Development of a Lithium-Ion Battery Lifetime Prediction Model Using

Deep Learning for Short-Term Learning

Yu Fujitaki and Hiroyuki Kobayashi

Osaka Institute of Technology, Osaka, Japan

Keywords:

Lithium-Ion Batteries, Deep Learning, Cnn, Prediction.

Abstract:

We will use the open data utilized in Severson’s research. This data consists of cycle data obtained from

repeated charging and discharging of lithium-ion batteries, which will be analyzed.One issue is that the amount

of cycle data is limited, which could lead to inadequate training. To address this problem, we have adopted a

method that extracts multiple data points from a single battery dataset, thereby improving prediction accuracy.

In this experiment, we compared data from 100 charge-discharge cycles with data from just 1 charge-discharge

cycle.

1 INTRODUCTION

I am surrounded by a wide range of portable elec-

tronic devices, such as smartphones, all of which de-

pend on lithium-ion batteries (LIBs). LIBs have been

in use for about 25 years and are now a common

component in our daily lives due to their high energy

and power density, long lifespan, cost-effectiveness,

and reliable safety compared to other commercially

available batteries. Despite these advantages, the fre-

quency of accidents involving LIBs has been increas-

ing in recent years. To ensure their safe and reliable

use in electric vehicles and other devices, it is essen-

tial to monitor various parameters including voltage,

temperature, state of charge (SOC), state of health

(SOH), remaining capacity, and cycle life. While

parameters like voltage and temperature can be di-

rectly measured with sensors, others such as SOC and

SOH require estimation through algorithms based on

measurement characteristics. Predicting battery cy-

cle life is crucial, but traditional methods for doing

so are complex, relying on physics-based models and

needing to account for a wide range of operating con-

ditions and device variability, even among batteries

from the same manufacturer. Recently, there has been

increasing interest in using machine learning-based

methods to predict battery behavior. These methods

offer the potential for accurate early predictions of

battery cycle life, which not only speeds up the vali-

dation of new manufacturing processes but also helps

end-users detect performance degradation and replace

failing batteries in time.(Schmush, 2018).

2 OBJECTIVE

In the prediction of the lifespan of lithium-ion batter-

ies (LIBs), methods include not only simple empirical

approaches but also physical models that hypothesize

degradation phenomena and numerically solve elec-

trochemical reaction equations. More recently, data-

driven models that use machine learning to predict

lifespan based on charge-discharge cycle data have

emerged. The data-driven method that I am particu-

larly focused on has been highly regarded for its abil-

ity to estimate LIB capacity, remaining life, and cycle

life. However, this method is based on limited empir-

ical or mechanical test data and does not take degra-

dation modes into account.

In Severson’s study, they predicted LIB cycle life

using data from the ﬁrst 100 cycles, during which

degradation is minimal, achieving a very low test er-

ror of 9.1%. This research demonstrates the potential

of data-driven approaches in predicting the behavior

of complex nonlinear systems. However, a limitation

of this method is that only one piece of data is gen-

erated per battery, leading to a shortage of data. Ad-

ditionally, when predicting the performance of a new

battery, it requires 100 charge-discharge cycles before

a prediction can be made, which is another issue.

Therefore, my goal is to develop a model that can

achieve more accurate predictions with fewer charge-

discharge cycles by training on single-cycle data and

extracting multiple data points from a single bat-

tery, rather than relying on multiple cycles.(Severson,

2019)

418

Fujitaki, Y. and Kobayashi, H.

Development of a Lithium-Ion Battery Lifetime Prediction Model Using Deep Learning for Short-Term Learning.

DOI: 10.5220/0013072100003822

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 21st International Conference on Informatics in Control, Automation and Robotics (ICINCO 2024) - Volume 2, pages 418-422

ISBN: 978-989-758-717-7; ISSN: 2184-2809

3 DATA SPLITTING METHOD

First, the data obtained from charging and discharging

a single battery includes values such as voltage during

charging and discharging, as well as discharge capac-

ity. This data is recorded over the course of hundreds

of charge-discharge cycles. Data collection continues

until the battery’s maximum capacity falls below 80%

of its initial value, indicating signiﬁcant degradation.

The data obtained in this way is often used for predict-

ing battery life and evaluating performance. By ana-

lyzing the changes in voltage and capacity for each

cycle, degradation trends can be identiﬁed.

However, instead of using this dataset as-is, em-

ploying a data splitting method is crucial for efﬁcient

analysis. By appropriately splitting the large amount

of data obtained from a single battery, we can increase

the diversity of data used for training machine learn-

ing models, resulting in more accurate predictions.

Next, the data is divided by each cycle. For exam-

ple, when using cycle data from a battery with a lifes-

pan of 800 cycles, the data from the ﬁrst cycle is used

as the training data(Figure1). At this time, the corre-

sponding target data for the ﬁrst cycle is ”800.” This

means that the battery is still capable of 800 more

charge-discharge cycles.

Figure 1: Data Splitting Method.

Similarly, when using the data from the second cy-

cle as training data, the corresponding target data is

”799.” This indicates that at the second cycle, the bat-

tery has 799 charge-discharge cycles remaining. In

this way, for each cycle’s data, the remaining number

of charge-discharge cycles (i.e., the remaining cycle

life) is set as the target data, and the model is trained

accordingly.

By using this method, the data obtained from

a single battery can be split into multiple training

datasets, increasing the overall amount of data. As

a result, the accuracy of the machine learning model

is expected to improve.

4 LEARNING MODELS

In this study, we utilized a one-dimensional convo-

lutional neural network (1D-CNN) to analyze bat-

tery charge-discharge cycle data. The 1D-CNN is

particularly effective for processing time-series data

due to its ability to perform convolutional operations

along a single dimension.(Figure2) This characteris-

tic makes it especially suitable for handling sequential

data, such as the voltage, current, temperature, and ca-

pacity measurements taken during battery cycles. Un-

like traditional two-dimensional convolutional neural

networks (2D-CNN), which are designed for spatial

data, the 1D-CNN focuses on extracting features from

a linear sequence of data points(Pelletier et al., 2019).

Figure 2: 1D-CNN.

The dataset for each cycle is represented as a two-

dimensional array, where one dimension corresponds

to the time steps within the cycle, and the other di-

mension represents various features related to the bat-

tery’s performance. By applying convolutional oper-

ations across the feature dimension, the 1D-CNN ef-

fectively captures and learns signiﬁcant patterns that

are indicative of battery degradation and remaining

lifespan. This approach allows the model to iden-

tify and understand the temporal dynamics and trends

within the charge-discharge cycles.Table 1 lists the

parameters of the 1D-CNN model used in this study.

Table 1: Model Architecture and Parameters.

Layer (type) Param #

Conv1D (64 ﬁlters, kernel size=50) 19,264

MaxPooling1D (pool size=2) 0

Conv1D (64 ﬁlters, kernel size=50) 204,864

MaxPooling1D (pool size=2) 0

Flatten 0

Dense (512 units, ReLU) 918,016

Dense (1 unit, Linear) 513

Development of a Lithium-Ion Battery Lifetime Prediction Model Using Deep Learning for Short-Term Learning

419

CNNs, in general, are renowned for their ability

to learn hierarchical features from data. They ex-

cel in recognizing patterns through convolutional lay-

ers, which apply ﬁlters to the input data to detect lo-

cal features, followed by pooling layers that reduce

the dimensionality while retaining critical informa-

tion. Fully connected layers then integrate these fea-

tures to perform the ﬁnal predictive tasks. Although

CNNs are widely used in image processing, their

adaptation to one-dimensional data sequences proves

valuable for predicting battery performance, as they

can capture temporal relationships and nuances in the

data(Simonyan and Zisserman, 2014).

In summary, by employing a 1D-CNN, our re-

search aims to leverage the model’s capability to pro-

cess and analyze time-series data from battery cycles,

thereby enhancing the accuracy of predictions related

to battery degradation and lifespan. This methodol-

ogy offers a promising avenue for improving the re-

liability and efﬁciency of battery performance assess-

ments.

5 EXPERIMENT

In this study, we used the dataset from the ”Data-

driven prediction of battery cycle life before capac-

ity degradation” (Schmush, 2018). This dataset con-

sists of data from 136 commercial lithium-ion batter-

ies that were cycled between 150 and 2,300 times un-

der 72 different fast-charging conditions, totaling ap-

proximately 96,700 cycles. The dataset includes 15

variables, such as voltage, charge capacity, discharge

capacity, charge energy, discharge energy, and inter-

nal battery temperature(Table 2). The batteries are

Table 2: Data set variables.

Data Point Test Time

Date Time Step Time

Step Index Cycle Index

Current Voltage

Charge Capacity Charge Energy

Discharge Capacity Discharge Energy

dV/dt Temperature

fff Internal Resistance

lithium iron phosphate (LFP)/graphite cells manufac-

tured by A123 Systems (APR18650M1A), cycled us-

ing a 48-channel Alvin LBT potentiometer in a forced

convection temperature chamber set at 30°C. The pur-

pose of this study was to optimize fast charging. All

cells were charged using a one-step or two-step fast

charging policy in the format of ”C1(Q1)-C2.” Charg-

ing was conducted at 1C CC-CV up to 80% SOC,

with upper and lower cutoff voltages set at 3.6V and

2.0V, respectively. All cells were discharged at 4C.

The dataset is divided into three batches, each

based on the ”batch date” and containing approxi-

mately 48 cells per batch. Temperature was measured

using a T-type thermocouple; however, it is important

to note that measurement accuracy varied, and con-

tact may have been lost during cycling. Internal re-

sistance was measured during charging at 80% SOC

using ±3.6C pulses.

Additionally, the data used in this study was ob-

tained from the publicly available Severson dataset,

which contains data from 127 batteries with lifespans

ranging from 450 to 1,325 cycles. The dataset in-

cludes six attributes: voltage, charge capacity, dis-

charge capacity, charge energy, discharge energy, and

internal battery temperature. For training purposes,

data from the ﬁrst 100 cycles of each battery was

used, yielding a total of 12,700 data points (100 cy-

cles × 127 batteries).

6 COMPARISON

To evaluate the proposed method, we compared it

with the results of a previous study on battery life pre-

diction. In that previous research, one data point was

obtained from each battery using data from cycles 1

to 100. The learning model employed was a CNN, as

in the current study.

In the proposed method, a 1D CNN was used,

but when training on data from cycles 1 to 100, the

convolutional operations on the features remained un-

changed. However, since the current data consists of

two-dimensional data with time steps and 100 cycles,

a 2D CNN was utilized. In the next section, we will

compare the prediction accuracy and data volume be-

tween the previous study and the proposed method.

7 RESULTS

For training data, 84 batteries were used, with 100 cy-

cles of data collected from each battery, resulting in a

total of 8,400 data points. The remaining 43 batter-

ies were used for validation data, yielding a total of

4,300 data points from 100 cycles of data each.The

current study(Figure 3) and the previous study(Figure

4) show graphs with the actual measured life cycles

on the x-axis and the predicted life cycles on the y-

axis.The closer the results are to the red line in the

graph, the better. However, due to the difference in

the amount of data between this study and the previ-

ous one, it may be difﬁcult to discern from the graph.

ICINCO 2024 - 21st International Conference on Informatics in Control, Automation and Robotics

420

Figure 3: Current.

Figure 4: Previous.

To evaluate the battery life prediction results, we

used the Mean Absolute Percentage Error (MAPE).

MAPE =

∑

i=1



− ˆy



× 100 (1)

MAPE is an effective metric for assessing prediction

accuracy, as it expresses the error between actual val-

ues and predicted values as a percentage, allowing

us to understand the overall error trend.The results of

the proposed method(table 3) are compared with the

predictions from the previous research described ear-

lier(table 4).

Table 3: Current.

Model 1D-CNN

Number of training samples 8400

Number of charge-discharge cycles 1

Number of features 6

MAPE 8.43%

By examining both the tables, it is evident that the

proposed method in this study provides better accu-

racy compared to the previous approach.

Table 4: Previous.

Model 2D-CNN

Number of training samples 84

Number of charge-discharge cycles 100

Number of features 6

MAPE 17.40%

8 CONCLUSION

The results conﬁrm that the proposed method of ex-

tracting multiple data points from a single battery cy-

cle effectively increases the amount of training data.

This approach also enables high-accuracy predictions

for the lifespan of new batteries, even with a lim-

ited number of charge-discharge cycles. Furthermore,

the improved prediction accuracy compared to con-

ventional methods suggests that this technique could

be highly useful in practical battery management and

lifespan prediction. However, despite the improve-

ment in accuracy, there is still a noticeable gap be-

tween the predicted and actual battery life, as shown

in the graphs.

9 PERSPECTIVES

In future research, we will focus on developing a

hybrid model that combines data-driven approaches

with physics-based models that numerically solve

electrochemical reaction equations, taking battery

degradation mechanisms into account. In this study,

we recognized the limitations of using only a data-

driven approach, which led us to consider integrating

a model-based approach. This hybrid model will in-

corporate predictions from the physics-based model

as auxiliary data, feeding them back into the learn-

ing process of the machine learning model. Since

many studies on model-based lifetime prediction es-

timate changes with each charge-discharge cycle, we

are exploring methods that align with the single-cycle

charge-discharge data proposed in this study. By

doing so, we aim to leverage the strengths of both

the physical and machine learning models to achieve

greater accuracy and broaden the application of bat-

tery lifetime prediction.

REFERENCES

Pelletier, C., Webb, G. I., and Petitjean, F. (2019). Tem-

poral convolutional neural network for the classiﬁca-

Development of a Lithium-Ion Battery Lifetime Prediction Model Using Deep Learning for Short-Term Learning

421

tion of satellite image time series. Remote Sensing,

11(5):523.

Schmush, Richard, e. a. (2018). Performance and cost

of materials for lithiumbased rechargeable automotive

batteries. Nature Energy 3.4, page 267.

Severson, K. (2019). Data-drivenpredic- tion of battery cy-

cle life before capacity degradation. Nature Energy

4.5, pages 383–391.

Simonyan, K. and Zisserman, A. (2014). Very deep con-

volutional networks for large-scale image recognition.

arXiv preprint arXiv:1409.1556.

ICINCO 2024 - 21st International Conference on Informatics in Control, Automation and Robotics

422