Lithium battery state-of-charge transfer learning estimation method based on data enhancement

Abstract

The invention discloses a lithium battery state-of-charge transfer learning estimation method based on data enhancement, which belongs to the technical field of battery detection, and comprises the following steps: S1, selecting a CALCE lithium battery as a source data set, and selecting novel lithium battery experimental data as a target data set; s2, performing data enhancement by combining a recursive least square method with a forgetting factor according to the equivalent circuit model and current and voltage in the data set; and S3, improving the long short-term memory network to enhance the data long short-term feature capture capability, designing network framework configuration parameters, and determining an optimal SOC estimation transfer learning scheme. According to the lithium battery state-of-charge transfer learning estimation method based on data enhancement, data enhancement is realized through an equivalent circuit model, and a network architecture is designed to capture a characteristic relation between measurable data and SOC; and the transfer learning method is applied to quickly adapt to lithium battery models with different working conditions and different chemical characteristics / models and realize accurate estimation of the SOC, so that manpower and time costs are saved.

Description

Technical Field

[0001] The present invention relates to the technical field of battery detection, and in particular to a lithium battery state of charge transfer learning estimation method based on data enhancement. Background Art

[0002] Lithium batteries have the advantages of small size, high energy density, and wide operating temperature range. They have been widely used in new energy electric vehicles, drones, mobile phones and other intelligent systems. The state of charge (SOC) is the percentage of the current remaining power of the battery, which is an important state of lithium batteries that cannot be directly measured. Therefore, accurately estimating the SOC of lithium-ion batteries is crucial to the safe operation of their power supply systems.

[0003] At present, the estimation methods of lithium battery SOC are mainly divided into model-based methods and data-driven methods. Model-based methods require the design / selection of appropriate models, including empirical models, electrochemical models, and equivalent circuit models. The table lookup method estimates the state of charge by recording the corresponding state of charge and battery voltage in a large number of experimental tests, but its error is often large. Electrochemical modeling describes the chemical process inside the battery through mathematical partial differential equations, but the parameter solution process is quite complicated. Equivalent circuit models are widely used in online state estimation due to their relatively simple circuit structure, but it is necessary to consider that their parameters are affected by operating conditions and aging. Therefore, they need to be integrated with other tools, such as Kalman filters, particle filters, etc.

[0004] Data-driven methods such as machine learning and deep learning methods use a large amount of existing experimental data to learn system dynamics through non-parametric methods, thereby achieving battery SOC estimation. However, this method has two main problems. On the one hand, the lithium battery experimental data used for data-driven method training is not completely consistent with the lithium battery operation data of the actual working conditions (charge and discharge current, temperature, health status, etc.); specifically, the battery data obtained in the experiment may be obtained under the conditions of charge and discharge current = 5A, temperature = 20℃, and health status = 20Ah, while the actual operating conditions may be that the charge and discharge current is dynamically changing, the temperature = -20℃, and the health status = 15Ah. On the other hand, with the continuous update and iteration of batteries, it takes a lot of time to obtain experimental data for batteries of all chemical characteristics / models.

[0005] Therefore, it is difficult to directly apply data-driven SOC estimation methods. A reasonable solution is to quickly adapt the existing data-driven model to learn the dynamic characteristics of the new battery, thereby saving manpower and time costs. Summary of the invention

[0006] The purpose of the present invention is to provide a lithium battery state of charge transfer learning estimation method based on data enhancement, which realizes data enhancement of existing battery measurable data through the method of equivalent circuit model, actively learns the hidden relationship between SOC and measurable data, adopts an improved dynamic jump long short-term memory network to design a network architecture to capture the long-term and short-term characteristic relationship between measurable data and SOC, and applies a transfer learning method to quickly adapt to different working conditions and different chemical properties / model lithium battery models and realize accurate estimation of SOC, thereby saving manpower and time costs.

[0007] To achieve the above object, the present invention provides a lithium battery state of charge transfer learning estimation method based on data enhancement, comprising the following steps:

[0008] S1, select CALCE lithium battery as the source data set, and select a new type of lithium battery experimental data as the target data set;

[0009] S2, based on the equivalent circuit model and the current and voltage in the data set, the data is enhanced by combining the recursive least squares method with a forgetting factor;

[0010] S3. Improve the long short-term memory network to enhance the ability to capture long-term and short-term features of data, design the network framework configuration parameters, and determine the optimal SOC estimation transfer learning scheme.

[0011] Preferably, the new lithium battery in S1 is a 40Ah lithium cobalt oxide battery, wherein the source data set and the target data set both include a training set and a test set, the training set of the source data set is the Beijing dynamic stress test BJDST and dynamic stress test DST conditions at 0°C, 25°C and 45°C, the test set is the data under the federal city driving test FUDS and high-speed driving test US06 conditions at the same temperature, and at each temperature there are two types of full discharge data with initial SOC of 80% and 50%; the training set of the target data set is collected by cyclic charge and discharge under constant current conditions of 15A charging and 20A discharging at -10°C, 0°C, 10°C, 20°C, 30°C and 40°C, and the test set is collected by cyclic charge and discharge at the same charge and discharge current at -20°C and 50°C.

[0012] Preferably, the real SOC value of the battery in S1 is calculated by the ampere-hour integration method, and the specific calculation formula is as follows:

[0013]

[0014] Where I represents the measured current, t represents the time, SOC0, SOC k ,η and Q N They represent the initial state SOC, SOC at time k, charge and discharge efficiency, and current capacity of the battery respectively.

[0015] Preferably, the equivalent circuit model in S2 is a first-order RC differential equivalent circuit model of the battery, and the first-order RC differential equivalent circuit model of the battery is as follows:

[0016]

[0017] Among them, R0,R p and C p is the equivalent ohmic resistance, polarization resistance and polarization capacitance of the lithium battery; U t ,U oc and U p are the terminal voltage, open circuit voltage and polarization voltage of the battery, I is the measured current of the battery, and the current direction of the battery discharge is defined as positive, and charging is defined as negative;

[0018] The above time-domain differential equivalent circuit model is discretized by Laplace transform, and its s-plane transfer function G(s) is as follows:

[0019]

[0020] Among them, U oc (s) represents the battery open circuit voltage U oc The Laplace transform of U t (s) represents the battery terminal voltage U t The Laplace transform of I(s) represents the Laplace transform of the battery measurement current I;

[0021] By bilinear transformation The transfer function of the above s-plane is mapped to the z-plane as follows:

[0022]

[0023] Among them, T s is a discrete sampling interval, c1, c2 and c3 are equivalent circuit model parameters and sampling interval T s The relevant constant coefficient is expressed as follows:

[0024]

[0025] Therefore, the discrete form of the first-order RC equivalent circuit model is:

[0026]

[0027] Among them, U t,k ,U oc,k and I k They represent the terminal voltage, open circuit voltage and measured current at time k respectively; further, they are written in the least squares form:

[0028]

[0029] in, is the estimated terminal voltage at time k, Φ k is the measurement vector, θ k =[θ1θ2θ3θ4] is the parameter vector;

[0030] The recursive equation of the FFRLS algorithm is set as follows, where K k ,P k are the gain matrix and covariance matrix at time k, respectively, λ is the adjustable forgetting factor, y k Actual measurement terminal voltage U t,k :

[0031]

[0032] Finally, the FFRLS algorithm is used to identify θ k , and further the open-loop voltage U at time k oc,k , polarization voltage U p,k and the voltage U across the ohmic resistor 0,k =I k R 0,k as follows:

[0033]

[0034] At this time, the data at time k is composed of measurable data (I k ,T k ,U t,k ) is enhanced to Represents the open circuit voltage, polarization voltage, and ohmic resistance response estimated at time k, thereby actively learning the hidden relationship between SOC and measurable data through the FFRLS algorithm as the input of the S3 deep learning framework to reduce the mapping complexity of deep learning.

[0035] Preferably, S3 specifically includes the following steps:

[0036] S3.1. Combined with the data enhanced in S2, a network architecture for deep learning of SOC estimation is designed. The architecture includes an input layer, two BiLSTM layers, four fully connected layers and an output layer, wherein the BiLSTM layer is designed with a residual connection structure.

[0037] S3.2, determine the hyperparameters involved in the deep learning network architecture through a search algorithm based on random search;

[0038] S3.3. Apply transfer learning to quickly adapt existing data-driven models to learn the dynamic characteristics of new batteries;

[0039] S3.4. By selecting different value combinations, different transfer learning schemes are formed.

[0040] Preferably, the input layer in S3.1 mainly completes the normalization processing of the data, wherein the normalization processing refers to mapping all input data into the range of [0,1], and the specific method is as follows:

[0041]

[0042] Among them, x, x', x max and x min Represent the original data, normalized data, the maximum value in the original data, and the minimum value in the original data, respectively.

[0043] The BiLSTM layer consists of two sequential BiLSTM networks, and each BiLSTM network is composed of two LSTM networks in opposite directions. It captures the long-term and short-term relationship of data by considering past and future data. The calculation formula is as follows:

[0044]

[0045] in, Respectively represent the forget gate, input gate, output gate, candidate cell state, cell state and hidden state of the LSTM network at time t; W and b are the corresponding weight matrix and bias vector; σ and tanh are the sigmoid function and hyperbolic tangent function. In addition, for the residual connection, assume that the input of the i-th layer BiLSTM network is J i , the output is H i , then the residual connection output of this layer is [J i H i ];

[0046] The fully connected layer consists of four dense layers. Assume that the input of each layer is F xi , the output is F yi , then the calculation formula is F yi =act(W Fi F xi +b Fi ), where W Fi ,b Fi is the weight matrix and bias vector corresponding to the i-th layer; act is the gelu activation function; the output layer is essentially composed of a dense layer, but its output represents the estimated SOC.

[0047] Therefore, the present invention adopts the above-mentioned lithium battery state of charge transfer learning estimation method based on data enhancement, realizes data enhancement of existing battery measurable data through the method of equivalent circuit model, actively learns the hidden relationship between SOC and measurable data, and adopts an improved dynamic jump long short-term memory network to design a network architecture to capture the long-term and short-term characteristic relationship between measurable data and SOC. The transfer learning method is applied to quickly adapt to different working conditions and different chemical properties / model lithium battery models and realize accurate estimation of SOC, thereby saving manpower and time costs.

[0048] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a flow chart of an embodiment of a lithium battery state of charge transfer learning estimation method based on data enhancement of the present invention;

[0050] Figure 2 It is a diagram of SOC estimation effects on a target data set under different migration schemes of an embodiment of a lithium battery state of charge transfer learning estimation method based on data enhancement of the present invention. DETAILED DESCRIPTION

[0051] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0052] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.

[0053] Embodiment 1

[0054] like Figure 1 As shown, the present invention provides a lithium battery state of charge transfer learning estimation method based on data enhancement, comprising the following steps:

[0055] S1, select CALCE lithium battery as the source data set, and select a new type of lithium battery experimental data as the target data set;

[0056] The new lithium battery is a 40Ah lithium cobalt oxide battery, of which the source data set D S and the target dataset D TBoth include training sets and test sets. The training set of the source data set is the Beijing dynamic stress test BJDST and dynamic stress test DST conditions at 0℃, 25℃ and 45℃, and the test set is the data under the federal city driving test FUDS and high-speed driving test US06 conditions at the same temperature. At each temperature, there are two types of full discharge data with initial SOC of 80% and 50%; the training set of the target data set is collected under cyclic charge and discharge conditions of 15A charging and 20A discharging at constant current conditions at -10℃, 0℃, 10℃, 20℃, 30℃ and 40℃, and the test set is collected under cyclic charge and discharge conditions with the same charge and discharge current at -20℃ and 50℃.

[0057] The data collected from the above two data sets include the measured current I, terminal voltage U t and temperature T e , the real SOC value of the battery is calculated by the ampere-hour integration method. The specific calculation formula is as follows:

[0058]

[0059] Where I represents the measured current, t represents the time, SOC0, SOC k ,η and Q N They represent the initial state SOC, SOC at time k, charge and discharge efficiency, and current capacity of the battery, respectively. The current and terminal voltage I at time k used for S2 data enhancement can be obtained. k and U t,k .

[0060] S2, based on the equivalent circuit model and the current and voltage in the data set, the data is enhanced by combining the recursive least squares method with a forgetting factor;

[0061] First, the equivalent circuit model is the first-order RC differential equivalent circuit model of the battery. The first-order RC differential equivalent circuit model of the battery is as follows:

[0062]

[0063] Among them, R0,R p and C p is the equivalent ohmic resistance, polarization resistance and polarization capacitance of the lithium battery; U t ,U oc and U p are the terminal voltage, open circuit voltage and polarization voltage of the battery, I is the measured current of the battery, and the current direction of the battery is specified as positive for discharge and negative for charge.

[0064] The above time-domain differential equivalent circuit model is discretized by Laplace transform, and its s-plane transfer function G(s) is as follows:

[0065]

[0066] Among them, U oc (s) represents the battery open circuit voltage U oc The Laplace transform of U t (s) represents the battery terminal voltage U t The Laplace transform of , I(s) represents the Laplace transform of the battery measurement current I.

[0067] By bilinear transformation The transfer function of the above s-plane is mapped to the z-plane as follows:

[0068]

[0069] Among them, T s is a discrete sampling interval, c1, c2 and c3 are equivalent circuit model parameters and sampling interval T s The relevant constant coefficient is expressed as follows:

[0070]

[0071] Therefore, the discrete form of the first-order RC equivalent circuit model is:

[0072]

[0073] Among them, U t,k ,U oc,k and I k They represent the terminal voltage, open circuit voltage and measured current at time k respectively; further, they are written in the least squares form:

[0074]

[0075] in, is the estimated terminal voltage at time k, Φ k is the measurement vector, θ k =[θ1θ2θ3θ4] is the parameter vector;

[0076] The recursive equation of the FFRLS algorithm is set as follows, where K k ,P k are the gain matrix and covariance matrix at time k, respectively, λ is the adjustable forgetting factor, y k Actual measurement terminal voltage U t,k :

[0077]

[0078] Finally, the FFRLS algorithm is used to identify θ k , and further the open-loop voltage U at time k oc,k , polarization voltage Up,k and the voltage U across the ohmic resistor 0,k =I k R 0,k as follows:

[0079]

[0080] At this time, the data at time k is composed of measurable data (I k ,T k ,U t,k ) is enhanced to Represents the open circuit voltage, polarization voltage, and ohmic resistance response estimated at time k. The hidden relationship between SOC and measurable data is actively learned through the FFRLS algorithm as the input of the S3 deep learning framework, reducing the mapping complexity of deep learning.

[0081] S3. Improve the long short-term memory network to enhance the data long-term and short-term feature capture capability, design the network framework configuration parameters, and determine the optimal SOC estimation transfer learning scheme;

[0082] The specific steps include:

[0083] S3.1. Combined with the enhanced data in S2, a deep learning network architecture for SOC estimation was designed. The architecture includes an input layer, two BiLSTM layers, four fully connected layers and an output layer, among which the BiLSTM layer is designed with a residual connection structure. The designed deep learning framework was trained on the source dataset DS, and the network configuration parameters and hyperparameters were determined by adjusting the hyperparameters and comparing the training effects. The configuration parameters include the filter size, number of filters, weight parameters and bias parameters of each layer. The hyperparameters include batch size, time step, number of features, iteration rounds, activation function and learning rate.

[0084] The input layer mainly completes the normalization of data, where normalization refers to mapping all input data to the range of [0,1]. The specific method is as follows:

[0085]

[0086] Among them, x, x', x max and x min Represent the original data, normalized data, the maximum value in the original data, and the minimum value in the original data, respectively.

[0087] The BiLSTM layer consists of two sequential BiLSTM networks, and each BiLSTM network is composed of two LSTM networks in opposite directions. It captures the long-term and short-term relationship of data by considering past and future data. The calculation formula is as follows:

[0088]

[0089] in, Respectively represent the forget gate, input gate, output gate, candidate cell state, cell state and hidden state of the LSTM network at time t; W and b are the corresponding weight matrix and bias vector; σ and tanh are the sigmoid function and hyperbolic tangent function. In addition, for the residual connection, assume that the input of the i-th layer BiLSTM network is J i , the output is H i , then the residual connection output of this layer is [J i H i ];

[0090] The fully connected layer consists of four dense layers. Assume that the input of each layer is F xi , the output is F yi , then the calculation formula is F yi =act(W Fi F xi +b Fi ), where W Fi ,b Fi is the weight matrix and bias vector corresponding to the i-th layer; act is the gelu activation function; the output layer is essentially composed of a dense layer, but its output represents the estimated SOC. Finally, the parameters of the designed deep learning framework for SOC estimation are as follows:

[0091] Table 1 Configuration parameters of deep learning framework

[0092]

[0093] S3.2, determine the hyperparameters involved in the deep learning network architecture through a search algorithm based on random search;

[0094] The hyperparameters are shown in Table 2:

[0095] Table 2 Hyperparameters of deep learning frameworks

[0096] parameter Target dataset Batch size 128 Time step 30 Number of features 6 Iteration round 1000 Activation Function gelu Learning Rate 0.0001

[0097] S3.3. Apply transfer learning to quickly adapt existing data-driven models to learn the dynamic characteristics of new batteries;

[0098] Assume that after the deep learning model designed above is trained on the source data set, the parameter sets of each layer are input layer parameters S1, two-layer BiLSTM layer parameters B2 and B3, four-layer fully connected layer parameters F1, F2, F3 and F4, and output layer parameters O1. The specific implementation scheme of transfer learning is to initialize the model parameters with the training model parameters on the source data set, freeze the parameters of some layers and do not train them on the target data set, and only train the parameters of the remaining layers. Therefore, the i-th type of transfer learning applied can be described as the following formula:

[0099] TL i =α i,1 S1+α i,2 B1+α i,3 B2+α i,4 F1+α i,5 F2+α i,6 F3+α i,7 F4+α i,8 O1

[0100] Among them, α i,j (j=1,2,...,8) has only two values: 0 and 1. i,j =1, it means that in the i-th migration scheme, the network layer participates in the training of the target data set, that is, the training can still be changed after the parameters are initialized; on the contrary, when α i,j =0, it means that the network parameters of this layer remain consistent with the model parameters trained on the source data set and do not participate in the training of the target data set, that is, training changes cannot be made after the parameters are initialized.

[0101] S3.4, by selecting different α i,j The value combinations of (j=1,2,...,8) constitute different transfer learning schemes;

[0102] Select the corresponding migration scheme and train it on the target data set to compare the migration effect. Select the root mean square error (RMSE) as the criterion for selecting the optimal migration scheme. The expression of RMSE is as follows:

[0103]

[0104] Among them, SOC k is the true value, SOC' k is the estimated value, and N is the total number of samples for SOC estimation.

[0105] The simulation results of the embodiment of the present invention are as follows: Figure 2 As shown in Table 3, Figure 2The SOC estimation effect on the target data set under different migration schemes is shown. It can be seen that under the data enhancement and migration learning method proposed in the present invention, the battery SOC estimation problem under different characteristics, different temperatures and different working conditions is solved, which can effectively save a lot of experimental costs and time costs. As shown in Table 3 below, the SOC mean square estimation error (RMSE) of different migration schemes is specifically counted. Among them, Scheme 9 summarizes that the RMSE of the first layer BiLSTM and all layers except the last dense layer is the smallest, and can be used as the optimal migration learning scheme.

[0106] Table 3. Mean square error of SOC estimation of different migration schemes on the target dataset

[0107]

[0108] Therefore, the present invention adopts the above-mentioned lithium battery state of charge transfer learning estimation method based on data enhancement, realizes data enhancement of existing battery measurable data through the method of equivalent circuit model, actively learns the hidden relationship between SOC and measurable data, and adopts an improved dynamic jump long short-term memory network to design a network architecture to capture the long-term and short-term characteristic relationship between measurable data and SOC. The transfer learning method is applied to quickly adapt to different working conditions and different chemical properties / model lithium battery models and realize accurate estimation of SOC, thereby saving manpower and time costs.

[0109] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A lithium battery state of charge transfer learning estimation method based on data enhancement, characterized by: The following steps are involved: S1, select CALCE lithium battery as the source data set, and select a new type of lithium battery experimental data as the target data set; S2, based on the equivalent circuit model and the current and voltage in the data set, the data is enhanced by combining the recursive least squares method with a forgetting factor; S3. Improve the long short-term memory network to enhance the ability to capture long-term and short-term features of data, design the network framework configuration parameters, and determine the optimal SOC estimation transfer learning scheme.

2. The method for estimating the state of charge of a lithium battery based on data enhancement transfer learning according to claim 1, characterized in that: The new lithium battery in S1 is a 40Ah lithium cobalt oxide battery, wherein the source data set and the target data set both include a training set and a test set, the training set of the source data set is the Beijing dynamic stress test BJDST and dynamic stress test DST conditions at 0°C, 25°C and 45°C, the test set is the data under the federal city driving test FUDS and high-speed driving test US06 conditions at the same temperature, and at each temperature there are two types of full discharge data with initial SOC of 80% and 50%; the training set of the target data set is collected by cyclic charge and discharge under constant current conditions of 15A charging and 20A discharging at -10°C, 0°C, 10°C, 20°C, 30°C and 40°C, and the test set is collected by cyclic charge and discharge at the same charge and discharge current at -20°C and 50°C.

3. The method for estimating the state of charge of a lithium battery based on data enhancement transfer learning according to claim 1, characterized in that: The actual SOC value of the battery in S1 is calculated by the ampere-hour integration method, and the specific calculation formula is as follows: Where I represents the measured current, t represents the time, SOC0, SOC k ,η and Q N They represent the initial state SOC, SOC at time k, charge and discharge efficiency, and current capacity of the battery respectively.

4. The method for estimating the state of charge of a lithium battery based on data enhancement transfer learning according to claim 1, characterized in that: The equivalent circuit model in S2 is a first-order RC differential equivalent circuit model of the battery. The first-order RC differential equivalent circuit model of the battery is as follows: Among them, R0,R p and C p is the equivalent ohmic resistance, polarization resistance and polarization capacitance of the lithium battery; U t ,U oc and U p are the terminal voltage, open circuit voltage and polarization voltage of the battery, I is the measured current of the battery, and the current direction of the battery discharge is defined as positive, and charging is defined as negative; The above time-domain differential equivalent circuit model is discretized by Laplace transform, and its s-plane transfer function G(s) is as follows: Among them, U oc (s) represents the battery open circuit voltage U oc The Laplace transform of U t (s) represents the battery terminal voltage U t The Laplace transform of I(s) represents the Laplace transform of the battery measurement current I; By bilinear transformation The transfer function of the above s-plane is mapped to the z-plane as follows: Among them, T s is a discrete sampling interval, c1, c2 and c3 are equivalent circuit model parameters and sampling interval T s The relevant constant coefficient is expressed as follows: Therefore, the discrete form of the first-order RC equivalent circuit model is: Among them, U t,k ,U oc,k and I k They represent the terminal voltage, open circuit voltage and measured current at time k respectively; further, they are written in the least squares form: in, is the estimated terminal voltage at time k, Φ k is the measurement vector, θ k =[θ1θ2θ3θ4] is the parameter vector; The recursive equation of the FFRLS algorithm is set as follows, where K k ,P k are the gain matrix and covariance matrix at time k, respectively, λ is the adjustable forgetting factor, y k Actual measurement terminal voltage U t,k : Finally, the FFRLS algorithm is used to identify θ k , further the open-loop voltage U at time k oc,k , polarization voltage U p,k and the voltage U across the ohmic resistor 0,k =I k R 0,k as follows: At this time, the data at time k is composed of measurable data (I k ,T k ,U t,k ) is enhanced to Represents the open circuit voltage, polarization voltage, and ohmic resistance response estimated at time k, thereby actively learning the hidden relationship between SOC and measurable data through the FFRLS algorithm as the input of the S3 deep learning framework to reduce the mapping complexity of deep learning.

5. The method for estimating the state of charge of a lithium battery based on data enhancement transfer learning according to claim 1, characterized in that: The S3 specifically includes the following steps: S3.

1. Combined with the data enhanced in S2, a network architecture for deep learning of SOC estimation is designed. The architecture includes an input layer, two BiLSTM layers, four fully connected layers and an output layer, wherein the BiLSTM layer is designed with a residual connection structure. S3.2, determine the hyperparameters involved in the deep learning network architecture through a search algorithm based on random search; S3.

3. Apply transfer learning to quickly adapt existing data-driven models to learn the dynamic characteristics of new batteries; S3.

4. By selecting different value combinations, different transfer learning schemes are formed.

6. The method for estimating the state of charge of a lithium battery based on data enhancement transfer learning according to claim 5, characterized in that: The input layer in S3.1 mainly completes the normalization of data, where normalization refers to mapping all input data to the range of [0,1]. The specific method is as follows: Among them, x, x', x max and x min Represent the original data, normalized data, the maximum value in the original data, and the minimum value in the original data, respectively. The BiLSTM layer consists of two sequential BiLSTM networks, and each BiLSTM network is composed of two LSTM networks in opposite directions. It captures the long-term and short-term relationship of data by considering past and future data. The calculation formula is as follows: Among them, f t ,i t ,o t , C t ,h t Respectively represent the forget gate, input gate, output gate, candidate cell state, cell state and hidden state of the LSTM network at time t; W and b are the corresponding weight matrix and bias vector; σ and tanh are the sigmoid function and hyperbolic tangent function. In addition, for the residual connection, assume that the input of the i-th layer BiLSTM network is J i , the output is H i , then the residual connection output of this layer is [J i H i ]; The fully connected layer consists of four dense layers. Assume that the input of each layer is F xi , the output is F yi , then the calculation formula is F yi =act(W Fi F xi +b Fi ), where W Fi ,b Fi is the weight matrix and bias vector corresponding to the i-th layer; act is the gelu activation function; the output layer is essentially composed of a dense layer, but its output represents the estimated SOC.

Images