Hybrid Deep Learning with Temporal Data Augmentation for Accurate Remaining Useful Life Prediction of Lithium-Ion Batteries

Yun Tian; Guili Wang; Jian Bi; Kaixin Han; Chenglu Wu; Zhiyi Lu; Chenhao Li; Liangwang Sun

cs.LGMar 28, 2026

Hybrid Deep Learning with Temporal Data Augmentation for Accurate Remaining Useful Life Prediction of Lithium-Ion Batteries

Authors: Yun Tian, Guili Wang, Jian Bi, Kaixin Han, Chenglu Wu, Zhiyi Lu, Chenhao Li, Liangwang Sun,

Organizations: School of Physics and Electronic Information, Anhui Normal University, Wuhu 241002, China · College of Mechanical and Vehicle Engineering, Hunan University, Changsha 410082, China · School of Design, NingboTech University, Ningbo 315100, China · The Chinese University of Hong Kong, Shatin, Hong Kong SAR, China ·

Abstract

Accurate prediction of lithium-ion battery remaining useful life (RUL) is essential for reliable health monitoring and data-driven analysis of battery degradation. However, the robustness and generalization capabilities of existing RUL prediction models are significantly challenged by complex operating conditions and limited data availability. To address these limitations, this study proposes a hybrid deep learning model, CDFormer, which integrates convolutional neural networks, deep residual shrinkage networks, and Transformer encoders extract multiscale temporal features from battery measurement signals, including voltage, current, and capacity. This architecture enables the joint modeling of local and global degradation dynamics, effectively improving the accuracy of RUL prediction.To enhance predictive reliability, a composite temporal data augmentation strategy is proposed, incorporating Gaussian noise, time warping, and time resampling, explicitly accounting for measurement noise and variability. CDFormer is evaluated on two real-world datasets, with experimental results demonstrating its consistent superiority over conventional recurrent neural network-based and Transformer-based baselines across key metrics. By improving the reliability and predictive performance of RUL prediction from measurement data, CDFormer provides accurate and reliable forecasts, supporting effective battery health monitoring and data-driven maintenance strategies.

Figures & tables

Figure 1: Workflow of the CDFormer prediction pipeline, including temporal data augmentation, feature extraction with CNN and DRSN, temporal dependency modeling via Transformer encoders, and RUL estimation.

Figure 2: Overall architecture of the 1D-CNN model used for time-series feature extraction.

Figure 3: Architecture of the DRSN module with convolution, soft-thresholding, and residual connections.

Panel of Figure 4 — Figure 4: Two architectures: (a) Transformer encoder block with multi-head attention and feed-forward layers; (b) Regression head.

Figure 5: Three Temporal Augmentation Techniques

Parameter	Description / Value
Battery Type	18650 Cylindrical Li-ion Cell
Manufacturer	Panasonic
Rated Capacity	2.0 Ah
Nominal Voltage	3.6 V
Charge Mode	CC-CV 1.5 A constant current cutoff voltage 4.2 V
Discharge Cutoff Voltage	B0005: 2.7 V; B0006: 2.5 V; B0007: 2.2 V; B0018: 2.5 V

Table 1: Detailed Information of NASA Lithium-Ion Battery Dataset

Parameter	Description / Value
Battery Type	Lithium-ion pouch cell
Manufacturer	Not explicitly specified varies
Rated Capacity	1.1 Ah nominal
Nominal Voltage	3.7 V
Charge Mode	CC-CV 0.5C constant current cutoff voltage 4.2 V
Discharge Cutoff Voltage	2.7 V

Table 2: Detailed Information of CALCE Lithium-Ion Battery Dataset

Input Feature	NASA	CALCE
Average voltage (Voltage_avg)	✓
Average current (Current_avg)	✓
Average temperature (Temp_avg)	✓
Discharge capacity (Capacity)	✓	✓
Constant-current charge time (derived from CC-phase duration)		✓
State of health (SOH) $\left(SOH=\frac{C_{\text{current}}}{C_{\text{initial}}}\times 100\%\right)$		✓

Table 3: Input features used in the NASA and CALCE datasets.

Models	NASA			CALCE
Models	RMSE $\downarrow$	MAE $\downarrow$	RE $\downarrow$	RMSE $\downarrow$	MAE $\downarrow$	RE $\downarrow$
RNN [ 4 ]	0.1435	0.1205	0.3692	0.0534	0.0423	0.1418
LSTM [ 51 ]	0.1573	0.1356	0.3692	0.0720	0.0487	0.1466
GRU [ 22 ]	0.0863	0.0675	0.3002	0.0973	0.0756	0.1646
DeTransformer [ 7 ]	0.0863	0.0746	0.2609	0.0616	0.0481	0.1282
SA-LSTM [ 42 ]	0.0849	0.0751	0.3725	0.0324	0.0263	0.1097

Table 4: Quantitative evaluations of CDFormer and baseline models on the NASA and CALCE datasets.

Panel of Figure 6 — Figure 6: Evaluation of CDFormer and baseline models on NASA and CALCE datasets. Left column: RMSE, MAE, and RE metrics (top) and predicted vs true capacity curves (bottom) for NASA. Right column: same metrics and predicted vs true curves for CALCE.

Models	NASA			CALCE
Models	RMSE $\downarrow$	MAE $\downarrow$	RE $\downarrow$	RMSE $\downarrow$	MAE $\downarrow$	RE $\downarrow$
Baseline-FC	0.1952	0.1822	0.5572	0.0309	0.0184	0.0980
CNN-FC	0.1117	0.0981	0.2322	0.0209	0.0127	0.0828
CNN-Transformer	0.1059	0.0904	0.2264	0.0225	0.0133	0.0896
CDFormer	0.0671	0.0586	0.2365	0.0179	0.0106	0.0877

Table 5: Quantitative evaluations of CDFormer and other deep learning-based variant models.

Panel of Figure 7 — Figure 7: RMSE, MAE, and RE metrics (top) and predicted vs true capacity curves (bottom) of different deep learning-based variant models on each battery in the NASA and CALCE datasets.

temporal data augmentation	NASA			CALCE
temporal data augmentation	RMSE $\downarrow$	MAE $\downarrow$	RE $\downarrow$	RMSE $\downarrow$	MAE $\downarrow$	RE $\downarrow$
No Augmentation	0.0671	0.0586	0.2365	0.0179	0.0106	0.0877
Gaussian Noise	0.0538	0.0447	0.1548	0.0166	0.0101	0.0869
Time Warping	0.0585	0.0492	0.1162	0.0181	0.0111	0.0819
Time Resampling	0.0666	0.0570	0.1184	0.0210	0.0125	0.0759
All Combined (Ours)	0.0371	0.0301	0.1327	0.0169	0.0111	0.0817

Table 6: Impact of temporal data augmentation on NASA and CALCE datasets.

Panel of Figure 8 — Figure 8: Visualization of temporal data augmentation results on the NASA and CALCE datasets. Top: normalized and inverted radar and bar charts of RMSE, MAE, and RE metrics; Bottom: predicted versus true capacity curves of CDFormer under different augmentation settings. Left column: NASA dataset; Right column: CALCE dataset.

Models	Before temporal data augmentation			After temporal data augmentation
Models	RMSE $\downarrow$	MAE $\downarrow$	RE $\downarrow$	RMSE $\downarrow$	MAE $\downarrow$	RE $\downarrow$
Baseline-FC	0.1952	0.1822	0.5572	0.1814	0.1677	0.4005
CNN-FC	0.1117	0.0981	0.2322	0.0748	0.0677	0.1990
CNN-Transformer	0.1059	0.0904	0.2264	0.0725	0.0606	0.1670
CDFormer	0.0671	0.0586	0.2365	0.0371	0.0301	0.1327

Table 7: Performance of model variants before and after temporal data augmentation on the NASA dataset

Models	Before temporal data augmentation			After temporal data augmentation
Models	RMSE $\downarrow$	MAE $\downarrow$	RE $\downarrow$	RMSE $\downarrow$	MAE $\downarrow$	RE $\downarrow$
Baseline-FC	0.0309	0.0184	0.0980	0.0312	0.0188	0.0870
CNN-FC	0.0209	0.0127	0.0828	0.0213	0.0126	0.0799
CNN-Transformer	0.0225	0.0133	0.0896	0.0220	0.0134	0.0724
CDFormer	0.0179	0.0106	0.0877	0.0169	0.0111	0.0817

Table 8: Performance of model variants before and after temporal data augmentation on the CALCE dataset

Panel of Figure 9 — Figure 9: Comparison of the impact of data augmentation on different model variants for both NASA and CALCE datasets.

Models	Before temporal data augmentation			After temporal data augmentation
Models	RMSE $\downarrow$	MAE $\downarrow$	RE $\downarrow$	RMSE $\downarrow$	MAE $\downarrow$	RE $\downarrow$
RNN [ 4 ]	0.1435	0.1205	0.3692	0.1429	0.1205	0.2880
LSTM [ 51 ]	0.1573	0.1356	0.3692	0.1574	0.1357	0.3692
GRU [ 22 ]	0.0792	0.0675	0.3002	0.0793	0.0675	0.3002
DeTransformer [ 7 ]	0.0863	0.0746	0.2609	0.0865	0.0747	0.2586
SA-LSTM [ 42 ]	0.0897	0.0751	0.2013	0.0898	0.0753	0.1990

Table 9: Impact of temporal data augmentation on single temporal feature models on the NASA dataset.

Models	Before temporal data augmentation			After temporal data augmentation
Models	RMSE $\downarrow$	MAE $\downarrow$	RE $\downarrow$	RMSE $\downarrow$	MAE $\downarrow$	RE $\downarrow$
RNN [ 4 ]	0.0534	0.0423	0.1418	0.0535	0.0423	0.1418
LSTM [ 51 ]	0.0720	0.0487	0.1466	0.072	0.0487	0.1470
GRU [ 22 ]	0.0973	0.0756	0.1646	0.0975	0.0757	0.1646
DeTransformer [ 7 ]	0.0616	0.0481	0.1282	0.0619	0.0484	0.1271
SA-LSTM [ 42 ]	0.0324	0.0263	0.1097	0.0323	0.0262	0.1103

Table 10: Impact of temporal data augmentation on single temporal feature models on the CALCE dataset.

RUL Estimation

Remaining useful life estimation predicts how long machines, batteries, and other assets can operate before failure or end of life.

See all 57 papers

A Multi-Task Targeted Learning Framework for Lithium-Ion Battery State-of-Health and Remaining Useful Life
Mar 20, 2026Chenhan Wang, Zhengyi Bao, Huipin Lin +2RUL Estimation Lithium-Ion Batteries
School of Electronics and Information Engineering, Hangzhou Dianzi University, Hangzhou 310018, China · School of Information Technology and Artificial Intelligence, Zhejiang University of Finance and Economics, Hangzhou 310018, China · Engineering Training Center, China Jiliang University, Hangzhou 310018, China
Joint Remaining Useful Life Prediction and Capacity Estimation of Lithium-Ion Batteries Using Partial-Charging Data
Sep 18, 2026Khoa Tran, Ho-Si-Hung Nguyen, Phone Wai Yan Moe +2RUL Estimation Multi-Task Learning
Data Science Laboratory, Faculty of Information Technology, Ton Duc Thang University, Ho Chi Minh City 70000, Vietnam · Faculty of Electrical Engineering, The University of Danang - University of Science and Technology (DUT), 54 Nguyen Luong Bang Street, Lien Chieu District, Da Nang 550000, Vietnam · AIWARE Limited Company, 17 Huynh Man Dat Street, Hoa Cuong Ward, Hai Chau District, Da Nang 550000, Vietnam +2
Dynamic Loss Balancing for Joint SOH and RUL Prediction of Lithium-Ion Batteries via a Rotary SOH-Injected Prior Battery Transformer
Jul 19, 2026Shuhao Chen, Tianyu Shi, Yiwen Huang +1RUL Estimation Predictive Maintenance
School of Economics and Management, Zhejiang Sci-Tech University, Hangzhou, 310018, China.

Mar 20, 2026J/K move · Enter open · S save

A Multi-Task Targeted Learning Framework for Lithium-Ion Battery State-of-Health and Remaining Useful Life

Chenhan Wang, Zhengyi Bao, Huipin Lin, Jiahao Nie+1

School of Electronics and Information Engineering, Hangzhou Dianzi University, Hangzhou 310018, China · School of Information Technology and Artificial Intelligence, Zhejiang University of Finance and Economics, Hangzhou 310018, China · Engineering Training Center, China Jiliang University, Hangzhou 310018, China

RUL Estimation Lithium-Ion Batteries Multi-Task Learning

Open paper AskPDF arXiv

Accurately predicting the state-of-health (SOH) and remaining useful life (RUL) of lithium-ion batteries is crucial for ensuring the safe and efficient operation of electric vehicles while minimizing associated risks. However, current deep learning methods are limited in their ability to selectively extract features and model time dependencies for these two parameters. Moreover, most existing methods rely on traditional recurrent neural networks, which have inherent shortcomings in long-term time-series modeling. To address these issues, this paper proposes a multi-task targeted learning framework for SOH and RUL prediction, which integrates multiple neural networks, including a multi-scale feature extraction module, an improved extended LSTM, and a dual-stream attention module. First, a feature extraction module with multi-scale CNNs is designed to capture detailed local battery decline patterns. Secondly, an improved extended LSTM network is employed to enhance the model's ability to retain long-term temporal information, thus improving temporal relationship modeling. Building on this, the dual-stream attention module-comprising polarized attention and sparse attention to selectively focus on key information relevant to SOH and RUL, respectively, by assigning higher weights to important features. Finally, a many-to-two mapping is achieved through the dual-task layer. To optimize the model's performance and reduce the need for manual hyperparameter tuning, the Hyperopt optimization algorithm is used. Extensive comparative experiments on battery aging datasets demonstrate that the proposed method reduces the average RMSE for SOH and RUL predictions by 111.3% and 33.0%, respectively, compared to traditional and state-of-the-art methods.

Abstract

RUL Estimation

Explore similar work

A Multi-Task Targeted Learning Framework for Lithium-Ion Battery State-of-Health and Remaining Useful Life

Joint Remaining Useful Life Prediction and Capacity Estimation of Lithium-Ion Batteries Using Partial-Charging Data

Dynamic Loss Balancing for Joint SOH and RUL Prediction of Lithium-Ion Batteries via a Rotary SOH-Injected Prior Battery Transformer