Data-Driven Remaining Life Prediction for EV Battery Pack

Over the past decade, the rapid growth of electric vehicles has highlighted the indispensable role of the traction battery pack. The EV battery pack is expected to deliver reliable performance for many years, yet its capacity gradually fades due to complex electrochemical and operational processes. Accurately predicting the remaining useful life (RUL) of an EV battery pack has become one of the most critical tasks in battery management system design. In this thesis, I focus on a data-driven framework that combines advanced signal decomposition, feature selection, and deep learning algorithms to estimate the capacity trajectory of lithium-ion batteries. The proposed method effectively reduces the impact of capacity regeneration and measurement noise, thereby improving the overall prediction accuracy for an EV battery pack .

Why is RUL prediction so important for an EV battery pack? The battery pack is the most expensive and safety-critical component in an electric vehicle. If the degradation state cannot be assessed accurately, the vehicle may suffer from insufficient power, unexpected shutdown, thermal runaway, or premature maintenance. A reliable RUL prediction algorithm allows users to schedule battery replacement proactively, reduces the risk of failure, and contributes to the sustainable life-cycle management of an EV battery pack . Moreover, under the pressure of carbon neutrality, extending the service life of an EV battery pack by accurate prognosis is an effective way to reduce the total carbon footprint and the ownership cost of electric vehicles.

1. Fundamentals of Lithium-Ion Battery Aging

Lithium-ion cells are the core energy storage units inside an EV battery pack. They operate by repeatedly shuttling lithium ions between the positive electrode and the negative electrode during charge and discharge. Although the electrochemical reaction appears reversible in an ideal sense, many side-reactions gradually degrade the cell components. The aging of an EV battery pack originates from several intrinsic factors:

First, the electrode materials suffer from structural degradation. Repeated lithium-ion insertion and extraction causes lattice expansion and contraction, leading to particle cracking, active material loss, and increasing impedance. Second, the electrolyte may decompose or be oxidized at high voltage, producing gases and solid deposits that block ion transport. Third, the solid electrolyte interphase (SEI) film at the anode continuously forms and repairs; while it prevents further electrolyte decomposition, its growth consumes active lithium and increases diffusion resistance. External factors such as high temperature, high charge/discharge rates, over-charge, and over-discharge also accelerate these degradation mechanisms. In practice, the capacity fade observed in an EV battery pack is the superposition of all these internal and external aging modes.

The state-of-health (SOH) and the remaining useful life (RUL) of an EV battery pack are defined on the basis of the current capacity. I adopt the capacity-based definition:

$$SOH_i = \frac{C_i}{C_0} \times 100\%$$

where \(C_i\) is the maximum available capacity at cycle \(i\), and \(C_0\) is the rated capacity. In the RUL task, the battery is considered to reach its end-of-life (EOL) when the SOH falls below a certain threshold, usually 70%–80%. Therefore, the RUL at a starting cycle \(\alpha\) is computed as:

$$RUL_{\alpha} = \beta – \alpha$$

where \(\beta\) is the cycle number corresponding to the EOL threshold. The precise prediction of \(RUL_{\alpha}\) for an EV battery pack relies on the ability to extrapolate the future capacity fading trajectory from the observed historical data.

2. Data Sources and Health Indicator Construction

To validate the proposed estimation algorithm for an EV battery pack, I employed two publicly available lithium-ion battery aging datasets: the NASA battery dataset and the CALCE battery dataset. Both datasets contain repeated charge/discharge experiments on commercial 18650 or prismatic cells, but they differ in cycle numbers, cut-off voltages, and capacity retention patterns. Table 1 summarizes the basic information of the tested batteries.

Dataset Battery ID Nominal capacity Cut-off voltage Number of cycles
NASA B05, B06, B07, B18 2 Ah 2.7 V / 2.5 V / 2.2 V / 2.5 V ≈ 166
CALCE CS2_35 1.1 Ah 2.7 V ≈ 960

In the raw records, the direct-charge and discharge channels contain many operational parameters, including terminal voltage, current, temperature, and time stamps. However, simply feeding these raw high-dimensional features into a learning model would introduce severe redundancy and might not reflect the actual health state of an EV battery pack . Therefore, I reconstructed the battery capacity from the discharge process using the coulomb-counting method:

$$C_i = \int_{t_1}^{t_2} i(t)\, dt$$

where \(t_1\) and \(t_2\) are the start and end instants of the constant-current discharge, and \(i(t)\) is the discharge current. Since the measured data are discrete, the integral is approximated by cumulative summation. The resulting capacity sequence inherently describes the capacity retention trend of the cycled batteries and serves as the target health indicator (HI).

Even after the coulomb-counting reconstruction, the capacity sequence still contains sharp spikes caused by slight temperature changes, current fluctuations, and capacity-recovery phenomena. These artifacts may lead to unstable predictions for an EV battery pack . To deal with abnormal outliers, I used a sliding-window 3\(\sigma\) rule. For each window, the mean \(\mu\) and standard deviation \(\sigma\) are computed, and any sample falling outside the interval \((\mu – 3\sigma, \mu + 3\sigma)\) is replaced by the window median. Then, a Savitzky–Golay (S–G) filter is applied to further smooth the sequence. The S–G filter preserves the overall descending trend while removing minor local oscillations. The principle is to fit a polynomial in a sliding window using least squares; the fitted value is given by

$$f_i = \sum_{k=0}^{n} a_k i^k$$

where \(n\) is the polynomial order, and the coefficients \(a_k\) are determined from the data points within the window. In my experiments, a window length of 15 and polynomial order of 3 performed well for all tested batteries. The smoothed capacity curve retains the main degradation shape of an EV battery pack , but significantly reduces the unwanted capacity-recovery jumps. Finally, I applied Min-Max normalization to scale the capacity values into [0, 1] before model construction.

3. CEEMDAN-Based Multi-Scale Decomposition

Lithium-ion batteries often show capacity regeneration at certain cycles. The brief plateau or even rise in capacity makes it difficult for a single neural network to capture both the long-term degradation trend and the short-term recoveries. To separate these components, I first employed the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN). CEEMDAN is an advanced version of empirical mode decomposition (EMD) that adaptively adds Gaussian white noise to avoid mode mixing.

The decomposition process for a normalized capacity sequence \(y(t)\) can be summarized as follows:

  • Create \(N\) noisy copies: \(y_n(t) = y(t) + \delta_0 \eta_n(t)\), where \(\eta_n(t)\) is the \(n\)th added white noise sequence.
  • Apply EMD to each \(y_n(t)\) and average their first intrinsic mode functions to obtain the first CEEMDAN mode \(\overline{c_1}(t)\).
  • Subtract \(\overline{c_1}(t)\) from \(y(t)\) to get the first residue \(r_1(t)\).
  • For each stage \(m\), add noise to the current residue, perform EMD, and average the first IMF to obtain \(\overline{c_{m+1}}(t)\).
  • Continue until the residue becomes monotonic or has fewer than two extrema, and the original signal is reconstructed as

$$y(t) = \sum_{m=1}^{M} \overline{c_m}(t) + r_M(t)$$

After CEEMDAN, the capacity series is decomposed into several intrinsic mode functions (IMFs) and a residual term. The IMFs correspond to different oscillation scales, including the capacity-recovery fluctuations, while the residual \(r_M(t)\) captures the global capacity fade tendency of an EV battery pack . In my study, the NASA B05/B06/B07 batteries produced four IMFs, the B18 battery produced three IMFs, and the CALCE CS2_35 battery produced six IMFs because of its longer cycle history. Figure 1 illustrates the decomposition results for several batteries. The residual components from all batteries show very similar monotonically decreasing trends, which confirms that the CEEMDAN preprocessing can separate stable degradation information from noise.

4. Feature Contribution Evaluation Using GBRT

Not all IMFs contribute equally to the final capacity trend. The high-frequency IMFs often contain noise or very weak local fluctuations, while the low-frequency residue is the most important part. To quantify the contribution of each sub-sequence, I trained a Gradient Boosting Regression Tree (GBRT) regressor. In GBRT, the response variable is the original capacity series and the features are the CEEMDAN components. The model learns to combine these features through an ensemble of decision trees. Each tree node split that uses a certain feature reduces the loss; the total loss reduction attributed to each feature is summed over all trees and normalized to obtain the feature importance.

The feature importance score \(I(f)\) is computed as:

$$I(f) = \frac{\sum_{t=1}^{T} Gain(f,t)}{\sum_{f’ \in F} \sum_{t=1}^{T} Gain(f’,t)}$$

where \(Gain(f,t)\) is the total reduction in loss brought by feature \(f\) in tree \(t\), and \(T\) is the total number of trees. Table 2 reports the calculated importance scores for the components of each battery. The model \(R^2\) was above 0.999 for all cases, indicating that the CEEMDAN components could almost perfectly reconstruct the original capacity data.

Battery B05 B06 B07 B18 CS2_35
\(R^2\) 0.999981 0.999959 0.999984 0.999930 0.999888
IMF1 0.000932 0.002295 0.001471 0.005258 0.000035
IMF2 0.000944 0.002475 0.000576 0.006381 0.000321
IMF3 0.003718 0.000735 0.033981 0.988359 0.003323
IMF4 0.994408 0.944942 0.994555 — 0.000736
IMF5 — — — — 0.010203
IMF6 — — — — 0.985381

It can be observed that the residual component (the last IMF in the table) is generally the dominant contributor, while the importance of higher-frequency IMFs varies among cells. To improve prediction stability, I retained those components with an importance score above a predefined low threshold and assigned weights proportional to their GBRT importance. In the final weighted reconstruction, the residual term was normalized with a fixed weight of 1 because it uniquely represents the irreversible aging progression of an EV battery pack .

5. Long Short-Term Memory Networks for Subsequence Prediction

Recurrent neural networks (RNNs) are commonly used for time-series forecasting, but vanilla RNNs suffer from gradient vanishing when learning long-term dependencies. The Long Short-Term Memory (LSTM) architecture overcomes this problem by introducing a cell state and three gating units: forget gate, input gate, and output gate. I employed LSTM networks to model each retained subsequence. The LSTM update equations are:

$$f_t = \sigma(W_f x_t + U_f h_{t-1} + b_f)$$
$$i_t = \sigma(W_i x_t + U_i h_{t-1} + b_i)$$
$$o_t = \sigma(W_o x_t + U_o h_{t-1} + b_o)$$
$$\tilde{C}_t = \tanh(W_c x_t + U_c h_{t-1} + b_c)$$
$$C_t = f_t \odot C_{t-1} + i_t \odot \tilde{C}_t$$
$$h_t = o_t \odot \tanh(C_t)$$

where \(x_t\) is the input subsequence value at time \(t\), \(h_t\) is the hidden state, \(C_t\) is the cell state, and \(\odot\) denotes element-wise multiplication. The LSTM can capture both short-term fluctuations and long-term tendencies, which is exactly what is required for an EV battery pack capacity series.

For each battery, I built a two-layer LSTM network with dropout layers to avoid overfitting. I applied grid search on the hyperparameters including learning rate, number of epochs, window size, batch size, and the number of hidden units. Mean square error was chosen as the loss function, and the Adam optimizer was adopted with an initial learning rate of 0.001. The hyperparameters for the NASA and CALCE datasets are given in Table 3.

Hyperparameter NASA CALCE CS2_35
Loss function MSE MSE
Optimizer Adam Adam
Learning rate 0.001 0.001
Epochs 200 300
Window size 8 32
Batch size 16 64
Hidden units [64, 32] [64, 64]

To evaluate the RUL prediction performance, I used three metrics: mean absolute error (MAE), root mean square error (RMSE), and relative error (RE). These are defined as:

$$MAE = \frac{1}{n} \sum_{t=1}^{n} |x_t – \hat{x}_t|$$
$$RMSE = \sqrt{ \frac{1}{n} \sum_{t=1}^{n} (x_t – \hat{x}_t)^2 }$$
$$RE = \frac{|x_t^{\text{RUL}} – \hat{x}_t^{\text{RUL}}|}{x_t^{\text{RUL}}}$$

where \(x_t\) is the observed capacity, \(\hat{x}_t\) is the predicted capacity, and \(x_t^{\text{RUL}}\) and \(\hat{x}_t^{\text{RUL}}\) are the actual and predicted numbers of remaining cycles, respectively.

6. Experimental Results and Comparison

I validated the proposed CEEMDAN-GBRT-LSTM model on the NASA B05, B06, B07, B18 batteries and on the CALCE CS2_35 battery. The starting points for prediction (SP) were set to cycle 80 for the four NASA cells and cycle 400 for the CS2_35 cell. For comparison, I also tested a single LSTM model and a CEEMDAN-LSTM model without the GBRT-based weighting. Table 4 summarizes the numerical results obtained from these models.

Battery SP Model MAE RMSE RE
B05 80 LSTM 0.0685 0.0739 0.0923
CEEMDAN-LSTM 0.0291 0.0324 0.0783
CEEMDAN-GBRT-LSTM 0.0169 0.0203 0.0137
B06 80 LSTM 0.0308 0.0358 0.0754
CEEMDAN-LSTM 0.0229 0.0289 0.0283
CEEMDAN-GBRT-LSTM 0.0209 0.0263 0.0094
B07 80 LSTM 0.0291 0.0314 0.0566
CEEMDAN-LSTM 0.0251 0.0283 0.0342
CEEMDAN-GBRT-LSTM 0.0167 0.0193 0.0240
B18 80 LSTM 0.0423 0.0551 0.0864
CEEMDAN-LSTM 0.0281 0.0351 0.0638
CEEMDAN-GBRT-LSTM 0.0224 0.0269 0.0106
CS2_35 400 LSTM 0.0539 0.0684 0.0902
CEEMDAN-LSTM 0.0421 0.0539 0.0747
CEEMDAN-GBRT-LSTM 0.0419 0.0507 0.0342

The results show that the single LSTM model has the worst prediction performance in all cases because it cannot separate the capacity-regeneration fluctuations from the actual degradation trend. After applying CEEMDAN, the prediction errors are substantially reduced, especially in terms of MAE and RMSE. However, CEEMDAN-LSTM still includes some low-importance IMFs, which introduces unnecessary noise to the model. The proposed CEEMDAN-GBRT-LSTM model uses the GBRT importance ranking to assign appropriate weights to each component and discards the components with negligible contribution. Consequently, it yields the best matching with the true capacity curve.

From the comparison table, the relative error of the CEEMDAN-GBRT-LSTM method is less than 3.4% for all evaluated batteries, while the single LSTM relative errors are 5.7% to 9.2%. More specifically, the proposed algorithm reduces the RE by more than 50% compared with the LSTM method on the NASA B05, B06, B18, and CS2_35 batteries, and by about 57.5% on the B07 battery. In addition, the RMSE values of the proposed method remain below 0.051 for all cells, indicating a high robustness not only for regular 18650 cells but also for a longer-cycling prismatic battery in the CALCE dataset. These findings demonstrate that the proposed data-driven framework is highly effective for RUL prediction of an EV battery pack under different capacity-degradation patterns.

7. Visualization Software Platform

To make the proposed method practically accessible, I developed a visualization software platform for RUL prediction of an EV battery pack using Python and PyQt5. The platform architecture consists of three layers: the data layer, the model layer, and the visualization layer. The data layer manages user credentials and battery datasets through a MySQL database. The model layer performs the CEEMDAN decomposition, GBRT importance weighting, and LSTM network training and prediction. The visualization layer presents the raw capacity curves, the smoothed curves, and the prediction results with intuitive charts and tables.

The platform has four main functional modules:

  • User authentication: ensures secure login and logout with usernames and passwords.
  • Data management: allows importing NASA or CALCE datasets and displaying the original capacity data in tabular form.
  • Feature extraction: transforms the raw recorded variables to a condition-monitoring index through the coulomb-counting integration and S–G smoothing. The platform allows the user to select a specific battery and plot the extracted capacity degradation curve.
  • Model prediction: users can select a battery dataset and a starting prediction point, and then click a button to run the trained prediction model. The interface displays the predicted RUL curve of the EV battery pack together with the MAE, RMSE, and RE of different algorithms.

During the system functional tests, the login module successfully rejected incorrect credentials and displayed warning messages. The data import operation was completed with 100% accuracy. The prediction module operated stably without any crashes or interface blocking. The software platform therefore offers a simple, interactive and reliable tool for demonstrating RUL predictability and for supporting maintenance decisions on an EV battery pack .

8. Conclusion and Outlook

In this work, I developed a complete data-driven framework for the remaining useful life prediction of lithium-ion batteries used in an EV battery pack . The main contributions are summarized in three aspects. First, I constructed a clean battery health indicator by applying coulomb counting, 3\(\sigma\) outlier correction, and Savitzky–Golay smoothing. This reduced the negative effect of measurement noise and capacity regeneration. Second, I proposed a hybrid CEEMDAN-GBRT-LSTM algorithm. The CEEMDAN operation separates the global capacity fading trend from local regenerations; the GBRT algorithm ranks the contribution of each decomposed component; the LSTM network predicts each selected component and combines them by weighting. Third, I designed a graphical software platform that integrates all the procedures into a user-friendly interface for practical battery prognosis.

The experimental evaluation on NASA and CALCE battery datasets demonstrates that the CEEMDAN-GBRT-LSTM model outperforms both the individual LSTM model and the CEEMDAN-LSTM model in terms of MAE, RMSE, and RE. The relative error is improved by more than 50% compared to LSTM, which confirms the effectiveness of this hybrid method for the RUL prediction of an EV battery pack .

Future work may focus on adapting the proposed framework to partial charge/discharge cycles, real-world vehicle driving profiles, and more complex battery pack configurations where cells exhibit different aging paths. Advanced neural architectures such as transformers and attention networks may also be introduced to further increase the prediction accuracy. Online learning mechanisms can be embedded in the software platform to continuously update the model when new data from an operating EV battery pack become available.

Scroll to Top