The rapid proliferation of the new energy sector has placed lithium-ion batteries at the forefront of technological advancement, serving as the indispensable power core for electric vehicles and large-scale energy storage systems. Within this context, the precise monitoring of performance degradation in these energy storage units emerges as a paramount technical challenge, critical for ensuring the safe, reliable, and efficient operation of modern energy infrastructure. Accurately estimating the State of Health (SOH) of a li ion battery is not merely a diagnostic task; it is a fundamental enabler for optimizing energy management strategies, proactively delaying the aging process, significantly reducing operational and maintenance costs, and ultimately minimizing resource waste and environmental impact. The pursuit of high-fidelity SOH evaluation systems is of particular engineering significance in the era of global decarbonization, as it underpins the realization of full life-cycle value for power batteries, contributing directly to sustainable energy goals.
Traditional methodologies for constructing SOH estimation models for li ion battery systems often encounter significant practical hurdles. Electrochemical models, which are grounded in the fundamental physics of battery operation, involve complex sets of partial differential equations. These models, while highly descriptive, demand substantial computational resources, making them unsuitable for real-time, embedded applications. On the other hand, Equivalent Circuit Models (ECMs) simplify the internal dynamics but require precise parameter identification. This identification process is notoriously sensitive to measurement noise, temperature fluctuations, and battery history, imposing stringent requirements on the operating environment and limiting their robustness across diverse application scenarios. These limitations have catalyzed the exploration of data-driven paradigms, where algorithms learn the complex, non-linear relationships between measurable battery signals and its underlying state of health.
This article explores the application of Long Short-Term Memory (LSTM) networks, a powerful variant of Recurrent Neural Networks (RNNs), for constructing a high-precision SOH estimation model for li ion battery. The core advantage of this approach lies in its ability to learn long-term temporal dependencies from historical operational data—such as voltage, current, and temperature sequences—without requiring an explicit understanding of the intricate internal electrochemical aging mechanisms or the formulation of complex mathematical expressions. The model autonomously discovers the mapping between evolving operational features and the battery’s capacity fade. We will delve into the architectural principles of LSTM networks, detail the design of an estimation framework utilizing publicly available battery aging datasets, and present a comprehensive simulation-based performance verification, comparing its efficacy against other traditional and data-driven methods.
1. The Operational Context and Degradation of Lithium-Ion Batteries
A li ion battery is a sophisticated electrochemical device where energy is stored and released through the movement of lithium ions between two electrodes. The quintessential structure, as illustrated below, comprises a positive electrode (cathode, typically a lithium metal oxide), a negative electrode (anode, typically graphite), a porous separator that prevents physical contact while allowing ionic conduction, and an electrolyte solution that facilitates ion transport.
The State of Health (SOH) is a critical metric quantifying the degree of degradation of a li ion battery relative to its pristine, fresh condition. It is most commonly defined as the ratio of the battery’s current maximum available capacity (C_actual) to its nominal or initial rated capacity (C_rated):
$$ SOH (\\%) = \\frac{C_{\\text{actual}}}{C_{\\text{rated}}} \\times 100\\% $$
As a li ion battery undergoes successive charge-discharge cycles, a multitude of irreversible side reactions occur. These include the growth of the Solid Electrolyte Interphase (SEI) layer on the anode, active material dissolution and structural disordering at the cathode, lithium plating, and electrolyte decomposition. These phenomena collectively lead to a gradual loss of cyclable lithium ions and active host materials, manifesting as a monotonic, yet often non-linear, decrease in available capacity and an increase in internal resistance. The end-of-life for a traction li ion battery is typically declared when its SOH degrades to 70-80% of its initial capacity, at which point its energy storage capability becomes insufficient for the intended application.
2. The LSTM Network: Architecture for Learning Long-Term Dependencies
Recurrent Neural Networks (RNNs) are a natural fit for sequential data like time-series from a li ion battery. However, standard RNNs suffer from the vanishing/exploding gradient problem, making it exceedingly difficult for them to learn and retain information from distant time steps—a crucial requirement for capturing the slow, long-term degradation trends in battery aging. The Long Short-Term Memory network, introduced by Hochreiter and Schmidhuber, ingeniously addresses this limitation through a gated cell structure.
The key innovation of the LSTM is its memory cell (\(c_t\)), which can maintain information over long periods. The flow of information into, within, and out of this cell is regulated by three specialized gates:
- Forget Gate (\(f_t\)): Determines what proportion of the previous cell state (\(c_{t-1}\)) should be discarded.
- Input Gate (\(i_t\)): Controls how much of the new candidate information (\(\tilde{c}_t\)) should be stored into the cell state.
- Output Gate (\(o_t\)): Governs how much of the current cell state (\(c_t\)) is used to compute the output activation (\(h_t\)).
These gates are implemented using sigmoid activation functions (\(\sigma\)), which output values between 0 and 1, representing “keep” or “discard” ratios. The mathematical formulation of an LSTM cell at time step \(t\) is as follows:
Gate Calculations:
$$ i_t = \\sigma(W_i \\cdot [h_{t-1}, x_t] + b_i) $$
$$ f_t = \\sigma(W_f \\cdot [h_{t-1}, x_t] + b_f) $$
$$ o_t = \\sigma(W_o \\cdot [h_{t-1}, x_t] + b_o) $$
Candidate Cell State:
$$ \\tilde{c}_t = \\tanh(W_c \\cdot [h_{t-1}, x_t] + b_c) $$
Cell State Update:
$$ c_t = f_t \\odot c_{t-1} + i_t \\odot \\tilde{c}_t $$
Hidden State (Output):
$$ h_t = o_t \\odot \\tanh(c_t) $$
Here, \(x_t\) is the input vector, \(h_{t-1}\) is the previous hidden state, \(W\) and \(b\) are learned weight matrices and bias vectors, \(\sigma\) is the sigmoid function, \(\tanh\) is the hyperbolic tangent function, and \(\odot\) denotes the Hadamard (element-wise) product. This architecture allows the LSTM network to dynamically learn which temporal features in the li ion battery operational history (e.g., voltage relaxation patterns, charge/discharge curve shapes over cycles) are relevant for predicting the long-term trajectory of SOH, effectively memorizing significant degradation events and filtering out noise.
3. Methodology: Designing an LSTM-Based SOH Estimation Framework
3.1 Data Source and Preprocessing
To validate the proposed approach, we utilize the publicly available lithium-ion battery aging dataset from the NASA Prognostics Center of Excellence. This benchmark dataset provides full charge-discharge cycle data for 18650 format li ion battery cells (LiCoO2 cathode, 2.1 Ah nominal capacity). The tests involve repeated cycles under controlled environmental conditions until the battery capacity falls below its end-of-life threshold. The data records time, voltage, current, and temperature at high frequency. The actual capacity (and thus SOH) for each cycle is calculated via coulomb counting during the discharge phase.
For this study, cells are grouped based on their operating conditions to evaluate model robustness:
| Group | Cell ID | Discharge Current (A) | Ambient Temperature (°C) | Discharge Cut-off Voltage (V) |
|---|---|---|---|---|
| A (Room Temp) | B05, B07, B18 | 2.0 | 24 | 2.7, 2.2, 2.5 |
| B (High Current) | B33, B34 | 4.0 | 24 | 2.0, 2.2 |
| C (Low Temp) | B46, B47, B48 | 1.0 | 4 | 2.2, 2.5, 2.7 |
Raw data often contains anomalies or missing points due to sensor errors. To ensure robust model training, we employ the Interquartile Range (IQR) method for outlier detection and removal. For a given feature vector, we calculate the first quartile (Q1, 25th percentile) and the third quartile (Q3, 75th percentile). Any data point lying outside the range \([Q1 – 1.5 \times IQR, Q3 + 1.5 \times IQR]\), where \(IQR = Q3 – Q1\), is considered an outlier and is discarded. This preprocessing step enhances the stability and generalization capability of the LSTM model trained on the li ion battery data.
3.2 Model Architecture and Training Strategy
The SOH estimation is formulated as a supervised sequential regression problem. The input to the LSTM model is a time-series window constructed from features extracted from one or several recent charge/discharge cycles. These features can include statistics from voltage curves, temperature profiles, time intervals for constant-voltage charging, etc. The output is the estimated SOH value for the subsequent cycle or for the end of the current window.
The specific LSTM model used here has an input layer that accepts sequential data, followed by one or more LSTM layers for temporal feature extraction. A final Dense (fully connected) layer outputs the single SOH estimate. The model is trained using past cycle data, with the goal of minimizing the difference between its predictions and the actual, calculated SOH values. We adopt a progressive training and testing scheme: the model is first trained on data from the early life of the li ion battery (e.g., the first 50% or 70% of cycles) and is then tasked with predicting the SOH for all remaining cycles in the battery’s life.
4. Simulation Results and Performance Analysis
The proposed LSTM-based framework was implemented in Python using the TensorFlow/Keras libraries. Simulations were conducted for cells across all three condition groups (A, B, C). For each cell, two training scenarios were evaluated: using the first 50% of cycle data for training and predicting the latter 50%, and using the first 70% for training to predict the remaining 30%. This tests the model’s ability to extrapolate the degradation trajectory with varying amounts of historical knowledge.
4.1 SOH Estimation Results
The estimation results consistently demonstrated the LSTM model’s capability to capture the non-linear capacity fade trajectory of a li ion battery. For cells in Group A (room temperature), the predicted SOH closely followed the actual measurements, especially when the training data ratio was increased to 70%. The model successfully learned the general aging trend despite minor capacity regeneration fluctuations. For Group B cells (high current), the initial convergence was slightly slower, reflecting the more strenuous operating condition’s impact on the degradation dynamics, yet the model adapted well with sufficient training data. For Group C (low temperature), the results confirmed that while low-temperature operation presents a distinct aging pattern, the data-driven LSTM model could effectively learn and predict it, with accuracy improving notably as more training cycles were provided.
4.2 Quantitative Performance Evaluation
To quantitatively assess the performance of the li ion battery SOH estimation model, three standard metrics are employed: Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and the Coefficient of Determination (R²).
$$ \\text{RMSE} = \\sqrt{\\frac{1}{N} \\sum_{i=1}^{N} (SOH_{\\text{actual}}^{(i)} – SOH_{\\text{estimated}}^{(i)})^2 } $$
$$ \\text{MAE} = \\frac{1}{N} \\sum_{i=1}^{N} | SOH_{\\text{actual}}^{(i)} – SOH_{\\text{estimated}}^{(i)} | $$
$$ R^2 = 1 – \\frac{\\sum_{i=1}^{N} (SOH_{\\text{actual}}^{(i)} – SOH_{\\text{estimated}}^{(i)})^2}{\\sum_{i=1}^{N} (SOH_{\\text{actual}}^{(i)} – \\bar{SOH}_{\\text{actual}})^2} $$
Where \(N\) is the number of test cycles, and \(\\bar{SOH}_{\\text{actual}}\) is the mean of the actual SOH values. RMSE penalizes larger errors more heavily, MAE provides a linear error score, and R² indicates the proportion of variance explained by the model (with 1.0 being a perfect fit).
The following tables summarize the performance metrics for selected cells under the 50% and 70% training scenarios:
| Group | Cell ID | RMSE | MAE | R² |
|---|---|---|---|---|
| A | B05 | 0.0223 | 0.0112 | 0.9914 |
| A | B18 | 0.0416 | 0.0228 | 0.9812 |
| B | B33 | 0.0178 | 0.0086 | 0.9803 |
| C | B46 | 0.0133 | 0.0068 | 0.9952 |
| Group | Cell ID | RMSE | MAE | R² |
|---|---|---|---|---|
| A | B05 | 0.0126 | 0.0092 | 0.9951 |
| A | B18 | 0.0081 | 0.0053 | 0.9972 |
| B | B33 | 0.0144 | 0.0091 | 0.9933 |
| C | B46 | 0.0063 | 0.0038 | 0.9972 |
The results clearly show that increasing the amount of training data consistently improves estimation accuracy across all metrics (lower RMSE/MAE, higher R²). The R² values are notably high (often >0.99), indicating that the LSTM model explains almost all the variance in the SOH degradation data for the li ion battery.
4.3 Comparative Analysis with Other Neural Network Architectures
To highlight the specific advantages of LSTM for this temporal prediction task, a comparison was made against two other common neural network architectures: Convolutional Neural Networks (CNN) and standard RNNs. CNNs are excellent at extracting spatial features but lack inherent mechanisms for modeling long-range sequential dependencies. Standard RNNs are designed for sequences but are hampered by the vanishing gradient problem. The comparison of RMSE for the 70% training data scenario is illustrative:
| Cell ID | CNN | Standard RNN | LSTM (Proposed) |
|---|---|---|---|
| B05 | 0.0152 | 0.0112 | 0.0126 |
| B18 | 0.0098 | 0.0075 | 0.0081 |
| B46 | 0.0076 | 0.0056 | 0.0063 |
While the performance differences can be subtle, the LSTM consistently shows robust and reliable performance. Its gated architecture provides a more stable learning process for the long-term dependencies inherent in li ion battery aging data compared to a standard RNN, and it is fundamentally more suited to the task than a CNN which requires careful engineering of the sequential input. The LSTM’s strength lies in its systematic approach to preserving and forgetting information over hundreds of cycles, which is essential for accurate SOH estimation.
5. Discussion and Future Perspectives
The exploration presented herein substantiates the significant potential of Long Short-Term Memory networks as a powerful, data-driven tool for the State of Health estimation of lithium-ion batteries. By effectively learning the non-linear mapping from operational time-series data to capacity fade, the LSTM approach circumvents the need for intricate physicochemical modeling or the sensitive parameter identification required by equivalent circuit models. This offers a flexible and adaptive pathway for SOH monitoring across diverse li ion battery applications.
However, the practical deployment of such models also presents challenges. The performance and generalization capability of any data-driven model, including LSTM, are intrinsically linked to the quality, quantity, and variety of the training data. Developing models that can generalize across different li ion battery chemistries, formats, and usage profiles remains an active area of research. Furthermore, the computational complexity of LSTM inference, though manageable on modern hardware, must be considered for resource-constrained embedded Battery Management Systems (BMS). Future work may focus on model compression techniques, transfer learning strategies to adapt pre-trained models to new battery types with limited data, and hybrid approaches that combine the interpretability of physics-based models with the adaptive learning power of deep neural networks. The integration of real-time model updating mechanisms to account for unforeseen operating conditions will also be crucial for ensuring long-term estimation robustness in the dynamic environment of a real-world li ion battery system.