A Reproducible Deep Learning Workflow for Predicting Lithium-Ion Battery Performance Using Multichannel Cycling Data

Summary

This protocol presents a reproducible deep learning workflow using multichannel cycling data and engineered features to predict lithium-ion battery state of health, remaining useful life, and energy-storage performance.

Abstract

This protocol presents a reproducible deep-learning workflow for predicting lithium-ion battery performance using multichannel cycling data. To address limitations in existing data-driven models related to multisource information integration and feature representation, the method utilizes five publicly available battery datasets. The inputs include raw operational signals, including voltage, current, capacity, temperature, internal resistance, and coulombic and energy efficiency, together with engineered features such as incremental capacity, differential voltage, and energy throughput. The protocol details standardized preprocessing procedures for constructing battery-level temporal tensors across multiple observation windows. To improve reproducibility, the workflow evaluates candidate neural-network architectures, including deep neural networks, convolutional neural networks (CNNs), long short-term memory (LSTM) networks, CNN–LSTM, and transformer-based models, using fixed random seeds, repeated evaluation runs, and environment tracking. Representative results demonstrated that the CNN–LSTM architecture with integrated multichannel data and engineered features achieved favorable predictive performance among the evaluated models. The selected configuration achieved a mean absolute error of 0.024 and an R² value of 0.955 for state-of-health estimation while maintaining high predictive performance for remaining useful life and dischargeable-energy estimation. Ablation and robustness analyses further demonstrated the contribution of multichannel feature integration and the stability of the reproducibility-focused workflow across repeated evaluations. This standardized methodology provides a reproducible framework for battery lifecycle management and predictive modeling under diverse operating conditions.

Introduction

Lithium-ion batteries are central to the global transition toward renewable energy because of their high energy density, extended cycle life, and low self-discharge rates across applications ranging from portable electronics to grid-scale energy storage systems. However, prolonged operation inevitably leads to gradual capacity degradation and increased internal resistance¹. This performance decline not only reduces economic viability but also introduces severe safety hazards, including thermal runaway. As the deployment of large-scale energy storage systems continues to accelerate, battery lifecycle management has become increasingly complex². Traditional threshold-based monitoring and manual inspection protocols are insufficient for dynamic operating environments³. Furthermore, reliance on a single parameter is inadequate for accurately predicting battery state of health (SOH) or remaining useful life (RUL) under fluctuating charge-discharge rates, ambient temperatures, and discharge voltage variations⁴. Although the current workflow was primarily validated using laboratory-controlled cycling datasets, the multichannel architecture and temporal modeling framework may support future adaptation to dynamic operating conditions encountered in electric-vehicle applications and variable environmental environments.

Battery degradation is a highly complex, nonlinear electrochemical process driven by solid electrolyte interphase growth, loss of active lithium, and structural deterioration of electrode materials. Historically, state estimation has relied on equivalent-circuit models or electrochemical-physical partial differential equations (PDEs)⁵. Although these physical models provide valuable mechanistic insights, they suffer from substantial computational overhead and parameter-identification challenges when processing multichannel coupled signals. Consequently, their real-time implementation in embedded battery-management systems (BMSs) is severely constrained by complex diffusion kinetics and environmental variability under high-rate discharge conditions⁶. In contrast, the proposed deep-learning (DL) workflow concentrates most computational demand within the offline training phase, while the trained model performs comparatively rapid inference during deployment. This framework may therefore support future implementation in resource-constrained BMSs and edge-computing environments.

Recently, data-driven DL approaches have emerged as powerful tools because of their exceptional nonlinear fitting capabilities. Nonetheless, existing DL methodologies exhibit two major limitations. First, many approaches rely primarily on single-capacity degradation curves or specific discharge-voltage thresholds while neglecting the inherent spatiotemporal correlations among multichannel cycling responses, such as voltage, current, and temperature. This limitation reduces model robustness under multicondition operating scenarios⁷. Previous studies have shown that neglecting thermal states during complex operations can accelerate the accumulation of predictive errors^8,9. Second, a persistent reproducibility challenge exists in applied machine learning (ML). Variations in experimental conditions, insufficiently described preprocessing pipelines, and arbitrarily selected neural network hyperparameters frequently cause otherwise robust models to fail during cross-dataset validation¹⁰. To reduce the risk of information leakage during preprocessing and early-cycle forecasting, the proposed workflow derives normalization scaling factors and engineered-feature baselines exclusively from the training datasets prior to external validation and cross-dataset evaluation.

To address these critical gaps, the overall goal of this protocol is to present an automated, highly reproducible DL workflow for predicting lithium-ion battery energy-storage performance using multichannel cycling data. The rationale behind this method is to explicitly capture the complex long-term dependencies associated with battery degradation by integrating cross-modal time-series data within a high-dimensional neural network architecture. A key advantage of this protocol over existing physics-based and DL approaches is its holistic standardization, which establishes a complete pipeline spanning raw-signal preprocessing, feature engineering, model training, and uncertainty evaluation. Although hybrid neural-network architectures have been previously investigated, the present workflow emphasizes standardized multichannel integration by aligning electrochemical cycling variables with engineered degradation features within a reproducibility-focused training framework. Unlike many previously reported workflows, this protocol incorporates standardized parameter configurations, fixed random seeds, repeated evaluation runs, and comprehensive environment tracking to improve reproducibility across heterogeneous battery datasets and laboratory settings. In addition, the workflow systematically compares multiple neural-network architectures to evaluate predictive stability and performance across different input configurations. Within the evaluated datasets and experimental settings, the convolutional neural network–long short-term memory (CNN–LSTM) architecture demonstrated favorable predictive performance relative to the evaluated standalone CNN and transformer-based models by combining local feature extraction with long-range temporal sequence modeling while maintaining moderate computational complexity. This protocol is well suited for researchers and engineers seeking to develop reproducible and environment-adaptive predictive models for battery lifecycle management.

Protocol

Institutional ethics approval was not required for this study because the research exclusively utilized publicly available lithium-ion battery cycling datasets and did not involve human participants, animal subjects, or biological samples.

1. Data Acquisition and Organization

1.  Acquire multichannel cycling datasets
   1. Download lithium-ion battery cycling datasets from publicly available battery repositories containing charge-discharge cycling records. Acquire the CALCE, SANYO, PANASONIC, KOKAM, and GOTION datasets directly from their respective institutional open-access repositories or digital object identifier links to maintain dataset-version consistency.
   2. Select battery samples containing complete cycle numbers, capacity measurements, and at least one continuous charge-discharge curve. Verify continuity by confirming that no interruption exceeding two hours occurs within a single cycle and that both constant-current and constant-voltage phases are fully recorded.
2. Filter raw cycling records
   1. Filter the raw cycling records to retain samples containing temperature, internal resistance, coulombic efficiency, and energy-efficiency measurements. Exclude battery cells lacking these mandatory state channels from the analysis workflow because subsequent multichannel feature fusion requires a complete set of physical input variables.
   2. Remove cycles containing incomplete records, inconsistent timestamps, undefined testing procedures, or unclear charge-discharge termination conditions. Define anomalous cycles as those exhibiting sudden capacity drops exceeding 5% between adjacent cycles, voltage measurements outside the chemistry-specific operating range, or nonmonotonic timestamp intervals.
3. Construct a temporal battery database
   1. Organize the retained cycling records chronologically to construct a battery-level time-series database.
   2. Store the processed datasets using Hierarchical Data Format version 5 (HDF5). Maintain a standardized directory structure organized by dataset source, chemical system, and individual battery-cell identifier.
   3. Assign unique identifiers to individual battery cells before dataset partitioning. Implement grouping based on the battery-cell identifier during train-test splitting to ensure that all sequential cycles from a single physical battery remain exclusively within one dataset subset.
      ​NOTE: Preserve the original raw datasets separately before preprocessing, normalization, resampling, or feature engineering.
4. Define observation windows
   1. Construct multichannel observation windows using the first 50 cycles, the first 100 cycles, and the first 20% of the total battery lifecycle. Calculate the 20% lifecycle threshold by multiplying the total cycle life before end-of-life capacity degradation by 0.2.
   2. Align voltage, current, capacity, temperature, internal resistance, and efficiency measurements within each observation window before tensor construction.
   3. Format the final input structure as a three-dimensional tensor consisting of the observation-window length, 128 standardized intra-cycle sampling points, and the ordered multichannel feature set.

2. Data Preprocessing and Feature Engineering

1. Preprocess raw cycling data
   1. Perform uniform preprocessing on the raw cycling datasets to standardize data quality across all battery sources. Execute all preprocessing, synchronization, and standardization procedures using Python 3.10 with the Pandas library for time-series manipulation and SciPy for numerical interpolation.
   2. Remove outliers, detect missing values, calibrate timestamps, identify charge-discharge phases, and normalize sampling frequencies before downstream analysis. Apply a modified Z-score thresholding method with a rejection criterion greater than 3.0 for anomaly detection.
   3. Detect missing values automatically using NaN flag identification across all numerical arrays. Differentiate charging and discharging phases according to current polarity by assigning positive current values to charging cycles and negative current values to discharging cycles.
   4. Reconstruct missing signal segments using piecewise cubic Hermite interpolating polynomials to preserve electrochemical temporal dynamics during phase-transition alignment.
2. Resample cycling curves
   1. Resample cycling curves when variations in sampling density or signal drift are detected across datasets. Trigger automated resampling whenever the interval between consecutive measurements deviates by more than 10% relative to the median sampling frequency of the corresponding cycling phase.
   2. Standardize the number of sampling points within each cycle to reduce platform-specific recording variation. Map all temporal sequences onto a fixed spatial grid using one-dimensional linear interpolation.
   3. Compress or expand each charge-discharge profile into exactly 128 equidistant sampling points before tensor construction.
3. Normalize continuous variables
   1. Normalize all continuous variables before model training to stabilize network convergence and reduce scale-related bias. Apply Min-Max scaling to transform all continuous variables into a numerical range between 0 and 1.
   2. Compute all normalization parameters exclusively from the training dataset. Apply the frozen scaling parameters to the validation and testing datasets to reduce the risk of forward temporal data leakage during early-cycle prediction.
4. Construct multichannel input tensors
   1. Divide the continuous cycling records for each battery into fixed-length temporal segments. Structure the temporal segments according to the predefined observation windows corresponding to the first 50 cycles, the first 100 cycles, or the first 20% of the battery lifecycle.
   2. Align voltage, current, capacity, temperature, internal-resistance, and efficiency measurements within each temporal segment before tensor construction.
   3. Concatenate the aligned feature matrices into a three-dimensional tensor formatted as (N × 128 × C), where N represents the observation-window length, 128 denotes the standardized intra-cycle sampling points, and C represents the total number of physical and engineered feature channels.
5. Extract engineered features
   1. Extract engineered electrochemical features from the raw cycling signals to improve degradation-pattern identification. Calculate incremental-capacity and differential-voltage features by computing dQ/dV and dV/dQ across the standardized sampling grid.
   2. Apply Savitzky–Golay filtering to generate smoothed incremental-capacity and differential-voltage curves.
   3. Extract degradation indicators including peak-voltage shifts, phase-transition voltage differences, and cycle-to-cycle energy-throughput variations from the filtered electrochemical features.

3. DL Model Construction and Training

1. Construct candidate prediction models
   1. Construct a standardized training and evaluation framework for all prediction models shown in Figure 1A. Implement the DL workflow using PyTorch version 2.0 or higher.
   2. Execute the training workflow using modular Python scripts for data loading, model initialization, optimization, and evaluation.
   3. Prepare traditional ML baseline models including linear regression, random forest, and gradient-boosted decision-tree models. Configure the random-forest model with 100 trees and a maximum depth of 10.
   4. Configure the gradient-boosted decision-tree framework using a learning rate of 0.1 and a maximum tree depth of 5 through the Scikit-learn library.
   5. Prepare DL candidate architectures including multilayer perceptron, one-dimensional CNN, LSTM network, CNN–LSTM, and transformer-based models.
   6. Standardize the core hyperparameters across all DL architectures by applying the AdamW optimizer, Gaussian error linear unit activation functions, a hidden dimension of 128, and a dropout rate of 0.2.
2. Configure the CNN–LSTM architecture
   1. Configure the CNN–LSTM architecture to fuse multichannel cycling information from the input tensors (Figure 1B). Construct the hybrid architecture by cascading the convolutional feature extractor into two stacked recurrent layers containing 128 hidden units each.
   2. Terminate the architecture using three parallel fully connected regression heads for simultaneous prediction of SOH, RUL, and dischargeable energy.
   3. Process the multichannel input tensors through three consecutive one-dimensional convolutional layers with kernel sizes of 3, 5, and 7 to extract multiscale localized features.
   4. Apply batch normalization after each convolutional layer to improve gradient stability during optimization.
   5. Reshape the convolutional feature maps into continuous temporal sequences before sequence modeling.
   6. Feed the encoded temporal sequences into stacked recurrent layers to capture long-range degradation dependencies across the predefined observation windows.
3. Partition datasets for model training
   1. Partition the processed battery-cell datasets into training, validation, and testing subsets using a 70:15:15 ratio. Execute the dataset split using a fixed random seed value of 42.
   2. Apply stratified sampling based on battery chemistry to maintain balanced degradation distributions across all dataset subsets.
   3. Perform dataset partitioning strictly at the battery-cell level rather than randomizing individual cycles to preserve sequential degradation trajectories.
   4. Keep each battery cell as an indivisible unit during partitioning to prevent information leakage between datasets.
   5. Verify dataset isolation programmatically by computing the intersection of battery-cell identifiers across the training, validation, and testing subsets and confirming that no overlap exists.
   6. Compile an external test dataset using the complete GOTION lithium iron phosphate (LFP) dataset for out-of-distribution evaluation and cross-dataset generalization assessment.
   7. Make the source code, environment-configuration files, and processed dataset partitions publicly available through a dedicated GitHub repository to support methodological transparency and reproducibility.
4. Initialize model training
   1. Set the regression loss function to mean squared error before training. Compute the global training objective by aggregating the loss values from each prediction target using equal weighting.
   2. Set the initial learning rate to 1 × 10⁻³ and select a batch size of 32 or 64 according to dataset size.
   3. Use a batch size of 64 for datasets containing more than 5,000 cycles to improve hardware utilization. Use a batch size of 32 for smaller datasets to improve optimization stability.
      NOTE: Select these hyperparameters through preliminary grid-search evaluation to balance convergence speed and gradient stability across different observation windows.
5. Apply training controls
   1. Apply a dynamic learning-rate scheduler based on validation-loss changes. Reduce the learning rate by a factor of 0.5 if the validation loss fails to improve over 5 consecutive epochs.
   2. Maintain a minimum learning-rate threshold of 1 × 10⁻⁵ during optimization.
   3. Stop model training after a maximum of 100 epochs to reduce overfitting.
   4. Apply early stopping if the validation loss does not decrease over 15 consecutive epochs. Restore the model weights corresponding to the best-performing validation iteration.
6. Maintain reproducibility records
   1. Set a fixed random seed value of 42 before model initialization and training.
   2. Apply the fixed random seed consistently across the Python random module, NumPy, and the PyTorch backend while disabling nondeterministic CUDA operations.
   3. Record the software environment, dependency versions, parameter settings, data partitions, and training outputs for each experimental run.
   4. Execute all DL workflows using Ubuntu 22.04 with CUDA 11.8 acceleration on a workstation equipped with an NVIDIA RTX 3090 graphics processing unit or equivalent hardware.

Figure 1. Reproducible deep-learning (DL) workflow for multichannel lithium-ion battery performance prediction. (A) End-to-end reproducible workflow illustrating multichannel data acquisition, standardized preprocessing, tensor construction, model training, and prediction of battery performance indicators. Reproducibility controls include fixed random seeds, repeated training runs, parameter tracking, and environment logging. (B) CNN–LSTM hybrid architecture used for multichannel feature extraction, temporal sequence modeling, feature fusion, and prediction of state of health (SOH), remaining useful life (RUL), and dischargeable energy. (C) Evaluation and deployment pipeline showing candidate model screening, robustness evaluation, hyperparameter optimization, and final model selection. Please click here to view a larger version of this figure.

4. Model Evaluation and Target Prediction

1. Generate model predictions
   1. Process the target test-battery datasets through each trained model. Load the preprocessed test datasets as serialized PyTorch tensors before inference.
   2. Execute the forward inference pass using a gradient-disabling context manager to prevent unintended parameter updates during evaluation.
   3. Generate concurrent predictions for each performance indicator using the trained regression heads.
   4. Extract simultaneous outputs from the three parallel regression heads configured for SOH, RUL, and dischargeable-energy prediction.
2. Calculate prediction targets
   1. Calculate SOH by comparing the present available capacity with the initial battery capacity.
   2. Define the initial battery capacity as the maximum discharge capacity recorded during the first three stabilization cycles.
   3. Compute SOH as the ratio between the present-cycle capacity and the established initial-capacity baseline.
   4. Estimate RUL by identifying the operational threshold at which stable battery capacity ceases.
   5. Define end of life as the cycle number at which battery capacity irreversibly decreases below 80% of the nominal rated capacity.
   6. Calculate RUL as the remaining cycle count from the current observation point to the defined end-of-life threshold.
   7. Forecast dischargeable energy from the recorded discharge conditions.
   8. Compute the ground-truth dischargeable energy by numerically integrating the product of discharge voltage and discharge current across the discharge interval using the trapezoidal integration method.
3. Evaluate predictive performance
   1. Calculate mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²) for each prediction target.
   2. Compute all evaluation metrics using the Scikit-learn metrics module with standardized statistical formulations.
   3. Apply identical metric calculations across all models, datasets, and input configurations to maintain evaluation consistency.
   4. Aggregate all performance metrics across the pooled battery samples within the independent testing dataset to evaluate overall model generalization.
4. Report stratified performance
   1. Calculate the mean and standard deviation for each evaluation metric across repeated experimental runs.
   2. Derive the reported statistical measures from five independent model initializations.
   3. Report 95% confidence intervals together with the mean performance metrics to quantify predictive uncertainty.
   4. Compare model performance across different observation-window lengths and input-channel configurations.
   5. Stratify the comparative analysis using observation windows corresponding to 50 cycles, 100 cycles, and 200 cycles.
   6. Evaluate predictive performance across multiple input configurations ranging from capacity-only inputs to full multichannel tensor integration. Refer to Figure 1C for the evaluation, robustness-screening, and deployment-selection workflow.

5. Reproducibility, Robustness, and Sensitivity Analysis

1. Verify reproducibility and robustness
   1. Maintain identical data partitions, preprocessing workflows, and experimental procedures across all reproducibility evaluations shown in Figure 2A.
   2. Compute the coefficient of variation for the primary evaluation metrics across repeated trials to assess reproducibility consistency.
   3. Repeat all evaluation procedures using the same standardized training and testing workflow.
   4. Initialize each repeated evaluation using randomized network weights while preserving the same deterministic data-loading and shuffling procedures.
2. Perform repeated training experiments
   1. Train the prediction models repeatedly under identical experimental conditions using predefined fixed random-seed values to ensure reproducible weight initialization.
   2. Perform five independent training runs using identical dataset partitions to reduce data-driven variability.
   3. Calculate the final performance metrics by averaging the evaluation metrics obtained from the five repeated training runs.
   4. Aggregate the final robustness metrics using the averaged evaluation scores rather than combining raw prediction tensors.
3. Evaluate hyperparameter sensitivity
   1. Modify the learning rate, batch size, hidden-layer dimension, and dropout ratio individually during sensitivity analysis.
   2. Evaluate learning rates ranging from 1 × 10⁻⁴ to 1 × 10⁻², batch sizes ranging from 32 to 128, hidden dimensions ranging from 32 to 256, and dropout ratios ranging from 0.1 to 0.5.
   3. Record the corresponding changes in predictive accuracy after each hyperparameter adjustment.
   4. Quantify predictive-performance variation using RMSE and R².
4. Perform feature-ablation analysis
   1. Remove selected input channels, including temperature, internal resistance, and engineered-feature vectors, from the multichannel input stream.
   2. Perform feature ablation using a leave-one-out strategy by removing one feature group at a time while preserving the remaining multichannel input structure.
   3. Compare the resulting predictive performance after each feature-removal experiment to evaluate the contribution of individual feature groups to model stability and accuracy.
   4. Quantify the relative contribution of each feature group by calculating the percentage increase in mean absolute error relative to the fully integrated multichannel baseline model. Refer to Figure 2B for the robustness-evaluation and sensitivity-analysis workflow

Figure 2. Standardized preprocessing, robustness evaluation, and sensitivity-analysis workflow. (A) Standardized multichannel preprocessing workflow illustrating raw cycling-signal processing, timestamp calibration, outlier removal, resampling, normalization, channel alignment, and tensor construction. Input-representation strategies compare single-channel inputs, multichannel raw inputs, and multichannel inputs combined with engineered features. (B) Robustness and sensitivity-analysis pipeline showing systematic hyperparameter perturbation, repeated-run evaluation, feature-ablation analysis, cross-validation, and external validation of the CNN–LSTM workflow. Please click here to view a larger version of this figure.

Results

Multichannel Cycling Signals and Data Distribution

Representative multichannel cycling signals and performance distributions were analyzed to characterize the cycling-response behavior of lithium-ion batteries. The study utilized five lithium-ion battery datasets representing multiple chemical systems, including lithium cobalt oxide (LCO), nickel cobalt aluminum oxide (NCA), nickel manganese cobalt oxide (NMC), and LFP (Table 1). Core cycling variables, including voltage, current, capacity, temperature, internal resistance, and coulombic or energy efficiency, were integrated with engineered features such as incremental-capacity and differential-voltage characteristics to establish a multichannel characterization framework. Capacity-voltage profiles and SOH distributions for the CALCE dataset (Figure 3A), mean voltage profiles and SOH distributions for the SANYO dataset (Figure 3B), aging trajectories for the PANASONIC dataset (Figure 3C), voltage-evolution heatmaps for the KOKAM dataset (Figure 3D), and voltage-distribution profiles for the GOTION dataset (Figure 3E) demonstrated distinct degradation behaviors across the evaluated battery chemistries.. For example, the GOTION IFP20100140A dataset exhibited step-like voltage characteristics commonly associated with LFP systems, whereas the PANASONIC NCR18650BD dataset demonstrated progressive capacity degradation accompanied by voltage-profile shifts during long-term cycling.

Dataset ID	Battery Type / Model	Chemistry	Format	No. of Cells	Total Cycles Analyzed	Early-Cycle Windows Used	Sampling Points per Cycle	Core Cycling Channels	Additional State Channels	Derived Features	Prediction Targets
Dataset #1	CALCE LCO battery	LCO	Pouch	38	4,216	First 50, first 100, first 20% life	128	Voltage, current, time, capacity	Temperature, internal resistance	dQ/dV, dV/dQ, charge duration, discharge energy	SOH, RUL, dischargeable energy
Dataset #2	SANYO UR18650E	NCA	Cylindrical 18650	24	2,487	First 50, first 100, first 20% life	128	Voltage, current, time, capacity	Temperature, coulombic efficiency	dQ/dV, voltage plateau slope, energy efficiency	SOH, RUL, dischargeable energy
Dataset #3	PANASONIC NCR18650BD	NCA	Cylindrical 18650	42	5,134	First 50, first 100, first 20% life	128	Voltage, current, time, capacity	Temperature, internal resistance, coulombic efficiency	dQ/dV, dV/dQ, charge duration, energy throughput	SOH, RUL, dischargeable energy
Dataset #4	KOKAM SLPB533459H4	NMC	Pouch	19	1,968	First 50, first 100, first 20% life	128	Voltage, current, time, capacity	Temperature, internal resistance	dQ/dV, discharge duration, voltage slope	SOH, RUL, dischargeable energy
Dataset #5	GOTION IFP20100140A	LFP	Prismatic	57	6,482	First 50, first 100, first 20% life	128	Voltage, current, time, capacity	Temperature, internal resistance, energy efficiency	dQ/dV, plateau duration, charge/discharge energy	SOH, RUL, dischargeable energy

Table 1: Characteristics of the lithium-ion battery datasets, multichannel cycling variables, engineered features, and prediction targets used in the reproducible DL workflow. The table summarizes the lithium-ion battery datasets included in this study, including battery chemistry, cell format, number of cells, total analyzed cycles, observation windows, sampling density, multichannel cycling variables, engineered electrochemical features, and prediction targets used for model development and evaluation. Abbreviations: SOH = state of health; RUL = remaining useful life; LCO = lithium cobalt oxide; NCA = nickel cobalt aluminum oxide; NMC = nickel manganese cobalt oxide; LFP = lithium iron phosphate.

Figure 3. Representative multichannel cycling profiles and battery-health distributions across lithium-ion battery datasets. (A) Capacity-voltage profiles and SOH distributions for the CALCE lithium cobalt oxide dataset. (B) Mean voltage profiles and SOH distributions for the SANYO nickel cobalt aluminum oxide dataset. (C) Aging trajectories and cumulative SOH distributions for the PANASONIC nickel cobalt aluminum oxide dataset. (D) Voltage-evolution heatmaps and SOH sample distributions for the KOKAM nickel manganese cobalt oxide dataset. (E) Voltage-distribution profiles and SOH versus internal-resistance relationships for the GOTION lithium iron phosphate dataset. Please click here to view a larger version of this figure.

Overall Predictive Performance for Energy Storage

The proposed workflow demonstrated low prediction errors and consistent predictive performance across all evaluation tasks (Table 2). For SOH estimation, the model achieved a MAE of 0.021 on the validation dataset and 0.024 on the test dataset, with corresponding R² values of 0.962 and 0.955. These results indicate stable predictive performance for continuous battery-health estimation under the evaluated conditions. The workflow also demonstrated reliable predictive capability for RUL and dischargeable-energy estimation, achieving test-dataset R² values of 0.927 and 0.941, respectively. To further evaluate predictive stability, all performance metrics were calculated across five independent model initializations using identical dataset partitions. The resulting standard deviations and coefficient-of-variation values remained low across repeated evaluations, indicating limited sensitivity to stochastic initialization and stable predictive behavior under repeated training conditions. In addition, 95% confidence intervals demonstrated narrow variability across repeated runs, supporting the robustness and reproducibility of the proposed multichannel workflow.

Prediction Task	Dataset Split	MAE	RMSE	MAPE (%)	R²
SOH estimation	Validation	0.021	0.029	2.47	0.962
SOH estimation	Test	0.024	0.033	2.83	0.955
RUL prediction	Validation	18.6	25.9	8.74	0.938
RUL prediction	Test	21.4	29.7	9.63	0.927
Dischargeable energy prediction	Validation	0.087	0.121	3.18	0.948
Dischargeable energy prediction	Test	0.094	0.129	3.56	0.941

Table 2: Prediction performance of the reproducible DL workflow for lithium-ion battery performance estimation across validation and test datasets. The table summarizes the predictive performance of the proposed workflow for SOH estimation, RUL prediction, and dischargeable-energy prediction using validation and test datasets. Performance metrics include mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²). Abbreviations: SOH = state of health; RUL = remaining useful life; MAE = mean absolute error; RMSE = root mean square error; MAPE = mean absolute percentage error.

Performance Comparison Across DL Architectures

Evaluation across different network architectures and input configurations demonstrated substantial variation in predictive performance (Table 3). RMSE distributions before and after reproducibility-based architecture selection are shown in Figure 4A. The deep neural network architecture produced comparatively higher MAE and RMSE values relative to sequence-based and hybrid architectures. Comparisons across different multichannel input configurations are summarized in Figure 4B. Models incorporating both local feature extraction and cross-cycle temporal modeling demonstrated improved predictive accuracy under the evaluated conditions. Models exhibiting lower prediction variability generally clustered at higher mean-performance scores with reduced standard deviations (Figure 4C). Mean-performance and standard-deviation relationships across evaluated input configurations are illustrated in Figure 4D. Among the tested configurations, the CNN–LSTM architecture achieved the lowest MAE (0.024) and the highest R² value (0.955) within the evaluated datasets and experimental settings. In addition, integrating engineered features with full multichannel inputs reduced the MAE to 0.022 with an R² value of 0.961, whereas capacity-only inputs demonstrated reduced predictive performance.

Category	Model / Input Configuration	MAE	RMSE	MAPE (%)	R²
Deep learning architecture	DNN	0.031	0.042	3.58	0.928
	CNN	0.027	0.037	3.14	0.941
	LSTM	0.026	0.036	3.05	0.944
	CNN–LSTM	0.024	0.033	2.81	0.955
	Transformer	0.025	0.034	2.89	0.951
Input configuration	Capacity only	0.034	0.046	3.92	0.916
	Voltage + current	0.029	0.039	3.29	0.936
	Full multichannel input	0.024	0.033	2.83	0.955
	Multichannel + engineered features	0.022	0.031	2.61	0.961

Table 3: Comparison of prediction performance across DL architectures and input configurations. The table compares the predictive performance of different DL architectures and input configurations used in the reproducible multichannel workflow. Performance metrics include mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²). Abbreviations: DNN = deep neural network; CNN = convolutional neural network; LSTM = long short-term memory; MAE = mean absolute error; RMSE = root mean square error; MAPE = mean absolute percentage error.

Figure 4. Reproducibility-driven comparison of model architectures and input configurations. (A) Root mean square error distributions before and after reproducibility-based architecture selection. (B) Root mean square error distributions across different input configurations before and after stable model selection. (C) Mean-performance versus standard-deviation comparison of candidate model architectures, with color indicating root mean square error and open circles indicating selected stable models. (D) Mean-performance versus standard-deviation comparison of input configurations, with color indicating root mean square error and open circles indicating selected optimal configurations. Please click here to view a larger version of this figure.

Error Distribution and Consistency Evaluation

Absolute-error distributions were analyzed to evaluate prediction stability across conventional ML baselines and DL architectures. Traditional ML baselines exhibited broader error distributions and extended high-error tails, indicating reduced stability for complex degradation-pattern prediction tasks. Among the evaluated DL models, the CNN–LSTM architecture demonstrated comparatively narrower error distributions and shorter high-error tails, suggesting improved integration of local feature extraction and temporal sequence modeling (Figure 5). Incorporating temperature, internal-resistance, and efficiency-related variables was associated with progressive reductions in prediction-error dispersion across the evaluated datasets.

Figure 5. Absolute-error distributions for conventional machine-learning (ML) baselines and multichannel DL models. Violin plots comparing the distribution of absolute prediction errors across conventional ML baselines, DL architectures, and multichannel input configurations. Black lines indicate mean absolute error, and colored lines indicate mean ± standard deviation. Please click here to view a larger version of this figure.

Reproducibility and Robustness Across Validation Scenarios

The workflow maintained consistent predictive performance across repeated training runs and multiple validation strategies (Table 4). Repeated runs performed using fixed data partitions and fixed random-seed initialization produced stable performance metrics, with R² values consistently remaining above 0.95 and low standard deviations in MAE. Additional evaluation using five-fold cross-validation and external validation datasets demonstrated limited performance variation across repeated experimental conditions. The external validation phase exclusively utilized the GOTION dataset, which remained entirely excluded from model training and hyperparameter optimization procedures. Because this dataset represents a distinct LFP chemistry and prismatic-cell configuration, the external evaluation provided a rigorous assessment of cross-dataset generalization capability under heterogeneous electrochemical and structural conditions. In addition, coefficient-of-variation analyses and repeated-run evaluations demonstrated low predictive dispersion across independent experiments, supporting the robustness and reproducibility of the proposed multichannel workflow.

Validation Setting	Performance Metric	Mean	SD	Minimum	Maximum	Coefficient of Variation (%)
Repeated runs with fixed data split (n = 5)	MAE	0.023	0.0021	0.021	0.026	9.13
	RMSE	0.032	0.0028	0.029	0.036	8.75
	MAPE (%)	2.76	0.24	2.48	3.11	8.7
	R²	0.957	0.006	0.949	0.964	0.63
Repeated runs with different random seeds (n = 5)	MAE	0.024	0.0024	0.021	0.027	10
	RMSE	0.033	0.0031	0.029	0.037	9.39
	MAPE (%)	2.84	0.27	2.52	3.19	9.51
	R²	0.954	0.007	0.946	0.962	0.73
Five-fold cross-validation	MAE	0.025	0.003	0.022	0.029	12
	RMSE	0.034	0.0034	0.03	0.039	10
	MAPE (%)	2.95	0.31	2.57	3.41	10.51
	R²	0.951	0.009	0.94	0.961	0.95
External validation	MAE	0.027	0.0032	0.023	0.031	11.85
	RMSE	0.037	0.0038	0.032	0.042	10.27
	MAPE (%)	3.18	0.35	2.76	3.67	11.01
	R²	0.944	0.011	0.931	0.956	1.17

Table 4: Reproducibility and robustness analysis of the proposed multichannel DL workflow across repeated training runs and validation strategies. The table summarizes the reproducibility and robustness performance of the proposed workflow under repeated runs with fixed data partitions, repeated runs with different random seeds, five-fold cross-validation, and external validation. Performance metrics include mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), coefficient of determination (R²), standard deviation (SD), and coefficient of variation. Abbreviations: MAE = mean absolute error; RMSE = root mean square error; MAPE = mean absolute percentage error; SD = standard deviation.

Hyperparameter Sensitivity and Feature Ablation

Hyperparameter sensitivity analysis demonstrated that prediction performance varied across observation-window sizes, activation functions, hidden-dimension sizes, and sequence-model depths. The effect of observation-window size on prediction-error distributions is illustrated in Figure 6A. Under the evaluated conditions, observation windows between 100 and 200 cycles were associated with lower prediction errors, whereas longer observation windows demonstrated diminishing improvements in predictive accuracy. Activation-function-dependent prediction variability is summarized in Figure 6B. Rectified linear unit (ReLU) and Gaussian error linear unit (GELU) activation functions demonstrated lower prediction errors relative to alternative activation functions under the evaluated conditions. The sensitivity of prediction performance to hidden-dimension size and sequence-model depth is illustrated in Figure 6C. Hidden dimensions of 128 and two to three sequence-modeling layers were associated with improved predictive performance, whereas shallow architectures and reduced hidden dimensions were associated with increased prediction error. Feature-ablation analysis further demonstrated that removal of engineered features or efficiency-related variables increased the MAE relative to the full multichannel configuration (Table 5), suggesting that combined multichannel inputs and engineered electrochemical features contributed to improved predictive performance.

Figure 6. Sensitivity of prediction performance to hyperparameter variation. (A) Effect of input-window size on prediction error distribution. (B) Effect of activation-function selection on prediction error distribution. (C) Heatmap showing prediction sensitivity to hidden-dimension size and sequence-model layer number. Black lines indicate mean absolute error, and colored lines indicate mean ± standard deviation where applicable. Please click here to view a larger version of this figure.

Category	Setting	MAE	RMSE	MAPE (%)	R²
Input window size	20 cycles	0.031	0.043	3.74	0.927
	50 cycles	0.026	0.036	3.08	0.944
	100 cycles	0.023	0.032	2.71	0.957
	150 cycles	0.022	0.031	2.63	0.96
	200 cycles	0.022	0.031	2.59	0.961
	250 cycles	0.023	0.032	2.67	0.958
	300 cycles	0.024	0.033	2.79	0.955
Activation function	ReLU	0.023	0.032	2.74	0.957
	GELU	0.022	0.031	2.61	0.961
	Tanh	0.026	0.036	3.05	0.946
	Sigmoid	0.034	0.047	3.96	0.918
Hidden Dimension / Channels per Layer	32	0.028	0.039	3.22	0.939
	64	0.024	0.033	2.81	0.954
	128	0.022	0.031	2.58	0.961
	256	0.023	0.032	2.69	0.958
Number of Sequence Modeling Layers	1	0.027	0.038	3.14	0.942
	2	0.023	0.032	2.73	0.956
	3	0.022	0.031	2.6	0.961
	4	0.024	0.034	2.86	0.952
Feature ablation	Full multichannel + engineered features	0.022	0.031	2.61	0.961
	Without engineered features	0.024	0.033	2.84	0.955
	Without temperature	0.025	0.034	2.93	0.951
	Without internal resistance	0.026	0.035	3.01	0.949
	Without efficiency-related variables	0.025	0.034	2.95	0.95
	Voltage + current only	0.029	0.039	3.29	0.936
	Capacity only	0.034	0.046	3.92	0.916

Table 5: Sensitivity analysis of model performance under different hyperparameter settings and feature-ablation configurations. The table summarizes the effects of input-window size, activation function, hidden-dimension size, sequence-model depth, and feature-ablation strategies on prediction performance within the proposed multichannel DL workflow. Performance metrics include mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²). Abbreviations: MAE = mean absolute error; RMSE = root mean square error; MAPE = mean absolute percentage error; ReLU = rectified linear unit; GELU = Gaussian error linear unit.

Discussion

The reproducible DL workflow described in this protocol demonstrated accurate prediction of lithium-ion battery performance using multichannel cycling data. The workflow achieved a test MAE of 0.024 and a R² value of 0.955 for SOH estimation under the evaluated experimental conditions, supporting its potential utility for battery lifecycle management and energy-storage monitoring applications^11,12,13.

A major advantage of this protocol is the reduction in computational complexity relative to traditional electrochemical-physical PDE models⁵. Although physics-based approaches provide mechanistic insight into battery degradation, they often encounter parameter-identification challenges under high-rate discharge conditions and multichannel operating environments. In contrast, the present data-driven workflow models nonlinear degradation behavior directly from cycling data without requiring explicit physics-based parameterization of cell chemistry¹⁴. Within the evaluated datasets and experimental settings, the CNN–LSTM architecture demonstrated improved predictive performance relative to the evaluated standalone CNN, LSTM, and transformer-based models. Recent studies have demonstrated the effectiveness of transformer architectures for degradation prediction with uncertainty estimation¹⁵, as well as the capability of CNNs for privacy-preserving battery-fault prediction using heterogeneous charging data¹⁶. However, the CNN–LSTM architecture was prioritized in the present workflow because it provided a balanced combination of predictive accuracy, computational efficiency, and training stability under the evaluated laboratory conditions. The convolutional layers extracted localized morphological features from the cycling curves, whereas the LSTM layers modeled long-range temporal dependencies across repeated cycling patterns^17,18. Although the hybrid architecture increases model complexity relative to single-network structures, the computational requirements remained compatible with practical battery-management-system deployment scenarios under the evaluated experimental conditions. Because the majority of computational cost occurs during offline model training, the inference phase remains comparatively lightweight, supporting the feasibility of real-time onboard prediction and edge-level battery-state monitoring without continuous dependence on cloud-based computation.

The performance of this protocol depends strongly on several critical preprocessing and input-configuration steps. First, rigorous sequence-alignment and resampling procedures (Step 2.2) are essential for preserving temporal relationships among multichannel cycling signals. Misaligned multichannel sequences were associated with reduced predictive stability and increased model error during preliminary evaluations. Second, integrating engineered electrochemical features, such as differential-voltage characteristics, with raw cycling signals contributed to improved predictive performance. Feature-ablation analysis demonstrated lower mean absolute error values when engineered features were retained together with temperature and internal-resistance variables¹⁹. In addition, combining engineered electrochemical descriptors with multichannel temporal features improved the model’s ability to capture complex degradation patterns across diverse operating conditions^20,21,22. A critical methodological safeguard implemented in this study was the strict isolation of normalization scaling factors and feature-engineering baselines. All statistical normalization parameters and engineered-feature calculations were derived exclusively from the training dataset before being applied to the validation and testing datasets. This procedure reduced the risk of information leakage during early-cycle forecasting and cross-dataset evaluation, thereby improving the reliability and reproducibility of the predictive workflow.

When executing or modifying this protocol, training instability or gradient-related issues may occur during CNN–LSTM optimization. Loss functions that fail to converge or generate NaN values may indicate gradient explosion within the recurrent sequence-modeling layers. In these situations, gradient clipping, reduction of the initial learning rate from 1 × 10⁻³ to 1 × 10⁻⁴, or moderate increases in batch size may improve training stability by smoothing gradient updates. In addition, ReLU and GELU activation functions were associated with improved gradient propagation relative to sigmoid or hyperbolic tangent activation functions during deep temporal-sequence modeling. The hyperparameter-sensitivity analyses presented in this study may therefore assist researchers in optimizing workflow stability and predictive performance under different training conditions²³.

Despite the predictive performance achieved under the evaluated conditions, this protocol has several limitations. The current workflow was validated primarily using laboratory-controlled constant-current/constant-voltage cycling datasets. In practical electric-vehicle applications, batteries are exposed to stochastic charge-discharge behavior, variable user operating patterns, and fluctuating environmental conditions that may alter degradation characteristics. Under such dynamic operating conditions, data-driven battery models may encounter out-of-distribution scenarios in which previously unobserved degradation trajectories or environmental conditions reduce predictive reliability. To improve robustness in practical deployment environments, future implementations of this workflow may incorporate uncertainty-quantification strategies, such as evidential DL or quantile-regression frameworks, to provide calibrated confidence bounds alongside battery-state predictions. In addition, domain-adaptation and transfer-learning approaches may help bridge the distribution gap between laboratory-generated degradation trajectories and real-world cycling behavior²⁴. Federated-learning or dynamically weighted transfer-learning strategies may further support model pretraining using large laboratory datasets followed by fine-tuning with limited real-world operational data while preserving user-data privacy²⁵. Although the current framework demonstrated stable predictive performance under controlled laboratory conditions, practical BMSs in electric vehicles must operate under highly dynamic load profiles and fluctuating environmental temperatures. Future adaptations of this workflow may therefore benefit from incorporating realistic driving-cycle data and online model-updating strategies to improve predictive reliability during real-world deployment.

By standardizing data partitioning, preprocessing, feature extraction, and random-seed initialization procedures, this protocol improves reproducibility within applied battery ML workflows²⁶. The multichannel feature-integration strategy and reproducibility-focused training framework presented here may support future development of predictive BMSs for renewable-energy storage, electric-vehicle monitoring, and long-term battery lifecycle assessment.

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
CALCE battery dataset	Center for Advanced Life Cycle Engineering (CALCE), University of Maryland	https://calce.umd.edu/battery-data	Source of multichannel lithium-ion battery cycling data for model training and validation
CUDA Toolkit	NVIDIA	Version 11.8	GPU acceleration for DL model training
GOTION IFP20100140A dataset	Mendeley Data	https://doi.org/10.17632/vpw4t7ytbx.2	External validation and multichannel battery degradation analysis
KOKAM SLPB533459H4 dataset	Mendeley Data	https://doi.org/10.17632/7w4y4fzzbb.1	Multichannel battery degradation modeling and validation
Matplotlib	Matplotlib Development Team	Version 3.7.1	Visualization of model performance, sensitivity analysis, and reproducibility results
NumPy	NumPy Developers	Version 1.24.3	Numerical computation and tensor construction
PANASONIC NCR18650BD dataset	Mendeley Data	https://doi.org/10.17632/wykht8y7tg.1	Source of multichannel cycling data and engineered-feature extraction
Pandas	Pandas Developers	Version 2.0.3	Data organization, preprocessing, and tabular-data management
PyTorch	PyTorch Foundation	Version 2.0.1	Construction and training of deep-learning architectures, including the CNN–LSTM model
Python	Python Software Foundation	Version 3.10.12	Data preprocessing, feature engineering, model training, and evaluation
SANYO UR18650E dataset	NASA Prognostics Data Repository	https://ti.arc.nasa.gov/tech/dash/groups/pcoe/prognostic-data-repository/	Source of multichannel cycling data for reproducibility evaluation and model validation
Scikit-learn	Scikit-learn Developers	Version 1.2.2	Baseline-model implementation, dataset partitioning, and performance-metric calculation