Accurate prediction of the Remaining Useful Life (RUL) of a lithium ion battery is paramount for ensuring the safety, reliability, and economic viability of energy storage systems in applications ranging from electric vehicles to grid-scale storage. The degradation of a lithium ion battery is a complex electrochemical process, making precise modeling challenging. Data-driven methods, particularly those leveraging artificial intelligence, have shown significant promise as they can learn intricate degradation patterns directly from operational data without requiring explicit physical models. In this work, we propose a novel hybrid prediction framework that synergistically combines feature engineering, a deep neural network architecture, and an advanced metaheuristic optimization algorithm to achieve high-accuracy RUL estimation.
The core of our methodology involves three key innovations: the extraction and selection of high-correlation Health Indicators (HIs) using Neighborhood Component Analysis (NCA), the construction of a Deep Echo State Network (DESN) prediction model, and the optimization of the DESN’s hyperparameters using a novel Hybrid Differential Evolution-Sparrow Search Algorithm (DESSA). The overarching workflow begins with processing raw battery cycling data to construct a comprehensive set of potential HIs. These HIs are then refined via NCA to eliminate redundancy and retain the most informative features. These selected health factors serve as the input to a DESN, whose critical parameters are meticulously tuned by the DESSA optimizer to create the final DESSA-DESN prediction model.
The performance and generalization capability of the proposed DESSA-DESN model are rigorously validated using publicly available aging datasets from NASA and the CALCE center. Comparative analyses against established and contemporary methods, including standard DESN, SSA-DESN, GPR, and others, demonstrate the superior predictive accuracy and robustness of our approach. The results confirm that the proposed framework can effectively track the nonlinear degradation trajectory of a lithium ion battery, yielding RUL predictions with minimal error.
Health Indicator Construction and Selection
The selection of appropriate Health Indicators (HIs) is a critical foundational step for any data-driven RUL prediction model for a lithium ion battery. Effective HIs must strongly correlate with the battery’s capacity fade and be derivable from readily measurable operational data, such as voltage, current, and temperature during charge/discharge cycles. Relying solely on direct capacity measurements is often impractical in real-world applications. Therefore, we construct a diverse set of indirect HIs from the charging phase of the lithium ion battery cycle, which typically follows a standardized Constant Current-Constant Voltage (CC-CV) protocol.
From the voltage and current profiles during charging, we extract ten candidate features that exhibit observable trends as the lithium ion battery ages. These features are categorized into four groups related to different aspects of the charge curve:
- Current Curve Features: Constant Current charging time (\(T_{cc}\)), Constant Voltage charging time (\(T_{cv}\)), and the area under the current curve (\(A_{cc}\)).
- Voltage Curve Features: Area under the voltage curve (\(A_{cv}\)), time for equal voltage rise (\(T_{dcc}\)), and the ratio \(R_{cc} = T_{cc} / (T_{cc}+T_{cv})\).
- Incremental Capacity (IC) Curve Features: The peak value of the IC curve (\(ICP\)) and the voltage at which this peak occurs (\(ICPL\)). The IC curve is derived from \(dQ/dV\), where \(Q\) is charge capacity.
- Differential Voltage (DV) Curve Features: The location (in terms of capacity \(Q\)) of the second peak in the DV curve (\(DVPL\)), and the distance between the first and second DV peaks (\(DVPLD\)). The DV curve is derived from \(dV/dQ\).
We first compute the Pearson correlation coefficient between each of these ten candidate HIs and the actual measured capacity for multiple lithium ion battery cells. HIs with a correlation magnitude generally above 0.9 are considered strongly indicative of capacity fade. However, simply using all high-correlation features can lead to redundancy and increased model complexity. To address this, we apply Neighborhood Component Analysis (NCA), a supervised dimensionality reduction technique. NCA learns a distance metric that maximizes the performance of a nearest-neighbor classifier, effectively projecting the high-dimensional HI data into a lower-dimensional space while preserving the most discriminative information related to the target (capacity). We apply NCA separately within each of the four feature groups, reducing each group to a single, most representative Health Factor (HF). The process yields four final HFs (HI1, HI2, HI3, HI4) that serve as the concise input vector for our prediction model. The high correlation of these selected HFs with capacity is validated across different cells, as summarized in Table 1.
| Cell ID | HI1 | HI2 | HI3 | HI4 |
|---|---|---|---|---|
| B5 (NASA) | 0.9882 | 0.9935 | 0.9891 | 0.9965 |
| B6 (NASA) | 0.9896 | 0.9845 | 0.9904 | 0.9856 |
| B7 (NASA) | 0.9912 | 0.9981 | 0.9895 | 0.9962 |
| CS2-35 (CALCE) | 0.9885 | 0.9871 | 0.9956 | 0.9972 |
| CS2-36 (CALCE) | 0.9851 | 0.9886 | 0.9969 | 0.9989 |
| CS2-37 (CALCE) | 0.9774 | 0.9659 | 0.9894 | 0.9980 |
Prediction Model: Deep Echo State Network (DESN)
To model the complex temporal degradation of the lithium ion battery, we employ a Deep Echo State Network (DESN). The DESN is a recurrent neural network architecture renowned for its efficiency in training and strong performance in modeling nonlinear dynamical systems. Unlike standard RNNs which suffer from the vanishing gradient problem and costly training, the DESN features a fixed, randomly initialized, and sparsely connected reservoir (or multiple reservoirs) that acts as a dynamic memory of input history. Only the output weights are trained via linear regression, making the process computationally efficient and stable.
Our DESN structure consists of an input layer, multiple stacked reservoir layers, and an output layer. Let \(K\), \(M\), \(N\), and \(L\) denote the number of input neurons, reservoir layers, neurons per layer, and output neurons, respectively. At time \(t\), the input vector is \(\mathbf{U}(t)=[u_1(t), u_2(t), \ldots, u_K(t)]^T\), the state of the \(m\)-th reservoir layer is \(\mathbf{x}^m(t) \in \mathbb{R}^N\), and the output is \(\mathbf{Y}(t)=[y_1(t), y_2(t), \ldots, y_L(t)]^T\). The collective state of all reservoirs is \(\mathbf{X}(t) = [\mathbf{x}^1(t); \mathbf{x}^2(t); \ldots; \mathbf{x}^M(t)]\).
The state update for each reservoir layer is governed by:
$$
\mathbf{x}^m(t+1) = \tanh\left( \mathbf{W}^{m}_{in} \mathbf{z}^m(t+1) + \mathbf{W}^{m}_{res} \mathbf{x}^m(t) \right)
$$
where \(\mathbf{z}^1(t+1) = \mathbf{U}(t)\) for the first layer, and \(\mathbf{z}^m(t+1) = \mathbf{x}^{m-1}(t+1)\) for subsequent layers (\(m > 1\)). Here, \(\mathbf{W}^{m}_{in} \in \mathbb{R}^{N \times K}\) (or \( \mathbb{R}^{N \times N}\) for \(m>1\)) is the input weight matrix for layer \(m\), and \(\mathbf{W}^{m}_{res} \in \mathbb{R}^{N \times N}\) is its sparse internal recurrent weight matrix. The \(\tanh(\cdot)\) function provides the nonlinear activation.
The network output is a linear combination of all reservoir states:
$$
\mathbf{Y}(t+1) = \mathbf{W}_{out} \mathbf{X}(t+1)
$$
where \(\mathbf{W}_{out} \in \mathbb{R}^{L \times (M \cdot N)}\) is the output weight matrix. The key training task is to find the optimal \(\mathbf{W}_{out}\) that minimizes the prediction error. Given a sequence of \(l_{tr}\) training inputs and corresponding target outputs \(\mathbf{D}\), \(\mathbf{W}_{out}\) is computed via ridge regression:
$$
\mathbf{W}_{out}^T = (\mathbf{Q}^T\mathbf{Q} + \gamma \mathbf{I})^{-1} \mathbf{Q}^T \mathbf{D}
$$
where \(\mathbf{Q}\) is the state matrix collecting \(\mathbf{X}(t)\) for all \(t\) in the training set, \(\mathbf{I}\) is the identity matrix, and \(\gamma\) is a regularization parameter that prevents overfitting. The depth provided by multiple reservoir layers allows the DESN to learn hierarchical temporal features, making it highly suitable for capturing the progressive degradation patterns of a lithium ion battery.
Hyperparameter Optimization: The Hybrid DESSA Algorithm
The performance of the DESN is highly sensitive to its hyperparameters, such as the spectral radius of the reservoir matrices, input scaling, and regularization coefficient \(\gamma\). Manually tuning these parameters is inefficient and suboptimal. We propose a novel Hybrid Differential Evolution-Sparrow Search Algorithm (DESSA) to automate and optimize this process.
The DESSA algorithm synergistically combines the global exploration strength of Differential Evolution (DE) with the efficient local exploitation capability of the Sparrow Search Algorithm (SSA). The SSA simulates the foraging and anti-predation behaviors of sparrow flocks, consisting of producers, scroungers, and scouts. While SSA converges quickly, it can get trapped in local optima. DE, through its mutation and crossover operations, promotes population diversity and global search. DESSA integrates these mechanisms.
The DESSA procedure is as follows:
- Initialization with Opposition-Based Learning: An initial population of \(N_p\) candidate parameter vectors is generated. To enhance quality, an opposition population is created where each individual \(\mathbf{X}_{i,j}\) has a counterpart \(\mathbf{X}^{*}_{i,j} = ub_{i,j} + lb_{i,j} – \mathbf{X}_{i,j}\). The fittest \(N_p\) individuals from the combined set form the initial population.
- SSA Update Phase: The population is updated according to standard SSA rules for producers and scroungers, improving local fitness.
- DE Enhancement Phase: To prevent premature convergence, DE operations are applied:
- Mutation: For each target individual \(\mathbf{X}_i\), a donor vector \(\mathbf{V}_i\) is created: \(\mathbf{V}_i = \mathbf{X}_{r1} + F \cdot (\mathbf{X}_{r2} – \mathbf{X}_{r3})\), where \(r1, r2, r3\) are distinct random indices and \(F\) is the scaling factor.
- Crossover: A trial vector \(\mathbf{U}_i\) is generated by mixing components from the donor \(\mathbf{V}_i\) and the target \(\mathbf{X}_i\) based on a crossover probability \(CR\).
- Selection: The trial vector \(\mathbf{U}_i\) replaces the target \(\mathbf{X}_i\) in the next generation if it has better (lower) fitness \(f(\cdot)\): \(\mathbf{X}_i^{new} = \mathbf{U}_i \text{ if } f(\mathbf{U}_i) \leq f(\mathbf{X}_i) \text{, else } \mathbf{X}_i\).
- Termination: Steps 2 and 3 repeat until a maximum iteration count is reached. The individual with the best fitness (lowest prediction error on a validation set) defines the optimal DESN hyperparameters.
The fitness function for DESSA is the Root Mean Square Error (RMSE) of the DESN model’s capacity predictions on a validation segment of the training data. The final, optimized model is termed DESSA-DESN.
Experimental Validation and Results
We evaluate the proposed DESSA-DESN framework on two distinct public datasets to assess its accuracy and generalization ability. Dataset A is from NASA Ames, comprising three 18650 LiCoO2 cells (B5, B6, B7) cycled under a CC-CV charge and constant current discharge regime. Dataset B is from CALCE, comprising three LiCoO2 cells (CS2-35, CS2-36, CS2-37) cycled under different but standardized protocols. The capacity degradation curves for these cells exhibit typical nonlinear fade with some variability, providing a robust testbed.
For each lithium ion battery, we use data from the first 80 cycles (NASA) or 400 cycles (CALCE) as the training set to predict the RUL from that point forward. The failure threshold is set at 1.4 Ah for NASA cells and 0.88 Ah (80% of nominal capacity) for CALCE cells. The model’s performance is quantified using standard metrics: Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and the absolute error in predicting the cycle of failure (AE).
$$
\text{RMSE} = \sqrt{\frac{1}{K}\sum_{t=1}^{K}(Y_t – \hat{Y}_t)^2}, \quad \text{MAE} = \frac{1}{K}\sum_{t=1}^{K}|Y_t – \hat{Y}_t|, \quad \text{AE} = |\text{RUL}_{true} – \text{RUL}_{pred}|
$$
where \(Y_t\) and \(\hat{Y}_t\) are the actual and predicted capacities at cycle \(t\), and \(K\) is the total number of cycles in the prediction horizon.
Results on NASA Dataset
The prediction results for the NASA cells are presented visually, showing the DESSA-DESN model’s trajectory closely aligning with the true capacity fade. The quantitative comparison against baseline models is decisive. We compare DESSA-DESN against the standard DESN, a DESN optimized only with DE (DE-DESN), and one optimized only with SSA (SSA-DESN). As shown in Table 2, DESSA-DESN achieves the lowest prediction errors across all cells.
| Cell | Model | RMSE (%) | MAE (%) | AE (Cycles) |
|---|---|---|---|---|
| B5 | DESN | 3.691 | 2.988 | 5 |
| DE-DESN | 2.541 | 1.931 | 3 | |
| SSA-DESN | 1.983 | 1.542 | 1 | |
| DESSA-DESN | 0.914 | 0.783 | 0 | |
| B6 | DESN | 4.471 | 3.664 | 3 |
| DE-DESN | 3.565 | 2.731 | 3 | |
| SSA-DESN | 2.535 | 1.999 | 1 | |
| DESSA-DESN | 1.466 | 0.689 | 1 | |
| B7 | DESN | 3.071 | 2.613 | 5 |
| DE-DESN | 2.159 | 1.683 | 3 | |
| SSA-DESN | 1.954 | 1.332 | 2 | |
| DESSA-DESN | 1.269 | 0.728 | 1 |
For cell B5, DESSA-DESN reduces the RMSE by 53.9% and 64.0% compared to SSA-DESN and DE-DESN, respectively. This significant improvement demonstrates the effectiveness of the hybrid DESSA optimizer in finding a superior parameter configuration for the DESN model, leading to more accurate tracking of the lithium ion battery‘s degradation.
Results on CALCE Dataset and Generalization
To test generalization, we applied the identical DESSA-DESN methodology to the CALCE dataset. The results, summarized in Table 3, confirm the model’s robustness. DESSA-DESN consistently outperforms the other DESN-based variants, with RMSE values all below 1.3% and MAE values below 1.0%. The successful application to a different lithium ion battery chemistry (LCO), form factor, and cycling protocol underscores the adaptability of our feature engineering (NCA) and modeling framework.
| Cell | Model | RMSE (%) | MAE (%) | AE (Cycles) |
|---|---|---|---|---|
| CS2-35 | DESN | 3.402 | 2.692 | 2 |
| DE-DESN | 2.299 | 1.680 | 2 | |
| SSA-DESN | 1.968 | 1.442 | 1 | |
| DESSA-DESN | 1.253 | 0.923 | 0 | |
| CS2-36 | DESN | 3.048 | 2.319 | 2 |
| DE-DESN | 2.819 | 2.112 | 1 | |
| SSA-DESN | 1.998 | 1.492 | 1 | |
| DESSA-DESN | 1.284 | 0.797 | 0 | |
| CS2-37 | DESN | 3.376 | 2.686 | 3 |
| DE-DESN | 2.251 | 1.565 | 2 | |
| SSA-DESN | 1.767 | 1.134 | 1 | |
| DESSA-DESN | 1.075 | 0.793 | 1 |
Comparative Analysis with State-of-the-Art Methods
We further benchmark DESSA-DESN against other prominent optimization algorithms and published methods. Table 4 compares DESSA-DESN with models using Particle Swarm Optimization (PSO) and Genetic Algorithm (GA) to optimize the DESN (PSO-DESN, GA-DESN) on the NASA dataset. Our method shows clear superiority.
| Cell | Model | RMSE (%) | MAE (%) | AE (Cycles) |
|---|---|---|---|---|
| B5 | PSO-DESN | 3.563 | 3.230 | 3 |
| GA-DESN | 2.869 | 2.250 | 1 | |
| DESSA-DESN | 0.914 | 0.783 | 0 | |
| B6 | PSO-DESN | 3.941 | 3.135 | 3 |
| GA-DESN | 2.763 | 2.112 | 2 | |
| DESSA-DESN | 1.466 | 0.689 | 1 | |
| B7 | PSO-DESN | 2.626 | 2.026 | 5 |
| GA-DESN | 2.042 | 1.504 | 3 | |
| DESSA-DESN | 1.269 | 0.728 | 1 |
Finally, we compare DESSA-DESN’s performance on cell B5 against several advanced methods reported in the literature, including Gaussian Process Regression (GPR), Grey Wolf Optimizer-LSTM (GWO-LSTM), Autoregressive-Relevance Vector Machine (AR-RVM), and a Self-adaptive Differential Evolution optimized Echo State Network (SADE-MESN). The results, consolidated in Table 5, demonstrate that our proposed framework achieves the lowest prediction errors, reducing RMSE by 28% to 77% compared to these benchmarks. This conclusively validates the effectiveness of our integrated approach for RUL prediction of a lithium ion battery.
| Model | RMSE (%) | MAE (%) | AE (Cycles) |
|---|---|---|---|
| GPR | 1.84 | 1.46 | 2 |
| GWO-LSTM | 3.95 | 3.35 | – |
| AR-RVM | 1.27 | 0.92 | – |
| SADE-MESN | 1.29 | 1.72 | 1 |
| DESSA-DESN (Proposed) | 0.91 | 0.78 | 0 |
Conclusion
In this work, we have developed and validated a comprehensive, high-precision framework for predicting the Remaining Useful Life of a lithium ion battery. The framework’s strength lies in its synergistic combination of sophisticated techniques for each stage of the prediction pipeline. First, we systematically construct a set of health indicators from standard charge data and employ Neighborhood Component Analysis to distill them into a concise, highly informative set of four health factors. Second, we leverage the powerful temporal modeling capability of a Deep Echo State Network to learn the complex degradation dynamics. Third, and crucially, we introduce a novel hybrid metaheuristic, the DESSA algorithm, which expertly optimizes the DESN’s hyperparameters by balancing the exploratory power of Differential Evolution with the exploitative efficiency of the Sparrow Search Algorithm.
Extensive experimentation on two distinct and publicly available lithium ion battery aging datasets (NASA and CALCE) confirms the superiority of the proposed DESSA-DESN model. It consistently achieves lower prediction errors (RMSE < 1.5%, MAE < 1.0%) compared to models using component algorithms individually (SSA-DESN, DE-DESN) or other optimization techniques (PSO-DESN, GA-DESN). Furthermore, benchmarking against several state-of-the-art methods from the literature demonstrates its competitive edge and higher accuracy.
The proposed method provides a reliable, data-driven tool for lithium ion battery prognostics and health management. Its ability to deliver accurate RUL predictions from early-cycle data can significantly enhance maintenance scheduling, improve system safety, and optimize the economic utilization of battery assets. Future work will focus on extending this methodology to account for more variable operating conditions, such as different temperatures and dynamic load profiles, to further enhance its practicality for real-world lithium ion battery management systems.