CN101504443B

CN101504443B - Prediction method for discharge capacity of lithium ion battery

Info

Publication number: CN101504443B
Application number: CN2008100653411A
Authority: CN
Inventors: 王勇; 吴光麟; 沈晞
Original assignee: BYD Co Ltd
Current assignee: Henan Senyuan Heavy Industry Co Ltd
Priority date: 2008-02-05
Filing date: 2008-02-05
Publication date: 2012-03-07
Anticipated expiration: 2028-02-05
Also published as: CN101504443A

Abstract

The invention provides a method for predicting discharge capacity of a lithium ion battery through partial discharge process by a BP neural network. The terminal voltage of the lithium ion battery in the at least previous 10 minute constant-current discharge process is taken as input, and a BP neural network model outputs the discharge capacity of the battery. The method solves the technical problems of long test period and high energy consumption in the conventional industrial method, also overcomes the defect that a laboratory method has complicated steps and is not suitable for massive industrial production, and guarantees that the average forecasting error is about 2.0 percent which is less than the error range of about 5percent allowable in the industrial production.

Description

A kind of Forecasting Methodology of discharge capacity of lithium ion battery

Technical field

The present invention relates to a kind of Forecasting Methodology of capacity of lithium ion battery, especially a kind of method of utilizing feedforward neural network prediction discharge capacity of lithium ion battery.

Background technology

Lithium ion battery has obtained application more and more widely with its excellent performance in fields such as communication, electronics, automobiles.Yet; The lithium ion battery of domestic production at present is because the restriction of aspects such as technology, battery material causes the actual capacity and the rated capacity and inequality of battery; And, even, the inconsistent defective of electrochemical properties also can occur with the battery of series-produced same model.Therefore, in order to guarantee the quality of battery, must carry out the sorting of discharge capacity, internal resistance before dispatching from the factory to each piece battery.Traditional capacity check method is that battery is full of electricity according to the IEC standard with constant current constant voltage at present, and constant-current discharge is to final voltage then, and the product of steady current and discharge time is just represented the discharge capacity of this battery.

Yet the constant-current discharge time of above-mentioned classic method is long, and partial volume process energy consumption is high.In addition,, cause that the lead time of accomplishing detection is bigger, reduced usage ratio of equipment even battery of the same type actual discharge capacity also has difference.

Feedforward neural network is a kind of of artificial neural network, also is the BP neural network, is widely used in the battery capacity prediction field." new method of capacity of lithium ion battery fast prediction " is (in intelligence Dragon King mighty force; Power technology, 2007.9, vol.31; No.9; P744～746) on the basis of having analyzed lithium ion battery open-circuit voltage and internal resistance and battery capacity relation, through partial discharge, utilization neural network prediction discharge capacity of lithium ion battery.After lithium ion battery constant current constant voltage under the different operating positions is full of electricity; With the input of the open-circuit voltage of battery under the discharge condition of 10%, 20%, 30% rated capacity, internal resistance as artificial neural network; Final discharge capacity is as output; The training sample that constitutes neural network is right, thereby utilizes the discharge capacity of neural network prediction battery, and error is about 3%.The model of neural network is as shown in Figure 1.This method so the step more complicated is fit to laboratory study, is not suitable for the application of large-scale industrial production owing to will test the open-circuit voltage and the internal resistance of the battery under the different discharge conditions.

Summary of the invention

In order to solve the problems referred to above of prior art; The present invention provides a kind of BP of utilization neural network model; With lithium ion battery at least before the wire-end voltage of 10min constant-current discharge process as the input of BP neural network, discharge capacity of the cell is the method for prediction of output lithium ion discharge capacity.This method comprises the steps:

(1) in mesuring battary, randomly draws at least 20 batteries, test the wire-end voltage of the preceding at least 10min constant-current discharge process of these batteries

(2) wire-end voltage that step (1) is obtained is as the input of Training of Feedforward Neural Networks sample, training feedforward neural network, output discharge capacity of the cell

(3) with the battery of remainder at least before the wire-end voltage of 10min constant-current discharge process as input, export discharge capacity of the cell with the feedforward neural network that step (2) trains

The model of BP neural network of the present invention is as shown in Figure 2.

The lithium ion battery that uses among the present invention is SL412454 flexible packing lithium ion battery and the LP053450ARUL lithium cobalt battery that BYD company produces.

In the above-mentioned steps, battery at least before the wire-end voltage of 10min constant-current discharge process be meant the constant-current discharge process at least before pairing a series of wire-end voltages when battery is emitted different electric weight in the 10min.

Lithium ion battery in the above-mentioned steps is the lithium ion battery through discharging and recharging that do not have of same model.On the one hand, choose such battery, can get rid of the influence of the internal resistance of cell BP network output result as test sample book; On the other hand, because wire-end voltage in the battery discharge procedure or open-circuit voltage and discharge capacity of the cell have correlativity preferably, and in the commercial production, the test of wire-end voltage is convenient than open-circuit voltage, is the input end of BP network so select wire-end voltage.In the prior art, during with BP neural network prediction battery capacity, input end can have a variety of selections; And the inventor finds through a large amount of experiments; With the preceding at least 10min of at least 25 batteries randomly drawing, the wire-end voltage of preferred preceding at least 15min constant-current discharge process could guarantee to significantly reduce in operation steps as the input of BP network training sample; Under the significantly reduced situation of energy consumption, make predicated error satisfy industrial requirement.

The constant-current discharge test duration is long more, and the wire-end voltage value of the discharge process of electrical property detecting box record is just many more, and BP network input end data coverage is big more, and it is just accurate more to predict the outcome, and this is a general knowledge known in this field.So mesuring battary quantity is more, and the sample number of batteries that extracts relatively more after a little while, can guarantee precision of prediction through the proper extension constant-current discharge test duration.

In the above-mentioned steps, use the electrical property detecting box that battery is discharged and recharged control, and carry out data recording and collection automatically.The blue strange BK3512L detecting box that uses among the present invention is gathered at least 30 wire-end voltage data of the preceding at least 10min of constant-current discharge automatically.

With BP neural network match wire-end voltage U in the MATLAB software and the funtcional relationship of discharge capacity C, thus output discharge capacity of the cell C.

The BP neural network model can have multiple choices, and its optimization model is:

Hidden layer neuron transport function tansig;

Output layer neuron transport function logsig;

Training function trainlm;

Hidden layer neuron several 15;

Training step number net.trainParam.epochs=10～5500;

Training objective net.trainParam.goal=0.005～0.02

More preferably:

Hidden layer neuron transport function tansig;

Output layer neuron transport function logsig;

Training function trainlm;

Hidden layer neuron several 15;

Training step number net.trainParam.epochs=5000;

Training objective net.trainParam.goal=0.01;

Netinit net=init (net)

Description of drawings

Fig. 1: BP neural network model of the prior art

Fig. 2: the BP neural network model among the present invention

Fig. 3-1: according to the prediction discharge capacity of the method for embodiment 1

Fig. 3-2: according to the predicated error of the method for embodiment 1

Fig. 4-1: according to the prediction discharge capacity of the method for embodiment 2

Fig. 4-2: according to the predicated error of the method for embodiment 2

Embodiment

Use blue strange BK3512L detecting box that the lithium ion battery of a certain model is carried out charge-discharge test; And write down automatically the constant-current discharge process at least before the wire-end voltage of 10min; As input, training BP neural network is exported discharge capacity of the cell with these a series of wire-end voltage numerical value.

The following example is with further explain the present invention.

Embodiment 1

Randomly draw the SL412454 flexible packing lithium ion battery that 68 BYD produce, its rated capacity is 520mAh (the 1C electric current is 520mA), according to classic method battery is carried out the first charge-discharge experiment with blue strange BK3512L detecting box earlier; Be T=30 ℃ of first constant current 0.5C down, final voltage 4.2V, constant voltage 4.2V again; Stop electric current 0.02C charging; Constant current 0.2C final voltage 2.75V discharges then, tests and calculate the discharge capacity of these batteries, is designated as the actual measurement capacity.After then all 68 batteries being full of electricity according to the method described above; Randomly draw wherein 26 batteries, T=30 ℃ of following 0.5C discharge is with a series of wire-end voltage data of 15min discharge process before the blue strange BK3512L detecting box record; And with it as input; BP neural network in the training MATLAB software, prediction of output capacity, the error between predicted value and measured value is～5%; Think that the BP network model of this moment reaches best, the discharge capacity that best BP network model is exported is designated as the prediction capacity.Then a series of wire-end voltages of the preceding 15min of all the other 42 battery 0.5C discharge processes are done the input of sample to be tested, with the discharge capacity of said best these 42 batteries of BP neural network model output.Best BP neural network model of the present invention is following:

● netinit net=init (net)

● hidden layer neuron number: 15

● hidden layer neuron transport function: tansig

Output layer neuron transport function: logsig

Training function: trainlm

● training step number: 5000 steps (net.trainParam.epochs=5000);

● training objective: 0.01 (net.trainParam.goal=0.01);

Table 1 is the part experimental data of embodiment 1:

Table 1

Numbering	Actual measurement capacity (mAh)	Prediction capacity (mAh)	Error volume (mAh)	Percentage error (%)
					1	540.94	519.84	21.10	3.90
2	521.28	518.12	3.16	0.61
					3	523.37	519.21	4.16	0.79
4	541.18	519.75	21.43	3.96
					5	523.72	507.78	15.94	3.04
6	528.73	519.47	9.26	1.75
					7	531.36	513.12	18.24	3.43

8	534.26	519.59	14.67	2.75
					9	524.33	518.50	5.83	1.11
10	528.41	518.75	9.66	1.83
					11	526.56	518.90	7.66	1.45
12	517.76	517.74	0.02	0.0046
					13	519.97	517.78	2.19	0.42
14	515.61	516.16	0.55	0.11
					15	517.46	517.91	0.45	0.09
16	520.30	517.28	3.02	0.58
					17	526.21	517.75	8.46	1.61
18	522.17	518.01	4.16	0.80
					19	524.33	518.50	5.83	1.11
20	528.41	518.75	9.67	1.83
					21	523.99	517.83	6.16	1.18
22	512.96	514.79	1.83	0.36
					23	511.96	514.35	2.39	0.47
24	517.76	517.74	0.02	0.0046
					25	519.97	517.78	2.19	0.42
26	515.61	516.16	0.55	0.11
					27	513.33	513.92	0.59	0.11
28	522.29	504.25	18.04	3.45
					29	525.80	514.34	11.46	2.18
30	529.57	516.05	13.52	2.55
					31	420.29	432.97	12.68	3.02
32	523.57	519.11	4.46	0.85
					33	528.15	517.99	10.16	1.92

34	524.20	518.75	5.45	1.04
					35	530.24	519.44	10.80	2.04
36	520.08	511.53	8.55	1.64
					37	527.34	518.78	8.56	1.62
38	526.97	518.91	8.06	1.53
					39	522.58	519.14	3.44	0.66
40	520.76	519.07	1.69	0.32
					41	531.14	518.19	12.95	2.44
42	525.58	518.77	6.811	1.30

Error in this table is a relative error, and the computing method of error volume and percentage error are:

Error volume=| actual measurement capacity-prediction capacity | ... (1)

Error ratio=(| actual measurement capacity-prediction capacity |/actual measurement capacity) ... (2)

Percentage error=(| actual measurement capacity-prediction capacity |/actual measurement capacity) * % ... (3)

In conjunction with this table 1 and Fig. 3-2, find that the consensus forecast percentage error is 1.44%, largest prediction error is 3.90%, and when satisfying industrial needs, the classic method test duration shortens about 105min relatively, and energy consumption is reduction greatly also.

Embodiment 2

Randomly draw the LP053450ARUL lithium cobalt battery that 197 BYD produce, its rated capacity is 700mAh (the 1C electric current is 700mA).Discharge capacity with these batteries of classic method test is designated as the actual measurement capacity.Method of testing is first 0.3C constant current, final voltage 4.2V, and the 4.2V constant voltage stops electric current 0.02C charging again, 0.5C constant current then, final voltage 2.75V discharge.Randomly draw wherein 58 batteries then; T=30 ℃ of following 1C current discharge, with a series of wire-end voltages of the preceding 10min of blue strange BK3512L detecting box record discharge process as input, the BP neural network in the training MATLAB software; Prediction of output capacity; Error between predicted value and measured value thinks that～5% the BP network model of this moment reaches best, and the discharge capacity that best BP network model is exported is designated as the prediction capacity.Then a series of wire-end voltages of the preceding 10min of all the other 139 battery 1C discharge processes are made the input end of sample to be tested, with the discharge capacity of said best these 139 batteries of BP neural network model output.Said best BP neural network model is identical with embodiment 1.

Table 2 is the part experimental data of embodiment 2:

Table 2

Numbering	Actual measurement capacity (mAh)	Prediction capacity (mAh)	Error volume (mAh)	Percentage error (%)
					1	684	708.64	24.64	3.60
2	707	702.98	4.02	0.57
					3	677	702.2	25.2	3.72
4	696	695.33	0.67	0.10
					5	679	696.02	17.02	2.51
6	681	699.51	18.51	2.72
					7	679	696.55	17.55	2.58
8	684	697.36	13.36	1.95
					9	677	702.94	25.94	3.83
10	680	698.87	18.87	2.78
					11	681	701.56	20.56	3.02
12	680	704.33	24.33	3.58
					13	681	695.26	14.26	2.09
14	679	696.33	17.33	2.55
					15	676	700.46	24.46	3.62
16	683	696.47	13.47	1.97
					17	676	697.02	21.02	3.11
18	680	699.18	19.18	2.82
					19	695	700.61	5.61	0.81
20	677	700.42	23.42	3.46
					21	676	703.22	27.22	4.03

22	682	709.38	27.38	4.01
					23	687	704.84	17.84	2.60
24	678	697.27	19.27	2.84
					25	680	697.77	17.77	2.61
26	681	705.49	24.49	3.60
					27	681	703.41	22.41	3.29
28	705	707.67	2.67	0.38
					29	679	699.97	20.97	3.09
30	682	709.54	27.54	4.04
					31	677	707.31	30.31	4.48
...	...	?...	...	?...
					130	709	706.45	2.55	0.36
131	705	693.84	11.16	1.58
					132	694	692.11	1.89	0.27
133	704	694.46	9.54	1.36
					134	705	701.75	3.25	0.46
135	710	708.72	1.28	0.18
					136	704	701.2	2.8	0.40
137	705	706.54	1.54	0.22
					138	706	705.64	0.36	0.05
139	716	712.03	3.97	0.55

Identical in the computing method of error volume and percentage error and the table 1 in this table.

In conjunction with this table 2 and Fig. 4-2, find that the consensus forecast error is 2.03%, largest prediction error is 4.48%, and when satisfying industrial needs, the classic method test duration shortens about 50min relatively, and energy consumption is reduction greatly also.

Claims

1. the method with feedforward neural network prediction discharge capacity of lithium ion battery comprises the steps:

(1) in mesuring battary, randomly draws at least 20 batteries, test the wire-end voltage of the preceding at least 10min constant-current discharge process of these batteries; Battery at least before the wire-end voltage of 10min constant-current discharge process be meant the constant-current discharge process at least before pairing a series of wire-end voltages when battery is emitted different electric weight in the 10min; And the lithium ion battery of being selected for use is the lithium ion battery through discharging and recharging that do not have of same model;

(2) wire-end voltage that step (1) is obtained is as the input of Training of Feedforward Neural Networks sample, training feedforward neural network, output discharge capacity of the cell;

(3) with the remaining part battery at least before the wire-end voltage of 10min constant-current discharge process as input, the feedforward neural network output discharge capacity of the cell that trains with step (2), and with output valve as the discharge capacity of lithium ion battery of predicting.

2. according to the method for claim 1, with battery at least before the wire-end voltage of 15min constant-current discharge process as the input of feedforward neural network.

3. according to the method for claim 1, said constant-current discharge electric current is 0.2～1C.

4. according to the method for claim 1, said feedforward neural network model is:

The hidden layer neuron number is 15;

The hidden layer neuron transport function is tansig;

Output layer neuron transport function is logsig;

The training function is trainlm;

Training step number 10～5500;

Training objective 0.005～0.02.

5. according to the method for claim 1, said feedforward neural network model is:

The hidden layer neuron number is 15;

The hidden layer neuron transport function is tansig;

Output layer neuron transport function is logsig;

The training function is trainlm;

Training step number 5000;

Training objective 0.01;

Netinit net=init (net).