CN103558554A - On-line estimating method for SOH of new energy automobile power battery - Google Patents

Abstract

A on-line estimating method for the SOH of a new energy automobile power battery includes the steps that S1, an SOH predicating model of the battery is established and optimized, and an improved Elman network (OIF Elman network) algorithm having an output-input feedback mechanism is adopted; S2, input parameters of the SOH model of the battery are selected; internal resistance, currents and temperature of the battery serve as the input parameters of the model; S3, operation is performed under the Matlab7.1 environment; S4, verification is performed. When the number of sample points is few, the OIF Elman network is obviously superior to an Elman network in whether training speed or predication precision. The generalization capacity of the OIF Elman network is improved, the requirement for the number of training samples is lowered, besides, the predication precision can be improved, and the on-line estimating method can be successfully applied to predicating the SOH of the power battery.

Description

A kind of method of new-energy automobile power battery SOH estimation on line

Technical field

The present invention relates to a kind of new-energy automobile power battery SOH(battery cycle life) method of estimation on line.

Background technology

Electrokinetic cell is the core of new-energy automobile, is the maximum bottleneck on new-energy automobile technology and cost, is also a ring most crucial in new-energy automobile industrial chain.Along with a large amount of consumption of Global Oil resource equal energy source, atmospheric pollution, weather and ecological deterioration further aggravate, the external interdependency of the energy surpass 50% and the overall background of development low-carbon economy under, development new-energy automobile has been trend of the times.

Compare with orthodox car, electric automobile has unrivaled advantage, but is still faced with at present the problem of many technology and cost aspect, is mainly the once charging distance travelled, battery life, charging equipment (communal facility) of electric automobile etc.And electrokinetic cell system life prediction problem is one of the most key problem of current battery system, the factor that affects battery life is a lot, and battery life is subject to its temperature that discharges and recharges mode of operation, size of current, running environment, pressure, cell making process, the structure of itself and the impact of the many factors such as chemical characteristic.So the life-span that simply and rapidly dopes battery seems very important.

For the prediction of battery life, consider and mainly contain based on battery mechanism, battery characteristics and data-driven now.In actual applications, should utilize battery mechanism model as far as possible, but its shortcoming is battery mechanism model, need meticulous parameter, complexity is higher, and the agine mechaism model of Erecting and improving is very difficult; Key based on signatures to predict is the relationship description finding between sensitive features parameter, feature on-line testing method and feature and the cell health state showing with cell degradation; Principle based on data-driven should be to describe at mechanism model the thought that in irrealizable situation, auxiliary data drives again, from data, excavate rule, pure fitting formula or the neural network model obtaining based on data-driven has certain limitation in actual applications, should be on the basis of mechanism or characteristic model, to be combined with the thought of data-driven again.

Summary of the invention

Technical matters to be solved by this invention, be just to provide a kind of forecast model comparatively simple, without fine parameters and can reduce the demand to training sample number, can improve again the method for the electric automobile power battery SOH estimation on line of precision of prediction.

Solve the problems of the technologies described above, the technical solution used in the present invention is:

A method for electric automobile power battery SOH estimation on line, is characterized in that: comprise the following steps:

Foundation and the optimization of S1 battery SOH forecast model

Employing has the improved Elman network algorithm of output-input feedback mechanism: OIF Elman network:

The same with general neural network, OIF Elman is divided into 3 layers: input layer, hidden layer and output layer, wherein input layer mathematical model is

x _u(k)=u(k-1) （1）；

X _u(k) represent k input layer state constantly, etching system input quantity when u (k) represents k;

The mathematical model of hidden neuron is:

x(k)＝f(W ¹¹x _c(k)+W ¹²x _u(k)+W ¹⁴y _c(k)) (2)

x _c(k)=αx _c(k-1)+x(k-1) （3）

y _c(k)=γy _c(k-1)+y(k-1) （4）；

Wherein, x (k) represents the k state of hidden node constantly; x _c(k) represent the k state of associative unit 1 node constantly; y _c(k) represent the k state of associative unit 2 nodes constantly; W ¹¹the connection weight matrix that represents implicit node and associative unit; W ¹²represent the connection weight matrix between hidden node and input node; W ¹⁴represent the connection weight matrix between hidden node and associative unit 2;

The neuronic mathematical model of output layer is:

y(k)=W ¹³x(k) （5）；

Wherein, y (k) represents the k state of output layer node constantly; W ¹³represent the connection weight matrix between hidden node and output node;

Being input as of the hidden node of described OIF Elman network: input layer state, associative unit 1 state, associative unit 2 states, establish the actual y of being output as of k step system _d(k), definition error function is:

$E\left(k\right)=\frac{1}{2}{(y_d\left(k\right)-y\left(k\right))}^T(y_d\left(k\right)-y\left(k\right))---\left(6\right);$

By E to connection weight W ¹¹, W ¹², W ¹³, W ¹⁴ask respectively local derviation, by gradient descent method, can be obtained the learning algorithm of OIF Elman network:

ΔW _ij ¹³＝η ₃δ _i ⁰x _j(k) i＝1，2，…，m；j＝1，2，…，n (7)

ΔW _jq ¹²＝η ₂δ _j ^hu _q(k-1) j＝1，2，…，n；q＝1，2，…，r (8)

$\mathrm{\Δ}{W_{\mathrm{jl}}}^{11}={\mathrm{\η}}_l\underset{i=1}{\overset{m}{\mathrm{\Σ}}}\left({{\mathrm{\δ}}_i}^0{W_{\mathrm{ij}}}^{13}\right)\frac{{\∂x}_i\left(k\right)}{{{\∂W}_{\mathrm{jl}}}^{11}},j=\mathrm{1,2},...,n;l=\mathrm{1,2},...,n---\left(9\right)$

$\mathrm{\Δ}{W_{\mathrm{js}}}^{14}={\mathrm{\η}}_4\underset{i=1}{\overset{m}{\mathrm{\Σ}}}\left({{{\mathrm{\δ}}_i}^0{W}_{\mathrm{ij}}}^{13}\right)\frac{{\∂x}_i\left(k\right)}{{{\∂W}_{\mathrm{js}}}^{14}},j=\mathrm{1,2},...,n;s=\mathrm{1,2},...,m---\left(10\right)$

$\frac{{\∂x}_i\left(k\right)}{{{\∂W}_{\mathrm{js}}}^{14}}={f_j}^{\′}(\·)y_s(k-1)+\mathrm{\γ}\frac{{\∂x}_i(k-1)}{{{\∂W}_{\mathrm{js}}}^{14}},j=\mathrm{1,2},...,n;s=\mathrm{1,2},...,m---\left(11\right)$

Wherein, η ₁, η ₂, η ₃, η ₄be respectively W ¹¹, W ¹², W ¹³, W ¹⁴learning Step, 0≤γ < 1 is the feedback gain of the connection certainly factor of associative unit 2, δ _i ⁰and δ _j ^hfor error signal;

δ _i ⁰＝(y _di(k)-y _i(k)) (12)

${{\mathrm{\&delta;}}_j}^h=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}\left({{\text{\&delta;}}_i}^0{W_{\mathrm{ij}}}^{13}\right){f_j}^{\&prime;}(\&CenterDot;)---\left(13\right)$

$\frac{{\text{\&PartialD;x}}_i\left(k\right)}{{{\&PartialD;\mathrm{<figure-callout\; id="1"\; label="W\; j"\; filenames="HDA0000395999200000011.png"\; state="\{\{state\}\}">W</figure-callout>}}_j}^{11}}={f_j}^{\&prime;}(\&CenterDot;)x_l(k-1)+\mathrm{\&alpha;}\frac{{\&PartialD;x}_i(k-1)}{\&PartialD;{W_{\mathrm{jl}}}^{11}}---\left(14\right);$

Described OIF Elman Learning Algorithms adopts gradient descent method, it is adaptive learning speed momentum gradient decline back-propagation algorithm, it can improve the training speed of network, effectively Suppression network is absorbed in local minimum point again, the destination of study is to revise weights and bias by the difference of the real output value of network and output sample value, makes the error sum of squares of network output layer minimum.

The selection of S2 battery SOH mode input parameter

Select the internal resistance of cell, electric current and temperature as the input parameter of model;

Determine the number n of hidden neuron ₂with input layer number n ₁:

The number n of hidden neuron ₂with input layer number n ₁between have a following approximation relation:

n ₂=2n ₁+1 （15）

From sample data, network input layer has 3 neurons, and output layer has 1 neuron, therefore the neuronic number in middle layer is set to 7 respectively;

S3 moves under Matlab7.1 environment, for fear of numerical value in the computer simulation process in neural network, overflow, must carry out standardization to each cell value of network input layer and output layer, and make them in [0,1] or [1,1], in, the standardized method of the input parameter that the present invention adopts is:

$\hat{x}=\frac{x-x_{\min }}{x_{\max }-x_{\min }}---\left(16\right)$

In formula, x _maxfor the maximal value of this group variable, x _minminimum value for this group variable;

Concrete operation step is as follows:

S3-1 parameter initialization;

S3-2 batch input learning sample, and input and output amount is normalized;

S3-3 calculates each layer and associative unit output;

S3-4 calculates output error;

S3-5 judgement output error meets the demands no: if finish after the weights that storage has been trained that meet the demands; If do not meet the demands, according to formula 12-13, do not calculate each layer of error signal, according to formula 7-10, adjust each layer of weights, return to step S3-3;

The simulation analysis of S4 checking battery SOH forecast model

Select the battery of 6 groups of same model, adopt identical charging system to be full of electricity, under the condition of different electric currents, different temperatures, 3 Battery packs are carried out to discharge test, obtain 100 groups of measured datas, select wherein 80 groups of data to this model training, network training process is moved under Matlab7.1 environment, in emulation, with improving Elman and two kinds of neural net model establishings of Elman, predict, predicted the outcome respectively, comparative analysis both with the error of actual measurement SOH.

Beneficial effect: the present invention is chosen in the life-span of predicting battery on characteristic model basis in conjunction with the thought of data-driven, having proposed to have the improvement Elman(OIF Elman of output-input feedback) neural network algorithm predicts the residual life of electric automobile power battery (SOH), with traditional Elman network contrast, OIF(output-input feedback) Elman network has not only added the feedback of hidden node, and counts the feedback of output layer node.Predict the outcome and show, when sample point is less, no matter on training speed, or on precision of prediction, OIF Elman network is obviously better than Elman network.OIF Elman network has improved the generalization ability of network, has both reduced the demand to training sample number, can improve precision of prediction again, can successful Application in electrokinetic cell residual life (SOH) prediction.

Accompanying drawing explanation

Fig. 1 is OIF Elman neural networks principles process flow diagram of the present invention;

Fig. 2 is OIF Elman algorithm flow chart in Matlab7.1.

Embodiment

This electric automobile power battery SOH estimation on line method comprises following part: foundation and the optimization of battery SOH predictive model algorithm, the selection of forecast model input parameter, the simulation comparison of forecast model result and analysis.

First be foundation and the optimization of algorithm, Elman neural network can be regarded as a feedforward neural network with local mnemon and local feedback link, different from general neural network structure is, many associated layers in Elman neuromechanism, its effect is for remembering the former output valve constantly of Hidden unit, can think a time delay operator, it makes whole network have the function of dynamic memory.And the method that the present invention puts forward is the improved Elman network algorithm with output-input feedback mechanism, i.e. OIF Elman network.OIF Elman network has increased the feedback of output node than Elman network, to obtain more information from limited training sample.

Then be the selection of battery SOH mode input parameter, consider that to affect the factor of battery life a lot, should be using principal element as input value, through comparative analysis, select to affect battery life principal element---the internal resistance of cell, electric current and temperature are as the input parameter of model;

It is finally the simulation analysis of battery SOH forecast model, in experiment, select the battery of 6 groups of same model, adopt identical charging system to be full of electricity, under the condition of different electric currents, different temperatures, 3 Battery packs are carried out to discharge test, obtain 100 groups of measured datas, select wherein 80 groups of data to this model training, network training process is moved under Matlab7.1 environment, in emulation, with improving Elman and two kinds of neural net model establishings of Elman, predict, predicted the outcome respectively, comparative analysis both with actual measurement SOH error.

Below in conjunction with Fig. 1, Fig. 2 is described in further detail the online SOH Forecasting Methodology of this battery system each several part.

Fig. 1 is OIF Elman neural networks principles process flow diagram.

In Fig. 1 OIF Elman neural network, the selection of u (k) input parameter should be the major influence factors of SOH, the factor that affects SOH is a lot, by analysis the present invention above, select the internal resistance, electric current, temperature of battery as input value, thereby can adopt the OIF Elman neural network model of 3-N-1, output valve is SOH, be that input layer number is 3, hidden neuron number is undetermined, and output layer neuron number is 1.

Definite especially middle layer neuron number object of OIF Elman neural network practical structures determines it is an experimental problem, also needs a large amount of experiments.In three-layer network, the number n of hidden neuron ₂with input layer number n ₁between have a following approximation relation:

n ₂=2n ₁+1 （17）

From sample data, network input layer should have 3 neurons, and output layer should have 1 neuron, therefore the number of hidden neuron is set to 7 respectively.

Fig. 2 is OIF Elman algorithm flow chart in Matlab7.1, network training process is moved under Matlab7.1 environment, for fear of numerical value in the computer simulation process in neural network, overflow, must carry out standardization to each cell value of network input layer and output layer, and make them in [0,1] or in [1,1], the standardized method of the input parameter that the present invention adopts is:

$\hat{x}=\frac{x-x_{\min }}{x_{\max }-x_{\min }}---\left(18\right)$

In formula, x _maxfor the maximal value of this group variable, x _minminimum value for this group variable.

Claims (2)

1. a method for electric automobile power battery SOH estimation on line, is characterized in that: comprise the following steps:

Foundation and the optimization of S1 battery SOH forecast model

Employing has the improved Elman network algorithm of output-input feedback mechanism: OIF Elman network, OIF Elman is divided into 3 layers: input layer, hidden layer and output layer, wherein input layer mathematical model is

x _u(k)=u(k-1) （1）；

X _u(k) represent k input layer state constantly, etching system input quantity when u (k) represents k;

The mathematical model of hidden neuron is:

x(k)＝f(W ¹¹x _c(k)+W ¹²x _u(k)+W ¹⁴y _c(k)) (2)

x _c(k)=αx _c(k-1)+x(k-1) （3）

y _c(k)=γy _c(k-1)+y(k-1) （4）；

Wherein, x (k) represents the k state of hidden node constantly; x _c(k) represent the k state of associative unit 1 node constantly; y _c(k) represent the k state of associative unit 2 nodes constantly; W ¹¹the connection weight matrix that represents implicit node and associative unit; W ¹²represent the connection weight matrix between hidden node and input node; W ¹⁴represent the connection weight matrix between hidden node and associative unit 2;

The neuronic mathematical model of output layer is:

y(k)=W ¹³x(k) （5）；

Wherein, y (k) represents the k state of output layer node constantly; W ¹³represent the connection weight matrix between hidden node and output node;

Being input as of the hidden node of described OIF Elman network: input layer state, associative unit 1 state, associative unit 2 states, establish the actual y of being output as of k step system _d(k), definition error function is:

$E\left(k\right)=\frac{1}{2}{(y_d\left(k\right)-y\left(k\right))}^T(y_d\left(k\right)-y\left(k\right))---\left(6\right);$

By E to connection weight W ¹¹, W ¹², W ¹³, W ¹⁴ask respectively local derviation, by gradient descent method, can be obtained the learning algorithm of OIF Elman network:

ΔW _ij ¹³＝η ₃δ _i ⁰x _j(k) i=1，2，…，m；j=1，2，…，n (7)

ΔW _jq ¹²=η ₂δ _j ^hu _q(k-1) j＝1，2，…，n；q=1，2，…，r (8)

$\mathrm{\&Delta;}{W_{\mathrm{jl}}}^{11}={\mathrm{\&eta;}}_l\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}\left({{\mathrm{\&delta;}}_i}^0{W_{\mathrm{ij}}}^{13}\right)\frac{{\&PartialD;x}_i\left(k\right)}{{{\&PartialD;W}_{\mathrm{jl}}}^{11}},j=\mathrm{1,2},...,n;l=\mathrm{1,2},...,n---\left(9\right)$

$\mathrm{\&Delta;}{W_{\mathrm{js}}}^{14}={\mathrm{\&eta;}}_4\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}\left({{{\mathrm{\&delta;}}_i}^0{W}_{\mathrm{ij}}}^{13}\right)\frac{{\&PartialD;x}_i\left(k\right)}{{{\&PartialD;W}_{\mathrm{js}}}^{14}},j=\mathrm{1,2},...,n;s=\mathrm{1,2},...,m---\left(10\right)$

$\frac{{\&PartialD;x}_i\left(k\right)}{{{\&PartialD;W}_{\mathrm{js}}}^{14}}={f_j}^{\&prime;}(\&CenterDot;)y_s(k-1)+\mathrm{\&gamma;}\frac{{\&PartialD;x}_i(k-1)}{{{\&PartialD;W}_{\mathrm{js}}}^{14}},j=\mathrm{1,2},...,n;s=\mathrm{1,2},...,m---\left(11\right)$

Wherein, η ₁, η ₂, η ₃, η ₄be respectively W ¹¹, W ¹², W ¹³, W ¹⁴learning Step, 0≤γ < 1 is the feedback gain of the connection certainly factor of associative unit 2, δ _i ⁰and δ _j ^hfor error signal;

δ _i ⁰＝(y _di(k)-y _i(k)) (12)

${{\mathrm{\&delta;}}_j}^h=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}\left({{\text{\&delta;}}_i}^0{W_{\mathrm{ij}}}^{13}\right){f_j}^{\&prime;}(\&CenterDot;)---\left(13\right)$

$\frac{{\text{\&PartialD;x}}_i\left(k\right)}{{{\&PartialD;W}_{j1}}^{11}}={f_j}^{\&prime;}(\&CenterDot;)x_l(k-1)+\mathrm{\&alpha;}\frac{{\&PartialD;x}_i(k-1)}{\&PartialD;{W_{\mathrm{jl}}}^{11}}---\left(14\right);$

Described OIF Elman Learning Algorithms adopts gradient descent method, it is adaptive learning speed momentum gradient decline back-propagation algorithm, it can improve the training speed of network, effectively Suppression network is absorbed in local minimum point again, the destination of study is to revise weights and bias by the difference of the real output value of network and output sample value, makes the error sum of squares of network output layer minimum;

The selection of S2 battery SOH mode input parameter

Select the internal resistance of cell, electric current and temperature as the input parameter of model;

Determine the number n of hidden neuron ₂with input layer number n ₁:

The number n of hidden neuron ₂with input layer number n ₁between have a following approximation relation:

n ₂=2n ₁+1 （14）

From sample data, network input layer has 3 neurons, and output layer has 1 neuron, therefore the neuronic number in middle layer is set to 7 respectively;

S3 moves under Matlab7.1 environment

Each cell value to network input layer and output layer carries out standardization, and makes them in [0,1] or [1,1], and the standardized method of the input parameter of employing is:

$\hat{x}=\frac{x-x_{\min }}{x_{\max }-x_{\min }}---\left(15\right)$

In formula, x _maxfor the maximal value of this group variable, x _minminimum value for this group variable;

The simulation analysis of S4 checking battery SOH forecast model is selected the battery of 6 groups of same model, adopt identical charging system to be full of electricity, under the condition of different electric currents, different temperatures, 3 Battery packs are carried out to discharge test, obtain 100 groups of measured datas, select wherein 80 groups of data to this model training, network training process is moved under Matlab7.1 environment, in emulation, with improving Elman and two kinds of neural net model establishings of Elman, predict, predicted the outcome respectively, comparative analysis both with actual measurement SOH error.

2. the method for electric automobile power battery SOH estimation on line according to claim 1, is characterized in that: described step S3 comprises following sub-step:

S3-1 parameter initialization;

S3-2 batch input learning sample, and input and output amount is normalized;

S3-3 calculates each layer and associative unit output;

S3-4 calculates output error;

S3-5 judgement output error meets the demands no: if finish after the weights that storage has been trained that meet the demands; If do not meet the demands, according to formula 12-13, do not calculate each layer of error signal, according to formula 7-10, adjust each layer of weights, return to step S3-3.

Images