effects:

[0043] (1) An innovative method is proposed, which uses part of the fixed-length data to train the CNN model during the constant current (CC) charging stage and uses random variable-length data segments to estimate SOH online. This method greatly avoids the failure of SOH estimation due to the uncertainty of the charging start and end points caused by the user's random charging and discharging behavior.

[0044] (2) An enhanced CNN framework is established for SOH estimation. In the end-to-end estimation method, the input sample size of the model is very large. Therefore, the designed model adopts a deeper network structure, which enables it to learn more complex and abstract features. In order to effectively avoid network degradation, gradient vanishing and gradient exploding, and to accelerate the training process, shortcut connections are introduced in the deep convolutional network. In addition, the Hyperband algorithm is introduced to optimize the model hyperparameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is the D-CNN model architecture diagram;

[0046] Figure 2 Flowchart for SOH estimation;

[0047] Figure 3 This is a schematic diagram of sliding window sampling;

[0048] Figure 4 SOH decline trajectory: (a) NCA battery, (b) NCM battery.

[0049] Figure 5 Error distribution of TL test batteries using different training batteries (a) Error distribution of CY25-05_1 NCA battery, (b) Error distribution of CY25-05_1 NCM battery, (c) Error distribution of CY45-05_1 NCM battery;

[0050] Figure 6 The SOH estimation diagram of NCM CY25-05_1-#7 battery; (a) is the SOH estimation diagram of CNNDT, and (b) is the SOH estimation diagram of CNNTL;

[0051] Figure 7 (a) is the SOH estimation diagram of NCM CY25-05_1-#19 battery; (b) is the SOH estimation diagram of CNNDT and (c) is the SOH estimation diagram of CNNTL. DETAILED DESCRIPTION

[0052] The technical solutions in the embodiments of the present invention will be clearly and completely described below in combination with the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0053] 1. Enhanced framework for SOH estimation

[0054] In this paper, we use random and variable-length data from the CC charging phase to estimate SOH, and use statistical methods to improve the accuracy of the estimation. During the model construction process, we design a one-dimensional convolutional neural network (1D-CNN) to automatically extract features from the input data and estimate SOH. By introducing residual connections and superband optimization algorithms, the learning ability, training speed, and estimation accuracy of the model are greatly improved. Finally, transfer learning is also introduced to enhance the generalization ability of the model. This section will introduce the proposed method in detail.

[0055] 1.1 One-dimensional CNN model based on residual connection

[0056] CNN is a type of artificial neural network used for automatic feature extraction in deep neural networks. Because it adopts local connection and weight sharing, the number of weights is greatly reduced, the complexity of the model is reduced, the network is easy to optimize, and the risk of overfitting is reduced. In the present invention, the input data is a 10×3 matrix, representing a one-dimensional time series, with three channels, and only contains time information. Therefore, the present invention uses a 1D-CNN model to automatically extract health features from the data and establish a mapping relationship with SOH.

[0057] The design of 1D-CNN model is the core task to realize battery capacity estimation. Figure 4 As shown in the figure, the developed model mainly consists of convolutional layers, flattening layers, fully connected layers, and elimination layers. In addition, the model also introduces shortcut connections. The one-dimensional convolutional layer extracts the local area feature information of the sample layer by layer through multiple convolution kernels. After each convolutional layer, the output is nonlinearly transformed using the "tanh" activation function, which is expressed as

[0058]

[0059] Since the input sequence in the present invention is short, the pooling layer commonly used in convolutional neural networks is omitted. In order to rearrange the data of all channels into a one-dimensional sequence, a tiled layer connection is used between the convolution layer and the fully connected layer. Both fully connected layers use linear activation functions to combine and transform the feature outputs extracted by convolution, and the ultimate goal is to obtain the SOH estimation result. In order to effectively avoid overfitting and achieve regularization to a certain extent, a elimination layer is added after the convolution layer. The elimination layer was first proposed in AlexNet. By introducing the elimination layer, the dependency between neurons is reduced, the risk of overfitting is reduced, and the generalization ability of the model is improved. In addition, the model also introduces shortcut connections. By adopting shortcut connections, the problem of network degradation when the depth of the neural network is too large is effectively overcome, the occurrence of gradient dissipation and explosion is avoided, the symmetry of the network is broken, and the representation ability of the network is improved.

[0060] Compared with other CNN models for SOH estimation, the designed network has a deeper depth, i.e., more layers. In order to avoid lengthy data, a 10×3 matrix is ​​used as input, which is much shorter than other methods. However, by using a sliding window technique for each cycle, multiple sets of samples can be obtained from one cycle, resulting in a large number of input samples. Therefore, adopting a deeper network structure can enable the model to learn more complex and abstract features, thereby improving the representation ability.

[0061] 1.2 SOH Estimation Framework

[0062] Figure 2The SOH estimation framework proposed in the present invention mainly includes four aspects: data preprocessing, model training, hyperparameter optimization and online estimation. This section will explain the above four aspects in detail.

[0063] Data preprocessing

[0064] Before training the CNN model, data preprocessing must be performed first, including data outlier cleaning, data filling and normalization. Data cleaning is the process of identifying and correcting inaccurate records in a data set. Data filling uses linear interpolation technology to replace irrelevant charging voltage data. In the present invention, when the segmentation coefficient k is set to 0.3, constant current charging voltage data with a length of 30% of the rated capacity is taken. After cleaning the data, linear interpolation is used to extract voltage data every 0.3% of the rated capacity. The voltage sequence obtained by linear interpolation is used as the input of the next data preprocessing process, where 0.3% is obtained from the percentage of 30%.

[0065] In order to improve the stability of the one-dimensional convolutional neural network based on shortcut connection and speed up the training process, the voltage series extracted from each segment is normalized using formula (8). For DV data, when the charging capacity is small, its value is large, and in the subsequent CC charging stage, its value is small. In order to map the differential voltage data to 0-1 without making most of the values ​​too small, we use formula (9) to normalize the differential voltage. For the capacity series, formula (10) is used for normalization.

[0066]

[0067]

[0068]

[0069] v max and v min are the upper and lower cut-off voltages of the batteries mentioned in Table 1. max and q min Respectively represent the maximum and minimum values ​​of the capacity sequence in the segmented data.

[0070] Hyperparameter Optimization

[0071] Hyperband is a hyperparameter optimization method that improves efficiency by evaluating hyperparameter configurations in multiple resource budgets (i.e., different training cycles). The specific process is shown in Algorithm 1, where R is the type of hyperparameter combination with the most allocated resources, and η is the ratio of hyperparameter combinations retained in each round of control. The present invention uses the Hyperband optimization algorithm to optimize the model structure and some hyperparameters, with a maximum of 40 rounds. The initial learning rate is set to 0.001. When the duration is greater than 15, the learning rate is updated according to formula (11):

[0072] lr=0.001*e 0.1*(30-epoch) (11)

[0073] lr is the current learning rate and epoch is the current number of iterations.

[0074]

[0075] Model Training

[0076] Model training consists of two parts, namely model pre-training and model fine-tuning. Model pre-training is called basic model training, and model fine-tuning is called transfer learning model training.

[0077] In order to make full use of the training battery data, a sliding window is used to truncate the charging voltage curve of the CC charging stage in each cycle into multiple data segments of length ΔQ, with a step size of ΔQ / 10. Therefore, multiple corresponding samples are generated for each cycle.

[0078] 1) Training of basic models

[0079] The initial learning rate of the model is set to 0.001. In the first 30 rounds, the learning rate remains unchanged so that the model can find the global optimum as much as possible. After 30 rounds, the learning rate is changed according to formula (11).

[0080] The loss function of the model is defined as the mean absolute error of the SOH estimate, expressed as formula (12):

[0081]

[0082] Where N is the number of samples in the training set. is the estimated SOH value of the i-th sample, SOH i,real is the actual value of SOH. When the loss function on the validation set does not decrease for 20 consecutive iterations or reaches the set maximum round 60, the training ends and the model with the minimum loss function on the validation set is saved.

[0083] SOH estimation output

[0084] Considering the uncertainty of user charging and discharging behavior in real life, during the test, we randomly cut out the original data segment of no less than ΔQ from the voltage curve of the CC charging stage, and cut it into multiple data segments of different lengths using a sliding window. Different from the step size of the sliding window during training, when estimating a new test set, the step size of the sliding window is ΔQ / 100, which greatly increases the number of inputs to the model. Figure 6 As shown in Figure 1, multiple SOH prediction values ​​can be obtained through multiple input data. The interquartile range (IQR) method is then used to remove outliers in the prediction values, and the average of the remaining estimation results is taken as the final SOH estimation value.

[0085] The IQR method used in this paper is a common method for detecting outliers in statistics. It sorts the data by size to obtain the data of the upper quartile (Q3) and the lower quartile (Q1). The interquartile range is calculated according to formula (13). Then the threshold of the outlier is determined, that is, when the data exceeds the threshold, the value is regarded as an outlier and deleted. The present invention sets the threshold to -1.5IQR and +1.5IQR.

[0086] IQR=Q3-Q1 (13)

[0087] The battery aging data comes from the lithium-ion battery cycle aging data set provided by Wei Xiaoming's team at Tongji University. All experimental batteries are commercial 18650 batteries. In this data set, we selected batteries with two positive electrode materials: nickel cobalt aluminum (NCA) and nickel cobalt manganese (NCM). Each battery is named "CYX-Y_Z-#N", where X represents the ambient temperature, Y_Z represents the charge and discharge current rate, and #N represents the battery label. The present invention selects NCA and NCM material batteries with ambient temperatures of 25°C and 45°C, a charging current rate of 0.5C, and a discharge current rate of 1C. Detailed parameters and charge and discharge overview are shown in Table 1.

[0088] Table. Battery parameters.

[0089]

[0090]

[0091] The SOH decline curve of each battery used in the present invention is as follows: Figure 4 shown. Figure 4 (a) and (b) show the SOH decline curves of selected batteries in the NCA and NCM battery datasets, respectively. The same color represents the same test environment, and different colors represent different test environments. It can be seen that even for the same type of battery, there are differences in its SOH decline trajectory under the same test environment. There are also significant differences in the aging process of the same type of battery under different test conditions.

[0092] Verifying the effectiveness of transfer learning

[0093] The present invention selects CY45-05_1 battery with NCA as the positive electrode material as the source data for the pre-trained basic model. In order to prove the superiority of the TL model, the CNN model is directly trained without using transfer learning. The model using TL is denoted as CNNTL, and the model directly trained is denoted as CNNDT. A small amount of data of CY25-05_1 NCA battery, CY25-05_1 NCM battery and CY45-05_1 NCM battery are used as training sets respectively. The remaining batteries are used for testing. Figure 5 The error distributions of all tested batteries based on the above three TL models are shown.

[0094] We demonstrate the SOH estimation results of fine-tuning the base model using a small number of cells CY25-05_1 from the NCM dataset. Figure 6 The estimation results of the two test batteries using TL and the direct 1D-CNN model are compared. It can be seen that the curve estimated using CNNTL is closer to the actual curve than the curve estimated using CNNDT, especially when the battery SOH is low, which is extremely important for timely battery replacement. In addition, it can be seen that the distribution center of the estimated value of the CNNDT method deviates from the actual value of SOH, while the distribution center of the estimated value of CNNTL is almost close to the actual value of SOH, which indicates that CNNDT has an overfitting problem due to insufficient data, while CNNTL has better generalization ability than CNNDT. Table 2 lists the estimation errors of the two test batteries, and it can be seen that the SOH estimation error of CNNTL is smaller than that of CNNDT.

[0095] Table 2. SOH estimation errors of the tested batteries using the direct CNN model and the CNNTL model.

[0096]

[0097] Based on the above experimental results and analysis, it can be concluded that after the introduction of transfer learning, the method proposed in this study only needs a small amount of target domain data to be extended to SOH estimation under different environments and battery materials. Moreover, when only a small amount of target domain training data is used, CNNTL outperforms CNNDT in estimation accuracy, training speed, computational cost, and generalization.

[0098] The preferred embodiments of the present invention disclosed above are only used to help explain the present invention. The preferred embodiments do not describe all details in detail, nor do they limit the invention to only the specific implementation methods described.

Claims

1. A method for evaluating the health status of a lithium-ion battery for energy storage, characterized in that: The steps include: Step 1: Construct a one-dimensional CNN model based on residual connection; Step 2: Optimize the parameters of the one-dimensional CNN model by using the Hyperband hyperparameter optimization method to obtain an optimized one-dimensional CNN model; Step 3: Train the optimized one-dimensional CNN model: 3.1 Using a sliding window, the charging voltage curve of the CC charging stage in each cycle is truncated into multiple data segments with a charging capacity increment of ΔQ and a step size of ΔQ / 10; performing data preprocessing on the multiple data segments to obtain a training set; 3.2 Use the training set to train the optimized one-dimensional CNN model to obtain the one-dimensional CNN model after preliminary training; 3.3 Use the data segments of random length in the actual environment to re-learn the trained one-dimensional CNN model to obtain the trained one-dimensional CNN model; Step 4: Get the SOH curve of the lithium battery to be evaluated: Randomly cut out the original data segment with a charging capacity increment not less than ΔQ from the voltage curve of the CC charging stage of the lithium battery to be evaluated, and use a sliding window with a step size of ΔQ / 100 to intercept multiple data segments with a charging capacity increment of ΔQ, and perform preprocessing to obtain the original data set, and remove the abnormal data in the original data set to obtain the predicted data set; input the predicted data set into the trained one-dimensional CNN model to obtain the SOH curve of the lithium battery to be evaluated; Where, ΔQ = k × C0 (2) Among them, k∈(0,1) represents the segment length coefficient, C0 represents the nominal capacity; When intercepting the charging voltage curve, since the capacity value cannot be measured directly, the ampere-hour integration method is used for calculation: I(t) represents the charging current, t represents the charging time, t0 and t end are the start and end times of the selected voltage curve.

2. The method for evaluating the health status of a lithium-ion battery for energy storage as claimed in claim 1, characterized in that: The method of the pretreatment in step 3.1 is as follows: 3.1.1 Divide the data segment with a charging capacity increment of ΔQ into 10 voltage curves with equal charging capacity increments, take the voltage values ​​of the right endpoints of the segments on the 10 voltage curves, and use these 10 voltage values ​​as the input data of the first column; the calculation method of the charging capacity increment Δq is as follows: 3.1.2 The differential voltage data calculated by formula (4) is used as the input data of the second column of the input matrix; dv n =v n -vn-1 (4) v n and v n-1 Represent the sampling voltage at the nth moment and the (n-1)th moment respectively; dv n is the nth element in the DV sequence; 3.1.3: The charging capacity from the start point of the data segment to the current moment forms a capacity sequence as the input data of the third column; 3.1.4: The input data of the training set obtained is as follows:

3. The method for evaluating the health status of a lithium-ion battery for energy storage according to claim 1, characterized in that: The one-dimensional CNN model in step 1 includes 22 one-dimensional convolutional layers, 7 addition layers, a flattening layer, a discard layer, a first fully connected layer, a second fully connected layer, a flattening layer, and a fully connected layer in sequence along the data processing direction; Each one-dimensional convolutional layer is followed by an activation layer, which uses the "tanh" activation function to perform a nonlinear transformation on the output, expressed as: f(x) represents the output, x represents the input, and e is the natural logarithm.

4. The method for evaluating the health status of a lithium-ion battery for energy storage according to claim 2, wherein: In step 3.1, when the segmentation coefficient k in formula (2) is set to 0.3, the constant current charging voltage data of 30% of the rated capacity length is taken, and after cleaning the data, the voltage data is extracted every 0.3% of the rated capacity using the linear interpolation method, and the voltage sequence obtained by the linear interpolation is used as the input of the next data preprocessing process; The differential voltage is normalized using formula (9), and for the capacity series, formula (10) is used for normalization. v max and v min are the upper and lower cut-off voltages of the battery, v is the open circuit voltage of the battery, q max and q min They represent the maximum and minimum values ​​of the capacity sequence in the segmented data, q is the capacity of the battery; v n is the normalized battery voltage at time n, q n is the normalized battery capacity at time n.

5. The method for evaluating the health status of a lithium-ion battery for energy storage according to claim 1, characterized in that: In step 3.2, the loss function is minimized to obtain the trained one-dimensional CNN model: Where N is the number of samples in the training set. is the estimated SOH value of the i-th sample, SOH i,real is the actual value of SOH.

Images