CN111948563A - A method for predicting the remaining life of electric forklift lithium battery based on multi-neural network coupling - Google Patents

Abstract

In order to overcome the problems of complicated calculation and long time-consuming and low prediction accuracy when predicting the remaining life of a lithium battery in the prior art, the invention provides a method for predicting the remaining life of an electric forklift lithium battery based on multi-neural network coupling, which improves the prediction calculation accuracy. The training time of the prediction model is reduced, including the following steps: establishing an open-circuit voltage prediction model based on a long-short-term memory neural network, and using the RMSprop algorithm and the dropout regularization method to optimize the network to predict the open-circuit voltage value V of the lithium battery during the discharge cycle. _OC ; Divide the prediction results into multiple discharge cycles in sequence, count the number of open-circuit voltage samples N _S from the initial voltage to the minimum voltage in each discharge cycle, and use the same sampling time T _S to obtain the discharge in each discharge cycle. Time T _min to the minimum voltage; establish a capacity prediction model based on an artificial neural network to predict the lithium battery capacity C, thereby obtaining the predicted value RUL of the remaining life of the lithium battery.

Description

A prediction of remaining life of electric forklift lithium battery based on multi-neural network coupling method

technical field

The invention relates to the technical field of batteries, in particular to a method for predicting the remaining life of a battery based on multi-neural network coupling.

Background technique

Lithium batteries have high energy density, long cycle life and high discharge platform, which can provide a reliable power source for a variety of construction machinery. At present, most electric forklift trucks use lithium batteries as their power source. When the lithium battery is continuously charged and discharged, its electrochemical performance will gradually decrease, and the capacity will gradually decay. Increase, reduce safety, easily lead to unstable equipment operation and cause economic losses, and even lead to serious safety accidents. Therefore, the lithium battery needs to be maintained and replaced in time, and the remaining life prediction method can give the attenuation of the battery capacity in advance, so that a reasonable maintenance plan can be formulated in advance to ensure the stable operation of the electric forklift. Remaining life prediction refers to predicting the number of cycles that a battery can still run before reaching the end of its life, given the partial capacity decay data.

The traditional remaining life prediction method is a model-based method, which establishes a model based on the battery failure principle and electrochemical reaction, and predicts the battery life decay. The disadvantage of this method is that the calculation is complicated and the accuracy is low.

SUMMARY OF THE INVENTION

In order to overcome the problems of complicated calculation and long time-consuming and low prediction accuracy when predicting the remaining life of a lithium battery in the prior art, the invention provides a method for predicting the remaining life of an electric forklift lithium battery based on multi-neural network coupling, which improves the prediction calculation accuracy. Reduced predictive model training time.

In order to achieve the above object, the present invention adopts the following technical solutions, comprising the following steps:

S1: establish an open-circuit voltage prediction model based on a long-short-term memory neural network, and use the RMSprop algorithm and the dropout regularization method to optimize the network to predict the open-circuit voltage value V _OC of the lithium battery during the discharge cycle;

S2: Divide the prediction results into multiple discharge cycles in sequence, count the number of open-circuit voltage samples N _S from the initial voltage to the minimum voltage in each discharge cycle, and use the same sampling time T _S to obtain the discharge cycle in each discharge cycle. the time T _min to the minimum voltage;

S3: Establish a capacity prediction model based on an artificial neural network to predict the capacity C of the lithium battery, thereby obtaining the predicted value RUL of the remaining life of the lithium battery.

By combining the long-short-term memory neural network and artificial neural network, a prediction model for the remaining life of electric forklift lithium batteries is established, which makes full use of the long-short-term memory neural network's ability to fit and predict nonlinear data, and takes advantage of the simple structure of the artificial neural network. , and use the RMSprop algorithm and the dropout regularization method to optimize the prediction model, improve the calculation accuracy of the model, and reduce the model training time.

Preferably, the specific steps of the S1 are as follows:

S11: Set the number of input neurons to N _I1 , set the number of neural hidden layers to L _H1 and the number of hidden layer neurons to N _H1 , use the sliding window to establish the relationship mapping between input parameters and output parameters, set The length of the sliding window is L, and the length of the training data is M, then the mapping relationship between the input V _input and the output V _output is established during training, as shown in formula (1):

Among them, each row in V _input corresponds to an input sample, and the length is L;

S12: Use the model trained in S11 to predict the open-circuit voltage V _OC using the following method, and the input and output are shown in formula (2):

利用得到的预测值

对下一次预测的输入样本序列进行更新，即将预测值

加入到输入样本序列的末尾，并删掉输入样本序列的第一个数据，从而完成窗口的滑移操作，得到全部的开路电压预测值

作为优选，所述S2的具体步骤如下：Among them, each row in V _OC corresponds to a predicted input sample, and each prediction obtains a predicted value of open circuit voltage

use the predicted value

Update the input sample sequence for the next prediction, that is, the predicted value

Add to the end of the input sample sequence, and delete the first data of the input sample sequence, so as to complete the sliding operation of the window and obtain all the predicted values of open circuit voltage

As preferably, the concrete steps of described S2 are as follows:

S21: The method of dividing the prediction result into multiple discharge cycles in sequence is:

If time T _j makes equation (15) true:

Then T _j (j=0, 1, 2, ..., n; T ₀ =1) is the termination time of the jth cycle, and the open-circuit voltage sequence of the jth cycle is

S22: Count the number of open-circuit voltage samples N _S from the initial voltage to the minimum voltage in each discharge cycle, and use the same sampling time T _S to obtain the time T _min from discharging to the minimum voltage in each discharge cycle, as shown in formula (16 ) as shown:

T _min =N _S ×T _S (16)

As preferably, the concrete steps of described S3 are as follows:

S31: Set the number of input neurons as N _I2 , set the number of neural hidden layers as L _H2 and the number of hidden layer neurons as N _H2 , establish the time T _min from the discharge cycle to the minimum voltage to the capacity of the lithium battery The mapping relationship of C;

S32: The calculation process of each neuron in the hidden layer is shown in formula (17):

H _T =σ(W _HT T _min +b _H ); (17)

H _T represents the hidden layer calculation function, where W _HT and b _H are the weight matrix and bias parameter matrix of the input value of the function; after 20 iterations, the training ends and the model is saved. Input the predicted time T _min from discharging to the minimum voltage in each discharge cycle into the model to obtain the capacity prediction result;

S33: Define the end point EOL of the lithium battery life as the cycle number j corresponding to when the capacity C _j is equal to 70% of the rated capacity C ₀ ; define and calculate the remaining life RUL of the lithium battery as shown in formula (18):

RUL=UL-EOL; (18)

Among them, UL is the number of cycles that the current lithium battery has run, and EOL is the end-of-life point of the lithium battery.

The beneficial effects of the present invention

The invention establishes a long-short-term memory-artificial neural network multi-neural network coupling model, which effectively improves the prediction accuracy. The RMSprop algorithm is used for training, which improves the accuracy of model prediction; this model uses the dropout method to optimize the model, which avoids the problem of overfitting, reduces the time of model training, and improves the accuracy of model prediction. Using the open circuit voltage as the prediction input data, the accuracy of the prediction results is significantly improved compared with the long-short-term memory-artificial neural network coupling model using the discharge time as the characteristic parameter and the simple long-short-term memory neural network model, which can accurately predict the lithium ion of the electric forklift. remaining battery life.

Description of drawings

Fig. 1 is the general flow chart of the present invention;

Fig. 2 is the analysis diagram of the first group of prediction results of the present invention;

Fig. 3 is the analysis diagram of the second group of prediction results of the present invention;

Figure 4 shows the absolute error of remaining life prediction;

Figure 5 shows the root mean square error of capacity decay predictions.

Detailed ways

The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

The overall flow diagram of the method for predicting the remaining life of an electric forklift lithium battery based on multi-neural network coupling of the present invention is shown in Figure 1, and the specific operation process includes:

S1: Establish an open-circuit voltage prediction model based on a long-short-term memory neural network, and use the RMSprop algorithm and dropout regularization method to optimize the network to predict the open-circuit voltage value V _OC of the lithium battery during the discharge cycle. The specific steps are as follows:

S11: Set the number of input neurons to N _I1 , set the number of neural hidden layers to L _H1 and the number of hidden layer neurons to N _H1 , use the sliding window to establish the relationship mapping between input parameters and output parameters, set The length of the sliding window is L, and the length of the training data is M, then the mapping relationship between the input V _input and the output V _output is established during training, as shown in formula (1):

Among them, each row in V _input corresponds to an input sample, and the length is L;

S12: Use the model trained in S11 to predict the open-circuit voltage V _OC using the following method, and the input and output are shown in formula (2):

利用得到的预测值

对下一次预测的输入样本序列进行更新，即将预测值

加入到输入样本序列的末尾，并删掉输入样本序列的第一个数据，从而完成窗口的滑移操作，得到全部的开路电压预测值

Among them, each row in V _OC corresponds to a predicted input sample, and each prediction obtains a predicted value of open circuit voltage

use the predicted value

Update the input sample sequence for the next prediction, that is, the predicted value

Add to the end of the input sample sequence, and delete the first data of the input sample sequence, so as to complete the sliding operation of the window and obtain all the predicted values of open circuit voltage

The long-short-term memory neural network has good prediction and fitting ability for time series data. In order to prevent over-fitting, dropout regularization is performed on the network, that is, 0 and 1 vectors are randomly generated, thereby randomly discarding some neurons in the network. , as shown in equations (3) and (4):

Among them, R _T obeys the Bernoulli distribution function, and p is the dropout regularization ratio; the neurons of each long and short-term memory neural network include three kinds of information selection and conversion mechanisms: input gate, forget gate and output gate. The process is as follows:

The forget gate determines whether the information from the previous neuron is retained, and takes a value of 1 or 0. The former represents retention and the latter represents discard, as shown in formula (5):

where W _fv and W _f h are the weight matrix of the input value of the current neuron of the forget gate function and the output value of the previous neuron, respectively, and b _f is the bias parameter matrix;

The input gate calculates the retained information r _T and the new state information s _T of the current neuron input, as shown in equations (6) and (7):

Among them, W _rv and W _sv are the weight matrices of the current neuron input values of functions r _T and s _T in the input gate, respectively, and W _rh and W _sh are the output values of the previous neuron of functions r _T and s _T in the input gate, respectively The weight matrix of , b _r and b _s are the bias parameter matrices of the functions r _T and s _T in the input gate, respectively;

Afterwards, the state c _T of the long-short-term memory neural network neurons is updated as in Equation (8):

c _T =r _T ⊙s _T +c _T-1 ⊙f _T (8)

The output gate calculates the output information h _T of the current neuron, as shown in equations (9) and (10):

h _T =o _T ⊙tanh(c _T ) (10)

where W _ov and W _oh are the weight matrix of the input value of the current neuron and the output value of the previous neuron in the function o _T in the output gate, and b _o is the bias parameter matrix of the function o _T ;

Use the RMSprop algorithm to update the weight matrix ω and bias parameter matrix b of the long-short-term memory neural network as shown in Equation (11-14):

where J ₁ is the cost function for calculating the weight matrix, J ₂ is the cost function for calculating the bias parameter matrix, k represents the number of current iterations, β is the coefficient, usually 0.999, ε is used to prevent division by zero, usually 10 ^-8 , α is the learning rate, the algorithm belongs to the accelerated gradient descent method, so after 20 iterations, the training of the long-short-term memory neural network is over, and the model is saved; input the known data to get the open-circuit voltage prediction sequence.

S2: Divide the prediction results into multiple discharge cycles in sequence, count the number of open-circuit voltage samples N _S from the initial voltage to the minimum voltage in each discharge cycle, and use the same sampling time T _S to obtain the discharge cycle in each discharge cycle. The time T _min to the minimum voltage, the specific steps are as follows:

S21: The method of dividing the prediction result into multiple discharge cycles in sequence is:

If time T _j makes equation (15) true:

Then T _j (j=0, 1, 2, ..., n; T ₀ =1) is the termination time of the jth cycle, and the open-circuit voltage sequence of the jth cycle is

S22: Count the number of open-circuit voltage samples N _S from the initial voltage to the minimum voltage in each discharge cycle, and use the same sampling time T _S to obtain the time T _min from discharging to the minimum voltage in each discharge cycle, as shown in formula (16 ) as shown:

T _min =N _S ×T _S (16)

S3: Establish a capacity prediction model based on an artificial neural network to predict the capacity C of the lithium battery, thereby obtaining the predicted value RUL of the remaining life of the lithium battery. The specific steps are as follows:

S31: Set the number of input neurons as N _I2 , set the number of neural hidden layers as L _H2 and the number of hidden layer neurons as N _H2 , establish the time T _min from the discharge cycle to the minimum voltage to the capacity of the lithium battery The mapping relationship of C;

S32: The calculation process of each neuron in the hidden layer is shown in formula (17):

H _T =σ(W _HT T _min +b _H ); (17)

H _T represents the hidden layer calculation function, where W _HT and b _H are the weight matrix and bias parameter matrix of the input value of the function; after 20 iterations, the training ends and the model is saved. Input the predicted time T _min from discharging to the minimum voltage in each discharge cycle into the model to obtain the capacity prediction result;

S33: Define the end point EOL of the lithium battery life as the cycle number j corresponding to when the capacity C _j is equal to 70% of the rated capacity C ₀ ; define and calculate the remaining life RUL of the lithium battery, as shown in formula (18):

RUL=UL-EOL; (18)

Among them, UL is the number of cycles that the current lithium battery has run, and EOL is the end of life of the lithium battery.

The present invention uses Python language for modeling, respectively verifies the prediction results of the above methods for the lithium battery life decay data provided by NASA AmesPrognostics Center of Excellence (PCoE) and Center for Advanced Life CycleEngineering (CALCE), and uses the discharge time as a feature. The prediction results of the long-short-term memory-artificial neural network coupling model of parameters and the simple long-short-term memory neural network model were compared. Each group of experiments used 30% data for training, and the prediction results are shown in Figure 2 and Figure 3 ( Figure 2 is different from the lithium battery sample used in Figure 3), where M1 is the prediction result of the long-short-term memory-artificial neural network coupling model with open circuit voltage as the characteristic parameter, M2 is the long-short-term memory-artificial neural network with discharge time as the characteristic parameter The prediction result of the network coupling model, M3 is the prediction result of a simple long-short-term memory neural network model, it can be seen that the long-short-term memory-artificial neural network coupling model established by the present invention with open circuit voltage as a characteristic parameter has a better performance than other models. Better prediction effect. By predicting different types of batteries, it can be seen that the long-short-term memory-artificial neural network coupling model with open circuit voltage as a characteristic parameter established by the present invention has good generalization performance. For further explanation, the prediction results are evaluated by absolute error (AE, absolute error) and root mean squared error (RMSE, root mean squared error). The calculation process is shown in equations (19) and (20):

为电池剩余循环寿命的预测值，C_j为第j个循环电池容量的真实值，

为第j个循环电池容量的预测值。计算结果如图4和图5所示(图4为剩余寿命预测的绝对误差，图5为容量衰减预测的均方根误差)，Among them, RUL is the true value of the remaining cycle life of the battery,

is the predicted value of the remaining cycle life of the battery, C _j is the actual value of the battery capacity in the jth cycle,

is the predicted value of the battery capacity for the jth cycle. The calculation results are shown in Figure 4 and Figure 5 (Figure 4 is the absolute error of the remaining life prediction, Figure 5 is the root mean square error of the capacity decay prediction),

The long-short-term memory-artificial neural network coupling model with open circuit voltage as the characteristic parameter established by the present invention has smaller error in predicting the remaining life of lithium battery, and has higher fitting accuracy for the actual capacity decay curve, and has a very high residual life. Lifetime prediction accuracy.

The foregoing embodiments are only used to illustrate the technical solutions of the present invention, but not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that the foregoing embodiments can still be used for The recorded technical solutions are modified, or some or all of the technical features thereof are equivalently replaced; and these modifications or replacements do not make the essence of the corresponding technical solutions depart from the scope of the technical solutions of the embodiments of the present invention.

Claims (4)

1. A method for predicting the remaining life of a lithium battery of an electric forklift based on multi-neural network coupling is characterized by comprising the following steps:

s1: establishing an open-circuit voltage prediction model based on a long-time and short-time memory neural network, and adopting an RMSprop algorithm and a dOptimizing the network by using a ropout regularization method, thereby predicting the open-circuit voltage value V of the lithium battery in the discharge cycle_OC；

S2: dividing the prediction result into a plurality of discharge cycles in sequence, and counting the number N of open-circuit voltage samples from the initial voltage to the minimum voltage in each discharge cycle_SUsing the sampling time T_SThe same, the time T to discharge to the minimum voltage in each discharge cycle is obtained_min；

S3: and establishing a capacity prediction model based on an artificial neural network to predict the capacity C of the lithium battery so as to obtain a predicted value RUL of the residual life of the lithium battery.

2. The method for predicting the residual life of the lithium battery of the electric forklift based on the multi-neural-network coupling as claimed in claim 1, wherein the specific steps of S1 are as follows:

s11: setting the number of input neurons to N_I1Setting the number of the nerve hiding layers to be L_H1And number of hidden layer neurons N_H1Establishing a relation mapping between input parameters and output parameters by using a sliding window, and establishing an input V during training if the length of the sliding window is L and the length of training data is M_inputTo outputs of V respectively_outputThe mapping relation of (2) is shown in formula (1):

wherein, V_inputEach row in the set corresponds to an input sample with a length of L;

s12: model obtained by training S11 is utilized to split open-circuit voltage V_OCThe prediction is carried out by adopting the following method, wherein the input and the output are shown as the formula (2):

wherein, V_OCOne for each row inEach time of prediction to obtain a predicted value of open-circuit voltage

(t ═ M +1, M +2, …, N), using the predicted values obtained

Updating the input sample sequence for the next prediction, i.e. the predicted value

Adding the data to the end of the input sample sequence, and deleting the first data of the input sample sequence, thereby completing the window sliding operation and obtaining all predicted values of the open-circuit voltage

。

3. The method for predicting the residual life of the lithium battery of the electric forklift based on the multi-neural-network coupling as claimed in claim 1, wherein the specific steps of S2 are as follows:

s21: the method for sequentially dividing the prediction result into a plurality of discharge cycles is as follows:

if the time T_jSo that equation (15) holds:

then T_j(j＝0，1，2，…，n；T₀1) is the end time of the jth cycle, and the open circuit voltage sequence of the jth cycle is

S22: counting the number N of open-circuit voltage samples from the initial voltage to the minimum voltage in each discharge cycle_SUsing the sampling time T_SSame, get each discharge cycleTime T for discharging to minimum voltage_minAs shown in formula (16):

T_min＝N_S×T_S (16)

4. the method for predicting the residual life of the lithium battery of the electric forklift based on the multi-neural-network coupling as claimed in claim 1, wherein the specific steps of S3 are as follows:

s31: setting the number of input neurons to N_I2Setting the number of the nerve hiding layers to be L_H2And number of hidden layer neurons N_H2Establishing the time T from discharge to minimum voltage in the discharge cycle_minMapping relation to lithium battery capacity C;

s32: the calculation process of each neuron of the hidden layer is shown as the formula (17):

H_T＝σ(W_HTT_min+b_H)； (17)

H_Trepresenting a hidden layer computation function, where W_HTAnd b_HInputting a weight matrix and a bias parameter matrix of the function input value; after 20 times of iteration, the training is finished, and the model is saved. The predicted time T for discharging to the minimum voltage in each discharge cycle_minInputting a model to obtain a capacity prediction result;

s33: defining EOL as capacity C_jEqual to rated capacity C₀The number j of cycles corresponding to 70%; defining and calculating the residual life RUL of the lithium battery, as shown in formula (18):

RUL＝UL-EOL； (18)

wherein, UL is the number of the operating cycles of the current lithium battery, and EOL is the end point of the service life of the lithium battery.

Images