Battery module SOC estimation method based on deep learning and Kalman filtering algorithm

Abstract

The invention provides a battery module SOC estimation method based on deep learning and a Kalman filtering algorithm, and the method comprises the following steps: obtaining the working condition working data of a battery module, and carrying out the preprocessing of the working condition working data; taking the preprocessed working condition working data as characteristic parameters to train a double-layer gating cycle unit DGRU deep learning network; and integrating the trained double-layer gating cycle unit DGRU deep learning network with a Kalman filtering algorithm, and integrating the double-layer gating cycle unit and Kalman filtering into a prediction model DGRU-KF. According to the method, the prediction precision is improved through the double-layer gating cycle unit DGRU deep learning network, meanwhile, the double-layer gating cycle units can perform parallel and simultaneous calculation, the estimation time is saved on the basis of precision improvement, the double-layer gating cycle unit network and the Kalman filtering algorithm are combined, fluctuation output by a deep learning network model is reduced, and the prediction precision is improved. And the accuracy of the estimation model is further improved.

Description

Technical Field

[0001] The present invention relates to the technical field of battery SOC estimation, and specifically to a battery module SOC estimation method based on deep learning and Kalman filtering algorithm. Background Art

[0002] In recent years, with the rise of new energy vehicles and energy storage industries, lithium-ion batteries have become ideal power devices for power systems due to their long life, high energy density, and low self-discharge rate. In order to ensure the normal operation of the power system, efficient and safe battery management is particularly important. Since single lithium-ion batteries are limited by capacity and voltage, when used as a power source, hundreds of single batteries must be connected in series and parallel to form a battery pack to increase power and energy to meet the requirements of actual use. The state of the battery pack is related to the maximum mileage of electric vehicles, among which the battery SOC is an important part of the state estimation. Therefore, determining the battery pack SOC is crucial for estimating the remaining mileage of electric vehicles. Since the battery may have problems such as uneven materials, impurities and burr defects during the manufacturing process, batteries of the same specification may have inconsistencies when leaving the factory, and this inconsistency will continue to magnify during the long-term service of the battery. Therefore, it is very important to evaluate the consistency between single cells and then estimate the battery pack SOC to improve the safety of the battery pack.

[0003] At present, the common battery cell SOC estimation methods include direct measurement method, model estimation method and data-driven method. Among them, the data-driven method has been increasingly valued with the development of deep learning and artificial intelligence. It does not require the establishment of a complex battery model to display the battery status, nor does it require a specific understanding of the internal characteristics of the battery. It can simulate and predict the battery status well through data training, and can handle complex nonlinear systems quickly. In addition, the SOC estimation methods for battery modules are mostly migrations of traditional methods, with inaccurate predictions and time-consuming and labor-intensive. The data-driven method can effectively improve this problem, but currently there are few data-driven methods for battery module status estimation, and there is a lack of suitable directions. Summary of the invention

[0004] In response to the above technical problems, the present invention provides a battery module SOC estimation method based on deep learning and Kalman filtering algorithm, which improves the SOC estimation accuracy of the battery module.

[0005] Note that the description of these objectives does not prevent the existence of other objectives. One embodiment of the present invention does not need to achieve all of the above objectives. Objectives other than the above objectives can be extracted from the description of the specification, drawings, and claims.

[0006] The present invention achieves the above technical objectives through the following technical means.

[0007] A battery module SOC estimation method based on deep learning and Kalman filtering algorithm includes the following steps:

[0008] Step S1, obtaining the working data of the battery module, and preprocessing the working data;

[0009] Step S2, using the working condition data preprocessed in step S1 as feature parameters to train a double-layer gated recurrent unit DGRU deep learning network;

[0010] Step S3: Integrate the double-layer gated recurrent unit DGRU deep learning network trained in step S2 with the Kalman filter algorithm to form a double-layer gated recurrent unit and Kalman filter integrated prediction model DGRU-KF.

[0011] In the above scheme, in step S1, the operating data includes constant current discharge condition CDC, dynamic stress test DST, federal urban driving plan FUDS and Chinese light vehicle driving condition CLTC, and the current and voltage change values ​​of the battery during operation under each condition are collected respectively.

[0012] In the above scheme, in step S1, the operating condition data preprocessing is to obtain the voltage, total current, branch current and discharge capacity of the battery module monomer after the operating condition cycle through experimental monitoring to calculate the SOC value of the single battery in the module, and distinguish the minimum SOC min and maximum SOC max .

[0013] In the above scheme, in step S2, the double-layer gated recurrent unit DGRU deep learning network includes a first-layer gated recurrent unit GRU network and a second-layer gated recurrent unit GRU network.

[0014] In the above scheme, in step S2, the input layer of the first layer of the double-layer gated recurrent unit DGRU deep learning network includes two characteristic parameters, namely, battery performance parameter voltage U and main circuit current I, and the output layer includes two characteristic parameters, namely and

[0015] In the above scheme, in step S2, the input layer of the second layer of the double-layer gated recurrent unit DGRU deep learning network includes two characteristic parameters, namely, battery performance parameter voltage U and main circuit current I, and the output layer includes two characteristic parameters, namely and

[0016] In the above scheme, in step S2, the SOC estimation result of the double-layer gated recurrent unit DGRU deep learning network model is obtained by estimating the battery module SOC error through the second-layer gated recurrent unit GRU network of the double-layer gated recurrent unit DGRU deep learning network and compensating the SOC prediction value of the first-layer gated recurrent unit GRU network.

[0017] In the above scheme, in step S2, in the first layer of the gated recurrent unit GRU network of the double-layer gated recurrent unit DGRU deep learning network, the measured true value SOC obtained in step S1 is min and SOC max The input network is used as the training target, and then the first layer of the gated recurrent unit GRU network training is performed. The input layer of the first layer of the gated recurrent unit GRU network contains two characteristic parameters, namely the battery performance parameter voltage U and the main circuit current I, and the output layer includes two characteristic parameters, namely and The first-layer gated recurrent unit GRU network model is obtained through data training, and then the current and voltage data are input to obtain the output and value;

[0018] Since there is an error between the predicted value and the actual measured value, the error is compensated by the second-layer gated recurrent unit GRU network. and The measured true SOC min and SOC max The error true values ​​Er1 and Er2 are obtained by difference, and then the error true values ​​Er1 and Er2 are used as the training targets of the second-layer gated recurrent unit GRU network. The second-layer gated recurrent unit GRU network is trained based on the input network, where the input layer contains two characteristic parameters, namely, battery performance parameter voltage U and main circuit current I, and the output layer includes two characteristic parameters, namely and The trained second-layer gated recurrent unit GRU network model is obtained through data training, and then the current and voltage data are input and output to obtain and value;

[0019] Finally, the outputs of the first-layer gated recurrent unit GRU network and the second-layer gated recurrent unit GRU network are added to obtain the final estimation result of the double-layer gated recurrent unit DGRU deep learning model. and

[0020] In the above scheme, in step S3, the final estimated SOC value of the double-layer gated recurrent unit DGRU deep learning model in step S2 is input into the Kalman filter algorithm, and the SOC estimated at time t+1 is (t+1) As the observed value, the SOC at time t t As the state value, the final SOC estimation result is obtained.

[0021] In the above scheme, step S3 specifically includes the following steps:

[0022] State initialization: x0, P0, R, Q, the estimated SOC values ​​in step S2 are arranged in time series as SOC (t) And as the initial state value x0 for initialization input, after which SOC (t) As the state value, SOC (t+1) As the actual observed value;

[0023] Then enter the prediction stage, first update the state, and get the predicted value and observed value at the next moment through the state value at the current moment: x t+1 =A t x t +B t u t +w t ,y t =H t x t +v t ,

[0024] Re-forecast error covariance matrix:

[0025] Finally, we enter the update phase and calculate the Kalman gain matrix:

[0026] Then, the residual is calculated by subtracting the actual observed value from the predicted value for a posteriori estimation, and then the system state correction is updated based on the Kalman gain, the observed value, and the residual:

[0027] Update error covariance correction: P0 = (1-K t H t ) t ′,

[0028] Determine whether the time series meets the data time. If so, it is completed. Otherwise, it returns to the prediction stage.

[0029] Among them, x0, P0, R, and Q represent the initial state of the system, the initial error covariance, the initial measurement noise covariance, and the initial process noise covariance, respectively; x t is the system input state at time t, x t+1is the predicted value at the next moment, u t is the input information of the system, y t is the observed value, w t is the process noise, v t is the observation noise, A t is the state matrix, B t is the input matrix, H t is the state observation matrix, is the current state of the system, is the predicted state of the system, P t ′ is the prediction error covariance matrix, K t is the Kalman gain matrix, Q t is the process noise covariance at time t, R t is the covariance matrix of the observation noise.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] The double-layer gated cyclic unit and Kalman filter integrated prediction model DGRU-KF of the present invention estimates the battery module SOC error by using the second layer gated cyclic unit GRU network of the double-layer gated cyclic unit DGRU deep learning network and compensates the SOC prediction value of the first layer gated cyclic unit GRU network, thereby obtaining the SOC estimation result of the double-layer gated cyclic unit DGRU deep learning network model, thereby improving the prediction accuracy. At the same time, the double-layer gated cyclic unit can be calculated in parallel and simultaneously, saving estimation time on the basis of improved accuracy. The double-layer gated cyclic unit network and the Kalman filter algorithm are combined to reduce the fluctuation of the output of the deep learning network model and further improve the accuracy of the estimation model. In addition, the double-layer gated cyclic unit and Kalman filter integrated prediction model DGRU-KF of the present invention estimates and evaluates the data at the current moment, determines the data update situation at the next moment, thereby obtaining the time dependency between the data, effectively predicting the future data trend, improving the prediction accuracy of the overall model, and applying it to battery module state estimation, providing new ideas for the battery module research field.

[0032] Note that the description of these effects does not prevent the existence of other effects. One embodiment of the present invention does not necessarily have all of the above effects. Effects other than the above can be clearly seen and extracted from the description of the specification, drawings, claims, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a flow chart of a battery module SOC estimation method based on deep learning and Kalman filter algorithm in an embodiment of the present invention;

[0034] Figure 2is the initial data collected under different working conditions in the embodiment of the present invention, where Figure 2 (a) is the constant current working condition data, Figure 2 (b) is the dynamic stress test data, Figure 2 (c) is the Federal Urban Driving Program cycle data, Figure 2 (d) driving condition data of Chinese light vehicles;

[0035] Figure 3 is a flow chart of a double-layer gated recurrent unit (DGRU) deep learning network in an embodiment of the present invention;

[0036] Figure 4 is a schematic diagram of a flow chart of a Kalman filter algorithm in an embodiment of the present invention;

[0037] Figure 5 : is a diagram showing the battery SOC estimation results and error of different models under the same data in an embodiment of the present invention, wherein Figure 5 (a) is SOC min The estimated result diagram of Figure 5 (b) SOC min The error diagram, Figure 5 (c) SOC max The estimated result diagram of Figure 5 (d) SOC max The error diagram of

[0038] Figure 6 is a graph of battery SOC estimation evaluation indicators for different models under different operating conditions in the embodiment of the present invention, where Figure 6 (a) is the SOC under CDC conditions min Multi-model estimation evaluation index diagram, Figure 6 (b) is the SOC under CDC conditions max Multi-model estimation evaluation index diagram, Figure 6 (c) is the SOC under DST condition min Multi-model estimation evaluation index diagram, Figure 6 (d) is the SOC under DST condition max Multi-model estimation evaluation index diagram, Figure 6 (e) is the SOC under FUDS condition min Multi-model estimation evaluation index diagram, Figure 6 (f) is the SOC under FUDS condition max Multi-model estimation evaluation index diagram, Figure 6 (g) is the SOC under CLTC conditions min Multi-model estimation evaluation index diagram, Figure 6 (h) is the SOC under CLTC conditions max Plot of evaluation indicators for multiple model estimation. DETAILED DESCRIPTION

[0039] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0040] Figure 1 A battery module SOC estimation method based on deep learning and Kalman filtering algorithm according to an embodiment of the present invention comprises the following steps:

[0041] Step S1, obtaining the working data of the battery module, and preprocessing the working data;

[0042] In this embodiment, the working condition data in step S1 is obtained by testing and monitoring the aged battery module through a battery testing system. The working conditions used include constant current discharge condition (CDC), dynamic stress test (DST), Federal Urban Driving Scheme (FUDS) and China Light Vehicle Driving Condition (CLTC). The specific initial data is as follows: Figure 2 As shown, Figure 2 (a) Figure 2 (b) Figure 2 (c) Figure 2 (d) corresponds to four operating conditions: constant current discharge condition (CDC), dynamic stress test (DST), Federal Urban Driving Program (FUDS) and China Light Vehicle Driving Condition (CLTC). For each operating condition, the current and voltage change values ​​during the battery operation are collected.

[0043] In this embodiment, in step S1, the preprocessing is mainly to provide the data required for subsequent model training. The SOC value of the single battery in the module is calculated by experimentally monitoring the voltage, total current, branch current and discharge capacity of the battery module monomer after the working cycle, and the minimum SOC is distinguished. min and maximum SOC max .

[0044] Step S2, using the working condition data preprocessed in step S1 as feature parameters to train a double-layer gated recurrent unit DGRU deep learning network;

[0045] In this embodiment, the double-layer gated recurrent unit DGRU deep learning network in step S2 includes a first-layer gated recurrent unit GRU network and a second-layer gated recurrent unit GRU network. The input layer of the first-layer gated recurrent unit GRU network of the double-layer gated recurrent unit DGRU deep learning network includes two characteristic parameters, namely, battery performance parameter voltage U and main circuit current I, and the output layer includes two characteristic parameters, namely and The second layer of the double-layer gated recurrent unit DGRU deep learning network, the input layer of the gated recurrent unit GRU network contains two characteristic parameters, namely the battery performance parameter voltage U and the main circuit current I, and the output layer includes two characteristic parameters, namely and

[0046] In step S2, the SOC estimation result of the double-layer gated recurrent unit DGRU deep learning network model is obtained by using the second-layer gated recurrent unit GRU network of the double-layer gated recurrent unit DGRU deep learning network to estimate the battery module SOC error and compensate the SOC prediction value of the first-layer gated recurrent unit GRU network.

[0047] The double-layer gated recurrent unit DGRU deep learning network is trained using working condition data. The algorithm flow of the double-layer gated recurrent unit DGRU deep learning network is as follows: Figure 3 As shown. The total current and voltage of the initial data obtained in step S1 and the SOC obtained after preprocessing are input into the first layer of the gated recurrent unit GRU network, and the predicted SOC value is output. After error comparison, it is input into the second layer of the gated recurrent unit GRU network for error correction to obtain the final estimation result of the double-layer gated recurrent unit DGRU deep learning network.

[0048] The calculation process of the gated recurrent unit GRU network is as follows: The candidate hidden state is determined by controlling the information transmission through the reset gate and the update gate: The current and voltage measured in step S1 are used as the network input x t , the hidden state is obtained through network calculation to get the predicted value. In the first layer of the gated recurrent unit GRU network, the hidden state is the SOC predicted value, and in the second layer of the gated recurrent unit GRU network, the hidden state is the predicted value error.

[0049] z t =Sigmoid(W z [h t-1 ,x t ])

[0050] r t =Sigmoid(W r [h t-1 ,x t])

[0051]

[0052] Among them, x t is the system input state at time t, h t-1 is the hidden state at the previous time step, is the candidate hidden state for the current time step, h t is the hidden state at the current moment, z t and r t They are update gate and reset gate, W9, W > and w C They are the weight matrices of the update gate, reset gate and hidden state respectively. The matrix dimensions depend on the input vector and the hidden state. ⊙ represents element-by-element multiplication.

[0053] In the first layer of the double-layer gated recurrent unit DGRU deep learning network, the measured true value SOC obtained in step S1 is min and SOC max The input network is used as the training target, and then the first layer of the gated recurrent unit GRU network training is performed. The input layer of the first layer of the gated recurrent unit GRU network contains two characteristic parameters, namely the battery performance parameter voltage U and the main circuit current I, and the output layer includes two characteristic parameters, namely and The first-layer gated recurrent unit GRU network model is obtained through data training, and then the current and voltage data are input to obtain the output and value;

[0054] Since there is an error between the predicted value and the actual measured value, the error is compensated by the second-layer gated recurrent unit GRU network. and The measured true SOC min and SOC max The error true values ​​Er1 and Er2 are obtained by difference, and then the error true values ​​Er1 and Er2 are used as the training targets of the second-layer gated recurrent unit GRU network. The second-layer gated recurrent unit GRU network is trained based on the input network, where the input layer contains two characteristic parameters, namely, battery performance parameter voltage U and main circuit current I, and the output layer includes two characteristic parameters, namely and The trained second-layer gated recurrent unit GRU network model is obtained through data training, and then the current and voltage data are input and output to obtain and value;

[0055] Finally, the outputs of the first-layer gated recurrent unit GRU network and the second-layer gated recurrent unit GRU network are added to obtain the final estimation result of the double-layer gated recurrent unit DGRU deep learning model. and

[0056] Step S3, integrating the double-layer gated recurrent unit deep learning network with the Kalman filter algorithm, and the double-layer gated recurrent unit and Kalman filter integrated prediction model (DGRU-KF) for battery module SOC estimation: input the final estimated SOC value of the double-layer gated recurrent unit DGRU deep learning model in step S2 into the Kalman filter algorithm, and input the estimated SOC value at time t+1 (t+1) As the observed value, the SOC at time t t As the state value, the final SOC estimation result is obtained.

[0057] In this embodiment, the Kalman filter algorithm process used is as follows: Figure 4 As shown, the specific steps are as follows:

[0058] State initialization: x0, P0, R, Q, the estimated SOC values ​​in step S2 are arranged in time series as SOC (t) And as the initial state value x0 for initialization input, after which SOC (t) As the state value, SOC (t+1) As the actual observed value;

[0059] Then enter the prediction stage, first update the state, and get the predicted value and observed value at the next moment through the state value at the current moment: x t+1 =A t x t +B t u t +w t ,y t =H t x t +v t ,

[0060] Re-forecast error covariance matrix:

[0061] Finally, we enter the update phase and calculate the Kalman gain matrix:

[0062] Then, the residual is calculated by subtracting the actual observed value from the predicted value for a posteriori estimation, and then the system state correction is updated based on the Kalman gain, the observed value, and the residual:

[0063] Update error covariance correction: P0 = (1-Kt H t ) t ′,

[0064] Determine whether the time series (the calculation time of the current model) meets the data time (the total data time to be input). If so, it is completed, otherwise it returns to the prediction stage;

[0065] Among them, x0, P0, R, Q0 represent the initial state of the system, the initial error covariance, the initial measurement noise covariance, and the initial process noise covariance respectively; x t is the system input state at time t, x t+1 is the predicted value at the next moment, u t is the input information of the system, y t is the observed value, w t is the process noise, v t is the observation noise, A t is the state matrix, B t is the input matrix, H t is the state observation matrix, is the current state of the system, is the predicted state of the system, P t ′ is the prediction error covariance matrix, K t is the Kalman gain matrix, Q t is the process noise covariance at time t, R t is the covariance matrix of the observation noise.

[0066] The double-layer gated recurrent unit and Kalman filter integrated prediction model DGRU-KF of the present invention estimates and evaluates the data at the current moment and determines the data update status at the next moment, thereby obtaining the time dependency between the data, effectively predicting the future data trend, and improving the prediction accuracy of the overall model.

[0067] In this embodiment, for the dual-layer gated recurrent unit and Kalman filter integrated prediction model (DGRU-KF), the root mean square error (RMSE), the determination coefficient (R-Square, R 2 ), Mean Bias Error (MBE) and Mean Absolute Error (MAE). The calculation formulas for these evaluation indicators are as follows:

[0068]

[0069] Where SOC(t) is the original state of charge of the battery pack. is the average value of the original charge state, is the predicted value, and N is the number of estimated data obtained by the prediction model. For the RMSE, MAE, and MBE indicators, the closer they are to zero, the higher the prediction accuracy of SOC. 2 , a value close to 1 indicates that the predicted result matches the actual curve very well.

[0070] In order to more comprehensively verify the role of the DRGU-KF model in SOC estimation, the present invention uses Long Short-Term Memory (LSTM), Dual Long Short-Term Memory (DLSTM), Dual Long Short-Term Memory-Kalman Filter (DLSTM-KF), Gated Recurrent Unit (GRU), Dual Gated Recurrent Unit (DGRU) and Dual Dual Gated Recurrent Unit-Kalman Filter (DGRU-KF) and compares the estimation accuracy of this model under constant current discharge conditions (CDC), dynamic stress test (DST), Federal Urban Driving Program (FUDS) and China Light Vehicle Driving Conditions (CLTC) conditions, and verifies its applicability in battery modules. The verification results are as follows: Figure 5-6 As shown:

[0071] Figure 5 The battery SOC estimation of the battery module under DST conditions is shown. Figure 5 (a) and Figure 5 As can be seen in (b), in the estimation of the minimum SOC value, the error of the GRU prediction result has a certain periodic change trend, and the fluctuation amplitude is also increasing with time, with a peak value of 0.7%, and an absolute error between -0.2% and 0.7%. After the introduction of the error compensation mechanism, although DGRU has an error of 0.5% in the initial prediction, which is larger than the initial error predicted by GRU, it can well identify the trend of error changes as time goes by, and the error is significantly reduced. The absolute error is generally maintained between -0.2% and 0.2%. After adding the Kalman filter algorithm, the DGRU-KF estimates that the oscillation of SOC and the estimation error of SOC will be significantly reduced, effectively eliminating the influence of transient signal oscillation and reducing the fluctuation of model output.

[0072] from Figure 5 (c) and Figure 5As can be seen in (d), in the estimation of the maximum SOC value, the error of the GRU prediction result has no obvious regularity and fluctuates violently, with an absolute error between -0.5% and 0.7%. Similar to the SOC minimum value estimation result, the DGRU estimation prediction has an error of 0.5% at the beginning, but the error decreases rapidly over time and generally remains between -0.2% and 0.2%. Similarly, the prediction of DGRU-KF effectively eliminates the influence of transient signal oscillation on the basis of DGRU.

[0073] Figure 6 They are the SOC of the battery module under different models and different working conditions. min and SOC max Estimation error. Figure 6 (a) is the SOC under CDC conditions min Multi-model estimation evaluation index diagram, Figure 6 (b) is the SOC under CDC conditions max Multi-model estimation evaluation index diagram, Figure 6 (c) is the SOC under DST condition min Multi-model estimation evaluation index diagram, Figure 6 (d) is the SOC under DST condition max Multi-model estimation evaluation index diagram, Figure 6 (e) is the SOC under FUDS condition min Multi-model estimation evaluation index diagram, Figure 6 (f) is the SOC under FUDS condition max Multi-model estimation evaluation index diagram, Figure 6 (g) is the SOC under CLTC conditions min Multi-model estimation evaluation index diagram, Figure 6 (h) is the SOC under CLTC conditions max Multi-model estimation evaluation index diagram. As can be seen from the figure, the average R 2 The average R of the GRU network is 0.86, the average RMSE is 0.68%, the average MAE is 0.57%, and the average R 2 The average R of DGRU-KF is 0.94, the average RMSE is 0.59%, and the average MAE is 0.52%. 2 The average RMSE is 0.32%, and the average MAE is 0.29%. The comparison shows that the DGRU-KF model proposed in the present invention has a relatively good effect on battery module SOC estimation. In addition, the estimation time of this model is shortened by 59% compared with the forward and reverse gated recurrent unit (Bidirectional GRU).

[0074] The present invention obtains the working condition data of the battery module and pre-processes the working condition data; uses the pre-processed working condition data as feature parameters to train a double-layer gated recurrent unit (DGRU) deep learning network; integrates the double-layer gated recurrent unit deep learning network with the Kalman filter algorithm, and the double-layer gated recurrent unit and Kalman filter integrated prediction model DGRU-KF is used for battery module SOC estimation. The present invention applies the time series model to battery module state estimation through the double-layer gated recurrent unit and Kalman filter integrated prediction model DGRU-KF, saving estimation time on the basis of improved accuracy. At the same time, the double-layer gated recurrent unit network and the Kalman filter algorithm are combined to reduce the fluctuation of the DGRU model output and further improve the accuracy of the estimation model.

[0075] It should be understood that although this specification is described according to various embodiments, not every embodiment contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

[0076] The series of detailed descriptions listed above are only specific descriptions of feasible embodiments of the present invention. They are not intended to limit the scope of protection of the present invention. All equivalent embodiments or changes that do not deviate from the technical spirit of the present invention should be included in the scope of protection of the present invention.

Claims

1. A battery module SOC estimation method based on deep learning and Kalman filter algorithm, characterized in that: The following steps are involved: Step S1, obtaining the working data of the battery module, and preprocessing the working data; Step S2, using the working condition data preprocessed in step S1 as feature parameters to train a double-layer gated recurrent unit DGRU deep learning network; Step S3: Integrate the double-layer gated recurrent unit DGRU deep learning network trained in step S2 with the Kalman filter algorithm to form a double-layer gated recurrent unit and Kalman filter integrated prediction model DGRU-KF.

2. The battery module SOC estimation method based on deep learning and Kalman filtering algorithm according to claim 1 is characterized in that: In step S1, the operating data includes constant current discharge condition CDC, dynamic stress test DST, federal urban driving program FUDS and Chinese light vehicle driving condition CLTC, and the current and voltage change values ​​of the battery during operation under each condition are collected respectively.

3. The battery module SOC estimation method based on deep learning and Kalman filter algorithm according to claim 1 is characterized in that: In step S1, the operating condition data is preprocessed to obtain the voltage, total current, branch current and discharge capacity of the battery module monomer after the operating condition cycle through experimental monitoring, calculate the SOC value of the single battery in the module, and distinguish the minimum SOC min and maximum SOC max .

4. The battery module SOC estimation method based on deep learning and Kalman filter algorithm according to claim 1, characterized in that: In step S2, the double-layer gated recurrent unit DGRU deep learning network includes a first-layer gated recurrent unit GRU network and a second-layer gated recurrent unit GRU network.

5. The battery module SOC estimation method based on deep learning and Kalman filtering algorithm according to claim 4 is characterized in that: In step S2, the input layer of the first layer of the gated recurrent unit GRU network of the double-layer gated recurrent unit DGRU deep learning network includes two characteristic parameters, namely, battery performance parameter voltage U and main circuit current I, and the output layer includes two characteristic parameters, namely and 6. The battery module SOC estimation method based on deep learning and Kalman filter algorithm according to claim 5, characterized in that: In step S2, the input layer of the second layer of the gated recurrent unit GRU network of the double-layer gated recurrent unit DGRU deep learning network includes two characteristic parameters, namely, battery performance parameter voltage U and main circuit current I, and the output layer includes two characteristic parameters, namely, and 7. The battery module SOC estimation method based on deep learning and Kalman filtering algorithm according to claim 6, characterized in that: In step S2, the SOC estimation result of the double-layer gated recurrent unit DGRU deep learning network model is obtained by using the second-layer gated recurrent unit GRU network of the double-layer gated recurrent unit DGRU deep learning network to estimate the battery module SOC error and compensate the SOC prediction value of the first-layer gated recurrent unit GRU network.

8. The battery module SOC estimation method based on deep learning and Kalman filter algorithm according to claim 7, characterized in that: In step S2, in the first layer of the gated recurrent unit GRU network of the double-layer gated recurrent unit DGRU deep learning network, the measured true value SOC obtained in step S1 is min and SOC max As the training target, the input network is used, and then the first layer of gated recurrent unit GRU network training is performed. The first layer of gated recurrent unit GRU network model completed by data training is obtained, and then the current and voltage data output is obtained. and value; The error compensation is performed through the second layer of gated recurrent unit GRU network, and the prediction value obtained by the first layer of gated recurrent unit network is firstly and The measured true SOC min and SOC max The error true values ​​Er1 and Er2 are obtained by difference, and then the error true values ​​Er1 and Er2 are used as the training targets of the second-layer gated recurrent unit GRU network. The second-layer gated recurrent unit GRU network is trained based on the input network. The trained second-layer gated recurrent unit GRU network model is obtained through data training, and then the current and voltage data are input to obtain the output. and value; Finally, the outputs of the first-layer gated recurrent unit GRU network and the second-layer gated recurrent unit GRU network are added to obtain the final estimation result of the double-layer gated recurrent unit DGRU deep learning model. and 9. The battery module SOC estimation method based on deep learning and Kalman filter algorithm according to claim 1, characterized in that: In step S3, the final estimated SOC value of the double-layer gated recurrent unit DGRU deep learning model in step S2 is input into the Kalman filter algorithm, and the SOC estimated at time t+1 is converted into (t+1) As the observed value, the SOC at time t t As the state value, the final SOC estimation result is obtained.

10. The battery module SOC estimation method based on deep learning and Kalman filter algorithm according to claim 9, characterized in that: The step S3 specifically comprises the following steps: State initialization: x0, P0, R, Q, the SOC values ​​estimated in step S2 are arranged in time series as SOC (t) And as the initial state value x0 for initialization input, after which SOC (t) As the state value, SOC (t+1) As the actual observed value; Then enter the prediction stage, first update the state, and get the predicted value and observed value at the next moment through the state value at the current moment: x t+1 =A t x t +B t u t +w t ,y t =H t x t +v t , Re-forecast error covariance matrix: Finally, we enter the update phase and calculate the Kalman gain matrix: Then, the residual is calculated by subtracting the actual observed value from the predicted value for a posteriori estimation, and then the system state correction is updated based on the Kalman gain, the observed value, and the residual: Update error covariance correction: P0 = (1-K t H t ) t ′ , Among them, x0, P0, R, and Q represent the initial state of the system, the initial error covariance, the initial measurement noise covariance, and the initial process noise covariance, respectively; x t is the system input state at time t, x t+1 is the predicted value at the next moment, u t is the input information of the system, y t is the observed value, w t is the process noise, v t is the observation noise, A t is the state matrix, B t is the input matrix, H t is the state observation matrix, is the current state of the system, is the predicted state of the system, P t ′ is the prediction error covariance matrix, K t is the Kalman gain matrix, Q t is the process noise covariance at time t, R t is the covariance matrix of the observation noise.

Images