The accurate prediction of the Remaining Useful Life (RUL) for lithium-ion batteries is a cornerstone for ensuring the reliability and safety of modern energy systems, from portable electronics to electric vehicles and grid-scale storage. As a critical energy storage component, the performance degradation of a li ion battery directly impacts system efficiency and operational safety. Traditional model-based approaches often struggle with the complex, nonlinear electro-chemical aging processes inherent to li ion batteries. Consequently, data-driven methods have gained significant prominence, leveraging historical operational data to model the battery’s fading trajectory without requiring intricate physical models. However, these methods face substantial hurdles: the scarcity of extensive, high-quality aging datasets, the presence of significant noise and capacity regeneration phenomena in recorded data, and the distinct, nonlinear degradation rates across different life stages of a li ion battery. This work addresses these interconnected challenges by proposing a holistic framework that synergizes data augmentation, integrated denoising and prediction, and strategic transfer learning to significantly enhance the accuracy and robustness of li ion battery RUL forecasts.
The core of our proposed methodology is a sophisticated neural architecture designed to learn from noisy, limited data. We begin by confronting the data scarcity issue. Publicly available li ion battery aging datasets, such as those from CALCE and NASA, typically contain only a handful of complete life-cycle records for a given battery type. Training deep learning models on such small datasets risks severe overfitting and poor generalization. To overcome this, we develop a hybrid generative model that combines the principled latent space learning of a Variational Autoencoder (VAE) with the high-fidelity data generation capability of a Generative Adversarial Network (GAN), termed VAE-GAN. This model ingests the limited real-world capacity fade curves and learns their underlying probability distribution. The encoder network maps an input capacity sequence x to parameters of a latent distribution (mean μ and variance σ²). The loss function for this encoding process includes a Kullback-Leibler (KL) divergence term to regularize the latent space:
$$ L_E = KL\left[ N(\mu, \sigma^2) \ || \ N(0,1) \right] = \frac{1}{2} \sum (1 + \log(\sigma^2) – \mu^2 – \sigma^2) $$
The decoder (acting as the GAN’s generator) then reconstructs data from this structured latent space, while a discriminator network judges the authenticity of the generated samples. The adversarial training ensures the generated capacity curves are realistic and diverse. The combined loss functions for the generator (G) and discriminator (D) are:
$$ L_G = \| x_i – \hat{x}_i \|^2 + E_{z \sim p(z)}[\log(1 – D(G(z)))] $$
$$ L_D = \frac{1}{2}E_{x \sim p_{data}}[(D(x)-1)^2] + \frac{1}{2}E_{z \sim p(z)}[(D(G(z)))^2] $$
Applying this VAE-GAN model to small public datasets allows for the generation of multiple, realistic full life-cycle degradation curves, effectively augmenting the training data available for the subsequent prediction model. The quality of generated data is quantified using a similarity metric, with results consistently above 0.92 when compared to real data, confirming their utility.
The second major challenge is the noisy nature of li ion battery capacity data. Fluctuations due to measurement errors, environmental variations, and the well-known capacity rebound effect can obscure the true degradation trend, leading to poor model learning. Instead of treating denoising and prediction as separate, sequential tasks, we propose a unified Denoising Autoencoder-Long Short-Term Memory (DAE-LSTM) network. This architecture performs feature extraction, noise suppression, and temporal modeling within a single end-to-end trainable model. The input is a windowed sequence of recent capacity measurements. A Denoising Autoencoder (DAE) first processes this sequence. Critically, its input is deliberately corrupted with additive Gaussian noise, forcing the encoder to learn robust feature representations that are invariant to such perturbations. The encoder compresses the noisy input x' into a lower-dimensional hidden vector z:
$$ z = a(\mathbf{W} \hat{x’} + \mathbf{b}) $$
The decoder then attempts to reconstruct a clean version of the input from this hidden vector:
$$ \hat{x} = f'(\mathbf{W’} z + \mathbf{b’}) $$
The key innovation is that this hidden vector z, which embodies a denoised and compressed representation of the input window, is not used for final reconstruction alone. Instead, it is fed directly into a stacked LSTM network. LSTMs are exceptionally suited for this task due to their gated memory cells, which can learn long-range temporal dependencies in the sequential degradation data of a li ion battery. The state updates for an LSTM cell are governed by:
$$
\begin{aligned}
i_t &= \sigma_g(\mathbf{W}_i x_t + \mathbf{U}_i h_{t-1} + \mathbf{b}_i) \\
f_t &= \sigma_g(\mathbf{W}_f x_t + \mathbf{U}_f h_{t-1} + \mathbf{b}_f) \\
o_t &= \sigma_g(\mathbf{W}_o x_t + \mathbf{U}_o h_{t-1} + \mathbf{b}_o) \\
\tilde{c}_t &= \tanh(\mathbf{W}_c x_t + \mathbf{U}_c h_{t-1} + \mathbf{b}_c) \\
c_t &= f_t \odot c_{t-1} + i_t \odot \tilde{c}_t \\
h_t &= o_t \odot \tanh(c_t)
\end{aligned}
$$
where i, f, o, c, h represent the input gate, forget gate, output gate, cell state, and hidden state, respectively. The final layers of the network map the LSTM’s output to a predicted capacity value for the next time step. Most importantly, the entire DAE-LSTM model is trained using a composite loss function that jointly optimizes for both clean reconstruction and accurate prediction:
$$ L = \alpha \sum_{t=1}^{n} \ell(x_t, \hat{x}_t) + \sum_{t=T}^{n} \ell(y_t, \hat{y}_t) + \lambda \Omega(\theta) $$
where the first term is the DAE reconstruction loss, the second term is the LSTM prediction loss (e.g., Mean Squared Error), α balances the two objectives, and the final term is a regularization penalty. This shared optimization ensures the features learned by the DAE are directly relevant and beneficial for the prediction task, creating a powerful symbiotic relationship within the model.
The final strategic component is transfer learning, which addresses the mismatch between data domains (synthetic vs. real) and the non-stationarity of degradation behavior across a li ion battery’s lifespan. We employ a two-stage training regimen. First, the DAE-LSTM model is pre-trained on the larger corpus of diverse, synthetic life-cycle data generated by the VAE-GAN. This allows the model to learn a robust, generalized representation of typical li ion battery degradation patterns, including various curve shapes and noise characteristics. In the second stage, the model is fine-tuned on the limited set of real, experimental data from the target battery or similar batteries. Crucially, we fine-tune all layers of the network (both DAE and LSTM components) with a very low learning rate. This allows the model to adapt its generalized knowledge to the specific noise distribution and precise degradation kinetics of the real-world data, significantly boosting prediction accuracy on the target task without requiring large amounts of real data.
| Component | Model | Key Function | Output |
|---|---|---|---|
| Data Augmentation | VAE-GAN | Learns data distribution & generates synthetic full-cycle curves | Multiple realistic capacity vs. cycle datasets |
| Denoising & Feature Extraction | DAE Encoder | Compresses noisy input window to robust latent vector | Lower-dimensional hidden vector `z` |
| Temporal Modeling & Prediction | LSTM Network | Learns sequential dependencies from latent vectors | Predicted capacity for next time step |
| Training Strategy | Transfer Learning | Pre-train on synthetic data, fine-tune on real data | Generalized & adapted model parameters |
We rigorously evaluated the proposed framework on two standard public datasets featuring different li ion battery chemistries: the CALCE CS2 series (LCO) and the NASA PCoE series (NCA). The prediction task used a rolling window approach: given capacity data up to cycle t, predict the capacity at cycle t+1, iteratively generating a full future trajectory until a predefined End-of-Life (EOL) threshold was crossed. We compared our full method (M3) against two ablated versions: a standard LSTM (M1) and the DAE-LSTM without transfer learning (M2). The hyperparameters for the DAE-LSTM model were optimized via grid search, with key settings summarized below:
| Hyperparameter | Search Space | Optimal Value (CALCE) | Optimal Value (NASA) |
|---|---|---|---|
| Input Window Size | [16, 32, 64] | 64 | 48 |
| Learning Rate | [1e-3, 1e-4, 1e-5] | 1e-4 | 1e-4 |
| LSTM Layers | [1, 2, 4] | 2 | 2 |
| Hidden Dimension | [16, 32, 64] | 64 | 64 |
| Noise Level (DAE) | [0.001, 0.01, 0.1] | 0.01 | 0.008 |
The results were assessed using standard metrics: Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and critically for RUL prediction, the Relative Error (RE) and Prediction Accuracy (Pre) defined as:
$$ RE = |RUL_{pred} – RUL_{true}| $$
$$ Pre = 1 – \frac{|RUL_{pred} – RUL_{true}|}{RUL_{true}} $$
On the CALCE dataset with an early prediction start point (cycle 364), our full method (M3) demonstrated superior performance. For instance, for battery CS2_37, M3 achieved an RE of only 3 cycles and a Pre of 0.989, significantly outperforming M1 (RE=21, Pre=0.894) and M2 (RE=20, Pre=0.891). The denoising capability was particularly evident for batteries with high fluctuations, like CS2_36. While all methods faced challenges with this atypical curve, M3 maintained the best robustness. The results for later prediction starts (e.g., cycle 464) further confirmed that prediction accuracy improves as more of the accelerated degradation phase is included in the training history, and our method consistently provided the most accurate and stable forecasts.
| Battery | Method | RE (cycles) | MAE | RMSE | Pre |
|---|---|---|---|---|---|
| CS2_37 | M1 (LSTM) | 21 | 0.0157 | 0.0221 | 0.894 |
| M2 (DAE-LSTM) | 20 | 0.0365 | 0.0416 | 0.891 | |
| M3 (Proposed) | 3 | 0.0090 | 0.0106 | 0.989 | |
| B5 (NASA) | M1 (LSTM) | 35 | 0.0594 | 0.0684 | 0.438 |
| M2 (DAE-LSTM) | 9 | 0.0287 | 0.0341 | 0.888 | |
| M3 (Proposed) | 2 | 0.0224 | 0.0286 | 0.971 |
The experiment on the NASA dataset, which has even shorter and noisier records, strongly validated the framework’s generalization ability. Our method successfully handled significant capacity rebounds and noise, achieving high prediction accuracy (Pre > 0.94 for all batteries) where baseline methods often failed. A statistical analysis of 100 independent runs for the NASA batteries yielded narrow 90% confidence intervals for the predicted RUL (e.g., [133, 139] for B5), underscoring the high robustness and reliability of the proposed approach. The total computational time for a complete prediction pipeline, including data generation, pre-training, fine-tuning, and testing, was under 200 seconds for the CALCE dataset and under 70 seconds for the NASA dataset, which is acceptable for offline prognostic applications.
In conclusion, this work presents a comprehensive, data-centric framework for enhancing the prediction of remaining useful life in li ion batteries. By systematically addressing the key practical challenges of data scarcity, signal noise, and domain adaptation, the proposed integration of VAE-GAN based data augmentation, a jointly optimized DAE-LSTM prediction model, and a two-stage transfer learning strategy delivers marked improvements in prediction accuracy, robustness, and generalizability. The method’s effectiveness is rigorously demonstrated on two distinct, benchmark li ion battery datasets. Future work will focus on architectural optimization to reduce computational latency for potential real-time applications and extending the framework to incorporate additional predictive features such as voltage and temperature profiles for an even more comprehensive health assessment of the li ion battery.