CN110703120A - Lithium ion battery service life prediction method based on particle filtering and long-and-short time memory network - Google Patents

Abstract

Based on the statistical learning theory, the invention realizes a method combining particle filtering and long-short-term memory network, and effectively solves the problems existing in the single particle filtering-based lithium-ion battery life prediction. Aiming at the shortcomings of the lithium-ion battery life prediction method based on particle filtering, the present invention realizes an iterative prediction algorithm of particle filtering without measurement value update by analyzing and comparing time series prediction strategies, and applies the above fusion prediction algorithm to lithium-ion battery. Battery life is predicted. The invention effectively improves the accuracy of the multi-step estimation of the life of the lithium ion battery.

Description

Lithium-ion battery life prediction method based on particle filter and long-short-term memory network

technical field

The invention relates to the technical field of health prediction and diagnosis of a battery management system, in particular to an indirect prediction method for the remaining life of a power battery.

Background technique

Lithium-ion batteries are the source of primary or auxiliary power in many devices, and they are fast becoming the most commonly used power source for electric vehicles (EVs). Remaining useful life (RUL) of a battery is defined as the useful life left on the battery for a specific operating time. RUL estimation is critical for condition-based maintenance (CBM) and battery health management. It is important to find a reliable and accurate way to monitor Li-ion battery health (SOH) and predict RUL for timely maintenance and replacement of battery systems. At present, the lithium battery life prediction mainly adopts the method based on model modeling and data modeling to predict the capacity attenuation of lithium battery. The current research method only predicts the battery life. However, the change of the internal state of the battery is the end of the battery life. main factor.

Model-based and data-driven prediction methods have their own disadvantages in actual usage scenarios. The present invention comprehensively considers the above two methods, and integrates model-based and data-driven methods. Based on the particle filter algorithm, combined with long and short-term memory network, to achieve a fusion prediction method

SUMMARY OF THE INVENTION

The invention proposes a lithium-ion battery life prediction method based on particle filtering and long-short-term memory network algorithm, so as to realize the battery cycle life prediction.

A lithium-ion battery cycle life prediction method based on particle filtering and long-short-term memory network algorithm, the specific steps are as follows:

Step 1: monitor the lithium-ion battery to be predicted, and obtain the monitoring data of the remaining capacity of the lithium battery;

Step 2: Establish the state space equation of the capacity exponential decay model describing battery aging;

The battery capacity exponential decay model:

Q=a·exp(b·k)+c·exp(d·k) (1)

Where a, b, c and d are the parameters of the battery capacity exponential decay model, a and c represent the initial value of the battery capacity, c and d represent the battery capacity decay rate, k is the number of charge-discharge cycles, and Q is the actual battery capacity at time k capacity.

The state transition equation of the model at time k is:

x _k = [ _ak , b _k , c _k , d _k ] (2)

The model's measurement equation at time k is:

Q _k = _ak ·exp(b _k ·k)+c _k ·exp(d _k ·k)+v _k v _k ~N(0,σ _v ) (4)

Among them, w _a , w _b , w _c , and w _d are process noise, and v _k is measurement noise.

Step 3: Set up the prediction starting point, the number of particles in the particle filter model and the battery life threshold, initialize the parameters a, b, c, d in the exponential decay model of the battery capacity, and the covariance σ _a～d of the process noise w _a ～d and the covariance σ _d of the measurement noise v;

Step 4: Use the particle filter algorithm and the long-short-term memory network model to predict the remaining service life of the battery in one step or multiple steps.

The long-short-term memory network model is a process of predicting z _k+1 , z _{k +2} , . . . , z _k+P by measuring time series z ₁ , z ₂ , . Through regression analysis, the relationship between data and time is established, so as to realize the prediction of future data. For the time series {z ₁ , z ₂ , ..., z _k }, according to the mapping relationship between the input {z _k-1 , z _k-2 , ..., z _km } and y={z _k }, the length and length are obtained. The learning samples of the time memory network are obtained, and then the battery capacity prediction of the next P steps is obtained.

The specific algorithm steps of the particle filter are as follows:

Step 31: Initialization. k=0, sample the particle set according to the known prior probability density p(x ₀ )

采样

更新粒子权值：Step 32: Calculation of importance weights: according to

sampling

Update particle weights:

Normalized importance weights:

得到新的等权样本

Step 33: Resampling. According to the original weight sample

get new equal weight samples

Step 34: State Estimation:

Step 35: Determine whether to end. If so, exit the program; otherwise, set k=k+1, and return to step 32 .

As shown in Figure 1, the specific steps of the single-step prediction of the cycle life of the lithium-ion battery are expressed as:

Step 11: Use the long-short-term memory network model to train and model the battery historical capacity data z _1:k ;

Step 12: Use the trained model to predict the measurement at time k+1

Step 13: Substitute the predicted measurement value into the state space model of the system, use it to update the particle importance weight, and then perform resampling to obtain the predicted state at time k+1

Step 14: Determine whether the battery reaches the capacity threshold according to the state prediction, if so, end the prediction, if otherwise, set k=k+1, and return to step 12.

As shown in Figure 2, the specific steps of the multi-step prediction of the cycle life of the lithium-ion battery are expressed as:

Step 21: Use the long-short-term memory network model to train and model the battery historical capacity data z _1:k ;

Step 22: Use the model obtained by training to predict the measurement value of P steps in the future;

代入到系统的状态空间模型中，用其更新粒子重要性权值，然后进行重采样，获得系统k+1时刻的状态

Step 23: Combine the above predicted measurements

Substitute it into the state space model of the system, use it to update the particle importance weight, and then resample to obtain the state of the system at time k+1

Step 24: Determine whether the capacity threshold is reached according to the prediction state, if so, end the prediction, if otherwise, set k=k+1, and return to step 23 .

The advantages of the present invention are: in view of the deficiencies in the lithium-ion battery life prediction method based on particle filtering, by analyzing and comparing time series prediction strategies, an iterative multi-step prediction algorithm of particle filtering without measurement value update is realized, and the above The fusion prediction algorithm predicts the life of lithium-ion batteries, which effectively improves the accuracy of multi-step prediction.

Description of drawings

Figure 1 is a schematic diagram of the single-step prediction of battery capacity decay based on particle filter and long-short-term memory network algorithm.

Figure 2 is a schematic diagram of multi-step prediction of battery capacity decay based on particle filter and long-short-term memory network algorithm.

Figure 3 shows the single-step prediction results of battery capacity fading based on particle filter and long-short-term memory network algorithm.

Figure 4 shows the multi-step prediction results of battery capacity fading based on a single particle filter algorithm.

Figure 5 shows the multi-step prediction results of battery capacity fading based on particle filter and long-short-term memory network algorithm.

Detailed ways

The method provided by the present invention specifically comprises the following steps:

Step 1: monitor the lithium-ion battery to be predicted, and obtain the monitoring data of the remaining capacity of the lithium battery;

Step 2: Establish the state space equation of the capacity exponential decay model describing battery aging;

Step 3: Set up the prediction starting point, the number of particles in the particle filter model and the battery life threshold, initialize the parameters a, b, c, d in the exponential decay model of the battery capacity, and the covariance σ _a～d of the process noise w _a ～d and the covariance σ _d of the measurement noise v;

Step 4: Use the particle filter algorithm and the long-short-term memory network model to predict the remaining service life of the battery in one step or multiple steps.

Figure 1 shows the schematic diagram of the single-step prediction algorithm for battery capacity decay based on particle filter and long-short-term memory network algorithm.

Figure 2 shows a schematic diagram of a multi-step prediction algorithm for battery capacity fading based on particle filter and long-short-term memory network algorithm.

The effectiveness of the present invention is demonstrated below in conjunction with an example. The test data set comes from the capacity data obtained by the accelerated life test of lithium ion batteries carried out by the Advanced Life Cycle Engineering Research Center of the University of Maryland in the United States. The dataset includes 4 finished 18650 round lithium-ion batteries from the same manufacturer, numbered A1, A2, A3, and A4, respectively. The A4 battery is selected from the test sample, and the threshold for judging whether the battery has reached the service life is Q<0.72.

Figure 3 shows the single-step prediction results of battery capacity fading based on particle filter and long-short-term memory network algorithms. Select the first 39 capacity data of battery A4 as the known data, start one-step prediction from the cycle period k=39, select the number of particles as 200, the predicted failure time of the battery is the 50th cycle, and the actual failure time of the battery is the 47th cycle cycle, and the prediction error is 3 cycles.

Figure 4 shows the multi-step prediction results of battery capacity decay based on the single particle filter algorithm. The first 39 capacity data of battery A4 are selected as known data, and the one-step prediction starts from the cycle period k=39, and the number of particles is selected as 200. The predicted failure time of the battery is the 61st cycle, while the actual failure time of the battery is the 47th cycle, the prediction error is 14 cycles, and the multi-step prediction results are not ideal.

Figure 5 shows the multi-step prediction results of battery capacity decay by particle filter and long-short-term memory network algorithm. Similarly, the first 39 capacity data of battery A4 are selected as known data, and the one-step prediction starts from the cycle period k=39. If 200 is selected, the predicted failure time of the battery is the 50th cycle, while the actual failure time of the battery is the 47th cycle, and the prediction error is 3 cycles. The simulation results show that the above based on particle filtering and long-term memory network The prediction algorithm predicts the life of lithium-ion batteries, which effectively improves the accuracy of multi-step prediction.

Claims (1)

1. The lithium ion battery cycle life prediction method based on the particle filter and the long-time and short-time memory network algorithm is characterized by comprising the following steps of:

the method comprises the following steps: monitoring a lithium ion battery to be predicted to obtain monitoring data of the residual capacity of the lithium ion battery;

step two: establishing a state space equation of a capacity exponential decay model for describing battery aging;

the battery capacity exponential decay model is as follows:

Q＝a·exp(b·k)+c·exp(d·k) (1)

wherein a, b, c and d are parameters of the exponential decay model of the battery capacity, a and c represent initial values of the battery capacity, c and d represent decay rates of the battery capacity, k is the number of charge-discharge cycles, and Q is the actual capacity of the battery at the moment k

The state transition equation of the model at the moment k is as follows:

x_k＝[a_k,b_k,c_k,d_k](2)

the measurement equation of the model at the time k is as follows:

Q_k＝a_k·exp(b_k·k)+c_k·exp(d_k·k)+v_kv_k～N(0,σ_v) (4)

wherein, w_a，w_b，w_c，w_dIs process noise, v_kFor measuring noise

Step three: setting a prediction starting point, the number of particles in the particle filter model and a service life threshold value of the battery, initializing parameters a, b, c and d in the battery capacity exponential decay model and process noise w_a～dOf (2) covariance σ_a～dAnd measuring the covariance σ of the noise v_v；

Step four: performing single-step or multi-step prediction on the remaining service life of the battery by utilizing a particle filter algorithm and a long-time and short-time memory network model;

the long-time memory network model measures the time sequence z by giving a history₁,z₂,...,z_kPredicting z_k+1,z_k+2,...,z_k+PFor a time series z₁，z₂，…，z_kAccording to the input { z }_k-1，z_k-2，…，z_k-mZ and y ═ z_kObtaining learning samples of a long-time and short-time memory network according to the mapping relation between the battery capacity and the battery capacity, and further obtaining battery capacity prediction of a future P step;

the single-step prediction comprises the following specific steps:

step 11: historical capacity data z of battery by utilizing long-time and short-time memory network model_1：kTraining and modeling;

step 12: predicting measurement value at k +1 moment by using trained model

Step 13: and substituting the predicted measured value into a state space model of the system, updating the particle importance weight value by using the predicted measured value, and then resampling to obtain the predicted state at the k +1 moment

Step 14: judging whether the battery reaches a capacity threshold value according to the state prediction, if so, ending the prediction, otherwise, enabling k to be k +1, and returning to the step 12;

the specific steps of the multi-step prediction are as follows:

step 21: historical capacity data z of battery by utilizing long-time and short-time memory network model_1：kTraining and modeling;

step 22: predicting the measured value of the future P step by using the model obtained by training;

step 23: measuring the above predicted values

Substituting the value into a state space model of the system, updating the particle importance weight value by using the value, and then resampling to obtain the predicted state at the k +1 moment

And step 24, judging whether the capacity threshold is reached according to the prediction state, if so, ending the prediction, otherwise, enabling k to be k +1, and returning to the step 23.

Images