A Deep Learning-Based Joint Estimation Method for Lithium-ion Battery State of Charge and Health

Abstract

This invention discloses a deep learning-based method for jointly estimating the state of charge (SOC) and state of health (SOH) of lithium batteries. This invention fully considers the self-discharge effect of lithium batteries and employs an improved ampere-hour integral method to calculate SOC and SOH, eliminating the influence of self-discharge and significantly improving the accuracy of SOC and SOH calculations, as well as the precision of training data, thereby enhancing the accuracy of SOC and SOH estimation. The TCN-LSTM neural network model is applied for the first time to the joint estimation of the SOC and SOH of lithium batteries, achieving higher accuracy compared to other models. The hyperparameters of the TCN-LSTM neural network model are optimized using a Bayesian optimization algorithm to reduce labor costs and improve model performance. Kalman filtering is used to smooth the estimation results, reducing fluctuations and improving the accuracy of SOC and SOH estimation.

Description

Technical Field

[0001] This invention relates to the field of power battery management, specifically a method for jointly estimating the state of charge and state of health of lithium batteries based on deep learning. Background Technology

[0002] Excessive carbon dioxide emissions are a major contributor to climate change. To control carbon emissions, vigorously promoting the development of new energy sources is a key direction. Among numerous available power sources, lithium batteries have attracted widespread attention due to their advantages such as high energy density, high energy ratio, wide operating temperature range, long storage life, and zero emissions.

[0003] Due to individual differences between batteries, overcharging and over-discharging can easily occur during actual use. To ensure safe battery use, the estimation of SOC (State of Charge) and SOH (State of Health) are core management indicators.

[0004] State of Charge (SOC) is the ratio of current remaining charge to current maximum usable capacity, influenced by factors such as temperature, depth of discharge, and battery health. During storage and use, batteries age due to self-discharge and the characteristics of battery materials, resulting in a decrease in battery capacity. State of Harm (SOH) is an assessment of the degree of battery aging, defined as the ratio of current maximum usable capacity to rated capacity. Due to the highly nonlinear nature of batteries, SOC and SOH cannot be directly measured and must be estimated. Currently, methods for estimating SOC and SOH mainly include experimental methods, model methods, and data-driven methods. Experimental methods can only be performed in experimental environments and cannot be generalized to real-world applications; model methods are difficult to implement, have poor model universality, and are costly; data-driven methods have strict requirements on the quantity and quality of data.

[0005] Currently, the most commonly used method for estimating SOC and SOH is the neural network method within the data-driven approach. This method has strict requirements for the data, as noise is prevalent in the measurement data, and existing methods do not perform noise reduction. Furthermore, current methods using the ampere-hour integration method to calculate SOC and SOH do not consider the self-discharge characteristics of lithium-ion batteries, leading to inaccurate data and increasing the estimation error. Hyperparameters in the neural network model have a significant impact on its performance; currently, hyperparameter tuning is mostly done manually, resulting in wasted time and manpower. Due to the characteristics of neural network models, the results fluctuate drastically during online estimation, a problem that current methods do not address.

[0006] In summary, current methods for estimating the state of charge and state of health of lithium batteries suffer from problems such as data noise, self-discharge, and fluctuations. Therefore, it is necessary to propose a comprehensive method for estimating the state of charge and state of health of lithium batteries that takes into account various factors. Summary of the Invention

[0007] To address the shortcomings of existing technologies, the technical problem this invention aims to solve is to provide a method for jointly estimating the state of charge and health status of lithium batteries based on deep learning.

[0008] The technical solution of this invention to solve the aforementioned technical problem is to provide a method for jointly estimating the state of charge and health status of a lithium battery based on deep learning, characterized in that the method includes the following steps:

[0009] Step 1: Establish the SOC estimation model and the SOH estimation model, and train the models offline to obtain the trained SOC estimation model and SOH estimation model respectively;

[0010] S101. Determine the rated capacity of the lithium battery;

[0011] S102. Set the ambient temperature and leave the lithium battery in the environment until the lithium battery temperature stabilizes. Record the resting time T. r Then, the lithium battery is charged or discharged according to different power configuration curves until the lithium battery voltage reaches the charging cutoff voltage or discharging cutoff voltage, thus completing one charging cycle or discharging cycle under each power configuration curve; each charging cycle and discharging cycle includes several time steps; during the battery's charging or discharging cycle, data of the lithium battery is collected at the required frequency to obtain charging data for one charging cycle or discharging data for one discharging cycle. Both charging and discharging data include three parameters: voltage, current, and temperature, which are collected at multiple steps; then, the SOC and SOH are calculated using the improved ampere-hour integration method based on the charging and discharging data.

[0012] The calculation formula for the improved ampere-hour integration method is shown in equation (2):

[0013]

[0014] In equation (2), C max δ represents the current maximum usable capacity of the lithium battery; δ represents the self-discharge rate of the lithium battery every thirty days.

[0015] According to equation (2), the formula for calculating SOC is shown in equation (3):

[0016]

[0017] According to equation (2), the formula for calculating SOH is shown in equation (4):

[0018]

[0019] In equation (4), C TThe amount of electricity a lithium battery can hold in one charge or discharge cycle; C ori This refers to the rated capacity of the lithium battery.

[0020] S103. Perform one charge-discharge cycle according to the maximum allowable charge-discharge rate of the lithium battery, without collecting any data during the cycle; each charge-discharge cycle includes several charging cycles and discharging cycles.

[0021] S104, repeat S102 and S103 until the SOH of the lithium battery drops to 80%, and complete the collection of charging and discharging data for all cycles.

[0022] S105. Perform equal-length sampling on the charging and discharging data used for SOH estimation to make the data lengths equal;

[0023] S106. Build the EMD module; then use the EMD method to decompose all parameter data into components to achieve noise reduction; then save the used EMD module as an EMD noise reduction tool.

[0024] S107. Establish a correlation analysis module to perform correlation analysis on all components obtained in S106 with SOC and SOH respectively, and obtain the component with the highest correlation; then use the component with the highest correlation to replace the parameter data in all loops to obtain the estimated SOC data and estimated SOH data; then save the used correlation analysis module as a correlation analysis tool.

[0025] S108. Normalize the SOC and SOH estimation data from S107 to obtain the SOC estimation dataset and SOH estimation dataset, and record the maximum value x of each parameter during the normalization process. max and minimum value x min ;

[0026] S109. Dataset partitioning: Divide the SOC estimation dataset and SOH estimation dataset obtained in S108 into their respective training sets and test sets;

[0027] S110, Bayesian Hyperparameter Optimization: Two TCN-LSTM neural network models are built to obtain a SOC estimation neural network model and a SOH estimation neural network model for SOC estimation and SOH estimation, respectively; the input and output forms of the two models and the value range of each hyperparameter are set, and the Bayesian optimization algorithm is used to achieve adaptive optimization of the model hyperparameters to obtain the optimal hyperparameters of the SOC estimation model and the SOH estimation model.

[0028] S111. Train the SOC estimation model and the SOH estimation model respectively:

[0029] The SOC estimation training set data is input into the SOC estimation model in the form of the SOC estimation model input in S110 for training; then the trained model is tested using the SOC estimation test set data, and a threshold is set according to the accuracy requirements; if the test error is greater than the threshold, the process returns to S110 for retraining; if the test error is less than or equal to the threshold, the trained SOC estimation model is obtained.

[0030] Similarly, the trained SOH estimation model is obtained;

[0031] Step 2: Based on the trained SOC estimation model and SOH estimation model obtained in Step 1, perform online estimation of SOC and SOH;

[0032] S201, Signal Monitoring: Using sensors to monitor the voltage, current, and temperature signals of the lithium battery;

[0033] S202, Signal Acquisition: For data used for SOC estimation, the voltage, current, and temperature parameters are acquired at each time step in the cycle according to the required frequency; for data used for SOH estimation, the voltage, current, and temperature parameters are acquired at the required frequency after each charge-discharge cycle.

[0034] S203. The charging and discharging data used for SOH estimation are sampled at the same length as the data length in S105;

[0035] S204. The EMD noise reduction tool of S106 is used to decompose the collected voltage, current and temperature parameter data into components to achieve noise reduction.

[0036] S205. Using the correlation analysis tool in S107, perform correlation analysis on all components obtained from the decomposition in S204 with SOC and SOH respectively, and obtain the component with the highest correlation; then use the component with the highest correlation to replace the parameter data in all cycles to obtain the estimated SOC data and estimated SOH data.

[0037] S206. The SOC and SOH estimates from S205 are taken from the maximum value x in S108. max and minimum value x min Normalization is performed to obtain normalized SOC estimates and normalized SOH estimates;

[0038] S207, SOH and SOC estimates:

[0039] The normalized SOC estimation data is input into the trained SOC estimation model in S111 in the form of the SOC estimation model input in S110, and the output of the SOC estimation model is obtained; then the maximum value x in S108 is used.max and minimum value x min The output of the SOC estimation model is inversely normalized to obtain the SOC estimate.

[0040] Similarly, the estimated value of SOH is obtained;

[0041] S208. Smooth the SOC and SOH estimates respectively to filter out the instantaneous fluctuations of the SOC and SOH estimates, and obtain the smoothed SOC estimate for the current time step. m and the actual SOH value of the current charge / discharge cycle

[0042]

[0043] S209, SOH Update: The first SOH update uses the actual SOH value from the first charge / discharge cycle. Replace the initial SOH value; subsequent SOH updates will use the actual SOH value for the current charge / discharge cycle. Replace the actual SOH value from the previous charge / discharge cycle. To perform SOH updates; n represents the nth charge / discharge cycle;

[0044] S210, SOC Update: First, based on the smoothed estimate of SOC at the current time step obtained in S208, SOC... m The true SOC value at the current time step is calculated using equation (11). Then perform a SOC update; the first SOC update uses the actual SOC value at the current time step. The smoothed estimate of SOC at the current time step obtained by replacing S208 m Subsequent SOC updates will be based on the actual SOC value at the current time step. Replace the previous time step SOC estimated true value

[0045]

[0046] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0047] (1) This invention fully considers the self-discharge effect of lithium batteries and uses an improved ampere-hour integration method to calculate SOC and SOH. By removing the influence of self-discharge, the accuracy of SOC and SOH calculation is greatly improved, the accuracy of training data is improved, and thus the accuracy of SOC and SOH estimation is enhanced.

[0048] (2) This invention uses EMD method and Pearson correlation analysis to denoise the data, optimize the model input form, and improve the model accuracy.

[0049] (3) This invention applies the TCN-LSTM neural network model to the joint estimation of the state of charge and health of lithium batteries for the first time, and has higher accuracy than other models.

[0050] (4) This invention utilizes the Bayesian optimization algorithm to optimize the hyperparameters of the TCN-LSTM neural network model, thereby reducing labor costs and improving model performance. For each hyperparameter, a range is given, and an evaluation equation is set to evaluate the performance of the neural network model. The Bayesian optimization algorithm will automatically evaluate the performance of the neural network model under different hyperparameters and output the hyperparameters when the neural network model has the best performance, thus optimizing the hyperparameters of the neural network model.

[0051] (5) The present invention uses Kalman filtering to smooth the estimation results, which reduces the fluctuation of the results and improves the estimation accuracy of SOC and SOH. Attached Figure Description

[0052] Figure 1 This is a flowchart of the offline training process for the joint estimation method of SOC and SOH of the present invention.

[0053] Figure 2 This is an online estimation flowchart of the joint SOC and SOH estimation method of the present invention;

[0054] Figure 3 This is a schematic diagram of the joint estimation model of SOC and SOH based on TCN-LSTM of the present invention. Detailed Implementation

[0055] Specific embodiments of the present invention are given below. These specific embodiments are only used to further illustrate the present invention in detail and do not limit the scope of protection of the claims of the present invention.

[0056] This invention provides a deep learning-based method for jointly estimating the state of charge and state of health of lithium batteries (hereinafter referred to as the method), characterized by the following steps:

[0057] Step 1: Establish the SOC estimation model and the SOH estimation model, and train the models offline to obtain the trained SOC estimation model and SOH estimation model respectively;

[0058] S101. Determine the rated capacity of the lithium battery;

[0059] Preferably, in step S101, the lithium battery is fully charged and fully discharged, and the rated capacity of the lithium battery is determined using the ampere-hour integration method, as shown in equation (1):

[0060]

[0061] In equation (1), Cori I represents the rated capacity of the lithium battery, and I represents the current.

[0062] S102. Set the ambient temperature and leave the lithium battery in the environment until the lithium battery temperature stabilizes. Record the resting time T. r Then, the lithium battery is charged or discharged according to different power configuration curves until the lithium battery voltage reaches the charging cutoff voltage or discharging cutoff voltage, thus completing one charging cycle or discharging cycle under each power configuration curve. Each charging cycle and discharging cycle includes several time steps, which are related to the acquisition frequency. During the charging or discharging cycle of the battery, the lithium battery data is acquired at the required frequency (1Hz in this embodiment) to obtain charging data (referred to as charging data) for one charging cycle or discharging data (discharging data) for one discharging cycle. Both charging data and discharging data include multiple steps of voltage, current, and temperature data. Then, the SOC and SOH are calculated using the improved ampere-hour integration method based on the charging data and discharging data.

[0063] Because lithium batteries undergo self-discharge during rest, resulting in a decrease in battery capacity, the ampere-hour integration method commonly used to calculate SOC and SOH does not consider this phenomenon. This leads to inaccurate calculations of SOC and SOH, thus affecting the accuracy of SOC and SOH estimates. Therefore, this invention proposes an improved ampere-hour integration method that considers the self-discharge effect of lithium batteries to calculate SOC and SOH. The calculation formula of the improved ampere-hour integration method is shown in equation (2):

[0064]

[0065] In equation (2), C max δ represents the current maximum usable capacity of the lithium battery; δ is the self-discharge rate of the lithium battery every thirty days, measured by a lithium battery discharge experiment; T r Settling time (in hours);

[0066] According to equation (2), the formula for calculating SOC is shown in equation (3):

[0067]

[0068] According to equation (2), the formula for calculating SOH is shown in equation (4):

[0069]

[0070] In equation (4), C T The amount of electricity a lithium battery can hold in one charge or discharge cycle; C ori This refers to the rated capacity of the lithium battery.

[0071] Preferably, in step S102, the ambient temperature is -20℃, -10℃, 0℃, 10℃, 25℃ and 35℃ to cover the ambient temperature during the actual use of the lithium battery.

[0072] Preferably, in step S102, the power configuration includes constant current charging, constant current discharging, DST driving cycle, US06 driving cycle, FUDS driving cycle, CLTC driving cycle, LA92 driving cycle, and driving cycle randomly mixed from the above cycles.

[0073] S103. Since the SOH of a lithium battery changes slowly, a charge-discharge cycle is performed based on the maximum allowable charge-discharge rate and specific accuracy requirements of the lithium battery, during which no data is collected; each charge-discharge cycle includes several charging cycles and discharging cycles.

[0074] S104. It is generally believed in the industry that the battery life ends when the lithium battery capacity decays to 80% of the nominal capacity. Therefore, repeat S102 and S103 until the SOH of the lithium battery drops to 80% to complete the collection of charging and discharging data for all cycles.

[0075] S105. Perform equal-length sampling on the charging and discharging data used for SOH estimation: Since the neural network model requires the input data to be of the same length, and the charging and discharging data of all cycles of the lithium battery obtained in S104 are of different lengths, all charging and discharging data are sampled to make the data lengths equal.

[0076] Preferably, in step S105, the data length is determined according to the performance of the mounted device; in this embodiment, it is 100 to 1000 time steps of data.

[0077] S106. Build an EMD (Empirical Mode Decomposition) module; then use the EMD method to decompose the parameter data in all charging and discharging data into components to achieve noise reduction; then save the used EMD module as an EMD noise reduction tool for online estimation in step 2.

[0078] Preferably, in step S106, the decomposition involves decomposing the parameter data into a series of intrinsic mode components (IMFs) and signal residual components, as follows:

[0079]

[0080] In equation (5), x(t) represents the parameter data to be decomposed; imf i For the i-th IMF obtained from the decomposition; r k This refers to the residual component of the signal.

[0081] S107. Using all components obtained in S106 as input data for the neural network model would increase computational costs. Therefore, a correlation analysis module is established to analyze all components obtained in S106. i (t) and r k (t) Perform correlation analysis with SOC and SOH respectively to obtain the component with the highest correlation; then use the component with the highest correlation to replace the parameter data in all loops to obtain the estimated SOC data and estimated SOH data; then save the used correlation analysis module as a correlation analysis tool for online estimation in step 2;

[0082] Preferably, in steps S106 and S107, the EMD module and the correlation analysis module are built in the Python platform.

[0083] Preferably, in step S107, the correlation analysis uses Pearson correlation analysis, and the calculation formula for Pearson correlation analysis is as follows:

[0084]

[0085] In equation (6), r represents the correlation coefficient; N is the data length; and X is the component obtained from the decomposition. Y represents the component average, where SOC or SOH data is used. The mean value of SOC or SOH is given; 0.0≤r<0.2 indicates no correlation, 0.2≤r≤0.4 indicates a weak correlation, 0.4<r≤0.6 indicates a moderate correlation, 0.6<r≤0.8 indicates a strong correlation, and 0.8<r≤1.0 indicates a very strong correlation.

[0086] S108. Normalize the SOC and SOH estimation data obtained in S107 to obtain the SOC estimation dataset and SOH estimation dataset, and record the maximum value x of each parameter during the normalization process. max and minimum value x min , used for online estimation in step 2;

[0087] Preferably, in step S108, for the SOC estimation data, each parameter data is normalized to [0,1], as shown in the formula:

[0088]

[0089] In equation (7), x SOC For the normalized SOC estimation data, x min x is the minimum value of this parameter data. max This represents the maximum value of the parameter data.

[0090] Because the input data for SOH estimation is large and varies widely, each parameter in the SOH estimation data is normalized to [-1, 1], as shown in the formula:

[0091]

[0092] In equation (8), x SOH The data are normalized SOH estimates.

[0093] S109. Dataset partitioning: Divide the SOC estimation dataset and SOH estimation dataset obtained in S108 into their respective training sets and test sets;

[0094] Preferably, in step S109, the training set is used to train the model, accounting for 70% of the entire dataset; the test set is used to verify the reliability of the model, accounting for 30% of the entire dataset.

[0095] S110. Bayesian Hyperparameter Optimization: Two TCN-LSTM (Temporal Convolutional Network-Long Short-Term Memory) neural network models are constructed to obtain a SOC estimation neural network model (hereinafter referred to as the SOC estimation model) and a SOH estimation neural network model (hereinafter referred to as the SOH estimation model), which are used for SOC estimation and SOH estimation, respectively. The input and output forms of the two models and the value range of each hyperparameter are set, and the Bayesian optimization algorithm is used to achieve adaptive optimization of the model hyperparameters to obtain the optimal hyperparameters of the SOC estimation model and the SOH estimation model, so as to achieve the best performance of the neural network model.

[0096] Preferably, in step S110, the SOC estimation model and the SOH estimation model have the same structure, specifically: the first layer is a TCN layer for extracting data features; after the TCN layer, several LSTM layers are sequentially connected to learn the temporal relationships in the data, and each LSTM layer is sequentially connected; after the last LSTM layer, a fully connected layer with tanh as the activation function is sequentially connected, and then another fully connected layer is sequentially connected as the output layer, and the optimizer is the Adam optimizer.

[0097] Preferably, in step S110, on the Python platform, a hyperparameter optimization module for neural network models is built using the hyperopt backend, and two TCN-LSTM neural network models are constructed respectively.

[0098] Preferably, in step S110, the hyperparameters of the SOC estimation model and the SOH estimation model and their value ranges are as follows:

[0099] Hyperparameters include: number of TCN filters, TCN kernel size, TCN inflation factor, number of LSTM layers, number of LSTM units, dropout rate, learning rate, number of fully connected layer units, number of training epochs, and batch size; details are as follows:

[0100] TCN layer: Number of filters N f (lnN f ~N(ln(40),ln(1.15))), number of kernels K s (lnK s ~N(ln(7),ln(1.12))), stack size N st (N s ~N(4,0.5)), expansion coefficient d i ((1,2));

[0101] LSTM layer number N L (N L ~U(1,3));

[0102] First LSTM layer: Number of cells U1(lnU1~N(ln(50),ln(1.3)));

[0103] The second LSTM layer has the following number of cells: U2(lnU2~N(ln(50),ln(1.3)));

[0104] The third LSTM layer has the following number of cells: U3 (lnU3 ~ N (ln(50), ln(1.3))), and a discard rate D. r (D r ~U(0.01,0.1));

[0105] First fully connected layer: Number of elements U4(lnU1~N(ln(50),ln(1.3)));

[0106] Last fully connected layer: Number of cells U5(1);

[0107] Adam optimizer: learning rate L r (lnL r ~N(ln(0.0011),ln(1.2))), first momentum B1(0.9), second momentum B2(0.999).

[0108] Preferably, in step S110, the optimization evaluation equation of the TCN-LSTM neural network model is:

[0109]

[0110] In equation (9), y i With y i 'These are the actual value and the estimated value, respectively.

[0111] Preferably, in step S110, the input to the SOC estimation model is the voltage, current, and temperature data for each step of each loop in the SOC estimation dataset, in the form of [V n ,I n ,T n The output is SOC; the input to the SOH estimation model is the voltage, current, and temperature data for each charge-discharge cycle in the SOH estimation dataset, in the form of [V1, V2, ..., V]. n ,I1,I2,…,I n ,T1,T2,…,T n The output is SOH.

[0112] S111. Train the SOC estimation model and the SOH estimation model respectively:

[0113] The hyperparameters of the SOC estimation model are set according to the optimal hyperparameters in S110. The SOC estimation training set data divided in S109 is input into the SOC estimation model for training, using the input form of the SOC estimation model in S110. The model optimization evaluation equation is set as MSE. Then, the trained model is tested using the SOC estimation test set data. The test error of the SOC estimation model is calculated using MSE, and a threshold is set according to the accuracy requirements. If the test error is greater than the threshold, the process returns to S110 for retraining. If the test error is less than or equal to the threshold, the trained SOC estimation model is obtained.

[0114] Similarly, to obtain the trained SOH estimation model, the following steps are taken: The hyperparameters of the SOH estimation model are set according to the optimal hyperparameters in S110. The SOH estimation training set data divided in S109 is input into the SOH estimation model for training, using the input format of the SOH estimation model in S110. The model optimization evaluation equation is set as MSE. Then, the trained model is tested using the SOH estimation test set data. The test error of the SOH estimation model is calculated using MSE, and a threshold is set according to the accuracy requirements. If the test error > the threshold, the process returns to S110 for retraining. If the test error ≤ the threshold, the trained SOH estimation model is obtained.

[0115] Preferably, step S111 is performed on the Python platform.

[0116] Step 2: Based on the trained SOC estimation model and SOH estimation model obtained in Step 1, perform online estimation of SOC and SOH;

[0117] S201, Signal Monitoring: Using sensors to monitor the voltage, current, and temperature signals of the lithium battery;

[0118] S202, Signal Acquisition: For data used for SOC estimation, the voltage, current, and temperature parameters are acquired at each time step in the cycle according to the required frequency; for data used for SOH estimation, the voltage, current, and temperature parameters are acquired at the required frequency after each charge-discharge cycle.

[0119] S203. The charging and discharging data used for SOH estimation are sampled at the same length as the data length in S105;

[0120] S204. The EMD noise reduction tool of S106 is used to decompose the collected voltage, current and temperature parameter data into components to achieve noise reduction.

[0121] S205. Using the correlation analysis tool in S107, perform correlation analysis on all components obtained from the decomposition in S204 with SOC and SOH respectively, and obtain the component with the highest correlation; then use the component with the highest correlation to replace the parameter data in all cycles to obtain the estimated SOC data and estimated SOH data.

[0122] S206. The SOC and SOH estimates from S205 are taken from the maximum value x in S108. max and minimum value x min Normalization is performed to obtain normalized SOC estimates and normalized SOH estimates;

[0123] Preferably, in step S206, the SOC estimation data is normalized to [0,1], and the SOH estimation data is normalized to [-1,1].

[0124] S207, SOH and SOC estimates:

[0125] The normalized SOC estimation data is input into the trained SOC estimation model in S111 in the form of the SOC estimation model input in S110, and the output of the SOC estimation model is obtained; then the maximum value x in S108 is used. max and minimum value x min The output of the SOC estimation model is inversely normalized to obtain the SOC estimate.

[0126] Similarly, to obtain the SOH estimate, the normalized SOH estimate data is input into the trained SOH estimate model in S111 using the input form of the SOH estimate model in S110, to obtain the output of the SOH estimate model; then the maximum value x in S108 is used. max and minimum value x min The output of the SOH estimation model is inversely normalized to obtain the SOH estimate;

[0127] S208. Smooth the SOC and SOH estimates respectively to filter out their instantaneous fluctuations, making them more stable and improving accuracy, thus obtaining the smoothed SOC estimate for the current time step. m and the actual SOH value of the current charge / discharge cycle

[0128] Preferably, in step S208, the smoothing process employs Kalman filtering, and the specific process is as follows:

[0129]

[0130]

[0131]

[0132]

[0133]

[0134] In equation (10), This represents the SOC estimate for the current time step or the SOH estimate for the current period. Let A be the SOC estimate of the previous time step or the SOH estimate of the previous cycle, and B be the state transition matrix and the control matrix. k-1 To control the input, P is the covariance between the smoothed SOC estimate and the SOC estimate, or between the actual SOH value and the SOH estimate. k-1 The covariance between the smoothed SOC estimate and the SOC estimate at the previous time step, or between the actual SOH value and the SOH estimate at the previous period, where Q is the process noise variance, and K is the variance. k Let H be the Kalman gain, H be the state-observation matrix, R be the observation noise variance, and z be the Kalman gain. k The observations are for the state matrix; the values ​​of Q and R are set according to the fluctuations of the estimated SOC or SOH values.

[0135] S209, SOH Update: The first SOH update uses the actual SOH value from the first charge / discharge cycle. Replace the initial SOH value, which is set to 100%. Subsequent SOH updates will use the actual SOH value for the current charge / discharge cycle. Replace the actual SOH value from the previous charge / discharge cycle. To perform SOH updates; n represents the nth charge / discharge cycle;

[0136] S210, SOC Update: First, based on the smoothed estimate of SOC at the current time step obtained in S208, SOC... mThe true SOC value at the current time step is calculated using equation (11). Then perform a SOC update; the first SOC update uses the actual SOC value at the current time step. The smoothed estimate of SOC at the current time step obtained by replacing S208 m Subsequent SOC updates will be based on the actual SOC value at the current time step. Replace the previous time step SOC estimated true value

[0137]

[0138] Any aspects not covered in this invention are applicable to existing technologies.

Claims

1. A method for jointly estimating the state of charge and state of health of a lithium battery based on deep learning, characterized in that, The method includes the following steps: Step 1: Establish the SOC estimation model and the SOH estimation model, and train the models offline to obtain the trained SOC estimation model and SOH estimation model respectively; S101. Determine the rated capacity of the lithium battery; S102. Set the ambient temperature and leave the lithium battery in the environment until the lithium battery temperature stabilizes. Record the resting time T. r Then, the lithium battery is charged or discharged according to different power configuration curves until the lithium battery voltage reaches the charging cutoff voltage or discharging cutoff voltage, thus completing one charging cycle or discharging cycle under each power configuration curve; each charging cycle and discharging cycle includes several time steps; during the battery's charging or discharging cycle, data of the lithium battery is collected at the required frequency to obtain charging data for one charging cycle or discharging data for one discharging cycle. Both charging and discharging data include three parameters: voltage, current, and temperature, which are collected at multiple steps; then, the SOC and SOH are calculated using the improved ampere-hour integration method based on the charging and discharging data. The calculation formula for the improved ampere-hour integration method is shown in equation (2): In equation (2), C max δ represents the current maximum usable capacity of the lithium battery; δ represents the self-discharge rate of the lithium battery every thirty days. According to equation (2), the formula for calculating SOC is shown in equation (3): According to equation (2), the formula for calculating SOH is shown in equation (4): In equation (4), C T The amount of electricity a lithium battery can hold in one charge or discharge cycle; C ori This refers to the rated capacity of the lithium battery. S103. Perform one charge-discharge cycle according to the maximum allowable charge-discharge rate of the lithium battery, without collecting any data during the cycle; each charge-discharge cycle includes several charging cycles and discharging cycles. S104, repeat S102 and S103 until the SOH of the lithium battery drops to 80%, and complete the collection of charging and discharging data for all cycles. S105. Perform equal-length sampling on the charging and discharging data used for SOH estimation to make the data lengths equal; S106. Build the EMD module; then use the EMD method to decompose all parameter data into components to achieve noise reduction; then save the used EMD module as an EMD noise reduction tool. S107. Establish a correlation analysis module to perform correlation analysis on all components obtained in S106 with SOC and SOH respectively, and obtain the component with the highest correlation; then use the component with the highest correlation to replace the parameter data in all loops to obtain the estimated SOC data and estimated SOH data; then save the used correlation analysis module as a correlation analysis tool. S108. Normalize the SOC and SOH estimation data from S107 to obtain the SOC estimation dataset and SOH estimation dataset, and record the maximum value x of each parameter during the normalization process. max and minimum value x min ; S109. Dataset partitioning: Divide the SOC estimation dataset and SOH estimation dataset obtained in S108 into their respective training and test sets; S110, Bayesian Hyperparameter Optimization: Two TCN-LSTM neural network models are built to obtain a SOC estimation neural network model and a SOH estimation neural network model for SOC estimation and SOH estimation, respectively; the input and output forms of the two models and the value range of each hyperparameter are set, and the Bayesian optimization algorithm is used to achieve adaptive optimization of the model hyperparameters to obtain the optimal hyperparameters of the SOC estimation model and the SOH estimation model. S111. Train the SOC estimation model and the SOH estimation model respectively: The SOC estimation training set data is input into the SOC estimation model in the form of the SOC estimation model input in S110 for training; then the trained model is tested using the SOC estimation test set data, and a threshold is set according to the accuracy requirements; if the test error is greater than the threshold, the process returns to S110 for retraining; if the test error is less than or equal to the threshold, the trained SOC estimation model is obtained. Similarly, the trained SOH estimation model is obtained; Step 2: Based on the trained SOC estimation model and SOH estimation model obtained in Step 1, perform online estimation of SOC and SOH; S201, Signal Monitoring: Using sensors to monitor the voltage, current, and temperature signals of the lithium battery; S202, Signal Acquisition: For data used for SOC estimation, the voltage, current, and temperature parameters are acquired at each time step in the cycle according to the required frequency; for data used for SOH estimation, the voltage, current, and temperature parameters are acquired at the required frequency after each charge-discharge cycle. S203. The charging and discharging data used for SOH estimation are sampled at the same length as the data length in S105; S204. The EMD noise reduction tool of S106 is used to decompose the collected voltage, current and temperature parameter data into components to achieve noise reduction. S205. Using the correlation analysis tool in S107, perform correlation analysis on all components obtained from the decomposition in S204 with SOC and SOH respectively, and obtain the component with the highest correlation; then use the component with the highest correlation to replace the parameter data in all cycles to obtain the estimated SOC data and estimated SOH data. S206. The SOC and SOH estimates from S205 are taken from the maximum value x in S108. max and minimum value x min Normalization is performed to obtain normalized SOC estimates and normalized SOH estimates; S207, SOH and SOC estimates: The normalized SOC estimation data is input into the trained SOC estimation model in S111 in the form of the SOC estimation model input in S110, and the output of the SOC estimation model is obtained; then the maximum value x in S108 is used. max and minimum value x min The output of the SOC estimation model is inversely normalized to obtain the SOC estimate. Similarly, the estimated value of SOH is obtained; S208. Smooth the SOC and SOH estimates respectively to filter out the instantaneous fluctuations of the SOC and SOH estimates, and obtain the smoothed SOC estimate for the current time step. m and the actual SOH value of the current charge / discharge cycle S209, SOH Update: The first SOH update uses the actual SOH value from the first charge / discharge cycle. Replace the initial SOH value; subsequent SOH updates will use the actual SOH value for the current charge / discharge cycle. Replace the actual SOH value from the previous charge / discharge cycle. To perform SOH updates; n represents the nth charge / discharge cycle; S210, SOC Update: First, based on the smoothed SOC estimate obtained in S208 for the current time step, SOC... m The true SOC value at the current time step is calculated using equation (11). Then perform a SOC update; the first SOC update uses the actual SOC value at the current time step. The smoothed estimate of SOC at the current time step obtained by replacing S208 m Subsequent SOC updates will be based on the actual SOC value at the current time step. Replace the previous time step SOC estimated true value 2. The method for jointly estimating the state of charge and state of health of a lithium battery based on deep learning according to claim 1, characterized in that, In step S102, the ambient temperature is -20℃, -10℃, 0℃, 10℃, 25℃ and 35℃ to cover the ambient temperature during the actual use of lithium batteries.

3. The method for jointly estimating the state of charge and state of health of a lithium battery based on deep learning according to claim 1, characterized in that, In step S102, the power configuration includes constant current charging, constant current discharging, DST driving cycle, US06 driving cycle, FUDS driving cycle, CLTC driving cycle, LA92 driving cycle, and driving cycle randomly mixed from the above cycles.

4. The method for jointly estimating the state of charge and state of health of a lithium battery based on deep learning according to claim 1, characterized in that, In step S106, the decomposition involves decomposing the parameter data into a series of intrinsic mode components (IMFs) and signal residual components, as shown below: In equation (5), x(t) represents the parameter data to be decomposed; imf i For the i-th IMF obtained from the decomposition; r k This refers to the residual component of the signal.

5. The method for jointly estimating the state of charge and state of health of a lithium battery based on deep learning according to claim 1, characterized in that, In step S107, the correlation analysis uses Pearson correlation analysis, and the calculation formula for Pearson correlation analysis is as follows: In equation (6), r represents the correlation coefficient; N is the data length; and X is the component obtained from the decomposition. Y represents the component average, where SOC or SOH data is used. The mean value of SOC or SOH is given; 0.0≤r<0.2 indicates no correlation, 0.2≤r≤0.4 indicates a weak correlation, 0.4<r≤0.6 indicates a moderate correlation, 0.6<r≤0.8 indicates a strong correlation, and 0.8<r≤1.0 indicates a very strong correlation.

6. The method for jointly estimating the state of charge and state of health of a lithium battery based on deep learning according to claim 1, characterized in that, In step S108, for the SOC estimation data, each parameter data is normalized to [0,1], as shown in the formula: In equation (7), x SOC For the normalized SOC estimation data, x min x is the minimum value of this parameter data. max This represents the maximum value of the parameter data. For the SOH estimation data, each parameter is normalized to [-1, 1], as shown in the formula: In equation (8), x SOH The data are normalized SOH estimates.

7. The method for jointly estimating the state of charge and state of health of a lithium battery based on deep learning according to claim 1, characterized in that, In step S110, the SOC estimation model and the SOH estimation model have the same structure, specifically: the first layer is a TCN layer used to extract data features; After the TCN layer, several LSTM layers are sequentially connected to learn the temporal relationships in the data. The LSTM layers are sequentially connected to each other. After the last LSTM layer, a fully connected layer with tanh as the activation function is sequentially connected, and then another fully connected layer is sequentially connected as the output layer. The optimizer is the Adam optimizer.

8. The method for jointly estimating the state of charge and state of health of a lithium battery based on deep learning according to claim 1, characterized in that, In step S110, the optimization evaluation equation for the TCN-LSTM neural network model is: In equation (9), y i With y i 'These are the actual value and the estimated value, respectively.

9. The method for jointly estimating the state of charge and state of health of a lithium battery based on deep learning according to claim 1, characterized in that, In step S110, the input to the SOC estimation model is the voltage, current, and temperature data for each step of each loop in the SOC estimation dataset, in the form of [V n ,I n ,T n The output is SOC; the input to the SOH estimation model is the voltage, current, and temperature data for each charge-discharge cycle in the SOH estimation dataset, in the form of [V1, V2, ..., V]. n ,I1,I2,…,I n ,T1,T2,…,T n The output is SOH.

10. The method for jointly estimating the state of charge and state of health of a lithium battery based on deep learning according to claim 1, characterized in that, In step S208, the smoothing process uses Kalman filtering, and the specific process is as follows: In equation (10), This represents the SOC estimate for the current time step or the SOH estimate for the current period. Let A be the SOC estimate of the previous time step or the SOH estimate of the previous cycle, and B be the state transition matrix and the control matrix. k-1 To control the input, P is the covariance between the smoothed SOC estimate and the SOC estimate, or between the actual SOH value and the SOH estimate. k-1 The covariance between the smoothed SOC estimate and the SOC estimate at the previous time step, or between the actual SOH value and the SOH estimate at the previous period, where Q is the process noise variance, and K is the variance. k Let H be the Kalman gain, H be the state-observation matrix, R be the observation noise variance, and z be the Kalman gain. k The observations are for the state matrix; the values ​​of Q and R are set according to the fluctuations of the estimated SOC or SOH values.

Images