In the era of rapid renewable energy integration, the role of energy storage systems has become paramount. Among various technologies, lithium-ion energy storage batteries stand out due to their high energy density, excellent cycle life, and environmental friendliness. They are critical for stabilizing grid operations, enabling peak shaving, and facilitating the transition to a low-carbon economy. However, the degradation of energy storage batteries over time poses significant challenges for safety, reliability, and economic viability. Accurately assessing the State of Health (SOH) of energy storage batteries is essential for predicting remaining useful life, optimizing maintenance schedules, and preventing catastrophic failures. Traditional methods for SOH evaluation, such as electrochemical or equivalent circuit models, often struggle with complexity, robustness, and real-time applicability. Recently, data-driven approaches have gained traction, leveraging machine learning algorithms to infer SOH from operational data like voltage, current, and temperature. Yet, many existing methods face limitations in accuracy, especially when relying on incomplete or noisy data. In this study, we address these challenges by proposing a novel data-driven framework for SOH evaluation of energy storage batteries. Our approach extracts key health factors from charging voltage profiles and employs a Genetic Algorithm-optimized Long Short-Term Memory (GA-LSTM) neural network for precise SOH estimation. We validate the method through cyclic aging experiments on commercial lithium iron phosphate (LiFePO4) energy storage batteries, demonstrating superior performance compared to baseline models. This work aims to contribute to the advancement of intelligent battery management systems, enhancing the sustainability and efficiency of energy storage infrastructure worldwide.
The SOH of an energy storage battery is a metric that quantifies its current capacity relative to its nominal capacity, typically expressed as a percentage. It reflects the battery’s aging degree and ability to store energy. Formally, SOH is defined as:
$$SOH = \frac{Q_0}{Q_N} \times 100\%$$
where \(Q_0\) is the actual capacity measured under standardized conditions, and \(Q_N\) is the nominal capacity provided by the manufacturer. As an energy storage battery undergoes charge-discharge cycles, its capacity gradually fades due to mechanisms like solid electrolyte interface growth, lithium plating, and active material loss. Accurate SOH estimation enables proactive management, but it is hindered by the nonlinear and time-dependent nature of battery degradation. Data-driven methods offer a promising alternative by learning patterns from historical data without requiring explicit physical models. Common techniques include Support Vector Regression (SVR), Gaussian Process Regression (GPR), and neural networks. However, these methods often rely on handcrafted features or full charge-discharge cycles, which may not be available in real-world applications. Our research focuses on utilizing fragmentary operational data—specifically, voltage segments during charging—to derive health indicators. This approach aligns with practical scenarios where energy storage batteries operate under partial cycles or variable conditions, making it highly applicable for grid-scale energy storage systems.
To understand the degradation behavior of energy storage batteries, we conducted cyclic aging experiments on five soft-pack LiFePO4 energy storage batteries with a nominal capacity of 20 Ah. These energy storage batteries were subjected to repeated charge-discharge cycles in a controlled environment at 25°C. Each cycle involved charging at a constant current of 0.5C to a cut-off voltage of 3.65 V, followed by discharging to 2.50 V at the same rate. After every 100 cycles, the capacity was calibrated using a standard protocol: three consecutive cycles of constant-current charge and discharge at 10 A, with the final discharge capacity recorded as the current capacity. This process simulates long-term operation in energy storage applications, such as frequency regulation or load shifting. Over 4,000 cycles, the energy storage batteries exhibited a near-linear capacity fade, with capacity retention around 90.5%. The voltage profiles during charging revealed consistent patterns, particularly in the plateau region between 3.25 V and 3.40 V, where the primary electrochemical reactions occur. This observation guided our selection of health factors from operational data.
The extraction of health factors is crucial for effective SOH evaluation. We analyzed charging voltage curves and their derivative Incremental Capacity Analysis (ICA) curves to identify sensitive indicators. The ICA curves, obtained by differentiating capacity with respect to voltage (\(dQ/dV\)), show distinct peaks corresponding to phase transitions in the electrode materials. For LiFePO4 energy storage batteries, these peaks lie within the 3.25–3.40 V range, confirming this region’s relevance for degradation monitoring. We computed voltage change values over specific time intervals during charging: after reaching 3.25 V, before reaching 3.35 V, and before reaching 3.40 V. These intervals were chosen based on preliminary correlation analysis with SOH. For each interval, we calculated the voltage difference over 10, 20, and 30 minutes, resulting in nine candidate health factors. To refine our selection, we computed the Pearson correlation coefficient between each factor and the actual SOH values. The Pearson coefficient measures linear correlation, with values close to 1 indicating strong positive correlation. The results are summarized in Table 1.
| Time Interval (minutes) | Voltage Threshold | Pearson Coefficient |
|---|---|---|
| 10 | After 3.25 V | 0.6406 |
| 20 | After 3.25 V | 0.2267 |
| 30 | After 3.25 V | 0.3857 |
| 10 | Before 3.35 V | 0.1855 |
| 20 | Before 3.35 V | 0.6358 |
| 30 | Before 3.35 V | 0.3615 |
| 10 | Before 3.40 V | 0.4912 |
| 20 | Before 3.40 V | 0.0333 |
| 30 | Before 3.40 V | 0.9028 |
Based on the correlation analysis, we selected three health factors with the highest Pearson coefficients: the voltage change over 10 minutes after reaching 3.25 V (0.6406), over 20 minutes before reaching 3.35 V (0.6358), and over 30 minutes before reaching 3.40 V (0.9028). These factors capture distinct aspects of the charging dynamics and exhibit strong linear relationships with SOH, making them suitable inputs for our predictive model. The mathematical formulation for each health factor is as follows:
$$\Delta V_1 = V(t_{3.25V} + 10\text{ min}) – V(t_{3.25V})$$
$$\Delta V_2 = V(t_{3.35V}) – V(t_{3.35V} – 20\text{ min})$$
$$\Delta V_3 = V(t_{3.40V}) – V(t_{3.40V} – 30\text{ min})$$
where \(V(t)\) denotes the voltage at time \(t\), and \(t_{3.25V}\), \(t_{3.35V}\), \(t_{3.40V}\) are the times when the voltage first crosses 3.25 V, 3.35 V, and 3.40 V during charging, respectively. These health factors are derived solely from operational data, requiring no full discharge or invasive testing, which aligns with practical constraints in energy storage battery management systems.
For SOH evaluation, we employed a Long Short-Term Memory (LSTM) neural network, a type of recurrent neural network adept at handling sequential data. The LSTM architecture addresses the vanishing gradient problem through gating mechanisms, allowing it to capture long-term dependencies in time-series data—a key advantage for modeling the gradual degradation of energy storage batteries. The LSTM cell comprises three gates: forget gate (\(f_t\)), input gate (\(i_t\)), and output gate (\(o_t\)), along with a cell state (\(C_t\)) that retains memory over time. The equations governing an LSTM cell are:
$$f_t = \sigma(W_f \cdot [h_{t-1}, x_t] + b_f)$$
$$i_t = \sigma(W_i \cdot [h_{t-1}, x_t] + b_i)$$
$$\tilde{C}_t = \tanh(W_C \cdot [h_{t-1}, x_t] + b_C)$$
$$C_t = f_t \odot C_{t-1} + i_t \odot \tilde{C}_t$$
$$o_t = \sigma(W_o \cdot [h_{t-1}, x_t] + b_o)$$
$$h_t = o_t \odot \tanh(C_t)$$
where \(x_t\) is the input vector at time \(t\) (e.g., health factors), \(h_t\) is the hidden state, \(W\) and \(b\) are weight matrices and bias terms, \(\sigma\) is the sigmoid function, \(\odot\) denotes element-wise multiplication, and \(\tanh\) is the hyperbolic tangent function. The LSTM network processes sequences of health factors extracted from consecutive cycles to predict SOH values. However, the performance of LSTM networks depends heavily on hyperparameters such as the number of hidden units, learning rate, and training epochs. Manual tuning of these parameters is often suboptimal. To enhance accuracy, we integrated a Genetic Algorithm (GA) for hyperparameter optimization. GA is a metaheuristic inspired by natural selection, which evolves a population of candidate solutions through selection, crossover, and mutation operations. In our GA-LSTM framework, each individual in the population represents a set of hyperparameters. The fitness function is defined as the negative mean squared error (MSE) on a validation set, encouraging the selection of configurations that minimize prediction error. The GA optimization process involves the following steps:
- Initialization: Generate a random population of hyperparameter sets.
- Evaluation: For each set, train an LSTM model and compute fitness (e.g., -MSE).
- Selection: Select individuals with higher fitness for reproduction.
- Crossover: Combine pairs of individuals to produce offspring.
- Mutation: Randomly alter some genes (hyperparameters) to maintain diversity.
- Termination: Repeat until a maximum number of generations is reached, then output the best hyperparameters.
This automated optimization ensures that the LSTM model is tailored to the specific characteristics of energy storage battery degradation data, leading to improved generalization and robustness.
To validate our method, we used data from five energy storage batteries, each subjected to 4,000 cycles. The dataset included health factors (\(\Delta V_1\), \(\Delta V_2\), \(\Delta V_3\)) and corresponding SOH values for each cycle. We partitioned the data into training and testing sets: data from three energy storage batteries (approximately 12,000 cycles) were used for training, while data from one battery (approximately 4,000 cycles) was held out for testing, and data from the remaining battery was used for validation during GA optimization. The training set comprised 80% of the cycles from the three batteries, and the testing set contained the remaining 20%. The GA-LSTM model was implemented in Python using TensorFlow and the DEAP library for genetic algorithms. Key hyperparameters optimized by GA included the number of LSTM layers (1 to 3), hidden units (32 to 256), learning rate (0.001 to 0.1), dropout rate (0 to 0.5), and batch size (16 to 128). The GA population size was set to 50, with 30 generations of evolution. For comparison, we also trained a standard LSTM model with manually chosen hyperparameters (e.g., 1 layer, 128 hidden units, learning rate 0.01) and a GA-optimized Backpropagation (GA-BP) neural network, a common baseline for SOH estimation. All models were evaluated using two metrics: Mean Squared Error (MSE) and Mean Absolute Percentage Error (MAPE), defined as:
$$MSE = \frac{1}{n} \sum_{i=1}^{n} (y_i – \hat{y}_i)^2$$
$$MAPE = \frac{1}{n} \sum_{i=1}^{n} \left| \frac{y_i – \hat{y}_i}{y_i} \right| \times 100\%$$
where \(y_i\) is the actual SOH, \(\hat{y}_i\) is the predicted SOH, and \(n\) is the number of samples. Lower values of MSE and MAPE indicate higher accuracy. The experimental results are summarized in Table 2.
| Model | MSE | MAPE (%) |
|---|---|---|
| Standard LSTM | 0.000548 | 2.11 |
| GA-BP | 0.002820 | 1.33 |
| GA-LSTM (Proposed) | 0.000142 | 1.09 |
The GA-LSTM model achieved the lowest MSE (0.000142) and MAPE (1.09%), outperforming the standard LSTM by reductions of 74.1% in MSE and 48.3% in MAPE. It also surpassed the GA-BP model, which had higher errors, demonstrating the superiority of LSTM’s sequential processing for time-series data from energy storage batteries. The predicted SOH curves closely followed the actual degradation trends, with deviations mostly within ±1% across cycles. This high accuracy validates the effectiveness of our selected health factors and the GA optimization process. Moreover, the method’s reliance on fragmentary charging data makes it suitable for real-world energy storage applications where full cycles are not always available. For instance, in grid-scale energy storage systems, batteries often operate in partial state-of-charge ranges, and our approach can still extract meaningful health indicators from short voltage segments.
Beyond numerical results, we conducted an in-depth analysis of the health factors’ physical significance. The voltage change after 3.25 V correlates with the initial stage of lithium intercalation into the graphite anode, which is sensitive to anode degradation. The voltage change before 3.35 V relates to the middle phase of the LiFePO4 cathode’s two-phase reaction, affected by cathode aging. The voltage change before 3.40 V corresponds to the final stage of charging, where polarization effects become prominent, reflecting overall internal resistance increase. By combining these factors, our model captures multiple degradation mechanisms, enhancing robustness against varying operating conditions. We also explored the impact of data noise and measurement errors by adding Gaussian noise to the voltage signals. The GA-LSTM model maintained stable performance with up to 5% noise, indicating good tolerance to real-world data imperfections. Furthermore, we tested the model’s generalizability by applying it to a different batch of energy storage batteries with similar chemistry but varying initial capacities. The results showed consistent accuracy, with MAPE below 1.5%, suggesting that the method can be adapted to diverse energy storage battery populations with minimal recalibration.
The proposed framework has several advantages for energy storage battery management. First, it uses operational data that is readily available from battery management systems, avoiding the need for additional sensors or invasive tests. Second, the GA optimization automates hyperparameter tuning, reducing the reliance on expert knowledge and making the method accessible for industrial deployment. Third, the LSTM’s ability to handle sequences allows for continuous SOH monitoring over the battery’s lifespan, enabling early detection of anomalies. However, there are limitations to consider. The method assumes consistent charging protocols; variations in charging rates or temperatures may affect health factor extraction. Future work could integrate adaptive preprocessing to handle such variability. Additionally, the model was validated on LiFePO4 energy storage batteries; extending it to other chemistries like NMC or LTO would require retraining with relevant data. Another direction is to incorporate more health factors, such as temperature rise or impedance spectra, to further improve accuracy. From a practical standpoint, implementing the GA-LSTM model in edge computing devices for real-time SOH estimation could enhance the autonomy of energy storage systems. As the global demand for energy storage batteries grows, such intelligent diagnostic tools will be crucial for maximizing lifespan, ensuring safety, and reducing lifecycle costs.
In conclusion, this study presents a robust data-driven method for SOH evaluation of energy storage batteries based on operational data. By extracting key health factors from charging voltage profiles and leveraging a GA-optimized LSTM neural network, we achieved accurate SOH estimation with MSE as low as 0.000142 and MAPE of 1.09%. The method significantly outperforms standard LSTM and GA-BP models, demonstrating the value of automated hyperparameter optimization and sequential learning for battery degradation modeling. Our approach is practical, relying on fragmentary data that aligns with real-world energy storage operations, and it provides insights into underlying degradation mechanisms through carefully selected health factors. As energy storage batteries become increasingly integral to renewable energy grids, advanced SOH assessment techniques like this will play a vital role in optimizing performance and longevity. Future research will focus on extending the framework to diverse battery types, integrating real-time adaptation, and exploring hybrid models that combine data-driven and physics-based elements for even greater reliability. Through continued innovation, we can ensure that energy storage batteries contribute effectively to a sustainable energy future.