As a researcher working on energy power systems, I have become increasingly convinced that the safety, reliability, and lifetime management of the electric vehicle battery pack are among the most important technical challenges for the clean transportation transition. The electric vehicle battery pack is not a static component. It undergoes continuous charge and discharge cycles, thermal gradients, mechanical vibration, and electrochemical side reactions. These processes cause gradual aging. Once aging passes a certain threshold, the risk of failure rises sharply. Therefore, I focused my work on how machine learning can predict health information for the electric vehicle battery pack, including current cycle number, remaining cycle number, state of health, and state of charge.
The motivation is straightforward. Global warming has intensified the pressure to reduce greenhouse gas emissions. The transportation sector accounts for a large share of carbon dioxide emissions, and electric vehicles are widely regarded as a practical pathway toward lower emissions. However, the electric vehicle battery pack remains the most expensive and safety-critical subsystem. Its degradation is coupled with manufacturing variation, usage patterns, ambient temperature, charging strategy, and driver behavior. Even batteries of the same model can age differently. For this reason, an accurate health information prediction method for the electric vehicle battery pack must combine data processing, algorithm selection, hyperparameter tuning, and model validation.
In this article, I present my research on machine learning methods for predicting health information of the electric vehicle battery pack. I discuss model-driven and data-driven approaches, feature importance evaluation, data preprocessing, algorithm selection, hyperparameter optimization, ensemble learning, neural networks, Bayesian probability models, and a fusion model that combines neural networks with gradient boosting decision trees. I also use several public battery degradation datasets, but I refer to them generically as Dataset A, Dataset B, and Dataset C to avoid relying on external identifiers.
1. Background and Significance
The electric vehicle battery pack is central to the performance and safety of electric vehicles. Its state of health is often defined by capacity fade or resistance increase. According to widely used engineering criteria, when the state of health falls to about 80%, the electric vehicle battery pack should be considered for replacement. However, the exact end-of-life condition depends on the application. For a vehicle, a gradual capacity loss may be acceptable for some time; for a stationary energy storage system, the requirement may be different.
Health information prediction for the electric vehicle battery pack includes several related tasks. The first is state of health estimation, which describes the current capacity relative to the initial capacity. The second is state of charge estimation, which describes the remaining charge available under current conditions. The third is remaining useful life prediction, which estimates how many cycles remain before the electric vehicle battery pack reaches a defined failure threshold. The fourth is current cycle number prediction, which is useful for warranty analysis and second-life assessment.
Model-driven methods include electrochemical models and equivalent circuit models. Electrochemical models describe internal reactions, diffusion, and aging mechanisms. Equivalent circuit models use resistors, capacitors, and voltage sources to approximate terminal behavior. These methods have clear physical meaning and can be accurate under controlled conditions. However, they require many parameters, are computationally expensive, and may not generalize well when the electric vehicle battery pack is used under diverse conditions.
Data-driven methods learn a mapping from measured signals to health information. They can use current, voltage, temperature, impedance, time, and other features. With the growth of sensing technology, more internal and external variables can be measured. This makes data-driven prediction for the electric vehicle battery pack increasingly attractive. The main idea is to train a model that fits the relationship between input features and health outputs. The quality of the model depends on both data and algorithm. Data determines the upper bound of achievable accuracy, while the algorithm determines how close the model can get to that upper bound.
2. Datasets and Modeling Methodology
I used three public battery datasets in my experiments. Dataset A contains commercial lithium-ion cells cycled under fast-charging conditions until failure. Dataset B contains cycling and impedance measurements for a set of lithium-ion cells. Dataset C contains small pouch cells tested under an urban driving profile. Although these datasets differ in chemistry, size, and test protocol, they all provide time-series measurements related to the electric vehicle battery pack. Table 1 summarizes the datasets.
| Dataset | Cell type | Main measurements | Purpose in my work |
|---|---|---|---|
| Dataset A | Commercial lithium-ion cells | Maximum charge power, maximum discharge power, internal resistance, temperature, charge time, discharge power sequence | Main training and testing for ensemble learning, neural networks, Bayesian models, and fusion model |
| Dataset B | Lithium-ion cells with impedance tests | Inductive branch current, battery branch current, impedance, electrolyte resistance, charge transfer resistance | Universality test of the fusion model |
| Dataset C | Small pouch cells | Charge time, discharge time, maximum charge temperature, maximum discharge temperature, charge voltage, discharge voltage | Additional universality test of the fusion model |
For Dataset A, I used both summary features and cycle-level features. The summary features include maximum discharge power, maximum charge power, internal resistance, maximum temperature, average temperature, minimum temperature, and charge time. The cycle-level feature is a discharge power sequence of length 1000. This sequence is ordered in time, so it is naturally suited to recurrent neural networks. For Dataset B, I used impedance-related variables. For Dataset C, I used charge and discharge voltage, temperature, and time variables.
The general modeling pipeline for the electric vehicle battery pack is shown below in conceptual form:
| Step | Task | Output |
|---|---|---|
| 1 | Select battery parameters and define prediction target | Input feature matrix and output vector |
| 2 | Evaluate feature importance | Ranked features |
| 3 | Preprocess data | Clean, scaled, and formatted dataset |
| 4 | Split data | Training set, validation set, test set |
| 5 | Select machine learning algorithm | Candidate model family |
| 6 | Tune hyperparameters | Optimized model configuration |
| 7 | Train and evaluate | Prediction model and error metrics |
| 8 | Apply model to health management | State of health, state of charge, remaining useful life, cycle number |
2.1 Feature Importance Evaluation
Feature importance is essential for the electric vehicle battery pack because not all measurable signals are equally useful. I considered three practical indicators: acquisition difficulty, coverage, and accuracy. Acquisition difficulty describes how hard it is to measure a variable in real operation. Coverage describes how well the training data represent the operating range. Accuracy describes how reliable the measured value is. A feature with low acquisition difficulty, high coverage, and high accuracy is preferable.
I used a random forest to evaluate feature importance for Dataset A. Random forest is an ensemble of decision trees. It can capture nonlinear relationships and provide importance scores. I used 200 trees and default settings for other parameters. The input included all selected features, and the output was the current cycle number. Table 2 shows the relative importance when five features were used.
| Feature | Relative importance |
|---|---|
| Maximum discharge power | 0.52 |
| Internal resistance | 0.18 |
| Charge time | 0.15 |
| Maximum charge power | 0.12 |
| Average temperature | 0.11 |
When maximum temperature was added, the ranking changed. Maximum temperature became the most important feature, followed by charge time and maximum discharge power. This result confirms that feature importance is context-dependent. Different combinations of input variables can change the ranking. Therefore, I do not treat feature importance as a fixed property. Instead, I use it as a guide for iterative feature selection. For the electric vehicle battery pack, maximum temperature is often highly informative because thermal stress accelerates aging.
Table 3 shows a more detailed importance ranking when different features were excluded. I observed that internal resistance, which appeared less important in one setting, became more important when other features were removed. This mutual influence is important. It means that dropping a feature because it ranks low in one model may reduce accuracy in another model. For the electric vehicle battery pack, feature selection must be validated experimentally.
| Excluded feature | Top feature | Second feature | Third feature |
|---|---|---|---|
| No maximum discharge power | Internal resistance | Maximum temperature | Maximum charge power |
| No maximum charge power | Internal resistance | Maximum temperature | Average temperature |
| No internal resistance | Maximum temperature | Charge time | Maximum discharge power |
| No average temperature | Charge time | Maximum temperature | Maximum discharge power |
| No charge time | Maximum temperature | Internal resistance | Maximum discharge power |
| No maximum temperature | Maximum discharge power | Internal resistance | Charge time |
2.2 Data Preprocessing
After feature selection, I preprocessed the data. The electric vehicle battery pack dataset may contain missing values, outliers, different scales, and repeated records. Table 4 lists the preprocessing methods I considered.
| Method | Purpose | Application to electric vehicle battery pack data |
|---|---|---|
| Data cleaning | Handle missing, abnormal, and duplicate values | Remove impossible temperature or voltage readings |
| Feature scaling | Standardize or normalize features | Scale resistance, power, temperature, and time |
| Feature encoding | Convert categorical variables | Encode charging protocol or cell group |
| Data transformation | Log, power, or normal transformation | Reduce skewness of degradation-related variables |
| Feature construction | Create new features from raw signals | Build slope, variance, or cumulative features |
| Data balancing | Adjust imbalanced distributions | Balance early-life and late-life samples |
| Dimensionality reduction | Reduce feature count | Compress high-dimensional impedance spectra |
| Anomaly detection | Identify abnormal samples | Detect sensor faults in the electric vehicle battery pack |
| Time-series processing | Create lag or window features | Use discharge power sequences for recurrent models |
For tree-based models such as random forest, XGBoost, LightGBM, and CatBoost, feature scaling is not strictly required. For neural networks, scaling is important. For Bayesian models such as hidden Markov models, discretization or state mapping is often required. I therefore prepared multiple versions of the same dataset depending on the algorithm. This is a key lesson for health prediction of the electric vehicle battery pack: preprocessing should be algorithm-aware.
2.3 Machine Learning Algorithm Selection
I compared several algorithm families for the electric vehicle battery pack. Table 5 summarizes their strengths and weaknesses.
| Algorithm family | Examples | Strengths | Weaknesses |
|---|---|---|---|
| Ensemble learning | Random forest, XGBoost, LightGBM, CatBoost | High accuracy on tabular data, robust to feature scaling, interpretable feature importance | Hyperparameter tuning, risk of overfitting, limited handling of raw sequences |
| Neural networks | DNN, CNN, LSTM, hybrid DNN-LSTM | Flexible input formats, strong nonlinear mapping, good for large data | Many hyperparameters, long training time, risk of local minima |
| Bayesian probability models | Hidden Markov model | Probabilistic interpretation, useful for sequence states | Sensitive to discretization, limited generalization |
| Kernel methods | Support vector regression, Gaussian process regression | Strong theoretical basis, uncertainty estimates | Scaling issues, high computational cost |
I selected ensemble learning, neural networks, and Bayesian probability models for detailed study. The reason is that these three families represent different views of the electric vehicle battery pack data. Ensemble learning treats the data as tabular features. Neural networks treat the data as flexible numerical arrays and can process sequences. Bayesian probability models treat the data as observations generated by hidden states. Comparing them gives a broader understanding of health prediction for the electric vehicle battery pack.
2.4 Hyperparameter Optimization with Chaotic Sparrow Search
Machine learning models have hyperparameters that control model capacity, regularization, and training dynamics. For the electric vehicle battery pack, different datasets and feature sets require different hyperparameters. I used a chaotic sparrow search algorithm to tune hyperparameters. The sparrow search algorithm simulates producer, follower, and scout behavior. To improve global search and avoid local optima, I introduced Tent chaotic mapping and Gaussian mutation.
The producer position update is:
$$ x_{i,d}^{t+1} = \begin{cases} x_{i,d}^{t} \cdot (1+Q), & R_2 < ST \\ x_{i,d}^{t} + Q, & R_2 \ge ST \end{cases} $$
Here, \(x_{i,d}^{t}\) is the \(d\)-th dimension of the \(i\)-th individual at iteration \(t\), \(Q\) is a standard normal random number, \(R_2\) is a random number in [0,1], and \(ST\) is a warning threshold. The follower update is:
$$ x_{i,d}^{t+1} = x_{b,d}^{t} + \frac{1}{D}\sum_{d=1}^{D}\left(\text{rand}(-1,1) \cdot \left|x_{b,d}^{t} – x_{i,d}^{t}\right|\right) $$
where \(x_{b,d}^{t}\) is the best position in the current population. The scout update is:
$$ x_{i,d}^{t+1} = \begin{cases} x_{i,d}^{t} + \beta \left|x_{i,d}^{t} – x_{b,d}^{t}\right|, & f_i \ne f_g \\ x_{i,d}^{t} + \beta \left|x_{w,d}^{t} – x_{b,d}^{t}\right|, & f_i = f_g \end{cases} $$
where \(\beta\) is a standard normal random number, \(f_g\) is the best fitness, and \(x_{w,d}^{t}\) is the worst position. To improve diversity, I used Gaussian mutation when an individual was better than the population average:
$$ \text{mutation}(x) = x(1 + N(0,1)) $$
When an individual was worse than or equal to the average, I used Tent chaotic perturbation. The Tent map was modified as:
$$ z_{i+1} = (2z_i) \bmod 1 + \text{rand}(0,1)\cdot \frac{1}{N_T} $$
The chaotic perturbation was then computed as:
$$ X_{\text{new}}^{d} = d_{\min} + (d_{\max} – d_{\min})Z_d $$
$$ X_{\text{new}}’ = \frac{X’ + X_{\text{new}}}{2} $$
I used the validation set error as the fitness function. The algorithm proposed a set of hyperparameters, trained the model, evaluated validation error, and updated the population. After a fixed number of iterations, I selected the hyperparameter set with the lowest validation error. In my tests, the chaotic sparrow search escaped local optima effectively and found suitable hyperparameters for the electric vehicle battery pack prediction models.
3. Health Information Prediction Methods
In this section, I describe my experiments on predicting the current cycle number of the electric vehicle battery pack. I used the same input-output structure where possible: features from charge and discharge cycles as input, and the current cycle number as output. I split the data into training and test sets. For ensemble models, I used the same training and test partitions. For neural networks and Bayesian models, I used separate partitions because their input formats differ.
3.1 Ensemble Learning Results
I trained random forest, XGBoost, LightGBM, and CatBoost. Table 6 lists representative hyperparameters. These values were obtained through experience and chaotic sparrow search. I used mean squared error as the loss function because large prediction errors are especially undesirable for the electric vehicle battery pack.
| Algorithm | Hyperparameter | Value |
|---|---|---|
| Random forest | n_estimators | 200 |
| XGBoost | num_boost_round / max_depth / eta / gamma | 6000 / 10 / 0.0785 / 0.1361 |
| LightGBM | num_boost_round / learning_rate / num_leaves / max_depth | 26621 / 0.3 / 21 / 5 |
| CatBoost | iterations / depth / learning_rate / random_seed | 28000 / 12 / 0.02 / 23 |
Table 7 shows the prediction results on the test set. The test set contained 11,470 samples with a mean cycle number of 488. The columns for relative error show the percentage of predictions within different error ranges.
| Algorithm | MAE | MSE | Relative error > 20% | Relative error 10%-20% | Relative error < 10% |
|---|---|---|---|---|---|
| Random forest | 22.20 | 3375.61 | 10.75% | 7.94% | 81.31% |
| XGBoost | 27.66 | 3447.43 | 12.89% | 11.00% | 76.11% |
| LightGBM | 31.53 | 3933.45 | 15.30% | 12.98% | 71.72% |
| CatBoost | 25.43 | 2926.77 | 11.87% | 11.04% | 77.10% |
Random forest had the highest percentage of predictions with relative error below 10%. CatBoost had the lowest MSE. LightGBM had the lowest mean relative error when large errors occurred, which suggests stable behavior in some cases. However, all algorithms had too many predictions with relative error above 20% for practical use. I found that the largest errors occurred when the true cycle number was very small, often between 1 and 10. Because a new electric vehicle battery pack is rarely predicted purely from data in real applications, I discarded samples with cycle number below 85 for further analysis.
Table 8 shows the results after removing early-life samples. The test set then contained 10,318 samples with a mean cycle number of 537.64.
| Algorithm | MAE | MSE | Relative error > 20% | Relative error 10%-20% | Relative error < 10% |
|---|---|---|---|---|---|
| Random forest | 21.81 | 3354.76 | 6.70% | 7.15% | 86.14% |
| XGBoost | 27.60 | 3490.52 | 7.81% | 10.44% | 81.75% |
| LightGBM | 31.50 | 3985.01 | 9.59% | 12.92% | 77.50% |
| CatBoost | 25.33 | 2957.23 | 6.77% | 10.32% | 82.90% |
Removing early-life samples reduced large relative errors, but the MAE and MSE did not improve dramatically. This shows that early-life prediction is only one source of error. To improve accuracy further, I studied feature selection. Table 9 compares CatBoost with different feature sets.
| Feature set | MAE | MSE | Relative error > 20% | Relative error < 10% |
|---|---|---|---|---|
| Five features without internal resistance | 34.39 | 4602.80 | 11.44% | 74.77% |
| Six features | 18.60 | 1787.00 | 4.66% | 88.44% |
| Five features without average temperature | 24.12 | 2781.88 | 6.33% | 84.74% |
The six-feature set included maximum charge power, maximum discharge power, maximum temperature, charge time, average temperature, and internal resistance. Adding maximum temperature improved accuracy substantially. Removing internal resistance hurt accuracy. This confirms the mutual influence of features in the electric vehicle battery pack. When many candidate features are available, feature selection is necessary, but it must be done carefully.
I also performed an error analysis by grouping predictions according to relative cycle position. Table 10 shows the results.
| Relative cycle range | Number of predictions with >20% error | Total | Percentage |
|---|---|---|---|
| > 80 | 107 | 2351 | 4.55% |
| 20 to 80 | 284 | 6872 | 4.13% |
| < 20 | 91 | 1112 | 8.18% |
The largest absolute errors tended to occur when the true cycle number was small or when the input data were abnormal. In practice, I would mitigate this by taking the median or average of multiple predictions from consecutive cycles. This reduces the influence of noisy samples and improves reliability for the electric vehicle battery pack.
3.2 Neural Network Results
Neural networks offer flexible input handling. For Dataset A, I designed a hybrid network. Seven scalar features were passed through a DNN. A sequence of length 1000 was passed through an LSTM. The outputs were concatenated and passed through another DNN. This allowed the model to use both tabular and sequential information. The amount of input data was about 144 times larger than the ensemble learning input, but the model still did not outperform CatBoost in my tests. Table 11 compares the neural network and a hidden Markov model.
| Model | MSE | MAE | Test set mean |
|---|---|---|---|
| Neural network | 41515.08 | 50.84 | 484.26 |
| Hidden Markov model | 86396.39 | 213.59 | 500.26 |
Although CatBoost was more accurate on this dataset, the neural network is still valuable. It can handle different data formats and can be combined with other algorithms. For a large electric vehicle battery pack fleet, neural networks may become more competitive as data volume grows. The LSTM part is especially useful for charge and discharge sequences. The DNN part is useful for scalar features. This flexibility makes neural networks a natural component in fusion models.
3.3 Bayesian Probability Model Results
I also studied hidden Markov models for the electric vehicle battery pack. A hidden Markov model assumes that a hidden state generates observations and that the hidden state evolves according to a Markov process. I used two learning modes: unsupervised and supervised.
In unsupervised learning, the six variables were treated as observations, and the hidden state number was set to 5. I used the expectation-maximization algorithm to learn model parameters. Then I predicted the next observation in a sequence. Table 12 shows representative results. The predicted values followed the true trends for most variables.
| Variable | True value | Predicted value |
|---|---|---|
| Maximum discharge power | 1.07665 | 1.07582 |
| Maximum charge power | 1.07641 | 1.07475 |
| Internal resistance | 0.01656 | 0.01664 |
| Average temperature | 31.60815 | 31.59645 |
| Charge time | 13.34155 | 13.38827 |
| Cycle number | 22 | 18 |
In supervised learning, the cycle number was treated as the hidden state, and the other variables were observations. I discretized each variable into intervals. For cycle number, I mapped ranges such as 1-100 to state 1, 101-200 to state 2, and so on. Table 13 shows the prediction results for several sub-datasets. The model was stable and reasonably accurate.
| Sub-dataset | MAE | Error interval | Maximum cycle |
|---|---|---|---|
| 1 | 0.16 | [0.16, 115] | 690 |
| 2 | 0.17 | [0.17, 116] | 533 |
| 3 | 0.24 | [4.2, 123] | 901 |
| 4 | 0.24 | [1.23, 123] | 616 |
| 5 | 0.26 | [1.25, 125] | 598 |
The unsupervised model can supplement experimental data and shorten test cycles. Its limitation is that a single model does not generalize widely. The supervised model performs well but depends heavily on data quality and discretization. For the electric vehicle battery pack, hidden Markov models are best used when the health state can be naturally represented as a finite set of degradation stages.
3.4 General Method for Health Prediction
From the above experiments, I can summarize a general method for machine learning prediction of electric vehicle battery pack health information.
| Stage | Recommended action | Reason |
|---|---|---|
| Data selection | Use same cell type and operating range | The electric vehicle battery pack behavior is chemistry and protocol dependent |
| Feature evaluation | Use random forest and domain knowledge | Feature importance is context-dependent |
| Preprocessing | Clean, scale, and format according to algorithm | Tree models and neural networks have different requirements |
| Algorithm selection | Start with CatBoost or random forest | They are strong on tabular battery data and easy to tune |
| Hyperparameter tuning | Use chaotic sparrow search with validation error | It finds good parameters efficiently |
| Evaluation | Use MAE, MSE, and relative error bins | Large errors are critical for safety |
| Deployment | Take median or average of multiple predictions | Reduces noise and improves reliability |
For the electric vehicle battery pack, I recommend starting with regression trees because they are faster to train and easier to tune than neural networks. CatBoost is a strong first choice. If sequential data are available, a neural network can be added to extract additional features. This leads to the fusion model described in the next section.
4. Fusion Model for Improved Prediction
My ensemble experiments showed that CatBoost achieved good accuracy. My neural network experiments showed that neural networks can process different input formats, including sequences. I therefore designed a fusion model that combines a neural network with CatBoost. The idea is to use the neural network as a feature extractor and then use CatBoost as the final predictor. This is especially relevant for the electric vehicle battery pack because it may have both scalar measurements and time-series measurements.
4.1 Neural Network Architecture for Feature Extraction
The neural network had two main parts. The first part was a parallel structure: a DNN for scalar features and an LSTM for sequence features. The second part was another DNN that combined the outputs. Table 14 describes the cases I tested.
| Case | Training input format | Test input format | Model |
|---|---|---|---|
| Case 1 | 91678 × 7 | 22920 × 7 | DNN |
| Case 2 | 91678 × 1007 | 22920 × 1007 | DNN + LSTM |
| Case 3 | 91678 × 7 | 22920 × 7 | CatBoost |
The seven scalar features were maximum discharge power, maximum charge power, internal resistance, maximum temperature, average temperature, minimum temperature, and charge time. The 1000-length sequence was the discharge power sequence. The neural network processed the scalar features through a DNN and the sequence through an LSTM. The outputs were concatenated and passed through another DNN.
I used batch normalization, linear layers, ELU activation, and dropout. Batch normalization is defined as:
$$ \mu_\beta = \frac{1}{m}\sum_{i=1}^{m} x_i $$
$$ \sigma_\beta^2 = \frac{1}{m}\sum_{i=1}^{m}(x_i – \mu_\beta)^2 $$
$$ \hat{x}_i = \frac{x_i – \mu_\beta}{\sqrt{\sigma_\beta^2 + \epsilon}} $$
$$ y_i = \gamma \hat{x}_i + \beta $$
A linear layer is defined as:
$$ y = xA^T + b $$
The ELU activation function is:
$$ f(x) = \begin{cases} x, & x > 0 \\ \alpha(\exp(x)-1), & x \le 0 \end{cases} $$
Dropout was used between linear layers to reduce overfitting. The LSTM hidden size was 64. The final DNN had 32 output features before the prediction layer. The model output was the current cycle number of the electric vehicle battery pack.
4.2 CatBoost Prediction Characteristics
CatBoost uses symmetric decision trees as base learners. Each tree splits nodes in a symmetric way. The output of a tree is determined by the input feature values. Once the model is trained, the leaf values are fixed. This gives CatBoost stability and high accuracy on tabular data. However, CatBoost cannot directly process raw sequences in the same flexible way as an LSTM. This is why the fusion model is useful.
Table 15 shows the prediction results for the three cases. The test set mean was 484.26.
| Case | MSE | MAE |
|---|---|---|
| Case 1: DNN | 12074.13 | 70.80 |
| Case 2: DNN + LSTM | 3863.29 | 43.07 |
| Case 3: CatBoost | 1525.03 | 19.49 |
CatBoost was the best single model. The DNN-LSTM was better than the DNN because it used the sequence feature. This confirms that additional valid features improve prediction accuracy for the electric vehicle battery pack.
I also observed some negative predictions. Negative cycle numbers are physically impossible. Table 16 shows two examples.
| Input condition | Case 1 | Case 2 | Case 3 | True value |
|---|---|---|---|---|
| Abnormal Qd and Qc | -4284.45 | -367.76 | -9.56 | 40 |
| Normal input | -9.49 | Positive | Positive | 35 |
When the input data were abnormal, the DNN produced highly unreliable negative values. CatBoost was more stable. When the input data were normal, a negative value indicated a large relative error but could still be interpreted as a prediction close to zero. This highlights the importance of data quality for the electric vehicle battery pack.
4.3 Fusion Model Design
The fusion model combines the neural network and CatBoost as follows. I first trained the DNN-LSTM model on the scalar and sequence features. Then I extracted the inputs to the last fully connected layer, which produced 8 new features. These 8 features were concatenated with the original 7 scalar features, forming a 15-feature input for CatBoost. The new CatBoost model was then trained and evaluated.
Table 17 shows the prediction performance of the fusion model.
| Relative error range | Number | Percentage | MSE | MAE |
|---|---|---|---|---|
| > 20% | 993 | 4.33% | 641.44 | 11.55 |
| 10% to 20% | 3855 | 16.82% | ||
| < 5% | 18072 | 78.85% |
The fusion model reduced MAE from 19.49 to 11.55, a reduction of about 41%. This is a substantial improvement. The 8 new features are not physically interpretable, but they encode information from the sequence and scalar data. They help CatBoost split the feature space more effectively.
To test whether the new features were meaningful, I replaced them with random numbers in a control experiment. Table 18 compares the training and test behavior.
| Training iteration | Control training RMSE | Fusion training RMSE | Control test MSE | Fusion test MSE |
|---|---|---|---|---|
| 1 | 366.79 | 366.86 | Higher | Lower |
| 75 | 168.70 | 168.27 | Higher | Lower |
| 12000 | 9.94 | 17.52 | Higher | Lower |
The control model with random features overfit more easily and generalized worse. This confirms that the 8 neural network features carry useful information for predicting the health of the electric vehicle battery pack.
4.4 Universality and Cause Analysis
To test whether the fusion model was a special case, I applied it to Dataset B and Dataset C. Table 19 summarizes the results.
| Dataset | Model | MSE | MAE |
|---|---|---|---|
| Dataset B | DNN | 800.72 | 20.82 |
| Dataset B | CNN | 528.77 | 16.57 |
| Dataset B | DNN + CNN | 318.31 | 14.04 |
| Dataset B | CatBoost | 730.81 | 17.77 |
| Dataset B | Fusion | 162.65 | 9.49 |
| Dataset C | DNN | 80266.47 | 217.59 |
| Dataset C | LSTM + LSTM | 103736.17 | 230.22 |
| Dataset C | LSTM + LSTM + DNN | 70354.51 | 198.75 |
| Dataset C | CatBoost | 64814.20 | 140.50 |
| Dataset C | Fusion | 33783.78 | 138.36 |
The fusion model improved performance on both datasets. This supports the generality of the approach for the electric vehicle battery pack. I also tested cases where the neural network and CatBoost used the same input features. Table 20 shows the comparison.
| Dataset | Model | MSE | MAE |
|---|---|---|---|
| Dataset A | CatBoost | 1525.03 | 19.49 |
| Dataset A | CatBoost with extracted features | 1468.44 | 17.32 |
| Dataset B | CatBoost | 730.81 | 17.77 |
| Dataset B | CatBoost with extracted features | 807.00 | 19.10 |
| Dataset C | CatBoost | 64814.20 | 140.50 |
| Dataset C | CatBoost with extracted features | 21135.97 | 97.57 |
The results are mixed but mostly positive. In Dataset A and Dataset C, the extracted features improved CatBoost. In Dataset B, the extracted features did not improve the result. This suggests that the benefit depends on data distribution and model alignment. The new features may provide additional split options for the decision trees. They may also make the data distribution more distinctive, helping CatBoost classify and regress more effectively.
4.5 Application to Electric Vehicle Battery Pack Health Management
The fusion model can be used in an electric vehicle battery pack management system. The workflow is as follows. First, collect scalar features and sequence features from the battery management system. Second, preprocess the data and ensure sensor reliability. Third, pass the data through the trained DNN-LSTM feature extractor. Fourth, concatenate the extracted features with the original features. Fifth, use CatBoost to predict the health information. Finally, apply the prediction to state of health estimation, remaining useful life estimation, and maintenance scheduling.
| Application | Prediction target | Benefit for electric vehicle battery pack |
|---|---|---|
| Warranty analysis | Current cycle number | Verify usage and detect abnormal aging |
| Remaining useful life | Cycles to 80% state of health | Plan replacement and second-life use |
| State of charge | Available charge | Improve range estimation |
| State of health | Capacity fade | Support safety and maintenance decisions |
| Fault warning | Abnormal degradation | Reduce risk of thermal runaway |
Table 21 compares my fusion model with selected published results. The comparison is only indicative because datasets and test protocols differ, but it shows that the fusion model is competitive.
| Method family | Reported error range | Observation |
|---|---|---|
| Gaussian process regression | About 1.2% to 6.9% | Strong uncertainty quantification |
| Hybrid decomposition and neural network | About 0.006 to 0.11 | High accuracy but complex preprocessing |
| LSTM | About 1.8 to 1.95 | Good for sequences |
| My single models | MAE 19.49 to 70.80 | CatBoost best among single models |
| My fusion model | MAE 11.55 | Best result in my experiments |
For the electric vehicle battery pack, the fusion model offers a practical balance. It uses the neural network to handle complex and sequence data, and it uses CatBoost to perform accurate tabular regression. It can be trained on historical data and deployed for online prediction. The main requirement is high-quality data and careful validation.
5. Conclusions and Future Work
I have studied machine learning methods for predicting health information of the electric vehicle battery pack. My work covers data processing, feature importance, algorithm selection, hyperparameter optimization, ensemble learning, neural networks, Bayesian probability models, and a fusion model. The main conclusions are as follows.
First, feature importance for the electric vehicle battery pack is context-dependent. Maximum temperature and maximum discharge power are often highly informative. Internal resistance may appear less important in one feature set but becomes important when other features are removed. Random forest is a useful tool for evaluating feature importance, but the results must be interpreted carefully.
Second, data preprocessing must match the algorithm. Tree-based models can handle unscaled features, while neural networks require scaling and appropriate formatting. Time-series features such as discharge power sequences should be processed with recurrent or convolutional layers.
Third, CatBoost is a strong first choice for tabular health prediction of the electric vehicle battery pack. In my experiments, it achieved lower MSE than random forest, XGBoost, and LightGBM. Neural networks are more flexible but require more data and tuning. Bayesian models are useful for sequence states and probabilistic interpretation but are sensitive to discretization.
Fourth, the fusion model improved accuracy significantly. By extracting features from a DNN-LSTM network and feeding them into CatBoost, I reduced MAE by about 41%. The method worked on multiple datasets, which suggests that it is not a special case. The extracted features help CatBoost learn from sequence information and complex nonlinear relationships.
Fifth, practical deployment should include multiple-prediction aggregation. Taking the median or average of predictions from consecutive cycles reduces noise and improves reliability. This is important for the electric vehicle battery pack because a single abnormal measurement can cause a large error.
For future work, I plan to expand the feature engineering methods. I will test fast Fourier transform, wavelet scattering, and other signal processing techniques. I will also test more machine learning algorithms, including support vector regression, k-nearest neighbors, Gaussian process regression, and other ensemble methods. In the fusion model, I will replace CatBoost with other regressors to see whether the neural network feature extraction approach generalizes. Finally, I will investigate online learning and transfer learning for the electric vehicle battery pack, so that a model trained on one battery type or operating condition can adapt to another with limited new data.
In summary, my research shows that machine learning can predict health information for the electric vehicle battery pack with good accuracy. The best results come from combining careful data processing, appropriate algorithm selection, hyperparameter optimization, and model fusion. As sensing technology and computing power continue to improve, I expect that data-driven health prediction for the electric vehicle battery pack will become more accurate, more robust, and more widely used in real-world battery management systems.