Lithium ion battery hybrid charging state estimation method based on EMD decomposition and LSTM-KF

Abstract

The invention provides a lithium ion battery hybrid charging state estimation method based on EMD decomposition and LSTM-KF, and the hybrid method can adapt to various temperatures and working conditions to realize more accurate and stable SOC real-time estimation. The method comprises the following steps: acquiring a battery data set, EM) decomposing a voltage signal for processing, normalizing the decomposed voltage signal and current signal, constructing a long-short term memory network with an input feature dimension of 8, and performing Kalman filtering for data post-processing to obtain a final SOC estimation result. According to the method, only the original data of the battery needs to be directly used for the neural network through a preprocessing method, the calculation amount of the neural network is greatly reduced, detailed model research is avoided, hyper-parameters of the network do not need to be accurately set, the SOC estimation result of the neural network is improved through a filtering method, the method has high expansion capacity, and the method is suitable for popularization and application. The SOC can be stably and effectively estimated, and the method has reliable generalization performance and can adapt to various driving working conditions and various temperature conditions.

Description

Technical Field

[0001] The present invention belongs to the field of lithium-ion battery state monitoring and management, and in particular relates to a lithium-ion battery hybrid charging state estimation method based on EMD decomposition and LSTM-KF. Background Art

[0002] Lithium-ion batteries have become the preferred power source for many modern electronic devices and electric vehicles due to their high energy density, long cycle life, low self-discharge rate, and wide operating temperature range. They are also the most promising power batteries for electric vehicles. Battery Management System (BMS) is one of the core technologies of power batteries. Its main function is to monitor and protect batteries to ensure the safe, reliable and efficient operation of battery systems. BMS can monitor the operating status of batteries in real time and detect potential safety hazards in a timely manner, such as overcharging, over-discharging or overheating. It not only improves driving safety, but also helps to better manage battery use and reduce long-term costs.

[0003] State of Charge (SOC) is one of the most important states in a battery management system. It provides an important reference for preventing overcharging and over-discharging of the battery, and helps to improve charging efficiency and extend battery life. Therefore, real-time online estimation of the state of charge of the power battery is one of the key issues of BMS. In practical applications, lithium-ion batteries are nonlinear and time-varying, and are easily affected by multiple factors such as temperature and aging. SOC cannot be measured directly and can only be estimated through measurable external parameters such as voltage, current, and temperature. Therefore, it is particularly important to use advanced algorithms and strategies to quickly and accurately estimate SOC.

[0004] At present, there are three main methods for estimating battery SOC: traditional methods, model-based methods, and data-driven methods. Traditionally, the measurement of battery SOC adopts direct measurement methods, such as open circuit voltage method and ampere-hour integration method. The open circuit voltage method estimates SOC by measuring the voltage of the battery after a long period of static state. This method is simple and easy to implement, but it requires the battery to be in a static equilibrium state and cannot reflect the battery SOC in real time. The ampere-hour integration method estimates SOC by accumulating the amount of electricity charged and discharged. The principle is intuitive and the calculation process is simple. This method does not require complex equipment and only requires current sensors to monitor the battery current. Therefore, it is easy to implement in various types of battery management systems. Due to the inaccurate current measurement caused by sensor errors, the existence of sampling intervals, and changes in the battery's own characteristics (such as temperature changes, aging, etc.), the calculation errors will continue to accumulate. Therefore, the traditional method cannot meet the needs of real-time monitoring in terms of accuracy and time.

[0005] In order to solve this problem, it is necessary to integrate data-driven methods based on models. Model-based methods are mainly equivalent circuit models and electrochemical models. Equivalent circuit models use circuit networks composed of circuit elements such as resistors, capacitors, and voltage sources to express the relationship between various physical quantities of the battery. They do not reflect the real structure and complex electrochemical reactions inside the battery. In recent years, with the continuous development of deep learning and artificial intelligence technology, neural network theory and technology have developed rapidly. This network can automatically extract and identify complex patterns and relationships by learning a large number of data samples, and is very suitable for dealing with nonlinear problems. Neural networks can process a large amount of historical data and real-time data, which helps to improve the accuracy and reliability of SOC estimation. Once the neural network model is trained, it can be applied to different types of batteries and different working conditions, and has strong versatility. Summary of the invention

[0006] The present invention provides a hybrid charging state estimation method for lithium-ion batteries based on EMD decomposition and LSTM-KF. The hybrid method can adapt to a variety of temperature and operating conditions to achieve more accurate and stable SOC real-time estimation.

[0007] The present invention is implemented through the following technical solutions:

[0008] Step 1: Charge the lithium-ion battery with constant current and constant voltage, discharge it with constant current to 80% SOC after it is fully charged, and then perform pulse discharge under different working conditions and temperatures, and stop when the discharge reaches the cut-off voltage. During the discharge, collect the time series of current and voltage.

[0009] Step 2: Preprocess the voltage signal data obtained in step 1 to improve the memory capacity of the neural network. Empirical mode decomposition (EMD) is based on the concepts of instantaneous frequency and intrinsic mode function, and can decompose complex signals into several IMF components, each of which represents the local characteristics of the signal. The empirical mode decomposition method can stabilize non-stationary data, and then perform Hilbert transform to obtain the time-frequency spectrum and obtain physically meaningful frequencies.

[0010] Step 3: Normalize the current signal and the IMF components and voltage residual sequences decomposed in step 2 to eliminate the influence of different feature scales and make their features comparable. Use the mapminmax function to map the data to [-1, 1]. After numerical normalization, the convergence speed can be accelerated, making it easier to find the optimal solution of gradient descent.

[0011] Step 4: Construct a long short-term memory (LSTM) network. Use the training set data to establish a lithium-ion battery SOC estimation model based on LSTM. By setting the network structure and parameters, Adam is used as the network training optimization algorithm and ReLu is used as the network activation function. The root mean square error (RMSE) value is used as the fitness value, and the SOC estimation model parameters are obtained by training.

[0012] Step 5: Kalman filter (KF) can effectively filter out the noise contained in the SOC curve. Since the output of the neural network model is a nonlinear function of time t, the Kalman filter is used to optimally estimate the system state by using the linear system state equation and the system input and output observation data. Finally, the output value is filtered by solving the state matrix and covariance matrix to obtain a smooth and stable SOC estimation curve.

[0013] Compared with the prior art, the present invention has the following beneficial effects:

[0014] 1) Only the raw data of the battery (voltage and current) are required. In addition, they can be directly used in LSTM through preprocessing methods, which greatly reduces the computational workload of the neural network and realizes fast real-time estimation of SOC.

[0015] 2) The KF method was used to improve the SOC estimation results of the LSTM network, reducing the estimation error and significantly improving the estimation accuracy.

[0016] 3) The EMD-LSTM-KF method has reliable generalization performance and can adapt to various driving conditions and temperature conditions.

[0017] 4) This method avoids detailed model research and does not require precise setting of network hyperparameters. This method has strong scalability and can estimate the battery state of charge stably and effectively. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to make the technical solutions of the prior art clearer and the specific implementation methods easier to understand, the accompanying drawings are introduced.

[0019] Figure 1 The overall flow chart of the hybrid charging state estimation method for lithium-ion batteries based on EMD decomposition and LSTM-KF.

[0020] Figure 2 Current, voltage, and SOC diagrams at 25°C for different working conditions of the hybrid charging state estimation method for lithium-ion batteries based on EMD decomposition and LSTM-KF.

[0021] Figure 3It is a flowchart based on EMD decomposition.

[0022] Figure 4 This is the LSTM network structure diagram.

[0023] Figure 5 This is a comparison chart of the results of lithium-ion battery hybrid charging state estimation methods based on EMD decomposition and LSTM-KF. DETAILED DESCRIPTION

[0024] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0025] Example

[0026] The present invention uses a public data set from the data warehouse of the Center for Advanced Life Cycle Engineering (CALCE) at the University of Maryland. The parameters of 18650 batteries are shown in Table 1 below:

[0027] Table 1 Basic parameters of batteries

[0028] The experiment conducted open circuit voltage tests at three different temperatures (0°C, 25°C and 45°C). In addition, there were four dynamic tests: Dynamic Stress Test (DST), US06 Highway Driving Schedule (US06), Beijing Dynamic Stress Test (BJDST) and Federal Urban Driving Schedule (FUDS). The present invention uses DST condition data as the training set, and US06, FUDS and BJDST as the test sets.

[0029] The present invention is a hybrid charging state estimation method for lithium-ion batteries based on EMD decomposition and LSTM-KF. The specific flow chart is as follows: Figure 1 , the specific implementation steps are:

[0030] Step 1: The charge and discharge current and voltage of electric vehicle batteries are not constant in actual use. It is of practical significance to study the charge and discharge characteristics of batteries under different working conditions. In order to make the data cover the actual load conditions as much as possible, the data used in this article is tested at 0℃, 25℃ and 45℃ with 80% battery capacity. Figure 2 The current, voltage and SOC diagrams at 25°C under different working conditions.

[0031] Step 2: Before inputting the data of the lithium-ion battery into the long short-term memory network (LSTM), it is usually necessary to preprocess the obtained data set and convert it into features that can be input into the LSTM network model. Since LSTM has certain limitations in processing long sequences, the present invention decomposes the voltage signal during the long-term discharge process of the battery to improve the memory capacity of LSTM. Empirical mode decomposition is based on the concepts of instantaneous frequency and intrinsic mode function, and can decompose complex signals into several IMF components, each IMF characterizing the local characteristics of the signal. The empirical mode decomposition method can stabilize non-stationary data, and then perform Hilbert transform to obtain a time-frequency spectrum to obtain a physically meaningful frequency. Compared with methods such as short-time Fourier transform and wavelet decomposition, this method is intuitive, direct, a posteriori and adaptive because the basis function is decomposed from the data itself. Since the decomposition is based on the local characteristics of the signal sequence time scale, it is adaptive. The EMD decomposition flow chart is as follows Figure 3 As shown, the specific steps are:

[0032] Step 2.1: Find all extreme points of the original voltage signal U(t), and connect the local maximum points into the upper envelope curve e through the cubic spline curve max (t), connect the local minimum points to form the lower envelope e min (t), the upper and lower envelopes contain all the data points.

[0033] Step 2.2: Calculate the average value m1(t) of the upper envelope and the lower envelope as shown in formula (1):

[0034]

[0035] Step 2.3: "Screen" the original voltage signal, that is, subtract the mean envelope m1(t) from U(t) to obtain the intermediate signal C 1,1 (t):

[0036] C 1，1 (t) = U (t) - m1 (t) (2)

[0037] Step 2.4: Determine the intermediate signal C 1,1(t) whether it meets the two conditions of IMF, namely: 1) In the entire data segment, the number of extreme points and the number of zero-crossing points must be equal or the difference cannot exceed one. 2) At any time, the average value of the upper envelope formed by the local maximum point and the lower envelope formed by the local minimum point is zero, that is, the upper and lower envelopes are locally symmetrical with respect to the time axis. If it meets the requirements, the signal is an IMF component; if it does not meet the requirements, then repeat steps 2.1 to 2.4 based on the signal to continue "screening" until the signal obtained after decomposition k times meets the two conditions of IMF. The acquisition of IMF components generally requires several iterations. At this time, the first IMF component I1(t) of the original signal is obtained, and the formula is as follows:

[0038] C 1,k (t) = C 1，k-1 (t)-m k (t) (3)

[0039] I1(t)=C 1，k (t) (4)

[0040] I1(t) represents the highest frequency IMF component in the original voltage signal U(t). The residual component r1(t) is obtained by subtracting I1(t) from the original voltage signal U(t).

[0041] Step 2.5: Perform the above “screening” process on r1(t) to obtain the second IMF component I2(t), and then subtract I2(t) from the above residual component r1(t) to obtain the residual component r2(t), and so on until the last residual signal r n (t) Until it cannot be decomposed:

[0042]

[0043] After the above steps, the original voltage signal U(t) is decomposed into n IMF components and a residual sequence r n (t), expressed as:

[0044]

[0045] LSTM is a special variant of a recurrent neural network (RNN) with a "gate" structure. The logic control of the gate unit determines whether the data is updated or discarded, which overcomes the shortcomings of RNN weights that are too influential and prone to gradient vanishing and explosion, allowing the network to converge better and faster, and can effectively improve prediction accuracy. LSTM has three gates, namely the forget gate, input gate, and output gate, which determine whether information is remembered or forgotten at each moment. The input gate determines how much new information is added to the cell, the forget gate controls whether the information will be forgotten at each moment, and the output gate determines whether there is information output at each moment. The input is the terminal current I(t) at sampling time t, the battery voltage U1(t)~U6(t) and the residual U7(t) after empirical mode decomposition, and the output is SOC(t). The specific steps are:

[0046] Step 4.1: In a separate LSTM cell, include the current input X t and the input H at the previous moment t-1 They are processed by 3 fully connected sigmoid layers and 1 tanh layer. The sigmoid layer is represented by σ in the figure, and its value is 0 to 1. The value of the tanh layer is -1 to 1, which is used to create a new candidate unit vector and add it to the cell state C at the previous moment. t-1 First, when data enters the current neural unit, it will be changed from the cell state C at the previous moment t-1 The information to be discarded is determined by the forget gate. The forget gate collects the input X at the current moment. t and the input H at the previous moment t-1 , and nonlinearly map it through the sigmoid layer to obtain the required F t , as shown in formula (7):

[0047] F t =σ(X i W xf +H t-1 W hf +b f ) (7)

[0048] Step 4.2: The neural unit uses the input gate to determine what new information is added to the current cell state. The sigmoid layer and tanh layer determine the current input X t and the input H at the previous moment t-1 The sigmoid layer also determines whether specific information is remembered or forgotten. In addition, the tanh layer t and H t-1 Create a new candidate cell vector C in t, it will be added to the cell state. As shown in equations (8) to (9):

[0049] I t =σ(X t W xi +H t-1 W hi +b i ) (8)

[0050]

[0051] Step 4.3: When the input gate and forget gate obtain the required data, the next step is to update the cell state with the obtained data, as shown in formula (10):

[0052]

[0053] Where: ⊙ represents the multiplication of each element; I t is the state of the input gate at time t; the forget gate data F t The cell state C at the previous moment t-1 At this time, some information in the cell state has been selected to be forgotten, and the new cell state is obtained by adding the input gate data and the data multiplied by the new candidate unit. Finally, the required value is output through the output gate, as shown in equations (11) to (12):

[0054] 0 t =σ(X t W xo +H t-1 W ho +b o ) (11)

[0055] H t =O t ⊙tanh(C t ) (12)

[0056] Similarly, the sigmoid layer determines what data to add to the new output. Then, based on the cell state C at the previous moment t , and get the output H t , that is, the output data is output from the output gate data O tDetermined by the forget gate and input gate data. The above gates and cell units determine that LSTM selectively retains the current cell content. If the detected feature is considered important, the forget gate will be closed and the storage content will be retained for a long time, which is equivalent to capturing a long-term feature dependency. On the other hand, the unit can decide to reset the memory content by opening the forget gate. Since these two operating modes can occur simultaneously on different LSTM units, RNNs with multiple gates can capture the correlation relationship of long time series. The LSTM network structure is as follows: Figure 4 shown.

[0057] Step 4.4: Normalization. In order to eliminate the influence of different feature scales and make the features comparable, it is necessary to normalize the data. In this paper, the mapminmax function is used to map the data to [-1, 1]. After numerical normalization, the convergence speed can be accelerated and it is easier to find the optimal solution of gradient descent.

[0058] Step 4.5: Build LSTM network. Use the training set data to build a lithium-ion battery SOC estimation model based on LSTM. By setting the network structure and parameters, Adam is used as the network training optimization algorithm, ReLu is used as the network activation function, and the root mean square error value is used as the fitness, and then the SOC estimation model parameters are obtained by training.

[0059] Step 4.6: Model evaluation. The test set is used as the input of the LSTM model to obtain the estimated SOC value, and then the model accuracy is measured by the error between the predicted value and the actual value. The mean absolute error (MAE), mean square error (MSE), and root mean square error (RMSE) are used. The calculation formulas are:

[0060]

[0061]

[0062]

[0063] In the formula, is the predicted value at the i-th moment; y i is the actual value at the i-th moment; n is the amount of data.

[0064] The Kalman filter can handle the noise and interference in the measurement, estimate the state of the dynamic system from a series of data with measurement noise, and can be applied to a variety of fields. The Kalman filter can effectively filter out the noise contained in the SOC curve. Since the output of the neural network model is a nonlinear function of time t, KF is used to use the linear system state equation to optimally estimate the system state through the system input and output observation data. Finally, the output value is filtered by solving the state matrix and covariance matrix to obtain a smooth and stable SOC estimation curve. The specific steps are as follows:

[0065] Step 5.1: Initialization: Initialize the state variables and error covariance.

[0066] Step 5.2: Take the SOC output of EMD-LSTM as the actual input at the current moment.

[0067] Step 5.3: Calculate the SOC value at the current moment as the filter output. The calculation formula is (16):

[0068]

[0069] Where x is the output of EMD-LSTM, is the output of the Kalman filter, and K is the Kalman gain.

[0070] Step 5.4: Update the Kalman gain K:

[0071]

[0072] Where P is the error covariance, Q and R are process noise and measurement noise respectively.

[0073] Step 5.5: Update the error covariance P:

[0074]

[0075] The data set used in the present invention is the voltage and current data of 18650 lithium-ion batteries at three different temperatures (0°C, 25°C, and 40°C) and four different working conditions (DST as a training set, US06, FUDS, and BJDST as a test set), with a sampling time of 1s. The experimental results of EMD-LSTM-KF are compared with several other complex neural network models, namely the CNN-LSTM model based on convolutional neural network and long short-term memory neural network and the BILSTM model based on bidirectional long short-term memory network. The comparison results are shown in Table 2 below.

[0076] Table 2 RMSE comparison of different methods

[0077] The RMSE of the EMD-LSTM-KF method is lower than that of the other two methods at different temperatures and working conditions, indicating that the EMD-LSTM-KF method still has good robustness in estimating SOC, and can adapt to a variety of temperatures and working conditions to achieve accurate and rapid estimation of SOC. EMD-LSTM-KF can meet the expected accuracy at different temperatures and working conditions, and has good accuracy and robustness. The comparison of the results of the hybrid charge state estimation method of lithium-ion batteries based on EMD decomposition and LSTM-KF is shown in the figure below. Figure 5 Shown: (a) is the FUDS working condition at 0℃; (b) is the US06 working condition at 25℃; (c) is the BJDST working condition at 40℃.

[0078] The present invention proposes a new battery SOC estimation method based on EMD-LSTM-KF, and uses dynamic tests to verify the accuracy and generalization ability of the method in estimating SOC. The experimental results show that only using the battery measurement voltage and current data as the original data to decompose the voltage signal through EMD into the LSTM network model improves the memory capacity of LSTM and reduces the workload. The output result combined with filtering reduces the output fluctuation, further stabilizes the output of SOC, and improves the estimation accuracy. Specifically, the RMSE of the battery SOC estimated by the EMD-LSTM-KF model proposed in this paper is less than 1.95%, and the RMSE at room temperature and high temperature is less than 0.97%, which has good estimation accuracy and generalization ability. Through comparative experiments with existing methods on the same data set, the EMD-LSTM-KF model only requires the voltage and current data of the battery, and does not need to specifically set the hyperparameters of the network. The EMD decomposition signal improves the ability of the neural network to capture information and converges quickly. The experimental results prove that the method can adapt to different driving conditions and different temperature conditions. The method can achieve fast and accurate estimation in SOC estimation, which is beneficial to battery health and prolongs the battery life.

[0079] Any matters not described in the present invention are applicable to the prior art.

Claims

1. A hybrid charging state estimation method for lithium-ion batteries based on EMD decomposition and LSTM-KF. The hybrid method can adapt to a variety of temperature and operating conditions to achieve more accurate and stable SOC real-time estimation. The specific steps are: Step 1: Charge the lithium-ion battery with constant current and constant voltage, and after fully charged, pulse discharge it under different working conditions and temperatures, and stop discharging when the discharge reaches the cut-off voltage. During the discharge, collect the time series of current and voltage; Step 2: Preprocess the voltage signal data obtained in step 1 to improve the memory capacity of the neural network. Empirical mode decomposition (EMD) is based on the concepts of instantaneous frequency and intrinsic mode function, and can decompose complex signals into several IMF components. Each IMF represents the local characteristics of the signal. The empirical mode decomposition method can stabilize non-stationary data, and then perform Hilbert transform to obtain the time-frequency spectrum and obtain physically meaningful frequencies. Step 3: Normalize the current signal and the IMF components and voltage residual sequence decomposed in step 2 to eliminate the influence of different feature scales and make their features comparable. Use the mapminmax function to map the data to [-1, 1]. After numerical normalization, the convergence speed can be accelerated, making it easier to find the optimal solution of gradient descent. Step 4: Construct an LSTM network and use the training set data to establish a lithium-ion battery SOC estimation model based on LSTM. By setting the network structure and parameters, Adam is used as the network training optimization algorithm, ReLu is used as the network activation function, and the root mean square error (RMSE) is used as the fitness value, and then the SOC estimation model parameters are obtained by training; Step 5: The Kalman filter can effectively filter out the noise contained in the SOC curve. Since the output of the neural network model is a nonlinear function of time t, the Kalman filter is used to use the linear system state equation to optimally estimate the system state through the system input and output observation data. Finally, the output value is filtered by solving the state matrix and covariance matrix to obtain a smooth and stable SOC estimation curve.

2. The method for estimating the hybrid charge state of a lithium-ion battery based on EMD decomposition and LSTM-KF according to claim 1 is characterized in that In the step 1, the lithium-ion battery is charged with constant current and constant voltage to a cut-off voltage, then discharged with constant current to 80% SOC, and then pulse discharged to collect voltage and current signals during the discharge process.

3. The method for estimating the hybrid charging state of a lithium-ion battery based on EMD decomposition and LSTM-KF according to claim 1 is characterized in that The step 2 is specifically as follows: Step 2.1: Find all extreme points of the original voltage signal U(t), and connect the local maximum points into the upper envelope curve e through the cubic spline curve max (t), connect the local minimum points to form the lower envelope e min (t), the upper and lower envelopes contain all the data points; Step 2.2: Calculate the average value m1(t) of the upper envelope and the lower envelope; Step 2.3: "Screen" the original voltage signal, that is, U(t) minus the mean envelope m1(t) to obtain the intermediate signal C 1,1 (t); Step 2.4: Determine the intermediate signal C 1,1 (t) Whether it meets the two conditions of IMF: 1) In the entire data segment, the number of extreme points and the number of zero-crossing points must be equal or the difference cannot exceed one; 2) At any time, the average value of the upper envelope formed by the local maximum point and the lower envelope formed by the local minimum point is zero, that is, the upper and lower envelopes are locally symmetrical with respect to the time axis. If it meets the conditions, the signal is an IMF component; if it does not meet the conditions, then repeat steps 1 to 4 based on the signal to continue "screening" until the signal obtained after decomposition k times meets the two conditions of IMF. The acquisition of IMF components generally requires several iterations, and then the first IMF component I1(t) of the original signal is obtained; Step 2.5: Perform the above "screening" process on r1(t) to obtain the second IMF component I2(t), and then subtract I2(t) from the above residual component r1(t) to obtain the residual component r2(t), and so on until the last residual signal r n (t) cannot be decomposed. After the above steps, the original voltage signal U(t) is decomposed into n IMF components and a residual sequence r n (t).

4. The method for estimating the hybrid charging state of a lithium-ion battery based on EMD decomposition and LSTM-KF according to claim 1 is characterized in that The step 5 is specifically as follows: Step 5.1: Initialize the state variables and error covariance before the initialization; Step 5.2: Take the SOC output of EMD-LSTM as the actual input at the current moment; Step 5.3: Calculate the SOC value at the current moment as the filter output; Step 5.4: Update the Kalman gain K; Step 5.5: Update the error covariance P.

Images