TWI794787B - Method for predicting life of battery online - Google Patents

Abstract

An online battery life prediction method is provided and includes the following steps. The health status of the battery is detected to obtain historical data. A backtracking training point and a prediction starting point are set, and the historical data between the backtracking training point and the predicted starting point is used as the training set. The training set is input to the machine learning model to obtain multiple prediction parameters at multiple time points in the future after the prediction starting point. The prediction parameters are input to the deep learning model, thereby calculating the residual life.

Description

Online Battery Life Prediction Method

The present invention relates to a service life prediction method, and in particular to an online (Online) battery life prediction method.

At present, the application products of fuel cells include stationary generators, transportation power systems, and portable power systems. After the fuel cell is used, its capacity will gradually decrease due to charging and discharging, so it needs to be replaced in a timely manner. Failure to know in advance when the exact timing of fuel cell replacement will expose consumers to the risk of loss. Therefore, it is necessary to predict the remaining life of the battery through prediction methods, so that consumers can take preventive maintenance actions in advance.

At present, the prediction methods for the service life of fuel cells on the market are mainly to analyze the fitness of the model. However, this method is only applicable to known data, that is, it needs to use the dependent variables of known future data to make predictions, and judge whether the model is suitable for predicting the service life of the battery. At present, the existing forecasting methods need to wait for all future data to be collected before starting to forecast. Therefore, these methods are not suitable for the real world, but only for model research or for products with exactly the same life cycle.

Existing forecasting methods all use single forecasting for predictive analysis. The so-called single forecasting refers to the use of only a single model for forecasting analysis. The future dependent variable is known, so there is no need to predict the dependent variable. However, existing techniques using known future dependent variables have the following problems. First, the high accuracy of the model using known data does not mean that the model is really suitable. Second, it is necessary to wait until all future data are collected before making predictions, which does not conform to the logic of true service life prediction. In practice, it is impossible to obtain future data, and the real known data should only be time data.

The invention provides an online battery life prediction method, which combines a deep learning model and a machine learning model to predict the failure point of the battery life, so as to provide enterprises with early implementation of preventive and maintenance measures before the battery life fails.

The online battery life prediction method of the present invention includes: detecting the state of health of the battery to obtain historical data; setting the retrospective training point and the starting point of prediction, and using the historical data between the retrospective training point and the starting point of prediction as the training set; inputting the training set into the machine learning model to obtain a plurality of prediction parameters at multiple time points in the future after the starting point of prediction; and inputting the prediction parameters into the deep learning model, thereby calculating the remaining life .

In an embodiment of the present invention, the step of inputting the prediction parameters into the deep learning model to calculate the remaining life includes: inputting the prediction parameters into the deep learning model to obtain multiple energy prediction values; judge whether the energy prediction value is less than the threshold value; take one of the energy prediction values that is less than the threshold value and closest to the threshold value among the energy prediction values, and use its corresponding time point as the final life time; and finally The difference between the lifetime time and the predicted onset point is taken as the residual lifetime.

In an embodiment of the present invention, the above online battery life prediction method further includes: setting the number of prediction steps as N, and setting each step as 0 hours; and setting N future time points after the starting point of prediction. After judging whether the energy prediction value is less than the threshold, it further includes: if none of the energy prediction values is less than the threshold, resetting the future N time points after N×O hours from the prediction starting point, so as to use machine learning model with a deep learning model to recalculate the residual life.

In an embodiment of the present invention, after obtaining the prediction parameters, the method further includes: returning the prediction parameters to the training set.

In an embodiment of the present invention, the above-mentioned machine learning model adopts one of support vector regression algorithm, regression algorithm and random forest algorithm, and the deep learning algorithm adopts convolutional neural network and sparse autoencoder (Sparse autoencoder) ) combined with one of the deep neural networks.

In an embodiment of the present invention, the prediction parameters include a predicted voltage value and a predicted current value at each time point.

In an embodiment of the present invention, the above battery life prediction method further includes: obtaining a prediction interval through a Monte Carlo (MC) dropout approach.

Based on the above, this disclosure proposes an online battery life prediction method, which can use deep learning models and machine learning models, combined with a rolling multi-step prediction method, to predict the failure point of battery life, thereby providing enterprises with the premise of battery life failure. Implement preventive maintenance early.

When a completely unused battery is used for a period of time, the battery will degrade and other conditions, and the data collected from the battery and the state of the parameters at the time of weakening cannot directly provide consumers with help. Therefore, it is necessary to use a The prediction method is used to predict the remaining life of the battery, so that consumers can take preventive maintenance actions in advance. Accordingly, the present disclosure provides an online battery life prediction method, which can realize online life prediction.

In the following embodiments, the online battery life prediction method can be implemented by an electronic device with a computing function. The electronic device includes a processor and a memory. A processor can be hardware (such as chipsets, processors, etc.), software components (such as operating systems, application programs, etc.), or a combination of hardware and software components with computing capabilities. The image processor 110 is, for example, a central processing unit (Central processing unit, CPU), a graphics processing unit (Graphics processing unit, GPU), or other programmable microprocessor (Microprocessor), digital signal processor (Digital signal processor) , DSP), programmable controller, application specific integrated circuits (Application specific integrated circuits, ASIC), programmable logic device (Programmable logic device, PLD) or other similar devices.

The storage is, for example, any type of fixed or removable random access memory, read only memory, flash memory, secure digital card, hard disk, or other similar devices or combinations thereof. At least one program code segment is stored in the memory, and the above program code segment is executed by the processor to realize the online battery life prediction method after being installed. The "battery" mentioned in the following examples includes fuel cells.

FIG. 1 is a flowchart of an online battery life prediction method according to an embodiment of the present invention. Referring to FIG. 1 , in step S105 , the state of health of the battery is detected to obtain historical data. Here, the state of health of the battery is detected through the charging and discharging experiments of the battery, so as to collect historical data such as voltage value, current value, and energy (voltage value × current value). For example, a Prognostic and health management (PHM) tool is used to collect historical data on batteries. Assume that there are 127,340 pieces of historical data collected originally, and then convert the data into hours by taking the average. After conversion, the test data ranges from 0 hours to 1021 hours. Next, perform data preprocessing, such as using Locally estimated scatterplot smoothing (LOESS) to smooth the original data.

In step S110 , set the retrospective training point and the forecast starting point, and use the historical data between the retrospective training point and the forecast starting point as a training set. In this step, set the independent variables of the first stage prediction, that is, the starting point of prediction and the backtracking training point. The training length between the prediction starting point and the retrospective training point is used as the training length of the prediction model, and the historical data between the retrospective training point and the predicted starting point is used as the training set. Here, the retrospective training point is earlier than the forecast start point, and there is no historical data of the battery recorded after the forecast start point.

FIG. 2A is a graph of voltage data according to an embodiment of the present invention. FIG. 2B is a graph of current data according to an embodiment of the present invention. Please refer to FIG. 2A and FIG. 2B , in this embodiment, only the historical data of 0-550 hours are detected. Therefore, 550 hours is set as the forecast starting point (Starting point, SP), and the future time T2 after the forecast starting point SP does not include any known data. And, set a backtracking point (Backtracking point, BP) between 0-550. For example, the user can set the backtracking training point BP according to his experience and battery type. Alternatively, the retrospective training point BP can be set according to the amount of required historical data. The historical data included in the training time length T1 between the prediction starting point SP and the backtracking training point BP are used as the training set.

Next, in step S115, the training set is input into the machine learning model to obtain multiple prediction parameters at multiple time points in the future after the prediction starting point SP. The machine learning model adopts a support vector regression (Support vector regression, SVR) algorithm.

Specifically, after the training time length T1 is determined, the predicted voltage value and the predicted current value at the future time T2 are predicted using the training set in the training time length T1. In this embodiment, a machine learning model based on the SVR algorithm is used. SVR model training speed is fast, the time cost is low, and the short-term prediction effect is good.

（1）

（2） The first stage forecast takes time as an independent variable, and the algorithm of the SVR model is shown in the following formulas (1) and (2):

(1)

(2)

與

是邊際線外的點與邊際的距離。即，將訓練時間長度T1中的訓練集作為輸入值x _i，來獲得輸出值y _i（即，預測參數）。 Among them, w is the coefficient matrix of the SVR model, C is the regular factor of L2 regularization (L2 Regularization), and C>0, x _i is the input value, y _i is the output value, ε is the hyperplane (Hyperplane) and the margin ( Margin) line distance,

and

is the distance from the point outside the boundary line to the boundary. That is, the training set in the training time length T1 is used as the input value _xi to obtain the output value y _i (ie, the prediction parameter).

After obtaining a plurality of prediction parameters at a plurality of time points in the future after the prediction start point SP, in step S120, the prediction parameters are input into the deep learning model, thereby calculating the remaining lifespan. The deep learning algorithm uses one of the convolutional neural network (CNN) and the sparse autoencoder (SAE) combined with the deep neural network (DNN). Here, an algorithm combining SAE with DNN is adopted.

（3）

（4） SAE algorithm formulas are shown in formulas (3) and (4):

(3)

(4)

是估計的輸出值、W是在SAE演算法使用的係數矩陣、b是誤差項、σ是非線性激活函數（Rectified linear unit，ReLU），SAE演算法中的損失函數F使用均方誤差（Mean square error，MSE）MSELoss(∙)加上L1正規化，藉此讓SAE演算法具有稀疏的特性。 where h(∙) is the hidden layer, x' _i is the input value,

is the estimated output value, W is the coefficient matrix used in the SAE algorithm, b is the error term, σ is the nonlinear activation function (Rectified linear unit, ReLU), and the loss function F in the SAE algorithm uses the mean square error (Mean square error, MSE) MSELoss(∙) plus L1 regularization, so that the SAE algorithm has sparse characteristics.

（5）

（6） The DNN algorithm is shown in formulas (5) and (6):

(5)

(6)

是估計的輸出值、W'是在DNN演算法使用的係數矩陣、b'是誤差項、σ'是非線性激活函數，DNN演算法中的損失函數F'使用均方誤差MSELoss(∙)。 where h'(∙) is the hidden layer, x" _i is the input value,

is the estimated output value, W' is the coefficient matrix used in the DNN algorithm, b' is the error term, σ' is the nonlinear activation function, and the loss function F' in the DNN algorithm uses the mean square error MSELoss(∙).

。之後，將SAE的輸出值

作為DNN的輸入值x" _i，進而獲得估計的輸出值

。輸出值

為能量預測值。 That is, the prediction parameters obtained in step S115 are used as the input value x' _i of SAE, and then the estimated output value is obtained

. After that, the output value of SAE

As the input value x" _i of DNN, and then obtain the estimated output value

. output value

is the predicted energy value.

Further, the steps for calculating the remaining lifetime are as follows. The prediction parameters obtained in step S115 are input into the deep learning model to obtain multiple energy prediction values at multiple time points in the future after the prediction starting point SP. Afterwards, it is judged whether the predicted energy value is smaller than a threshold. One of the predicted energy values that is smaller than the threshold and closest to the threshold is taken from the plurality of predicted energy values, and the corresponding time point is used as the final life time. Then, the difference between the final life time and the predicted starting point is taken as the residual life.

FIG. 3 is a graph of energy prediction values according to an embodiment of the invention. Please refer to Figure 3. Hours 0 to 550 are the real energy values obtained from historical data, and hours after the 550th hour are energy prediction values obtained using deep learning models and deep neural network models. A threshold (Threshold, TH) is set to find an energy prediction value that is smaller than the threshold and closest to the threshold.

In addition, a rolling forecast can be further added to improve the accuracy of the forecast. FIG. 4 is a flowchart of an online battery life prediction method according to an embodiment of the present invention. Please refer to FIG. 4, in step S405, historical data is obtained. Next, in step S410, parameters are set. That is, set the backtracking training point and prediction starting point, and set the number of prediction steps (Prediction step) as N, and set each step as O hours.

Next, in step S410, execute the SVR model. Here, time is used as an independent variable (Independent variable) to predict the dependent variable (Controlled variable) of future data, that is, the prediction parameter. Take N=3, O=1 and the prediction start point SP is 550 hours to illustrate, predict the prediction parameters at three future time points after the prediction start point SP, that is, the prediction parameters of 551, 552, and 553 hours .

And, in step S415, the training set is updated. That is, return the prediction parameters obtained by the SVR model each time to the training set, and record the parameters of each SVR model, such as Gamma (Gamma) parameters, penalty coefficient (Cost), etc.

Afterwards, in step S420, the SAE-DNN model is executed. For data with few variables, the SAE-DNN model has fast calculation speed and good learning tendency. The predicted energy value is obtained by using the predicted parameters (predicted voltage value and predicted current value) and time (predicted starting point SP and backtracking training point BP) predicted in step S410. For example, energy prediction values for hours 551, 552, and 553 are obtained by using prediction parameters for hours 551, 552, and 553.

In step S425, it is judged whether the predicted energy value is smaller than a threshold. If none of the predicted energy values is less than the threshold, the rolling prediction is continued according to the number of prediction steps, and the rolling prediction is stopped when the predicted energy value is lower than the threshold. That is, by taking N=3, O=1 and the forecast start point SP as 550 hours, reset the future 3 time points after 3×1 hours from the forecast start point SP (that is, 554, 555, 556 hours ), re-execute steps S410 to S425.

If the loop stops, that is, one of the predicted energy values is less than the threshold, then one of the predicted energy values that is smaller than the threshold and closest to the threshold is taken from the predicted energy values, and the time point corresponding to the predicted energy value is taken as the final life time . Taking Figure 3 as an example, the difference (EP-SP) between the final life time (End point, EP) and the predicted starting point SP is regarded as the residual life.

Before the reliability estimation is performed, the obtained results are predicted values, which may cause deviations. Therefore, a prediction interval (prediction interval) (for example, a 95% confidence level) must be provided to judge the deviation of the prediction results. If the prediction result falls within the prediction interval, it means its reliability is high.

（7） Therefore, in the SAE-DNN model, the Monte Carlo (MC) dropout approach can be further used to establish a prediction interval and calculate the uncertainty of the neural network under the same architecture (prediction interval). The MC method of randomly closing neurons is shown in formula (7):

(7)

（8）

（9） Where f _h is the random closing neuron (Dropout) function, t is the input value, d _i is the mask (Dropout mask), and M is the number of predictions. Use the formula (8) to calculate the mean value p, and use the formula (9) to calculate the estimated value u as the predicted variance, and then get the prediction interval.

(8)

(9)

,殘餘壽命+

]。 The prediction interval refers to a confidence interval (Confidence interval), which can be determined according to the user's acceptance of different products. For example, at a confidence level of 95%, assuming a Z-value of 1.96 for the normal distribution, the prediction interval is: [Residual life -

, remaining life +

].

Table 1 below illustrates the prediction results of performing a verification procedure according to the above-mentioned embodiment.

Table 1 Forecast starting point (Hours) Actual measured residual life (Hours) Remaining life (Hours) obtained through the above examples error prediction interval 550 385 392.0059 7.01 [390.37,394.37] 600 335 336.8737 1.87 [336.45,337.27] 650 285 286.6364 1.64 [284.43,287.15] 700 235 236.3693 1.36 [236.08,236.69] 750 185 186.2191 1.22 [186.57,185.57] 800 135 135.2503 0.25 [134.44,135.65] 850 85 85.7036 0.70 [85.52,85.74]

In summary, the difference between the method of the present invention and the traditional method is that the present invention utilizes double forecasting, wherein the first forecast uses time to predict the dependent variable of future data, and the second forecast uses the first The predicted future dependent variables are used to calculate the residual life. In practical operation, future data cannot be known. The present invention provides rolling multi-step forecasting. The length of rolling steps can be set to one hour, one day, one month, depending on different product life cycles. The rolling forecast of the number of steps forward to the future is more in line with practical applications.

The prediction time and prediction interval of the remaining battery life can be obtained through the present invention, which can be used to assist the owner in evaluating how much time is needed to prepare a new battery in the future, and to determine when to replace the relevant criteria in advance to prevent shutdown problems, or how to increase the loaded battery. control system to increase and prolong battery life remaining life.

S105～S120: steps of online battery life prediction method S405～S430: Steps of Online Battery Life Prediction Method BP: backtracking training point EP: end life time SP: prediction starting point T1: training time length T2: future time

FIG. 1 is a flowchart of an online battery life prediction method according to an embodiment of the present invention. FIG. 2A is a graph of voltage data according to an embodiment of the present invention. FIG. 2B is a graph of current data according to an embodiment of the present invention. FIG. 3 is a graph of energy prediction values according to an embodiment of the invention. FIG. 4 is a flowchart of an online battery life prediction method according to an embodiment of the present invention.

S105~S120: steps of online battery life prediction method

Claims (6)

An online battery life prediction method, comprising: detecting a battery's state of health to obtain a historical data; setting a backtracking training point and a forecast starting point, and using the distance between the backtracking training point and the forecast starting point historical data as a training set; input the training set into a machine learning model to obtain a plurality of forecast parameters at multiple time points in the future after the forecast starting point; and input these forecast parameters into a A deep learning model, whereby a residual life is calculated, wherein the prediction parameters are input into the deep learning model, whereby the step of calculating the residual life includes: inputting the prediction parameters into the deep learning model to obtain the A plurality of energy prediction values at a time point; judging whether these energy prediction values are less than a threshold value; taking one of the energy prediction values that is smaller than the threshold value and closest to the threshold value among the energy prediction values, and predicting one of the energy values The time point corresponding to the value is taken as the final life time; and the difference between the final life time and the predicted starting point is taken as the residual life. The online battery life prediction method as described in claim 1 further includes: setting a number of prediction steps to be N, and setting each step to be O hours; and setting N time points in the future after the starting point of the prediction; wherein, After judging whether the predicted energy values are less than the threshold, it further includes: If none of the predicted energy values is less than the threshold, reset N time points in the future after N×O hours from the starting point of the prediction, so as to recalculate the residual life through the machine learning model and the deep learning model . The online battery life prediction method according to claim 1, after obtaining the prediction parameters, further includes: returning the prediction parameters to the training set. The online battery life prediction method as described in claim item 1, wherein the machine learning model adopts one of support vector regression algorithm, regression algorithm and random forest algorithm, and the deep learning algorithm adopts convolutional neural network and sparse automatic The encoder (Sparse autoencoder) combines one of the deep neural networks. The online battery life prediction method according to claim 1, wherein the prediction parameters include a predicted voltage value and a predicted current value at each of the time points. The online battery life prediction method described in Claim 1 further includes: obtaining a prediction interval by a Monte Carlo method of randomly closing neurons.

Images