Lithium ion battery residual life prediction method and model based on random error compensation

Abstract

The embodiment of the invention provides a lithium ion battery residual life prediction method and model based on random error compensation, and the method comprises the steps: obtaining a data set, carrying out the preprocessing of the data set, and dividing the preprocessed data set into a training set, a verification set and a test set; constructing a deep learning model based on a CNN-LSTM network architecture; training the deep learning model, and adjusting model parameters to obtain an optimized deep learning model and optimized model parameters; based on the optimized model parameters and the test set, performing performance evaluation on the optimized deep learning model to obtain a trained deep learning model and updated model parameters; and inputting a to-be-predicted lithium ion battery data set into the trained deep learning model, and predicting the state of charge, the remaining service life and the peak power of the lithium ion battery. According to the method, the limitation that a traditional neural network algorithm seriously depends on training data precision is relieved, and the service life of the power battery can be prolonged to the maximum extent.

Description

Technical Field

[0001] The present application relates to the technical field of battery remaining life prediction, and in particular to a method and model for predicting the remaining life of a lithium-ion battery based on random error compensation. Background Art

[0002] To solve the increasingly severe energy shortage and environmental pollution problems, the promotion and application of new energy electric vehicles has become a global development trend. Improving the performance of electric vehicle power battery management system (BMS) has become a key factor in ensuring the safety of electric vehicles and extending the life of lithium-ion batteries. The core functions of BMS include accurate estimation of battery state of charge (SOC), remaining useful life (RUL) and peak power state (SOP), providing guarantee for energy management and power distribution of the whole vehicle.

[0003] However, the above state estimation is affected by many complex factors such as discharge current, self-discharge effect, internal temperature, external ambient temperature, battery aging, etc., making it difficult to accurately estimate the state. The presence of random errors in the original data will interfere with the model's learning of the data, making it difficult for the model to accurately capture the key features and patterns in the data, thereby reducing the model's prediction performance. Secondly, during the prediction process, random errors will increase the uncertainty of the model's prediction results, thereby reducing the model's robustness and reliability. Therefore, obtaining accurate and reliable battery state estimation results remains a major challenge. Summary of the invention

[0004] The embodiment of the present application provides a lithium-ion battery remaining life prediction method and model based on random error compensation, which reduces the influence of approximate SOC value on peak power capacity estimation through the high robustness and efficiency of processing time series tasks of an adaptive extended Kalman filter (AEKF) model and a convolutional neural network and long short-term memory network (CNN-LSTM) algorithm.

[0005] To solve the above technical problems, in the first aspect, an embodiment of the present application provides a method for predicting the remaining life of a lithium-ion battery based on random error compensation, the method comprising the following steps: first, obtaining a data set, and preprocessing the data set, dividing the preprocessed data set into a training set, a validation set, and a test set; then, based on the CNN-LSTM network architecture, building a deep learning model; next, based on the training set and the validation set, training the deep learning model, adjusting the model parameters, and obtaining an optimized deep learning model and optimized model parameters; then, based on the optimized model parameters and the test set, performing a performance evaluation on the optimized deep learning model, and obtaining a trained deep learning model. model and updated model parameters; finally, the lithium-ion battery data set to be predicted is input into the trained deep learning model to predict the state of charge, remaining service life and peak power of the lithium-ion battery; wherein, predicting the peak power of the lithium-ion battery includes: based on the updated model parameters and the predicted state of charge of the lithium-ion battery, using constraints to realize peak current estimation based on the battery model and peak current estimation based on the battery state of charge; combining the peak current estimation result based on the battery model and the peak current estimation result based on the battery state of charge, deriving the estimated value of the battery's continuous peak current under multiple constraints, and obtaining the estimated value of the battery's continuous dynamic peak power under multiple constraints.

[0006] In some exemplary embodiments, the data set is an electric vehicle driving condition data set including current, voltage and temperature data, and the current, voltage and temperature data are collected by sensors of the electric vehicle lithium-ion battery management system; the data set is preprocessed, including: based on the data set, a battery model is established and parameter identification is performed; an adaptive extended Kalman filter model is constructed, and the data collected by the sensor is preprocessed based on the adaptive extended Kalman filter model to remove random errors; the preprocessing process includes: extracting characteristic variables related to battery discharge in a driving cycle from the data set, fusing and randomly arranging the characteristic variables to form an input data set; the characteristic variables include power battery current, voltage, capacity and temperature.

[0007] In some exemplary embodiments, a battery model is established based on a data set and parameter identification is performed, including: in order to track the dynamic characteristics of the battery and accurately reflect the various working states of the battery, a second-order resistance-capacitance circuit model, i.e., a dual-polarization model, is established according to the data set; based on the dual-polarization model and the current and voltage data in the data set, a forgetting factor recursive least squares method is used to perform parameter identification; the parameter identification process includes: performing a dynamic stress test on the battery at 25°C and recording the sensor signal every 1 second; the sensor signal includes a current signal and a voltage signal.

[0008] In some exemplary embodiments, in the dual polarization model, the relationship between the open circuit voltage and the battery state of charge is fitted and the forgetting factor recursive least squares method is used for parameter identification; the relationship between the open circuit voltage and the battery state of charge is represented by a monotonically increasing function, and the monotonically increasing function is:

[0009] u oc (soc) = α1 × soc 8 +α2×soc 7 +α3×soc 6 +α4×soc 5 +α5×soc 4 +α6×soc 3 +

[0010] α7×soc 2 +α8×soc 1 +α9

[0011] Where, the voltage source u oc It represents the open circuit voltage related to the battery state of charge SOC, and α is the electrochemical coefficient.

[0012] In some exemplary embodiments, an adaptive extended Kalman filter model is constructed according to the parameter identification results and the data set type, and the filter parameters of the model are determined, and its state equation is as follows:

[0013] x k =f(x k-1 ,u k-1 )+ω k-1

[0014] y k =h(x k -u k )+v k

[0015] Where x is the time t k,l is the system state vector, u is the system input vector, t is the fixed sampling interval between two adjacent measurement points, k and l are the time scale indicators of the macroscopic system parameters and the microscopic system state, L represents the time scale separation level, x k,0 =x k-1,L ; y is the system output vector, ω is the system noise, and v is the actual measurement noise; ω and v are independent of each other; the state estimation and covariance estimation are corrected using real-time measurement values, and the results are expressed as and The measurement update equation of the adaptive extended Kalman filter is expressed as:

[0016] d k =u OC,k +Ck x k -i k R0

[0017]

[0018] C k =[-1 -1 0]

[0019] Among them, d k is the terminal voltage under the posterior state value, e k is the terminal voltage error under the posterior state value; the system state and error covariance correction is:

[0020]

[0021] Where K is the Kalman gain.

[0022] In some exemplary embodiments, a deep learning model is constructed based on a CNN-LSTM network architecture, and the deep learning model is trained based on a training set and a validation set, and the model parameters are adjusted to obtain an optimized deep learning model and optimized model parameters, including: constructing a CNN network main structure; wherein the CNN network main structure includes a convolution layer, a pooling layer, a normalization layer, and a fully connected layer; constructing an LSTM network structure: based on the LSTM network structure, a recurrent connection is introduced into the CNN network, so that the CNN network can capture the time dependency in the sequence, and the training set data is input into the CNN network main structure for training; adjusting the model parameters, optimizing the model, and obtaining an optimized deep learning model and optimized model parameters based on the minimum loss value obtained by inputting the validation set.

[0023] In some exemplary embodiments, based on the multiple constraints of battery design current, voltage, and power, and the operating range limit of the state of charge of the electric vehicle control system, a multi-constraint dynamic battery peak power capability estimation method is established to predict the peak power of the lithium-ion battery; wherein, the multi-constraint dynamic battery peak power capability estimation method is established, including: first, based on the maximum and minimum limits of the battery state of charge, an expression of the peak current of the battery is obtained; then, a terminal voltage constraint is set; finally, based on the expression of the peak current of the battery, the terminal voltage constraint, and the maximum discharge current and minimum charging current specified when the battery leaves the factory, a continuous dynamic peak power estimation expression under multiple constraints is obtained; the peak current of the battery at time l is:

[0024]

[0025] Among them, Z max and Z min are the maximum and minimum values ​​of the battery state of charge, respectively, and Q is the maximum capacity;

[0026] The terminal voltage constraints are:

[0027]

[0028] The continuous dynamic peak power estimation expression under multiple constraints is as follows:

[0029]

[0030] in, and are the estimated values ​​of the continuous discharge peak current and continuous charge peak current at time l, P max and P min are the maximum power and minimum power set when the power battery is designed; and are the peak discharge power and peak charging power estimated by the dynamic continuous peak power estimation method under multiple constraints at time l; ΔL is the length of the continuous sampling time; i max and i min They are the maximum discharge current and minimum charge current specified when the battery leaves the factory.

[0031] In some exemplary embodiments, the data sets include: a CALCE data set, an Oxford data set, and a NASA data set; the CALCE data set includes charging, discharging, and measured resistance information of four 18650LiCoO2 lithium batteries, namely CS35, CS36, CS37, and CS38; the Oxford data set includes battery aging measurement data collected from eight small lithium-ion batteries; the NASA data set includes a test data set of charging, discharging, and impedance of four lithium-ion batteries in three different operating modes; the data sets of the four lithium-ion batteries are B0005, B0006, B0007, and B0018, respectively.

[0032] In some exemplary embodiments, the preprocessed data set is divided into a training set, a validation set, and a test set, including: in the CALCE data set, batteries CS35, CS36, and CS37 are selected as training sets, and CS38 is selected as a test set; in the Oxford data set, the measurement data of batteries with serial numbers 1, 3, and 7 are selected as training sets, and the measurement data of batteries with serial numbers 2, 4, 6, and 8 are selected as test sets; in the NASA data set, data sets B0005, B0006, and B0007 are selected as training sets, and B0018 is selected as a test set.

[0033] In the second aspect, the embodiment of the present application also provides a lithium-ion battery remaining life prediction model based on random error compensation, which uses the lithium-ion battery remaining life prediction method based on random error compensation as in the above embodiment for prediction, including: a data set module, a deep learning model module, a model training module, a model optimization module and a prediction module connected in sequence; wherein the data set module is used to obtain a data set, and preprocess the data set, and divide the preprocessed data set into a training set, a validation set and a test set; the deep learning model module is used to construct a deep learning model according to the network architecture of CNN-LSTM; the model training module is used to train the deep learning model according to the training set and the validation set, adjust the model parameters, and obtain an optimized deep learning model and optimized model parameters; the model optimization module is used to perform performance evaluation on the optimized deep learning model according to the optimized model parameters and the test set, and obtain a trained deep learning model and updated model parameters; the prediction module is used to calculate the lithium-ion battery data to be predicted. The data set is input into the trained deep learning model to predict the state of charge, remaining service life and peak power of the lithium-ion battery; wherein the prediction module includes a battery state of charge prediction module, a battery remaining service life prediction module and a battery peak power prediction module; the battery state of charge prediction module is used to predict the state of charge of the lithium-ion battery; the battery remaining service life prediction module is used to predict the remaining service life of the lithium-ion battery; the battery peak power prediction module includes a peak current estimation module and a continuous dynamic peak power estimation module; the peak current estimation module is used to realize the peak current estimation based on the battery model and the peak current estimation based on the battery state of charge according to the updated model parameters and the predicted lithium-ion battery state of charge by using constraints; the continuous dynamic peak power estimation module is used to combine the peak current estimation results based on the battery model and the peak current estimation results based on the battery state of charge, derive the estimated value of the battery continuous peak current under multiple constraints, and obtain the estimated value of the battery continuous dynamic peak power under multiple constraints.

[0034] The technical solution provided by the embodiments of the present application has at least the following advantages:

[0035] The embodiment of the present application provides a method and model for predicting the remaining life of a lithium-ion battery based on random error compensation. The method comprises the following steps: first, obtaining a data set, preprocessing the data set, and dividing the preprocessed data set into a training set, a validation set, and a test set; then, building a deep learning model based on the CNN-LSTM network architecture; next, training the deep learning model based on the training set and the validation set, adjusting the model parameters, and obtaining an optimized deep learning model and optimized model parameters; then, based on the optimized model parameters and the test set, performing a performance evaluation on the optimized deep learning model, and obtaining a trained deep learning model and an updated model. model parameters; finally, the lithium-ion battery data set to be predicted is input into the trained deep learning model to predict the state of charge, remaining service life and peak power of the lithium-ion battery; wherein, predicting the peak power of the lithium-ion battery includes: based on the updated model parameters and the predicted state of charge of the lithium-ion battery, using constraints to realize peak current estimation based on the battery model and peak current estimation based on the battery state of charge; combining the peak current estimation results based on the battery model and the peak current estimation results based on the battery state of charge, deriving the estimated value of the battery's continuous peak current under multiple constraints, and obtaining the estimated value of the battery's continuous dynamic peak power under multiple constraints.

[0036] This application proposes a method and model for predicting the remaining life of lithium-ion batteries based on random error compensation. First of all, the estimation of SOC and SOP is crucial for BMS to predict its remaining life. BMS can monitor the SOC and SOP of the battery in real time to determine the current state of the battery and adjust the charging and discharging strategies accordingly, thereby extending the service life of the battery. More accurate SOC and SOP estimates can provide better input parameters for the battery life prediction model. Therefore, this application adopts the AEKF-CNN-LSTM combined model for training and prediction, aiming to minimize random errors through three different data sets and verify the accuracy and generalization ability of the model. Experimental results show that this method achieves high-precision estimation of battery SOC and SOP at different temperatures, as well as accurate prediction of battery life under different types and discharge rates. In addition, the combination of the AEKF algorithm and the CNN-LSTM network adapts to the influence of random sensor errors and enhances the generalization ability of the algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] One or more embodiments are exemplarily described by pictures in the corresponding drawings, and these exemplifications do not constitute limitations on the embodiments. Unless otherwise stated, the pictures in the drawings do not constitute proportional limitations.

[0038] Figure 1 A schematic flow chart of a method for predicting the remaining life of a lithium-ion battery based on random error compensation provided in an embodiment of the present application.

[0039] Figure 2 A second-order resistance-capacitance (RC) circuit model of a lithium-ion battery remaining life prediction method based on random error compensation is provided in an embodiment of the present application.

[0040] Figure 3 A network architecture diagram of the LSTM provided in the embodiment of the present application.

[0041] Figure 4 A structural diagram of a multifunctional prediction model based on AEKF-CNN-LSTM provided in an embodiment of the present application.

[0042] Figure 5 A schematic diagram of the structure of a lithium-ion battery remaining life prediction model based on random error compensation provided in an embodiment of the present application.

[0043] FIG. 6A to FIG. 6D A schematic diagram of the prediction results of SOC, SOP and RUL for the lithium-ion battery remaining life prediction model based on random error compensation provided in an embodiment of the present application. DETAILED DESCRIPTION

[0044] As can be seen from the background technology, the presence of random errors in the original data will interfere with the model's learning of the data, making it difficult for the model to accurately capture the key features and patterns in the data, thereby reducing the model's prediction performance. Secondly, in the prediction process, random errors will increase the uncertainty of the model's prediction results, thereby reducing the robustness and reliability of the model. Therefore, in order to reduce the impact of random errors, methods such as multi-model fusion and noise filtering are needed.

[0045] The core functions of BMS include accurately estimating battery SOC, RUL and SOP to provide guarantee for energy management and power distribution of the whole vehicle. However, the above state quantity estimation is affected by many complex factors such as discharge current, self-discharge effect, internal temperature, external ambient temperature, battery aging, etc., making it difficult to accurately estimate the state. Among them, the measurement of current and voltage is largely affected by the accuracy of the sensor, resulting in inaccurate state estimation.

[0046] In view of the deficiencies in the prior art, the embodiment of the present application provides a method and model for predicting the remaining life of a lithium-ion battery based on random error compensation. The method comprises the following steps: first, obtaining a data set, preprocessing the data set, and dividing the preprocessed data set into a training set, a validation set, and a test set; then, building a deep learning model based on the CNN-LSTM network architecture; next, training the deep learning model based on the training set and the validation set, adjusting the model parameters, and obtaining an optimized deep learning model and optimized model parameters; then, based on the optimized model parameters and the test set, performing a performance evaluation on the optimized deep learning model, and obtaining a trained deep learning model. Model and updated model parameters; finally, the lithium-ion battery data set to be predicted is input into the trained deep learning model to predict the lithium-ion battery state of charge, remaining service life and peak power; wherein, predicting the peak power of the lithium-ion battery includes: based on the updated model parameters and the predicted lithium-ion battery state of charge, using constraints to achieve peak current estimation based on the battery model and peak current estimation based on the battery state of charge; combining the peak current estimation results based on the battery model and the peak current estimation results based on the battery state of charge, deriving an estimated value of the battery's continuous peak current under multiple constraints, and obtaining an estimated value of the battery's continuous dynamic peak power under multiple constraints. The lithium-ion battery remaining life prediction method and model based on random error compensation provided in this application reduces the impact of approximate SOC values ​​on peak power capacity estimation through the high robustness of the AEKF-CNN-LSTM algorithm and the efficiency of processing time series tasks.

[0047] The following will describe the various embodiments of the present application in detail with reference to the accompanying drawings. However, it will be appreciated by those skilled in the art that in the various embodiments of the present application, many technical details are provided in order to enable the reader to better understand the present application. However, even without these technical details and various changes and modifications based on the following embodiments, the technical solution claimed in the present application can be implemented.

[0048] See also Figure 1 The present application provides a method for predicting the remaining life of a lithium-ion battery based on random error compensation, the method comprising the following steps:

[0049] Step S101: obtain a data set, preprocess the data set, and divide the preprocessed data set into a training set, a validation set, and a test set.

[0050] Step S102: construct a deep learning model based on the CNN-LSTM network architecture.

[0051] Step S103: Based on the training set and the validation set, the deep learning model is trained, and the model parameters are adjusted to obtain an optimized deep learning model and optimized model parameters.

[0052] Step S104: Based on the optimized model parameters and the test set, a performance evaluation is performed on the optimized deep learning model to obtain a trained deep learning model and updated model parameters.

[0053] Step S105: input the lithium-ion battery data set to be predicted into the trained deep learning model to predict the state of charge, remaining service life and peak power of the lithium-ion battery.

[0054] Wherein, in step S105, predicting the peak power of the lithium-ion battery includes:

[0055] Step S1051 : Based on the updated model parameters and the predicted state of charge of the lithium-ion battery, peak current estimation based on the battery model and peak current estimation based on the battery state of charge are implemented using constraint conditions.

[0056] Step S1052: Combine the peak current estimation result based on the battery model and the peak current estimation result based on the battery state of charge to derive an estimated value of the battery's continuous peak current under multiple constraints, and obtain an estimated value of the battery's continuous dynamic peak power under multiple constraints.

[0057] In some embodiments, the data set in step S101 is an electric vehicle driving condition data set including current, voltage and temperature data, and the current, voltage and temperature data are collected by sensors of the electric vehicle lithium-ion battery management system. In step S101, the data set is preprocessed, including:

[0058] Step S1011: Based on the data set, a battery model is established and parameter identification is performed.

[0059] Step S1012: construct an adaptive extended Kalman filter model, and pre-process the data collected by the sensor based on the adaptive extended Kalman filter model to remove random errors.

[0060] In step S1012, the preprocessing process includes: extracting characteristic variables related to battery discharge during the driving cycle from the data set, fusing and randomly arranging the characteristic variables to form an input data set; the characteristic variables include power battery current, voltage, capacity and temperature.

[0061] In some embodiments, in step S1011, a battery model is established based on the data set and parameter identification is performed, including:

[0062] Step S1011a: In order to track the dynamic characteristics of the battery and accurately reflect the various working states of the battery, a second-order resistance-capacitance circuit model, namely, a dual-polarization model, is established according to the data set.

[0063] Step S1011b: perform parameter identification using the Forgetting Factor Recursive Least Squares (FFRLS) method based on the dual polarization model and the current and voltage data in the data set.

[0064] The parameter identification process in step S1011b includes: performing a dynamic stress test on the battery at 25°C and recording a sensor signal every 1 second; the sensor signal includes a current signal and a voltage signal.

[0065] Because the chemical reaction inside the battery is very complicated, and the external characteristics are highly nonlinear. Therefore, in order to track the dynamic characteristics of the battery and accurately reflect its various working states, this application adopts a second-order resistor-capacitor (RC) circuit model, i.e., a dual polarization model. The model uses a voltage source to represent the thermodynamic equilibrium of the battery and uses an RC network to describe the dynamic characteristics of the battery. In addition, BMS can monitor the SOC and SOP of the battery in real time to determine the current state of the battery and adjust the charging and discharging strategies accordingly, thereby extending the service life of the battery. More accurate SOC and SOP estimates can provide better input parameters for the battery life prediction model. Therefore, this application uses the AEKF-CNN-LSTM combined model for training and prediction, and all the three most important parameters of the power battery (battery design current, voltage, power) are taken into account, effectively improving the reliability of the estimation results.

[0066] Specifically, in step S1011, in order to track the dynamic characteristics of the battery and accurately reflect its various working states, a dual polarization equivalent circuit model is established, and the significance of the key system parameters is explained. Afterwards, the relationship between the calibrated open circuit voltage OCV and the battery state of charge SOC is fitted to identify the unknown parameters of the system using the forgetting factor recursive least squares method. Afterwards, by establishing a general dual estimation method for battery parameters and states, the system matrix and the parameter matrix are connected through the time parameter matrix, which reduces the influence of random noise while improving the computational efficiency, laying the foundation for the subsequent input into the time domain convolutional network as a training set for learning. Then the training set, validation set and test set are divided, and the cross-validation method is used. In order to simulate the real driving conditions as much as possible, this application uses mixed operating condition data collected from the New European Driving Cycle (NEDC) and Urban Urban Road Cycle (UDDS) conditions as training data. The test data must be different from the training data to prove the generalization ability of the AEKF-CNN-LSTM network. Therefore, this application uses the dynamic stress test (DST), the Japanese driving cycle (JDC), and the latest world light vehicle test cycle (WLTC) as test data. Training and test data are collected at 10°C, 25°C, and 40°C to improve the versatility of the model. In each experiment, the subject went through a complete charge and discharge cycle, in which the test curve was applied to simulate the actual discharge process. During the test, the sensor signals, including current, voltage, and temperature, were recorded every 1s.

[0067] It should be noted that in step S1011a, for the second-order resistance-capacitance model, the open circuit voltage OCV is a high-value capacitor or voltage source. The rest of the model includes the contact resistance of the battery and two other RC pairs, which describe different battery kinetic characteristics, namely double-layer kinetics and diffusion kinetics. Related experiments show that the terminal voltage of the equivalent circuit model is close to the actual battery model, and the second-order model can fit the experimental data well under all test conditions. In addition, compared with the time constant of battery dynamics, the model parameters change slowly and can be regarded as a local constant that changes gradually over time. Therefore, the dependence of the model parameters on temperature, SOC and battery degradation can be handled by the forgetting factor.

[0068] In some embodiments, the Forgetting Factor Recursive Least Squares (FFRLS) theory is a commonly used model parameter identification and data mining method. When the system model and parameters vary greatly and are easily affected by the uncertain application environment, the recursive least squares method overcomes the uncertainty of the model parameters by regularly correcting and updating the parameters and accurately captures the real-time characteristics of the system.

[0069] In some embodiments, in the dual polarization model, FFRLS is used for parameter identification by fitting the relationship between the calibrated open circuit voltage and the battery state of charge; the relationship between the open circuit voltage and the battery state of charge is represented by a monotonically increasing function. According to the basic characteristics of the power battery, OCV and SOC always maintain a monotonically increasing relationship. Therefore, the SOC estimation error can be corrected by using the mapping relationship between OCV and SOC. Through the monotonically increasing function of OCV and SOC, the monotonic relationship between the battery terminal voltage and SOC can be obtained, which improves the optimization efficiency and convergence speed of the filter state feedback regulation; the monotonically increasing function of OCV and SOC is:

[0070] u oc (soc) = α1 × soc 8 +α2×soc 7 +α3×soc 6 +α4×soc 5 +α5×soc 4 +α6×soc 3 +α7

[0071] ×soc 2 +α8×soc 1 +α9

[0072] Where, the voltage source u oc It represents the open circuit voltage related to the battery state of charge SOC, and α is the electrochemical coefficient.

[0073] Afterwards, the battery pulse discharge data is used to verify the accuracy of the model and validate the parameter identification results.

[0074] For the state of the state space model based on Gaussian white noise system, Kalman filtering technology can obtain the most accurate estimation results. The standard EKF treats both the process noise covariance and the measurement noise covariance as constants, which is highly inconsistent with the actual situation. Therefore, the AEKF with adaptive update process and measurement noise covariance matrix is ​​used to improve the accuracy of the algorithm.

[0075] In some embodiments, based on the parameter identification results and the data set type, an adaptive extended Kalman filter model is constructed, and the filter parameters of the model are determined. The specific steps include:

[0076] Step S10121: For any nonlinear discrete system, f(x k-1 ,u k-1 ) is the system state equation function, h(x k ,u k ) is the observation equation function of the system. The general form of the state equation and observation equation is as follows:

[0077]

[0078] In the formula, x is the state vector of the n-dimensional system, u is the input vector of the r-dimensional system, y is the output vector (observation value) of the m-dimensional system, ω is the white noise of the system with a mean of 0 and a covariance of Q k , v has a mean of 0 and a variance of R k The measured white noise sequence, ω and v are independent of each other.

[0079] Step S10122, then, f(x k ,u k ) and h(x k ,u k ) performs a first-order Taylor expansion linearization at each moment and defines and On this basis, the optimal state estimation process of nonlinear discrete systems is established using the AEKF algorithm:

[0080]

[0081] Step S10123: Initialize and set the initial value of the state observation.

[0082] The a priori estimation is performed in step S10121, including:

[0083] Step S10121a: For k=1, 2, ..., perform the following a priori estimation operation to calculate the state estimation and covariance estimation from the previous time (k-1) to the current time (k). The specific expression is as follows:

[0084]

[0085] in,

[0086] Step S10121b, perform a posteriori estimation and correction.

[0087] Using real-time measurement values k The state estimate and covariance estimate are corrected. The results are expressed as and The measurement update equation of the adaptive extended Kalman filter is expressed as:

[0088] Update sequence:

[0089]

[0090] Where, d k is the terminal voltage under the a posteriori state value, e k is the terminal voltage error under the a posteriori state value.

[0091] The derivative of the terminal voltage with respect to the state value [u1 u2 SOC]:

[0092]

[0093] Kalman gain:

[0094]

[0095] System state and error covariance correction:

[0096]

[0097] In some embodiments, in step S102, building a deep learning model based on the network architecture of CNN-LSTM includes the following steps:

[0098] Step S1021, construct the main network framework of the convolutional neural network (CNN).

[0099] The main network structure of CNN includes convolutional layers, pooling layers, normalization layers and fully connected layers; CNN can visualize the features learned by the neural network to help understand the prediction process of the model, and then combine LSTM to achieve high-precision SOC estimation. In terms of RUL prediction, CNN can learn battery characteristics and identify battery degradation patterns, thereby achieving high-precision life prediction.

[0100] Step S1022: construct a network architecture of a long short-term memory neural network (LSTM).

[0101] The main purpose of LSTM is to solve the gradient vanishing and gradient exploding problems in recurrent neural networks (RNNs), as well as to handle long-term dependencies. It selectively passes information into the network using a mechanism called a "gate" to control the flow of information. The principle of the LSTM recurrent layer is to save information for later use to prevent the early information from disappearing as the number of layers gradually increases during the processing of the sequence, which can effectively alleviate the problem of gradient vanishing. In short, the problem of gradient vanishing is solved by allowing past information to re-enter the cycle later.

[0102] Step S1023, based on the LSTM network architecture constructed in step S1022, first, use AEKF to denoise the raw data containing random errors observed in each cycle, and input the processed results into the CNN-LSTM network to predict the final results. The "processing results" use two layers (16 and 16 hidden layer neurons, respectively) of independent networks, trained on the data for more than 100 time steps, where the kernel size is 3×1, the number of convolution kernels is 256, and the expansion factor is [1,2,4,8,16,32,64]. In the GPU environment, the calculation time for each step is about 0.24ms. Adjust the model parameters to achieve the optimal model, and obtain the deep learning model based on the lowest loss value obtained by the input validation set.

[0103] In some embodiments, the data sets of the present application include: CALCE data set, Oxford data set and NASA data set; CALCE data set includes charging, discharging and measured resistance information of four 18650LiCoO2 lithium batteries, namely CS35, CS36, CS37 and CS38. In the experiment, all batteries were charged in constant current and constant voltage modes. The Oxford data set includes battery aging measurement data collected from eight small lithium-ion batteries; the NASA data set includes a test data set of charging, discharging and impedance of four lithium-ion batteries in three different operating modes; the data sets of the four lithium-ion batteries are B0005, B0006, B0007 and B0018 respectively.

[0104] Specifically, this application uses multiple data sets to evaluate the estimation performance of the proposed method. First, the accuracy of battery voltage and state of charge estimation in the joint estimation algorithm is analyzed under dynamic stress test conditions at different temperatures. Secondly, based on SOC estimation under different discharge durations, the effect of temperature on SOP estimation is verified. Finally, using the denoised training data, three independent data sets are used to predict the remaining service life of lithium-ion batteries. All training processes are performed on a computer with a single GeForce RTX 4080 GPU.

[0105] In some embodiments, preprocessing the time series of current, voltage, and capacity usage by AEKF can provide more accurate input parameters for the CNN-LSTM prediction model. This enables the model to learn more accurate state information and predict the next value based on the operating conditions of the battery, including: In the training phase, three different data sets are used. For each data set, a portion of the battery data is extracted as a training set, and the remaining battery data is used as a test set for the model to provide a prediction of the remaining percentage. In the CALCE data set, batteries CS35, CS36, and CS37 are selected as training sets, and CS38 is used as a test set. In the Oxford data set, the measurement data of battery numbers 1, 3, and 7 are selected as training sets, and the measurement data of battery numbers 2, 4, 6, and 8 are selected as test sets. In the NASA data set, data sets B0005, B0006, and B0007 are selected as training sets, and B0018 is selected as a test set.

[0106] The model was subsequently used as a predictive tool to estimate the RUL of lithium-ion batteries. Ongoing predictions were performed using a test set by generating a series of capacity values ​​from the current cycle to the end-of-life threshold. The remaining cycle count, or RUL, was calculated by determining the number of cycles between the current cycle and the point at which the battery reaches its life limit. The life threshold was defined as the point at which the capacity drops to 80% of its original value, and the number of cycles remaining until this point was predicted.

[0107] In some embodiments, in step S105, when predicting the peak power of a lithium-ion battery, a multi-constraint dynamic battery peak power capability estimation method is established based on multiple constraints of battery design current, voltage, and power, and the operating range limit of the state of charge of the electric vehicle control system to predict the peak power of the lithium-ion battery.

[0108] Among them, a multi-constraint dynamic battery peak power capability estimation method is established, including: first, based on the maximum and minimum limits of the battery state of charge, an expression for the peak current of the battery is obtained; then, a terminal voltage constraint is set; finally, based on the expression for the peak current of the battery, the terminal voltage constraint, and the maximum discharge current and minimum charge current specified when the battery leaves the factory, an expression for continuous dynamic peak power estimation under multi-constraint conditions is obtained.

[0109] In some embodiments, in step S1051, based on the updated model parameters and the more accurate SOC estimation result, the peak current estimation based on the battery model and the peak current estimation based on the battery SOC are realized by using the constraint conditions. Then, in step S1052, the two are combined to derive the estimated value of the continuous peak current under the battery's multiple constraint conditions, and then the SOP estimated value of the battery under the multiple constraint conditions is obtained.

[0110] Among them, the terminal voltage constraint is:

[0111]

[0112] Where Q is the maximum capacity of the battery, Z k,l t k,l The polarization voltage and state of charge of the battery at all times.

[0113] The maximum current that the battery can withstand under current conditions depends on the maximum and minimum limits of the battery SOC. According to the SOC constraint, the peak current of the battery at time l is:

[0114]

[0115] Where Z max and Z min are the maximum and minimum values ​​of the battery state of charge, respectively.

[0116] In some embodiments, this method ensures the safety of the battery in the actual charging and discharging process to a certain extent, taking into account the peak current of the battery within the safe operating time. However, when the design and operating range of the battery SOC is wide, using only this method as a constraint can easily lead to an overly large peak current estimation result. Therefore, this method should be combined with the peak current estimation method of voltage constraints to achieve better results.

[0117] Based on the peak current expression of the battery, the terminal voltage constraint, and the maximum discharge current and minimum charge current (maximum discharge current i max , minimum charging current i min ) The multi-constrained dynamic peak current expression obtained is as follows:

[0118]

[0119] Combining the three constraints, we can get the expression of peak power SOP under multiple constraints:

[0120]

[0121] in, and are the estimated values ​​of the continuous discharge peak current and continuous charge peak current at time l, P max and P min are the maximum power and minimum power set when the power battery is designed. and are the peak discharge power and peak charging power estimated by the dynamic continuous peak power estimation method under multiple constraints at time l, and ΔL is the length of the continuous sampling time.

[0122] Specifically, when the terminal voltage of the battery is high or low, it will affect the size of the real-time peak power accordingly. But at the same time, the change of SOC will also affect the voltage and capacity of the battery. Therefore, through the corresponding coupling relationship and combined with the rated current limit of the power battery itself, a joint estimation method of terminal voltage, SOC and rated current can be established to obtain a more accurate SOP estimation result.

[0123] See also Figure 2 In order to track the dynamic characteristics of the battery and accurately reflect its various working states, the embodiment of the present application adopts a second-order resistance-capacitance circuit model, namely a dual-polarization model. The model uses a voltage source to represent the thermodynamic equilibrium of the battery and an RC network to describe the dynamic characteristics of the battery. The model consists of three parts. (1) Voltage source u oc Indicates the open circuit voltage (OCV) related to SOC. (2) Ohmic internal resistance R0 represents the battery electrode material, electrolyte, diaphragm resistance and contact resistance of each component. (3) RC network: I t is the working current, which is positive during discharge and negative during charge; u T is the terminal voltage. Symbols R1 and C1 represent electrochemical polarization resistance and capacitance respectively. R2 and C2 are the second polarization resistance and capacitance respectively. They are responsible for describing the dynamic characteristics of the battery, including polarization characteristics and diffusion effect. Polarization characteristics refer to the phenomenon that the voltage of the battery presents a nonlinear change due to the potential difference generated inside the battery during the charge and discharge process of the battery due to factors such as chemical reactions and changes in electrolyte concentration. This nonlinear change will affect the performance of the battery, such as the discharge capacity and cycle life of the battery. The polarization characteristics of lithium-ion batteries mainly include two types: electrochemical polarization and concentration polarization. Electrochemical polarization is polarization caused by electrochemical reactions generated on the electrode surface, mainly including overpotential, activation polarization and concentration polarization. Concentration polarization is polarization caused by changes in ion concentration in the electrolyte. The diffusion effect refers to the fact that in lithium-ion batteries, lithium ions diffuse from the positive electrode through the electrolyte to the negative electrode, and are embedded in the negative electrode material at the negative electrode to form lithium intercalation compounds. The battery releases electrons during the discharge process to complete the electrochemical reaction. During the charging process, lithium ions are deintercalated from the negative electrode material and diffuse to the positive electrode through the electrolyte. The diffusion effect of lithium-ion batteries is affected by many factors, such as the pore size and porosity of the electrode material, the ion concentration in the electrolyte, the temperature of the battery, etc. The adverse effects of the diffusion effect will manifest as problems such as reduced discharge capacity of the battery, increased internal resistance, and shortened cycle life. In order to reduce the impact of the diffusion effect, a variety of measures can be taken, such as optimizing the pore structure of the electrode material, improving the electrolyte formulation, and controlling the temperature of the battery. The voltages u1 and u2 represent the voltages of the electrochemical capacitor C1 and the polarization capacitor C2, respectively.

[0124] Specifically, in this model, OCV is a high-value capacitor or voltage source. The rest of the model includes the contact resistance of the battery and two other RC pairs, which describe different battery dynamics characteristics, namely double-layer dynamics and diffusion dynamics. Related experiments show that the terminal voltage of this equivalent circuit model is close to the actual battery model, and the second-order model can fit the experimental data well under all test conditions. In addition, compared with the time constant of battery dynamics, the model parameters change slowly and can be regarded as a local constant that changes gradually over time. Therefore, the dependence of the model parameters on temperature, SOC and battery degradation can be handled by the forgetting factor.

[0125] According to Kirchhoff's law and Laplace transform, the mathematical relationship between the battery output voltage and input current is as follows:

[0126]

[0127] The output equation is:

[0128] u T =u oc (SOC)-u1-u2-R0I T

[0129] See also Figure 3 and Figure 4 , the embodiment of the present application provides an LSTM network architecture diagram and a sliding window arrangement diagram. The middle part, including 5 blocks and 4 main lines, shows the steps that each neuron cycle in the LSTM layer must go through. Among them,

[0130]

[0131] Where W represents the weight. Since the previous memory should be forgotten, F t It determines how much information can be retained after the previous cycle. The last information will only be retained when the door is opened. g The input gate I determines the percentage of information retained. The closer its value is to 1, the more previous information is retained. t Determines the new input information to be memorized. Memory Gate C t , together with the previous input gate and forget gate, determines how much memory should be retained. Each gated layer is actually a fully connected layer. The rest of the LSTM unit uses the tanh function as the activation function because its second-order derivative can last for a long time before going to 0, which can effectively solve the problem of gradient disappearance. Finally, the output gate O t Determines the final output value of the next cycle, where h t Represents the new value obtained after each loop operation.

[0132] The input data set can be obtained by Figure 4 The sliding windows shown are arranged.

[0133] In this method, AEKF is used to eliminate random errors in the raw data collected in each cycle. The cleaned data is then fed into the CNN-LSTM network for final prediction, and the model parameters are adjusted to achieve the optimal model. The best trained deep learning model is selected based on the lowest loss value obtained by inputting the validation set.

[0134] See also Figure 5 The present application also provides a lithium-ion battery remaining life prediction model based on random error compensation, which uses the lithium-ion battery remaining life prediction method based on random error compensation of the above embodiment for prediction, including: a data set module 101, a deep learning model module 102, a model training module 103, a model optimization module 104 and a prediction module 105 connected in sequence; wherein the data set module 101 is used to obtain a data set, and pre-process the data set, and divide the pre-processed data set into a training set, a validation set and a test set; the deep learning model module 102 is used to construct a deep learning model according to the network architecture of CNN-LSTM; the model training module 103 is used to train the deep learning model according to the training set and the validation set, adjust the model parameters, and obtain an optimal prediction module 105. The deep learning model and the optimized model parameters are obtained by the training module 105. The model optimization module 104 is used to perform performance evaluation on the optimized deep learning model according to the optimized model parameters and the test set, so as to obtain the trained deep learning model and the updated model parameters. The prediction module 105 is used to input the lithium-ion battery data set to be predicted into the trained deep learning model, and predict the state of charge, remaining service life and peak power of the lithium-ion battery. The prediction module 105 includes a battery state of charge prediction module 1051, a battery remaining service life prediction module 1052 and a battery peak power prediction module 1053. The battery state of charge prediction module 1051 is used to predict the state of charge of the lithium-ion battery. The battery remaining service life prediction module 1052 is used to predict the remaining service life of the lithium-ion battery. The battery peak power prediction module 1053 includes a peak current estimation module and a continuous dynamic peak power estimation module; the peak current estimation module is used to realize peak current estimation based on the battery model and peak current estimation based on the battery state of charge according to the updated model parameters and the predicted lithium-ion battery state of charge using constraints; the continuous dynamic peak power estimation module is used to combine the peak current estimation results based on the battery model and the peak current estimation results based on the battery state of charge, derive the estimated value of the battery's continuous peak current under multiple constraints, and obtain the estimated value of the battery's continuous dynamic peak power under multiple constraints.

[0135] The lithium-ion battery remaining life prediction model based on random error compensation provided in an embodiment of the present application is a multifunctional prediction model based on AEKF-CNN-LSTM, and the data parameters collected by this framework are denoised by advanced AEKF. These denoised parameters are used as training inputs for the CNN-LSTM network, enabling it to capture the nonlinear dynamics between input and output. Once the training phase is completed and the optimal model is determined, the estimation process can be performed to generate more accurate SOC and battery remaining life prediction results. Subsequently, the peak current is estimated based on the updated model and SOC value. By combining the two sets of estimates, the final estimate of the battery's multi-constrained continuous peak current is obtained, and then the multi-constrained peak power is calculated. Finally, a comprehensive prediction result of SOC, RUL and peak power is obtained.

[0136] The multifunctional prediction model of lithium-ion batteries based on AEKF-CNN-LSTM provided in the present application is described in detail below with respect to specific embodiments.

[0137] First of all, the multifunctional prediction method and model of lithium-ion batteries based on AEKF-CNN-LSTM proposed in the embodiment of the present application is based on the Windows 11 system and implemented using Python 3.9 and MATLAB 2019.

[0138] In addition, it should be noted that the embodiments of the present application are analyzed and evaluated by comparing different working conditions, methods and data sets.

[0139] First, the current and voltage values ​​in the original data set are extracted through the data set module 101, and then the preliminary SOC value is calculated by the ampere-hour integration method for fitting, and the equivalent circuit model and AEKF framework are established. The un-denoised current, voltage and SOC are input into the framework for random error compensation.

[0140] Then, the preprocessed current and voltage data sets are divided into training set, validation set and test set. Based on the network architecture of AEKF-CNN-LSTM, a deep learning model is constructed through the deep learning model module 102. The training set, validation set and test set are converted into random sequence data, input into the constructed network main body architecture, trained through the model training module 103, and the model parameters are adjusted to achieve the optimal model. The trained optimal deep learning model is selected through the model optimization module 104 according to the minimum loss value obtained by inputting the validation set. The NEDC condition uses a network with a CNN-LSTM layer to train on the data for more than 100 epochs. The tracking results and estimated errors are calculated as the actual value minus the estimated value. Then, the above experimental process is repeated using the AEKF-CNN-LSTM network. The MAE, RMSE and maximum error are obtained. The performance of the network is then tested on data sets at 10°C, 25°C and 40°C, respectively. The results are FIG. 6A to FIG. 6D Show.

[0141] Then, based on the test set, the performance of the deep learning model is evaluated by the prediction module 105. The evaluation indicators include mean absolute error (MAE) and root mean square error (RMSE), which are defined as follows:

[0142]

[0143] Overall, the AEKF-CNN-LSTM network shows commendable performance in estimating SOC. The RMSE is 1.52% at 25°C. However, as the temperature increases, the estimation accuracy gradually decreases, with RMSE values ​​ranging from 1.18% to 1.89%. The most significant performance degradation is at 40°C. This degradation can be attributed to the intensification of chemical reactions of the battery at high temperatures, which complicates the dynamic characteristics and hinders the estimation of SOC. In general, the trained network is adaptable to temperatures outside the training set. In addition, the CNN-LSTM network demonstrates excellent generalization capabilities under various driving cycles and temperature conditions. However, during the training stage, random errors caused by inaccurate sensors can seriously affect the accuracy of the training data, thereby affecting the accuracy and stability of the CNN-LSTM network estimation. AEKF can effectively mitigate the impact of random noise, thereby improving the estimation accuracy.

[0144] Table 1 SOC estimation results at different temperatures in the embodiment

[0145]

[0146] On the other hand, for SOP estimation, peak power estimation considering multiple constraints can significantly reduce the risk of overcharge and over-discharge, thereby improving the safety and reliability of electric vehicles. The results of the multi-constraint dynamic persistent peak current estimation method can be divided into two parts: model-based peak current estimation and SOC value-based peak current estimation. When the actual SOC of the battery is close to the maximum limit defined by the design parameters, the BMS can work with the vehicle control unit (VCU) to reduce or completely disable the peak charging current to avoid overcharging. At this critical moment, the peak charging power estimation mainly depends on the SOC. Conversely, when the actual SOC is close to the minimum design threshold, the BMS will also reduce or stop the peak discharge current to prevent over-discharge. If the estimated value exceeds the design constraint, the output will default to these limit values. Therefore, it becomes crucial to obtain accurate estimates of SOC and voltage. As can be seen from the figure, the estimation error of peak discharge current and peak discharge power within 2min and 5min can be effectively reduced by adopting the multi-constraint model. Similarly, it can be seen from the results that the estimation error increases at the end of battery discharge. The reason here is that the internal resistance of the battery increases sharply due to polarization reaction. The increase in internal resistance will hinder the battery current circulation, reduce current output, and affect battery life and safety. Therefore, the accurate estimation of peak current and power proposed in the embodiments of the present application can effectively improve the safety and reliability of electric vehicles.

[0147] Table 2 SOP estimation results at different discharge times in the embodiment

[0148]

[0149]

[0150] Finally, since RUL, together with the SOC and peak power of the lithium battery, provides information about the current state of the battery, it helps to determine whether the battery should be maintained, charged or replaced to prevent any failures and decide on measures to maximize the battery life. Therefore, accurate estimation and prediction of the above state parameters can allow manufacturers to extend the warranty period and help users avoid battery failures. The embodiment of the present application provides the results of a fusion model for predicting RUL. For battery RUL prediction, preprocessing the time series of current, voltage and capacity usage by AEKF can provide more accurate input parameters for the CNN-LSTM prediction model. This enables the model to learn more accurate state information and predict the next value based on the operating conditions of the battery. In the training phase, three different data sets are used. For each data set, a portion of the battery data is extracted as a training set, and the remaining battery data is used as a test set for the model to provide a prediction of the remaining percentage. In the CALCE data set, batteries CS35, CS36 and CS37 are selected as training sets, and CS38 is selected as a test set. In the Oxford data set, batteries 1, 3 and 7 are selected as training sets, and batteries 2, 4, 6 and 8 are selected as test sets. In the NASA dataset, datasets B0005, B0006, and B0007 are selected as training sets, and B0018 is selected as the test set. It can be seen that the AEKF-CNN-LSTM model can effectively track the capacity degradation curve, and the prediction error on the training set is very small. In addition, it also demonstrates high-level fitting performance and robustness for different types of datasets at different discharge rates.

[0151] Table 3 RUL estimation results under different data sets in the embodiment

[0152]

[0153] Here, MAPE is the mean absolute percentage error.

[0154] The present application provides a method for predicting the remaining life of a lithium-ion battery based on random error compensation. First, by introducing a second-order RC equivalent circuit model and using FFRLS to identify battery parameters online, AEKF denoises the original data of random errors observed by the sensor in each cycle, and divides the denoised data into a training set, a validation set, and a test set; second, based on the CNN-LSTM network architecture, a deep learning model is constructed; then, the denoised data is input into the CNN-LSTM neural network as a training sample; third, based on the training set and the validation set, the deep learning model is trained. Adjust the model parameters to obtain the optimized deep learning model and optimized model parameters; the fourth step is to evaluate the performance of the optimized deep learning model based on the optimized model parameters and test set, and obtain the trained deep learning model and updated model parameters, which can effectively reduce the impact of noise in the observed data on the data processing process, further improve the estimation performance, and ultimately achieve accurate prediction of SOC, SOP, and RUL; finally, based on the constraints of the battery's design current, voltage, power, SOC operating range, and power continuous output time, a multi-parameter constrained dynamic continuous peak power capability estimation method is established. Combined with the high robustness of the TCN algorithm and the efficiency of processing time series tasks, the impact of the approximate SOC value on the peak power capability estimation is reduced. Finally, the data set used for life prediction is divided and input into the trained deep learning model to predict RUL.

[0155] Based on the above technical solution, the embodiment of the present application provides a method and model for predicting the remaining life of a lithium-ion battery based on random error compensation. The method includes the following steps: first, obtaining a data set, preprocessing the data set, and dividing the preprocessed data set into a training set, a validation set, and a test set; then, based on the CNN-LSTM network architecture, building a deep learning model; next, based on the training set and the validation set, training the deep learning model, adjusting the model parameters, and obtaining an optimized deep learning model and optimized model parameters; then, based on the optimized model parameters and the test set, performing a performance evaluation on the optimized deep learning model, and obtaining a trained deep learning model and Updated model parameters; finally, the lithium-ion battery data set to be predicted is input into the trained deep learning model to predict the state of charge, remaining service life and peak power of the lithium-ion battery; wherein, predicting the peak power of the lithium-ion battery includes: based on the updated model parameters and the predicted state of charge of the lithium-ion battery, using constraints to realize peak current estimation based on the battery model and peak current estimation based on the battery state of charge; combining the peak current estimation results based on the battery model and the peak current estimation results based on the battery state of charge, deriving the estimated value of the battery's continuous peak current under multiple constraints, and obtaining the estimated value of the battery's continuous dynamic peak power under multiple constraints.

[0156] This application proposes a lithium-ion battery remaining life prediction method and model based on random error compensation, which can better estimate the SOC and SOP in lithium-ion batteries and predict their remaining life. BMS can monitor the SOC and SOP of the battery in real time to determine the current state of the battery and adjust the charging and discharging strategies accordingly, thereby extending the service life of the battery. More accurate SOC and SOP estimates can provide better input parameters for the battery life prediction model. Therefore, this application adopts the AEKF-CNN-LSTM combined model for training and prediction, aiming to minimize random errors through three different data sets and verify the accuracy and generalization ability of the model. Experimental results show that this method achieves high-precision estimation of battery SOC and SOP at different temperatures, as well as accurate prediction of battery life under different types and discharge times. In addition, the combination of the AEKF algorithm and the CNN-LSTM network adapts to the influence of random sensor errors and enhances the generalization ability of the algorithm.

[0157] Those skilled in the art can understand that the above-mentioned embodiments are specific examples for implementing the present application, and in practical applications, various changes can be made to them in form and detail without departing from the spirit and scope of the present application. Any person skilled in the art can make their own changes and modifications without departing from the spirit and scope of the present application, so the scope of protection of the present application shall be based on the scope defined in the claims.

Claims

1. A method for predicting the remaining life of a lithium-ion battery based on random error compensation, characterized in that: The following steps are involved: Obtaining a data set, preprocessing the data set, and dividing the preprocessed data set into a training set, a validation set, and a test set; Build a deep learning model based on the CNN-LSTM network architecture; Based on the training set and the validation set, the deep learning model is trained and model parameters are adjusted to obtain an optimized deep learning model and optimized model parameters; Based on the optimized model parameters and the test set, a performance evaluation is performed on the optimized deep learning model to obtain a trained deep learning model and updated model parameters; Input the lithium-ion battery data set to be predicted into the trained deep learning model to predict the state of charge, remaining service life and peak power of the lithium-ion battery; Among them, the peak power of lithium-ion batteries is predicted, including: Based on the updated model parameters and the predicted state of charge of the lithium-ion battery, the peak current estimation based on the battery model and the peak current estimation based on the battery state of charge are realized by using the constraint conditions; The peak current estimation results based on the battery model and the peak current estimation results based on the battery state of charge are combined to derive the estimated value of the battery's continuous peak current under multiple constraints, and obtain the estimated value of the battery's continuous dynamic peak power under multiple constraints.

2. The method for predicting the remaining life of a lithium-ion battery based on random error compensation according to claim 1, characterized in that: The data set is an electric vehicle driving condition data set including current, voltage and temperature data, and the current, voltage and temperature data are collected by sensors of the electric vehicle lithium-ion battery management system; The data set is preprocessed, including: Based on the data set, a battery model is established and parameter identification is performed; Constructing an adaptive extended Kalman filter model, and preprocessing the data collected by the sensor based on the adaptive extended Kalman filter model to remove random errors; The preprocessing process includes: Characteristic variables related to battery discharge during a driving cycle are extracted from the data set, and the characteristic variables are merged and randomly arranged to form an input data set; the characteristic variables include power battery current, voltage, capacity and temperature.

3. The method for predicting the remaining life of a lithium-ion battery based on random error compensation according to claim 2, characterized in that: Based on the data set, a battery model is established and parameter identification is performed, including: In order to track the dynamic characteristics of the battery and accurately reflect the various working states of the battery, a second-order resistance-capacitance circuit model, namely the dual-polarization model, is established based on the data set; According to the dual polarization model and the current and voltage data in the data set, a forgetting factor recursive least square method is used to perform parameter identification; The parameter identification comprises: The battery was subjected to a dynamic stress test at 25° C., and the sensor signal was recorded every 1 second; the sensor signal included a current signal and a voltage signal.

4. The method for predicting the remaining life of a lithium-ion battery based on random error compensation according to claim 3, characterized in that: In the dual polarization model, the relationship between the open circuit voltage and the battery state of charge is fitted and calibrated, and the forgetting factor recursive least squares method is used for parameter identification; the relationship between the open circuit voltage and the battery state of charge is represented by a monotonically increasing function, and the monotonically increasing function is: u oc (soc)=α1×soc 8 +α2×soc 7 +α3×soc 6 +α4×soc 5 +α5×soc 4 +α6×soc 3 +α7×soc 2 +α8×soc 1 +α9 Among them, the voltage source u oc It represents the open circuit voltage related to the battery state of charge SOC, and α is the electrochemical coefficient.

5. The method for predicting the remaining life of a lithium-ion battery based on random error compensation according to claim 2, characterized in that: According to the parameter identification results and data set type, an adaptive extended Kalman filter model is constructed, and the filter parameters of the model are determined. Its state equation is as follows: x k =f(x k-1 ,u k-1 )+ω k-1 y k =h(x k -u k )+v k Where x is the time t k，l is the system state vector, u is the input vector of the system, t is the fixed sampling interval between two adjacent measurement points, k and l are the time scale indicators of the macroscopic system parameters and the microscopic system state, L represents the time scale separation level, x k，0 =x k-1，L ; y is the system output vector, ω is the system noise, and v is the actual measurement noise; ω and v are independent of each other; The state estimation and covariance estimation are corrected using real-time measurements, and the results are expressed as and The measurement update equation of the adaptive extended Kalman filter is expressed as: d k =u OC，k +C k x k -i k R0 C k ＝[-1 -1 0] Among them, d k is the terminal voltage under the posterior state value, e k is the terminal voltage error under the posterior state value; The system state and error covariance correction are: Where K is the Kalman gain.

6. The method for predicting the remaining life of a lithium-ion battery based on random error compensation according to claim 1, characterized in that: Based on the CNN-LSTM network architecture, a deep learning model is constructed, and based on the training set and the validation set, the deep learning model is trained, and the model parameters are adjusted to obtain an optimized deep learning model and optimized model parameters, including: Construct the main network structure of CNN; the main network structure of CNN includes convolution layer, pooling layer, normalization layer and fully connected layer; construct the network structure of LSTM: Based on the LSTM network architecture, a recurrent connection is introduced into the CNN network, so that the CNN network can capture the time dependency in the sequence, and the training set data is input into the main network architecture of the CNN for training; the model parameters are adjusted to optimize the model, and the optimized deep learning model and optimized model parameters are obtained based on the minimum loss value obtained by inputting the validation set.

7. The method for predicting the remaining life of a lithium-ion battery based on random error compensation according to claim 1, characterized in that: Based on the multi-constraint conditions of battery design current, voltage, and power, as well as the operating range limit of the state of charge of the electric vehicle control system, a multi-constraint dynamic battery peak power capability estimation method is established to predict the peak power of lithium-ion batteries. A multi-constrained dynamic battery peak power capability estimation method is established, including: Based on the maximum and minimum limits of the battery state of charge, the expression of the peak current of the battery is obtained; Set terminal voltage constraints; Based on the peak current expression of the battery, the terminal voltage constraint condition, and the maximum discharge current and minimum charge current specified when the battery leaves the factory, the continuous dynamic peak power estimation expression under multiple constraints is obtained; The peak current of the battery at time l is: Among them, Z max and Z min are the maximum and minimum values ​​of the battery state of charge, respectively, and Q is the maximum capacity; The terminal voltage constraints are: The continuous dynamic peak power estimation expression under multiple constraints is as follows: in, and are the estimated values ​​of the continuous discharge peak current and continuous charge peak current at time l, P max and P min are the maximum power and minimum power set when the power battery is designed; and are the peak discharge power and peak charging power estimated by the dynamic continuous peak power estimation method under multiple constraints at time l; ΔL is the length of the continuous sampling time; i max and i min They are the maximum discharge current and minimum charge current specified when the battery leaves the factory.

8. The method for predicting the remaining life of a lithium-ion battery based on random error compensation according to claim 1, characterized in that: The data set includes: CALCE dataset, Oxford dataset and NASA dataset; CALCE dataset includes charging, discharging and measured resistance information of four 18650LiCoO2 lithium batteries, namely CS35, CS36, CS37 and CS38; Oxford dataset includes battery aging measurement data collected from eight small lithium-ion batteries; NASA dataset includes a collection of test data of charging, discharging and impedance of four lithium-ion batteries in three different operating modes; the datasets of the four lithium-ion batteries are B0005, B0006, B0007 and B0018.

9. The method for predicting the remaining life of a lithium-ion battery based on random error compensation according to claim 8, characterized in that: The preprocessed dataset is divided into training set, validation set and test set, including: In the CALCE dataset, batteries CS35, CS36, and CS37 are selected as training sets, and CS38 is selected as a test set; in the Oxford dataset, the measurement data of batteries with serial numbers 1, 3, and 7 are selected as training sets, and the measurement data of batteries with serial numbers 2, 4, 6, and 8 are selected as test sets; in the NASA dataset, datasets B0005, B0006, and B0007 are selected as training sets, and B0018 is selected as a test set.

10. A lithium-ion battery remaining life prediction model based on random error compensation, using the lithium-ion battery remaining life prediction method based on random error compensation as described in any one of claims 1 to 9 for prediction, characterized in that: include: The dataset module, deep learning model module, model training module, model optimization module and prediction module are connected in sequence; among them, The data set module is used to obtain a data set, preprocess the data set, and divide the preprocessed data set into a training set, a validation set, and a test set; The deep learning model module is used to build a deep learning model according to the network architecture of CNN-LSTM; The model training module is used to train the deep learning model according to the training set and the validation set, adjust the model parameters, and obtain an optimized deep learning model and optimized model parameters; The model optimization module is used to perform performance evaluation on the optimized deep learning model according to the optimized model parameters and the test set, to obtain a trained deep learning model and updated model parameters; The prediction module is used to input the lithium-ion battery data set to be predicted into the trained deep learning model to predict the state of charge, remaining service life and peak power of the lithium-ion battery; wherein, The prediction module includes a battery state of charge prediction module, a battery remaining service life prediction module and a battery peak power prediction module; the battery state of charge prediction module is used to predict the state of charge of the lithium-ion battery; the battery remaining service life prediction module is used to predict the remaining service life of the lithium-ion battery; The battery peak power prediction module includes a peak current estimation module and a continuous dynamic peak power estimation module; The peak current estimation module is used to realize peak current estimation based on the battery model and peak current estimation based on the battery state of charge by using constraint conditions according to the updated model parameters and the predicted state of charge of the lithium-ion battery; The continuous dynamic peak power estimation module is used to combine the peak current estimation result based on the battery model and the peak current estimation result based on the battery state of charge, derive the estimated value of the battery's continuous peak current under multiple constraints, and obtain the continuous dynamic peak power estimation value of the battery under multiple constraints.

Images