CN103176139B - The charge state evaluation method that electrokinetic cell Non-smooth surface lagging characteristics compensates and system - Google Patents

Abstract

The present invention is charge state evaluation method and the system of the compensation of electrokinetic cell Non-smooth surface lagging characteristics, this law first step gathers cell output voltage and electric current, the relational expression being obtained each parameter by battery equivalent-circuit model builds neural network OCV (k) prediction model, solve wherein parameter, open-circuit terminal voltage OCV (k) On-line Estimation.Second step SDH model and RBF2 are composed in series Dynamic Hysteresis mixture model.SDH model is input with first step gained OCV (k), the input that its y (k) exported and OCV (k), OCV (k-1) are RBF2, the parameter of RBF2 weighting study Indirect method SDH model, approach the sluggish relation of actual complexity, the final SOC (k) exporting estimation on line.Native system is made up of microprocessor and the electric current, voltage sensor etc. being installed on battery circuit, stores the program performing this method, obtains SOC(k) estimated value.The present invention uses for reference neural network, compensate for electrokinetic cell complicated Non-smooth surface Hysteresis Nonlinear characteristic, improves SOC (k) estimation on line precision.

Description

The charge state evaluation method that electrokinetic cell Non-smooth surface lagging characteristics compensates and system

Technical field

The present invention relates to the state of charge estimating techniques field of automobile power cell, be specially charge state evaluation method and the system of the compensation of a kind of electrokinetic cell Non-smooth surface lagging characteristics.

Background technology

In New-energy electric vehicle, electrokinetic cell is one of its three large gordian technique.Accumulator develops into lithium ion battery from lead-acid battery, Ni-MH battery, and the theoretical and technological studies that efficient, the safety of battery and long-life use receives much attention always.Automobile power cell is completely different from the battery charge and discharge process for other equipment such as mobile phones, electric automobile power battery is under stochastic and dynamic replaces charge and discharge state, overcharge and deep discharge all may cause the irreversible damage of battery, also relate to safety problem.State of charge SOC(State of Charge in battery operation) be that dynamic charge and discharge process is efficient, one of key parameter in safety management.So, accurate estimation on line or measure effective guarantee that state of charge is power battery pack safe operation, discharge and recharge optimum management and control.

State of charge is the isoparametric function of electrokinetic cell electric current, voltage, temperature and internal resistance, and state of charge can not directly obtain, and need be obtained by the measurement of various indirect method.Existing state of charge assay method and defect are:

1) AH(Ampere Hour) metering method: the identification that still there is model parameter and state, the problems such as original state.

2) SOC estimation method that is combined with model of mind of AH method, is considered as general nonlinearity characteristic by the Hysteresis Nonlinear characteristic of SOC, has carried out approximate processing.

3) impedance method: operand is large, canbe used on line difficulty.

4) based on the SOC estimation method of battery equivalent-circuit model, its model can not describe the complicated lagging characteristics that battery shows at discharge and recharge dynamic changing process fully.

5) based on the SOC estimation method of intelligent modeling, the complicated lagging characteristics comprehensively considering that SOC occurs in charging and discharging two processes all failed by the model adopted.

Capacity is large, energy density is high, service life cycle is long, have the first-selection that the lithium ion battery of wide application prospect and NiMH battery become new-energy automobile power battery.The battery terminal voltage of lithium ion battery charge and discharge process and state of charge SOC show the complicated Hysteresis Nonlinear characteristic of Non-smooth surface, and as shown in Figure 1, wherein solid line is charging process, and dotted line is discharge process.Within 2012, current paper shows, analyzes its complicated sluggish certainty existed from inside battery ion motion.So the compensation of complicated lagging characteristics, becomes the factor that the estimation of SOC high precision be can not ignore.Such as: for NiMH battery, if do not compensated the impact of Hysteresis Nonlinear, SOC can be caused to measure the error of about 40%.

Multiple secondary ring phenomenons that electrokinetic cell has occurred except main ring at the dynamic process that random discharge and recharge replaces, be illustrated in figure 2 the open terminal voltage OCV(Open Circuit Voltage of ni-mh (NiMH) battery under charge and discharge process) with state of charge SOC main ring and secondary ring lagging characteristics curve, wherein ● line represents discharge process, zero line represents charging process.

Document provides the Non-smooth surface characteristic at forward and inverse journey lagging characteristics and switching point place, and there is inconsistency in battery terminal voltage and SOC, namely with repeatedly charging, there is drifting problem, the battery terminal voltage that its 1500th time and the 3000th time cycle charge-discharge shows and SOC lagging characteristics, there is obvious difference, be illustrated in figure 3 lithium ion battery terminal voltage and the sluggish graph of a relation of SOC in different serviceable life; Wherein solid line is terminal voltage and the sluggish relation of SOC of new battery, dot-and-dash line is through the battery terminal voltage of 1500 charge and discharge cycles and the sluggish relation of SOC, dotted line is the sluggish relation of battery terminal voltage after 3000 charge and discharge cycles and SOC, as can be seen from Figure 3, if ignore the compensation of lagging characteristics, directly cause error about 3%.

Battery open circuit terminal voltage OCV and SOC shows stable lagging characteristics to be had and well repeats comformity relation, as shown in Figure 4 ◆ line represent new battery, zero line represent the 1500th cycle charge-discharge, ▲ line represents the lagging characteristics curve of OCV and the SOC of 3000 cycle charge-discharges, visible new battery and show stable consistance at the lagging characteristics of OCV and the SOC of the 1500th time, the 3000th time cycle charge-discharge.

Existing battery equivalent electrical circuit linear model, describes its lagging characteristics in a simple manner decoupled roughly, fails fully to describe the Dynamic Hysteresis nonlinear characteristic of electrokinetic cell Non-smooth surface, many rings.The complicacy of electric automobile power battery electrochemical process, result in consistency problem and the drifting problem of SOC estimation; The complicated Hysteresis Nonlinear characteristic of SOC and battery open circuit terminal voltage OCV and compensation problem are there is; With SOC-OCV sluggish relation estimation SOC, to the online estimation problem fast of OCV in requisition for solution.These three key problems are the key issues needing in the estimation of SOC high precision to break through, and are also the bottleneck problems of SOC high precision estimation.

Summary of the invention

The object of the invention is the charge state evaluation method designing the compensation of a kind of electrokinetic cell Non-smooth surface lagging characteristics, this method is estimated the state of charge SOC of electrokinetic cell in two steps, and the first step is the On-line Estimation to electrokinetic cell open terminal voltage OCV; Second step be according to OCV to state of charge Hysteresis Nonlinear compensation of error, estimate current SOC value.

Another object of the present invention is the state of charge estimating system that a kind of electrokinetic cell Non-smooth surface lagging characteristics of design compensates, and by above-mentioned evaluation method, design a calculating machine signal processing system, realizes the electrokinetic cell state of charge of display estimation on line in real time.

The charge state evaluation method that design motivation battery Non-smooth surface lagging characteristics of the present invention compensates, estimate discrete digital amount SOC (k) of the state of charge of electrokinetic cell in two steps, the first step is the discrete digital amount voltage that digital collection obtains electrokinetic cell with current i (k), On-line Estimation is carried out to discrete digital amount OCV (k) of battery open circuit terminal voltage; Second step, for compensate state of charge Hysteresis Nonlinear error according to discrete digital amount OCV (k), estimates current discrete digital amount SOC (k) value.

The first step, discrete digital amount OCV (k) On-line Estimation to battery open circuit terminal voltage OCV (t)

Have to stablize according to the relation of electrokinetic cell (comprising lithium ion battery and Ni-MH battery) SOC and open terminal voltage OCV and repeat conforming complicated lagging characteristics, the first step first carries out On-line Estimation to discrete digital amount OCV (k) of battery open circuit terminal voltage, avoids to obtain battery open circuit terminal voltage and long Holding Problem;

In order to carry out On-line Estimation to discrete digital amount OCV (k), the present invention adopts representative battery equivalent electrical circuit Thevenin model.In this equivalent electrical circuit, resistance R ₂resistance-capacitance circuit is formed, open terminal voltage OCV (t) the resistance in series R of battery with electric capacity C parallel connection ₁with above-mentioned resistance-capacitance circuit, battery equivalent electrical circuit output voltage is V (t), by resistance R ₁electric current be i (t), the terminal voltage of electric capacity C is u _c(t).

Its mathematic(al) representation is as follows:

$\left\{\begin{array}{c}\mathrm{OCV}\left(t\right)=V\left(t\right)+R_{1}i\left(t\right)+u_c\left(t\right)\\ i\left(t\right)=\frac{u_c\left(t\right)}{R_2}+C\frac{{\mathrm{du}}_c\left(t\right)}{\mathrm{dt}}\end{array}\right.---\left(1\right)$

(1) formula is equivalent to:

$\left\{\begin{array}{c}V\left(t\right)=\mathrm{OCV}\left(t\right)+R_{1}i\left(t\right)+u_c\left(t\right)\\ i\left(t\right)=\frac{u_c\left(t\right)}{R_2}+C\frac{{\mathrm{du}}_c\left(t\right)}{\mathrm{dt}}\end{array}\right.---\left(2\right)$

Wherein: OCV (t) is the function of SOC (t).

The charge and discharge process of battery is process more slowly, within the short time (in several seconds), OCV (t) is relatively stable, OCV (t) can be used as constant value process, under metastable state, and OCV (t), resistance R during above-mentioned equation (2) is stable ₁, R ₂and electric capacity C exists with determined value, there is stable unique solution in corresponding quaternary dynamic equation (2).

Output voltage V (k) corresponding according to the above-mentioned math equation (2) obtained by battery equivalent-circuit model and battery beginning voltage OCV (k), current i (k) and capacitance terminal voltage u _cthe discrete magnitude relation of (k):

V(k)=OCV(k)-R ₁i(k)-u _c(k) （3）

The capacitance terminal voltage u corresponding according to the above-mentioned math equation (2) obtained by battery equivalent-circuit model _cthe discrete magnitude relation of (k) and i (k):

u _c(k)=k ₂u _c(k-1)+k ₁i(k) （4）

Wherein:

$k_1=\frac{R_{2}T}{T+R_{2}C},$ $K_2=\frac{R_{2}C}{T+R_{2}C}$

T is the sampling period.

According to (3), (4) formula, corresponding structure neural network OCV (k) prediction model, use for reference neural network learning, in complete battery pair equivalent-circuit model, parameter solves.

Neural network OCV(k) discrete digital amount OCV(k in prediction model), V (k), i (k) and u _canalog quantity OCV (t), V (t), i (t) and u in (k) difference corresponding equation (2) _c(t).

Corresponding battery equivalent-circuit model builds neural network OCV(k of the present invention) prediction model, comprise three neuron nodes and the first radial basis function neural network RBF(Radical Basis Function), be expressed as RBF1, three neuron nodes calculate by (3), (4) formula respectively, export and are respectively capacitance terminal voltage u _c(k), resistance R ₁terminal voltage and equivalent electrical circuit output voltage V (k).First radial basis function neural network RBF1 completes the On-line Estimation to battery open circuit terminal voltage OCV (k).Digital collection obtains the discrete digital amount voltage of the actual output of electrokinetic cell with current i (k), as the input value of this model.

Peripheral sensory neuron node asks capacitance terminal voltage u _c(k).Z ^-1for to back operator, the u that peripheral sensory neuron node exports _ck () passes through z ^-1obtain corresponding u _c(k-1).Peripheral sensory neuron node is according to formula u _c(k)=k ₂u _c(k-1)+k ₁i (k) is by weighting coefficient k ₁and k ₂respectively to the discrete digital amount i(k gathering gained) and u _c(k-1) be weighted summation, obtain exporting u _c(k).

Nervus opticus unit node obtains i (k) and Model Parameter k according to digital collection ₃obtain R ₁on voltage be k ₃× i (k), k ₃represent R ₁.

Third nerve unit node calculate equivalent electrical circuit output voltage V (k) estimated value, V (k)=OCV (k)-R ₁i (k)-u _c(k).

The output OCV(k of the first radial basis function neural network RBF1) be the kinematic function of current i (k), equivalent electrical circuit output voltage V (k).OCV(k) z is passed through ^-1oCV (k-1) is obtained to back operator, OCV (k-1) is as external feedback, with the last sampling instant value V (k-1) of V (k) and i (k) input signal as the first radial basis function neural network RBF1, first radial basis function neural network RBF1 is to voltage V (k), current i (k) and OCV(k) kinematic nonlinearity characteristic be described, supplement battery equivalent electrical circuit linear model and expand, the output of the first radial basis function neural network RBF1 is battery open circuit terminal voltage OCV (k).

Neural network OCV(k) the weights k of prediction model ₁, k ₂and k ₃study, estimate the cell output voltage V(k obtained with this model) with the discrete digital amount of the battery terminal voltage of actual measurement difference, form neural network OCV(k) the learning objective function of prediction model, adopt ripe method of steepest descent, i.e. gradient method, by neural network OCV(k) prediction model, parameter R in the electrokinetic cell of identification simultaneously equivalent-circuit model ₁, R ₂and C, and state OCV(k), realize the On-line Estimation of OCV (k).Its neural network OCV(k) the Parameter Self-learning adjustment capability of prediction model, the electrokinetic cell otherness characteristically of different capabilities can be adapted to.

Second step, according to OCV (k) to the complicated Hysteresis Nonlinear error compensation of state of charge, estimate current SOC(k) value

The Dynamic Hysteresis mixture model that this step adopts simple dynamic hysteresis model-SDH model (Simple Dynamic Hysteresis) and the second radial basis function neural network RBF2 to be composed in series.

The expression formula of SDH model is as follows:

$\left\{\begin{array}{c}y\left(t\right)=\mathrm{\η}\left(t\right)-\frac{1}{k_4}\mathrm{OCV}\left(t\right)\\ \stackrel{\·}{\mathrm{\η}}\left(t\right)=k_4(\mathrm{OCV}\left(t\right)-\mathrm{\Δ}\left(\mathrm{\η}\left(t\right)\right))\end{array}\right.---\left(5\right)$

Wherein

$\mathrm{\&Delta;}\left(\mathrm{\&eta;}\left(t\right)\right)=\left\{\begin{array}{cc}{k}_4(\mathrm{\&eta;}\left(t\right)-1)& \mathrm{\&eta;}\left(t\right)>1\\ 0& \left|\mathrm{\&eta;}\left(t\right)\right|\&le;1\\ k_4(\mathrm{\&eta;}\left(t\right)+1)& \mathrm{\&eta;}\left(t\right)<-1\end{array}\right.---\left(6\right)$

Wherein, OCV (t) is the On-line Estimation value that the first step obtains, and be the input of SDH model, y (t) is the output of SDH model.

Parameter k in SDH model ₄determine the retardant curve of SDH model and actual SOC(t)-OCV(t) sluggish relation curve at the similarity degree of contour shape, also determine the on-line study speed of the learning process of the second radial basis function neural network RBF2 be connected on after SDH model.

Corresponding discrete model can be obtained by SDH model.Analog quantity OCV (t) and y (t) respectively corresponding discrete digital amount OCV (k) and y (k), T is the sampling period.

The expression formula of SDH discrete model is as follows:

$\left\{\begin{array}{c}y\left(k\right)=\mathrm{\&eta;}\left(k\right)-\frac{1}{k_4}\mathrm{OCV}\left(k\right)\\ \frac{\mathrm{\&eta;}\left(k\right)-\mathrm{\&eta;}(k-1)}{T}=k_4(\mathrm{OCV}\left(k\right)-\mathrm{\&Delta;}\left(\mathrm{\&eta;}\left(k\right)\right))\end{array}\right.$

After arrangement:

$\left\{\begin{array}{c}y\left(k\right)=\mathrm{\&eta;}\left(k\right)-\frac{1}{k_4}\mathrm{OCV}\left(k\right)\\ \mathrm{\&eta;}\left(k\right)=\mathrm{\&eta;}(k-1)+T\&times;k_4(\mathrm{OCV}\left(k\right)-\mathrm{\&Delta;}\left(\mathrm{\&eta;}\left(k\right)\right))\end{array}\right.---\left(7\right)$

Wherein

$\mathrm{\&Delta;}\left(\mathrm{\&eta;}\left(k\right)\right)=\left\{\begin{array}{cc}{k}_4(\mathrm{\&eta;}\left(k\right)-1)& \mathrm{\&eta;}\left(k\right)>1\\ 0& \left|\mathrm{\&eta;}\left(k\right)\right|\&le;1\\ k_4(\mathrm{\&eta;}\left(k\right)+1)& \mathrm{\&eta;}\left(k\right)<-1\end{array}\right.---\left(8\right)$

SDH model has the Non-smooth surface lagging characteristics describing monocycle, i.e. outer shroud, also has description secondary ring, i.e. many rings lagging characteristics of inner ring.

The present invention is using the preamble part of SDH model as Dynamic Hysteresis mixture model that be simple for structure, that can describe many rings lagging characteristics.Connect the second radial basis function neural network RBF2 after SDH model, builds Dynamic Hysteresis mixture model.The OCV (k) that the y (k) that SDH model obtains and the first step obtain, the last sampling instant value OCV (k-1) of OCV (k) are as the input of the second radial basis function neural network RBF2, by the weighting study in the second radial basis function neural network RBF2, realize the Nonlinear Mapping that any monodrome is corresponding, the k of Indirect method SDH model ₄parameter, to approach actual SOC(k) and OCV(k) complicated sluggish relation, the final SOC (k) exporting estimation on line.

Described Dynamic Hysteresis mixture model not only can represent sluggish outer shroud, i.e. main ring, also can describe sluggish multiple secondary ring, i.e. inner ring.When being used for the electric automobile in dynamic operation when electrokinetic cell, this Dynamic Hysteresis mixture model can describe the outer shroud and multiple interior ring property that the complicated lagging characteristics of Non-smooth surface in charging, discharging electric batteries Stochastic Dynamic Process shows simultaneously.

SDH model has Non-smooth surface characteristic, makes SOC(k)-OCV(k) Dynamic Hysteresis mixture model can describe the Non-smooth surface characteristic of battery.The latter part of second radial basis function neural network RBF2 of sluggish mixture model of connecting just realizes Nonlinear Mapping.The Non-smooth surface characteristic of electrokinetic cell, by the mode of this Dynamic Hysteresis mixture model, avoids the problem directly cannot process Non-smooth surface signal in differential or the local derviation neural network modeling approach that is Fundamentals of Mathematics dexterously.

This model, first according to experimental data, adopts ripe steepest decline technology to complete the study of the second radial basis function neural network RBF2, adopts the output estimation SOC(k of the Dynamic Hysteresis mixture model trained) value.

According to the charge state evaluation method that the electrokinetic cell Non-smooth surface lagging characteristics of the invention described above compensates, devise the state of charge estimating system that electrokinetic cell Non-smooth surface lagging characteristics compensates, comprise microprocessor, current sensor, voltage sensor, analog to digital converter, program storage, programmable storage, timer and display.Current sensor and voltage sensor export through analog to digital converter access microprocessor, microprocessor linker storer, programmable storage, timer and display.

Neural network OCV(k is stored in program storage) prediction model and the SDH model calculation procedure of Dynamic Hysteresis mixture model of connecting with the second radial basis function neural network RBF2, the parameter in OCV (k) prediction model and SDH model is stored in programmable storage, current sensor and voltage sensor are installed in electrokinetic cell and load connecting circuit, the load current of current sensor and the electrokinetic cell measured by voltage sensor and terminal voltage are by analog to digital conversion, and the digital quantity obtaining corresponding load current and voltage sends into microprocessor.SOC(k in timer control program storer) estimation program start and interrupt operation, the current SOC(k of the operation result of microprocessor) estimated value, shown in real time by display.

Compared with prior art, the advantage of the charge state evaluation method that electrokinetic cell Non-smooth surface lagging characteristics of the present invention compensates and system is: 1, according to battery equivalent-circuit model structure, build neural network OCV (k) prediction model, use for reference neural network to solve an equation method, realize electrokinetic cell open terminal voltage OCV(k) On-line Estimation; Utilize Neural Network Self-learning ability, solve electrokinetic cell because of capacity different with batch grade, the variability issues caused, On-line Estimation battery open circuit terminal voltage OCV(k), avoided to obtain open terminal voltage OCV(k) and the problem that waits as long for; 2, SOC(k is utilized) and the stable and consistent of OCV(k) relation, build the Dynamic Hysteresis mixture model of SDH model and the second radial basis function neural network RBF2 series connection, realize compensating to the complicated Non-smooth surface Hysteresis Nonlinear characteristic of electrokinetic cell, improve the estimation on line precision of SOC (k); 3, first OCV(k is obtained), again by Dynamic Hysteresis mixture model estimation SOC(k) two-stage process, solve the complicacy of electrokinetic cell electrochemical reaction and the drifting problem of repeatedly charge and discharge, obtain SOC(k) consistance estimation, effectively improve SOC(k) estimation precision; 4, on this method basis, realize with active computer software and hardware the system that SOC (k) estimates, only need measure the parameter such as electric current, voltage of electrokinetic cell, the state of charge of electrokinetic cell can be shown in real time, be convenient to implement to use.

Accompanying drawing explanation

Fig. 1 is the complicated Hysteresis Nonlinear performance diagram of Non-smooth surface of electrokinetic cell terminal voltage and state of charge SOC (t) under charge and discharge process;

Fig. 2 is open terminal voltage OCV (t) of ni-mh (NiMH) battery under charge and discharge process and state of charge SOC (t) main ring and secondary ring lagging characteristics curve map;

Fig. 3 is lithium ion battery terminal voltage and the SOC(t in different serviceable life) sluggish graph of a relation;

Fig. 4 is electrokinetic cell open terminal voltage OCV (t) and the SOC(t in different serviceable life) sluggish graph of a relation;

Fig. 5 is the model schematic that charge state evaluation method embodiment state of charge SOC (k) that carries out electrokinetic cell in two steps that this electrokinetic cell Non-smooth surface lagging characteristics compensates is estimated;

Fig. 6 is electrokinetic cell equivalent-circuit model schematic diagram used in the charge state evaluation method embodiment of this electrokinetic cell Non-smooth surface lagging characteristics compensation;

Fig. 7 is the neural network model schematic diagram of open terminal voltage OCV (k) On-line Estimation used in the charge state evaluation method embodiment of this electrokinetic cell Non-smooth surface lagging characteristics compensation;

Fig. 8 is that in the charge state evaluation method embodiment that compensates of this electrokinetic cell Non-smooth surface lagging characteristics, SDH model used is connected the Dynamic Hysteresis mixture model schematic diagram of the second radial basis function neural network RBF2;

Fig. 9 is the monocycle lagging characteristics curve map of SDH model in the charge state evaluation method embodiment of this electrokinetic cell Non-smooth surface lagging characteristics compensation;

Figure 10 is many rings lagging characteristics curve map of SDH model in the charge state evaluation method embodiment of this electrokinetic cell Non-smooth surface lagging characteristics compensation;

Figure 11 is the state of charge estimating system example structure schematic diagram that this electrokinetic cell Non-smooth surface lagging characteristics compensates.

Embodiment

The charge state evaluation method embodiment that electrokinetic cell Non-smooth surface lagging characteristics compensates

This example is estimated the state of charge SOC of electrokinetic cell in two steps, and as shown in Figure 5, the first step is digital collection electrokinetic cell, obtains the discrete digital amount of the battery terminal voltage of actual measurement with discrete digital amount current i (k), On-line Estimation is carried out to discrete digital amount OCV (k) of battery open circuit terminal voltage OCV (t); Second step, for compensate state of charge Hysteresis Nonlinear error according to OCV (k), estimates current SOC (k) value.

The first step, to battery open circuit terminal voltage OCV(t) discrete digital amount OCV (k) On-line Estimation

Battery equivalent electrical circuit Thevenin model as shown in Figure 6, in equivalent electrical circuit, resistance R ₂resistance-capacitance circuit is formed, OCV (t) the contact resistance R successively of battery with electric capacity C parallel connection ₁with above-mentioned resistance-capacitance circuit, battery equivalent electrical circuit output voltage is V (t), by resistance R ₁electric current be i (t), the terminal voltage of electric capacity C is u _c(t).

Its mathematic(al) representation is as follows:

$\left\{\begin{array}{c}\mathrm{OCV}\left(t\right)=V\left(t\right)+R_{1}i\left(t\right)+u_c\left(t\right)\\ i\left(t\right)=\frac{u_c\left(t\right)}{R_2}+C\frac{{\mathrm{du}}_c\left(t\right)}{\mathrm{dt}}\end{array}\right.---\left(1\right)$

(1) formula is equivalent to:

$\left\{\begin{array}{c}V\left(t\right)=\mathrm{OCV}\left(t\right)+R_{1}i\left(t\right)+u_c\left(t\right)\\ i\left(t\right)=\frac{u_c\left(t\right)}{R_2}+C\frac{{\mathrm{du}}_c\left(t\right)}{\mathrm{dt}}\end{array}\right.---\left(2\right)$

Wherein: OCV (t) is the function of SOC (t).

According to the above-mentioned math equation (2) obtained by battery equivalent-circuit model, corresponding output voltage V (k) and capacitance terminal voltage u can be obtained _cthe discrete magnitude relational expression of (k):

V(k)=OCV(k)-R ₁i(k)-u _c(k) （3）

u _c(k)=k ₂u _c(k-1)+k ₁i(k) （4）

Wherein:

$k_1=\frac{R_{2}T}{T+R_{2}C},$ $K_2=\frac{R_{2}C}{T+R_{2}C}$

T is the sampling period.

Neural network OCV(k) prediction model, as shown in Figure 7, comprise three neuron node J ₁, J ₂and J ₃, and the first radial basis function neural network RBF1.Discrete digital amount OCV(k in this model), V (k), i (k) and u _canalog quantity OCV (t), V (t), i (t) and u in (k) difference corresponding equation (2) _c(t).The output of three neuron nodes is respectively capacitance terminal voltage u _c(k), resistance R ₁terminal voltage and equivalent electrical circuit output voltage V (k).First radial basis function neural network RBF1 completes the On-line Estimation to battery open circuit terminal voltage OCV (k).Digital collection electrokinetic cell, obtains the discrete digital amount of the battery terminal voltage of actual measurement with discrete digital amount current i (k), as the input value of this model.

Peripheral sensory neuron node J ₁ask for capacitance terminal voltage u _c(k).Z ^-1for to back operator, the u that peripheral sensory neuron node exports _ck () passes through z ^-1obtain corresponding u _c(k-1).Peripheral sensory neuron node is according to formula u _c(k)=k ₂u _c(k-1)+k ₁i (k) is by weighting coefficient k ₁and k ₂respectively to the i(k gathering gained) and u _c(k-1) be weighted summation, obtain exporting u _c(k).

Nervus opticus unit node J ₂according to collecting i (k) and Model Parameter k ₃obtain R ₁on voltage be k ₃× i (k), k ₃represent R ₁.

Third nerve unit node J ₃calculate equivalent electrical circuit output voltage V (k) estimated value, V (k)=OCV (k)-R ₁i (k)-u _c(k), namely the weighting coefficient of this node summation operation is respectively 1, and-1 and-1.

The output OCV(k of the first radial basis function neural network RBF1) be the kinematic function of current i (k), equivalent electrical circuit output voltage V (k).OCV(k) z is passed through ^-1oCV (k-1) is obtained to back operator, OCV (k-1) is as external feedback, with the last sampling instant value V (k-1) of V (k) and i (k) input signal as the first radial basis function neural network RBF1, first radial basis function neural network RBF1 is to voltage V (k), current i (k) and OCV(k) kinematic nonlinearity characteristic be described, supplement battery equivalent-circuit model and expand, the output of the first radial basis function neural network RBF1 is battery open circuit terminal voltage OCV (k).

Neural network OCV(k) the weights k of prediction model ₁, k ₂and k ₃study, estimate the cell output voltage V(k obtained with this model) with the battery terminal voltage of actual measurement difference, form neural network OCV(k) the learning objective function of prediction model.Adopt ripe method of steepest descent, i.e. gradient method, by neural network OCV(k) prediction model, parameter R in the electrokinetic cell of identification simultaneously equivalent-circuit model ₁, R ₂with C and state OCV(k), realize the On-line Estimation of OCV (k).Its neural network OCV(k) the Parameter Self-learning adjustment capability of prediction model, the electrokinetic cell otherness characteristically of different capabilities can be adapted to.

Second step, according to OCV(k) to the complicated Hysteresis Nonlinear error compensation of state of charge, estimate current SOC(k) value

This step devises the Dynamic Hysteresis mixture model that simple dynamic hysteresis model-SDH model and the second radial basis function neural network RBF2 are composed in series, as shown in Figure 8.

The expression formula of SDH model is as follows:

$\left\{\begin{array}{c}y\left(t\right)=\mathrm{\&eta;}\left(t\right)-\frac{1}{k_4}\mathrm{OCV}\left(t\right)\\ \stackrel{\&CenterDot;}{\mathrm{\&eta;}}\left(t\right)=k_4(\mathrm{OCV}\left(t\right)-\mathrm{\&Delta;}\left(\mathrm{\&eta;}\left(t\right)\right))\end{array}\right.---\left(5\right)$

Wherein

$\mathrm{\&Delta;}\left(\mathrm{\&eta;}\left(t\right)\right)=\left\{\begin{array}{cc}{k}_4(\mathrm{\&eta;}\left(t\right)-1)& \mathrm{\&eta;}\left(t\right)>1\\ 0& \left|\mathrm{\&eta;}\left(t\right)\right|\&le;1\\ k_4(\mathrm{\&eta;}\left(t\right)+1)& \mathrm{\&eta;}\left(t\right)<-1\end{array}\right.---\left(6\right)$

Wherein, OCV (t) is the On-line Estimation value that the first step obtains, and be the input of SDH model, y (t) is the output of SDH model.

Parameter K4 in SDH model determines the retardant curve of SDH model and actual SOC(t)-OCV(t) sluggish relation curve at the similarity degree of contour shape, also determine the on-line study speed of the learning process of the second radial basis function neural network RBF2 be connected on after SDH model.

Corresponding discrete model can be obtained by SDH model.Analog quantity OCV (t) and y (t) respectively corresponding discrete digital amount OCV (k) and y (k), T is the sampling period.

The expression formula of SDH discrete model is as follows:

$\left\{\begin{array}{c}y\left(k\right)=\mathrm{\&eta;}\left(k\right)-\frac{1}{k_4}\mathrm{OCV}\left(k\right)\\ \frac{\mathrm{\&eta;}\left(k\right)-\mathrm{\&eta;}(k-1)}{T}=k_4(\mathrm{OCV}\left(k\right)-\mathrm{\&Delta;}\left(\mathrm{\&eta;}\left(k\right)\right))\end{array}\right.$

After arrangement:

$\left\{\begin{array}{c}y\left(k\right)=\mathrm{\&eta;}\left(k\right)-\frac{1}{k_4}\mathrm{OCV}\left(k\right)\\ \mathrm{\&eta;}\left(k\right)=\mathrm{\&eta;}(k-1)+T\&times;k_4(\mathrm{OCV}\left(k\right)-\mathrm{\&Delta;}\left(\mathrm{\&eta;}\left(k\right)\right))\end{array}\right.---\left(7\right)$

Wherein

$\mathrm{\&Delta;}\left(\mathrm{\&eta;}\left(k\right)\right)=\left\{\begin{array}{cc}{k}_4(\mathrm{\&eta;}\left(k\right)-1)& \mathrm{\&eta;}\left(k\right)>1\\ 0& \left|\mathrm{\&eta;}\left(k\right)\right|\&le;1\\ k_4(\mathrm{\&eta;}\left(k\right)+1)& \mathrm{\&eta;}\left(k\right)<-1\end{array}\right.---\left(8\right)$

SDH model has the Non-smooth surface lagging characteristics describing monocycle, i.e. outer shroud, as shown in Figure 9, its figure below horizontal ordinate is time t, its ordinate is the amplitude of SDH mode input signal OCV (t), represent the time dependent curve of size of input signal OCV (t) of SDH model, upper figure horizontal ordinate is the amplitude of SDH mode input signal OCV (t), its ordinate is SDH model output signal y (t), represents the input of SDH model, the relation curve of output signal.SDH model also has description secondary ring, i.e. many rings lagging characteristics of inner ring, and as shown in Figure 10, the parameter that represents of coordinate is identical with Fig. 9 in length and breadth for upper figure below of Figure 10.

SDH model is as the preamble part of Dynamic Hysteresis mixture model, and connect the second radial basis function neural network RBF2 after SDH model, builds Dynamic Hysteresis mixture model.The OCV (k) that the y (k) that SDH model obtains and the first step obtain, the previous moment value OCV (k-1) of OCV (k) are as the input of the second radial basis function neural network RBF2, by the weighting study in RBF2, realize the Nonlinear Mapping that any monodrome is corresponding, the k of Indirect method SDH model ₄parameter, to approach actual SOC(k) and OCV(k) complicated sluggish relation, the final SOC (k) exporting estimation on line.

Described Dynamic Hysteresis mixture model not only can represent sluggish outer shroud, i.e. main ring, also can describe sluggish multiple secondary ring, i.e. inner ring.When being used for the electric automobile in dynamic operation when electrokinetic cell, this Dynamic Hysteresis mixture model can describe the outer shroud and multiple interior ring property that the complicated lagging characteristics of Non-smooth surface in charging, discharging electric batteries Stochastic Dynamic Process shows simultaneously.

This routine Dynamic Hysteresis mixture model, first according to experimental data, adopts ripe steepest decline technology to complete the study of the second radial basis function neural network RBF2, adopts the output estimation SOC(k of the Dynamic Hysteresis mixture model trained) value.

The state of charge estimating system that electrokinetic cell Non-smooth surface lagging characteristics compensates

According to the charge state evaluation method that upper routine electrokinetic cell Non-smooth surface lagging characteristics compensates, set up the state of charge estimating system embodiment that this electrokinetic cell Non-smooth surface lagging characteristics compensates, system architecture as shown in figure 11, comprises microprocessor, current sensor, voltage sensor, analog to digital converter (AD converter), program storage, programmable storage, timer and display.Current sensor and voltage sensor export through analog to digital converter access microprocessor, microprocessor linker storer, programmable storage, timer and display.The display of this example is LCD display.

Neural network OCV(k is stored in program storage) prediction model, the SDH model calculation procedure of Dynamic Hysteresis mixture model of connecting with the second radial basis function neural network RBF2, the parameter in OCV (k) prediction model and SDH model is stored in programmable storage, current sensor and voltage sensor are installed in electrokinetic cell and load connecting circuit, the load current of current sensor and the electrokinetic cell measured by voltage sensor and terminal voltage are by analog to digital conversion, and the digital quantity obtaining corresponding load current and voltage sends into microprocessor.SOC(k with in timer control program storer) estimation program start and interrupt operation, the current SOC(k of the operation result of microprocessor) estimated value, shown in real time by display.

Above-described embodiment, be only the specific case further described object of the present invention, technical scheme and beneficial effect, the present invention is not defined in this.All make within scope of disclosure of the present invention any amendment, equivalent replacement, improvement etc., be all included within protection scope of the present invention.

Claims (5)

1. the charge state evaluation method of electrokinetic cell Non-smooth surface lagging characteristics compensation, estimates state of charge SOC (k) of electrokinetic cell in two steps, and the first step, for gathering electrokinetic cell, obtains the discrete digital amount of the battery terminal voltage of actual measurement with discrete digital amount current i (k), On-line Estimation is carried out to discrete digital amount OCV (k) of battery open circuit terminal voltage; Second step, for compensate state of charge Hysteresis Nonlinear error according to OCV (k), estimates current SOC (k) value;

The first step, On-line Estimation to discrete digital amount OCV (k) of battery open circuit terminal voltage

In battery equivalent electrical circuit, resistance R ₂resistance-capacitance circuit is formed, open terminal voltage OCV (t) the resistance in series R of battery with electric capacity C parallel connection ₁with above-mentioned resistance-capacitance circuit, battery equivalent electrical circuit output voltage is V (t), by resistance R ₁electric current be i (t), the terminal voltage of electric capacity C is u _c(t); Output voltage V (k) and battery open circuit terminal voltage OCV (k), current i (k) and capacitance terminal voltage u is obtained according to battery equivalent-circuit model _cthe discrete magnitude relation of (k) and capacitance terminal voltage u _cthe discrete magnitude relation of (k) and i (k):

V(k)＝OCV(k)-R ₁i(k)-u _c(k)

u _c(k)＝k ₂u _c(k-1)+k ₁i(k)

Wherein:

$k_1=\frac{R_{2}T}{T+R_{2}C},K_2=\frac{R_{2}C}{T+R_{2}C}$

T is the sampling period;

Corresponding above-mentioned formula builds neural network OCV (k) prediction model, comprises three neuron node (J ₁, J ₂, J ₃) and the first radial basis function neural network (RBF1), digital collection electrokinetic cell, the discrete digital amount of the battery terminal voltage of actual measurement with discrete digital amount current i (k), as the input value of this model;

Peripheral sensory neuron node (J ₁) ask capacitance terminal voltage u _c(k); z ^-1for to back operator, peripheral sensory neuron node (J ₁) u that exports _ck () passes through z ^-1obtain corresponding u _c(k-1), peripheral sensory neuron node (J ₁) according to formula u _c(k)=k ₂u _c(k-1)+k ₁i (k) is by weighting coefficient k ₁and k ₂respectively to i (k) and the u of digital collection gained _c(k-1) be weighted summation, obtain exporting u _c(k);

Nervus opticus unit node (J ₂) obtain i (k) and Model Parameter k according to digital collection ₃obtain R ₁on voltage be k ₃× i (k), k ₃represent R ₁;

Third nerve unit node (J ₃) calculate equivalent electrical circuit output voltage V (k) estimated value, V (k)=OCV (k)-R ₁i (k)-u _c(k), namely the weighting coefficient of this node summation operation is respectively 1 ,-1, and-1;

Output OCV (k) of the first radial basis function neural network (RBF1) is the kinematic function of current i (k), equivalent electrical circuit output voltage V (k); OCV (k) passes through z ^-1oCV (k-1) is obtained to back operator, OCV (k-1) is as external feedback, with the last sampling instant value V (k-1) of V (k) and i (k) input signal as the first radial basis function neural network (RBF1), first radial basis function neural network (RBF1) is described with the kinematic nonlinearity characteristic of OCV (k) voltage V (k), current i (k), supplement battery equivalent-circuit model and expand, the output of the first radial basis function neural network (RBF1) is battery open circuit terminal voltage OCV (k);

Second step, according to OCV (k) to the complicated Hysteresis Nonlinear error compensation of state of charge, estimate current SOC (k) value

This step adopts simple dynamic hysteresis model-SDH discrete model and the second radial basis function neural network (RBF2) to be composed in series Dynamic Hysteresis mixture model;

The expression formula of SDH discrete model is as follows:

$\left\{\begin{array}{c}y\left(k\right)=\mathrm{\&eta;}\left(k\right)-\frac{1}{k_4}\mathrm{COV}\left(k\right)\\ \mathrm{\&eta;}\left(k\right)=\mathrm{\&eta;}(k-1)+T\&times;k_4(\mathrm{OCV}\left(k\right)-\mathrm{\&Delta;}\left(\mathrm{\&eta;}\left(k\right)\right))\end{array}\right.$

$\mathrm{\&Delta;}\left(\mathrm{\&eta;}\left(k\right)\right)=\left\{\begin{array}{cc}{k}_4(\mathrm{\&eta;}\left(k\right)-1)& \mathrm{\&eta;}\left(k\right)>1\\ 0& \left|\mathrm{\&eta;}\left(k\right)\right|\&le;1\\ k_4(\mathrm{\&eta;}\left(k\right)+1)& \mathrm{\&eta;}\left(k\right)<-1\end{array}\right.$

Wherein, OCV (k) is the On-line Estimation value that the first step obtains, and be the input of SDH discrete model, y (k) is the output of SDH discrete model, T is the sampling period, be identical value with the sampling period in first step battery equivalent-circuit model formula, k ₄for the parameter of SDH discrete model;

The OCV (k) that the y (k) that SDH discrete model obtains and the first step obtain, the last sampling instant value OCV (k-1) of OCV (k) are as the input of the second radial basis function neural network (RBF2), by the weighting study in the second radial basis function neural network (RBF2), realize the Nonlinear Mapping that any monodrome is corresponding, the k of Indirect method SDH discrete model ₄parameter, to approach actual SOC (k) and the complicated sluggish relation of OCV (k), the final SOC (k) exporting estimation on line.

2. the charge state evaluation method of electrokinetic cell Non-smooth surface lagging characteristics compensation according to claim 1, is characterized in that:

The weights k of neural network OCV (k) prediction model in the described first step ₁, k ₂and k ₃study, estimate the battery terminal voltage of cell output voltage V (k) and the actual measurement obtained with this model difference, form the learning objective function of neural network OCV (k) prediction model; Adopt ripe method of steepest descent, by neural network OCV (k) prediction model, parameter R in the electrokinetic cell of identification simultaneously equivalent-circuit model ₁, R ₂and C, and state OCV (k), realize the On-line Estimation of OCV (k).

3. the charge state evaluation method of electrokinetic cell Non-smooth surface lagging characteristics compensation according to claim 1, is characterized in that:

The SDH discrete model of described second step has the Non-smooth surface lagging characteristics describing monocycle, i.e. outer shroud, also has description secondary ring, i.e. many rings lagging characteristics of inner ring.

4. the charge state evaluation method of electrokinetic cell Non-smooth surface lagging characteristics compensation according to claim 1, is characterized in that:

Described second step second radial basis function neural network (RBF2), first according to experimental data, adopts ripe method of steepest descent to complete study, adopts output estimation SOC (k) value of the Dynamic Hysteresis mixture model trained.

5. the charge state evaluation method of electrokinetic cell Non-smooth surface lagging characteristics compensation according to any one of claim 1 to 4, the state of charge estimating system of the electrokinetic cell Non-smooth surface lagging characteristics compensation of design, is characterized in that:

Comprise microprocessor, current sensor, voltage sensor, analog to digital converter, program storage, programmable storage, timer and display; Current sensor and voltage sensor export through analog to digital converter access microprocessor, microprocessor linker storer, programmable storage, timer and display;

Neural network OCV (k) prediction model is stored in program storage, the calculation procedure of the Dynamic Hysteresis mixture model that SDH discrete model is connected with the second radial basis function neural network (RBF2), the parameter in OCV (k) prediction model and SDH discrete model is stored in programmable storage, current sensor and voltage sensor are installed in electrokinetic cell and load connecting circuit, load current and the terminal voltage of current sensor and the electrokinetic cell measured by voltage sensor pass through analog to digital conversion, the digital quantity obtaining corresponding load current and voltage sends into microprocessor, the operation that SOC (k) estimation program in timer control program storer starts and interrupts, current SOC (k) estimated value of microprocessor operation result, is shown in real time by display.