Hybrid machine learning framework for battery remaining useful life prediction with time-series modelling and statistical validation

Data acquisition and pre-processing

The data set taken for the study should be clean and structured to make it suitable for hybrid learning. So, the data pre-processing is needed. Battery datasets (NASA Data set) collected comprises of charge/discharge cycles. These include sensor measurements like voltage, current, temperature, internal resistance, charge capacity, and discharge capacity. The steps involved in pre-processing are: Missing data handling, cycle alignment, normalization and labelling:

The first step is forward-filling for small gaps; remove or impute larger ones. Then align the features to each charge/discharge cycle to maintain sequential consistency. Apply Min-Max or Z-score normalization to scale features for better neural network convergence. RUL is defined as the number of remaining cycles until the battery falls below a predefined capacity threshold (e.g., 80% of initial capacity).

Feature selection

The data set used may be huge with so many parameters, but all the parameters may not contribute to determine the required output. So, the dimensionality of the parameters to be reduced and highlight only the most informative indicators using feature selection model.

In this work a Recursive Feature Elimination (RFE) with RF and Pearson Correlation Analysis (PCA) is used. The purpose of using RFE is to eliminate irrelevant or redundant features that may introduce noise or overfitting and PCA is used to avoid multicollinearity that can mislead regression-based models and inflate variance.

The steps involved are:

• Start with all features.

• Train a RF model.

• Rank features by importance.

• Iteratively remove the least important feature and retrain.

• Calculate Pearson correlation coefficient (|r|) between each pair of features.

• If |r| > 0.85, one of the two features are removed.

Temporal feature extraction

The objective of temporal feature extraction is to capture long-term degradation trends that span multiple cycles. The model used is LSTM neural networks. This is used due to its ability to retain long-term sequential dependencies. Here Hyperparameter tuning and k-fold cross-validation is used.

The steps involved are:

• Input the data like sequence of cycles (e.g., last 20–50 cycles) containing features like voltage, temperature, and charge/discharge capacity.

• Output to predicted is RUL value. The outcome of this step generates time-aware degradation representations that feed into the ensemble predictor.

Structured feature learning using random forest

The structed feature learning is used to model non-temporal, structured data like internal resistance and summary statistics from each cycle. RF is used because of its resistance to overfitting, ability to handle mixed data types and feature importance scoring. RF works in parallel with LSTM to process structured, non-sequential input and contributes feature importance for fusion and model interpretability.

Feature fusion and final RUL prediction

The feature fusion step merge outputs from LSTM and RF and perform the final RUL prediction using a powerful XGBoost ensemble model. The fusion strategy used is LSTM latent representations and selected structured features from RF are concatenated and combined feature vector is used as input to XGBoost. The gradient boosted decision trees is known for handling structured/tabular data efficiently. Also learns complex interactions and nonlinear patterns. Instead of averaging outputs, a stacking ensemble strategy is used which allows XGBoost to learn how to weight inputs from LSTM and RF adaptively. The final RUL prediction is made, leveraging both time-series and structured insights.

Statistical validation and evaluation

To rigorously validate the robustness, accuracy, and generalizability of the proposed hybrid framework, statistical validation and evaluation is needed. The following are the statistical validation tool used:

• a.

  Confidence interval estimation: It is a Bootstrap resampling or standard error-based methods. It calculates 95% confidence intervals for metrics like Mean Absolute Error (MAE) and Root Mean Square Error (RMSE).

• b.

  Residual error analysis: The tool used is Kernel Density Estimation plots for prediction errors to visualize the model bias and variance.

• c.

  Wilcoxon signed-rank test: It is used to compare the proposed model’s performance against baseline models like Support Vector Machine (SVM) and Recurrent Neural Network (RNN).

• d.

  Training time vs. model complexity: The parameters are evaluated are runtime, memory usage, and model complexity to determine the deployment and feasibility of the model.

• e.

  Learning curve analysis: A plot between training and validation loss over epochs is used to ensure the model converges smoothly and avoids overfitting.

• f.

  Bland-Altman plot: It is used to compare predicted RUL to actual RUL to visualize agreement and systematic bias.

Mathematical analysis of the proposed method

Although LSTM networks can be implemented in an end-to-end fashion, their performance can degrade in the presence of redundant or noisy features. In this study, RF was first applied to rank feature importance and select the most informative predictors. This reduces the input dimensionality, accelerates training, and enhances the model’s robustness. The selected features are subsequently used by the LSTM to capture temporal dependencies. The combination of RF and LSTM therefore leverages both interpretability and deep sequence modeling, yielding improved accuracy and stability compared with conventional end-to-end LSTM approaches. The mathematical formulation of the RF–LSTM framework is discussed:

Problem setup and notation

Let a battery degradation time series be:

$$\{{\mathrm{x}}_{\mathrm{t}},{\mathrm{y}}_{\mathrm{t}}{\}}^{\mathrm{T}}\mathrm{t}=1,\mathrm{w}\mathrm{h}\mathrm{e}\mathrm{r}\mathrm{e}{\mathrm{x}}_{\mathrm{t}}\in {\mathrm{R}}^{\mathrm{d}}\mathrm{a}\mathrm{n}\mathrm{d}{\mathrm{y}}_{\mathrm{t}}\in \mathrm{R}.$$where x_t contains d telemetry/derived features (e.g., voltage, current, etc.) and y_t is the target (SoH).

The sliding windows of length w is used to predict horizon h:

(1) $${\mathrm{X}}_{\mathrm{t}}=[{\mathrm{x}}_{\mathrm{t}-\mathrm{w}+1,\dots ,}{x}_t]\in {\mathrm{R}}^{\mathrm{w}\times \mathrm{d}},{\hat{\mathrm{y}}}_{\mathrm{t}+\mathrm{h}}={\mathrm{f}}_{\mathrm{\theta }}({\mathrm{X}}_{\mathrm{t}}).$$

Before modeling, features are standardized

(2) $${\tilde{\mathrm{x}}}^{(\mathrm{j})}=({\mathrm{x}}^{(\mathrm{j})}-{\mathrm{\mu }}_{\mathrm{j}})/{\sigma }_{\mathrm{j}},\mathrm{j}=1,\dots ,\mathrm{d}.$$

RF-based feature importance and selection

Train a Random Forest regressor F with T_f trees. The importance of feature j is:

(3) $$\mathrm{I}\mathrm{m}\mathrm{p}(\mathrm{j})=(1/\mathrm{T}\mathrm{f})\sum _{l=1}^{Tf}\sum _{\mathrm{s}\in \mathrm{S}\ell \left(\mathrm{j}\right)}^n(\mathrm{N}\mathrm{s}/\mathrm{N}\mathrm{r}\mathrm{o}\mathrm{o}\mathrm{t})\mathrm{\Delta }{\mathrm{I}}_s$$where s $\ell$ (j) are the split nodes using feature j in tree $\ell$, N_s is the sample count at node s, N_root is the root count.

Permutation importance for robustness:

(4) $$\mathrm{P}\mathrm{I}\mathrm{m}\mathrm{p}(\mathrm{j})=(1/\mathrm{B})\sum _{\mathrm{b}=1}^B{\mathrm{L}}_{\mathrm{v}\mathrm{a}\mathrm{l}}({\mathrm{\pi }}_{\mathrm{j}}^{(\mathrm{b})}\mathrm{X})-{\mathrm{L}}_{\mathrm{v}\mathrm{a}\mathrm{l}}(\mathrm{X}))$$where π_j^(b) randomly permutes column j on the validation set and L_val is the validation loss.

Select the top-m features:

(5) $${\mathrm{J}}_{\mathrm{m}}=\mathrm{a}\mathrm{r}\mathrm{g}.\mathrm{t}\mathrm{o}\mathrm{p}.{\mathrm{m}}_{\mathrm{j}}\mathrm{I}\mathrm{m}\mathrm{p}(\mathrm{j}),\mathrm{w}\mathrm{i}\mathrm{t}\mathrm{h}\mathrm{m}\mathrm{a}\mathrm{s}\mathrm{k}\mathrm{M}\in \{0,1{\}}^{\mathrm{d}}.$$

Apply the selection mask to each window:

(6) $${{\mathrm{X}}_{\mathrm{t}}}^{(\mathrm{m})}={\mathrm{X}}_{\mathrm{t}}{\mathrm{D}}^{\mathrm{M}},{\mathrm{D}}_{\mathrm{M}}=\mathrm{d}\mathrm{i}\mathrm{a}\mathrm{g}(\mathrm{M}).$$

LSTM sequence model

Given reduced window ${\mathrm{X}}_{\mathrm{t}}^{(\mathrm{m})}=[{\tilde{\mathrm{x}}}_{\{\mathrm{t}-\mathrm{w}+1{\}}^{(\mathrm{m})}},\dots ,{\tilde{\mathrm{X}}}_{\mathrm{t}}^{(\mathrm{m})}]$, the LSTM recurrence for time step τ (within window) is:

(7) $$\begin{array}{c}{\mathrm{i}}_{\tau }=\sigma ({\mathrm{W}}_{\mathrm{i}}{\tilde{\mathrm{x}}}_{\tau }^{(\mathrm{m})}+{\mathrm{U}}_{\mathrm{i}}{\mathrm{h}}_{\{\tau -1\}}+\mathrm{b}\mathrm{i})\hfill \\ {\mathrm{f}}_{\tau }=\sigma ({\mathrm{W}}_{\mathrm{f}}{{\tilde{\mathrm{x}}}_{\tau }}^{(\mathrm{m})}+{\mathrm{U}}_{\mathrm{f}}{\mathrm{h}}_{\{\tau -1\}}+{\mathrm{b}}_{\mathrm{f}})\hfill \\ {\mathrm{o}}_{\tau }=\sigma ({\mathrm{W}}_{\mathrm{o}}{\tilde{\mathrm{x}}}_{\tau }^{(\mathrm{m})}+{\mathrm{U}}_{\mathrm{o}}{\mathrm{h}}_{\{\tau -1\}}+{\mathrm{b}}_{\mathrm{o}})\hfill \\ \tilde{\mathrm{c}}=\mathrm{t}\mathrm{a}\mathrm{n}\mathrm{h}({\mathrm{W}}_{\mathrm{c}}\tilde{\mathrm{x}}\mathrm{\_}{\tau }^{(\mathrm{m})}+{\mathrm{U}}_{\mathrm{c}}{\mathrm{h}}_{\{\tau -1\}}+{\mathrm{b}}_{\mathrm{c}})\hfill \\ {\mathrm{c}}_{\tau }={\mathrm{f}}_{\tau }\odot {\mathrm{c}}_{\{\tau -1\}}+{\mathrm{i}}_{\tau }\odot \mathrm{i}{\tilde{\mathrm{c}}}_{\tau }\hfill \\ {\mathrm{h}}_{\tau }={\mathrm{o}}_{\tau }\odot \mathrm{t}\mathrm{a}\mathrm{n}\mathrm{h}({\mathrm{c}}_{\tau })\hfill \end{array}$$where σ(.) is the logistic sigmoid, $\odot$ is the element wise product, and W_*, U_*, b_*, are learnable parameters.

A linear readout maps the last hidden state to the prediction:

(8) $${\hat{\mathrm{y}}}_{\{\mathrm{t}+\mathrm{h}\}}={\mathrm{w}}^{\mathrm{T}}{\mathrm{h}}_{\mathrm{w}}+\mathrm{b}.$$

Learning objective and training

Use mean-squared error with L₂ regularization:

(9) $$\mathrm{L}(\theta )=1/|{\mathrm{T}}_{\mathrm{t}\mathrm{r}\mathrm{a}\mathrm{i}\mathrm{n}}|.{\sum }_{\mathrm{t}\in \mathrm{T}\mathrm{\_}\mathrm{t}\mathrm{r}\mathrm{a}\mathrm{i}\mathrm{n}}({\hat{\mathrm{y}}}_{\{\mathrm{t}+\mathrm{h}\}}-{\mathrm{y}}_{\{\mathrm{t}+\mathrm{h}\}}{)}^2+\lambda \Vert \mathrm{\theta }{\Vert }^2.$$

Optimized by Adam. Here θ denotes all LSTM and readout parameter λ is L₂ regularization.

To mitigate randomness, average across R independent runs:

$$\overline{\mathrm{M}}=(1/\mathrm{R}){\sum }_{\{\mathrm{r}=1\}}^{\mathrm{R}}{\mathrm{M}}^{(\mathrm{r})},$$

(10) $$\mathrm{S}\mathrm{D}(\mathrm{M})=\mathrm{s}\mathrm{q}\mathrm{r}\mathrm{t}((1/\mathrm{R}-1)){\sum }_{\{\mathrm{r}=1\}}^{\mathrm{R}}{({\mathrm{M}}^{(\mathrm{r})}-\overline{\mathrm{M}})}^2)$$for each metric $\mathrm{M}\in \{\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E},\mathrm{M}\mathrm{A}\mathrm{E},{\mathrm{R}}^2\}$

Evaluation metrics

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\mathrm{s}\mathrm{q}\mathrm{r}\mathrm{t}((1/\mathrm{N})\sum ({\hat{\mathrm{y}}}_{\mathrm{i}}-{\mathrm{y}}_{\mathrm{i}}{)}^2),$$

(11) $$\mathrm{M}\mathrm{A}\mathrm{E}=(1/\mathrm{N}){\sum }_{\mathrm{i}=1}^{\mathrm{N}}|{\hat{\mathrm{y}}}_{\mathrm{i}}-{\mathrm{y}}_{\mathrm{i}}|$$

(12) $${\mathrm{R}}^2=1-({\sum }_{\mathrm{i}-1}^{\mathrm{N}}({\hat{\mathrm{y}}}_{\mathrm{i}}-{\mathrm{y}}_{\mathrm{i}}{)}^2/{\sum }_{\mathrm{i}=1}^{\mathrm{N}}({\mathrm{y}}_{\mathrm{i}}-\overline{\mathrm{y}}{)}^2).$$

End-to-end RF–LSTM pipeline (summary)

• Standardize features via Eq. (2).

• Train RF, compute importance (Eq. 3) or permutation importance (Eq. 4).

• Select top-m features and mask inputs (Eqs. (5), (6)).

• Train LSTM on reduced windows using Eqs. (7)–(9).

• Report mean ± SD across runs (Eq. 10) and metrics (Eqs. (11), (12)).

Data set used

The dataset used in this study is obtained from Saha & Goebel (2007) (NASA Prognostics Data Repository/Battery Data Set) which provides detailed charge-discharge cycles of lithium-ion batteries under varying operational conditions. The data includes voltage, current, temperature, cycle count, internal resistance, and capacity, with time-stamped logs used for RUL prediction. The dataset is pre-processed for missing values, normalization, and split into training and test sets in a 70:30 ratio. The pre-processing of the data is done as mentioned in ‘Data Acquisition and Pre-processing’. The features considered after pre-processing are: Voltage, Current, Temperature, Cycle Count, Internal Resistance, Charge Capacity, Discharge Capacity.

AI application

All code, documentation, and the dataset used in this work are provided as Supplemental Materials and publicly accessible to ensure reproducibility. The full experimental pipeline, including preprocessing, feature extraction, model training, and statistical evaluation, can be executed using the provided Python scripts and README guidelines.

Reproducibility

The complete codebase, data preprocessing steps, model training, evaluation, and statistical validation scripts are shared as Supplemental Materials. A detailed README is included, explaining how to run the code, model architecture, and environment dependencies. All experiments can be reproduced using the full_workflow.py script, which automates data loading, feature selection, model training (LSTM + RF + XGBoost), and evaluation.

Results of feature selection

To enhance predictive performance and reduce model complexity, a dual-stage feature selection process is applied. Recursive Feature Elimination (RFE) is used with a RF estimator to iteratively eliminate the irrelevant features and then Pearson Correlation Analysis is done. It filters highly correlated variables, ensuring that retained features contribute unique information and reduce redundancy.

a. Recursive Feature Elimination (RFE)

The RFE reduced the features of the battery from 9 to 5. The top features retained are: discharge capacity, charge capacity, temperature, internal resistance and cycle count. The feature importance scores are calculated before implementing RFE which is given in Table 1 and Fig. 2. The comparison of the proposed model performance before and after applying RFE is given in Tables 2 and 3.

The first five features cumulatively contribute over 89% of the total importance score. The feature important scores before and after RFE implementation indicates that RFE not only reduced the dimensionality but also improve the accuracy and reduced training time by ~29.4%.

b. Pearson correlation matrix

These features align well with known electrochemical degradation indicators. For the validation of the results, Table 4 showing pairwise Pearson correlation coefficients between RUL and all selected features.

The correlation strength is obtained based of the following ( |r| value)

If |r| ranges between

0.00–0.10: Negligible

0.10–0.30: Very Weak

0.30–0.50: Weak

0.50–0.70: Moderate

0.70–0.90: Strong

0.90–1.00: Very Strong

The discharge capacity has the highest correlation with RUL (|r| = +0.261), but still falls in the weak range. The charge capacity shows a weak negative correlation, indicating that as capacity degrades, RUL decreases. Most other features fall in the very weak or negligible range, supporting their elimination during feature selection.

The heatmap showing the Pearson correlation of each feature with the RUL is given in Fig. 3. Warmer colors indicate positive correlations, and cooler tones indicate negative correlations. This visual emphasizes that discharge capacity and charge capacity are the most relevant features, justifying their selection in the model. The Pearson correlation matrix of all the features is given in Fig. 4. It visually represents how each feature correlates with every other, helping to identify the redundancy by highly correlated pairs and key predictors of RUL.

Results of temporal feature extraction using LSTM

The temporal features are extracted using LSTM with different learning rates (0.01, 0.001, 0.005), batch size (32, 64) and epochs (50, 100, 200). The MAE, RMSE and R² is calculated. The results obtained using Grid search hyper parameter optimization are given in Table 5. The best performance is obtained for set 3 with MAE reduced by 28.9% compared to baseline LSTM and High R² (0.93) indicates strong time-series learning.

Results of structured feature learning using Random Forest

After extracting temporal features using LSTM, RF is applied to the structured features for interpretability and complementary learning. Table 6 is the evidence to support the claim. From Table 6 it is shown that the top three features cumulatively contribute 80% to the model’s decision-making process which is represented in Fig. 5.

The ablation study is carried out with all the five structured features and with top 3 features and without top features. The performances like MAE, RMSE and R² are compared and is given in Table 7. From the results, it is shown that the charge capacity plays a critical role in performance, on removing this feature MAE is increased by ~13.6%.

Results of final RUL prediction using XGBoost (ensemble layer)

The performance of the proposed hybrid model is benchmarked against standalone implementations of LSTM, RF, and XGBoost in Table 8. The ensemble strategy fuses the strengths of LSTM (temporal modeling), RF (feature interpretability), and XGBoost (nonlinear regression).

The proposed hybrid model achieves a lowest MAE of 0.084, lowest RMSE of 0.138 and highest R² (goodness of fit) Score of 0.95. The results of proposed model is compared with the best individual model (LSTM). It is evident that MAE improves by 7.7%, RMSE improves by 6.8%, R² increases from 0.93 to 0.95, showing better model fit. The results confirm that ensemble learning effectively balances interpretability, temporal learning, and predictive power. The results given in Table 8 is represented in graphical form in Fig. 6 to have better understanding and comparison of the performance gains against benchmarking standalone model. The results confirm that proposed hybrid ensemble model is an excellent data fit model for RUL prediction. Although the performance gains in terms of average RMSE and MAE are moderate, the RF-LSTM consistently outperformed baseline models in repeated experiments. Over five independent runs, the standard deviation of RMSE was reduced from 0.012 (LSTM) to 0.006 (RF-LSTM), demonstrating improved robustness. Such stability is crucial in practical BMS applications where noisy or variable conditions often degrade prediction accuracy.

Results for statistical validation

A statistical validation of the results is done to validate the robustness and accuracy of the proposed model. The following are the statistical validation tools used:

• Confidence interval estimation

• Residual error analysis

• Wilcoxon signed-rank test

• Training time vs. complexity

• Learning curve analysis

• Bland-Altman plot

a. Results of Confidence interval estimation

MAE is calculated with 95% CI interval and the proposed hybrid model is compared with SVM, RNM and LSTM. The CI estimation is given in Table 9.

b. Residual error distribution analysis

The residual error distribution of the proposed model is compared with other models in Fig. 7. It shows that the residual error distribution is tight, symmetrical and centered around zero. The residual spread is narrower when compared to other baseline models, indicating the consistent performance.

c. Results of Wilcoxon signed-rank test

The Wilcoxon signed-rank test is done and the proposed hybrid method is compared with other models like SVM, RNN and LSTM. The results presented in Table 10 shows the superior performance of the proposed model.

d. Results of training time and complexity

The training time and the complexity of the proposed model is computed and compared with other models in Table 11. It can be seen that the slight increase in complexity yields significantly better performance.

e. Learning curve analysis

The learning curve is obtained for different epochs and a comparison is made in Table 12 with the proposed model over other models. A learning curve plots shown in Fig. 8 gives the training loss and validation loss vs. epochs. The key sign of a good model is that training and validation loss decreases together and stabilize and no wide gap (overfitting) or upward trend in validation loss (underfitting).

The loss metric considered is normalized MSE. The proposed model shows the lowest training and validation loss even at 100 epochs, indicating early convergence and better generalization. The classical models (SVM) show no epoch-based updates, so losses are fixed. The proposed model shows a steady and parallel drop in both losses. The proposed model achieves lowest final validation loss of 0.036. But RNN has a persistent gap. It means slight overfitting whereas LSTM is close but still lags the hybrid model.

Figure 8 shows the consistent and fast convergence after 120 epochs (vs. 160 for LSTM, 200+ for RNN) and No sign of overfitting or under fitting.

f. Results of Bland-Altman plot

From Fig. 9, it is shown that Mean Bias = +0.002 is the average difference between predicted and actual RUL. A small positive bias (+0.002) means, on average, the model slightly over estimates the RUL, but only by 0.2%, which is negligible. Ideally, bias should be close to zero.

95% limits of agreement (–0.042 to +0.046): This means that 95% of prediction errors lie within the range of –4.2% to +4.6% of the actual RUL. The narrow interval indicates that the variation in prediction errors is very small. This suggests the model is highly consistent and reliable in its predictions across different battery samples.

A visual summary of these results: performance comparison plot, residual plots, Bland-Altman plot is shown in Fig. 9.

g. Ablation study

Ablation study is carried out to quantify the contribution of each model in the pipeline. The study is conducted by removing one model out of three and the performance on the original test split is measured and is given in Table 13 and Fig. 10. Removing RF-based feature selection increases error (worse RMSE/MAE). Similarly, removing the XGBoost ensemble (so only RF + LSTM) also slightly degrades performance. The ablation indicates that both RF and XGBoost contribute positively. RF reduces noise/redundancy and XGBoost improves nonlinear fusion and final regression.

h. Five-fold cross-validation (stability)

To measure generalization variance across unseen splits, five-fold cross-validation (RMSE ± SD) is performed. The results obtained are shown in Fig. 11. The figure shows that proposed RF–LSTM–XGBoost model has RMSE of 0.138 ± 0.006 and LSTM (standalone) has RMSE of 0.148 ± 0.012 and XGBoost has RMSE of 0.161 ± 0.010. The proposed model not only achieves lower error, but also displays smaller run-to-run variance, indicating better stability and generalization.

Results for experimental validation

An extended set of experiments are performed to evaluate (a) model performance across different charge–discharge profiles and temperatures variations (b) the contribution of each model in the hybrid pipeline (RF, LSTM, XGBoost) via ablation study and (c) generalization via cross validation and perturbed (simulated) test splits. The additional physical testbeds are not available to emulate realistic complex operating conditions, so a reproducible simulation driven augmentation of the NASA dataset is used.

The NASA battery dataset used are perturbed as follows to emulate realistic variations:

• Charge and discharge profile variation is done by scaling the current time-series by factors sampled from the range ±10–30% to emulate fast/slow charge-discharge regimes.

• Temperature fluctuation is introduced by temperature ramps and cycle-level random offsets in the range ±5–15 °C to simulate thermal variability.

• Data degradation is considered by injecting the Gaussian noise with standard deviation up to 3% of the signal range and randomly set to 8% for missing entries.

These perturbations are applied only to the test data (not the training split) to measure out-of-distribution robustness.

The RMSE is evaluated with original and perturbed data and is provided in Table 14 and Fig. 12. From the table it is evident that the proposed RF–LSTM–XGBoost model exhibits the smallest relative degradation under perturbations approximately 2.5% increase in RMSE demonstrating improved robustness when compared to other single baseline model. This supports the claim that RF feature selection reduces noise propagation to the LSTM, and the ensemble layer stabilizes predictions under perturb/new conditions.

Key contributions

The hybrid RF–LSTM–XGBoost framework offers a statistically validated, generalizable, and computationally efficient solution for lithium-ion battery RUL prediction. It consistently outperforms conventional models, maintains accuracy under complex operating scenarios, and demonstrates clear innovation by combining interpretable feature selection with deep temporal learning and ensemble fusion. The proposed method achieves an optimal balance of accuracy, interpretability, and computational efficiency, making it suitable for integration into real-time BMS applications.