CN111680848A - Battery life prediction method and storage medium based on prediction model fusion - Google Patents

Abstract

The invention provides a battery life prediction method and storage medium based on the fusion of prediction models. The battery life prediction method embeds a long-short memory network model in a particle filter model. The fusion model has a simple structure and is trained with existing historical data. The long-short memory network model obtains the degradation trend equation to determine the state transition equation of the particle filter model, which solves the problem that the particle filter model relies too much on the empirical model. In addition, the new samples obtained online are added to the original training sample set to retrain the model, so that the model parameters are updated in time and have better adaptability, which can realize the prediction of the remaining cycle life of the nickel-cadmium battery.

Description

Battery life prediction method and storage medium based on prediction model fusion

technical field

The invention belongs to the technical field of battery life prediction, and in particular relates to a battery life prediction method and storage medium based on prediction model fusion.

Background technique

Regardless of whether it is an electric locomotive or a diesel locomotive, the battery and the charger are connected in parallel to form the energy source of the locomotive control circuit. Once the battery fails, it is impossible to maintain the normal use of the interior lighting, wireless communication devices and emergency devices, and the safety of passengers' lives and property. will bring a great threat. Through investigation, it is found that most of the batteries used in high-speed railway vehicles are alkaline nickel-cadmium batteries, which are generally replaced according to the number of kilometers or years of use in practical applications. At this time, the battery life often has a large margin, and the early replacement will undoubtedly increase the operating cost of the EMU. Therefore, it is urgent to study accurate and reliable life prediction models. At present, battery life prediction methods are roughly divided into two categories: model-driven and data-driven.

The model-driven method establishes a life prediction model based on the internal structure principle and degradation mechanism of the battery. The model-driven method applies battery tomography measurement technology and electrochemical performance measurement technology as disclosed in the prior art, and builds a power battery cycle life prediction model according to the internal structure of the lithium battery, but it is affected by factors such as battery type and model. difficult to apply in practice. The model-driven method disclosed in the prior art 2 is a degradation model, which uses extended Kalman filtering to estimate the health and remaining life of a fuel cell (PEMFC) online, and the model is robust to operating conditions. The model-driven method provided by the prior art is a model based on a new particle filter (PF) framework, which uses the current measurement value to resample the state particles, which can prevent the degeneracy of the particles, and can also adaptively adjust the number of particles. for online applications. The experimental results show that compared with other standard models, the model can obtain more accurate prediction results in a shorter time. Although the prediction performance of the model-driven method is getting better and better, the model-driven method is too dependent on the failure mechanism, and the accuracy of the prediction depends to a large extent on the state model used, and the battery working environment factors are complex and changeable, so an accurate degradation model can be established. more difficult.

The data-driven method obtains the battery performance degradation law by mining and analyzing the failure data, and then predicts the battery life. Data-driven methods such as: prior art 4 proposes a real-time remaining service life RUL estimation method based on support vector machine (SVM), analyzes the cycle data of lithium batteries under different working conditions, extracts key features from voltage and temperature curves, and uses These features train the model, so as to achieve the purpose of RUL prediction of lithium-ion batteries; the prior art five combines the parameters of the equivalent circuit model with the aging process data, and uses the correlation vector machine (RVM) to improve the PF prognosis framework, which further improves the prediction. The accuracy of the prediction is reduced, and the uncertainty of prediction is reduced; the existing technology 6 uses the elastic mean square backpropagation method to adaptively optimize the long short-term memory network (LSTM) for life prediction. Recurrent Neural Networks predict results more accurately. The data-driven model based on neural network has relatively better performance in the existing technology, but although neural network has good learning ability for historical data, the network structure is difficult to determine, and the sample size and quality of data are very high. And does not have the uncertainty expression of the output.

In addition, the battery life research in the prior art is mainly aimed at lithium batteries and fuel cells, while the life test of nickel-cadmium batteries is time-consuming and the test conditions are harsh, so there is no relevant life research at present. The cycle life of batteries used in existing research institutes is less than 1,000 times, while the life cycle of nickel-cadmium batteries for a certain type of EMU is as high as 2,000 times, and the battery capacity will only drop below the standard. With the increase of the number of cycles, the offline method cannot update the model, the errors accumulate, and it is difficult to obtain better accuracy, while the online prediction model can update the model with the update of the data, and the prediction accuracy of the model will be higher. In addition, nickel-cadmium batteries have the characteristics of "memory effect", and it is difficult to obtain better prediction results with general prediction methods.

SUMMARY OF THE INVENTION

In view of this, the present invention provides a battery life prediction method and storage medium based on prediction model fusion, so as to solve the problem of establishing an accurate battery life due to excessive dependence on the failure mechanism when predicting the battery life based on a single particle filter model. It is difficult to degenerate the model and solve the uncertainty of the output of the prediction model based on a single neural network. At the same time, it can also solve the technical problem that the existing technology cannot update the model online.

A battery life prediction method based on the fusion of prediction models, the prediction model includes a particle filter model and a long-short memory network model, and the battery life prediction method includes:

Step 1: Get historical data of the battery,

Step 2: Predict the capacity state of the battery according to the historical data through the particle filter model to obtain the predicted capacity value of the battery,

Step 3: Compare the predicted capacity value with the set capacity failure threshold of the battery, and when the predicted capacity value reaches the capacity failure threshold, determine that the battery has failed, and make the particle filter The iteration of the model ends, and the remaining life of the battery is obtained according to the number of iterations,

The long-short memory network model is embedded in the particle filter model, and the long-short-term memory network model uses the battery capacity state value data in the historical data as the number of training samples for training and learning, so that the long-short-term memory network model is trained and learned. The model predicts the capacity state of the battery according to the historical data of the battery capacity, and constructs the state transition equation of the particle filter model according to the output value of the trained long-short memory network model and obtains the capacity state of the battery a priori predicted value.

Preferably, the historical data includes data on the number of times of charging and discharging of the battery and data on the capacity state value of the battery corresponding to each time the charging and discharging are performed,

The battery life prediction method further includes:

Before training and learning the long-short memory network model through the battery capacity data, pre-processing the historical data, the pre-processing includes removing useless data in the historical data and normalizing the historical data processing.

Preferably, the step 2 includes:

Step 21: A set of randomly generated capacity values of the battery at the predicted starting point is used as the initial particle group of the particle filter model, and an initial weight coefficient is assigned to each of the initial particle groups,

Step 22: Use the training sample data to train and learn the long-short-term memory network model, so that the long-short-term memory network model predicts the capacity state of the battery according to the battery capacity historical data,

Step 23: constructing the state equation according to the output of the long-short memory network model and obtaining the a priori predicted value at the current moment,

Step 24: Generate a new particle group according to the prior value at the current moment and the particle group at the previous moment,

Step 25: Update the weight of the new particle group,

Step 26: Modify the prior prediction value at the current moment according to the new particle group and the weight of the new particle group to obtain the a posteriori prediction value at the current moment.

Preferably, the battery life prediction method further includes step 4, step 5 and step 6. If the judgment result in step 3 is that the battery has not failed, then step 4 and step 5 are executed in sequence,

The step 4 is to obtain the measured value of the battery capacity online,

The step 5 is to judge whether the difference between the predicted value at the current moment and the new measurement value is greater than the set error value. If the judgment result is yes, then step 6 is executed; otherwise, the Increase the number of iterations of the particle filter model by 1, and return to step 23,

Step 6 is to add the measured value obtained in step 4 to the training sample to update the training sample, and after updating the training sample, go to step 22 to retrain the training sample through the updated training sample. The long short-term memory network model.

Preferably, overfitting of the long short-term memory network model is prevented by setting a dropout module in the long-short-term memory network model.

Preferably, the capacity state value data in the historical data is

A time series is established, and the long short-term memory network model is trained with the time series.

Preferably, the step of determining the state equation and the a priori predicted value according to the trained long-short-term memory network includes:

Make the trained long-short memory network model predict the state value of the battery capacity at the current moment according to the capacity state value data at the first m moments of the current moment,

The output value of the long-short memory network model at the current moment and the process noise of battery capacity degradation at the previous moment at the current moment are superimposed as the state prediction value at the current moment in the state equation to construct the state equation,

The a priori predicted value of the current moment is obtained by calculating according to the state equation and the a posteriori predicted value of the previous moment at the current moment.

Preferably, importance sampling is used to update the weight of the new particle group, so that the particle that is closer to the predicted value of the battery capacity state corresponds to a larger weight coefficient.

Preferably, the battery is a nickel-cadmium battery, and the long-short memory network model includes an input and output layer, a long-short memory network layer, a dropout layer and a fully connected layer.

A storage medium, characterized in that the storage medium is a computer-readable storage medium, and when a computer program stored on the readable storage medium is executed by a processor, the method for predicting battery life as described in any of the above is implemented .

Beneficial effects obtained by the present invention: the online prediction method of battery remaining life based on the fusion of particle filter model and long-short memory network model provided by the present invention can realize the remaining cycle life prediction of nickel-cadmium battery, and the long-short memory network model is nested in the particle filter model. The structure of the fusion model is simple. The long-short-term memory network model is trained with existing historical data to obtain the degradation trend equation to determine the state transition equation of the particle filter model, which solves the problem that the particle filter model relies too much on the empirical model. The particle filter model uses the weighting of particles. In addition, the new samples obtained online are added to the original training sample set to retrain the model, so that the model parameters are updated in a timely manner and have better adaptability.

Description of drawings

1 is a schematic flowchart of a method for predicting battery life based on prediction model fusion provided according to an embodiment of the present invention;

2 is a comparison chart of the prediction effect of battery life prediction according to the fusion prediction model LSTM-P model provided by the present invention under the setting condition that the prediction starting point is T=1100cycle and the number of cycles is RUL=1742cycle;

Figure 3 is a comparison chart of the prediction effect of battery life using the standard PF model under the setting conditions that the prediction starting point is T=1100cycle and the number of cycles is RUL=1742cycle;

Figure 4 is a comparison chart of the prediction effect of battery life prediction using the standard LSTM model under the setting conditions that the prediction starting point is T=1100cycle and the number of cycles is RUL=1742cycle;

5 is a comparison chart of the prediction effect of battery life prediction according to the fusion prediction model LSTM-P model provided by the present invention under the setting condition that the prediction starting point is T=200cycle and the number of cycles is RUL=842cycle;

Figure 6 is a comparison chart of the prediction effect of battery life using the standard PF model under the setting conditions that the prediction starting point is T=200cycle and the number of cycles is RUL=842cycle;

Figure 7 is a comparison chart of the prediction effect of battery life prediction using the standard LSTM model under the setting conditions that the prediction starting point is T=200cycle and the number of cycles is RUL=842cycle.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments generated by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention. In addition, it should be noted that, in the content of the specific embodiments, "the..." refers only to the technical attributes or features of the present invention.

In order to solve the above problems existing in the prior art, the present invention provides a battery life prediction method based on the fusion of prediction models, which mainly fuses two prediction models, the particle filter (PF) model and the long short-term memory network (LSTM) model. , that is, the two prediction models are fused by embedding the long-short memory network model in the particle filter model, thereby forming a fusion prediction model for predicting battery life. In the fusion prediction model, the long-short-term memory network model uses the battery capacity state value data in the historical data as the number of training samples for training and learning, so that the long-short-term memory network model performs training and learning based on the battery capacity historical data. The capacity state of the battery is predicted, and a state transition equation of the particle filter model is constructed according to the output value of the trained long-short memory network model, and a priori prediction value of the capacity state of the battery is obtained.

Specifically, the execution steps of the battery life prediction method based on the fusion prediction model provided by the present invention include:

Step 1: Get historical data of the battery. The historical data includes data on the number of times of charging and discharging of the battery and data on the capacity state value of the battery corresponding to each time the charging and discharging are performed.

Step 2: Predict the capacity state of the battery according to the historical data through the particle filter model, so as to obtain a capacity prediction value of the battery.

The step of obtaining the predicted capacitance value includes:

Step 21: Randomly generate a set of capacity values of the battery at the predicted starting point as an initial particle group of the particle filter model, and assign an initial weight coefficient to each of the initial particle groups.

Step 22 : using the training sample data to train and learn the long-short-term memory network model, so that the long-short-term memory network model predicts the capacity state of the battery according to the battery capacity historical data.

Step 23: Construct the state equation according to the output of the long short-term memory network model and obtain the prior prediction value at the current moment.

Step 24: Generate a new particle group according to the prior value at the current moment and the particle group at the previous moment.

Step 25: Update the weight of the new particle group.

Step 26: Modify the prior prediction value at the current moment according to the new particle group and the weight of the new particle group to obtain the a posteriori prediction value at the current moment.

Step 3: Compare the predicted capacity value with the set capacity failure threshold of the battery, and when the predicted capacity value reaches the capacity failure threshold, determine that the battery has failed, and make the particle filter The iteration of the model ends, and the remaining life of the battery is obtained according to the number of iterations.

The battery in the battery life prediction method provided by the present invention is mainly a cadmium-nickel battery. Because the number of charge and discharge cycles is as high as 2000 or more, the battery may fail, which is much higher than the number of charge and discharge cycles of a commonly used lithium battery. Therefore, the prediction model for the battery life of the nickel-cadmium battery needs to have new data to update the fusion prediction model online, so as to update the fusion prediction model adaptively, so that the prediction accuracy of the fusion prediction model will be higher.

Therefore, according to the battery life prediction method provided by the present invention, it further includes steps 4, 5 and 6. If the judgment result in the step 3 is that the battery has not failed, then the steps 4 and 6 are executed in sequence. Step 5,

The step 4 is to obtain the new state value and new measurement value of the battery capacity online. Specifically, the new state value and the new measurement value (observed value) generated in the online acquisition of the life test experiment can be added to the original value. Provides historical battery data in the established time series. The step 5 is to judge whether the difference between the predicted value at the current moment and the new measurement value is greater than the set error value. If the judgment result is yes, then step 6 is executed; otherwise, the The number of iterations of the particle filter model is incremented by 1, and the process returns to step 23. The step 6 is to add the new state value obtained in the step 4 to the training sample to update the training sample. After updating the training sample, go to step 22 to pass the updated training sample. Retrain the long short-term memory network model.

In addition, we also prevent overfitting of the long short memory network model by setting a dropout module in the long short memory network model. Therefore, the long-short-term memory network model provided according to the embodiment of the present invention includes an input and output layer, a long-short-term memory network layer, a dropout layer, and a fully connected layer.

Before further explaining the present invention in detail through specific embodiments, we first describe the related principles of the particle filter model and the long-short-term memory network model.

The algorithm of particle filtering is an algorithm that introduces Monte Carlo sampling to obtain estimates of posterior probability and random samples based on Bayesian filtering. Assuming a system (such as the battery system of the present invention), its state equation and observation equation are shown in formula (1) and formula (2):

x _k =f _k (x _k-1 , v _k-1 ) (1)

Y _k =h _k (x _k , n _k ) (2)

Where x _k , Y _k are the system state and observation value at time k, v _k-1 is the process noise (dynamic noise) of the system at time k-1, n _k is the observation noise at time k, and x _k-1 is k System state at time -1. In the application of battery life prediction, equation (1) is usually an empirical degradation equation, and it is difficult to obtain an accurate degradation equation in practical engineering. In order to obtain the optimal estimation of the target state, the particle filter obtains the posterior probability density p(x _k |Y _k ) of the system at time k through two processes of prediction and update. In the prediction stage, the formula for obtaining the prior probability p(x _k |Y _k-1 ) using the probability density p(x _k-1 |Y _k-1 ) at time k-1 is shown in formula (3):

p(x _k |Y _k-1 )=∫p(x _k |x _k-1 )p(x _k-1 |Y _k-1 )dx _k-1 (3)

The update stage uses the importance sampling method to introduce the importance probability density function q(x _k |Y _k ), generates sampled particles from it, and uses the weighted sum of the particles to approximate the posterior probability distribution p(x _k |Y _k ) to obtain the posterior The formula for calculating the test probability is formula (4)

为k时刻第i个粒子的状态，其权值为

权值的分配公式如公式(5)所示：in

is the state of the ith particle at time k, and its weight is

The weight distribution formula is shown in formula (5):

Recurrent neural network (RNN) can use its memory function to deal with nonlinear time series, but when the sequence is very long, it is prone to the problem of gradient explosion and gradient disappearance. Long short-term memory network (LSTM) is designed to solve this problem. A special kind of RNN. Compared with RNN, LSTM adds an information processing unit, that is, a cell, which consists of a forget gate, an input gate, and an output gate.

The forget gate can discard the redundant information of the upper layer with a certain probability, and its calculation formula is shown in formula (6):

f ^(t) =σ(W _f h ^(t-1) +U _f x ^(t) +b _f ) (6)

Where h ^(t-1) is the hidden state of the previous layer, x ^(t) is the current sequence position information, W _f , U _f , b _f are the weights and offsets of the linear relationship in the forget gate, and σ is the sigmoid activation function. The gate will output a value between 0 and 1 that determines how much information is lost, with 0 being "completely discarded" and 1 being "completely preserved".

The input gate can process the information of the current sequence position, and its calculation formula is shown in formula (7) and formula (8):

i ^(t) =σ(W _i h ^(t-1) +U _i x ^(t) +b _i ) (7)

a ^(t) = tanh(W _a h ^(t-1) +U _a x ^(t) +b _a ) (8)

Among them, Wi, U _i , _Wa , and U _a are the _weights of the linear relationship in the input gate, and _bi and _ba are the offsets. The results of the forget gate and the input gate will be used to update the cell state, and the update equation is shown in formula (9):

C ^(t) = C ^(t-1) ⊙f ^(t) +i ^(t) ⊙a ^(t) (9)

where C ^(t) is the updated cell state and ⊙ is the Hadamard product.

The output gate can process the information of the current sequence, the cell state and the hidden state of the upper layer, and output a new hidden state to the next layer. The calculation formulas corresponding to the output gate are shown in formulas (10) and (11):

o ^(t) = σ(W _o h ^(t-1) +U _o x ^(t) +b _o ) (10)

h ^(t) = o ^(t) ⊙tanh(C ^(t) ) (11)

Among them, W _o , U _o , and _bo are the weight and offset of the linear relationship in the output gate, and h ^(t) is the hidden state of the current layer, which is used as the output of the current layer and continues to be passed to the next layer.

In view of the good learning ability of long short-term memory network (LSTM), particle filter (PF) can well adapt to nonlinear and non-Gaussian systems, and can give uncertainty expression, so this paper proposes a fusion of LSTM and PF two algorithms With reference to the schematic flowchart of the method for predicting battery life based on the fusion prediction model shown in FIG. 1 , it is further explained in detail how the present invention integrates the particle filter model and the long-short memory network model to predict the of battery life.

As shown in Figure 1, during the use of the nickel-cadmium battery, the available capacity of the battery decreases due to the deactivation of the active material and the reduction of the electrolyte. TB_T3061-2016 stipulates that the capacity value is used as the basis for judging the failure. When the capacity is reduced to the rated capacity 70%, it is invalid. Therefore, the battery capacity is usually used as a performance degradation factor, and the lifespan is predicted according to the evolution law of the degradation factor. The battery life is affected by various factors such as temperature, charge-discharge rate, and working conditions, and the failure process is nonlinear and non-Gaussian. Particle filtering can be well applied to non-Gaussian nonlinear systems, and can obtain the inaccurate expression of prediction results, but standard particle filtering requires the state transition equation shown in formula (1). It is difficult to obtain a more accurate state equation. LSTM has a memory function and can learn time series with a long time span, but it cannot adapt to uncertain factors such as noise in the system, and cannot give uncertain expressions. Therefore, we integrate two prediction models and combine their advantages to better achieve Battery life prediction.

This paper selects the capacity as the degradation factor, and establishes the battery capacity data in the early stage (such as after 1000 battery charge-discharge cycles) as a time series (x ₁ , x ₂ , x ₃ ,...x _n ), through the established The time sequence sequence trains the LSTM model, so that the LSTM model can obtain the predicted value of the k time based on the information of the first m moments at the k time (the current time), and the long-short memory network model performs the capacitance of the k time. The formula of the capacity state prediction method is as formula (12)

为第k时刻LSTM在k时刻的输出，即通过LSTM模型预测的k时刻所述电池的容量状态，根据LSTM训练得到的容量退化模型式(12)来确定粒子滤波的状态转移方程：

is the output of the LSTM at time k at time k, that is, the capacity state of the battery at time k predicted by the LSTM model, and the state transition equation of the particle filter is determined according to the capacity degradation model formula (12) obtained by LSTM training:

根据状态转移方程的先验概率得到一组新的粒子的计算公式如公式(14)所示：ω _k-1 is the process noise, and x _k is the state prediction value in the particle filter model at the kth moment. The randomly generated capacity value at time k is used as the initialization particle

According to the prior probability of the state transition equation, the calculation formula of a new set of particles is obtained as shown in formula (14):

Use importance sampling to optimize the weight of new particles, the closer the particle is to the state prediction value x _k , the greater the weight, and the weighted particle sum is used to approximate the capacity prediction value at time k. The newly added time series is used to update the LSTM model parameters. The specific process is shown in Figure 1.

Step a: Preprocess the capacity data, remove the unavailable data, and perform normalization. This step is equivalent to the above step 1.

作为初始粒子，例如随机生成第k时刻电池的一组容量值作为所述初始粒子。Step b: Randomly generate N particles

As the initial particle, for example, a set of capacity values of the battery at the k-th time is randomly generated as the initial particle.

Step c: The processed data is used as a training sample to be put into the LSTM model to learn and train the LSTM model, and a dropout module is added to the LSTM to prevent overfitting. The LSTM model we choose in the present invention includes an input-output layer, an LSTM layer, and the dropout layer.

所述先验预测值与先验概率对应可以根据所述先验概率获得。Step d: use the capacitance degradation equation determined by the output of the trained LSTM model as the state equation in the particle filter model, as shown in formula (13), and obtain a priori predicted value of the observed value according to the state equation

The a priori predicted value corresponds to the prior probability and can be obtained according to the prior probability.

Step e: Generate a new particle group state transition equation according to the a priori predicted value at the current moment and the particle group at the previous moment to generate new particles

所述后验预测值为所述粒子滤波模型对所述电池容量的后验值。Step f: Update the particle weight, and verify the predicted value after getting it

The a posteriori predicted value is a posteriori value of the particle filter model for the battery capacity.

Step g: Compare the predicted battery capacity with 70% of the rated capacity. If the former is less than the latter, it is determined that the battery has failed, the prediction is over, and the remaining life is obtained, otherwise, go to step h.

与y_new的差值是否超出设定的误差范围PEB，若未超出范围，则LSTM模型无需更新，宽度为m的滑动窗口向前移动一步，继续进行预测；否则将新增时间序列加入到训练集中，转到步骤c重新训练LSTM模型参数，用重新训练的模型进行后续预测。Step h: The newly added time series (x _new , y _new ) are obtained online; x _new is the new state value of the battery currently obtained online, and the y _new is the current observed value of the battery capacity, and the posterior prediction value is judged

Whether the difference with y _new exceeds the set error range PEB, if it does not exceed the range, the LSTM model does not need to be updated, and the sliding window with a width of m moves one step forward to continue the prediction; otherwise, a new time series will be added to the training. To concentrate, go to step c to retrain the LSTM model parameters and use the retrained model for subsequent predictions.

In order to study the aging characteristics of nickel-cadmium batteries, we used several groups of exhaust-type nickel-cadmium batteries for EMU vehicles of the same type. The nominal voltage of the single battery is 1.2V, the rated capacity is 160A·h, and the high and low temperature test box is used for the maintenance test. Ambient temperature, battery pack test system is used to monitor parameters such as current and voltage.

According to the iron standard TB_T3061-2016, the cycle life test is carried out under the environment of 25℃±5℃, with 50 cycles as a group, the first cycle in each group is charged at 0.25I _t for 6 hours, and discharged at 0.25I _t for 2.5 hours h, 2~50 cycles of charging at 0.2I _t for 7h~8h, and discharging to 1.0V/cell at 0.2I _t , until the discharge time of any 50 cycles is less than 3.5h, and then carry out one more set at 0.2I _t Cycle, if the discharge time of the 50th cycle of two consecutive groups is less than 3.5h, indicating that the capacity drops below 70% of the rated capacity, the life test is terminated.

The capacity is calculated according to the ampere-hour integration theorem, and the capacity is used as the battery performance degradation characteristic. The formula of the time-integration theorem is shown in formula (15):

C _k is the capacity of the k-th charge-discharge cycle, I is the discharge current, and the time series of the capacity is obtained, and the data is preprocessed as shown in formula (16) using the normalization function:

The fitness evaluation function of the battery life prediction model is defined as formula (17):

为预测容量值。where n is the total number of predicted data points, C _k is the actual capacity value,

is the predicted capacity value.

In order to verify the prediction effect of the proposed method, the standard particle filter is used to predict the experimental data as a comparison. The state transition equation uses an exponential model, as shown in Equation (18):

C _k = η _c C _k-1 +β ₁ exp(-β ₂ /Δt _k-1 ) (18)

Where η _c is the Coulomb efficiency, generally taken as 0.998, Δt _k-1 =t _k -t _k-1 , which is the time interval between two adjacent cycles, and the remaining parameters are obtained by fitting the experimental data.

It was obtained from the experiment that the nickel-cadmium battery showed a low capacity phenomenon in the early stage due to its unique "memory effect". After several thorough charge-discharge cycles, the capacity returned to the rated value, and it failed in the 2842nd cycle. The capacity threshold for battery failure is 112A·h. T=1100cycle and T=2000cycle are used as the starting point of prediction, the experimental data before the starting point of prediction is used as the training set, and the data after the starting point of prediction is used as the test set. The LSTM model structure is an input and output layer, an LSTM layer, a dropout layer, and a fully connected layer, and the optimizer uses adam. Number of particles N=300, observation noise covariance Q=0.0001

Table 1 T=1100cycle, experimental results

Figures 2 to 4 are respectively a comparison chart of the prediction effect of the fusion prediction model LSTM-PF, the standard PF prediction model and the LSTM prediction model provided by the present invention under the setting conditions of the prediction starting point T=1100cycle and the actual RUL=1742cycle,

Table 1 shows the results evaluation of the three models, mainly including the prediction results, error and fit. According to the fit data, it can be seen that the error of the fusion model is smaller, and the prediction error is 27 cycles less than PF and 18 cycles less than LSTM.

Table 2 T=2000cycle, RUL predicted value

Figures 5-7 are the prediction comparison charts of the fusion model LSTM-PF, standard PF and LSTM model under the setting conditions of the prediction starting point T=2000cycle and the actual RUL=842cycle. Table 2 shows the result evaluation of the three models , the prediction error of the fusion model is 9 cycles less than that of PF and 5 cycles less than that of LSTM, which has a higher degree of fit.

The experimental results show that when starting from the same starting point, the fusion model has more accurate prediction results than the standard PF and LSTM models, and for all three models, when T = 2000 cycles, the prediction results are better than those when T = 1100 cycles. The effect is better, the later the starting point is, which means that more data can be used to train the model, and the model is more accurate. For the same model, as the observation data is updated, the prediction model continuously learns and updates the parameters, and the online prediction results are more accurate.

It can be seen from the above that the online battery remaining life prediction method based on the fusion of the particle filter model and the long-short memory network model provided by the present invention can realize the remaining cycle life prediction of the nickel-cadmium battery. The long-short memory network model is embedded in the particle filter model, and the fusion model The structure is simple, using the existing historical data to train the long-short memory network model to obtain the degradation trend equation to determine the state transition equation of the particle filter model, which solves the problem that the particle filter model relies too much on the empirical model. The predicted value can get the uncertain expression of the remaining life. In addition, the new samples obtained online are added to the original training sample set to retrain the model, so that the model parameters can be updated in a timely manner and have better adaptability.

In addition, the present invention also provides a storage medium, where the storage medium is a computer-readable storage medium, and when a computer program stored on the readable storage medium is executed by a processor, the computer program according to any embodiment of the present invention is implemented. Battery life prediction method.

Embodiments in accordance with the present invention are described above, but these embodiments do not exhaust all the details and do not limit the invention to only the specific embodiments described. Numerous modifications and variations are possible in light of the above description. This specification selects and specifically describes these embodiments in order to better explain the principle and practical application of the present invention, so that those skilled in the art can make good use of the present invention and modifications based on the present invention. The present invention is to be limited only by the claims and their full scope and equivalents.

Claims (10)

1. a battery life prediction method based on prediction model fusion, is characterized in that, described prediction model comprises particle filter model and long-short memory network model, and described battery life prediction method comprises: Step 1: Get historical data of the battery, Step 2: Predict the capacity state of the battery according to the historical data through the particle filter model to obtain the predicted capacity value of the battery, Step 3: Compare the predicted capacity value with the set capacity failure threshold of the battery, and when the predicted capacity value reaches the capacity failure threshold, determine that the battery has failed, and make the particle filter The iteration of the model ends, and the remaining life of the battery is obtained according to the number of iterations, The long-short memory network model is embedded in the particle filter model, and the long-short-term memory network model uses the battery capacity state value data in the historical data as the number of training samples for training and learning, so that the long-short-term memory network model is trained and learned. The model predicts the capacity state of the battery according to the historical data of the battery capacity, and constructs the state transition equation of the particle filter model according to the output value of the trained long-short memory network model and obtains the capacity state of the battery a priori predicted value. 2 . The battery life prediction method according to claim 1 , wherein the historical data includes charge and discharge times data of the battery and capacity state value data of the battery corresponding to each time the charge and discharge are performed, 2 . The battery life prediction method further includes: Before training and learning the long-short memory network model through the battery capacity data, pre-processing the historical data, the pre-processing includes removing useless data in the historical data and normalizing the historical data processing. 3. The battery life prediction method according to claim 1, wherein the step 2 comprises: Step 21: A set of randomly generated capacity values of the battery at the predicted starting point is used as the initial particle group of the particle filter model, and an initial weight coefficient is assigned to each of the initial particle groups, Step 22: Use the training sample data to train and learn the long-short-term memory network model, so that the long-short-term memory network model predicts the capacity state of the battery according to the battery capacity historical data, Step 23: constructing the state equation according to the output of the long-short memory network model and obtaining the a priori predicted value at the current moment, Step 24: Generate a new particle group according to the prior value at the current moment and the particle group at the previous moment, Step 25: Update the weight of the new particle group, Step 26: Modify the prior prediction value at the current moment according to the new particle group and the weight of the new particle group to obtain the a posteriori prediction value at the current moment. 4. The battery life prediction method according to claim 3, further comprising step 4, step 5 and step 6, if the judgment result in step 3 is that the battery has not failed, then execute the Step 4 and Step 5, The step 4 is to obtain the current measurement value of the battery capacity online, The step 5 is to judge whether the difference between the predicted value at the current moment and the new measurement value is greater than the set error value. If the judgment result is yes, then step 6 is executed; otherwise, the Increase the number of iterations of the particle filter model by 1, and return to step 23, Step 6 is to update the training sample by adding the measured value obtained in step 4 to the training sample. After updating the training sample, go to step 22 to retrain the training sample with the updated training sample. Long Short-Term Memory Network Model. 5 . The battery life prediction method according to claim 1 , wherein a dropout module is set in the long and short memory network model to prevent overfitting of the long and short memory network model. 6 . 6. The battery life prediction method according to claim 2, the capacity state value data in the historical data is A time series is established, and the long short-term memory network model is trained with the time series. 7. The battery life prediction method according to claim 6, wherein the step of determining the equation of state and the a priori predicted value according to the trained long short-term memory network comprises: Make the trained long-short memory network model predict the state value of the battery capacity at the current moment according to the capacity state value data at the first m moments of the current moment, The output value of the long-short memory network model at the current moment and the process noise of battery capacity degradation at the previous moment at the current moment are superimposed as the state prediction value at the current moment in the state equation to construct the state equation, The a priori predicted value of the current moment is obtained by calculating according to the state equation and the a posteriori predicted value of the previous moment at the current moment. 8 . The battery life prediction method according to claim 3 , wherein the weight of the new particle group is updated by using importance sampling, so that the weight coefficient corresponding to the particle that is closer to the predicted value of the battery capacity state is larger. 9 . . 9 . The battery life prediction method according to claim 1 , wherein the battery is a nickel-cadmium battery, and the long-short memory network model comprises an input and output layer, a long-short memory network layer, a dropout layer and a fully connected layer. 10 . 10. A storage medium, wherein the storage medium is a computer-readable storage medium, and the computer program stored on the readable storage medium is executed by a processor to realize any one of claims 1 to 9 The described battery life prediction method.

Images