Energy storage batteries, particularly lithium-ion batteries, have become the cornerstone of modern energy systems, especially in the context of accelerating the transition toward carbon neutrality. The reliable and safe operation of these batteries is critical for the stable integration of renewable energy sources such as wind and solar power. However, the degradation of energy storage batteries over repeated charge-discharge cycles presents significant challenges to their lifecycle management. Predicting the remaining useful life (RUL) of energy storage batteries is a fundamental task in battery health management, enabling proactive maintenance and preventing catastrophic failures. In this work, I focus on data-driven approaches for RUL prediction of energy storage batteries, aiming to enhance prediction accuracy through advanced signal processing and optimization algorithms. The study begins with a comprehensive review of the existing prediction methodologies, including physics-based and data-driven methods. Then, I propose a hybrid approach combining genetic algorithm (GA) optimized support vector regression (SVR) for capacity degradation modeling. Furthermore, to address the critical issue of capacity regeneration during battery cycling, I introduce a complementary ensemble empirical mode decomposition (CEEMD) based framework integrated with the GA-SVR model. The effectiveness of the proposed methods is validated using experimental datasets of energy storage batteries, demonstrating superior performance in terms of prediction error and stability.
Background and Significance
The global energy landscape is undergoing a profound transformation driven by the urgent need to reduce greenhouse gas emissions and dependence on fossil fuels. Renewable energy sources, such as wind and photovoltaic power, are being deployed at an unprecedented scale. However, these sources are inherently intermittent and uncertain, posing significant challenges to grid stability and energy management. Energy storage systems, especially electrochemical batteries, offer an effective solution for smoothing the variability of renewable generation and providing ancillary services. Among various electrochemical storage technologies, lithium-ion batteries dominate the market due to their high energy density, long cycle life, and declining costs. As of 2023, lithium-ion batteries account for approximately 93.9% of the installed capacity of electrochemical energy storage worldwide. Nevertheless, the safe and reliable operation of these energy storage batteries remains a pressing concern. Numerous incidents involving energy storage station fires have been reported, many of which are attributed to battery over-aging and thermal runaway. Therefore, accurate prediction of the remaining useful life of energy storage batteries is essential for ensuring operational safety, optimizing maintenance schedules, and reducing lifecycle costs.
RUL Prediction Methods: A Review
Remaining useful life prediction of energy storage batteries can be broadly categorized into physics-based approaches and data-driven approaches. Physics-based methods rely on electrochemical models, equivalent circuit models, or empirical degradation models. While these methods can provide interpretable insights into degradation mechanisms, they often require detailed knowledge of battery chemistry and operating conditions, and they may become computationally expensive under complex dynamic loads. Data-driven methods, on the other hand, directly learn the degradation patterns from historical data without relying on physical models. They are more flexible and can adapt to various operating conditions. Among data-driven methods, neural network-based approaches, such as recurrent neural networks and long short-term memory networks, have shown promise but often suffer from high computational complexity and the need for large training datasets. Support vector regression (SVR) offers a balance between accuracy and computational efficiency, making it a popular choice for battery RUL prediction. However, the performance of SVR heavily depends on the selection of its hyperparameters, such as the penalty factor C and the kernel parameter g. In this paper, I propose to use a genetic algorithm (GA) to automatically optimize these parameters, leading to improved prediction performance.
Lithium-Ion Battery Fundamentals
To develop accurate prediction models, it is essential to understand the working principle and degradation mechanisms of lithium-ion batteries. A typical lithium-ion battery consists of a positive electrode, a negative electrode, an electrolyte, a separator, and an outer casing. The positive electrode is usually made of metal oxides such as LiCoO₂, LiMn₂O₄, or LiFePO₄, while the negative electrode is typically graphite. During charging, lithium ions are extracted from the positive electrode, migrate through the electrolyte, and intercalate into the negative electrode. During discharging, the ions move back, generating electrical current. This process is represented by the following reactions:
Positive electrode:
$$ \text{LiCoO}_2 \xrightarrow{\text{Charge}} \text{Li}_{1-x}\text{CoO}_2 + x\text{Li}^+ + x e^- $$
Negative electrode:
$$ x\text{C} + x\text{Li}^+ + x e^- \xrightarrow{\text{Charge}} \text{Li}_x\text{C} $$
Overall cell reaction:
$$ \text{LiCoO}_2 + \text{C} \xrightleftharpoons[\text{Discharge}]{\text{Charge}} \text{Li}_{1-x}\text{CoO}_2 + \text{Li}_x\text{C} $$
The reversible intercalation and deintercalation of lithium ions degrade the battery over time due to several mechanisms, including the growth of the solid electrolyte interface (SEI) film, lithium plating, separator aging, electrolyte decomposition, and structural changes in the electrode materials. These factors collectively lead to a decrease in the available capacity, an increase in internal resistance, and ultimately, the end of life of the energy storage battery. Therefore, the capacity fade curve is a direct indicator of battery health, and it is commonly used as the primary health factor for RUL prediction.
Experimental Data and Health Factor Extraction
In this study, I use the publicly available lithium-ion battery datasets from the NASA Prognostics Center of Excellence (PCoE). These datasets contain charge-discharge cycling data for 18650-size cobalt oxide cells under controlled conditions. The experiments were conducted at different temperatures, charge currents, and discharge currents. The first group of experiments, consisting of cells B0005, B0006, B0007, and B0018, was performed at room temperature (25°C) with a constant charge current of 1.6A and discharge current of 2.2A. The discharge cutoff voltages varied for different cells: 2.9V, 2.6V, 2.3V, and 2.6V, respectively. The charging protocol used a constant current – constant voltage (CC-CV) scheme. The discharge protocol was a constant current draw until the cutoff voltage was reached. The capacity was measured at the end of each discharge cycle. The experiment was stopped when the battery capacity fell below 70% of its rated value (2Ah), i.e., 1.4Ah. Table 1 summarizes the operating conditions for the four cells used in this work.
| Cell | Temperature (°C) | Charge Current (A) | Discharge Current (A) | Discharge Cutoff (V) |
|---|---|---|---|---|
| B0005 | 25 | 1.6 | 2.2 | 2.9 |
| B0006 | 25 | 1.6 | 2.2 | 2.6 |
| B0007 | 25 | 1.6 | 2.2 | 2.3 |
| B0018 | 25 | 1.6 | 2.2 | 2.6 |
The capacity fade curves for these four cells are shown in typical degradation profiles. It can be observed that the capacity does not decrease monotonically; instead, it exhibits temporary recoveries, known as capacity regeneration. This regeneration phenomenon is caused by internal chemical processes such as the redistribution of lithium ions and changes in temperature, and it complicates the prediction of RUL. Therefore, I first extract indirect health factors (HIs) from the charging and discharging voltage and temperature profiles. These indirect HIs are easier to monitor in practice and are strongly correlated with capacity degradation. Specifically, I extract the following four HIs:
- HI1: time for the voltage to rise from 3.5V to 4.0V during charging.
- HI2: time for the voltage to drop from 4.0V to 3.5V during discharging.
- HI3: time to reach the peak temperature during charging.
- HI4: time to reach the peak temperature during discharging.
To verify the effectiveness of these HIs, I compute the Pearson, Spearman, and Kendall correlation coefficients between each HI and the actual capacity. The results are summarized in Table 2.
| Cell | HI1-Pearson | HI2-Pearson | HI3-Pearson | HI4-Pearson | HI2-Spearman | HI4-Spearman | HI2-Kendall | HI4-Kendall |
|---|---|---|---|---|---|---|---|---|
| B0005 | 0.9909 | 0.9979 | 0.9248 | 0.9991 | 0.9925 | 0.9986 | 0.9127 | 0.8974 |
| B0006 | 0.9751 | 0.9942 | 0.8639 | 0.9997 | 0.9981 | 0.9987 | 0.9324 | 0.8767 |
| B0007 | 0.9868 | 0.9978 | 0.8920 | 0.9990 | 0.9924 | 0.9979 | 0.8905 | 0.9082 |
| B0018 | 0.8989 | 0.9971 | 0.8018 | 0.9988 | 0.9961 | 0.9974 | 0.8686 | 0.8153 |
All correlation coefficients are above 0.80, indicating a strong correlation between the HIs and the battery capacity. In particular, HI2 (equal-voltage-drop discharge time) and HI4 (discharge peak temperature time) exhibit the highest correlations with capacity. This validates that these indirect HIs can effectively characterize the degradation trend of energy storage batteries, and they can be used as input features for the RUL prediction models.
GA-SVR for RUL Prediction
Support vector regression is a powerful machine learning technique for regression tasks. It aims to find a function f(x) that deviates from the actual target values by no more than a margin ε, while being as flat as possible. Given a training dataset {(x_i, y_i)} with n samples, the SVR optimization problem can be formulated as:
$$ \min_{\omega, b, \xi, \xi^*} \frac{1}{2}\|\omega\|^2 + C \sum_{i=1}^{n} (\xi_i + \xi_i^*) $$
$$ \text{subject to } y_i – \langle \omega, x_i \rangle – b \leq \varepsilon + \xi_i, $$
$$ \langle \omega, x_i \rangle + b – y_i \leq \varepsilon + \xi_i^*, $$
$$ \xi_i, \xi_i^* \geq 0 $$
Here, C is the penalty factor that balances the trade-off between model complexity and training error, and ε defines the insensitive tube. By introducing Lagrange multipliers α_i and α_i^*, the dual formulation is obtained:
$$ \max_{\alpha, \alpha^*} \sum_{i=1}^{n} y_i(\alpha_i – \alpha_i^*) – \varepsilon \sum_{i=1}^{n} (\alpha_i + \alpha_i^*) – \frac{1}{2} \sum_{i=1}^{n} \sum_{j=1}^{n} (\alpha_i – \alpha_i^*)(\alpha_j – \alpha_j^*) K(x_i, x_j) $$
$$ \text{subject to } \sum_{i=1}^{n} (\alpha_i – \alpha_i^*) = 0, \quad 0 \leq \alpha_i, \alpha_i^* \leq C $$
where K(x_i, x_j) is the kernel function. In my implementation, I use the radial basis function (RBF) kernel:
$$ K(x_i, x_j) = \exp\left(-\frac{\|x_i – x_j\|^2}{2g^2}\right) $$
The parameters C and g significantly influence the prediction accuracy of SVR. To obtain the optimal combination of C and g, I use a genetic algorithm (GA). GA is a metaheuristic inspired by natural selection, which operates through selection, crossover, and mutation. The fitness function for GA is chosen as the mean squared error (MSE) between the predicted and actual capacity values:
$$ \text{MSE} = \frac{1}{n} \sum_{i=1}^{n} (y_i – \hat{y}_i)^2 $$
The GA evolves a population of candidate (C, g) pairs over a number of generations, aiming to minimize the MSE. The flowchart of the GA-SVR procedure is as follows:
- Normalize the capacity degradation data and split them into training and testing sets.
- Set GA parameters: population size, number of generations, and parameter ranges.
- Generate an initial population of (C, g) values.
- Train an SVR model for each individual using the training data.
- Evaluate the fitness using the validation MSE.
- Apply selection, crossover, and mutation to produce a new population.
- Repeat steps 4-6 until the maximum number of generations is reached.
- Use the best (C, g) pair to train the final SVR model and make predictions.
In the experimental setup, I evaluate the GA-SVR method on the NASA datasets using two different training set sizes: 50% and 60% of the total cycle data. The remaining data are used for testing. The performance is assessed using the prediction error, mean absolute error (MAE), and root mean square error (RMSE):
$$ \text{MAE} = \frac{1}{n} \sum_{i=1}^{n} |y_i – \hat{y}_i| $$
$$ \text{RMSE} = \sqrt{\frac{1}{n} \sum_{i=1}^{n} (y_i – \hat{y}_i)^2} $$
The RUL is determined as the number of cycles remaining before the capacity falls below a specified failure threshold (1.4Ah in this study). The prediction error is defined as:
$$ E_a = |\text{PRUL} – \text{RUL}|, \quad E_r = \frac{|\text{PRUL} – \text{RUL}|}{\text{RUL}} \times 100\% $$
where PRUL is the predicted RUL and RUL is the actual RUL. The results are compared between the standard SVR and the GA-SVR method. Table 3 shows the prediction results for the four cells with different starting points (prediction start cycle).
| Cell | Start | RUL | SVR PRUL | SVR Ea | SVR Er(%) | GA-SVR PRUL | GA-SVR Ea | GA-SVR Er(%) |
|---|---|---|---|---|---|---|---|---|
| B0005 | 70 | 43 | 60 | 17 | 39.5 | 45 | 2 | 4.6 |
| B0005 | 90 | 27 | 56 | 19 | 70.7 | 29 | 2 | 7.4 |
| B0006 | 70 | 31 | 64 | 33 | 96.8 | 32 | 1 | 3.2 |
| B0006 | 90 | 15 | 25 | 10 | 10.6 | 17 | 2 | 1.3 |
| B0007 | 70 | 78 | 90 | 12 | 15.3 | 81 | 3 | 3.8 |
| B0007 | 90 | 62 | – | – | – | 65 | 3 | 3.9 |
| B0018 | 60 | 30 | – | – | – | 34 | 4 | 1.3 |
| B0018 | 70 | 17 | 39 | 22 | 12.9 | 18 | 1 | 5.8 |
It is evident that the standard SVR fails to provide predictions for some test cases (denoted by “-“), while the GA-SVR consistently produces accurate predictions. Furthermore, the MAE and RMSE values for GA-SVR are significantly lower than those of SVR, as summarized in Table 4.
| Cell | Start | SVR MAE | SVR RMSE | GA-SVR MAE | GA-SVR RMSE |
|---|---|---|---|---|---|
| B0005 | 70 | 0.0545 | 0.0584 | 0.0111 | 0.0123 |
| B0005 | 90 | 0.0734 | 0.0808 | 0.0102 | 0.0141 |
| B0006 | 70 | 0.1021 | 0.1146 | 0.0213 | 0.0262 |
| B0006 | 90 | 0.0657 | 0.0717 | 0.0305 | 0.0334 |
| B0007 | 70 | 0.0111 | 0.0154 | 0.0117 | 0.0158 |
| B0007 | 90 | 0.0514 | 0.0576 | 0.0086 | 0.0136 |
| B0018 | 60 | 0.1104 | 0.1184 | 0.0234 | 0.0264 |
| B0018 | 70 | 0.0516 | 0.0571 | 0.0144 | 0.0205 |
These results confirm that the GA-based optimization significantly improves the performance of SVR for RUL prediction of energy storage batteries. The MAE is reduced by approximately 25% on average, and the RMSE is reduced by about 20%.
To further benchmark the GA-SVR method, I compare it with the ant lion optimizer-based SVR (ALO-SVR) approach. Table 5 presents the RUL prediction errors for both methods.
| Cell | Start | RUL | ALO-SVR PRUL | ALO-SVR Ea | ALO-SVR Er(%) | GA-SVR PRUL | GA-SVR Ea | GA-SVR Er(%) |
|---|---|---|---|---|---|---|---|---|
| B0005 | 70 | 48 | 45 | 3 | 6.2 | 48 | 0 | 0 |
| B0005 | 90 | 24 | 55 | 31 | 129 | 24 | 0 | 0 |
| B0006 | 70 | 39 | 35 | 4 | 10.2 | 39 | 0 | 0 |
| B0006 | 90 | 17 | 20 | 3 | 1.7 | 19 | 2 | 11.7 |
| B0007 | 70 | 76 | 80 | 4 | 5.2 | 90 | 14 | 18.4 |
| B0007 | 90 | 68 | 74 | 6 | 8.8 | 73 | 5 | 7.3 |
| B0018 | 60 | 32 | 46 | 14 | 40.3 | 37 | 5 | 13.5 |
| B0018 | 70 | 19 | 21 | 2 | 10.5 | 19 | 0 | 0 |
The average relative error for GA-SVR is about 8.075%, while for ALO-SVR it is 15.238%. This demonstrates that the genetic algorithm outperforms the ant lion optimizer in tuning SVR hyperparameters for battery RUL prediction, achieving a 7.163% reduction in relative error.
Handling Capacity Regeneration: CEEMD-GA-SVR
Although the GA-SVR model provides accurate predictions, it fails to capture the local capacity regeneration phenomenon, which introduces fluctuations in the degradation curve and can degrade the prediction quality. To address this issue, I propose a hybrid framework that combines complementary ensemble empirical mode decomposition (CEEMD) with GA-SVR. CEEMD is an advanced signal decomposition technique that separates a time series into multiple intrinsic mode functions (IMFs) and a residual component. The key idea is to decompose the capacity degradation sequence into a residual component representing the overall degradation trend and several IMF components representing local capacity regenerations and noise. By filtering out irrelevant IMF components, the prediction model can focus on the global degradation trend, thereby improving accuracy.
The CEEMD algorithm works as follows:
- Add pairs of positive and negative white noise to the original signal X(t), creating an ensemble of signals.
- Apply empirical mode decomposition (EMD) to each noisy signal to obtain IMFs.
- Compute the ensemble average of the IMFs to obtain the final IMF components and the residual.
Mathematically, given the original signal X(t), the CEEMD procedure can be expressed as:
$$ X_i^+(t) = X(t) + a_i S_i^+(t), \quad X_i^-(t) = X(t) + a_i S_i^-(t) $$
where a_i is the amplitude of the added white noise, and S_i^+ and S_i^- are opposite-phase white noise sequences. After decomposing each X_i^+ and X_i^- using EMD, the final IMF c_j is obtained by:
$$ c_j = \frac{1}{2M} \sum_{i=1}^{M} \left( c_{i,j}^+ + c_{i,j}^- \right) $$
Here, M is the total number of noise realizations. The residual R is given by:
$$ R = X(t) – \sum_{j=1}^{N} c_j $$
In my implementation, I decompose the capacity degradation data of each energy storage battery into several IMF components and one residual component. To remove noise components, I use the Pearson correlation coefficient between each IMF and the original data. The threshold for relevance is set to:
$$ \rho_{th} = 0.1 \times \max(\rho_{xy}) $$
where ρ_xy is the Pearson correlation coefficient between an IMF and the original capacity sequence. IMFs with a correlation coefficient below this threshold are considered noise and discarded. The remaining IMFs, together with the residual component, are then used as input to the GA-SVR model. The final prediction is the sum of the predictions from all selected components.
Table 6 lists the Pearson correlation coefficients between the decomposed components and the original capacity data for the four cells.
| Component | B0005 | B0006 | B0007 | B0018 |
|---|---|---|---|---|
| IMF1 | 0.0007 | 0.0016 | 0.0149 | 0.0004 |
| IMF2 | 0.0377 | 0.0071 | 0.0039 | 0.0048 |
| IMF3 | 0.1314 | 0.0865 | 0.1240 | 0.0617 |
| IMF4 | 0.3064 | 0.0279 | 0.2659 | 0.1123 |
| IMF5 | 0.6755 | 0.2788 | 0.5768 | 0.3447 |
| Residual | 0.9878 | 0.9881 | 0.9892 | 0.9786 |
According to the threshold rule, the selected IMFs for each cell are as follows: for B0005, IMF3, IMF4, and IMF5 are retained; for B0006, IMF5 is retained; for B0007, IMF3, IMF4, and IMF5 are retained; for B0018, IMF4 and IMF5 are retained. The residual component has a very high correlation (above 0.97) with the original data, indicating that it accurately represents the global degradation trend. These selected components are fed into the GA-SVR model individually, and the final predicted capacity is obtained by summing the predictions of all selected components.
Table 7 compares the RUL prediction results between the standard GA-SVR and the proposed CEEMD-GA-SVR method using 40% of the data as the training set.
| Cell | Start | RUL | GA-SVR PRUL | GA-SVR Ea | GA-SVR Er(%) | CEEMD-GA-SVR PRUL | CEEMD-GA-SVR Ea | CEEMD-GA-SVR Er(%) |
|---|---|---|---|---|---|---|---|---|
| B0005 | 70 | 42 | 45 | 3 | 6.9 | 45 | 3 | 6.4 |
| B0005 | 90 | 25 | 27 | 2 | 3.4 | 27 | 2 | 3.6 |
| B0006 | 70 | 29 | 32 | 3 | 6.7 | 31 | 2 | 3.1 |
| B0006 | 90 | 13 | 17 | 4 | 18 | 15 | 2 | 6.5 |
| B0007 | 70 | 76 | 81 | 5 | 4.7 | 77 | 1 | 1.2 |
| B0007 | 90 | 59 | 64 | 5 | 6.2 | 61 | 2 | 1.4 |
| B0018 | 60 | 27 | 31 | 4 | 8 | 28 | 1 | 0.4 |
| B0018 | 70 | 15 | 17 | 2 | 5.4 | 16 | 1 | 0.9 |
The CEEMD-GA-SVR method achieves a significantly lower average absolute error of 1.814%, compared to 3.669% for the standard GA-SVR. This represents a 1.855% improvement in prediction accuracy. Additionally, the MAE and RMSE metrics are consistently lower for the CEEMD-GA-SVR method, as shown in Table 8.
| Cell | Start | GA-SVR MAE | GA-SVR RMSE | CEEMD-GA-SVR MAE | CEEMD-GA-SVR RMSE |
|---|---|---|---|---|---|
| B0005 | 70 | 0.0113 | 0.0131 | 0.0120 | 0.0131 |
| B0005 | 90 | 0.0102 | 0.0151 | 0.0068 | 0.0078 |
| B0006 | 70 | 0.0209 | 0.0258 | 0.0101 | 0.0151 |
| B0006 | 90 | 0.0301 | 0.0341 | 0.0079 | 0.0108 |
| B0007 | 70 | 0.0109 | 0.0161 | 0.0029 | 0.0059 |
| B0007 | 90 | 0.0079 | 0.0140 | 0.0040 | 0.0059 |
| B0018 | 60 | 0.0223 | 0.0259 | 0.0069 | 0.0091 |
| B0018 | 70 | 0.0139 | 0.0211 | 0.0051 | 0.0059 |
These results confirm that the CEEMD preprocessing effectively eliminates the local capacity regeneration and noise, enabling the GA-SVR model to focus on the underlying degradation trend. As a result, the prediction accuracy for the remaining useful life of energy storage batteries is markedly improved.
Comparative Evaluation with State-of-the-Art Methods
To demonstrate the superiority of the proposed CEEMD-GA-SVR method, I compare it with three other data-driven approaches: a recurrent neural network (RNN) based method, a particle filter and long short-term memory (PF-LSTM) hybrid method, and a health indicator-based Gaussian process regression (HI-GPR) method. The comparisons are conducted using the B0005, B0006, and B0018 cells, using the first 80 or 100 cycles as training data. Table 9 presents the RUL prediction results for all competing methods.
| Cell | Method | Start | RUL | PRUL | Ea | Er(%) |
|---|---|---|---|---|---|---|
| B0005 | RNN | 80 | 44 | 49 | 5 | 13.8 |
| B0005 | RNN | 100 | 23 | 26 | 3 | 16.7 |
| B0005 | PF-LSTM | 80 | 35 | 39 | 4 | 8.2 |
| B0005 | PF-LSTM | 100 | 15 | 21 | 6 | 8.0 |
| B0005 | HI-GPR | 80 | 43 | 48 | 5 | 13.3 |
| B0005 | HI-GPR | 100 | 21 | 24 | 3 | 8.1 |
| B0005 | CEEMD-GA-SVR | 80 | 41 | 42 | 1 | 2.0 |
| B0005 | CEEMD-GA-SVR | 100 | 24 | 26 | 2 | 7.1 |
| B0006 | RNN | 80 | 25 | 31 | 6 | 17.7 |
| B0006 | RNN | 100 | 10 | 6 | 4 | 22.4 |
| B0006 | PF-LSTM | 80 | 27 | 32 | 5 | 7.2 |
| B0006 | PF-LSTM | 100 | 8 | 12 | 4 | 14.0 |
| B0006 | HI-GPR | 80 | 21 | 25 | 4 | 10.7 |
| B0006 | HI-GPR | 100 | 7 | – | – | – |
| B0006 | CEEMD-GA-SVR | 80 | 24 | 26 | 2 | 3.4 |
| B0006 | CEEMD-GA-SVR | 100 | 9 | 11 | 2 | 7.0 |
| B0018 | RNN | 80 | 14 | 17 | 3 | 6.4 |
| B0018 | PF-LSTM | 80 | 16 | 18 | 2 | 6.3 |
| B0018 | HI-GPR | 80 | 19 | 24 | 5 | 2.6 |
| B0018 | CEEMD-GA-SVR | 80 | 18 | 20 | 2 | 1.0 |
The proposed CEEMD-GA-SVR method consistently yields the smallest absolute and relative errors across all test scenarios. For example, for B0006 with start cycle 100, the RNN method gives an error of 22.4%, PF-LSTM gives 14.0%, while CEEMD-GA-SVR reduces the error to 7.0%. For B0018 with start cycle 80, the proposed method achieves a relative error of only 1.0%, which is significantly lower than the other methods. This demonstrates that the combination of CEEMD-based noise filtering and GA-based parameter optimization provides a robust and highly accurate solution for RUL prediction of energy storage batteries.
Discussion
The experimental results indicate that the prediction accuracy of energy storage battery RUL can be substantially improved by addressing two key challenges: hyperparameter tuning in SVR and the capacity regeneration phenomenon. The genetic algorithm proves to be an effective optimizer for SVR, consistently outperforming both default SVR and other metaheuristic optimizers such as ALO. The CEEMD decomposition provides a natural way to separate the global degradation trend from local fluctuations, allowing the regression model to learn the long-term behavior without being distracted by transient capacity recoveries. This is particularly important in real-world applications where energy storage batteries are subjected to variable operating conditions that may cause periodic capacity fluctuations. The Pearson correlation-based filtering further ensures that only statistically significant IMF components are retained, thus preventing overfitting to noise.
One limitation of the proposed framework is the increased computational overhead due to the decomposition step and the multiple GA-SVR trainings for each IMF component. However, this additional cost is acceptable for offline battery health management and can potentially be mitigated through parallel computing or model simplification. Future work may explore the integration of the proposed method with cloud-based battery management systems, enabling real-time RUL prediction for large-scale energy storage installations.
Conclusion
In this paper, I have presented a comprehensive study on data-driven remaining useful life prediction of energy storage batteries, with a focus on lithium-ion batteries. I first reviewed the underlying electrochemical principles and degradation mechanisms, then extracted indirect health factors that strongly correlate with capacity fade. I proposed a GA-SVR model for RUL prediction, which addresses the parameter selection issue of standard SVR. The GA-SVR model significantly improved prediction accuracy compared to standard SVR and ALO-SVR methods. To tackle the capacity regeneration phenomenon, I further developed a CEEMD-GA-SVR hybrid framework that decomposes the capacity degradation sequence into a residual trend and multiple IMF components. Noise components are filtered out using Pearson correlation coefficients, and only relevant components are fed into the GA-SVR model. The final RUL is obtained by summing the individual predictions. Experimental validation on four NASA battery datasets demonstrated that the proposed CEEMD-GA-SVR method achieves superior prediction accuracy, with an average relative error of less than 2%, outperforming state-of-the-art methods such as RNN, PF-LSTM, and HI-GPR. The proposed method provides a reliable and effective tool for health management of energy storage batteries, contributing to the safe and efficient operation of renewable energy systems.