The accurate and reliable prediction of the Remaining Useful Life (RUL) of lithium-ion batteries is a cornerstone of modern Battery Management Systems (BMS), with profound implications for safety, performance optimization, and economic lifecycle management in applications ranging from electric vehicles to grid-scale energy storage. As a lithium-ion battery ages, complex internal degradation mechanisms, such as solid electrolyte interphase (SEI) layer growth, active material loss, and lithium plating, lead to capacity fade and power capability reduction. Predicting the point at which a battery will no longer meet operational requirements is therefore critical. Among various diagnostic techniques, Electrochemical Impedance Spectroscopy (EIS) has emerged as a powerful, non-destructive tool that provides a rich, frequency-domain fingerprint of the internal electrochemical processes. The EIS spectrum reflects kinetic and transport phenomena like charge transfer and lithium-ion diffusion, which are intrinsically linked to the battery’s State of Health (SOH) and its degradation trajectory. This article explores a data-driven approach to lithium-ion battery RUL prediction by leveraging the information-dense nature of EIS data.
Traditional model-based approaches for lithium-ion battery prognostics often rely on equivalent circuit models or detailed electrochemical models. While physically insightful, these models can be challenging to parameterize accurately across diverse operating conditions and throughout a battery’s entire lifespan. Data-driven methods, particularly those employing deep learning, offer a compelling alternative. They can learn complex, non-linear mappings from input features (like EIS spectra) to a target (like RUL) without requiring explicit prior knowledge of the degradation physics. However, EIS data presents unique challenges: it is high-dimensional, sequential (evolving over cycles), and contains multi-scale information from different frequency regions corresponding to different physical processes. Capturing the spatiotemporal dependencies within this data is key to building a robust prediction model.
This work proposes a hybrid deep learning architecture specifically designed to address these challenges for lithium-ion battery RUL prediction. The model synergistically combines a Convolutional Neural Network (CNN) for spatial/spectral feature extraction, a Bidirectional Long Short-Term Memory (BiLSTM) network for capturing temporal dependencies in the degradation sequence, and an Attention mechanism to dynamically weigh the importance of different time steps and features. This CNN-BiLSTM-Attention model is designed to automatically learn the deep, coupled relationship between the evolving EIS characteristics of a lithium-ion battery and its remaining useful life.
Methodology and Model Architecture
The core of our prognostic framework is the processing of sequential EIS data acquired at regular intervals throughout the aging test of a lithium-ion battery. Each EIS measurement, typically a vector of impedance magnitudes and phases across a log-sampled frequency range, serves as a snapshot of the battery’s internal state at a specific cycle number. The target value is the battery’s actual capacity or a derived RUL metric at that cycle. The goal is to train a model that can take a sequence of past EIS measurements and accurately predict the future capacity or RUL.
1. Data Preprocessing and Feature Engineering
Raw EIS data requires preprocessing to be suitable for neural network training. A common first step is normalization to ensure stable and faster convergence. We apply Z-score normalization to each feature (e.g., real impedance, imaginary impedance at each frequency) across the entire dataset:
$$ y_i = \frac{x_i – \mu}{\sigma} $$
where $x_i$ is the original feature value, $\mu$ is the mean of that feature across all samples, $\sigma$ is its standard deviation, and $y_i$ is the normalized value.
To structure the data for sequence modeling, we construct input samples using a sliding window approach. For a sequence length of $T$, an input sample $X_t$ at time $t$ is formed from the EIS measurements from cycle $t-T+1$ to cycle $t$. The corresponding target $y_t$ is typically the capacity at cycle $t$ or the calculated RUL. This format explicitly provides the model with recent historical context for making a prediction.
2. Hybrid CNN-BiLSTM-Attention Model
Our proposed model consists of three interconnected modules, each addressing a specific aspect of the learning problem. The advantages of different neural network components for lithium-ion battery data are summarized below.
| Network Component | Primary Function | Advantage for Lithium-Ion Battery Data |
|---|---|---|
| Convolutional Neural Network (CNN) | Spatial/spectral feature extraction | Automatically learns local patterns and hierarchical features from the EIS spectrum (e.g., semicircle diameters, slope of Warburg tail) without manual feature engineering. |
| Bidirectional LSTM (BiLSTM) | Temporal dependency modeling | Captures long-range dependencies in the degradation sequence by processing data in both forward and backward directions, understanding how past and future EIS states influence the current health. |
| Attention Mechanism | Dynamic feature weighting | Identifies and focuses on the most informative time steps or cycle intervals in the input sequence (e.g., rapid degradation phases), improving model interpretability and robustness to noise. |
2.1 Convolutional Feature Extractor
The normalized EIS sequence window $X_t$ is first passed through one or more 1D convolutional layers. A 1D CNN is well-suited for signal data like EIS. Each convolutional layer applies multiple filters (kernels) that slide across the frequency dimension of the input. This operation allows the network to detect local patterns—such as the shape of the high-frequency semicircle (related to charge transfer resistance) or the low-frequency Warburg region (related to diffusion)—that are indicative of specific degradation modes in the lithium-ion battery. The operation for a single filter at layer $l$ is:
$$ \mathbf{z}_t^l = \mathbf{W}^l * \mathbf{a}_t^{l-1} + \mathbf{b}^l $$
$$ \mathbf{a}_t^l = f(\mathbf{z}_t^l) $$
where $*$ denotes the convolution operation, $\mathbf{W}^l$ and $\mathbf{b}^l$ are the learnable kernel weights and bias for layer $l$, $\mathbf{a}_t^{l-1}$ is the activation from the previous layer (with $\mathbf{a}_t^0 = X_t$), and $f$ is a non-linear activation function like ReLU. Pooling layers may follow to reduce dimensionality and provide translational invariance. The output is a refined feature map $\mathbf{C}_t$ that encodes significant spectral patterns.
2.2 Bidirectional Sequence Encoder
The feature maps $\mathbf{C}_t$ for each time step in the window are then fed into a BiLSTM layer. A standard LSTM unit addresses the vanishing gradient problem in RNNs through gating mechanisms. For a single time step $t$, the computations are:
$$ \begin{aligned}
\text{Forget Gate: } & \mathbf{f}_t = \sigma(\mathbf{W}_f \cdot [\mathbf{h}_{t-1}, \mathbf{C}_t] + \mathbf{b}_f) \\
\text{Input Gate: } & \mathbf{i}_t = \sigma(\mathbf{W}_i \cdot [\mathbf{h}_{t-1}, \mathbf{C}_t] + \mathbf{b}_i) \\
\text{Candidate State: } & \tilde{\mathbf{c}}_t = \tanh(\mathbf{W}_c \cdot [\mathbf{h}_{t-1}, \mathbf{C}_t] + \mathbf{b}_c) \\
\text{Cell State Update: } & \mathbf{c}_t = \mathbf{f}_t \odot \mathbf{c}_{t-1} + \mathbf{i}_t \odot \tilde{\mathbf{c}}_t \\
\text{Output Gate: } & \mathbf{o}_t = \sigma(\mathbf{W}_o \cdot [\mathbf{h}_{t-1}, \mathbf{C}_t] + \mathbf{b}_o) \\
\text{Hidden State: } & \mathbf{h}_t = \mathbf{o}_t \odot \tanh(\mathbf{c}_t)
\end{aligned} $$
where $\sigma$ is the sigmoid function, $\odot$ is the Hadamard product, $\mathbf{W}$ and $\mathbf{b}$ are learnable weights and biases. A BiLSTM runs two independent LSTM layers—one processing the sequence forward and one backward—and concatenates their hidden states: $\overrightarrow{\mathbf{h}}_t$ and $\overleftarrow{\mathbf{h}}_t$. The final output for each time step is $\mathbf{H}_t = [\overrightarrow{\mathbf{h}}_t; \overleftarrow{\mathbf{h}}_t]$. This allows the model to integrate information from both past and future contexts within the window, which is crucial for understanding the degradation trajectory of the lithium-ion battery.
2.3 Attention Mechanism
The sequence of BiLSTM outputs $\mathbf{H} = [\mathbf{H}_1, \mathbf{H}_2, …, \mathbf{H}_T]$ contains encoded information for all time steps. However, not all cycles contribute equally to the current health state prediction. An attention mechanism computes a set of weights that signify the importance of each time step. First, a context vector $\mathbf{u}$ is often used to calculate alignment scores:
$$ e_t = \mathbf{v}^T \tanh(\mathbf{W}_a \mathbf{H}_t + \mathbf{b}_a) $$
where $\mathbf{W}_a$, $\mathbf{b}_a$, and $\mathbf{v}$ are learnable parameters. These scores are normalized using a softmax function to obtain the attention weights $\alpha_t$:
$$ \alpha_t = \frac{\exp(e_t)}{\sum_{j=1}^{T} \exp(e_j)} $$
A weighted context vector $\mathbf{c}$ is then computed as the summary of the entire input sequence:
$$ \mathbf{c} = \sum_{t=1}^{T} \alpha_t \mathbf{H}_t $$
This context vector $\mathbf{c}$, which emphasizes the most relevant historical states of the lithium-ion battery, is passed to the final output layer.
2.4 Output and Training
The final prediction for the lithium-ion battery’s RUL (or capacity) is generated by one or more fully connected (dense) layers:
$$ \hat{y} = \mathbf{W}_o \cdot \mathbf{c} + \mathbf{b}_o $$
The entire CNN-BiLSTM-Attention model is trained end-to-end by minimizing the difference between the predicted RUL $\hat{y}$ and the true RUL $y$. A common loss function is the Mean Squared Error (MSE):
$$ \mathcal{L} = \frac{1}{N} \sum_{i=1}^{N} (y_i – \hat{y}_i)^2 $$
The model is optimized using algorithms like Adam. The hyperparameters used in our implementation are summarized below.
| Hyperparameter | Value/Range |
|---|---|
| Input Sequence Length (T) | 10-50 cycles |
| CNN Filters | 32, 64 |
| Kernel Size | 3 |
| BiLSTM Units | 50-100 |
| Attention Dimension | Equal to BiLSTM output |
| Dropout Rate | 0.2-0.5 |
| Optimizer | Adam |
| Learning Rate | 1e-3 to 1e-4 |
| Batch Size | 32-64 |
Experimental Results and Model Evaluation
To validate the effectiveness of the proposed CNN-BiLSTM-Attention model for lithium-ion battery RUL prediction, we conducted experiments using aging datasets that include periodic EIS measurements and capacity fade records. The dataset was split into training and testing sets to evaluate both learning capability and generalization performance. We compared our model against several baseline data-driven models commonly used in time-series prediction.
1. Comparative Analysis of Prediction Performance
We benchmarked the CNN-BiLSTM-Attention model against a standard Backpropagation (BP) Neural Network (a simple multi-layer perceptron), a standalone Long Short-Term Memory (LSTM) network, and a CNN-LSTM model (without bidirectional layers or attention). The BP network treats each EIS sample independently, ignoring temporal dependencies. The LSTM model captures sequential information but lacks powerful feature extraction from the complex EIS spectra. The CNN-LSTM model improves feature learning but does not utilize bidirectional context or adaptive weighting. The prediction curves on the test set clearly demonstrate the superiority of the proposed hybrid approach in tracking the true degradation trajectory of the lithium-ion battery.
The quantitative performance was assessed using four standard regression metrics:
• Root Mean Square Error (RMSE): $RMSE = \sqrt{\frac{1}{N}\sum_{i=1}^{N}(y_i – \hat{y}_i)^2}$
• Mean Absolute Error (MAE): $MAE = \frac{1}{N}\sum_{i=1}^{N}|y_i – \hat{y}_i|$
• Mean Absolute Percentage Error (MAPE): $MAPE = \frac{100\%}{N}\sum_{i=1}^{N}\left|\frac{y_i – \hat{y}_i}{y_i}\right|$
• Coefficient of Determination (R²): $R^2 = 1 – \frac{\sum_{i=1}^{N}(y_i – \hat{y}_i)^2}{\sum_{i=1}^{N}(y_i – \bar{y})^2}$
A lower RMSE, MAE, and MAPE indicate better accuracy, while an R² value closer to 1 indicates a better fit. The performance comparison on the test set is presented below.
| Model | RMSE | MAE | MAPE (%) | R² |
|---|---|---|---|---|
| BP Neural Network | 4.057 | 3.110 | 13.8 | 0.851 |
| LSTM Network | 3.951 | 3.280 | 14.5 | 0.859 |
| CNN-LSTM | 3.878 | 3.296 | 13.9 | 0.864 |
| CNN-BiLSTM-Attention (Proposed) | 2.711 | 2.524 | 11.0 | 0.940 |
The results clearly show that the proposed CNN-BiLSTM-Attention model achieves the best performance across all metrics. It significantly reduces prediction error (RMSE reduced by ~30% compared to the next best model) and achieves a much higher R² value, explaining 94% of the variance in the test data. This demonstrates its superior ability to model the complex, non-linear relationship between the evolving EIS characteristics and the remaining useful life of the lithium-ion battery.
2. In-depth Evaluation of the Proposed Model
Further analysis of the CNN-BiLSTM-Attention model’s performance provides insights into its robustness. On the training set, the model achieved an R² of 0.999 and an extremely low RMSE, indicating an excellent fit to the learned data without significant overfitting, thanks to regularization techniques like dropout. The primary strength is evidenced on the test set. The scatter plot of predicted versus true RUL values shows a tight clustering around the ideal line of unity, confirming high accuracy. The histogram of prediction errors approximates a normal distribution centered near zero, indicating no systematic bias in the model’s predictions for the lithium-ion battery lifespan.
The attention weights produced by the model can be visualized over the input sequence window. Typically, higher weights are assigned to time steps corresponding to cycles where the EIS features show significant changes, such as during the knee-point region of capacity fade or after a high-stress event. This interpretability aspect is a key advantage, as it allows engineers to understand which periods in the battery’s history the model deems most critical for its current health assessment.
Conclusion
Predicting the remaining useful life of lithium-ion batteries is essential for ensuring the safety, reliability, and economy of energy storage systems. This work presents a novel, data-driven prognostic approach that capitalizes on the rich diagnostic information contained in Electrochemical Impedance Spectroscopy (EIS) data. The proposed CNN-BiLSTM-Attention model is specifically architected to overcome the challenges of processing high-dimensional, sequential EIS data. The CNN layer acts as an automatic feature extractor, discerning critical patterns from the impedance spectrum that correlate with internal degradation mechanisms of the lithium-ion battery. The BiLSTM layer effectively captures the long-term temporal dependencies in the battery’s aging trajectory. Finally, the attention mechanism provides dynamic focus, allowing the model to weigh historical states differently and enhance its predictive focus on critical degradation phases.
Experimental results demonstrate that this hybrid model significantly outperforms traditional neural network architectures like BP, LSTM, and CNN-LSTM. It achieves higher prediction accuracy (lower RMSE, MAE, MAPE) and superior explanatory power (higher R²) on unseen test data. The model exhibits strong generalization capability and robustness, making it a promising tool for practical BMS applications. By providing accurate early warnings of impending lithium-ion battery failure, this methodology can facilitate proactive maintenance, extend usable life, and prevent catastrophic failures, thereby contributing to the sustainable and safe deployment of lithium-ion battery technology across various sectors.