CN117783875A - Lithium battery state of charge prediction method and device based on model fusion - Google Patents
Lithium battery state of charge prediction method and device based on model fusion Download PDFInfo
- Publication number
- CN117783875A CN117783875A CN202311820312.2A CN202311820312A CN117783875A CN 117783875 A CN117783875 A CN 117783875A CN 202311820312 A CN202311820312 A CN 202311820312A CN 117783875 A CN117783875 A CN 117783875A
- Authority
- CN
- China
- Prior art keywords
- lithium battery
- data
- model
- soc
- equivalent circuit
- 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.)
- Granted
Links
- WHXSMMKQMYFTQS-UHFFFAOYSA-N Lithium Chemical compound [Li] WHXSMMKQMYFTQS-UHFFFAOYSA-N 0.000 title claims abstract description 86
- 229910052744 lithium Inorganic materials 0.000 title claims abstract description 86
- 238000000034 method Methods 0.000 title claims abstract description 74
- 230000004927 fusion Effects 0.000 title claims abstract description 29
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 22
- 238000012549 training Methods 0.000 claims abstract description 16
- 238000005457 optimization Methods 0.000 claims abstract description 15
- 238000013136 deep learning model Methods 0.000 claims abstract description 13
- 230000010354 integration Effects 0.000 claims abstract description 7
- 238000012545 processing Methods 0.000 claims abstract description 6
- 239000002245 particle Substances 0.000 claims description 57
- 239000013598 vector Substances 0.000 claims description 26
- 230000008569 process Effects 0.000 claims description 20
- 230000000739 chaotic effect Effects 0.000 claims description 19
- 230000006870 function Effects 0.000 claims description 11
- 230000010287 polarization Effects 0.000 claims description 9
- 238000004590 computer program Methods 0.000 claims description 6
- 238000002759 z-score normalization Methods 0.000 claims description 6
- 238000004364 calculation method Methods 0.000 claims description 5
- 238000001914 filtration Methods 0.000 claims description 5
- 238000007781 pre-processing Methods 0.000 claims description 4
- 238000012804 iterative process Methods 0.000 claims description 3
- 230000001131 transforming effect Effects 0.000 claims description 3
- HBBGRARXTFLTSG-UHFFFAOYSA-N Lithium ion Chemical compound [Li+] HBBGRARXTFLTSG-UHFFFAOYSA-N 0.000 claims description 2
- 229910001416 lithium ion Inorganic materials 0.000 claims description 2
- 230000011218 segmentation Effects 0.000 claims description 2
- 230000009286 beneficial effect Effects 0.000 description 3
- 238000011156 evaluation Methods 0.000 description 3
- 239000011159 matrix material Substances 0.000 description 3
- 230000032683 aging Effects 0.000 description 2
- 238000006243 chemical reaction Methods 0.000 description 2
- 239000013256 coordination polymer Substances 0.000 description 2
- 230000005253 gamme decay Effects 0.000 description 2
- 239000012633 leachable Substances 0.000 description 2
- 230000007246 mechanism Effects 0.000 description 2
- 238000012544 monitoring process Methods 0.000 description 2
- 238000010606 normalization Methods 0.000 description 2
- 238000002360 preparation method Methods 0.000 description 2
- 230000002159 abnormal effect Effects 0.000 description 1
- 230000004931 aggregating effect Effects 0.000 description 1
- 230000005262 alpha decay Effects 0.000 description 1
- 238000003491 array Methods 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 238000004140 cleaning Methods 0.000 description 1
- 230000000295 complement effect Effects 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000007599 discharging Methods 0.000 description 1
- 230000005264 electron capture Effects 0.000 description 1
- 230000005484 gravity Effects 0.000 description 1
- 230000036541 health Effects 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000007500 overflow downdraw method Methods 0.000 description 1
- 230000004044 response Effects 0.000 description 1
Landscapes
- Tests Of Electric Status Of Batteries (AREA)
Abstract
The invention discloses a lithium battery state of charge prediction method and device based on model fusion, which are used for acquiring a lithium battery data set, and processing and data division; establishing a second-order Thevenin equivalent circuit model of the lithium battery, and improving an energy valley optimization algorithm EVO; combining the mathematical form of the equivalent circuit with an improved energy valley optimization algorithm IEVO, calculating each parameter to be solved, and completing the parameter identification of the equivalent circuit model to obtain a complete second-order Thevenin equivalent circuit model of the lithium battery; calculating by using an ampere-hour integration method to obtain predicted data of the lithium battery; inputting the preprocessed stable lithium battery data set into a deep learning model for training to obtain corresponding lithium battery prediction data; and carrying out data fusion on the obtained lithium battery SOC output data of the two models by an entropy weight method to obtain a final lithium battery SOC state after fusion. Compared with the existing lithium battery state of charge prediction, the method has higher accuracy and stability, and realizes more accurate lithium battery state of charge prediction.
Description
Technical Field
The invention belongs to the technical field of lithium battery state-of-charge detection, and particularly relates to a lithium battery state-of-charge prediction method and device based on model fusion.
Background
State of charge is one of the core parameters of a battery management system (Battery Management System, BMS) that monitors and estimates the State of charge (SOC) of a battery to ensure that the battery operates efficiently while avoiding battery damage due to overcharge or overdischarge. The accurate estimation of the SOC is beneficial to reasonably utilizing the electric energy provided by the battery, improving the utilization rate of the battery and prolonging the service life of the battery pack, and meanwhile, the estimation of the SOC is influenced by various factors such as temperature, charge-discharge multiplying power, battery aging and the like. Therefore, the battery management system needs to continuously correct the SOC to maintain the accuracy thereof, and can avoid the battery operating in an excessively high or excessively low state of charge, thereby reducing the aging and self-discharge speed of the battery and prolonging the service life of the battery. Through real-time monitoring and prediction of the SOC, abnormal conditions such as overcharge and overdischarge of the battery can be timely found, so that corresponding measures are taken to avoid safety accidents, and accurate prediction of the state of charge of the lithium battery is of great significance in improving the battery performance, prolonging the service life of the battery, optimizing energy management, guaranteeing safety and the like.
The state of charge of a lithium battery is a ratio representing the remaining capacity of the battery to the capacity of the fully charged state, and this ratio ranges from 0 to 1. When SOC is 0, it indicates that the battery has been fully discharged; when the SOC is 1, the battery is completely full, in practical application, various algorithms and methods are often adopted to estimate the SOC, and accurate prediction of the SOC can help people reasonably utilize the battery energy in the charging and discharging processes, and overcharge or overdischarge is avoided, so that the utilization rate of the battery is improved. The method has the advantages that the method does not need to deeply study the complex chemical reaction mechanism in the lithium battery, and the charging state of the battery can be predicted only by monitoring parameters such as voltage, current and the like which are easy to measure and historical data; however, this method also has obvious disadvantages, namely, the single prediction model has larger contingency and poor robustness, and the information of the battery health condition cannot be comprehensively represented.
Disclosure of Invention
The invention aims to: aiming at the problems of low SOC prediction precision and poor stability of the existing single model prediction lithium battery, the method and the device for predicting the state of charge of the lithium battery based on model fusion are provided, and the precision of the lithium battery state of charge estimation model is remarkably improved.
The technical scheme is as follows: the invention discloses a lithium battery state of charge prediction method based on model fusion, which comprises the following steps:
(1) Acquiring initial data of a lithium ion battery, acquiring a training set required by a second-order Thevenin equivalent circuit model, and preprocessing the training set:
(2) Establishing a second-order Thevenin equivalent circuit model of the lithium battery, and transforming an equivalent circuit model equation into a mathematical form which contains terminal voltage, open-circuit voltage, capacitance and parameters to be solved and can be identified by a computer;
(3) Performing parameter identification on the second-order Thevenin equivalent circuit model based on an improved energy valley optimization algorithm IEVO to obtain an optimized second-order Thevenin equivalent circuit model of the lithium battery, and calculating by using an ampere-hour integration method to obtain SOC prediction data of the lithium battery; the improved energy valley optimization algorithm IEVO is used for carrying out local multi-element search on the position of the population in the iterative process by adopting a multi-kernel learning method; by adopting a chaotic local search strategy, finding out a better solution by searching a nearby area of each solution;
(4) Inputting the preprocessed lithium battery data set into a cross former deep learning model after random initialization parameters for training to obtain a trained deep learning model, thereby obtaining lithium battery SOC prediction data;
(5) And (3) carrying out data fusion on the SOC prediction data of the two models obtained in the step (3) and the step (4) through an entropy weight method to obtain the final lithium battery state of charge.
Further, the data set in the step (1) includes charge and discharge data of the lithium battery, an open-circuit voltage, a temperature of the lithium battery, an internal resistance, a capacity, a working current and corresponding state of charge indexes respectively.
Further, in the step (1), the data processing is to perform a z-score normalization after performing a filtering process on the training data by using a Savitzky-Golay filter.
Further, in the step (2), a second-order Thevenin equivalent circuit model is built, and a state space equation of the equivalent circuit model is shown in the following formula:
U O ＝U OC -U P1 -U P2 -IR
in U P1 And U P2 The voltages of two RC loops in the second-order Thevenin equivalent circuit model are respectively; r is R P1 ，R P2 2 polarization internal resistances of the battery; c (C) P1 ，C P2 Two polarization capacitances for the battery; u (U) OC Represents an open circuit voltage; u (U) O Is the terminal voltage; r is the internal resistance of the battery; i is the total current in the circuit; r is R P1 ，R P2 ，C P1 ，C P2 ，U OC Is the parameter to be identified.
Further, the implementation process of the step (3) is as follows:
the lithium battery state space equation is embedded into an energy valley optimization algorithm EVO, and R is identified through repeated iteration 1 ,R 2 ,C 1 ,C 2 Equivalent circuit parameters;
carrying out an initialization process to obtain a parameter X to be solved of a second-order Thevenin equivalent circuit model i Assume particles with different stability levels in the search space:
where n represents the total number of particles in the search space, d is the dimension of the problem under consideration,is the j-th decision variable determining the initial position of the i-th candidate,/or->And->Representing the lower and upper bounds, respectively, of j variables in the ith candidate variable, rand is at [0,1]Random numbers uniformly distributed in the range;
adding a chaotic local search strategy to generate a group of initial solutions with randomness and uniformity by using a chaotic phenomenon:
wherein,representing a new solution generated at the t+1st iteration using the chaotic local search strategy,/o>And->Respectively representing the random selection of 2 different individuals from the dominant group, C t Representing a chaotic value generated by the chaotic map of the t-th iteration;
obtaining an SOC value by an ampere-hour integration method:
wherein U is 1 (t)，U 2 And (t) is electrochemical polarization voltage of the lithium battery:
wherein C represents the capacitance of the lithium battery, i L And (t) is an operating current when the lithium battery is charged and discharged at the time t, and the SOC (0) represents an SOC1 value at the initial time of the lithium battery.
Further, the implementation process of the step (4) is as follows:
l2 regularization optimization is carried out on a CrossFormer deep learning model:
where L (w) is the loss function regularized by L2, loss (w, x) is the original loss function of the model, w is the weight vector of the model,is the square of the L2 norm of the weight vector, λ is the regularization parameter, used to control the strength of the regularization;
embedding a multivariate time sequence using dimension segmentation; the layered encoder-decoder structure predicts in a cross former deep learning model by using larger-scale information, adds the predicted values of different scales, and outputs a final result SOC2.
Further, the implementation process of the step (5) is as follows:
performing dimensionality removal treatment on each index, and performing data standardization: two data sets { SOC 1 ,SOC 2 }, wherein SOC i ＝{x 1 ,x 2 ,…,x n Normalized value for each dataset is { SOC } 1 * ,SOC 2 * -then:
the variance size of each dataset was determined:
the information entropy of each data set is calculated, and according to the definition of the information entropy, the information entropy of one group of data is as follows:
wherein E is i Not less than 0, if p ij =0, definition E i =0; determining the weight of each index, and calculating the information entropy of each data set to be E according to a calculation formula of the information entropy 1 ,E 2 ；
Calculating the weight of each index through information entropy:
the data fusion adopts an entropy weight method estimation algorithm, as shown in the formula:
in the formula, the SOC is the state of charge prediction of the final lithium battery after fusion.
The invention relates to a device and equipment, comprising a memory and a processor, wherein:
a memory for storing a computer program capable of running on the processor;
a processor for performing the steps of the model fusion based lithium battery state of charge prediction method as described above when running the computer program.
The beneficial effects are that: compared with the prior art, the invention has the beneficial effects that:
aiming at the problems of lower precision and poor parameter stability of the traditional second-order Thevenin equivalent circuit when the parameter identification is carried out, the invention combines an energy valley optimization algorithm, improves the global optimizing capability and the local searching capability when the parameter identification is carried out, ensures that the algorithm is finer in the process of searching a high-quality solution, provides more accurate parameters for SOC estimation, and improves the precision; the situation that parameters of the cross former depth prediction model are too many can lead to the need of a large amount of calculation resources and time when the model is trained, in addition, too many parameters can also lead to the phenomenon of overfitting of the model, so that the prediction performance is reduced, and L2 regularization is adopted to reduce the number of model parameters, so that the overfitting risk is reduced; the method adopts the model fusion method to comprehensively consider the internal chemical reaction of the battery during operation, accurately simulate the external dynamic response of the battery, and the multi-model prediction method can fully utilize the characteristics of different models, realize complementary advantages, reduce the risk of a single model and improve the reliability of prediction.
Drawings
FIG. 1 is a flow chart of a method for predicting state of charge of a lithium battery based on model fusion;
fig. 2 is a second order Thevenin equivalent circuit diagram provided by the invention.
Detailed Description
The invention is described in further detail below with reference to the accompanying drawings.
As shown in FIG. 1, the invention provides a lithium battery state of charge prediction method based on model fusion, which adopts a Savitzky-Golay filter to filter training data and then carries out z-score normalization, so that noise interference can be effectively reduced, data quality can be improved, and model efficiency can be improved; and establishing a second-order Thevenin equivalent circuit and a CrossFormer deep learning model. In order to improve EVO, improve searching efficiency and searching precision, a chaotic local searching strategy is introduced to improve EVO, an improved algorithm (IEVO) is obtained, the IEVO is utilized to conduct parameter identification on second-order Thevenin, then the lithium battery SOC is solved according to a calculation formula, finally, the entropy weight method is utilized to conduct weighted fusion on two groups of lithium battery SOCs, the charge state can be effectively predicted, and the model prediction precision is improved. The method comprises the following specific steps:
step 1, historical use data of a lithium battery are obtained, a data set comprises charge and discharge data, open-circuit voltage, lithium battery temperature, internal resistance, capacity, working current and corresponding charge state indexes of the lithium battery, a Savitzky-Golay filter is used for carrying out filtering treatment on training data, and then z-score normalization is carried out, so that noise interference can be effectively reduced, data quality is improved, data processing process is simplified, and performance and prediction accuracy of a model are improved.
The training data is filtered by adopting a Savitzky-Golay filter, and then the z-score normalization is carried out, comprising the following steps:
using an S-G filter to perform polynomial fitting on the data in the window, setting the width n=2m+1 of the filtering window, and setting the original signal length as N, wherein the data to be smoothed in the window: x is x i ＝(x -m ,x -m+1 ,…,x 0 ,x 1 ,…,x m-1 ,x m ) Polynomial fitting of the data within the window using the (k-1) th order polynomial:
y i ＝a 0 +a 1 x+a 2 x 2 +…+a k-1 x k-1 (1)
fitting parameters are determined using a least squares method. Such n equations constitute a system of k-element linear equations (n>k) Determining fitting parameter a by least square method 0 ,a 1 ,a 2 ,…,a k-1 . Namely:
the matrix expression form is as follows:
Y (2m+1)×1 ＝X (2m+1)×k ·A k×1 +E (2m+1)×1 (3)
least squares solution for AThe method comprises the following steps:
after filtering, the output value or predicted valueThe method comprises the following steps:
the z-score normalization process is specifically as follows:
wherein x is data to be normalized, x' is normalized data, mu is an average value of the data, and S is a standard deviation.
And 2, establishing a second-order Thevenin equivalent circuit model of the lithium battery, and transforming an equivalent circuit model equation into a mathematical form which contains terminal voltage, open-circuit voltage, capacitance and parameters to be solved of the lithium battery and can be identified by a computer, so that preparation is made for model identification.
Establishing a second order Thevenin as shown in fig. 2, the state space equation of which can be expressed by the following formula:
U O ＝U OC -U P1 -U P2 -IR (7)
wherein: u (U) P1 And U P2 Respectively the voltages of two RC loops in the Thevenin circuit model, R P1 ，R P2 For 2 polarization internal resistances of the cell, C P1 ，C P2 Two polarization capacitances for the battery; u (U) OC Represents an open circuit voltage, U O Is the terminal voltage; r is the internal resistance of the battery; i is the total current in the circuit, R P1 ，R P2 ，C P1 ，C P2 ，U OC Is the parameter to be identified.
And 3, improving an energy valley optimization algorithm EVO, optimizing the algorithm by adopting a chaotic local search strategy to obtain an IEVO algorithm, combining the mathematical form of the step with the improved energy valley optimization algorithm IEVO, calculating each parameter to be solved in the step 1, further completing parameter identification of an equivalent circuit model, obtaining a complete second-order Thevenin equivalent circuit model of the lithium battery, and calculating by using an ampere-hour integration method to obtain SOC prediction data SOC1 of the lithium battery.
The lithium battery state space equation is embedded into an energy valley optimization algorithm (EVO), and R is identified through repeated iteration 1 ,R 2 ,C 1 ,C 2 Equivalent circuit parameters.
Carrying out an initialization process to obtain a parameter X to be solved of a second-order Thevenin equivalent circuit model i Assume particles with different stability levels in the search space:
where n represents the total number of particles in the search space, d is the dimension of the problem under consideration,is the j-th decision variable determining the initial position of the i-th candidate,/or->And->Representing the lower and upper bounds, respectively, of j variables in the ith candidate variable, rand is at [0,1]Random numbers uniformly distributed within the range.
The optimal solution of each parameter in the second-order Thevenin equivalent circuit model can be found through an iterative process of simulating particles with different stable levels in a physical process to release energy through decay so as to reach the stable level.
Determining an enrichment boundary (EnrichmentBound, EB) of the particles for performing an objective function evaluation on each particle taking into account the difference between the neutron-rich particles and the neutron-poor particles, and determining a Neutron Enrichment Level (NEL) of the particles, expressed mathematically as follows:
wherein NEL i Is the neutron enrichment energy level of the ith particle, and EB is the enrichment boundary of the particles in the universe.
From the objective function evaluation, the stability level of the particles was determined as follows:
wherein is the SL th i The stability levels of the particles, BS and WS, are the best and worst particles for the stability levels in the universe, which correspond to the minimum and maximum values of the objective function values found at present, respectively.
In the main search loop of EVO, if the neutron enrichment level of the particles is higher than the enrichment limit (NEL i >EB) then it is interesting to use decay processes in alpha, beta or gamma format, assuming that the particles have a large N/Z ratio, in 0,1]Generating a random number in a range to simulate Stability Bound (SB) in the universe if the Stability level of the particles is higher than that of StabilityDelimitation (SL) i >SB), then alpha and gamma decay is considered to occur, expressed mathematically as one of the EVO's location update schemes, in which a new candidate solution is generated. Generating two random integers, i.e. in [1, d ]]Alpha Index1 (representing the number of rays emitted) and within [1, alpha Index1 ]]Alpha Index2 in the range (defining the Alpha rays to be emitted).
The rays emitted are the decision variables in the candidate solution, which are removed and are then processed by the method with the best stability level (X BS ) Is a particle or ray replacement in a candidate. The mathematical formulas for these aspects are as follows:
wherein the method comprises the steps ofIs a newly generated particle in the universe, X i Is the current position vector, X, of the ith particle (candidate solution) in the universe (search space) BS Is the position vector of the particle with the best stability level,/->Is the j-th decision variable or the ray emitted.
Furthermore, in Gamma decay, gamma rays are emitted to increase the stability level of the excited particles, so this aspect can be expressed mathematically as another location update process of EVO, in which new candidate solutions are generated, for which two random integers Gamma Index1 is in the range of [1, d ], representing the number of emitted photons; in this case, the total distance between the considered particle and the other particles is calculated as follows, and the nearest particle is utilized:
wherein the method comprises the steps ofIs the total distance between the ith particle and the kth adjacent particle, (x) 1 ,y 1 ) And (x) 2 ,y 2 ) Representing the coordinates of the particles in the search space, using these operations, the location update procedure to generate the second candidate solution at this stage is as follows:
wherein the method comprises the steps ofIs a newly generated particle in the universe, X i Is the current position vector, X, of the ith candidate solution in the search space Ng Is the position vector of adjacent particles around the ith particle,/->Is the j-th decision variable. Gamma Index2 is in [1, gamma Index1 ]]In the range, it is indicated which photons are considered in the particle. Photons in the particles are the determining variables in the candidate solution, they are removed and are either adjacent particles or candidate X Ng Instead, this mimics the interaction of the excited particles with other particles and even magnetic fields.
Adding a chaotic local search strategy to generate a group of initial solutions with randomness and uniformity by using a chaotic phenomenon:
wherein,representing a new solution generated at the t+1st iteration using the chaotic local search strategy,/o>And->Respectively representing the random selection of 2 different individuals from the dominant group, C t Representing the chaotic value generated by the chaotic map of the t-th iteration.
If the stability level of the particles is below the stability limit (S Li SB). Ltoreq.then beta.decay is considered to occur, the particle undergoes a site-updating process, for a particle having an optimum stable level X BS And particle center X CP These aspects of the algorithm simulate the tendency of the particles to reach a stable band, where most known particles are located near the band and most have higher stability, as follows:
by the current particle position X i Obtaining the current particle center position X CP ：
Expression typeRepresenting the new vector position, X BS For optimum particle position, SL i Is the stable level of the ith particle, r 1 ，r 2 Is [0,1]Two random numbers within the range determine the amount of movement of the particles.
If the neutron enrichment level of the particle is below the enrichment boundary (NEL) i And EB) the N/Z of the particle is considered to be relatively small, so the particle tends to move towards the stable band by electron capture or positron emission, in which respect random motion in the search space is determined to take account of these types of motion as follows:
in the middle ofAnd X i Is the future and current position vector of the ith candidate solution in the search space, r is [0,1 ] which determines the amount of particle motion]Random numbers within a range.
At the end of the EVO main cycle, if the enrichment level of particles is higher than the enrichment limit, only two new position vectors are generated per particle, respectivelyAnd->Whereas for particles with a lower enrichment level only +.>As a new position vector, in each state the newly generated vector is combined with the current population, the best particles participate in the next search cycle of the algorithm, boundary violations are determined for decision variables exceeding predefined upper and lower bounds, and the maximum number of objective function evaluations or the maximum number of iterations can be used as termination criteria.
And finally returning particles with optimal stability, so that an optimal solution of each parameter in the second-order Thevenin equivalent circuit model is obtained.
Step 3.3: obtaining an SOC value by an ampere-hour integration method:
wherein U is 1 (t)，U 2 And (t) is electrochemical polarization voltage of the lithium battery:
wherein C represents the capacitance of the lithium battery, i L And (t) is an operating current when the lithium battery is charged and discharged at the time t, and the SOC (0) represents an SOC1 value at the initial time of the lithium battery.
Step 4, the preprocessed stable lithium battery data set comprises the following steps: and the battery charge and discharge data, the internal resistance, the temperature, the capacity, the voltage, the current and the SOC data are input into a CrossFormer deep learning model after random initialization parameters for training, so that a trained deep learning model is obtained, and the lithium battery SOC prediction data SOC2 is obtained.
Establishing a CrossFormer model, and carrying out L2 regularization optimization on the CrossFormer deep learning model:
where L (w) is the loss function regularized by L2, loss (w, x) is the original loss function (e.g., mean square error or cross entropy) of the model, w is the weight vector of the model,is the square of the L2 norm of the weight vector, λ is the regularization parameter, used to control the strength of the regularization. The greater the regularization parameter λ, the greater the specific gravity of the regularization term in the loss function and the greater the penalty on the weights.
Embedding the preprocessed lithium battery history data set into a multi-element time series (MTS) using dimension-segment-wise (DSW): DSW embedding, dividing points in each dimension of a lithium battery history data set into segments with length Lseg, and then embedding:
length L in d dimension seg I-th section of (a).
Each line segment is embedded into a vector using linear projection plus position embedding:
representing a learnable projection matrix->The leachable locations representing the locations are embedded (i, d).
After embedding, a two-dimensional vector array is obtained:
wherein each h i,d Representing a univariate time series segment.
A two-stage attention (TSA) layer to effectively capture the dependency between the embedded segments of the processed lithium battery history data set. A Two-Stage Attention (TSA) layer captures data between different dimensions.
Given a two-dimensional arrayAs input to the TSA layer, L is the number of segments, D is the dimension, where Z can be the output of the DSW embedded or lower TSA layer i,: A vector representing all dimensions at time step i, Z :,d A vector representing all dimensions at time step d. In the cross-time phase, multi-headed self-attention (MSA) is directly applied to each dimension:
wherein D is equal to or greater than 1 and equal to or less than D, layerNorm represents widely used layer normalization, MLP represents a multilayer feedforward network, MSA (Q, K, V) represents a multi-headed self-care layer, wherein Q; k, performing K; v is used as inquiry, key and value, all dimensions 1-D share the same MSA layer,Z time representing the outputs of MSA and MLP, after this stage, at Z time The dependency relationship between time periods of the same dimension is captured, then Z time Becomes an input to the cross-dimensional phase to capture cross-dimensional dependencies.
And a second stage: in the latitudinal stage, a fixed number of leachable vectors (c < D) are set as routers for each time step i, the routers first aggregate messages from all dimensions by using the routers as queries in the MSA and using vectors of all dimensions as keys and values, and then the routers distribute received messages to the dimensions by using dimension vectors as queries and aggregating messages as keys and values, thus establishing a full-pair full-connection between the D dimensions:
in the middle of(c is a constant) as a router for a learnable vector array, < >>Is an aggregate message from all dimensions, +.>Representing the output of the router mechanism, all time steps (1.ltoreq.i.ltoreq.L) share the sameZ dim Representing a jump connection output and an MLP output.
The connection cross-time phase and cross-latitude phase equations can be obtained:
Y＝Z dim ＝TSA(Z) (32)
wherein Z is a group of the total number of the,representing input and output vector arrays of the TSA layer, respectively, every two segments (i.e. Z after cross-time and cross-dimension phases i1,d1 ,Z i2,d2 ) Are connected, thus capturing cross-time and cross-dimension dependencies in Y.
Layered encoder-decoder structures are used in a cross former model for multivariable lithium battery state-of-charge time prediction and to obtain information on different scales.
In each layer of the encoder (except the first layer), every two vectors that are temporally adjacent are combined to get a representation at a coarser level, and then the TSA layer is applied to capture the scale of the dependency, modeled as Z enc,l ＝Encoder(Z enc,l-1 )：
Wherein H is a two-dimensional array obtained by DSW embedding, Z enc,l Representing the output of the layer I encoder;representing a segment-merged leavable matrix, [ ·]Indicating the connection operation, L l-1 The number of segments representing each dimension in layer l-1, if not divisible by 2, will Z enc,l-1 Filling to a proper length; />Representing the combined array of the ith layer segment, assuming N layers in the encoder, Z is used enc,0 ,Z enc,1 ,…,Z enc,N Z enc,0 Representing the N +1 outputs of the encoder.
A decoder: an n+1 feature array of the encoder output is obtained, using an n+1 layer (index 0,1, …, N) for the predicted decoder. The layer 1 takes the array coded by the layer I as input, outputs the two-dimensional array decoded by the layer 1, and the process is summarized as follows: z is Z dec,l ＝Decoder(Z dec,l-1 ,Z enc,l )：
In the middle ofLearning position embedding representing decoder +.>For the output of TSA, MSA layer is provided with +.>As a query->As key and value, a connection between encoder and decoder is established, and the output of MSA is noted as Z dec,l Respectively represent the jumper output and the MLP output, and Z is used for dec,0 ,Z dec,1 ,…,Z dec,N Representing the decoder output, applying linear projection to the output of each layer to obtain a prediction for that layer, and summing the layer predictions to obtain a final prediction:
/>
wherein the method comprises the steps ofIs a learnable matrix for projecting vectors into time series segments.
Representing the ith segment of the prediction dimension d.
Rearranging all segments in layer l to obtain layer predictionsAdding the predictions of all layers to obtain the final prediction +.>
Hierarchical encoder-decoders construct a hierarchical encoder-decoder (HED) using proposed DSW embedding, TSA layer and segment merging, predict with larger scale information, add the predictors of different scales, and output the final result SOC2.
The emphasis of predicting the SOC of the lithium battery by using the cross former neural network is to learn the characteristics of the historical data of the lithium battery and fit a back nonlinear curve, and the process is as follows: data preparation: the method comprises the steps of collecting and sorting historical data of the lithium battery, wherein the data comprise time series data of multiple dimensions of voltage, current, temperature and the like; data preprocessing: preprocessing the collected historical data, including data cleaning, normalization processing, time sequence decomposition and the like, so as to ensure the quality and usability of the data; model training: inputting the preprocessed historical data into a cross former model for training, wherein the model learns cross dependencies among different dimensions, and the prediction capability of the model on the state of charge is built through the historical data; state of charge prediction: once the model training is completed, the model can be used for predicting the future state of charge, new time series data are input into a trained cross former model, and the model generates a prediction result of the future state of charge;
and 5, obtaining SOC output data of the two models in the step 3 and the step 4, and carrying out data fusion by an entropy weight method to obtain the final lithium battery state of charge.
Firstly, carrying out dimensionality removal treatment on each index, and carrying out data standardization. A total of three data sets { SOC 1 ,SOC 2 }, wherein SOC i ＝{x 1 ,x 2 ,…,x n Assume that the value normalized for each dataset is { SOC } 1 * ,SOC 2 * Then (S)
The variance size of each dataset was determined:
the information entropy of each data set is calculated, and according to the definition of the information entropy, the information entropy of one group of data is as follows:
wherein E is i And is more than or equal to 0. If p ij =0, definition E i ＝0。
Determining the weight of each index, and calculating the information entropy of each data set to be E according to a calculation formula of the information entropy 1 ,E 2 。
Calculating the weight of each index through information entropy:
the data fusion adopts an entropy weight method estimation algorithm, as shown in the formula:
in the formula, the SOC is the state of charge prediction of the final lithium battery after fusion.
The invention also provides an apparatus device comprising a memory and a processor, wherein the memory is for storing a computer program capable of running on the processor; a processor for performing the steps of the model fusion based lithium battery state of charge prediction method as described above when running the computer program.
The foregoing embodiments are merely illustrative of the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the present invention and to implement the same, not to limit the scope of the present invention. All equivalent changes or modifications made according to the spirit of the present invention should be included in the scope of the present invention.
Claims (8)
1. The lithium battery state of charge prediction method based on model fusion is characterized by comprising the following steps of:
(1) Acquiring initial data of a lithium ion battery, acquiring a training set required by a second-order Thevenin equivalent circuit model, and preprocessing the training set:
(2) Establishing a second-order Thevenin equivalent circuit model of the lithium battery, and transforming an equivalent circuit model equation into a mathematical form which contains terminal voltage, open-circuit voltage, capacitance and parameters to be solved and can be identified by a computer;
(3) Performing parameter identification on the second-order Thevenin equivalent circuit model based on an improved energy valley optimization algorithm IEVO to obtain an optimized second-order Thevenin equivalent circuit model of the lithium battery, and calculating by using an ampere-hour integration method to obtain SOC prediction data of the lithium battery; the improved energy valley optimization algorithm IEVO is used for carrying out local multi-element search on the position of the population in the iterative process by adopting a multi-kernel learning method; by adopting a chaotic local search strategy, finding out a better solution by searching a nearby area of each solution;
(4) Inputting the preprocessed lithium battery data set into a cross former deep learning model after random initialization parameters for training to obtain a trained deep learning model, thereby obtaining lithium battery SOC prediction data;
(5) And (3) carrying out data fusion on the SOC prediction data of the two models obtained in the step (3) and the step (4) through an entropy weight method to obtain the final lithium battery state of charge.
2. The method of claim 1, wherein the data set in step (1) includes charge and discharge data of the lithium battery, open circuit voltage, lithium battery temperature, internal resistance, capacity, operating current and respectively corresponding state of charge indexes.
3. The model fusion-based lithium battery state of charge prediction method of claim 1, wherein the data processing in the step (1) is to perform z-score normalization after performing filtering processing on training data by using a Savitzky-Golay filter.
4. The method for predicting the state of charge of a lithium battery based on model fusion according to claim 1, wherein in the step (2), a second-order Thevenin equivalent circuit model is established, and a state space equation of the equivalent circuit model is shown in the following formula:
U O ＝U OC -U P1 -U P2 -IR
in U P1 And U P2 The voltages of two RC loops in the second-order Thevenin equivalent circuit model are respectively; r is R P1 ，R P2 2 polarization internal resistances of the battery; c (C) P1 ，C P2 Two polarization capacitances for the battery; u (U) OC Represents an open circuit voltage; u (U) O Is the terminal voltage; r is the internal resistance of the battery; i is the total current in the circuit; r is R P1 ，R P2 ，C P1 ，C P2 ，U OC Is the parameter to be identified.
5. The method for predicting the state of charge of a lithium battery based on model fusion according to claim 1, wherein the implementation process of the step (3) is as follows:
the lithium battery state space equation is embedded into an energy valley optimization algorithm EVO, and R is identified through repeated iteration 1 ,R 2 ,C 1 ,C 2 Equivalent circuit parameters;
carrying out an initialization process to obtain a parameter X to be solved of a second-order Thevenin equivalent circuit model i Assume particles with different stability levels in the search space:
where n represents the total number of particles in the search space, d is the dimension of the problem under consideration,is the j-th decision variable determining the initial position of the i-th candidate,/or->And->Representing the lower and upper bounds, respectively, of j variables in the ith candidate variable, rand is at [0,1]Random numbers uniformly distributed in the range;
adding a chaotic local search strategy to generate a group of initial solutions with randomness and uniformity by using a chaotic phenomenon:
wherein,representing a new solution generated at the t+1st iteration using the chaotic local search strategy,/o>And->Respectively representing the random selection of 2 different individuals from the dominant group, C t Representing a chaotic value generated by the chaotic map of the t-th iteration;
obtaining an SOC value by an ampere-hour integration method:
wherein U is 1 (t)，U 2 And (t) is electrochemical polarization voltage of the lithium battery:
wherein C represents the capacitance of the lithium battery, i L And (t) is an operating current when the lithium battery is charged and discharged at the time t, and the SOC (0) represents an SOC1 value at the initial time of the lithium battery.
6. The method for predicting the state of charge of a lithium battery based on model fusion according to claim 1, wherein the implementation process of the step (4) is as follows:
l2 regularization optimization is carried out on a CrossFormer deep learning model:
where L (w) is the loss function regularized by L2, loss (w, x) is the original loss function of the model, w is the weight vector of the model,is the square of the L2 norm of the weight vector, λ is the regularization parameter, used to control the strength of the regularization;
embedding a multivariate time sequence using dimension segmentation; the layered encoder-decoder structure predicts in a cross former deep learning model by using larger-scale information, adds the predicted values of different scales, and outputs a final result SOC2.
7. The method for predicting the state of charge of a lithium battery based on model fusion according to claim 1, wherein the implementation process of the step (5) is as follows:
performing dimensionality removal treatment on each index, and performing data standardization: two data sets { SOC 1 ,SOC 2 }, wherein SOC i ＝{x 1 ,x 2 ,…,x n Normalized value for each dataset is { SOC } 1 * ,SOC 2 * -then:
the variance size of each dataset was determined:
the information entropy of each data set is calculated, and according to the definition of the information entropy, the information entropy of one group of data is as follows:
wherein E is i Not less than 0, if p ij =0, definition E i =0; determining the weight of each index, and calculating the information entropy of each data set to be E according to a calculation formula of the information entropy 1 ,E 2 ；
Calculating the weight of each index through information entropy:
the data fusion adopts an entropy weight method estimation algorithm, as shown in the formula:
in the formula, the SOC is the state of charge prediction of the final lithium battery after fusion.
8. An apparatus device comprising a memory and a processor, wherein:
a memory for storing a computer program capable of running on the processor;
a processor for performing the steps of the model fusion based lithium battery state of charge prediction method according to claims 1 to 7 when running the computer program.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311820312.2A CN117783875B (en) | 2023-12-26 | Lithium battery state of charge prediction method and device based on model fusion |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311820312.2A CN117783875B (en) | 2023-12-26 | Lithium battery state of charge prediction method and device based on model fusion |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN117783875A true CN117783875A (en) | 2024-03-29 |
| CN117783875B CN117783875B (en) | 2024-08-30 |
Family
ID=
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JPH06511310A (en) * | 1991-06-13 | 1994-12-15 | アライド シグナル インコーポレイテッド | How to charge and test rechargeable storage batteries |
| CN112116952A (en) * | 2020-08-06 | 2020-12-22 | 温州大学 | Gene selection method of wolf optimization algorithm based on diffusion and chaotic local search |
| US20230076118A1 (en) * | 2021-09-03 | 2023-03-09 | Electronics And Telecommunications Research Institute | Device and method for predicting state of battery |
| CN116186534A (en) * | 2022-12-23 | 2023-05-30 | 北京百度网讯科技有限公司 | Pre-training model updating method and device and electronic equipment |
| CN116679211A (en) * | 2023-05-18 | 2023-09-01 | 淮阴工学院 | Lithium battery health state prediction method |
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JPH06511310A (en) * | 1991-06-13 | 1994-12-15 | アライド シグナル インコーポレイテッド | How to charge and test rechargeable storage batteries |
| CN112116952A (en) * | 2020-08-06 | 2020-12-22 | 温州大学 | Gene selection method of wolf optimization algorithm based on diffusion and chaotic local search |
| US20230076118A1 (en) * | 2021-09-03 | 2023-03-09 | Electronics And Telecommunications Research Institute | Device and method for predicting state of battery |
| CN116186534A (en) * | 2022-12-23 | 2023-05-30 | 北京百度网讯科技有限公司 | Pre-training model updating method and device and electronic equipment |
| CN116679211A (en) * | 2023-05-18 | 2023-09-01 | 淮阴工学院 | Lithium battery health state prediction method |
Non-Patent Citations (3)
| Title |
|---|
| SHIXI HOU 等: "Neural Networks for State-of-Charge Estimation of Li-ion Batteries", 《SMART ELECTRICAL SYSTEMS ＆ TECHNOLOGY》, 31 December 2021 (2021-12-31), pages 729 - 732 * |
| YIDUO ZHU 等: "State of Charge Estimation Based on State of Health Correction for Lithium-ion Batteries", 《2018 IEEE INTELLIGENT VEHICLES SYMPOSIUM (IV)》, 31 December 2018 (2018-12-31), pages 1578 - 1583 * |
| 万广伟 等: "锂离子电池SOC评估方法研究进展", 《电源技术》, vol. 47, no. 9, 30 September 2023 (2023-09-30), pages 1122 - 1125 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Ma et al. | A novel method for state of health estimation of lithium-ion batteries based on improved LSTM and health indicators extraction | |
| CN109991542B (en) | Lithium ion battery residual life prediction method based on WDE optimization LSTM network | |
| Lu et al. | Battery degradation prediction against uncertain future conditions with recurrent neural network enabled deep learning | |
| CN113219343A (en) | Lithium battery health state prediction method, system, equipment and medium based on elastic network | |
| CN107132490B (en) | Method for estimating state of charge of lithium battery pack | |
| CN112711892A (en) | Cloud battery management system and method based on digital twin and block chain technology | |
| CN115586444A (en) | Lithium battery residual life prediction method based on VMD and BP neural network | |
| CN116106761A (en) | Real-time lithium ion battery electric quantity estimation method based on typical correlation analysis | |
| Bak et al. | Accurate estimation of battery SOH and RUL based on a progressive lstm with a time compensated entropy index | |
| CN116298936A (en) | Intelligent lithium ion battery health state prediction method in incomplete voltage range | |
| Chen et al. | Remaining useful life prediction of the lithium-ion battery based on CNN-LSTM fusion model and grey relational analysis. | |
| CN114757340A (en) | Lithium battery health state prediction method and system based on neural network integration | |
| Reza et al. | Recent advancement of remaining useful life prediction of lithium-ion battery in electric vehicle applications: A review of modelling mechanisms, network configurations, factors, and outstanding issues | |
| Tang et al. | Model parameter identification for lithium-ion batteries using adaptive multi-context cooperatively co-evolutionary parallel differential evolution algorithm | |
| Ghadbane et al. | Optimal parameter identification strategy applied to lithium-ion battery model for electric vehicles using drive cycle data | |
| Selvaraj et al. | A Bayesian optimized machine learning approach for accurate state of charge estimation of lithium ion batteries used for electric vehicle application | |
| CN117169743A (en) | Battery health state estimation method and device based on partial data and model fusion | |
| CN117783875B (en) | Lithium battery state of charge prediction method and device based on model fusion | |
| Zhou et al. | Remaining useful life prediction and state of health diagnosis of lithium-ion batteries with multiscale health features based on optimized CatBoost algorithm | |
| Channegowda et al. | An attention‐based synthetic battery data augmentation technique to overcome limited dataset challenges | |
| CN117783875A (en) | Lithium battery state of charge prediction method and device based on model fusion | |
| Udeogu et al. | Remaining useful life prediction for supercapacitors using an optimized end-to-end deep learning approach | |
| Li et al. | Probabilistic neural network-based flexible estimation of lithium-ion battery capacity considering multidimensional charging habits | |
| El Marghichi | Estimation of battery capacity using the enhanced self-organization maps | |
| Ma et al. | A Battery Prognostics and Health Management Technique Based on Knee Critical Interval and Linear Complexity Self-Attention Transformer in Electric Vehicles |
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 |