Li-Ion Battery Temperature Forecasting Method:

Case-Study

ARTI KHAPARDE, VAIDEHI DESHMUKH, VIDUSHI SHARMA, UTKARSH SINGH

Department of Electrical and Electronics Engineering,

Dr. Vishwanath Karad MIT World Peace University,

Kothrud, Pune,

INDIA

crossing the threshold value above which battery's health characteristics might get hampered. The growing

popularity of data-driven battery prognostics methods shows that ARIMA and LSTM are even when there

aren't many prior details available about the batteries. Have a unique dataset of 34 lithium-ion battery packs for

this challenge. In one way, the results imply that the existing ARIMA techniques offer interpreting data at

various batteries. Having said that, LSTM model outcome recommend that the developed Univariate and

Multivariate LSTM model provides finer prediction accuracy in the existence of greater diversification in data

for one battery. Thus, try to generalize one forecasting model for each battery type depending on the model's

performance.

Key-Words: - Lithium-ion batteries, ARIMA, Deep Learning, LSTM, Time Series, and battery temperature.

applications because of their high energy density,

high power density, low pollution, and prolonged

lifespan, [1]. While the current life test will take a

while, the battery life will inevitably be evaluated

thoroughly and frequently during development.

Long battery life means that performance feedback

is sometimes delayed by many months to years, as is

the case with many chemical, mechanical, and

electronic systems. Additionally, a battery’s

electrical performance will alter as its remaining

batteries could cause thermal runaway, which could

burn or explode devices, [2]. To use batteries

practically, it is crucial to forecast the temperature

change of the batteries in electronic devices and

electric vehicles and to study the thermal effect on

the batteries.

Recently, data-driven techniques are slowly

becoming more popular. Data-driven solutions

based on statistical theories or artificial intelligence

algorithms can deal directly with recorded data as a

result of the sensor's outstanding precision in real-

time battery temperature monitoring, [5].

Researchers provide various techniques, such as

artificial neural network (ANN) approaches, multi-

node equivalent circuit models, physic-chemical

expense associated with gathering and producing

data, as well as the decrease in prediction accuracy

due to extrapolation throughout the prediction

process due to a lack of training data, are other

drawbacks of conventional models. For this reason,

deep learning offers an appropriate way to simulate

the complex and non-linear relationships among

input and output values in the context of Li-ion

cells. For predicting battery temperature, an LSTM

is recommended in the suggested study due to these

performs better for most of the files, i.e., it gives a

lower RMSE and can be concluded as the best-fit

forecasting model for that battery type. ARIMA

models may, however, account for a variety of

patterns, including non-seasonal or seasonal

fluctuations, non-linear or linear trends, and

constant or changing volatility. Since ARIMA

models, at most require fewer variables and

assumptions, it becomes easy to put in the

application and understand. Non-stationary time

predicted using an ARIMA-based LSTM

under a variety of circumstances.

processed to improve the data quality for

effective prediction performance.

for the prediction process using LSTM, which

can remember information for extended

periods.

through the use of ARIMA, which uses the

series data to provide a better understanding

forest, Gaussian process, and artificial neural

network regression. However, in practical

applications, putting cells through equilibration

cycles is not feasible. The authors in [11], developed

a collection of rudimentary models to take the place

of the extreme learning machine's active functions

to improve generalization performance. Because the

model parameters and initial SOC in these "rough

models" are randomly selected within a

characteristics during its operational profile for

different cycles. Owing to this Pearson's Correlation

obtained, we consider the current charge as the other

independent variable for Multivariate LSTM

forecasting. To train our model, we split the dataset

in an 80-20% ratio. We keep aside 20% of the data

for validation. We resample the data frame to

contain only the features that concern, i.e., Time,

Temperature measured, and Current Charge. Figure

To create a forecast model using time series

analysis, the fitting ARIMA model is discovered

from an input y to the criterion Ϸ. This process is

known as integrated (I), autoregressive (AR), and

modeling average (MA).

By transforming the data, the time series can be

rendered stationary if its statistical characteristics,

such as its variance and mean are not constant.

Differentiating is a straightforward transformation

by

, are

computed from this stationary time series. The

model that is given implies  (, , ), where

 represents the AR-term,

 denotes the number of differencing operations

performed. An organized method for defining these

variables ought to be the interpretation of

autocorrelations and partial autocorrelations of

by the fact that

correlated with –2, and the remaining chained

– i)

applicability to compute the precision in the

practical implementation. RMSE is chosen as the

loss function because it is appropriate for the

evaluation of regression models and, consequently,

for the evaluation of time series prediction models.

To determine the value of



temperature readings up to the most recent

observation and estimator pair are considered.

Where  is the desired (correct) output and ŷ is the

estimator. The RMSE number is in the identical

To verify the implementation method efficacy, we

selected four sets of Lithium-ion batteries' (#B0045,

B0046, B0047, and B0048) 34 batteries worth of

seasonal ARIMA. However, due to the length of

observations not being long enough to ensure

seasonality or some clear pattern, we picked

fundamental ARIMA for model buildout. Mainly,

the seasonality is presently unclear for the

temperature measured. For implementing the

ARIMA model, we require optimal values of (p, d,

q). Instead of doing the laborious task of manually

checking the time series for stationarity (d) and

representing the PACF and ACF plots to get the

models comparatively. It is a solitary number score

that can be utilized to find out among several

models which is, in all likelihood, to be the finest

model for a particular data set. A lower AIC score is

However, as a precautionary measure to

graph shows the correlation of today's value with

the lag values. The lag value that gives the highest

correlation is the optimal value to be considered. By

trying different window size values for each battery

type, we chose the value of window size that, on

average gave the best performance. When we

observed the PACF graph, there were high

autocorrelations with many

We found out the best value for window size only by

size values that can be considered.

gives the lowest

This graph plots the RMSE values of all the files

Fig. 15: B0048 Forecasting models' comparison

B0048 comparison table for understanding. Figure

Fig. 16: Count for best B0048 forecasting models

Fig. 17: RMSE plot for B0048 charging files

Figure 17 shows the RMSE plot for B0048

charging files. This graph plots the RMSE values of

(DNN).

emulate patterns based on

the dataset to consider

available. A way to generalize a forecasting model

for each battery type has been discussed based on

the performance of the models on all the charging

files associated with a particular battery type. The

method's primary flaw is the very short time series

data, which may reveal seasonality or other

variations ultimately related to the battery

chemistry.

The use cases of this application are many in

inaccessible areas where a person cannot repeatedly

intervene to record the temperature. This

application is able to give caution warnings in case

the threshold temperature is going to be crossed in

x time steps in the future. Business value is that

provided there is some way to collect the

impact on more complex design and over fitting

issues in the prediction process, thus causing harder

prediction performance. In future work, a novel

technique will be integrated into the deep learning

model to avoid over fitting issues with improve its

performance.

References:

A. Massoud, “A Comprehensive Review of

Lithium-Ion Batteries Modeling, and State of

Health and Remaining Useful Lifetime

Prediction,” IEEE Access, 2022

study on thermal runaway mechanism of

lithium-ion battery with LiNixMnyCo1-x-yO2

cathode materials,” Battery Energy, vol. 1, no.

1, pp. 20210011, 2022.

predictive frequency control management of