The widespread adoption of lithium-ion batteries, the representative of modern energy storage devices, is propelled by their high energy density, long cycle life, and environmental friendliness. However, the inherent trade-off between lengthy charging times and accelerated degradation under high-rate charging remains a critical bottleneck for their further application, particularly in electric vehicles and grid storage. Consequently, the design of intelligent fast charging strategies has emerged as a pivotal research frontier in battery management. An effective strategy must navigate the conflicting demands of rapidity (minimizing charging time), safety (adhering to voltage and temperature limits), and sustainability (maximizing energy efficiency and minimizing aging). This article provides a comprehensive review of the research landscape, systematically examining the core pillars of fast charging strategy design: problem formulation, battery modeling, and optimization methodologies.
1. Problem Formulation for Fast Charging
The design of a fast charging strategy is fundamentally an optimization problem. Its formulation involves defining the design variables, setting the optimization objectives, and imposing necessary constraints, all tailored to specific application scenarios.
1.1 Optimization Objectives and Constraints
The user-centric requirements translate into quantifiable metrics for the optimization problem. Common objectives and constraints derived from these requirements are summarized below.
| Requirement Category | Typical Metrics | Common Role in Problem |
|---|---|---|
| Rapidity | Charging Time (t_chg) | Primary objective to minimize. |
| Safety | Cell Voltage (V_cell) | Constraint: $V_{min} \le V_{cell}(t) \le V_{max}$. |
| Cell Temperature (T_cell) | Constraint: $T_{cell}(t) \le T_{max}$. | |
| Sustainability | Energy Efficiency ($\eta$) | Objective to maximize or constraint. |
| Capacity Fade ($\Delta Q$) | Objective to minimize, often via aging models. | |
| SEI Growth / Li Plating | Constraint on overpotential or Li surface concentration. |
A multi-objective formulation is often necessary. For example, a strategy might aim to minimize charging time while also minimizing the degradation incurred during that charge. These are conflicting goals, as a shorter time typically requires higher currents, which exacerbate degradation. The specific formulation varies significantly across studies, as shown in the following summary of representative research focus areas.
| Study Focus | Primary Objective(s) | Key Constraint(s) |
|---|---|---|
| Time-Centric | Minimize $t_{chg}$ | $V_{cell}$, $T_{cell}$ |
| Awareness-Centric | Minimize $\Delta Q$ (Aging) | $t_{chg}$, $V_{cell}$, $T_{cell}$ |
| Thermal-Centric | Minimize $T_{max}$ or $\Delta T$ | $t_{chg}$, $V_{cell}$ |
| Multi-Objective | Pareto front of $t_{chg}$, $\Delta Q$, $T_{max}$ | Voltage limits |
1.2 Design Variables: Charging Protocols
The design variable is typically the charging current profile, $I_{chg}(t)$. Directly optimizing the current at every time step offers maximum flexibility but leads to a high-dimensional problem. A more practical approach is to parameterize the current profile using a charging protocol. The most common protocols include:
- Constant Current Constant Voltage (CCCV): This two-stage protocol is the industry baseline. A constant current $I_{CC}$ is applied until the voltage reaches $V_{max}$, followed by a constant voltage stage until the current tapers to a cutoff $I_{cut}$.
$$ I_{chg}(t) = \begin{cases} I_{CC}, & V_{cell}(t) < V_{max} \\ \text{(Variable to hold } V_{cell}=V_{max}) , & V_{cell}(t) = V_{max} \text{ and } I > I_{cut} \end{cases} $$
The design variables are $I_{CC}$ and $V_{max}$. - Multi-Stage Constant Current (MCC): This protocol uses $N$ constant current stages, offering a balance between flexibility and complexity. Stages can be defined by duration or State of Charge (SOC) limits.
$$ I_{chg}(t) = I_k \quad \text{for} \quad t \in [t_{k-1}, t_k], \quad k=1,…,N $$
The design variables are the current levels $\{I_1, I_2, …, I_N\}$ and the transition points $\{t_1, t_2, …, t_{N-1}\}$ or $\{SOC_1, SOC_2, …, SOC_{N-1}\}$. - Pulse Charging (PC): This protocol applies current in periodic pulses (e.g., rectangular, sinusoidal). A rectangular pulse is defined by amplitude $I_{pulse}$, frequency $f$, and duty cycle $D$.
$$ I_{chg}(t) = \begin{cases} I_{pulse}, & \text{if mod}(t, 1/f) < D/f \\ 0 \text{ (or a negative pulse)}, & \text{otherwise} \end{cases} $$
The design variables are $I_{pulse}$, $f$, and $D$.
2. Battery Modeling for Strategy Evaluation
Evaluating candidate charging strategies directly on physical lithium-ion batteries is time-consuming and causes irreversible aging. Therefore, high-fidelity yet computationally efficient battery models are indispensable for the design process. A comprehensive model must capture coupled electrical, thermal, and aging dynamics.
2.1 Physics-Based Modeling Approaches
Electrical Models: These predict terminal voltage $V_{cell}$ given current $I$ and internal states like SOC.
- Equivalent Circuit Models (ECM): Use resistors and capacitors to emulate dynamic behavior. A second-order RC model is common:
$$ V_{cell}(t) = OCV(SOC(t)) – I(t)R_0 – V_1(t) – V_2(t) $$
$$ \frac{dV_1}{dt} = -\frac{V_1}{R_1C_1} + \frac{I}{C_1}, \quad \frac{dV_2}{dt} = -\frac{V_2}{R_2C_2} + \frac{I}{C_2} $$
$$ \frac{dSOC}{dt} = -\frac{I(t)}{Q_{nom}} $$
where $OCV$ is the open-circuit voltage, $R_0$ is the ohmic resistance, and $R_1C_1$, $R_2C_2$ pairs represent short- and long-term polarization. ECMs are computationally light but lack physical insight into internal states crucial for aging prediction. - Electrochemical Models (e.g., Pseudo-Two-Dimensional – P2D): These are derived from first principles, describing Li-ion concentration and potential in electrodes and electrolyte via coupled partial differential equations (PDEs). They provide detailed internal states (e.g., surface concentration $c_s^{surf}$, overpotential $\eta_{anode}$) vital for mechanistic aging prediction but are computationally expensive. Reduced-Order Models (ROMs), like the Single Particle Model (SPM), simplify these PDEs for faster computation at the cost of accuracy at high rates.
Thermal Models: Predict temperature rise due to heat generation $Q_{gen}$. A lumped-parameter model is often coupled:
$$ Q_{gen} = I(V_{cell} – OCV) + I T \frac{dOCV}{dT} $$
$$ m C_p \frac{dT}{dt} = Q_{gen} – hA(T – T_{amb}) $$
where the first term is Joule heating and the second is reversible entropic heat.
Aging Models: Quantify capacity/power fade. Models range from empirical (e.g., $Q_{loss} = A \cdot exp(-E_a/RT) \cdot t^{z}$) to mechanistic, modeling side reactions like Solid Electrolyte Interphase (SEI) growth and Lithium Plating (LP). A simplified SEI growth rate might be:
$$ \frac{d\delta_{SEI}}{dt} \propto exp\left(\frac{-\alpha_c F}{RT} \eta_{SEI}\right) $$
where $\delta_{SEI}$ is SEI thickness and $\eta_{SEI}$ is the overpotential driving the reaction.
2.2 Machine Learning Enhanced Modeling
Data-driven approaches are increasingly used to enhance or replace physics-based models, offering high computational speed and the ability to model complex, poorly understood phenomena.
- Black-box Models: Neural Networks (NNs), Long Short-Term Memory (LSTM) networks, and Gaussian Processes (GPs) can directly map inputs ($I$, $T$, past states) to outputs ($V$, $SOC$, $T$, $SOH$). They require large datasets but are extremely fast for inference.
- Physics-Informed Neural Networks (PINNs): This hybrid approach embeds physical laws (PDEs of the P2D model, energy balance) directly into the loss function of a neural network:
$$ \mathcal{L} = \mathcal{L}_{data} + \lambda \mathcal{L}_{physics} $$
where $\mathcal{L}_{data}$ minimizes prediction error on measured data, and $\mathcal{L}_{physics}$ ensures the NN’s output satisfies the governing equations. This improves data efficiency and generalization for the lithium-ion battery system. - Early Prediction Models: Machine learning models trained on early-cycle data can predict the entire battery lifetime (RUL), drastically reducing the testing needed to evaluate a charging strategy’s long-term impact.
3. Optimization and Control Methodologies
The choice of optimization algorithm is critical and is often dictated by the problem formulation and model complexity. Methods can be broadly classified into offline (design-then-deploy) and online (real-time control) strategies.
3.1 Offline Design Methods
These methods find an optimal charging protocol (parameter set) beforehand. The optimized protocol is then deployed statically.
| Method | Key Idea | Advantages | Disadvantages | Typical Model Used |
|---|---|---|---|---|
| Dynamic Programming (DP) | Solves the multi-stage problem via backward recursion, guaranteeing global optimality for the discretized problem. | Conceptually clear, well-suited for MCC protocols. | Suffers from the “curse of dimensionality”; state and control grids must be coarse. | ROM, SPM, ECM |
| Evolutionary Algorithms (EA) (e.g., GA, PSO, NSGA-II) |
Population-based stochastic search inspired by natural evolution (selection, crossover, mutation). | Handles non-convex, multi-objective problems; derivative-free. | Requires a vast number of performance evaluations (model simulations); convergence can be slow. | ECM, Lumped Thermal |
| Bayesian Optimization (BO) | Builds a probabilistic surrogate model (e.g., Gaussian Process) of the objective and uses an acquisition function to guide sample-efficient search. | Extremely sample-efficient for expensive black-box functions (like P2D simulation). | Scalability to high dimensions (>20) is challenging. | P2D, High-fidelity models |
For a multi-objective problem minimizing time $t_{chg}$ and degradation $D$, BO would model the objective space and iteratively choose protocol parameters $\mathbf{x}$ to evaluate by optimizing an acquisition function $\alpha(\mathbf{x})$, such as Expected Improvement (EI):
$$ \mathbf{x}_{next} = \arg\max_{\mathbf{x}} \alpha(\mathbf{x}; \{\mathbf{x}_i, f(\mathbf{x}_i)\}_{i=1}^t) $$
where $f(\mathbf{x})$ could be a scalarized objective like $f = w_1 t_{chg} + w_2 D$.
3.2 Online Control Methods
These methods compute the optimal charging current in real-time based on the battery’s present state, enabling adaptation to varying conditions and model inaccuracies.
| Method | Key Idea | Advantages | Disadvantages | Typical Model Used |
|---|---|---|---|---|
| Model Predictive Control (MPC) | At each time step, solves a finite-horizon optimal control problem using a dynamic model, applies the first control action, and repeats. | Handles constraints explicitly; feedback compensates for model mismatch. | Online computational burden; requires a fast, sufficiently accurate model. | ECM, Linearized ROM, NDC |
| Deep Reinforcement Learning (DRL) | An AI agent (neural network policy) learns the optimal control policy through interaction with a simulated or real environment to maximize a cumulative reward. | Can learn complex policies from raw data; very fast online execution (just network inference). | Extensive training data/compute needed; safety and stability guarantees are hard. | Any (as part of simulation environment) |
In MPC, at time $k$, the following optimization is solved online:
$$ \begin{aligned} \min_{I_{k},…,I_{k+N-1}} & \quad \sum_{j=0}^{N-1} \left( \| t_{chg,j} \| + \| D_j \| \right) \\
\text{s.t.} & \quad \text{Battery Model Dynamics (ECM/ROM)} \\
& \quad V_{min} \le V_{cell,j} \le V_{max} \\
& \quad I_{min} \le I_j \le I_{max} \\
& \quad \text{Other constraints}
\end{aligned} $$
Only $I_k$ is applied before re-solving at step $k+1$.
In DRL, the agent learns a policy $\pi_\theta(a_t|s_t)$ parameterized by $\theta$ that maps state $s_t$ (e.g., $SOC$, $V$, $T$) to action $a_t$ (e.g., $I_{chg}$). Learning is performed by maximizing the expected cumulative reward $R_t = \sum_{i=t}^\infty \gamma^{i-t} r(s_i, a_i)$, where the reward $r$ encodes objectives (e.g., $r = – \Delta t + \beta_1 \cdot \text{violation} – \beta_2 \cdot \Delta Q$).
4. Current Challenges and Future Directions
Despite significant progress, several challenges persist in designing practical fast-charging strategies for lithium-ion batteries.
- Comprehensive Problem Formulation: Real-world charging involves dynamically shifting priorities (e.g., time vs. cost vs. battery health depending on trip urgency, electricity price, and battery state). Formulating strategies that can dynamically adapt to these multi-faceted, time-varying user preferences remains complex.
- Model Accuracy vs. Efficiency Trade-off: High-fidelity models (P2D) are too slow for online use, while fast models (ECM) lack the internal state details for accurate aging prediction. Developing physics-informed machine learning models that are both fast and predictive of internal critical states (like anode potential for Li plating) is a key direction.
- Adaptation to Battery Aging: An optimal strategy for a new battery is suboptimal for a degraded one. Future strategies need to be adaptive, incorporating real-time parameter and state estimation to adjust the charging policy as the lithium-ion battery ages.
- From Cell to Pack Level: Most research focuses on single cells. Extending optimization to battery packs must account for cell-to-cell variations, balancing, and pack-level thermal management, significantly increasing the problem’s complexity.
- Sample-Efficient and Safe Learning: For data-driven methods like DRL, reducing the required interaction data (through transfer learning, simulation-to-real transfer) and ensuring safety during exploration are critical for practical deployment.
- Integration with Grid and Thermal Management: Optimal charging of a lithium-ion battery should consider grid load, electricity prices, and the capabilities of the onboard thermal management system. Co-design of the charging current profile and cooling system actuation is a promising holistic approach.
In conclusion, the design of fast charging strategies for lithium-ion batteries is a rich, interdisciplinary field at the intersection of electrochemistry, control theory, optimization, and machine learning. While offline methods using advanced optimization with high-fidelity models have successfully identified theoretically superior protocols, the future lies in adaptive, online, and pack-level strategies. These will leverage hybrid modeling approaches and intelligent control to deliver safe, rapid, and battery-health-conscious charging, ultimately enhancing the usability and longevity of lithium-ion battery systems across all applications.