CN108872866B - Dynamic evaluation and long-acting prediction fusion method for charge state of lithium ion battery - Google Patents
Dynamic evaluation and long-acting prediction fusion method for charge state of lithium ion battery Download PDFInfo
- Publication number
- CN108872866B CN108872866B CN201810565223.0A CN201810565223A CN108872866B CN 108872866 B CN108872866 B CN 108872866B CN 201810565223 A CN201810565223 A CN 201810565223A CN 108872866 B CN108872866 B CN 108872866B
- Authority
- CN
- China
- Prior art keywords
- state
- soc
- battery
- ion battery
- charge
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- HBBGRARXTFLTSG-UHFFFAOYSA-N Lithium ion Chemical compound [Li+] HBBGRARXTFLTSG-UHFFFAOYSA-N 0.000 title claims abstract description 72
- 229910001416 lithium ion Inorganic materials 0.000 title claims abstract description 72
- 238000011156 evaluation Methods 0.000 title claims abstract description 41
- 238000007500 overflow downdraw method Methods 0.000 title claims abstract description 9
- 238000000034 method Methods 0.000 claims abstract description 68
- 238000013528 artificial neural network Methods 0.000 claims abstract description 16
- 239000011159 matrix material Substances 0.000 claims description 40
- 238000004422 calculation algorithm Methods 0.000 claims description 37
- 238000012549 training Methods 0.000 claims description 34
- 238000012360 testing method Methods 0.000 claims description 29
- 238000001914 filtration Methods 0.000 claims description 13
- 238000002790 cross-validation Methods 0.000 claims description 6
- 230000004927 fusion Effects 0.000 claims description 6
- 238000006073 displacement reaction Methods 0.000 claims description 5
- 238000001228 spectrum Methods 0.000 claims description 5
- 238000010998 test method Methods 0.000 claims description 4
- 238000001514 detection method Methods 0.000 abstract description 11
- 230000007774 longterm Effects 0.000 abstract description 10
- 230000008569 process Effects 0.000 description 13
- 230000010287 polarization Effects 0.000 description 10
- 238000007599 discharging Methods 0.000 description 5
- 230000000694 effects Effects 0.000 description 5
- 210000004027 cell Anatomy 0.000 description 4
- 230000006870 function Effects 0.000 description 4
- 238000009825 accumulation Methods 0.000 description 3
- 230000008859 change Effects 0.000 description 3
- 210000002569 neuron Anatomy 0.000 description 3
- 238000012545 processing Methods 0.000 description 3
- 238000004364 calculation method Methods 0.000 description 2
- 239000003990 capacitor Substances 0.000 description 2
- 230000005518 electrochemistry Effects 0.000 description 2
- 238000005259 measurement Methods 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 238000005070 sampling Methods 0.000 description 2
- WHXSMMKQMYFTQS-UHFFFAOYSA-N Lithium Chemical compound [Li] WHXSMMKQMYFTQS-UHFFFAOYSA-N 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000004146 energy storage Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000002349 favourable effect Effects 0.000 description 1
- 230000036541 health Effects 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 229910052744 lithium Inorganic materials 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 238000013021 overheating Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
- 230000004584 weight gain Effects 0.000 description 1
- 235000019786 weight gain Nutrition 0.000 description 1
Images
Landscapes
- Tests Of Electric Status Of Batteries (AREA)
Abstract
本发明公开一种锂离子电池荷电状态动态评估与长效预测融合方法，首先利用扩展卡尔曼滤波法对锂离子电池的电池荷电状态进行评估，得到锂离子电池荷电状态SOCKEF；然后利用回声状态神经网络对锂离子电池的电池荷电状态进行预测，得到锂离子电池荷电状态SOCESN；最后对锂离子电池荷电状态SOCKEF和锂离子电池荷电状态SOCESN进行加权融合，得到最终锂离子电池的电池荷电状态SOC。本发明提高了现有电池SOC检测方法的适应性和评估精度，克服单一方法进行SOC动态评估的局限性，针对性的选取基于模型和数据驱动的融合方法，兼顾SOC检测评估动态实时性和长期长效预测的需求。
The invention discloses a method for integrating dynamic evaluation and long-term prediction of the state of charge of a lithium ion battery. First, the extended Kalman filter method is used to evaluate the state of charge of the lithium ion battery, and the state of charge SOC KEF of the lithium ion battery is obtained; Using the echo state neural network to predict the battery state of charge of the lithium-ion battery, the lithium-ion battery state of charge SOC ESN is obtained; finally, the lithium-ion battery state of charge SOC KEF and lithium-ion battery state of charge SOC ESN are weighted and fused, Obtain the battery state of charge SOC of the final lithium-ion battery. The invention improves the adaptability and evaluation accuracy of the existing battery SOC detection method, overcomes the limitation of a single method for SOC dynamic evaluation, selects a model-based and data-driven fusion method in a targeted manner, and takes into account the dynamic real-time and long-term SOC detection and evaluation. Demand for long-term forecasting.
Description
Technical Field
The invention relates to the technical field of battery state of charge evaluation, in particular to a method for fusing dynamic evaluation and long-acting prediction of the state of charge of a lithium ion battery.
Background
The power battery is a key part of the new energy automobile, and the health condition and the management level of the power battery are related to the use efficiency and the safe operation of the whole automobile. The detection and estimation of battery state of charge (SOC) is the basis for battery thermal management, equalization management, and fail-safe management. Accurate SOC estimation can balance the differences between the cells, optimize the charge-discharge strategy, and prevent overheating and overcharging and overdischarging. The battery state estimation can fully and reasonably utilize the battery, prolong the service life of the battery and provide data basis for the energy management of the whole vehicle, so the battery state estimation has important significance for the battery management and even the energy management of the whole vehicle. The SOC estimation method and the realization technology of the power battery are deeply researched to generate important influence on the industrial development and social progress of the electric automobile.
The lithium ion battery is a typical dynamic, time-varying and nonlinear electrochemical system, internal parameters of the lithium ion battery are difficult to directly measure during online operation, a great challenge still exists for state identification and state estimation of the lithium ion battery at present, and meanwhile, most advanced methods for SOC prediction are still in the theoretical and method level, and the research of a method system is not mature. Although the SOC estimation method of the existing method has various characteristics, the SOC estimation method is generally subject to selection of an initial value of an estimation algorithm, errors caused by constant estimation model parameters along with changes of battery working conditions and service time, and the like, so that the detection accuracy and adaptability of the existing algorithm are reduced, and the like.
Disclosure of Invention
The invention provides a dynamic evaluation and long-acting prediction fusion method for the state of charge of a lithium ion battery, aiming at the problem that the detection precision and adaptability of the state of charge of the battery are reduced because the existing battery SOC estimation method depends on the fixation of model parameters and can not be adjusted along with the reappearance of the actual working condition of the battery.
In order to solve the problems, the invention is realized by the following technical scheme:
a dynamic evaluation and long-acting prediction fusion method for the charge state of a lithium ion battery specifically comprises the following steps:
step 1, estimating the battery charge state of the lithium ion battery by using an extended Kalman filtering method to obtain the lithium ion battery charge state SOCKEF；
Step 2, predicting the battery charge state of the lithium ion battery by using the echo state neural network to obtain the lithium ion battery charge state SOCESN；
Step 3, carrying out SOC (state of charge) on the lithium ion battery obtained in the step 1KEFAnd the state of charge SOC of the lithium ion battery obtained in the step 2ESNCarrying out weighted fusion to obtain the final lithium ion batteryState of charge SOC, wherein
SOC＝SOCKEF×w+SOCESN×(1-w)
Wherein w is weight, and w is more than or equal to 0 and less than or equal to 1.
The specific process of the step 2 is as follows:
step 2.1, establishing an echo state network model, and determining input and output nodes of the echo state network;
2.2, collecting voltage, current and temperature data of M groups of batteries, discretizing an SOC-OCV curve to obtain state of charge data of the corresponding M groups of batteries, and equally dividing the M groups of current, voltage, electric temperature and corresponding state of charge into K data sets, wherein each data set contains M/K groups of data, and M and K are set values;
step 2.3, dividing K data sets into K training sets and testing sets, and determining optimal parameters of the echo state network by adopting a cross validation method, namely the size N of a reserve pool, the spectrum radius SR, the input scaling IS and the input displacement IF; meanwhile, determining the sparsity SD by adopting a test method; thereby obtaining a parametrically determined echo state network;
step 2.4, initializing the echo state network obtained in the step 2.3, randomly setting an input weight matrix of the echo state network and an internal weight matrix of a reserve pool, and keeping the input weight matrix and the internal weight matrix unchanged;
2.5, selecting a training set and a test set from the K data sets, performing network learning on the echo state network by adopting a recursive least square method with forgetting factors, and updating an output weight matrix in real time;
step 2.6, judging whether the network learning reaches a termination condition, if not, returning to the step 2.5, continuing to load the training set and the test set for learning until the condition is met and terminated, and obtaining a trained echo state network;
step 2.7, inputting the voltage, the current and the temperature of the battery obtained by actual acquisition into the echo state network obtained in the step 2.6 for prediction, and outputting the state of charge (SOC) of the lithium ion batteryESN。
In step 2.4, the values of the elements in the input weight matrix and the internal weight matrix in the reserve pool are randomly generated between [ -1,1 ].
In step 2.6, the termination condition of the network learning is that the set Error or the step number is reached.
The specific process of the step 1 is as follows:
step 1.1, establishing a simplified GNL equivalent circuit model;
step 1.2, performing online identification on the simplified GNL equivalent circuit model parameters established in the step 1.1 by adopting a recursive least square method with forgetting factors, starting an online identification algorithm by taking actually acquired battery current as an input quantity and taking a battery terminal voltage and open-circuit voltage difference value as an observed quantity, and determining numerical values, namely parameter values, of all equivalent elements in the simplified GNL equivalent circuit model;
step 1.3, aiming at the simplified GNL equivalent circuit model parameter value obtained in the step 1.2, updating a corresponding parameter value in an extended Kalman filter algorithm state space equation, then taking the current and the temperature of the battery as input quantities, taking the terminal voltage of the battery as an observed quantity, starting the extended Kalman filter algorithm to realize the prediction of the state of charge of the battery, and obtaining the SOC of the lithium ion batteryKEF。
Compared with the prior art, the invention has the following characteristics:
1. the adaptability and the evaluation accuracy of the existing battery SOC detection method are improved, the limitation of a single method for SOC dynamic evaluation is overcome, a fusion method based on a model and data driving is selected in a targeted manner, and the requirements of SOC detection evaluation dynamic real-time performance and long-term prediction are met.
2. The method comprises the steps of establishing an equivalent circuit model capable of well describing the energy storage capacity of the battery, the electrochemistry of the battery and the concentration polarization effect, carrying out online dynamic parameter identification aiming at the parameters of the circuit model, carrying out state evaluation of a lithium ion battery complex dynamic nonlinear system based on an extended Kalman algorithm on the basis of the circuit model, establishing an accurate relation between the parameters of the power model and a target detection quantity SOC, and realizing online dynamic detection and evaluation of the SOC.
3. And establishing a long-term long-acting algorithm for lithium ion battery SOC evaluation prediction based on an ESN neural network prediction algorithm, wherein the terminal voltage, the charging and discharging current, the temperature, the charging and discharging multiplying power, the charging and discharging times and the working state of the battery of the lithium ion battery are used as input variables of the ESN neural network, and the SOC is used as the prediction output of the neural network. And (3) training the network by taking the factory charge-discharge curve of the lithium ion battery as an initial sample, further optimizing the network by acquiring state data on line, and finally realizing the data-based intelligent SOC algorithm for predicting the ESN.
4. A detection result weighting fusion mechanism is adopted, and organic combination of SOC real-time online dynamic evaluation and long-term long-acting state prediction is considered. Namely, a fusion method of a Kalman filtering algorithm based on a circuit model and a neural network prediction algorithm based on data is realized. The SOC evaluation result based on the model at the initial working stage of the lithium ion battery has a larger proportion in the final result, namely the weight takes a larger value; with the progress of the use, the ESN algorithm is gradually optimized and perfected, and the weight proportion can be adjusted to enable the ESN neural network to be used as a main basis of an evaluation result.
5. The integration method is used for organically combining SOC dynamic evaluation and long-acting prediction of the lithium ion battery, the advantages of the respective methods are exerted, the adaptability of the methods is improved, the evaluation prediction precision is finally improved, a detection theoretical support is provided for evaluation and prediction of the SOC of the lithium ion battery in practical application, and the evaluation accuracy and the reliability of maintenance suggestions are improved.
Drawings
Fig. 1 is a flow chart of a fusion method of dynamic state of charge evaluation and long-term prediction of a lithium ion battery.
Fig. 2 is a GNL equivalent circuit model.
Fig. 3 is a simplified equivalent circuit model of GNL.
Fig. 4 is an echo state network model.
Fig. 5 is a flow chart of echo state network prediction.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail below with reference to the accompanying drawings in conjunction with specific examples.
Referring to fig. 1, a method for fusing dynamic state of charge evaluation and long-term prediction of a lithium ion battery needs to undergo the following processes:
(1) and collecting the working state data of the lithium ion battery, such as the terminal voltage, the charging and discharging current, the temperature, the charging and discharging frequency and the like, in real time, and taking the working state data as the input of the SOC evaluation and prediction algorithm.
(2) According to the charge-discharge curve of the lithium ion battery, the relation between the open-circuit voltage and the SOC of the lithium ion battery is fitted, the relation is used as the basis for detecting the SOC by an open-circuit voltage method, and a corresponding SOC value is obtained by looking up a table according to the collected open-circuit voltage value during working.
(3) An equivalent circuit model of the lithium ion battery is established based on the working process characteristics of the lithium ion battery, and the state change of the battery can be accurately reflected. And performing online identification on the circuit model parameters by adopting a least square method with forgetting factors to dynamically adjust the circuit model parameters according to the identification result.
(4) And establishing a dynamic evaluation algorithm of the SOC of the lithium ion battery on the basis of the GNL equivalent circuit model of the lithium ion battery and the dynamic parameter identification result, and taking the algorithm as one of main bases of the online dynamic evaluation result of the SOC of the lithium ion battery.
(5) An echo state prediction neural network (ESN) is trained according to the relation between historical charge and discharge data of the lithium ion battery and the SOC, the trained neural network in the rebound state is used for processing battery parameters collected in real time to calculate to obtain an online SOC value in actual work, the online SOC value is used as a main basis for long-term prediction and employment release of the SOC of the lithium ion battery, and the SOC prediction precision of the method is gradually improved to a more ideal stable value along with the accumulation of the charge and discharge process.
(6) And (5) combining the predicted values of the SOC obtained in the step (4) and the step (5) and calculating a final SOC evaluation result by adopting a proper weighted proportion. The weight occupied by the algorithm evaluation result of the lithium ion battery in the initial service life stage (4) is large, and the secondary weight is taken as a main basis; with the advance of the using process, the prediction accuracy of the algorithm provided by the step is greatly improved due to the increase of data volume and the accumulation of experience, at the moment, the weight of the evaluation result can be properly improved (5), and the effects that the SOC can be dynamically evaluated on line and can be predicted for a long time can be achieved through method fusion.
The key steps of the present invention are described in detail below:
1. lithium ion battery SOC evaluation of extended Kalman filtering method
1.1, establishing an equivalent circuit model
The equivalent circuit model is based on the working principle of the battery, and elements such as a resistor, a capacitor, a constant voltage source and the like are used in the circuit to represent the dynamic characteristics of the battery. The equivalent circuit model is suitable for various working conditions; related mathematical formulas can be deduced according to the circuit principle, so that simulation analysis is facilitated; the effect of temperature is easily taken into account in the model.
The common battery equivalent circuit models at present mainly include: rint model, Thevenin model, PNGV model and GNL model. The PNGV model and the GNL model are obtained by improvement after the characteristics of direct current characteristics, polarization characteristics, self-discharge characteristics and the like of the battery are considered on the basis of the Rint model and the Thevenin model, and because the internal parameters of the battery are influenced by the characteristics in the working condition, the two equivalent circuit models have higher precision and are relatively complex in practical application.
The General Nonlinear (GNL) model is a complex high-order equivalent circuit model, and two RC links are used for simulating the polarization effect inside the battery, so that the precision is high. As shown in fig. 2, R in the circuit1And R2Electrochemical polarization resistance and concentration polarization resistance of the cell, respectively, C1And C2Respectively the electrochemical polarization capacitance and concentration polarization capacitance, R, of the celliIs the ohmic internal resistance, CbIs the inter-plate capacitance of the battery, the self-discharge internal resistance RsThe electric quantity loss in the self-discharge process of the battery can be reflected. The disadvantage of the GNL equivalent circuit model is that the parameters are too many and the identification process is complex.
The mathematical expression of the model is shown as the formula (1-1).
The precision of the Kalman filtering method depends on the precision of a dynamic model of a described system, and the relation between system observed quantity and state quantity is required to be accurately expressed by a mathematical formula. The PNGV model is formed by adding a capacitor C on the basis of Thevenin modelbThe SOC change caused by the accumulation of current in time can be simulated; the GNL model considers self-discharge influence factors on the basis of the PNGV model, and adds an RC link to simulate the influence of concentration polarization inside the battery, so that the model can reflect the working condition characteristics of the battery most. However, the GNL equivalent circuit model has many parameters and a complex structure, and the amount of computation of the identification algorithm and the SOC estimation algorithm based on the GNL equivalent circuit model is relatively large, which causes a large computational burden on the processor and is not favorable for engineering implementation. In engineering application, a GNL equivalent circuit model needs to be simplified to a certain extent.
In the invention, on the basis of the GNL model, the simplified GNL equivalent circuit model shown in FIG. 3 can be obtained by neglecting the self-discharge factor and the capacitance between the plates. In the figure VbAn open circuit voltage for the battery; riOhmic internal resistance of the battery; r1、R2Polarizing internal resistance of the battery; c1、C2Is the battery polarization capacitance. R1、C1And R2、C2The links describe the cell electrochemistry and concentration polarization effects of the cell respectively. Therefore, having six parameters in the simplified GNL equivalent circuit model requires identification.
1.2, identifying parameters of lithium battery circuit model on line
The equivalent elements in the equivalent circuit model of the battery have respective physical meanings and have close relation with the state of the battery. Therefore, the determination of the battery state requires the identification of parameters in the equivalent circuit model, i.e. the determination of the values of the equivalent elements. The current common battery model parameter identification methods include two types: offline parameter identification and online parameter identification.
The off-line parameter identification method is a method for roughly determining parameters in the battery model by fitting a voltage hysteresis curve through a pulse test. However, in the driving process of the electric vehicle, parameters in the battery equivalent circuit model may change along with the influence factors such as the current discharge rate, the ambient temperature, the battery SOC, and the like, and at this time, the model parameters determined by the offline parameter identification may not truly reflect the actual working state of the battery. Only by carrying out parameter identification on line in real time can the battery model parameters which best meet the actual state be obtained, and the accuracy of the SOC estimation method based on the model is improved.
The invention adopts a recursive least square method (FFRLS) with forgetting factors to carry out online identification on circuit model parameters, and the specific formula is shown in (1-2) - (1-4).
In the formula, the forgetting factor λ is selected to be a positive number close to 1, and is usually in the range of 0.95. ltoreq. λ.ltoreq.1. When λ is 1, the FFRLS algorithm degenerates to the common recursive least squares method.
For the simplified GNL equivalent circuit model shown in fig. 3, the equation of state in the frequency domain obtained from the circuit relationship is shown in equations (1-5).
Let τ be1＝R1C1，τ2＝R2C2Multiplication of (τ) at both ends of the equation1s+1)·(τ2s +1) gives:
let a be τ1·τ2，b＝τ1+τ2，c＝R1+R2+Ri，d＝R1τ2+R2τ1+Ri(τ1+τ2) Then equations (1-6) can be simplified as:
aVbs2+bVbs+Vb＝aRiIs2+dIs+cI+aVs2+bVs+V (1-7)
let s ═ x (k) -x (k-1)]/T，s2＝[x(k)-2x(k-1)+2x(k-2)]/T2Where T is the sampling period, s and s2And (3) carrying out difference operation on the formulas (1-7) to obtain the following results:
wherein:
namely:
θ＝[k1 k2 k3 k4 k5] (1-14)
by usingWhen the FFRLS algorithm is used for parameter identification, the coefficient matrix theta can be directly operated by using recursion formulas (1-2) - (1-4), and then the identification result is further operated to obtain a circuit model parameter R1、R2、C1、C2、Ri. The process is as follows:
let k0＝T2+ bT + a, then the formulae (1-9) - (1-13) can be deduced:
a＝-k0k2 (1-15)
by a ═ τ1·τ2，b＝τ1+τ2Then τ is2-b τ + a ═ 0, solved to:
by c ═ R1+R2+Ri，d＝R1τ2+R2τ1+Ri(τ1+τ2) Obtaining:
R2＝c-R1-Ri (1-21)
the simplified parameter R in the GNL equivalent circuit model can be obtained from the formulas (1-19) - (1-23)1、R2、C1、C2、RiAnd (6) estimating the value.
When FFRLS is used for identifying model parameters on line, the known parameters are V (k) and I (k) at the current moment, V (k-1), I (k-1) and SOC (k-1) at the previous moment, and V (k-2), I (k-2) and SOC (k-2) at the previous two moments.
1.3 expansion Kalman filtering algorithm lithium ion battery SOC dynamic evaluation method
An Extended Kalman Filter (EKF) is a state estimation method proposed for a nonlinear system, and was originally proposed by Stanley Schmidt to solve the problem of nonlinear navigation of a spacecraft. When the algorithm is used, firstly, the nonlinear function of the system is subjected to first-order Taylor expansion, and then the linear system equation is utilized to complete filtering tracking processing. The state space equations and observation equations of the nonlinear system are shown in equations (1-24).
In the formula, f (x)k，uk) Equation of state for a nonlinear system, g (x)k，uk) In the measurement equation of the nonlinear system, v and w are not related to each other and are Gaussian white noise which obeys normal distribution in the filtering calculation process.
When estimating the SOC of the battery by using the EKF, physical quantities such as the current and the temperature of the battery are generally used as an input vector U of the system, an operating voltage is used as an output Y of the system, and the SOC of the battery is included in a state quantity X of the system.
For the equivalent circuit model shown in fig. 3, it can be obtained from kirchhoff's law:
Vb(t)＝V+RiI(t)+V1(t)+V2(t) (1-27)
the equation of state obtained from equations (1-25), (1-26) and the ampere-hour metering equation is:
discretizing it can give:
the formula (1-29) is arranged in a matrix form to obtain:
where Δ t is the sampling time, η is the charge-discharge efficiency, w (k) is the process noise, and v (k) is the measurement noise, both of which are white gaussian noise with known variance and zero mean.
The coefficient matrix is:
therefore, the SOC evaluation value SOC of the lithium ion battery based on the extended Kalman algorithm can be obtainedKEF。
2. Prediction of lithium ion battery SOC by echo state neural network
2.1 establishment of echo State network model
As shown in fig. 4, an echo state network model is established, in which an input unit is K, a reserve pool is N, and an output unit is L, where an input signal is u (N), an output signal is y (N), an output signal of a neuron in the reserve pool is x (N), a weight matrix Win (N × K) is input, a weight matrix Win (N × N) in the reserve pool, and a feedback weight matrix Wback (N × L). The internal updating state of the reserve pool is shown as formulas (2-1) and (2-2):
X(n+1)＝f(Win×X(n)+Wres×U(n+1)+Wback×Y(n)) (2-1)
Y(n+1)＝fout(X(n+1),U(n+1)，Y(n)) (2-2)
wherein f is a neuron processing function inside the reserve pool, generally an S-shaped function, foutWhen the input signal is input into the echo state network model, the output signal X (n) of the neuron in the reserve pool is initialized to be X (0) 0, and the feedback weight matrix is an all-zero matrix. The input weight matrix, the internal weight matrix in the reserve pool and the feedback weight matrix are randomly generated, the value of matrix elements is between-1 and 1, and the network training only calculates the output weight matrix.
The echo state network model has a plurality of uncertain parameters, including reserve pool size N, spectrum radius SR, input scaling IS, input displacement IF and sparsity SD, and the parameters are independent and do not influence each other. The invention adopts a method based on K-fold cross validation to optimize the unknown parameters. Obtaining a relation curve of voltage and current (a corresponding curve is obtained under a certain discharge rate and represents charge-discharge current) and temperature and SOC from an OCV-SOC curve provided by a battery manufacturer, collecting real-time current I, voltage V and temperature T when a battery works by a collecting device, obtaining the SOC-OCV curve by an open-circuit voltage method, discretizing the curve to obtain a corresponding relation of the voltage V, the current I, the temperature T and the SOC, dividing the collected data I, V, T and the discretized SOC into K equal parts, firstly taking the K-th part as a test set, taking the rest data as a training set, then taking the K-1-th part as the test set, taking the rest data as the training set, and repeating the steps until the 1-th part is the test set and the rest parts are the training sets, and finishing K times of training and testing; in the cross validation process, firstly, input scaling IS and input displacement IF are fixed, the values of IS and IF are selected, the size N of a reserve pool and the spectrum radius SR are changed in a certain step length, training errors and testing errors are obtained, and when the sum of the obtained training errors and the obtained testing errors IS minimum, the corresponding N and SR are theoretically optimal parameters. In the same way, optimal parameters IS and IF can be obtained, the sparsity SD can be selected through an exhaustion method, and the establishment of the echo state network model IS completed. In order to find more accurate optimization parameters of the echo state network, theoretically, the more accurate the obtained optimization parameters are when the K value is larger, but the calculation disaster is also caused by the overlarge K value, so that the value should be taken by specifically combining the actual situation when the K value is selected.
2.2 training and prediction of echo State networks
In order to enable the echo state network to present the optimal prediction accuracy and generalization capability, 50%, 60%, 70%, 80% and 90% of the collected data are respectively used as training sets, and the rest are used as test sets. And when the training error and the testing error both meet the set error range and the sum of the two errors is minimum, the optimal training set and the testing set are corresponding to the training error and the testing error. And (3) learning the network by adopting a recursive least square method with forgetting factors for the optimal training set and the optimal test set, wherein in order to ensure that the error between the network output and a teacher supervision signal y (n) is minimum, an equation (2-3) is provided, and an output weight matrix Wout is obtained to ensure that E (k) is minimum:
introducing a forgetting factor lambda, then J (n) satisfies:
calculating an input weight matrix satisfying the formula (2-3) by calculating a partial derivative:
in order to avoid the network training from being trapped in local optimum, when the formula (2-5) is satisfied, a constraint Error is set, if the training Error is smaller than the constraint Error, a weight matrix is output, so when an Error function J (n) obtains an extreme value, the set Error constraint is reached, the obtained corresponding Wout can be approximately considered to be the output weight matrix satisfying the network optimum condition, an updating equation of an ESN output weight matrix can be obtained by a recursive least square method with forgetting factors, wherein Q is shown as (2-6), and Q iskIs a weight gain matrix.
Wout k+1＝Wout k+QkE(k) (2-6)
Referring to fig. 5, the implementation steps of the echo state network model prediction SOC are as follows:
step 1: data acquisition, namely determining input and output nodes and establishing an echo state network model;
step 2: dividing initial battery data into K training sets and testing sets, determining optimal parameters N, SR, IS and IF through cross validation, and determining SD through a test method;
and step 3: network initialization, randomly setting an input weight matrix, randomly generating matrix element values between [ -1,1] in a reserve pool, and keeping the matrix element values unchanged;
and 4, step 4: optimizing a training set and a test set, performing network learning by adopting a recursive least square method with forgetting factors, and updating an output weight matrix in real time;
and 5: judging whether the network learning reaches a termination condition (reaches a set Error or step number), if not, returning to the step 4, and continuing loading the training set and the test set for learning until the condition is met and terminated;
step 6: after the network training is finished, predicting according to the voltage V, the current I and the temperature T of the battery obtained by actual acquisition as network input, and outputting a predicted SOCESNThe value is obtained.
3. Lithium ion battery state of charge fusion prediction
The estimation and prediction numerical value SOC of the lithium ion battery is obtained by estimating and predicting the SOC based on the two algorithms of the extended Kalman filtering algorithm and the echo state neural networkKEFAnd SOCESNAdopting a weighting algorithm to predict and output a final result, namely SOC (state of charge) being equal to SOCKEF×w+SOCESNX (1-w), wherein w ranges from 0 to 1.
When w is 0, the final reproduction prediction result of the SOC is SOCESNNamely, completely adopting the evaluation result of the ESN neural network algorithm; when w is 1, SOC is SOCKEFNamely, the evaluation result of the extended kalman algorithm is completely adopted.
When the SOC of the lithium ion battery is estimated in an online prediction mode, w can be set between 0 and 1 and is set to a value close to 1, namely the proportion of the expanded Kalman filter algorithm in the estimation result is higher in the initial use stage of the battery, the ESN network estimation accuracy is gradually improved along with the extension of the use time, and the proportion of the W increasing ESN algorithm can be gradually reduced.
In summary, the method for fusing dynamic evaluation and long-term prediction of the state of charge of the lithium ion battery provided by the invention comprises the following steps:
step 1, estimating the battery charge state of the lithium ion battery by using an extended Kalman filtering method to obtain the lithium ion battery charge state SOCKEF. The method comprises the following specific steps:
step 1.1, establishing a simplified GNL equivalent circuit model;
step 1.2, performing online identification on the simplified GNL equivalent circuit model parameters established in the step 1.1 by adopting a recursive least square method with forgetting factors, starting an online identification algorithm by taking actually acquired battery current as an input quantity and taking a battery terminal voltage and open-circuit voltage difference value as an observed quantity, and determining numerical values, namely parameter values, of all equivalent elements in the simplified GNL equivalent circuit model;
step 1.3, aiming at the simplified GNL equivalent circuit model parameter value obtained in the step 1.2, updating the corresponding parameter value in the state space equation (1-30) of the extended Kalman filter algorithm, and then taking the battery current and the temperature as input quantities and the battery end electricityAnd (5) starting an extended Kalman filtering algorithm by taking the voltage as an observed quantity to realize the prediction of the state of charge of the battery and obtain the state of charge (SOC) of the lithium ion batteryKEF。
Step 2, predicting the battery charge state of the lithium ion battery by using the echo state neural network to obtain the lithium ion battery charge state SOCESN. The method comprises the following specific steps:
step 2.1, establishing an echo state network model, and determining input and output nodes of the echo state network;
2.2, collecting voltage, current and temperature data of M groups of batteries, discretizing an SOC-OCV curve to obtain state of charge data of the corresponding M groups of batteries, and equally dividing the M groups of current, voltage, electric temperature and corresponding state of charge into K data sets, wherein each data set contains M/K groups of data, and M and K are set values;
step 2.3, dividing K data sets into K training sets and testing sets, and determining optimal parameters of the echo state network by adopting a cross validation method, namely the size N of a reserve pool, the spectrum radius SR, the input scaling IS and the input displacement IF; meanwhile, determining the sparsity SD by adopting a test method; thereby obtaining a parametrically determined echo state network;
step 2.4, initializing the echo state network obtained in the step 2.3, randomly setting an input weight matrix of the echo state network and an internal weight matrix of a reserve pool, wherein the values of elements in the matrixes are randomly generated between [ -1,1] and are kept unchanged;
2.5, selecting a training set and a test set from the K data sets, performing network learning on the echo state network by adopting a recursive least square method with forgetting factors, and updating an output weight matrix in real time;
step 2.6, judging whether the network learning reaches a termination condition (the termination condition of the network learning is that the set Error or the step number is reached), if not, returning to the step 2.5, continuing to load the training set and the test set for learning until the condition is met and the trained echo state network is obtained;
step 2.7, inputting the voltage, the current and the temperature of the battery obtained by actual acquisition into the step 2.6Predicting and outputting the state of charge SOC of the lithium ion battery in an echo state networkESN。
Step 3, carrying out SOC (state of charge) on the lithium ion battery obtained in the step 1KEFAnd the state of charge SOC of the lithium ion battery obtained in the step 2ESNCarrying out weighted fusion to obtain the final battery state of charge (SOC) of the lithium ion battery, wherein
SOC＝SOCKEF×w+SOCESN×(1-w)
Wherein w is weight, and w is more than or equal to 0 and less than or equal to 1.
It should be noted that, although the above-mentioned embodiments of the present invention are illustrative, the present invention is not limited thereto, and thus the present invention is not limited to the above-mentioned embodiments. Other embodiments, which can be made by those skilled in the art in light of the teachings of the present invention, are considered to be within the scope of the present invention without departing from its principles.
Claims (3)
1. A lithium ion battery charge state dynamic evaluation and long-acting prediction fusion method is characterized by comprising the following steps:
step 1, estimating the battery charge state of the lithium ion battery by using an extended Kalman filtering method to obtain the lithium ion battery charge state SOCKEF(ii) a Namely:
step 1.1, establishing a simplified GNL equivalent circuit model;
step 1.2, performing online identification on the simplified GNL equivalent circuit model parameters established in the step 1.1 by adopting a recursive least square method with forgetting factors, starting an online identification algorithm by taking actually acquired battery current as an input quantity and taking a battery terminal voltage and open-circuit voltage difference value as an observed quantity, and determining numerical values, namely parameter values, of all equivalent elements in the simplified GNL equivalent circuit model;
step 1.3, aiming at the simplified GNL equivalent circuit model parameter value obtained in the step 1.2, updating a corresponding parameter value in an extended Kalman filtering algorithm state space equation, taking the current and the temperature of the battery as input quantities, taking the terminal voltage of the battery as an observed quantity, and starting the extended Kalman filteringThe algorithm realizes the prediction of the state of charge of the battery to obtain the state of charge SOC of the lithium ion batteryKEF；
Step 2, predicting the battery charge state of the lithium ion battery by using the echo state neural network to obtain the lithium ion battery charge state SOCESN(ii) a Namely:
step 2.1, establishing an echo state network model, and determining input and output nodes of the echo state network;
2.2, collecting voltage, current and temperature data of M groups of batteries, discretizing an SOC-OCV curve to obtain state of charge data of the corresponding M groups of batteries, and equally dividing the current, the voltage, the temperature and the corresponding state of charge of the M groups into K data sets, wherein each data set contains M/K groups of data, and M and K are set values;
step 2.3, dividing K data sets into K training sets and testing sets, and determining optimal parameters of the echo state network by adopting a cross validation method, namely the size N of a reserve pool, the spectrum radius SR, the input scaling IS and the input displacement IF; meanwhile, determining the sparsity SD by adopting a test method; thereby obtaining a parametrically determined echo state network;
step 2.4, initializing the echo state network obtained in the step 2.3, randomly setting an input weight matrix of the echo state network and an internal weight matrix of a reserve pool, and keeping the input weight matrix and the internal weight matrix unchanged;
2.5, selecting an optimal training set and a test set from the K data sets, namely respectively collecting training errors and test errors after network training and testing, when the training errors and the test errors both meet a set error range and the sum of the errors is minimum, the training errors and the test errors are corresponding to the optimal training set and the test set, and performing network learning on the echo state network by adopting a recursive least square method with forgetting factors, and updating an output weight matrix in real time;
step 2.6, judging whether the network learning reaches a termination condition, if not, returning to the step 2.5, continuing to load the training set and the test set for learning until the condition is met and terminated, and obtaining a trained echo state network;
step 2.7, the voltage, the current and the temperature of the battery obtained by actual acquisition are outputEntering the echo state network obtained in the step 2.6 for prediction and outputting the state of charge SOC of the lithium ion batteryESN；
Step 3, carrying out SOC (state of charge) on the lithium ion battery obtained in the step 1KEFAnd the state of charge SOC of the lithium ion battery obtained in the step 2ESNCarrying out weighted fusion to obtain the final battery state of charge (SOC) of the lithium ion battery, wherein
SOC＝SOCKEF×w+SOCESN×(1-w)
W is weight, w is more than or equal to 0 and less than or equal to 1, and the proportion of the expanded Kalman filtering method calculated in the evaluation result is larger at the initial stage of the use of the battery, namely the weight w takes a larger value; the evaluation accuracy of the echo state neural network is gradually improved along with the prolonging of the using time, and the weight of the echo state neural network in the evaluation result is increased by gradually reducing the weight w.
2. The method of claim 1, wherein in step 2.4, values of elements in the input weight matrix and the internal weight matrix in the reserve pool are randomly generated between [ -1,1 ].
3. The method according to claim 1, wherein in step 2.6, the termination condition of the network learning is that a set error or step number is reached.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810565223.0A CN108872866B (en) | 2018-06-04 | 2018-06-04 | Dynamic evaluation and long-acting prediction fusion method for charge state of lithium ion battery |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810565223.0A CN108872866B (en) | 2018-06-04 | 2018-06-04 | Dynamic evaluation and long-acting prediction fusion method for charge state of lithium ion battery |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108872866A CN108872866A (en) | 2018-11-23 |
| CN108872866B true CN108872866B (en) | 2021-02-05 |
Family
ID=64335055
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810565223.0A Active CN108872866B (en) | 2018-06-04 | 2018-06-04 | Dynamic evaluation and long-acting prediction fusion method for charge state of lithium ion battery |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108872866B (en) |
Families Citing this family (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109685207A (en) * | 2018-11-27 | 2019-04-26 | 电子科技大学 | A kind of multi core chip thermal management algorithm based on Recognition with Recurrent Neural Network |
| CN109598449B (en) * | 2018-12-14 | 2023-09-12 | 蓝谷智慧(北京)能源科技有限公司 | Lithium-ion battery pack performance evaluation method based on charging data |
| CN109581225A (en) * | 2018-12-28 | 2019-04-05 | 深圳市超思维电子股份有限公司 | The energy state evaluation method and battery management system of battery on-line parameter identification |
| CN110501652B (en) * | 2019-09-05 | 2021-11-19 | 上海毅信环保科技有限公司 | Rapid assessment method and device for available capacity of retired lithium battery |
| CN110531281B (en) * | 2019-09-09 | 2021-08-13 | 合肥工业大学 | Method and system for estimating state of health of power battery cells |
| KR102844868B1 (en) | 2019-10-18 | 2025-08-08 | 주식회사 엘지에너지솔루션 | Apparatus and method for estimating state of charge |
| CN111062170A (en) * | 2019-12-03 | 2020-04-24 | 广东电网有限责任公司 | Transformer top layer oil temperature prediction method |
| CN110888059B (en) * | 2019-12-03 | 2022-04-19 | 西安科技大学 | Charge state estimation algorithm based on improved random forest combined volume Kalman |
| CN113495213B (en) * | 2020-03-20 | 2025-02-21 | 北京京东乾石科技有限公司 | Electricity prediction model construction, handling equipment deployment method and device |
| CN112018465B (en) * | 2020-10-13 | 2021-01-29 | 北京理工大学 | Multi-physical-field-constrained intelligent quick charging method for lithium ion battery |
| CN112798967B (en) * | 2020-12-04 | 2022-03-15 | 电子科技大学 | Long-term and short-term prediction methods for solid oxide fuel cells |
| EP4512441A3 (en) * | 2020-12-30 | 2025-05-14 | Baxter International Inc. | System and method for generating battery alarms in infusion devices |
| CN113406500B (en) * | 2021-06-29 | 2022-06-14 | 同济大学 | Method for estimating residual electric quantity of power lithium battery |
| CN114705990B (en) * | 2022-03-31 | 2023-10-20 | 上海玫克生储能科技有限公司 | Method and system for estimating battery cluster state of charge, electronic equipment and storage media |
| CN115453376B (en) * | 2022-10-26 | 2023-04-07 | 国网浙江省电力有限公司宁波市北仑区供电公司 | State monitoring method and system for power battery of electric automobile and electric automobile |
| CN116359741B (en) * | 2023-03-14 | 2024-08-20 | 深圳市南霸科技有限公司 | Lithium battery safety detection method, device, computer equipment and storage medium |
| CN116381508A (en) * | 2023-04-12 | 2023-07-04 | 广东电网有限责任公司 | A battery state of charge estimation method and system based on multi-mode fusion |
| CN119535237A (en) * | 2025-01-17 | 2025-02-28 | 澄瑞电力科技(上海)股份公司 | A cloud SOC capacity correction method based on neural network |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102930173A (en) * | 2012-11-16 | 2013-02-13 | 重庆长安汽车股份有限公司 | SOC(state of charge) online estimation method for lithium ion battery |
| CN103293487A (en) * | 2013-06-28 | 2013-09-11 | 哈尔滨工业大学 | Lithium ion battery service life forecasting method based on integrated model |
| CN103295075A (en) * | 2013-04-01 | 2013-09-11 | 沈阳航空航天大学 | Ultra-short-term power load forecasting and early warning method |
| CN103336913A (en) * | 2013-07-25 | 2013-10-02 | 哈尔滨工业大学 | Method for monitoring state and predicting cut-off voltage of spatial lithium ion battery |
| CN105548896A (en) * | 2015-12-25 | 2016-05-04 | 南京航空航天大学 | Power-cell SOC online closed-loop estimation method based on N-2RC model |
| CN107402353A (en) * | 2017-06-30 | 2017-11-28 | 中国电力科学研究院 | A kind of state-of-charge to lithium ion battery is filtered the method and system of estimation |
| CN107748336A (en) * | 2017-11-06 | 2018-03-02 | 清华大学 | The state-of-charge On-line Estimation method and system of lithium ion battery |
-
2018
- 2018-06-04 CN CN201810565223.0A patent/CN108872866B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102930173A (en) * | 2012-11-16 | 2013-02-13 | 重庆长安汽车股份有限公司 | SOC(state of charge) online estimation method for lithium ion battery |
| CN103295075A (en) * | 2013-04-01 | 2013-09-11 | 沈阳航空航天大学 | Ultra-short-term power load forecasting and early warning method |
| CN103293487A (en) * | 2013-06-28 | 2013-09-11 | 哈尔滨工业大学 | Lithium ion battery service life forecasting method based on integrated model |
| CN103336913A (en) * | 2013-07-25 | 2013-10-02 | 哈尔滨工业大学 | Method for monitoring state and predicting cut-off voltage of spatial lithium ion battery |
| CN105548896A (en) * | 2015-12-25 | 2016-05-04 | 南京航空航天大学 | Power-cell SOC online closed-loop estimation method based on N-2RC model |
| CN107402353A (en) * | 2017-06-30 | 2017-11-28 | 中国电力科学研究院 | A kind of state-of-charge to lithium ion battery is filtered the method and system of estimation |
| CN107748336A (en) * | 2017-11-06 | 2018-03-02 | 清华大学 | The state-of-charge On-line Estimation method and system of lithium ion battery |
Non-Patent Citations (1)
| Title |
|---|
| 基于KF-ESN算法的储能电站磷酸铁锂电池SOC估算的研究;唐燕青;《中国优秀硕士学位论文全文数据库》;20161231;第18-34页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108872866A (en) | 2018-11-23 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108872866B (en) | Dynamic evaluation and long-acting prediction fusion method for charge state of lithium ion battery | |
| Xu et al. | A novel data-driven method for predicting the circulating capacity of lithium-ion battery under random variable current | |
| Chen et al. | State-of-charge estimation of lithium-ion batteries based on improved H infinity filter algorithm and its novel equalization method | |
| Xuan et al. | Real-time estimation of state-of-charge in lithium-ion batteries using improved central difference transform method | |
| He et al. | State-of-charge estimation of lithium ion batteries based on adaptive iterative extended Kalman filter | |
| CN114091282B (en) | Lithium ion battery state estimation method and system based on fractional order model | |
| CN113049962B (en) | LSTM-based energy storage device operation situation deduction method | |
| Zhang et al. | An on-line estimation of battery pack parameters and state-of-charge using dual filters based on pack model | |
| Wang et al. | A state-of-charge and state-of-health joint estimation method of lithium-ion battery based on temperature-dependent extended Kalman filter and deep learning | |
| Li et al. | An improved model combining machine learning and Kalman filtering architecture for state of charge estimation of lithium-ion batteries | |
| CN110850298B (en) | Data-driven SOH estimation method and system for lithium batteries | |
| CN113777510A (en) | Lithium battery state of charge estimation method and device | |
| CN111426957A (en) | SOC estimation optimization method for power battery under simulated vehicle working condition | |
| WO2019006995A1 (en) | Intelligent prediction system for power battery soc of electric vehicle | |
| Zhao et al. | Digital twin-driven estimation of state of charge for Li-ion battery | |
| CN112946481A (en) | Based on federation H∞Filtering sliding-mode observer lithium ion battery SOC estimation method and battery management system | |
| CN114609525A (en) | Power battery SOC estimation method based on fractional order cubature Kalman filtering | |
| CN117054892A (en) | An assessment method, device and management method for battery health status of an energy storage power station | |
| CN113466710A (en) | SOC and SOH collaborative estimation method for energy storage battery in receiving-end power grid containing new energy | |
| Wei et al. | State-of-charge estimation for lithium-ion batteries based on temperature-based fractional-order model and dual fractional-order Kalman filter | |
| Chen et al. | State of charge estimation for lithium-ion batteries using an adaptive cubature Kalman filter based on improved generalized minimum error entropy criterion | |
| CN115219911A (en) | A SOC-SOH joint estimation method based on lithium battery model and fractional order theory | |
| CN118275901B (en) | Lithium battery state of charge estimation method based on data and model filtering fusion | |
| Geng et al. | SOC Prediction of power lithium battery using BP neural network theory based on keras | |
| Li et al. | Real-world data-driven charging strategies for incorporating health awareness in electric buses |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |