Battery capacity online estimation method based on model parameters

Abstract

The invention discloses a battery capacity online estimation method based on model parameters, and belongs to the technical field of battery capacity estimation. Comprising the following steps: acquiring voltage and current data of a battery, and performing parameter identification and feature processing on the voltage and current data; performing correlation analysis on the battery model parameter and the maximum available capacity thereof, and extracting a battery health factor; building a bidirectional gating cycle unit model based on a double attention mechanism, and dividing battery data into a training data set and a test data set; and inputting the training data set into the bidirectional gating cycle unit model for training until the loss function converges and then carries out battery capacity online estimation together with the test data set. According to the method, a double attention mechanism is introduced, the correlation degree of battery model parameters, historical information and capacity is analyzed from the two aspects of features and time sequences, the features and time sequences with high contribution degree are highlighted through dynamic weight distribution, and the model precision is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of battery capacity estimation, and in particular to an online battery capacity estimation method based on model parameters. Background Art

[0002] Batteries will be worn out during use, causing the battery capacity to decrease slightly compared to the initial capacity. For different batteries, the decay rate of battery capacity is not exactly the same while ensuring that the driving distance and usage time remain consistent. This is because the internal resistance of different batteries is not consistent. The greater the internal resistance, the greater the battery decay, and there is a strong correlation between the two. In this case, the existing technology cannot accurately estimate the battery capacity online. Therefore, a battery capacity online estimation method based on model parameters is needed to solve the above problems. Summary of the invention

[0003] The purpose of the present invention is to propose a battery capacity online estimation method based on model parameters, comprising the following steps:

[0004] Step 1: Obtain the voltage and current data of the battery, and perform parameter identification and feature processing;

[0005] Step 2: Perform correlation analysis on battery model parameters and their maximum available capacity to extract battery health factors;

[0006] Step 3: Build a bidirectional gated recurrent unit model based on the dual attention mechanism, and divide the battery data into training data set and test data set;

[0007] Step 4: Input the training dataset into the bidirectional gated recurrent unit model for training until the loss function converges, and then perform online battery capacity estimation together with the test dataset.

[0008] In step 2, correlation analysis is performed using the Spearman coefficient and the maximum mutual information coefficient.

[0009] The bidirectional gated recurrent unit model in step 3 includes a feature attention mechanism encoder and a temporal attention mechanism decoder.

[0010] The formula for calculating the attention weight in the feature attention mechanism encoder is as follows:

[0011]

[0012] Among them, e t,m represents the attention weight coefficient corresponding to feature m at time step t; v e ∈R T , W e ∈R T×2pis the weight coefficient that the neuron needs to learn, where p is the size of the last hidden layer of the encoder; B e ∈R n is the bias coefficient.

[0013] The bidirectional gated recurrent unit model is:

[0014] e t =utanh(wh t +b a )

[0015]

[0016] In the formula, h t (t∈[0,n]) is the hidden state corresponding to each input, α t (t∈[0,n]) is the attention weight, y t is the output value of Bi-GRU after the attention mechanism is introduced. u and w are weight coefficients. b a is the bias coefficient, e t is the attention weight value at time t, α t is the normalized weight probability distribution value.

[0017] The beneficial effects of the present invention are:

[0018] 1. The present invention introduces a dual attention mechanism, which analyzes the correlation between battery model parameters, historical information and capacity from the aspects of features and timing respectively, and highlights the features and timing with high contribution by dynamically allocating weights, thereby improving the model accuracy.

[0019] 2. The present invention takes into account both battery health characteristics and time series data, has low dependence on data volume, and has strong model reliability, high prediction accuracy, fast convergence speed, and good online application capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a flow chart of the battery capacity online estimation method based on model parameters of the present invention;

[0021] Figure 2 The capacity battery degradation curve of Oxford battery;

[0022] Figure 3 (a) (b) are the three-dimensional surface and two-dimensional curve of the OCV-SOC-Cycles relationship respectively;

[0023] Figure 4 This is the error result diagram of Oxford data identification;

[0024] Figure 5 (a) and (b) are the variation curves of resistance and capacitance of battery model parameters during the whole life cycle;

[0025] Figure 6 is the curve of each parameter value after normalization;

[0026] Figure 7 Bi-GRU structure diagram;

[0027] Figure 8 This is the structure diagram of the feature attention mechanism encoder;

[0028] Fig. 9 This is the structure diagram of the temporal attention mechanism decoder;

[0029] Fig.10 DA-BiGRU structure diagram for battery capacity estimation;

[0030] Fig.11 An overall flow chart to verify effectiveness;

[0031] Fig.12 (a)~(h) are the prediction results on 8 cells respectively;

[0032] Fig.13 This is the Loss curve diagram during model training. DETAILED DESCRIPTION

[0033] The present invention proposes a battery capacity online estimation method based on model parameters, which is further described below in conjunction with the accompanying drawings and specific embodiments.

[0034] Figure 1 The flowchart of the method for online estimating battery capacity based on model parameters of the present invention comprises the following steps:

[0035] Step 1: Obtain the voltage and current data of the battery, and perform parameter identification and feature processing on it.

[0036] This example uses the Oxford battery data set for analysis. It covers the aging data of 8 lithium battery cells, and its capacity degradation curve is as follows: Figure 2 shown.

[0037] The battery used is the soft pack battery of Kokam Co., Ltd., model SLPB533459H4, with a rated capacity of 740mAh. The test environment temperature is controlled at 40℃ by a constant temperature box. The test process is constant current and constant voltage charging, Artemis urban working condition discharge, and performance test is performed every 100 cycles. The publicly available data is the performance test results of 8 batteries every 100 cycles. The performance test includes 1C cycle and pseudo OCV cycle.

[0038] from Figure 2The comparison of different curves in the figure can lead to the following conclusions: During the continuous use of the battery, each single cell has different degrees of loss, and the loss increases slightly with the increase in the number of cycles. This shows that the loss of the battery is continuous during use. The more times the battery is used, the more serious its aging will be, and the capacity will be lower and lower, resulting in worse and worse battery performance. Therefore, each battery has a certain service life and cannot be used indefinitely.

[0039] First, we use the full life cycle data of cell1 in the Oxford battery data. Cell1 has a total of 8200 charge and discharge cycles and 78 performance tests. A pseudo OCV test was performed in each performance test. Therefore, we first make a three-dimensional OCV-SOC-Cycles curve under the full life cycle. In order to more clearly and intuitively observe the changes of the OCV-SOC curve with life attenuation, we make three-dimensional and two-dimensional curves, such as Figure 3 As shown in (a)(b).

[0040] Comprehensive analysis Figure 3 (a)(b) It can be seen that although the battery life decay has a small impact on the change of the OCV-SOC curve, it still has a certain impact, which is mainly concentrated in the 10%~20% SOC range, while the OCV-SOC curve in the 30%~40% SOC range is almost completely overlapped. Therefore, when doing parameter identification, selecting voltage and current data in the 30%~40% SOC range can eliminate the OCV error caused by battery life decay to a certain extent.

[0041] After obtaining the OCV-SOC relationship, the parameters of a single cycle are first identified, and the identification error is as follows: Figure 4 As shown, the error is controlled within 0.25% after stabilization, and the algorithm has high accuracy.

[0042] Then, based on 78 performance tests, full life cycle parameter identification was performed. The results are as follows: Figure 5 (a)(b). As can be seen from the figure, the five battery model parameters change with the deepening of battery aging, and show different change trends. 1 , C 2 As the number of cycles increases, it generally shows a downward trend, which keeps pace with the battery capacity decline trend and is positively correlated with the battery aging degree. The three resistance parameters R 0 , R 1 , R 2 The change trend is opposite, and it generally shows an upward trend or remains basically unchanged as the number of cycles increases. The five model parameters are all correlated with the battery capacity decay, so they can be selected as the relevant factors of the subsequent capacity estimation model. 1 , C2 The volatility of these two model parameters is relatively large. In order to reduce the impact of data fluctuations on model accuracy, the identified model parameters R 0 , R 1 , R 2 , C 1 , C 2 Filtering is required. At the same time, the value ranges of the five model parameters are different, and the order of magnitude is quite different, so homogenization is required. Finally, the pre-processed model parameters are used for correlation analysis with the battery capacity, and the parameters with stronger correlation are selected as the input for subsequent capacity estimation.

[0043] In order to solve the above C 1 , C 2 For fluctuation problems, select the Hampel filter method, set a filter window and threshold, the threshold is generally 3 or 4 times the median absolute deviation (MAD) value, remove the numbers exceeding the threshold and replace them with the median of the window. The specific steps are as follows:

[0044] (1) Determine the size of the sliding window. The window is usually symmetrical, for example, taking k points before and after the current point.

[0045] (2) Calculate the median and median absolute difference (MAD) in the window and define a threshold (usually 3 or 4 times the MAD).

[0046] (3) Calculate the difference between each data point and the value in the window. If it is greater than the threshold, replace the value with the median.

[0047] (4) Repeat the above steps until the last data point is reached, indicating that the filtering is completed.

[0048] The formulas used in Hampel filtering are as follows: (1) and (2). k is the median value of the kth window; S k is the value of MAD; 3 times MAD is selected as the threshold.

[0049]

[0050] S k =1.4286median{|x k -m k ∣} (2)

[0051] After filtering, the parameters must be standardized to make the dimensions of each model parameter the same. In this way, the model will not give higher weights to some model parameters because of their larger values ​​during training, thereby ensuring that each feature contributes to the model in a balanced manner, and can also accelerate model convergence and enhance model stability. The min-max standardization method is used to scale the original linear data to between [0,1], and the formula is as follows:

[0052]

[0053] In the formula, x i 、x min and x max represent the actual value, theoretical maximum value, and theoretical minimum value of the ith model parameter, respectively. Represents the value after normalization. The normalized parameter values ​​are as follows Figure 6 shown.

[0054] Step 2: Perform correlation analysis between battery model parameters and their maximum available capacity to extract battery health factors.

[0055] In order to intuitively understand the strength of the correlation between the battery model parameters and its maximum available capacity, it is sufficient to directly calculate the Spearman coefficient and the maximum mutual information coefficient (MIC) between the two. The Spearman coefficient is a non-parametric method used to evaluate a single relationship between two variables. It does not assume the distribution form of the data and can be applied to non-normally distributed data. The MIC algorithm can not only detect linear relationships, but also identify arbitrary nonlinear relationships, and can reveal complex relationships that may be missed by traditional correlation coefficients. In order to ensure the rationality of the selected model parameters, these two coefficients are calculated separately for comprehensive analysis.

[0056] The steps for calculating the Spearman correlation coefficient are as follows:

[0057] (1) Sort and number each variable data to obtain the corresponding ranking. If there are data points with the same numerical value, the average ranking can be taken;

[0058] (2) Calculate the corresponding rank difference, that is:

[0059] d i = rank(x i ) - rank(y i ) (4)

[0060] (3) Calculate the Spearman correlation coefficient:

[0061]

[0062] Among them, x i represents the battery model parameters, y i represents the battery capacity; n represents the number of battery aging cycles; the calculated Spearman correlation coefficient value is between -1 and 1, the closer to -1, the stronger the negative correlation, the closer to 1, the stronger the positive correlation, and the closer to 0, the weaker the correlation.

[0063] The pseudo code of the MIC algorithm is shown in Table 1, where the capacity is C L ={x j , j = 1, 2, ..., N}, the battery model parameter is F H (i) = {y j , j = 1, 2, ..., N}, then the mutual information MI(C L ; F H (i)) can be calculated as follows:

[0064]

[0065] Where p(x,y) is C L and F H (i) joint probability density, p(x) is C L The marginal probability density of F, p(y) is H (i) The marginal probability density of j ,y j ), j = 1, 2, ..., N}, and at the same time, the variable C L and F H (i) The data range is divided into x and y segments, and the x*y grid G ​​is obtained. First, the mutual information between x and y is calculated in each grid, and then the mutual information under different grid divisions is calculated to obtain a feature matrix. Finally, the maximum value in the matrix is ​​found, which is the value of MIC. MIC can be calculated using formula (7).

[0066]

[0067] The calculated MIC coefficient value ranges from 0 to 1, and the closer it is to 1, the stronger the correlation.

[0068] Table 1 MIC algorithm pseudo code

[0069]

[0070] Tables 2 and 3 show the Spearman and MIC correlation coefficients between the model parameters and capacity of the Oxford University public data set cell1, cell2, cell3, cell4, cell5, cell6, cell7, and cell8 batteries obtained by analysis.

[0071] Table 2 Cell health factor and battery capacity Spearman coefficient

[0072]

[0073] Table 3 Cell health factor and battery capacity MIC coefficient

[0074]

[0075]

[0076] It can be seen from the table that R 1 The two coefficients of are relatively small numbers, and the rest of R 0 , R 2 , C 1 , C 2 The two coefficient values ​​of are both above 0.8, proving that the battery model parameter R 0 , R 2 , C 1 , C 2 They are strongly correlated with the battery capacity, so they are selected as the input for subsequent capacity estimation.

[0077] Step 3: Build a bidirectional gated recurrent unit model based on the dual attention mechanism, and divide the battery data into training data set and test data set.

[0078] Unlike traditional feedforward neural networks, recurrent neural networks can "memorize" previous input information by forming directed loops through internal units. This feature also gives recurrent neural networks a unique advantage in processing time series problems such as sentence translation and semantic recognition. However, when performing backpropagation, the traditional recurrent neural network, also known as the simple recurrent neural network (SRNN), may have exponential decay or growth in the time step due to repeated multiplication of the weight matrix, which makes it difficult for RNNs on long sequences to learn long-distance dependencies. Therefore, a series of improvements and variant structures such as the Long Short-Term Memory (LSTM) unit and the Gated Recurrent Unit (GRU) have been proposed.

[0079] Compared with LSTM units, GRU has a simple structure, the model is easier to converge during training, and performs better on some small data sets. Therefore, GRU is used to build a codec to realize the capacity estimation of lithium-ion batteries.

[0080] in is a candidate state, h t is the hidden state, and t is the time step.t is the update gate, r t is the reset gate. The entire GRU unit operation process is as follows.

[0081] (1) GRU receives the input sequence x t , and GRU also has a hidden state, the initial hidden state h 0 Typically a zero vector or random initialization.

[0082] (2) Update gate calculation. Update gate z t The calculation formula is shown in the following formula (8). Where σ is the sigmoid function, W z and b z are the weights and biases of the update gate, [h t-1 ,x t ] means concatenating the hidden state of the previous time step and the input of the current time step.

[0083] z t =σ(W z ·[h t-1 ,x t ]+b z ) (8)

[0084] (3) Reset gate calculation. t The calculation formula of is shown in the following formula (9).

[0085] r t =σ(W r ·[h t-1 ,x t ]+b r ) (9)

[0086] (4) Update the current memory unit. According to the result of the reset gate, calculate the candidate state of the current memory unit. The calculation formula is shown in the following formula (10).

[0087]

[0088] (5) Update the hidden state. The update formula is as shown in formula (11).

[0089]

[0090] Through the above steps, GRU is able to dynamically control the flow and memory of information at each time step, thereby capturing long-term dependencies in the sequence.

[0091] The gated recurrent unit can only learn from past information and may not be able to fully capture the global dependencies in the sequence. The Bi-GRU neural network model is selected. It combines the bidirectional recurrent structure on the basis of the above GRU gating mechanism to process the input from both directions of the sequence (forward and reverse). This design enables the model to consider both past and future information at the same time and better capture the contextual relationship in the sequence. Its structure is as follows Figure 7 As shown, G t represents the forward GRU, G t ' represents the reverse GRU.

[0092] Select the battery model parameter R 0 , R 2 , C 1 , C 2 As the input x of Bi-GRU t , construct the matrix X, whose details are shown in formula (12). In the formula, each row represents the set of all input features at a time step, and each column represents the set of all time steps of a feature.

[0093]

[0094] After the input, the hidden state of the network contains two, namely, forward propagation and back propagation These two are concatenated to form the hidden state h output by the Bi-GRU layer. t , the specific calculation formula is as follows:

[0095]

[0096]

[0097]

[0098] Among them, f GRU (·) represents the GRU network function, and the arrows represent the propagation direction.

[0099] The Bi-GRU model combines the advantages of the gating mechanism and the bidirectional recurrent structure, which enables it to perform well in many sequence modeling and prediction tasks. Therefore, Bi-GRU is selected as the key part of the capacity estimation model.

[0100] In order to explore the intrinsic characteristics of battery model parameters, highlight the influence of important feature data, and improve the accuracy of model operation, the attention mechanism is introduced. The attention mechanism allows the model to dynamically adjust the weights of each feature when processing the input of each time step, so as to focus more on task-critical information. Its introduction can increase the model's attention to important information in the sequence and further improve the performance of the model. On the other hand, the attention mechanism can also dynamically adjust the weight of each time step, so that the model pays more attention to those time steps that are most helpful for the current output prediction, thereby alleviating the problem of difficulty in capturing long-term dependencies caused by gradient disappearance in traditional recurrent neural networks. In response to these two aspects, the attention mechanism is introduced before and after the Bi-GRU, taking into account the different importance of the input and the influence of the historical state to maximize the performance of the model.

[0101] Among them, x t (t∈[0,n]) is the output at n time steps, h t (t∈[0,n]) is the hidden state corresponding to each input, α t (t∈[0,n]) is the attention weight, and y is the output value of Bi-GRU after the attention mechanism is introduced. The formula of this process is as follows:

[0102] e t =utanh(wh t +b a ) (16)

[0103]

[0104]

[0105] In the formula, u and w are weight coefficients, b a is the bias coefficient. e t is the attention weight value at time t, α t is the normalized weight probability distribution value. The purpose of normalization is to ensure that the sum of the weights is 1.

[0106] The Encoder-Decoder model is a deep learning architecture commonly used for sequence-to-sequence tasks. Its main goal is to take a sequence as input and generate another sequence as output. It consists of three parts: an encoder, an intermediate vector, and a decoder. The Encoder-Decoder model solves the defect that traditional neural networks can only encode input sequences in fixed dimensions and cannot be used for sequence-to-sequence mapping.

[0107] The encoder is responsible for converting the input sequence into a context vector or a fixed-length representation. It is usually composed of multiple recurrent units (RNN, GRU, etc.), each of which is responsible for processing an element in the input sequence and encoding its information into the hidden state.

[0108] The intermediate vector is the final output of the encoder after processing the entire input sequence and serves as the initial hidden state of the decoder. It encapsulates the semantic information of the input sequence and can be seen as the representation obtained after the encoder encodes the entire input sequence. It also provides the decoder with information about the input sequence and a starting point for generating the target sequence.

[0109] The decoder, like the encoder, consists of multiple recurrent cells. It receives the intermediate vector as the initial state and generates the target sequence based on this information. The decoder generates an output at each time step until the complete target sequence is generated. During the generation process, the decoder generates the output of the next time step by continuously receiving the output of the previous time step and the hidden state of the current time step.

[0110] Combining the attention mechanism and the Encoder-Decoder model, we get the encoder-decoder model based on the dual attention mechanism Bi-GRU, as shown in the following formula.

[0111] Encoder(x)=σ(W e x+B e ) (19)

[0112] Decoder(x)=σ(W s x+B s ) (20)

[0113] Where σ(·) is the activation function; W e,s is the weight matrix; B e,s is the bias coefficient matrix.

[0114] The encoder combines feature attention, and the input battery model parameters are features. It focuses on mining the correlation information between battery model parameters and dynamically assigns weights to them, so that the parameters that can most effectively reflect the battery aging state contribute more significant value. The decoder combines time attention, fully considers the impact of historical states, dynamically adjusts the weight of each time step, finds time steps that are more helpful to model prediction, and makes them contribute higher value.

[0115] Although correlation analysis has been performed and R values ​​with Spearman coefficient and MIC coefficient greater than 0.8 have been selected, 0 , R 2 , C 1 , C 2These four strongly correlated parameters are used as input. However, the importance of these four parameters to battery capacity estimation is not exactly the same. If the parameter weights can be dynamically and autonomously allocated at different time steps, it will definitely be a good improvement for model performance. Therefore, the attention mechanism is introduced on the model input side to highlight the influence of important model parameters and improve model accuracy, and then the encoder is combined for quantization encoding. Its structure is as follows Figure 8 As shown in the figure, the processing process of the feature attention encoder can be seen. First, the model parameters and the historical moment state are used as input, and the attention weights are calculated through the attention layer equations (21) and (22), where e t,m represents the attention weight coefficient corresponding to feature m at time step t; v e ∈R T , W e ∈R T×2p is the weight coefficient that the neuron needs to learn, where p is the size of the last hidden layer of the encoder; B e ∈R n is the bias coefficient; Represents the attention weight of the k-dimensional feature at time t.

[0116]

[0117]

[0118] Then the weight is multiplied by the input to get the output of the encoder, as shown in formula (23). Its output is used as the input of the Bi-GRU layer, and the encoder output at the final time step t is expressed as formula (24):

[0119]

[0120]

[0121] Where: f BiGRU (·) represents the Bi-GRU network function. The feature attention mechanism dynamically distributes feature weights by autonomous learning, highlighting the correlation between the main model parameters and the battery capacity.

[0122] The aging of the battery is correlated with its previous state, and the degree of influence of the state at different historical moments varies. Therefore, when performing battery aging analysis and predicting battery capacity, it is necessary to consider the differentiated impact of battery historical information. When making predictions with the Bi-GRU model alone, the historical hidden state is considered, but the difference in the degree of influence of different historical moments on the current moment is not considered. Therefore, temporal attention is introduced on the decoding side to dynamically adjust the weight of the historical state, and assign higher weights to the historical states with stronger correlation. The specific structure is as follows: Fig. 9 shown.

[0123] The figure shows the processing process of the temporal attention mechanism decoder. First, the encoder output h τ (τ=1,2,...,T) is used as input, and the temporal attention weights at time step t are calculated and normalized through equations (25) and (26) of the attention layer.

[0124]

[0125]

[0126] Among them, β t,τ analogy The quantization represents the relevance of the historical information at the t-th time step to the battery capacity prediction. Then the weight is multiplied by the hidden state at each moment to obtain the weighted sequence as shown in formula (27):

[0127] c t =[β 1 h 1 ,β 2 h 2 ,...,β T h T ] T (27)

[0128] The output of the decoder at the final time step t is h t As shown in formula (28):

[0129] h t '=f BiGRU (h t ' -1 ,c t ) (28)

[0130] The feature attention mechanism decoder highlights the impact of important time series on the current capacity estimation by adaptively and dynamically distributing time series weights, thereby improving the model accuracy.

[0131] In summary, the DA-BiGRU structure diagram for battery capacity estimation is as follows: Fig.10 shown.

[0132] Step 4: Input the training dataset into the bidirectional gated recurrent unit model for training until the loss function converges, and then perform online battery capacity estimation together with the test dataset.

[0133] Verify the effectiveness of the method of the present invention. The overall process is as follows Fig.11 shown.

[0134] In addition, when training the model, in order to make the input more suitable for the training of the neural network, the Box-Cox transformation is performed to adjust the distribution of the features. The Box-Cox transformation can stabilize the parameter variance and make the parameter data closer to the normal distribution. It is a common method for data processing, which is particularly useful for reducing data skew and improving model performance.

[0135] Considering the accuracy and scientificity of model evaluation, mean absolute error (MAE) and root mean square error (RMSE) are selected as model evaluation indicators. The calculation formulas are shown in formula (29) and formula (30).

[0136]

[0137]

[0138] Among them, x i and Represent the actual value and predicted value of capacity respectively.

[0139] The core parameter settings of the overall model are as follows: the number of Bi-GRU units is set to 64; the activation function of the encoder and decoder uses the ReLU unit; the Adam optimizer is uniformly adopted; MAE is used as the loss function for compilation; the learning rate is set to 0.001; and the epoch is set to 200.

[0140] The proposed DA-BiGRU model was verified using eight batteries (labeled as cell1, cell2, cell3, cell4, cell5, cell6, cell7, and cell8) in the Oxford University public battery dataset Oxford_Battery_Degradation_Dataset_1.

[0141] The number of cycles of these batteries are 8200, 7700, 8100, 5100, 5000, 5000, 8100 and 8100 respectively. The purpose is to evaluate the accuracy of the proposed model on different battery samples, examine its adaptability to the differences in the data set and its performance under different sample numbers. The data of a single battery is divided into a training data set and a test data set, and the first 50% and the last 50% of the battery cycle are used as the training set and the test set respectively. This division is intended to ensure that the model can fully consider the overall cycle history of the battery during training and testing, and effectively evaluate its performance in predicting the battery health status.

[0142] The test data set consists of all the data after the training samples to ensure the rigor and comprehensiveness of the test. By performing such analysis on the data of these batteries, the performance of the DA-BiGRU model can be fully evaluated and its reliability in practical applications can be fully supported.

[0143] Fig.12 (a) to (h) are the prediction results of 8 cells respectively. In the figure, the orange curve represents the prediction result of the proposed model for the capacity value, and the blue curve represents the actual capacity value. By observing the data trend in the figure, it can be clearly found that the proposed model shows a good effect in fitting the capacity data, and shows a trend that is relatively close to the actual capacity value. This good fitting effect shows that the proposed model has high prediction accuracy and generalization ability, and has strong credibility and reliability in practical applications.

[0144] In the field of machine learning, the convergence of the model is regarded as one of the important indicators of its training process. Its good convergence indicates the efficiency and accuracy of the model in learning data patterns and parameter adjustment. To observe the convergence of the model, draw the Loss curve during model training as shown in Fig.13 .

[0145] It can be clearly observed from the figure that the error of the DA-BiGRU model decreases rapidly with the increase of the number of training times and tends to be stable after about 18 training times. This phenomenon fully proves that the proposed DA-BiLSTM model has good convergence ability.

[0146] In order to systematically evaluate the scientificity and accuracy of the model's battery capacity estimation, the mean absolute error (MAE) and root mean squared error (RMSE) were selected as the main evaluation indicators. MAE measures the average absolute deviation between the model's predicted value and the true value, while RMSE takes into account the variance of the prediction error. Therefore, these two indicators can comprehensively evaluate the prediction accuracy and robustness of the model. In order to further intuitively compare the performance of the models, their percentage form is used as the evaluation indicator to more clearly show the relative size of the prediction error in different situations. The evaluation indicators of the proposed model on 8 cells are shown in Table 4 below.

[0147] Table 4 Performance indicators of capacity estimation

[0148]

[0149] From the results, it can be observed that the DA-BiGRU model showed a relatively satisfactory performance for the capacity estimation of the 8 cells. Among them, the MAE% and RMSE% of cell2 and cell6 are slightly higher than those of the other 6 cells. Further observation of the actual capacity degradation curves of the two cells revealed the reason for this difference: in the case of cell2, its actual capacity degradation curve showed an abnormal decrease in the middle stage of the cycle, while in cell6, its actual capacity degradation curve showed an abnormal decrease at the end of the cycle. These abnormal phenomena may cause the model's prediction effect to be inferior to that of the other 6 cells. For the remaining 6 cells, their evaluation indicators are almost all below 0.3%, which shows that the model has excellent prediction effects on these batteries. Therefore, considering the situation of the 8 cells comprehensively, the DA-BiGRU model shows relatively good performance in capacity prediction.

[0150] In view of the shortcomings of traditional neural networks in battery capacity estimation, the present invention introduces a dual attention mechanism, proposes a Bi-GRU encoder-decoder model based on the dual attention mechanism, and uses the Oxford battery data set to verify the constructed battery evaluation method. The dual attention mechanism introduced in this model analyzes the correlation between battery model parameters, historical information and capacity from the two aspects of features and timing, and highlights the features and timing with high contribution by dynamically allocating weights, thereby improving the accuracy of the model. The entire algorithm takes into account both battery health characteristics and timing data, has low dependence on data volume, and has strong model reliability, high prediction accuracy, fast convergence speed, and good online application capabilities.

Claims

1. A battery capacity online estimation method based on model parameters, characterized in that: The following steps are involved: Step 1: Obtain the voltage and current data of the battery, and perform parameter identification and feature processing; Step 2: Perform correlation analysis on battery model parameters and their maximum available capacity to extract battery health factors; Step 3: Build a bidirectional gated recurrent unit model based on the dual attention mechanism, and divide the battery data into training data set and test data set; Step 4: Input the training dataset into the bidirectional gated recurrent unit model for training until the loss function converges, and then perform online battery capacity estimation together with the test dataset.

2. The method for online battery capacity estimation based on model parameters according to claim 1, characterized in that: In step 2, correlation analysis is performed using the Spearman coefficient and the maximum mutual information coefficient.

3. The method for online estimating battery capacity based on model parameters according to claim 1, characterized in that: The bidirectional gated recurrent unit model in step 3 includes a feature attention mechanism encoder and a temporal attention mechanism decoder.

4. The method for online estimating battery capacity based on model parameters according to claim 3, characterized in that: The formula for calculating the attention weight in the feature attention mechanism encoder is as follows: Among them, e t,m represents the attention weight coefficient corresponding to feature m at time step t; v e ∈R T , W e ∈R T×2p is the weight coefficient that the neuron needs to learn, where p is the size of the last hidden layer of the encoder; B e ∈R n is the bias coefficient.

5. The method for online battery capacity estimation based on model parameters according to claim 4, characterized in that: The bidirectional gated recurrent unit model is: e t =utanh(wh t +b a ) In the formula, h t (t∈[0,n]) is the hidden state corresponding to each input, α t (t∈[0,n]) is the attention weight, y t is the output value of Bi-GRU after the attention mechanism is introduced. u and w are weight coefficients. b a is the bias coefficient, e t is the attention weight value at time t, α t is the normalized weight probability distribution value.

Images