Lithium ion battery SOC estimation method based on TCN-GRU neural network model and transfer learning

Abstract

The invention discloses a lithium ion battery SOC estimation method based on a TCN-GRU neural network model and transfer learning, which can effectively shorten the battery test period and reduce the BMS cost, especially when the training data of a target battery is insufficient and in the early stage of BMS development. The method comprises the following steps: acquiring measurement data of a source battery under different discharge conditions and environmental conditions as a source domain data set; for battery behaviors in the source domain, a TCN-GRU framework is adopted to take the current, SOC and temperature of the battery as model input and the voltage of the battery as model output to construct a CN-GRU neural network model, and the data in the source domain data set is used to pre-train the constructed TCN-GRU neural network model; aiming at a battery behavior of a target domain, expanding the TCN-GRU neural network model by adopting a transfer learning technology, and finely adjusting data from the target domain to adapt to specific characteristics of a target battery, so as to obtain an expanded TCN-GRU neural network model; and constructing a square root cubature Kalman filter (SRCKF) algorithm to estimate the SOC.

Description

Technical Field

[0001] The embodiments of the present application relate to the technical field of battery state of charge prediction based on a specific computing model, and in particular to a lithium-ion battery SOC estimation method based on a TCN-GRU neural network model and transfer learning, and a device thereof. Background Art

[0002] As the mainstream energy storage unit, lithium-ion batteries have been widely used in electric vehicles. In order to fully utilize the performance of batteries and ensure the safe use of batteries, it is particularly important to equip them with a battery management system (BMS) to manage the batteries during the battery charging and discharging process.

[0003] Battery state of charge (SOC) estimation is an important reference for determining the remaining mileage of electric vehicles. SOC is an important indicator for BMS to monitor the battery status and is also a key task for battery management. However, SOC is an implicit state that cannot be directly measured by sensors and can only be obtained by processing, analyzing and calculating the measured data.

[0004] SOC estimation methods based on specific computer models have attracted much attention in academia and industry, but SOC estimation requires a large number of battery tests, which greatly prolongs the development cycle of BMS and increases development costs. Changes in battery materials or versions require a large number of repeated tests, resulting in a significant increase in subsequent maintenance costs. Therefore, reducing reliance on a large number of offline tests or minimizing the amount of data required for model-based SOC estimation remains a key area for further research. Summary of the invention

[0005] In order to overcome the above technical defects, the present application embodiment provides a lithium-ion battery SOC estimation method based on TCN-GRU neural network model and transfer learning. The specific scheme is as follows:

[0006] S1. Data acquisition:

[0007] Obtaining measurement data of the source battery under different discharge conditions and environmental conditions as a source domain data set, the measurement data includes the current, SOC, temperature and voltage of the battery;

[0008] S2. TCN-GRU neural network model construction and pre-training:

[0009] According to the battery behavior in the source domain, the TCN-GRU framework is used to take the battery current, SOC and temperature as the model input, and the battery voltage as the model output to build the TCN-GRU neural network model, and the data in the source domain dataset is used to pre-train the constructed TCN-GRU neural network model;

[0010] S3, TCN-GRU neural network model expansion:

[0011] Aiming at the battery behavior in the target domain, the TCN-GRU neural network model is extended by using transfer learning technology, and the data from the target domain is fine-tuned to adapt to the specific characteristics of the target battery, thus obtaining the extended TCN-GRU neural network model.

[0012] S4. Construct a square root cubic Kalman filter (SRCKF) algorithm to estimate SOC.

[0013] Optionally, in a possible implementation, the source domain dataset includes two datasets, wherein the first dataset is test data obtained from laboratory tests, and the second dataset is test data from the Center for Advanced Lifecycle Engineering (CALCE) of the University of Maryland.

[0014] Optionally, in one possible implementation, the constructed TCN-GRU neural network model is defined as:

[0015] Optionally, in one possible implementation, through fine-tuning, the extended TCN-GRU neural network model is defined as:

[0016] In the formula, θ src Represents the pre-training parameters.

[0017] Optionally, in a possible implementation, a combined loss function including source domain loss and target domain loss is used to calculate the loss of the extended TCN-GRU neural network model. The combined loss function is: L = L src +λL tgt ; where λ is a hyperparameter that balances the loss contribution of the source domain and the target domain, L src and L tgt Denote the loss of the source domain and the target domain respectively, and are expressed as:

[0018]

[0019] Among them, N src and N tgt Represent the number of samples in the source dataset and the target dataset respectively.

[0020] The embodiments of the present application adopt the above technical solution to achieve the following technical effects:

[0021] TCN-GRU can effectively track voltage changes, while SRCKF can effectively correct SOC using battery model errors. This approach can effectively shorten the battery test cycle and reduce BMS costs, especially when the training data of the target battery is insufficient and in the early stages of BMS development. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The accompanying drawings exemplarily illustrate the embodiments and constitute a part of the specification, and together with the text description of the specification, are used to explain the exemplary implementation of the embodiments. The embodiments shown are for illustrative purposes only and do not limit the scope of the claims. In all drawings, the same reference numerals refer to similar but not necessarily identical elements.

[0023] Figure 1 A flow chart of a lithium-ion battery SOC estimation method based on a TCN-GRU neural network model and transfer learning in this application;

[0024] Figure 2 are the operating current and voltage of battery 1;

[0025] Figure 3 is the operating current and voltage of battery 2;

[0026] Figure 4 is the operating current and voltage of battery 3;

[0027] Figure 5 is the operating temperature of the battery 3;

[0028] Figure 6 Workflow diagram for estimating SOC;

[0029] Figure 7 SOC estimation curves under different algorithms;

[0030] Figure 8 SOC error curves under different algorithms;

[0031] Fig. 9 This is the SOC estimation result at 55°C;

[0032] Fig.10 It is the SOC estimation result under dynamic temperature;

[0033] Fig.11 The SOC estimation results of different machine learning methods at 0℃;

[0034] Fig.12 The SOC estimation results of different machine learning methods at 50°C;

[0035] Fig.13 SOC estimation error of different machine learning methods at 0℃;

[0036] Fig.14 The SOC estimation error of different machine learning methods at 50°C;

[0037] Fig.15 The SOC statistical error of different machine learning methods at 0°C;

[0038] Fig.16 SOC statistical error of different machine learning methods at 50℃. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solutions and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application.

[0040] It should be noted that the descriptions involving "first", "second", etc. in the embodiments of the present application are only for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In addition, the technical solutions between the various embodiments can be combined with each other, but they must be based on the ability of ordinary technicians in the field to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such combination of technical solutions does not exist and is not within the scope of protection required by this application.

[0041] In the description of the present application, it should be understood that the numerical labels before the steps do not indicate the order in which the steps are executed, but are only used to facilitate the description of the present application and to distinguish each step, and therefore should not be understood as a limitation on the present application.

[0042] First, the following terms are explained:

[0043] CC-CV charging strategy: Constant Current and Constant Voltage Discharge (CC-CV Discharge) is a combination of CC (constant current) discharge and CV (constant voltage) discharge. First, discharge with a constant current to the termination voltage, keep the voltage unchanged, and then gradually reduce the current to the set value, and the discharge ends;

[0044] DST: Daylight Saving Time is a time system used to save energy and adapt to seasonal changes. Usually, in spring, people adjust the clock forward one hour to enter the daylight saving time period; in autumn, the clock is adjusted back one hour to return to standard time. This is implemented in different regions around the world, such as the United States in North America and some countries in Europe. ;

[0045] FUDS: (Federal Urban Diesel Emissions Standard), this is a regulation formulated by the U.S. federal government for diesel vehicle emission standards in urban areas. It stipulates the emission limits of particulate matter (PM) and nitrogen oxides (NOx) allowed to be generated by diesel engines under certain conditions, with the aim of reducing air pollution;

[0046] US06: (USE Environmental Protection Agency (EPA) Test Procedure US06) This is a dynamic driving cycle test procedure developed by the US Environmental Protection Agency to evaluate the fuel economy and pollutant emission performance of light vehicles (including sedans, small multi-purpose vehicles, etc.). US06 simulates typical urban traffic conditions, including acceleration, climbing, stopping and other driving conditions, and is used to evaluate the performance of vehicles on actual roads;

[0047] Transfer learning is a method in machine learning that allows a model to apply knowledge learned from one task to another related task.

[0048] Kalman filtering is an algorithm that uses the linear system state equation to optimally estimate the system state through system input and output observation data.

[0049] Secondly, in order to facilitate those skilled in the art to understand the technical solutions provided in the embodiments of the present application, the relevant technologies are described below:

[0050] See also Figure 1 In the embodiment of the present application, the lithium-ion battery SOC estimation method based on the TCN-GRU neural network model and transfer learning includes:

[0051] S1. Data acquisition:

[0052] The measurement data of the source battery under different discharge conditions and environmental conditions are obtained as the source domain data set, and the measurement data includes the current, SOC, temperature and voltage of the battery.

[0053] Specifically, different discharge conditions refer to different operating conditions in the discharge mode, such as DST, FUDS and US06. The environmental conditions are mainly temperature, such as 0° C., 25° C. and 50° C. respectively.

[0054] The source domain dataset includes two datasets, the first of which is test data obtained from laboratory tests, and the second is test data from the Center for Advanced Lifecycle Engineering (CALCE) at the University of Maryland.

[0055] Specifically, the first data set includes an NCM battery with a nominal capacity of 4Ah and specifications as shown in Table 1. During the test, the battery was initially charged at a constant current (CC) and constant voltage (CV) of 0.5C until the current dropped to a predetermined cutoff value. After a 2-hour rest, the battery was repeatedly loaded with one of the dynamic current curves at different operating temperatures until the terminal voltage dropped to 2.75V, corresponding to a SOC of 0%. The dynamic current curves include the Federal Urban Driving Schedule (FUDS), the Urban Dynamometer Driving Schedule (UDDS), and the Dynamic Stress Test (DST), which are used to simulate 6 of the 22 actual operating conditions of road electric vehicle lithium-ion batteries. The specific test steps are shown in Table 2.

[0056] The second data set is described for LFP batteries with a nominal capacity of 1.1Ah, and the specifications are shown in Table 1. The battery is fully charged with a cut-off voltage and current of 3.65V and 0.01C using a constant current constant voltage (CC-CV) charging strategy at 1℃, and then discharged after idling for 2 minutes until the terminal voltage drops to the discharge cut-off voltage of 2V. The discharge mode includes three different operating conditions: DST, FUDS, and US06, and the operating temperature is controlled at 0, 25, and 50℃, respectively.

[0057] Table 1

[0058]

[0059] Table 2

[0060]

[0061] S2. TCN-GRU neural network model construction and pre-training:

[0062] Aiming at the battery behavior in the source domain, the TCN-GRU framework is used to take the battery current, SOC and temperature as the model input, and the battery voltage as the model output to construct the TCN-GRU neural network model, and the data in the source domain dataset is used to pre-train the constructed TCN-GRU neural network model.

[0063] The TCN-GRU neural network model has the following advantages compared with other models:

[0064] 1. Long sequence processing capability

[0065] By using dilated convolution, TCN can expand the receptive field without adding too many parameters, so that it can effectively capture long-range dependencies in long sequence data. GRU itself also has the ability to process long sequences. Its internal gating mechanism can selectively forget or retain historical information, and effectively store and utilize important information in long sequences. Compared with traditional recurrent neural networks (RNNs), TCN-GRU overcomes the gradient vanishing and gradient exploding problems that RNNs are prone to when processing long sequences. Because the convolution structure of TCN and the gating mechanism of GRU make information transmission more stable. Compared with simple convolutional neural networks (CNNs), TCN-GRU performs better on long sequences. Since ordinary CNNs have a fixed convolution kernel size, capturing long-range dependencies requires multiple layers of stacking, which increases the model complexity and information loss, while TCN-GRU can more flexibly handle dependencies of various time scales in long sequences.

[0066] 2. Parallel computing capabilities

[0067] The convolution operation of TCN is parallel, and multiple time steps can be calculated simultaneously when processing data. Although GRU is a sequence model, in the overall architecture of TCN-GRU, the parallel computing characteristics of TCN can reduce the time overhead caused by sequence calculation. Compared with traditional RNN and some models based on recursive structures, TCN-GRU does not need to process each time step sequentially, which greatly improves the computing efficiency. Especially in scenarios with large amounts of data and long sequences, the advantages of this parallel computing capability are more obvious, which can significantly shorten the training time of the model.

[0068] 3. Feature extraction capability

[0069] By stacking multiple convolutional layers, TCN can gradually extract features at different levels from the original data. When processing the feature sequence extracted by TCN, GRU further screens and integrates these features. Compared with the simple feature engineering method combined with the machine learning model (such as manually extracting features and then using models such as support vector machines), TCN-GRU can automatically learn the feature representation in the data, reducing the workload of manual feature engineering, and can mine more complex and abstract features. Compared with some simple deep learning models, TCN-GRU's multi-level feature extraction capability makes its representation of data richer and more accurate, thereby achieving better results in tasks such as classification and prediction.

[0070] 4. Model interpretability and stability

[0071] The structure of TCN-GRU is relatively clear, which makes it relatively easy to analyze and explain the model. During the training process, the convolution structure of TCN and the gating mechanism of GRU stabilize the input data, making the model less susceptible to excessive interference from abnormal data such as noise. Compared with some neural networks with unstable structures (such as some deep neural networks without regularization), TCN-GRU has better stability and relatively stronger generalization ability when facing input data with different distributions.

[0072] Specifically, the constructed TCN-GRU neural network model is defined as:

[0073]

[0074] In the formula, It represents the prediction result of the model in the source domain, TCN represents the time domain convolutional network, GRU represents the gated recurrent unit, I, V, T, S represent the variables input into the model, which are current, voltage, temperature and SOC value in historical data respectively.

[0075] After the construction is completed, the data in the source domain dataset, such as the first dataset and the second dataset mentioned above, is used to pre-train the constructed TCN-GRU neural network model to complete the basic model construction.

[0076] S3, TCN-GRU neural network model expansion:

[0077] Aiming at the battery behavior in the target domain, the transfer learning technology is used to expand the TCN-GRU neural network model, and the data from the target domain is fine-tuned to adapt to the specific characteristics of the target battery to obtain the expanded TCN-GRU neural network model.

[0078] The model extensions do the following:

[0079] 1. Performance improvement: By adding more layers or adjusting the network structure, the expressiveness of the model can be improved, which helps to capture more complex sequence patterns and thus improve prediction accuracy.

[0080] 2. Enhanced generalization ability: Increasing model capacity can help the model better adapt to the training data, reduce the risk of overfitting, and improve the generalization effect on unseen data.

[0081] 3. Feature fusion: You can try convolution kernels of different sizes or introduce an attention mechanism so that the model can focus on local and global information at the same time and improve the overall understanding of the input sequence.

[0082] 4. Modular design: The expanded model is usually easier to understand and debug because each new part has a clear function, which is easy to study and optimize.

[0083] 5. Application expansion: For problems in specific fields, such as adding positional encoding or other field-specific knowledge, the model can perform better in specific scenarios.

[0084] Specifically, the target domain may be various scenarios or data sets. Accordingly, when the target domain is data for predicting a certain time period in the future, the specific characteristic of the target battery is SOC.

[0085] After fine-tuning, the expanded TCN-GRU neural network model is defined as:

[0086]

[0087] In the formula, θ src Represents the pre-training parameters. Pre-training parameters are gated weights, including the weight matrices of the update gate and the reset gate. These parameters are used to control the updating and forgetting of information.

[0088] Furthermore, in order to ensure the effective transfer of knowledge in the TCN-GRU neural network model before and after expansion, this application defines a combined loss function including source domain loss and target domain loss to calculate the loss of the expanded TCN-GRU neural network model. The combined loss function is:

[0089] L=L src +λL tgt ; where λ is a hyperparameter that balances the loss contribution of the source domain and the target domain, L src and L tgt Denote the loss of the source domain and the target domain respectively, and are expressed as:

[0090]

[0091] Among them, N src and N tgt Represent the number of samples in the source dataset and the target dataset respectively.

[0092] S4. Construct the square root volumetric Kalman filter SRCKF algorithm to estimate SOC. Its workflow is as follows: Figure 6 shown.

[0093] Specifically, the specific steps of SOC estimation combined with the SRCKF algorithm include:

[0094] S4.1. Initialization: Give the initial state estimate and the square root of the initial state error covariance matrix.

[0095] The initial state estimate is a guess at the initial SOC value, usually determined based on the initial battery charge measurement or known state of charge; the size and form of the initial state error covariance matrix depends on the confidence level in the initial SOC estimate. The greater the confidence level, the larger the element value in the matrix. For example, when only the SOC state variable is considered, it is a 1×1 matrix, and the element value is the variance of the initial SOC estimate. The specific expression is as follows:

[0096]

[0097] In the formula, chol(·) represents Cholesky decomposition, E[·] is the expectation, and x 0 To initialize the initial state, S 0|0 is the initial value of the square root, n represents the state dimension, m=2n.

[0098] S4.2, Time Update: According to the state equation, the volume point is used to propagate the state estimate, and the calculation formula for the cultivation point construction is:

[0099] Among them, x i，k|k is the state cultivation point input for the i-th cultivation point, i represents the i-th sample point, S k|k is the square root of the state error covariance matrix, ξ g For the cultivation point, is the estimated value of the state vector at time k;

[0100] The formula for calculating the propagation points is:

[0101] in, represents the predicted state vector at time k for time k+1. The subscript g is related to the sample point, f() represents the function, and x g，k|k is the state vector at time k, u k+1 is the input vector at time k+1;

[0102] The calculation formula of the state prediction value is:

[0103] in, is the state prediction value at time k+1, m represents the total number of sample points, Represents the state of the sample points after m propagation Find the average;

[0104] The square root of the forecast covariance matrix is ​​expressed as: In the formula, Tria[] represents QR decomposition, Q represents process noise, R represents measurement noise, S Q，k The square root of the measurement noise;

[0105]

[0106] S4.3, Time Update: Update the state estimate using the predicted state and measurement data according to the measurement equation,

[0107] Among them, the construction training points are:

[0108] Update the measurement point of propagation: g，k+1|k =TCN-GRU(u k+1 , x g，k+1|k , T k , v k ;

[0109] The predicted value for the observed value is calculated as:

[0110] The square root of the innovation covariance matrix can be derived as: S yy，k+1|k =Tria[ζ k+1|k , S R，k+1 ];

[0111] In the formula, S R，k+1 is the square root matrix of the measurement noise covariance matrix R (at time k+1);

[0112] in:

[0113]

[0114] The cross covariance matrix is ​​obtained as:

[0115] in:

[0116]

[0117] The Kalman gain matrix can be solved as:

[0118]

[0119] Therefore, the state can be estimated as:

[0120]

[0121] In the formula, z k+1 The actual measurement value at time k+1;

[0122] Determine the square root of the error covariance matrix:

[0123] S k+1|k+1 = Tria[x k+1|k -K k+1 ζ k+1|k , Kk+1 S R,k+1 ].

[0124] In the embodiment of the present application, TCN-GRU can effectively track voltage changes, while SRCKF can effectively correct SOC using battery model errors. This method can effectively shorten the battery test cycle and reduce BMS costs, especially when the training data of the target battery is insufficient and in the early stages of BMS development.

[0125] Advantages: As shown in Table 3, compared with the other two battery models, the maximum error is the smallest, below 1.1%.

[0126] Next, the SOC estimation results and battery model in this application are verified from three perspectives as follows:

[0127] 1. Comparison of several traditional modeling methods and state of charge estimation methods

[0128] Several traditional modeling methods and state of charge estimation methods are compared. In addition, in order to simulate the ability of the on-board BMS to correct the initial SOC after an unexpected power failure in actual application scenarios, the initial SOC is deliberately set to 90%, introducing a 10% error. The SOC statistical results of different algorithms are shown in Table 3:

[0129] Table 3

[0130]

[0131]

[0132] Depend on Figure 7 and Figure 8 It can be seen that the estimation results of battery 1 by different algorithms and models are compared, where "Reference" is the reference value calculated by the Ah integration method in the laboratory scenario, "1RC" and "2RC" are battery models constructed using first-order RC (1RC) and second-order RC (2RC) equivalent circuit models, respectively, and then the model parameters are identified based on the particle swarm optimization (PSO) algorithm. In summary, TCN-GRU can effectively track voltage changes, while SRCKF can effectively correct SOC using battery model errors. The synergistic combination of the optimal model and algorithm enables high SOC estimation accuracy, with a maximum SOC error of less than 1.1%.

[0133] 2. Evaluating the SOC estimation value of this method at different ambient temperatures

[0134] The SOC estimation of the proposed method at different ambient temperatures was evaluated, and the generalizability of the proposed method at different temperatures of Unit 22, Unit 216 and Unit 3 was studied. Two dynamic driving cycles were conducted using DST at high temperature (55°C) and FUDS at a dynamic temperature range (9-53°C) to evaluate the performance of the developed battery model and SOC estimation method;

[0135] Depend on Fig. 9 and Fig.10 It can be seen that the SOC obtained by this method can track the reference value well under high temperature and variable temperature conditions, and the error curve reduces vibration. The MAE and RMSE of SOC estimation in the DST cycle at 55°C and the FUDS cycle at variable temperature are 1.44%, 1.27% and 1.18%, 0.87%, respectively. The verification results show that this method can tolerate the battery parameter deviation caused by temperature changes and has satisfactory estimation accuracy.

[0136] 3. Use lithium iron phosphate battery with different characteristics from the source battery as the target battery for cross-battery material verification.

[0137] In order to reveal the versatility of the proposed method, LFP batteries (Cell 4 and Cell 5) with different characteristics from the source domain batteries are used as target batteries for cross-battery material verification. In this study, only the first cycle data of the target battery under the DST cycle, a very small data set containing approximately 6700 seconds of operation, is selected to fine-tune the source domain battery model, thereby obtaining a target domain battery model for estimating the SOC of the LFP battery. This paper uses the discharge data of the target battery under FUDS and US06, 0℃ and 50℃ dynamic conditions as the target domain test set to test the established transfer learning model; the method in this paper can accurately estimate the state of charge of the target battery, such as Fig.11 and Fig.12 The errors in these cases are also plotted in Fig.13 and Fig.14 The MAEs of our method can be limited to less than 2%.

[0138] In order to demonstrate the advantages of the SOC estimation method based on cross-domain transfer learning proposed in this paper, a series of machine learning models without transfer learning were also designed for SOC estimation under the same test conditions. Note that the hyperparameter settings of these models are the same as those of the models with transfer learning. Unlike the cross-domain transfer learning method proposed in this study, these machine learning methods do not migrate the effective parameters of the source domain model, but randomly initialize the parameters and then retrain and optimize the model. Therefore, these methods require a richer training set, using two additional cycles at the same temperature (about 16700 seconds) as training sets. Figure 11-16The SOC estimation results and error statistics of different machine learning methods are given. Under two different operating cycles and temperature conditions, the traditional machine learning-based SOC estimation method showed poor performance due to the limitation of the training dataset size. The best RMSE, MAE and MAX were 0.80%, 2.20% and 0.43%, respectively. However, these values ​​were improved by 53.23%, 69.25% and 29.11%, respectively, compared with the method proposed in this study. Therefore, the results of the SOC estimation method based on traditional machine learning are not as good as those based on cross-domain transfer learning. The TCN-GRU source domain model provides better and more accurate weight parameters for the TCN-GRU transfer learning target domain model. Therefore, it can learn the characteristics of other types of batteries with a smaller training data sample, making up for its data deficiency. Compared with the traditional machine learning-based SOC estimation method, the cross-domain transfer learning method has a wider knowledge base, which reduces the required data amount by 60% while improving the estimation accuracy. The machine learning method without a transfer learning model has limited training sample data and lacks the support of the source domain model. Therefore, these datasets are not sufficient for model training, resulting in their estimated performance being inferior to that of cross-domain transfer learning models.

[0139] Obviously, those skilled in the art should understand that the modules or steps of the above-mentioned embodiments of the present application can be implemented by general-purpose computer devices, they can be concentrated on a single computer device, or distributed on a network composed of multiple computer devices, optionally, they can be implemented by executable program codes of computer devices, so that they can be stored in a storage device and executed by the computer device, and in some cases, the steps shown or described can be executed in a different order from that herein, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. In this way, the embodiments of the present application are not limited to any specific combination of hardware and software.

[0140] It should be noted that the above are only preferred embodiments of the present application, and the patent protection scope of the present application is not limited thereto. Any equivalent structure or equivalent process transformation made using the contents of the specification and drawings of the present application, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A lithium-ion battery SOC estimation method based on TCN-GRU neural network model and transfer learning, characterized in that: include: S1. Data acquisition: Obtaining measurement data of the source battery under different discharge conditions and environmental conditions as a source domain data set, the measurement data includes the current, SOC, temperature and voltage of the battery; S2. TCN-GRU neural network model construction and pre-training: According to the battery behavior in the source domain, the TCN-GRU framework is used to take the battery current, SOC and temperature as the model input, and the battery voltage as the model output to build the TCN-GRU neural network model, and the data in the source domain dataset is used to pre-train the constructed TCN-GRU neural network model; S3, TCN-GRU neural network model expansion: Aiming at the battery behavior in the target domain, the TCN-GRU neural network model is extended by using transfer learning technology, and the data from the target domain is fine-tuned to adapt to the specific characteristics of the target battery, thus obtaining the extended TCN-GRU neural network model. S4. Construct a square root cubic Kalman filter (SRCKF) algorithm to estimate SOC.

2. The lithium-ion battery SOC estimation method according to claim 1, characterized in that: The source domain dataset includes two datasets, the first of which is test data obtained from laboratory tests, and the second is test data from the Center for Advanced Lifecycle Engineering (CALCE) at the University of Maryland.

3. The lithium-ion battery SOC estimation method according to claim 1, characterized in that: The constructed TCN-GRU neural network model is defined as:

4. The lithium-ion battery SOC estimation method according to claim 3, characterized in that: After fine-tuning, the expanded TCN-GRU neural network model is defined as: In the formula, θ src Represents the pre-training parameters.

5. The lithium-ion battery SOC estimation method according to claim 4, characterized in that: The loss of the extended TCN-GRU neural network model is calculated using a combined loss function that includes source domain loss and target domain loss. The combined loss function is: L=L src +λL tgt Where λ is a hyperparameter that balances the loss contribution of the source domain and the target domain, and L src and L tgt Denote the loss of the source domain and the target domain respectively, and are expressed as: Among them, N src and N tgt Represent the number of samples in the source dataset and the target dataset respectively.

Images