CN108804800A - Lithium ion battery SOC on-line prediction methods based on echo state network - Google Patents
Lithium ion battery SOC on-line prediction methods based on echo state network Download PDFInfo
- Publication number
- CN108804800A CN108804800A CN201810565231.5A CN201810565231A CN108804800A CN 108804800 A CN108804800 A CN 108804800A CN 201810565231 A CN201810565231 A CN 201810565231A CN 108804800 A CN108804800 A CN 108804800A
- Authority
- CN
- China
- Prior art keywords
- network
- echo state
- state network
- input
- training
- 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.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 48
- HBBGRARXTFLTSG-UHFFFAOYSA-N Lithium ion Chemical compound [Li+] HBBGRARXTFLTSG-UHFFFAOYSA-N 0.000 title claims abstract description 14
- 229910001416 lithium ion Inorganic materials 0.000 title claims abstract description 14
- 238000012549 training Methods 0.000 claims abstract description 71
- 238000012360 testing method Methods 0.000 claims abstract description 62
- 238000002790 cross-validation Methods 0.000 claims abstract description 16
- 230000008569 process Effects 0.000 claims abstract description 11
- 230000000306 recurrent effect Effects 0.000 claims abstract description 5
- 239000011159 matrix material Substances 0.000 claims description 40
- 230000003595 spectral effect Effects 0.000 claims description 24
- 238000006073 displacement reaction Methods 0.000 claims description 22
- 210000002569 neuron Anatomy 0.000 claims description 22
- WHXSMMKQMYFTQS-UHFFFAOYSA-N Lithium Chemical compound [Li] WHXSMMKQMYFTQS-UHFFFAOYSA-N 0.000 claims description 8
- 229910052744 lithium Inorganic materials 0.000 claims description 8
- 230000005611 electricity Effects 0.000 claims description 4
- 240000002853 Nelumbo nucifera Species 0.000 claims 1
- 235000006508 Nelumbo nucifera Nutrition 0.000 claims 1
- 235000006510 Nelumbo pentapetala Nutrition 0.000 claims 1
- 238000013528 artificial neural network Methods 0.000 abstract description 7
- 238000004364 calculation method Methods 0.000 description 7
- 230000008859 change Effects 0.000 description 7
- 238000012795 verification Methods 0.000 description 6
- 230000006870 function Effects 0.000 description 5
- 210000005036 nerve Anatomy 0.000 description 5
- 230000008901 benefit Effects 0.000 description 3
- 210000004027 cell Anatomy 0.000 description 3
- 238000013480 data collection Methods 0.000 description 3
- 238000010998 test method Methods 0.000 description 3
- 230000004913 activation Effects 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 2
- 239000012141 concentrate Substances 0.000 description 2
- 238000001514 detection method Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000005284 excitation Effects 0.000 description 2
- NAWXUBYGYWOOIX-SFHVURJKSA-N (2s)-2-[[4-[2-(2,4-diaminoquinazolin-6-yl)ethyl]benzoyl]amino]-4-methylidenepentanedioic acid Chemical compound C1=CC2=NC(N)=NC(N)=C2C=C1CCC1=CC=C(C(=O)N[C@@H](CC(=C)C(O)=O)C(O)=O)C=C1 NAWXUBYGYWOOIX-SFHVURJKSA-N 0.000 description 1
- 241000208340 Araliaceae Species 0.000 description 1
- 201000004569 Blindness Diseases 0.000 description 1
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 description 1
- 235000003140 Panax quinquefolius Nutrition 0.000 description 1
- 238000009825 accumulation Methods 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 239000008358 core component Substances 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000003912 environmental pollution Methods 0.000 description 1
- 235000008434 ginseng Nutrition 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 238000009533 lab test Methods 0.000 description 1
- 238000012886 linear function Methods 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 230000001537 neural effect Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 239000000843 powder Substances 0.000 description 1
- 238000005096 rolling process Methods 0.000 description 1
- 239000007787 solid Substances 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 208000011580 syndromic disease Diseases 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computer Hardware Design (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Geometry (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Secondary Cells (AREA)
Abstract
The present invention discloses a kind of lithium ion battery SOC straight line prediction techniques based on echo state network, k folding cross-validation methods are applied to the preferred process of multiple uncertain parameters of echo state network, simplify the process for finding optimized parameter, simultaneously during finding suitable training set and test set, with certain multiple training sets of gradient difference spacing primary election and test set training and test network, according to the error size of training and test, consider and selects rational training set and test set, ensure to a certain extent so that network has stronger generalization ability, promote neural network forecast precision.In addition, also training echo state network using the recurrent least square method with forgetting factor, then according to most freshly harvested battery data, network is adjusted in real time and exports weights, it is ensured that the on-line prediction of network.
Description
Technical field
The present invention relates to battery performance electric powder predictions, and in particular to a kind of lithium-ion electric based on echo state network
Pond SOC on-line prediction methods.
Background technology
With increasingly serious, modern new energy the problems such as environmental pollution, energy crisis caused by orthodox car industry in recent years
The fast development of source electric car will constantly alleviate the above problem.Compared with orthodox car, New-energy electric vehicle is with no dirt
The advantages that dye discharge, clean energy.Core component and power source of the car lithium battery as electric vehicle, can store
Electricity determines the course continuation mileage of electric vehicle.The state-of-charge SOC of car lithium battery be one can not physics measured directly
Amount, existing technology can only indirectly detect SOC, and common detection method has following several：1) current integration method, just
Beginning SOC is difficult to determine have error accumulation.2) coupled circuit method depends on battery model, parameter identification relatively difficult.3) it discharges
Test method(s) is applicable in all batteries, and generally applicable battery maintenance or laboratory test are not suitable for real-time working condition monitoring.Based on
Upper analysis still has problems with using above method detection battery SOC：1) it is difficult to establish accurately battery model；2) join
Number identification calculates complexity, inefficient；3) it cannot achieve online fast prediction battery SOC.
Invention content
To be solved by this invention is the problem of existing method can only indirectly detect battery SOC, provides one kind
Lithium ion battery SOC on-line prediction methods based on echo state network.
To solve the above problems, the present invention is achieved by the following technical solutions：
Lithium ion battery SOC on-line prediction methods based on echo state network, specifically comprise the following steps：
Step 1, structure echo state network determine the uncertain of constructed echo state network using k folding cross validations
Parameter obtains echo state network model；
Step 1.1 acquires M group battery currents I, voltage V, battery pack temperature T with harvester, and provides battery producer
SOC-OCV curve discretizations obtain the state-of-charge SOC of corresponding M groups, by electric current I, voltage V, battery pack temperature T and corresponding
State-of-charge SOC is divided into k parts of data sets；
Step 1.2, the initial parameter for setting echo state network, including input dimension K export dimension L, reserve pool nerve
The variation range and step-length of first number N, the variation range and step-length of spectral radius SR, the variation range and step-length of input scaling IS,
The variation range and step-length and error threshold minerror of input displacement IF；
Step 1.3, echo state network input scaling IS arbitrarily chosen in its variation range and be kept fixed it is constant,
Input displacement IF arbitrarily chooses in its variation range and is kept fixed constant, and reserve pool neuron number N is in its variation range
Interior to be changed with its step-length, spectral radius SR is changed in its variation range with its step-length；And it will acquire in step 1.1
K parts of data collection are sequentially inputted to these echo state networks using k folding cross validation modes and are trained and test, when wherein certain
Each training error trainerror obtained by the valued combinations of one reserve pool neuron number N and spectral radius SR and corresponding
When test error testerror is satisfied by optimal conditions, then the value of the reserve pool neuron number N and spectral radius SR is back
The optimal reserve pool neuron number N and spectral radius SR found needed for sound state network；
Step 1.4, echo state network reserve pool neuron number N arbitrarily chosen in its variation range and keep solid
Fixed constant, spectral radius SR arbitrarily chooses in its variation range and is kept fixed constant, and input scaling IS is in its variation range
It is changed with its step-length, input displacement IF is changed in its variation range with its step-length；And it will acquire in step 1.1
K parts of data collection are sequentially inputted to these echo state networks using k folding cross validation modes and are trained and test, when wherein certain
Each training error trainerror and corresponding test obtained by the valued combinations of one input scaling IS and input displacement IF
When error testerror is satisfied by optimal conditions, then input scaling IS and input displacement IF is needed for echo state network
The optimal input found scales IS and input displacement IF；
Step 1.5, according to optimal reserve pool neuron number N and spectral radius SR determined by step 1.3 and step 1.4 institute
Determining optimal input scales IS and input displacement IF to build echo state network module；
Step 2, in real time acquisition lithium battery real-time current I, voltage V in the process of running and battery pack external skin temperatures T, and
Real-time current I, voltage V and battery pack external skin temperatures T are input in echo state network model, the echo state network model
The state-of-charge SOC of lithium battery that as predicts in real time of output.
In above-mentioned steps 1.2, when setting the initial parameter of echo state network, it is also necessary to initialize echo state network
Structural parameters, including input weight matrix, reserve pool internal state weight matrix and feedback weight matrix.
In above-mentioned steps 1.2, input weight matrix W dimensions are N × K, reserve pool internal weights matrix WinDimension is N × N,
Feedback weight matrix WbackDimension is N × L.
In above-mentioned steps 1.2, it is [0,1] that spectral radius SR, input, which scale the variation range of IS and input displacement IF,.
Optimal conditions in above-mentioned steps 1.3 and 1.4 is：
[(trainerror+testerror)/2]<minerror
Wherein, trainerror is training error, and testerror is test error, and minerror is initial setting error
Threshold value.
Above-mentioned steps 1.5 still further comprise：Using the recurrent least square method with forgetting factor to building echo state
Network module carries out the process of network training update output weight matrix.
Compared with prior art, k folding cross-validation methods are applied to multiple uncertain ginsengs of echo state network by the present invention
Several preferred process, provide it is a kind of choose network optimized parameter method, simplify find optimized parameter process, while
During finding suitable training set and test set, with certain multiple training sets of gradient difference spacing primary election and test set training and
Test network considers according to the error size of training and test and selects rational training set and test set, it is ensured that one
Determine in degree so that network has stronger generalization ability, promotion neural network forecast precision.It is minimum using the recurrence with forgetting factor
Square law trains echo state network, then according to most freshly harvested battery data, adjusts network in real time and exports weights, it is ensured that net
The on-line prediction of network.
Description of the drawings
Fig. 1 is that echo state network predicts battery SOC flow chart.
Fig. 2 is echo state network structure chart.
Fig. 3 is k folding fork verification flow charts.
Fig. 4 is that k rolls over cross validation training and test set divides figure.
Specific implementation mode
To make the objectives, technical solutions, and advantages of the present invention clearer, below in conjunction with specific example, and with reference to attached
Figure, the present invention is described in more detail.
Compared to traditional neural network, echo state network will uniformly be attributed to a storage except the other parts output and input
Standby pool structure, simplifies network structure, and reserve pool neuron number is much larger with respect to traditional neural member hidden neuron scale, storage
Standby pond is rich in abundant network world, solves the problems, such as that traditional neural network structure can not determine, and network calculations are only with instruction
Practice output weight matrix, other weight matrixs are randomly generated, immobilized, and which simplify network calculations, since reserve pool is rich
Rich network world, substantially increases network calculations efficiency and learning ability.In addition, echo state network is without setting up any electricity
Pool model, the network model is by the way that a large amount of analyses of battery history data and network training, battery is excavated and analyzed to depth
The influence that the factors such as electric current, voltage and temperature change battery SOC, so that network has certain extensive energy after training
Power, echo state network simplify network structure, reduce calculation amount, but also bring multiple incoherent unknown parameters simultaneously
(reserve pool scale N, spectral radius SR, input displacement IF, input scaling IS) can not accurately select this problem, tradition to choose optimal
Parametric technique be mainly to be chosen by the method for exhaustion and test method(s), method of exhaustion calculation amount is huge, carry out it is very difficult,
Test method(s) has many uncertain and blindness, therefore how fast accurate optimized parameter of choosing makes echo state network performance
The optimal emphasis for the present invention.
K rolls over cross-validation method, initial cells data is divided into k parts of data sets, an individual data subset, which is used as, to be tested
The test set of model of a syndrome is left k-1 data set sample as training set, and repeated overlapping is verified k times, and corresponding optimizing parameter is same
When with certain step change, the parameter obtained corresponding to effect best training set and test set is exactly optimized parameter.This
The advantage of method is, in given data collection and network model, while repeating to be trained with the data set randomly generated
And verification, obtain that result verification is primary every time, in this way when not knowing the particular kind of relationship between optimizing parameter and model, with big
The data training of amount and test model and the substitution model of optimizing parameter gradients variation are trained and test, and make in theory
Being optimal of model is obtained, corresponding is exactly required optimized parameter.
Referring to Fig. 1, the present invention proposes a kind of lithium ion battery SOC on-line prediction methods based on echo state network, profit
Cross-validation method is rolled over k to optimize four important parameters of echo state network：Inside reserve pool neuron number N, reserve pool
Weights spectral radius SR, network inputs scale IS and displacement IF, and this method effectively accelerates the speed that network chooses optimized parameter,
And the echo state network of certain generalization ability is obtained by multigroup different number of data training and test, it is adopted to fit
The certain Nonlinear Mapping relationship of the battery related data and battery SOC of collection, while ensureing that network has certain generalization ability
It predicts battery SOC, specifically includes following steps：
Step 1：Echo state network model is built, the uncertain parameter of network is chosen using k folding cross validations, determines network knot
Structure.Echo state network structure is as shown in Figure 2.
Step 1.1：Use harvester acquisition M group battery currents I, voltage V, battery pack temperature T as network inputs data,
And the SOC-OCV curves provided by battery producer, discretization obtain the state-of-charge SOC of corresponding M groups output data as a comparison,
Network inputs data and corresponding comparison output data are divided into k parts of data sets, every part of data concentrate the network inputs contained
Data and corresponding comparison output data are M/k groups.
Step 1.2：It goes to choose echo state net using the obtained data set of step 1.1, and using k folding cross-validation methods
Connection weight matrix spectral radius SR inside the optimized parameter of network, including reserve pool neuron number N, reserve pool, input displacement IF,
Input scaling IS.K folding fork verifications flow is as shown in Figure 3.
Step 1.2.1：It sets echo state network and inputs dimension as K, output dimension is L, and reserve pool neuron number N is pressed
According to 5 being change step shown in table 1, spectral radius SR, input scaling IS and input displacement IF are change step with 0.05.Setting is handed over
Fork verification training and test error threshold value minerror (error threshold can carry out adjustment appropriate according to actual conditions).It is set
Spectral radius SR, input scaling IS it is consistent with the variation range of input displacement IF, step-length is consistent or inconsistent.
Initialize echo state network structural parameters, including input weight matrix, reserve pool internal state weight matrix and
Feedback weight matrix, wherein input weight matrix W dimensions are N × K, reserve pool internal weights matrix WinDimension is N × N, feedback
Weight matrix WbackDimension is N × L.In [- 1,1], size and symbol randomly generate above-mentioned each weight matrix element value range.
Table 1
| Parameter | Range | Step-length | Step number |
| N | [30,150] | 5 | 24 |
| SR | [0,1] | 0.05 | 20 |
| IS | [0,1] | 0.05 | 20 |
| IF | [0,1] | 0.05 | 20 |
Step 1.2.2：It concentrates kth part as test set the k acquired in step 1.1 part data, is left k-1 parts as training
Collection, chooses the value of IS and IF arbitrarily between 0 to 1, and remains unchanged, and reserve pool neuron N and spectral radius SR are respectively with 5 Hes
0.05 step-length is changed, and change frequency is 24 times and 20 times respectively, therefore has obtained 480 groups in first time training and test
Training error and test error.
Training error：Network inputs data (i.e. electric current I, voltage V, battery pack temperature T) in training set are input to echo
Be trained obtained network output data (i.e. state-of-charge SOC) in state network, and by the network output data with it is right
The comparison output data answered is compared obtained error.
Test error：Network inputs data (i.e. electric current I, voltage V, battery pack temperature T) in test set are input to echo
Carry out testing obtained network output data (i.e. state-of-charge SOC) in state network, and by the network output data with it is right
The comparison output data answered is compared obtained error.
Step 1.2.3：It regard -1 part of kth in step 1.1 as test set, remaining data are as training set, IS and IF
It remains unchanged, N and SR are changed with 5 and 0.05 for step-length respectively, similarly obtain 480 groups of training errors and test error.
Step 1.2.4：Similarly using -2 parts of data of kth in step 1.1 as test set, remaining data are training set,
IS and IF are remained unchanged, and N and SR are changed with 5 and 0.05 for step-length respectively, obtain 480 groups of training errors and test error.
Step 1.2.5：And so on training set and test set be changed according to Fig. 4, while N and SR are respectively with 5 Hes
0.05 is changed for step-length, and the variation of training set and test set is until using first part of data as test set, with the 2nd part to kth
Until part data are training set, the change of k training set and test set is completed.
Step 1.2.6：All training sets and test set are calculated and completed, then obtained training error and test is added up to miss
Difference is 480*k groups.By in 480*k group data, by the valued combinations of the value and spectral radius SR of each reserve pool neuron N, institute
Determining echo state network obtained training error trainerror and test error testerror is classified as one group, if returning
Some training set and test set training error and test error of sound state network reach setting condition i.e. formula (1) when, then can be with
Think that optimizing parameter N and SR corresponding to the echo state network are the optimized parameters it needs to be determined that echo state network.
((trainerror+testerror)/2)<minerror (1)
Step 1.2.7：Using mode identical with step 1.2.2-1.2.6, the N preferably gone out and SR are immobilized, IS
It is changed with 0.05 step-length with IF, reuses the k folding preferred IS and IF of cross validation, obtain 20*20*k=400*k groups instruction
White silk and test error.
Step 1.2.8：So far, k folding cross-validation method preferred echo state network parameter N, SR, IS and IF are completed, is returned
Sound state network model foundation is completed.
Echo state network basic structure is as shown in Fig. 2.Network inputs number of nodes is K, and output node number is L, deposit
Pond intrinsic nerve member number is N.The input signal u (n) of input node=[u1(n), u2(n) ..., uK(n)], echo state net
Network reserve pool intrinsic nerve member handles signal x (n)=[x1(n), x2(n) ..., xN(n)], echo state network output node
Output signal y (n)=[y1(n), y2(n) ..., yL(n)], renewal equation such as formula (2) and (3) inside reserve pool：
X (n+1)=f (W × x (n)+Win×u(n+1)+Wback×y(n)) (2)
Y (n+1)=fout(Wout(u (n+1), x (n+1), y (n))) (3)
Wherein, f (*)=[f1, f2..., fN], N is neuron number, and f (*) is reserve pool intrinsic nerve member activation primitive,
Generally S type functions or SIN function, input signal obtain x (n), y under the excitation by reserve pool internal state excitation function
(n) it is n in output function treated network output, foutTo export activation primitive, generally linear function.
Step 2：It is preferred that training set and test set, network instruction is carried out using the recurrent least square method RLS with forgetting factor
Practice, acquires real-time current I, voltage V and the battery pack external skin temperatures of lithium battery in the process of running after the completion of network training in real time
T, and using real-time current I, voltage V and battery pack external skin temperatures T as the input of echo state network model, lithium electricity is predicted in real time
The SOC in pond.
Step 2.1：With harvester acquisition M group battery currents I, voltage V, battery pack temperature T and by SOC-OCV curves
Discretization obtains the state-of-charge SOC of corresponding M groups as network preprocessed data early period, and network training collection and test set are according to table
It is chosen shown in 2, the error obtained according to training and test and network operation time are come preferred optimal training and and test
Collection, under the principle minimized the error for ensureing training and test, the shortest training set of the network operation and test set are as optimal
Training set and test set；
Table 2
| Serial number | Training scale | Measurement scope |
| 1 | 50% | 50% |
| 2 | 60% | 40% |
| 3 | 70% | 30% |
| 4 | 80% | 20% |
Step 2.2：According to optimal training set and test set that selection obtains, network training is carried out, can be obtained from deposit
The x (n) of pond state space is collected into matrix A, and network output y (n) is collected into matrix B, calculates output weight matrix Wout, dimension L
× (K+N+L), then exporting weight matrix is：
(Wout)T=A-1B (4)
To ensure minimizing the error between network output and teacher's supervisory signals, then Constrained：
It is E (k) that output weight matrix, which must meet error between above-mentioned output and teacher's supervisory signals, then has
The forgetting factor λ for introducing RLS algorithm then has J (n) satisfactions：
The introducing of RLS algorithm forgetting factor weakens what historical data farther out adjusted weight matrix to a certain extent
Influence, distance n moment closer data, forgetting factor can then reinforce the network information of new data offer so that band forget because
The least square method of son can export weights to network training and make quick adjustment, that is, the introducing of the new data acquired can make RLS
Fast reaction is made to required output weight matrix so that network, which accelerates, seeks the weight matrix W so that J (n) minimumsout, then
The input weight matrix of formula (5) can be met by asking local derviation to calculate：
Step 2.3：To avoid network training from being absorbed in local optimum, when output weight matrix meets formula (8), setting constraint
Error MINERROR preserves output weight matrix if training error is less than constraint error.Therefore error function J (n) is obtained
When extreme value, the error constraints for having reached set can be approximately considered the corresponding W of gainedoutExactly meet network optimal conditions
Weight matrix is exported, the renewal equation such as formula that the output weight matrix of ESN adjusts in real time can be obtained by the RLS algorithm with forgetting factor
(9)：
Wout k+1=Wout k+QkE(k) (9)
yi(n)=Wi out×xi(n) (10)
Step 2.4：So far off-line training and the test for having completed echo state network model, will acquire lithium ion in real time
When electric current I, voltage V and the battery pack temperature T of battery in the process of running are inputted as echo state network, inside reserve pool
State renewal equation (4) and (5) and the output real-time adjustment type of weights (9) carry out recursion prediction, you can online according to formula (10)
The current SOC states of assessment prediction.
Echo state network of the present invention includes input layer, reserve pool and output layer, and the network is rolled over by K and intersected
Verification has chosen network optimized parameter, including reserve pool neuron number N, reserve pool intrinsic nerve member spectral radius SR, input scaling
IS and input displacement IF simplifies the process of parameter optimization, and training only exports weight matrix with calculating after netinit, with
Traditional neural network is compared, and network calculations efficiency is greatly improved, and is trained using different gradient training sets and test set,
Using least square in training network, comparing calculation comes out corresponding error, chooses optimal training set and test set, optimizes net
Network performance improves the generalization ability of network.Weights square is exported using the recurrent least square method real-time update with forgetting factor
Battle array, realizes the on-line prediction function of SOC.
The echo state network model and device of a kind of prediction battery SOC include data acquisition unit, modeling unit, initial
It is single to change unit, parameter selection unit, computing unit, judging unit, network training and test cell, neural network forecast unit and output
Member.
The data acquisition unit, for acquiring the real-time current I of battery, voltage V, temperature T and using open circuit voltage method
SOC-OCV curves are obtained, and discretization obtains state-of-charge SOC.
The modeling unit is K for establishing an input number of nodes, and reserve pool scale is N, and output node is the mark of L
Quasi- echo state network model.
The parameter selection unit, with least square in training network, chooses network for rolling over cross-validation method according to k
Optimized parameter N, SR, IS, IF.
Institute's initialization unit, for initializing network inputs weight matrix, reserve pool weight matrix, feedback weight matrix etc.
Parameter.
The computing unit calculates error for network training, chooses optimized parameter, and the calculating of output weights.
The judging unit is additionally operable to sentence for judging whether network training and test error reach setting constraints
Whether circuit network output weights can allow network reality output and the error of teacher's supervisory signals time to reach set fixed condition.
The network training and test cell, for choosing network optimized parameter, and the optimal training set of selection and test
Collection, is additionally operable to training and the test network generalization ability of whole network model.
The neural network forecast unit predicts the real output value of network for trained network model.
The network output unit, the prediction for exporting network export.
It should be noted that although the above embodiment of the present invention is illustrative, this is not to the present invention
Limitation, therefore the invention is not limited in above-mentioned specific implementation mode.Without departing from the principles of the present invention, every
The other embodiment that those skilled in the art obtain under the inspiration of the present invention is accordingly to be regarded as within the protection of the present invention.
Claims (6)
1. the lithium ion battery SOC on-line prediction methods based on echo state network, characterized in that including steps are as follows：
Step 1, structure echo state network determine the uncertain parameter of constructed echo state network using k folding cross validations,
Obtain echo state network model；
Step 1.1 acquires M group battery currents I, voltage V, battery pack temperature T with harvester, and battery producer is provided
SOC-OCV curve discretizations obtain the state-of-charge SOC of corresponding M groups, by electric current I, voltage V, battery pack temperature T and corresponding lotus
Electricity condition SOC is divided into k parts of data sets；
Step 1.2, the initial parameter for setting echo state network, including input dimension K, export dimension L, reserve pool neuron number
The variation range and step-length of mesh N, the variation range and step-length of spectral radius SR, the variation range and step-length of input scaling IS, input
The variation range and step-length and error threshold minerror of displacement IF；
The input of step 1.3, echo state network scales IS and is arbitrarily chosen in its variation range and be kept fixed constant, input
Displacement IF arbitrarily chosen in its variation range and be kept fixed it is constant, reserve pool neuron number N in its variation range with
Its step-length is changed, and spectral radius SR is changed in its variation range with its step-length；And by the k acquired in step 1.1 parts
Data set is sequentially inputted to these echo state networks using k folding cross validation modes and is trained and tests, when wherein a certain
Each training error trainerror and corresponding survey obtained by the valued combinations of reserve pool neuron number N and spectral radius SR
When examination error testerror is satisfied by optimal conditions, then the value of the reserve pool neuron number N and spectral radius SR is echo
The optimal reserve pool neuron number N and spectral radius SR found needed for state network；
Step 1.4, echo state network reserve pool neuron number N arbitrarily choose and be kept fixed not in its variation range
Become, spectral radius SR arbitrarily chosen in its variation range and be kept fixed it is constant, input scaling IS in its variation range with it
Step-length is changed, and input displacement IF is changed in its variation range with its step-length；And by the k acquired in step 1.1 parts
Data set is sequentially inputted to these echo state networks using k folding cross validation modes and is trained and tests, when wherein a certain
Each training error trainerror and corresponding test obtained by the valued combinations of input scaling IS and input displacement IF miss
When poor testerror is satisfied by optimal conditions, then input scaling IS and input displacement IF is to be sought needed for echo state network
The optimal input scaling IS and input displacement IF looked for；
Step 1.5 is determined according to optimal reserve pool neuron number N and spectral radius SR and step 1.4 determined by step 1.3
Optimal input scaling IS and input displacement IF build echo state network module；
Step 2, in real time acquisition lithium battery real-time current I, voltage V in the process of running and battery pack external skin temperatures T, and will be real
When electric current I, voltage V and battery pack external skin temperatures T be input in echo state network model, the echo state network model it is defeated
Go out the state-of-charge SOC as the lithium battery predicted in real time.
2. the lithium ion battery SOC on-line prediction methods according to claim 1 based on echo state network, feature
It is, in step 1.2, when setting the initial parameter of echo state network, it is also necessary to echo state network structural parameters are initialized,
Including input weight matrix, reserve pool internal state weight matrix and feedback weight matrix.
3. the lithium ion battery SOC on-line prediction methods according to claim 2 based on echo state network, feature
It is that in step 1.2, input weight matrix W dimensions are N × K, reserve pool internal weights matrix WinDimension is N × N, feedback weight
Matrix WbackDimension is N × L.
4. the lithium ion battery SOC on-line prediction methods according to claim 1 based on echo state network, feature
It is that in step 1.2, it is [0,1] that spectral radius SR, input, which scale the variation range of IS and input displacement IF,.
5. the lithium ion battery SOC on-line prediction methods according to claim 1 based on echo state network, feature
It is that the optimal conditions in step 1.3 and 1.4 is：
[(trainerror+testerror)/2]<minerror
Wherein, trainerror is training error, and testerror is test error, and minerror is initial setting error threshold.
6. the lithium ion battery SOC on-line prediction methods according to claim 1 based on echo state network, feature
It is that step 1.5 still further comprises：Using the recurrent least square method with forgetting factor to structure echo state network module into
The process of row network training update output weight matrix.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810565231.5A CN108804800A (en) | 2018-06-04 | 2018-06-04 | Lithium ion battery SOC on-line prediction methods based on echo state network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810565231.5A CN108804800A (en) | 2018-06-04 | 2018-06-04 | Lithium ion battery SOC on-line prediction methods based on echo state network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108804800A true CN108804800A (en) | 2018-11-13 |
Family
ID=64088498
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810565231.5A Pending CN108804800A (en) | 2018-06-04 | 2018-06-04 | Lithium ion battery SOC on-line prediction methods based on echo state network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108804800A (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110322072A (en) * | 2019-07-09 | 2019-10-11 | 程新宇 | A kind of economic forecasting method neural network based |
| CN110763830A (en) * | 2019-12-04 | 2020-02-07 | 济南大学 | Method for predicting content of free calcium oxide in cement clinker |
| 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 |
| CN111831955A (en) * | 2020-06-05 | 2020-10-27 | 南京航空航天大学 | Lithium ion battery residual life prediction method and system |
| CN112362276A (en) * | 2020-10-27 | 2021-02-12 | 南京林业大学 | Substructure mixing test method |
| CN115236535A (en) * | 2022-07-18 | 2022-10-25 | 湖北文理学院 | Battery SOC estimation method, device, equipment and storage medium |
| US20220373601A1 (en) * | 2019-11-07 | 2022-11-24 | Basf Se | Battery Performance Prediction |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102749199A (en) * | 2012-07-17 | 2012-10-24 | 哈尔滨工业大学 | Method for predicting residual service lives of turbine engines on basis of ESN (echo state network) |
| CN107766986A (en) * | 2017-10-31 | 2018-03-06 | 天津大学 | Leak integral form echo state network on-line study photovoltaic power Forecasting Methodology |
-
2018
- 2018-06-04 CN CN201810565231.5A patent/CN108804800A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102749199A (en) * | 2012-07-17 | 2012-10-24 | 哈尔滨工业大学 | Method for predicting residual service lives of turbine engines on basis of ESN (echo state network) |
| CN107766986A (en) * | 2017-10-31 | 2018-03-06 | 天津大学 | Leak integral form echo state network on-line study photovoltaic power Forecasting Methodology |
Non-Patent Citations (4)
| Title |
|---|
| 严其艳 等: "基于BP 神经网络的电池SOC 预测研究", 《数码世界》 * |
| 刘可 等: "一种基于激光诱导击穿光谱的塑料分类方法", 《光谱学与光谱分析》 * |
| 王永骥 等: "《神经元网络控制》", 28 February 1998, 机械工业出版社 * |
| 王红: "卫星锂离子电池剩余寿命预测方法及应用研究", 《中国优秀博硕士学位论文全文数据库(硕士)工程科技Ⅱ辑》 * |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110322072A (en) * | 2019-07-09 | 2019-10-11 | 程新宇 | A kind of economic forecasting method neural network based |
| US20220373601A1 (en) * | 2019-11-07 | 2022-11-24 | Basf Se | Battery Performance Prediction |
| 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 |
| CN110763830A (en) * | 2019-12-04 | 2020-02-07 | 济南大学 | Method for predicting content of free calcium oxide in cement clinker |
| CN111831955A (en) * | 2020-06-05 | 2020-10-27 | 南京航空航天大学 | Lithium ion battery residual life prediction method and system |
| CN112362276A (en) * | 2020-10-27 | 2021-02-12 | 南京林业大学 | Substructure mixing test method |
| CN112362276B (en) * | 2020-10-27 | 2022-04-15 | 南京林业大学 | Substructure mixing test method |
| CN115236535A (en) * | 2022-07-18 | 2022-10-25 | 湖北文理学院 | Battery SOC estimation method, device, equipment and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108804800A (en) | Lithium ion battery SOC on-line prediction methods based on echo state network | |
| CN108872866B (en) | Dynamic evaluation and long-acting prediction fusion method for charge state of lithium ion battery | |
| CN112269134B (en) | Battery SOC and SOH joint estimation method based on deep learning | |
| CN104849671B (en) | A kind of battery capacity detecting system based on combination neural net | |
| CN109324291A (en) | A kind of prediction technique for Proton Exchange Membrane Fuel Cells life prediction | |
| CN112986830A (en) | Lithium battery capacity estimation method based on convolution time memory neural network | |
| CN110363334B (en) | Grid line loss prediction method of photovoltaic grid connection based on gray neural network model | |
| CN110045292A (en) | Lithium ion battery SOC prediction technique based on big data and bp neural network | |
| CN114280490B (en) | Lithium ion battery state of charge estimation method and system | |
| CN112861436A (en) | Real-time prediction method for engine emission | |
| CN113125960A (en) | Vehicle-mounted lithium ion battery charge state prediction method based on random forest model | |
| CN110866366A (en) | XGboost algorithm-based island detection method for photovoltaic microgrid containing PHEV | |
| CN113093014B (en) | Online collaborative estimation method and system for SOH and SOC based on impedance parameters | |
| AU2021101964A4 (en) | Artificial intelligence based smart electric vehicle battery management system | |
| CN102567782A (en) | Neural-network-based automobile engine torque estimation method | |
| CN112686380A (en) | Neural network-based echelon power cell consistency evaluation method and system | |
| CN107247888A (en) | Sewage disposal water outlet total phosphorus TP flexible measurement methods based on reserve pool network | |
| CN118011211B (en) | Lithium battery state of charge and energy state estimation method | |
| CN116029183A (en) | Power battery temperature prediction method based on iPSO-LSTM model | |
| CN115730525A (en) | Rail transit UPS storage battery health state prediction method | |
| CN114839542A (en) | Sample data set generation method and SOC estimation method for power lithium battery | |
| CN113376541B (en) | Lithium ion battery health state prediction method based on CRJ network | |
| CN112327169B (en) | Lithium battery residual life prediction method | |
| CN117169743A (en) | Battery health state estimation method and device based on partial data and model fusion | |
| CN115031794B (en) | Novel gas-solid two-phase flow measuring method based on multi-feature graph convolution |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20181113 |