As a high-performance electrochemical energy storage device, the energy storage battery has become an indispensable component in modern green energy systems, particularly in grid-scale applications, renewable energy integration, and peak shaving. The production of energy storage battery cells involves multiple sophisticated stages, among which the formation and capacity grading processes are critical but time-consuming and energy-intensive. To measure the actual capacity of a cell, a complete charge-discharge cycle is required, which can take up to several hours for large-format cells. Moreover, during the operational phase, the complex internal electrochemical reactions within the energy storage battery lead to gradual capacity degradation and impedance growth, posing safety risks such as thermal runaway. Therefore, accurate prediction of both the split capacity during manufacturing and the remaining useful life (RUL) during service is of paramount importance. This thesis focuses on a 280 Ah prismatic aluminum-shell lithium iron phosphate (LFP) energy storage battery. A data-driven framework is developed to predict the discharge capacity from partial formation charge data and to predict the RUL using cyclic aging data. The proposed methods integrate advanced machine learning techniques with battery degradation analysis, significantly improving prediction accuracy and operational efficiency.
The main contributions of this work are twofold. First, a novel capacity prediction model is constructed using LightGBM, K-nearest neighbors (KNN), decision trees, and their weighted ensemble. This model predicts the final discharge capacity of the energy storage battery from only the first 20% of the formation charge curve, thereby drastically reducing the required processing time and cost. Second, a hybrid Gaussian process regression (GPR) model with optimized kernel functions and incremental capacity analysis (ICA) is developed for RUL prediction. The proposed HI-C-GPR model leverages health indicators (HIs) extracted from incremental capacity curves to estimate both the current state of health (SOH) and the future capacity trajectory. Extensive validation is performed on both the NASA battery dataset and the custom 280 Ah cell dataset, demonstrating superior prediction performance.
1. Introduction
Under the global carbon neutrality target, renewable energy sources such as wind and solar power have experienced rapid expansion. However, the intermittent nature of these sources necessitates efficient energy storage systems to balance supply and demand. Among various storage technologies, electrochemical energy storage using lithium-ion batteries has become the dominant choice due to its high energy density, long cycle life, and low self-discharge rate. The energy storage battery market has been growing exponentially, with many manufacturers focusing on large-format cells such as the 280 Ah LFP cell. The production process of these cells includes electrode mixing, coating, winding, assembly, formation, and capacity grading. The formation and grading processes are particularly resource-intensive because they require multiple full charge/discharge cycles to stabilize the solid electrolyte interface (SEI) and to measure the actual cell capacity. For a 280 Ah energy storage battery, a single grading cycle can take more than six hours, leading to significant energy consumption and reduced throughput.
To address this issue, data-driven capacity prediction models have been developed that use only partial formation data to estimate the final grading capacity. Such models can significantly reduce the production time and cost while maintaining acceptable accuracy. In addition to manufacturing efficiency, the operational safety and reliability of energy storage battery systems are critical. Battery degradation over cycling leads to capacity fade and increased internal resistance, which can ultimately cause system failure or even catastrophic events. Therefore, accurate RUL prediction is essential for proactive maintenance and safe operation. This thesis focuses on both aspects: predicting the split capacity from formation charge data and predicting the RUL from cyclic aging data, using advanced data-driven algorithms.
2. Fundamentals of Lithium-Ion Batteries and Machine Learning Models
2.1 Classification of Lithium-Ion Batteries
Lithium-ion batteries are broadly classified by their physical shape: cylindrical, prismatic, and pouch cells. Cylindrical cells, such as the 18650, are widely used in consumer electronics due to their mature manufacturing process and good mechanical stability. Prismatic cells, including the 280 Ah aluminum-shell cells, are commonly used in energy storage battery systems because of their compact structure, high capacity, and excellent thermal management. Pouch cells offer flexible packaging and high energy density, but they are more prone to swelling and damage. The 280 Ah prismatic cell used in this study is shown in the following hyperlink as a representative of large-format energy storage battery cells.
2.2 Working Principle and Aging Mechanisms
The energy storage battery operates on the principle of lithium-ion intercalation and deintercalation between positive and negative electrodes. During charging, lithium ions are extracted from the cathode material (e.g., LiFePO₄) and inserted into the anode graphite structure. The opposite occurs during discharge. The electrochemical reactions are accompanied by the flow of electrons through the external circuit. The capacity and voltage of the cell are determined by the thermodynamics and kinetics of these reactions.
Battery aging is an inevitable process caused by multiple factors, including SEI growth, electrode material degradation, electrolyte decomposition, lithium plating, and current collector corrosion. These mechanisms lead to loss of active lithium, increased impedance, and structural damage, all of which contribute to capacity fade and power capability reduction. Understanding these aging mechanisms is crucial for developing accurate RUL prediction models.
2.3 Algorithm Models
2.3.1 LightGBM
LightGBM (Light Gradient Boosting Machine) is a highly efficient gradient boosting decision tree algorithm. It utilizes histogram-based binning to speed up training and reduce memory usage, and a leaf-wise growth strategy to achieve lower loss. The model builds an ensemble of weak learners (decision trees) in a sequential manner, each fitting the residual errors of the previous trees. The prediction for a given input vector $\boldsymbol{x}$ is:
$$ \hat{y} = \sum_{m=1}^{M} \alpha_m h_m(\boldsymbol{x}) $$
where $h_m$ is the $m$-th weak learner and $\alpha_m$ is its weight. LightGBM is known for its scalability and robustness when dealing with large-scale datasets, making it suitable for mining the massive formation data of energy storage battery cells.
2.3.2 K-Nearest Neighbors (KNN)
KNN is a non-parametric algorithm that predicts outputs based on the weighted average of the $K$ closest training samples in feature space. The distance metric is often Euclidean:
$$ d(\boldsymbol{x}_i, \boldsymbol{x}_j) = \sqrt{\sum_{k=1}^{n} (x_{ik} – x_{jk})^2} $$
The output prediction is the mean of the target values of the $K$ nearest neighbors:
$$ \hat{y} = \frac{1}{K} \sum_{i=1}^{K} y_i $$
KNN is simple and effective, but its performance depends heavily on the choice of $K$ and the distance metric. In this thesis, KNN serves as a baseline model for capacity prediction.
2.3.3 Decision Tree
Decision tree regression partitions the feature space into regions and assigns a constant value to each region. The optimal split is determined by minimizing the sum of squared errors. For a region $R_j$, the predicted value is the mean of the target variables in that region:
$$ \hat{y}_{R_j} = \frac{1}{N_j} \sum_{i \in R_j} y_i $$
Although individual decision trees are prone to overfitting, they are powerful when integrated into ensemble models such as LightGBM or random forests.
2.3.4 Gaussian Process Regression (GPR)
GPR is a non-parametric Bayesian regression method that models the target variable as a stochastic process. A Gaussian process is defined by its mean function $m(\boldsymbol{x})$ and covariance function $k(\boldsymbol{x}, \boldsymbol{x}’)$:
$$ f(\boldsymbol{x}) \sim \mathcal{GP}(m(\boldsymbol{x}), k(\boldsymbol{x} ,\boldsymbol{x}’)) $$
The standard choice is a zero mean and a squared exponential covariance function:
$$ k(\boldsymbol{x}, \boldsymbol{x}’) = \sigma_f^2 \exp\left(-\frac{\|\boldsymbol{x} – \boldsymbol{x}’\|^2}{2l^2}\right) $$
For noisy observations $y = f(\boldsymbol{x}) + \epsilon$ with $\epsilon \sim \mathcal{N}(0, \sigma_n^2)$, the predictive distribution at test points $\boldsymbol{X}_*$ is:
$$ \hat{\boldsymbol{f}}_* = \boldsymbol{K}_*^\top [\boldsymbol{K} + \sigma_n^2 \boldsymbol{I}]^{-1} \boldsymbol{y} $$
$$ \text{cov}(\hat{\boldsymbol{f}}_*) = \boldsymbol{K}_{**} – \boldsymbol{K}_*^\top [\boldsymbol{K} + \sigma_n^2 \boldsymbol{I}]^{-1} \boldsymbol{K}_* $$
GPR provides not only point predictions but also uncertainty estimates, which is advantageous for battery RUL prediction where confidence bounds are valuable.
3. Split Capacity Prediction Based on Formation Data
3.1 Data Collection and Problem Statement
The study focuses on a 280 Ah prismatic LFP energy storage battery. The formation process consists of multiple stages, including rest periods and constant-current charging steps. A typical formation charging curve is illustrated conceptually in the original thesis, where the charging stages are labeled I, II, III, and IV. The total formation time can exceed several hours. The goal of this chapter is to predict the final discharge capacity (measured during the subsequent grading process) using only the first 20% of the formation charge data. The challenge arises from the complex SEI formation dynamics and the need to extract meaningful features from the early-stage data.
In total, more than 50,000 cell samples were collected from the production line at a sampling rate of 1 Hz. For each cell, the recorded variables include phase number, phase state, phase time, voltage, current, capacity, energy, temperature (positive and negative terminals), pressure, contact impedance, and auxiliary voltage. The target variable is the actual discharge capacity obtained from the grading process.
3.2 Data Preprocessing
Raw formation data often contain anomalies such as duplicate records, missing values, and outliers. A robust preprocessing pipeline was developed to ensure data quality.
3.2.1 Anomaly Removal
Duplicate cell barcodes and repeated phase records were identified and removed. Each data file was checked for consistency in the number of rows per phase. Cells with missing records or extra measurements were discarded.
3.2.2 Outlier Detection
Outliers were detected by plotting distributions of key variables. For example, the temperature values at a specific time point during charging were expected to lie within a narrow band. Points outside the range (45°C, 52°C) were treated as outliers and removed. This process was applied to all features to avoid extreme values that could distort model training.
3.2.3 Smoothing Using Savitzky-Golay Filter
The charging voltage curves exhibited significant noise due to sensor measurement and electrochemical fluctuations. To reduce noise while preserving the underlying trend, a Savitzky-Golay (SG) filter was applied. The SG filter performs local polynomial regression and is defined as:
$$ T_j^* = \frac{\sum_{i=-m}^{m} C_i T_{j+i}}{N} $$
where $T_j$ is the original time series, $T_j^*$ is the smoothed value, $C_i$ are the filter coefficients, $N$ is the convolution window size, and $m$ is the half-window length. The SG filter effectively smooths the voltage data without significantly distorting the peaks and valleys that carry important information about battery state.
3.2.4 Capacity Correction Based on Temperature
The measured discharge capacity is affected by the ambient and cell temperature. To standardize the capacity values, a temperature correction factor was applied:
$$ \eta_T = 1 + K (T – T_0) $$
where $T$ is the actual cell temperature, $T_0$ is the reference temperature (25°C), and $K$ is the temperature coefficient. The corrected capacity is then:
$$ Q_{\text{corr}} = Q_{\text{meas}} \times \eta_T $$
This correction reduces the variation caused by temperature differences during the grading process.
3.3 Feature Extraction and Selection
The original formation data recorded the following basic features: voltage ($V$), pressure ($P$), temperature ($T$), and contact impedance ($R$). To improve the prediction accuracy, derived features were also extracted. The final set of features is summarized in the table below.
| Type | Feature | Description |
|---|---|---|
| Basic | $V$ | Voltage measured at fixed intervals |
| Basic | $P$ | Pressure measured at fixed intervals |
| Basic | $T$ | Temperature measured at fixed intervals |
| Basic | $R$ | Contact impedance |
| Derived | $dQ/dV$ | Differential capacity slope |
| Derived | $V_{ave}$ | Average voltage during rest stages |
| Derived | $P_{ave}$ | Average pressure during rest stages |
| Derived | $T_{ave}$ | Average temperature during rest stages |
The $dQ/dV$ feature was computed for each charging stage. Due to the different shapes of the voltage curves, the time interval for calculating the slope was adapted for each stage. For instance, in stage I, a 20-second interval was used for the first 50 seconds, followed by a 100-second interval for the remainder. This adaptive sampling ensures that the derived feature captures the most relevant phase transitions.
3.4 Single Model Construction and Feature Analysis
Three base models were constructed: LightGBM, KNN, and decision tree. All features were normalized before training. The dataset was split into 70% training and 30% testing. Three different feature groups were evaluated to determine the optimal feature set:
- Group 1: $V$, $P$, $T$, $R$
- Group 2: $V$, $P$, $T$, $R$, $dQ/dV$
- Group 3: $V$, $P$, $T$, $R$, $V_{ave}$, $P_{ave}$, $T_{ave}$
The prediction performance was evaluated using Mean Absolute Percentage Error (MAPE) and Max Error. The results are shown below.
| Model | Metric | Group 1 | Group 2 | Group 3 |
|---|---|---|---|---|
| LightGBM | MAPE (%) | 0.465 | 0.437 | 0.453 |
| Max Error (%) | 2.223 | 2.170 | 2.217 | |
| KNN | MAPE (%) | 0.455 | 0.463 | 0.459 |
| Max Error (%) | 2.224 | 2.165 | 2.219 | |
| Decision Tree | MAPE (%) | 0.470 | 0.464 | 0.469 |
| Max Error (%) | 2.285 | 2.186 | 2.323 |
From the table, Group 2 yields the lowest MAPE for LightGBM and the lowest Max Error for KNN. The addition of $dQ/dV$ significantly improves the performance for all models. In contrast, adding the average rest-stage features (Group 3) does not provide further improvement; hence Group 2 was selected as the optimal feature set.
3.5 Ensemble Model and Prediction Results
To further enhance the prediction accuracy, a weighted average ensemble model was constructed using the predictions from LightGBM, KNN, and decision tree. The ensemble output is given by:
$$ \hat{y}_{\text{ensemble}} = w_1 \hat{y}_{\text{LGBM}} + w_2 \hat{y}_{\text{KNN}} + w_3 \hat{y}_{\text{DT}} $$
with $w_1 + w_2 + w_3 = 1$. The weights were heuristically assigned based on the individual model performance. LightGBM, having the lowest MAPE, was assigned the highest weight; KNN, having the lowest max error, received a moderate weight; decision tree, though less accurate, contributed to the diversity of the ensemble. The optimal weights were set to $w_1 = 0.5$, $w_2 = 0.3$, and $w_3 = 0.2$ after several trials. The ensemble model was tested on the held-out dataset, yielding a MAPE of 0.433% and a Max Error of 2.159%.
Table 3 summarizes the prediction performance of individual models and the ensemble for the optimal feature group (Group 2).
| Model | MAPE (%) | Max Error (%) |
|---|---|---|
| LightGBM | 0.437 | 2.170 |
| KNN | 0.463 | 2.165 |
| Decision Tree | 0.464 | 2.186 |
| Ensemble | 0.433 | 2.159 |
The ensemble model demonstrates that fusing complementary models can slightly but consistently improve the prediction accuracy. With a MAPE below 0.5%, the model is highly acceptable for industrial production, where a tolerance of ±2% is often required. The proposed method reduces the grading time by utilizing only partial formation data, thus significantly lowering the manufacturing cost for energy storage battery producers.
4. Remaining Useful Life Prediction Model Development and Validation
4.1 NASA Battery Dataset and IC Curve Analysis
The NASA PCOE dataset is widely used for battery prognosis research. It contains cycling data for four 18650 lithium-ion cells (#5, #6, #7, #18) with different discharge cutoff voltages. The detailed test conditions are listed in Table 4.
| Cell | Charge cutoff (V) | Discharge cutoff (V) | Discharge current (A) | Temperature (°C) |
|---|---|---|---|---|
| #5 | 4.2 | 2.7 | 2 | 24 |
| #6 | 4.2 | 2.5 | 2 | 24 |
| #7 | 4.2 | 2.2 | 2 | 24 |
| #18 | 4.2 | 2.5 | 2 | 24 |
The capacity fading curves of these cells exhibit a general downward trend, but with temporary increases due to the so-called capacity regeneration phenomenon. This makes RUL prediction challenging because the capacity sequence is not monotonic.
To extract meaningful degradation features, the incremental capacity (IC) curve is calculated from the constant-current charging stage. The IC curve is defined as $dQ/dV$, which can be computed as:
$$ \frac{dQ}{dV} = I \cdot \frac{dt}{dV} $$
where $I$ is the charging current. The IC curve reveals the phase transitions of the electrode materials. As the battery ages, the peaks of the IC curve decrease in magnitude, and the area under the curve decreases as well.
For the NASA cells, the initial IC curves were noisy, so a Savitzky-Golay filter was applied to obtain smooth IC curves. Figure (from the original thesis) shows the smoothed IC curves for the first cycle. Then, for different cycles, the IC curves were plotted to observe the degradation trend. Based on the analysis, two health indicators (HIs) were extracted:
- HI1: Peak value of the IC curve in the voltage range 3.8 V–4.1 V.
- HI2: Peak area under the IC curve in the same voltage range.
Figure (in thesis) displays the IC curves at different cycles for each cell. It is visible that both the peak and the area decline with increasing cycle number.
4.2 Correlation Analysis of Health Indicators
To evaluate the effectiveness of the extracted HIs, both qualitative and quantitative analyses were performed. Qualitative: normalized HI1 and HI2 were plotted against the actual normalized capacity over cycles, showing similar trends. Quantitative: Pearson and Spearman rank correlation coefficients were computed between each HI and the capacity. The correlation coefficients are listed in Table 5.
| Cell | Pearson (HI1) | Pearson (HI2) | Spearman (HI1) | Spearman (HI2) |
|---|---|---|---|---|
| #5 | 0.9921 | 0.9949 | 0.9818 | 0.9876 |
| #6 | 0.9884 | 0.9891 | 0.9925 | 0.9923 |
| #7 | 0.9836 | 0.9865 | 0.9735 | 0.9848 |
| #18 | 0.9644 | 0.9821 | 0.9633 | 0.9665 |
All correlation coefficients are above 0.9, indicating strong linear and monotonic relationships between the proposed HIs and the battery capacity. Therefore, these HIs can serve as reliable proxies for estimating the SOH and RUL of energy storage battery cells.
4.3 SOH Prediction Using Single GPR Model
Initially, a single GPR model was built using the two HIs as input features and capacity as the output. The dataset for each cell was divided into training and testing sets. Three training scenarios were considered: using the first 60, 80, or 100 cycles. The GPR model was trained to predict the SOH (capacity) for the remaining cycles. The performance was evaluated using MAE and RMSE.
| Metric | Training Cycles | #5 | #6 | #7 | #18 |
|---|---|---|---|---|---|
| MAE | 60 | 0.2393 | 0.0284 | 0.3362 | 0.0737 |
| 80 | 0.0211 | 0.0316 | 0.0735 | 0.0267 | |
| 100 | 0.0283 | 0.0499 | 0.0267 | 0.0367 | |
| RMSE | 60 | 0.2498 | 0.0395 | 0.3460 | 0.0775 |
| 80 | 0.0317 | 0.0464 | 0.0812 | 0.0384 | |
| 100 | 0.0324 | 0.0618 | 0.0303 | 0.0436 |
As shown, the single GPR model performs reasonably when the training set contains more than 50% of the data, but its stability is poor when the training set is small (e.g., 60 cycles). Moreover, this model can only predict the current SOH using measured HIs; it cannot forecast future SOH without additional extrapolation. Therefore, an improved model is needed to perform RUL prediction.
4.4 Optimized GPR Model and HI-C-GPR Framework
To address the limitations of the single GPR model, two key improvements were introduced. First, the kernel function of the GPR was optimized to better capture the capacity regeneration phenomenon. Specifically, the mean function was set to a linear function:
$$ m(\boldsymbol{x}) = a \boldsymbol{x} + b $$
and the covariance function was formulated as the sum of a squared exponential kernel and a periodic kernel:
$$ k(\boldsymbol{x}, \boldsymbol{x}’) = \sigma_{f1}^2 \exp\left(-\frac{\|\boldsymbol{x} – \boldsymbol{x}’\|^2}{2 l_1^2}\right) + \sigma_{f2}^2 \exp\left(-\frac{2 \sin^2(\pi \omega (\boldsymbol{x} – \boldsymbol{x}’))}{l_2^2}\right) $$
where $\omega$ is the periodicity parameter. This composite kernel allows the model to fit both the smooth degradation trend and the oscillatory capacity regeneration behavior.
Second, a dual GPR framework (HI-C-GPR) was constructed. The first GPR model (HI-GPR) predicts the future HIs from the cycle number, while the second GPR model (C-GPR) maps the HIs to the capacity. Specifically:
- Input to HI-GPR: cycle number $n, n+1, \dots$
- Output of HI-GPR: predicted HI1 and HI2.
- Input to C-GPR: predicted HI1 and HI2 (or measured HIs for current SOH).
- Output of C-GPR: predicted capacity (SOH).
The framework is illustrated conceptually in the thesis. This dual-stage structure enables multi-step-ahead forecasting of the capacity trajectory, thus allowing RUL prediction.
4.5 RUL Prediction Results on NASA Dataset
The optimized HI-C-GPR model was applied to all four NASA cells. Three prediction starting points were selected: 61, 81, and 101 cycles, corresponding to using the first 60, 80, and 100 cycles as training data. The failure threshold was defined as 70% of the rated capacity (1.4 Ah for #5, #6, and #18). Cell #7 did not reach the threshold within the available cycles. The prediction results are shown in the original thesis figures. The quantitative evaluation metrics include MAPE, RMSE, MAE, and the absolute error (AE) of the predicted RUL (difference between actual and predicted failure cycles).
| Metric | Start Point | #5 | #6 | #7 | #18 |
|---|---|---|---|---|---|
| MAPE | 61 | 0.0170 | 0.0171 | 0.0157 | 0.0186 |
| 81 | 0.0076 | 0.0131 | 0.0084 | 0.0145 | |
| 101 | 0.0063 | 0.0086 | 0.0081 | 0.0144 | |
| RMSE | 61 | 0.0290 | 0.0327 | 0.0393 | 0.0330 |
| 81 | 0.0153 | 0.0291 | 0.0176 | 0.0252 | |
| 101 | 0.0124 | 0.0160 | 0.0152 | 0.0263 | |
| MAE | 61 | 0.0246 | 0.0239 | 0.0246 | 0.0267 |
| 81 | 0.0108 | 0.0176 | 0.0128 | 0.0203 | |
| 101 | 0.0087 | 0.0113 | 0.0119 | 0.0201 | |
| AE (cycles) | 61 | 4 | 0 | — | 4 |
| 81 | 3 | 1 | — | 3 | |
| 101 | 3 | 1 | — | 4 |
From Table 7, all MAPE values are below 0.02, and RMSE/MAE values are below 0.04. The AE of the RUL prediction is within 5 cycles for all cells. These results demonstrate that the optimized HI-C-GPR model significantly improves prediction stability and accuracy, and effectively mitigates the influence of capacity regeneration. Therefore, this model is suitable for RUL prediction in energy storage battery systems.
5. RUL Prediction for a 280 Ah Energy Storage Battery
5.1 Experimental Data of 280 Ah Cell
The 280 Ah prismatic LFP energy storage battery under study was subjected to an accelerated aging test with a total of 1000 charge-discharge cycles. The test conditions are listed in Table 8.
| Parameter | Value |
|---|---|
| Charge cutoff voltage | 3.65 V |
| Discharge cutoff voltage | 2.5 V |
| Discharge current | 280 A |
| Temperature | 60 °C |
The capacity fade curve of the 280 Ah cell is not monotonic; it contains local fluctuations due to the complex electrochemical processes. Nevertheless, the overall trend is decreasing.
5.2 IC Curve and Health Indicator Extraction
Using the constant-current charging data, the IC curve was calculated for each cycle. The IC curve of the 280 Ah cell exhibits three characteristic peaks in the voltage range 3.20 V–3.50 V. The most significant peak is located around 3.41 V–3.45 V, which shows a clear decline with cycling. Thus, the peak value (HI1) and the area under this peak (HI2) were extracted as health indicators. The SG filter was applied to smooth the IC curves.
Qualitative analysis, as shown in the thesis figure, indicates that the normalized HI1 tracks the normalized capacity very closely, while HI2 shows a slightly lower but still acceptable correlation. The quantitative correlation coefficients are given in Table 9.
| Indicator | Pearson | Spearman |
|---|---|---|
| HI1 | 0.9924 | 0.9954 |
| HI2 | 0.8787 | 0.9688 |
HI1 exhibits extremely strong correlation (above 0.99) with capacity, while HI2 has lower Pearson but high Spearman correlation, indicating some nonlinearity. Both HIs are considered usable for RUL prediction.
5.3 SOH Prediction Results
The HI-C-GPR model was applied to predict the SOH of the 280 Ah cell. Three prediction starting points were evaluated: 401, 601, and 801 cycles, using the first 400, 600, and 800 cycles as training data, respectively. The SOH prediction results for the test cycles are shown in the original thesis figures. The prediction curves closely match the actual capacity trajectories, with the fit improving as the training data amount increases.
| Metric | Start Point 401 | Start Point 601 | Start Point 801 |
|---|---|---|---|
| MAPE | 0.0009 | 0.0008 | 0.0006 |
| RMSE | 0.3123 | 0.3461 | 0.1897 |
| MAE | 0.2331 | 0.2134 | 0.1607 |
All MAPE values are below 0.001, which is exceptionally accurate. The RMSE and MAE also decrease as the training dataset grows, confirming that more data improves model performance.
5.4 RUL Prediction Results
The ultimate goal is to predict the remaining useful life, i.e., the cycle number at which the battery capacity falls below a failure threshold. For the 280 Ah cell, the failure threshold was set to 70% of the rated capacity, which equals 196 Ah. Since the available data only covers 1000 cycles, the RUL prediction extends beyond the observed range. The proposed model predicts the future capacity trajectory until it intersects the threshold.
The RUL prediction results for the three starting points are shown in the original thesis figures. The predicted failure cycles are summarized in Table 11.
| Prediction Start Point | Predicted Failure Cycle |
|---|---|
| 401 | 2491 |
| 601 | 3259 |
| 801 | 3413 |
The predictions at start points 601 and 801 are relatively close (3259 vs. 3413 cycles), suggesting that the model has converged to a stable estimate. The difference between the 401 start point and the later ones indicates that the model benefits from observing more degradation patterns. The predicted RUL values provide valuable guidance for maintenance scheduling and replacement planning of energy storage battery systems.
To facilitate industrial use, the model can also output the SOH (remaining capacity) after each cycle up to the failure point. This information is presented in tables or curves, allowing operators to track the degradation level in real time.
6. Conclusion and Future Work
This thesis presented a comprehensive data-driven framework for predicting the split capacity and remaining useful life of energy storage battery cells, specifically focusing on a 280 Ah prismatic LFP battery. The main conclusions are as follows:
- For split capacity prediction, a weighted ensemble model combining LightGBM, KNN, and decision tree was developed. Using an optimal feature set that includes the derived $dQ/dV$ feature, the ensemble achieved a MAPE of 0.433% and a maximum error of 2.159% on the test dataset. This accuracy is sufficient for production line applications, significantly reducing the time and energy required for capacity grading.
- For RUL prediction, a health indicator extraction method based on incremental capacity analysis was proposed. Two HIs—the IC peak and the peak area—were extracted and validated through correlation analysis. An optimized GPR model with a composite kernel (squared exponential plus periodic) was integrated into a dual-stage HI-C-GPR framework. Validation on the NASA dataset showed MAPE below 0.02 and RMSE below 0.04, with RUL prediction errors within 5 cycles.
- The proposed method was successfully applied to the 280 Ah energy storage battery. The SOH prediction errors were below 0.001 in MAPE, and the RUL predictions were consistent across different starting points, demonstrating the model’s robustness and practical utility.
Future work can explore several directions:
- Incorporating more diverse features, such as temperature gradients and impedance spectra, to further improve prediction accuracy.
- Investigating deep learning models (e.g., LSTM, transformer) for capacity and RUL prediction, which may capture more complex nonlinear degradation dynamics.
- Studying the effects of different operating conditions (temperature, charge/discharge rates) on the aging of energy storage battery cells to build universal prediction models.
- Developing online learning algorithms that can continuously update the model as new data becomes available, improving adaptability to varying degradation modes.
In summary, the data-driven methods developed in this thesis provide reliable and cost-effective solutions for both manufacturing quality control and operational safety management of energy storage battery systems.