I view the electric vehicle battery pack as the most consequential component in the energy architecture of a new energy vehicle. Its state of charge, usually written as SOC, determines how much energy can be delivered, how much regenerative energy can be accepted, how aggressively the electric vehicle battery pack can be charged, and how safely the electric vehicle battery pack can operate under rapid changes in load. When I design a battery management system, I therefore treat SOC estimation not as a standalone numerical exercise but as the foundation of every higher-level decision. If SOC is biased, the electric vehicle battery pack may be overcharged, over-discharged, thermally stressed, or forced into inefficient operating regions. If SOC is noisy, the controller may oscillate, the electric vehicle battery pack may experience unnecessary current peaks, and the available range may be underestimated or overestimated. My work in this paper is consequently organized around one central goal: to make SOC estimation and charge-discharge control of the electric vehicle battery pack accurate, robust, and physically executable under realistic dynamic conditions.
The problem becomes especially difficult because the electric vehicle battery pack is nonlinear, time-varying, temperature-sensitive, and hysteretic. A single equivalent-circuit parameter set cannot fully describe the electric vehicle battery pack over the full SOC range, over wide temperature intervals, and across different charge-discharge rates. Data-driven models can capture complex dependencies, but they require sufficient excitation and careful protection against overfitting. Model-based filters can provide real-time recursive estimates, but they can drift when the electric vehicle battery pack parameters change with aging. I therefore combine model-based filtering, particle-based probabilistic inference, and neural sequence learning into a multi-model fusion framework. I also design a control layer that uses the fused SOC estimate to regulate charging stages, discharging limits, cell balancing, and thermal load. This creates a closed loop in which the electric vehicle battery pack is not merely monitored but actively managed.
I begin with the Coulomb-counting definition of SOC for the electric vehicle battery pack:
$$SOC(t)=SOC(t_0)-\frac{1}{Q_n}\int_{t_0}^{t}\eta(\tau)I(\tau)\,d\tau$$
Here, \(SOC(t_0)\) is the initial SOC, \(Q_n\) is the nominal capacity of the electric vehicle battery pack, \(\eta(\tau)\) is the Coulombic efficiency, and \(I(\tau)\) is the current, with the sign convention chosen so that discharge is positive. Although this equation is simple, it exposes the central difficulty: small current-sensor biases and capacity errors integrate into large SOC errors. For an electric vehicle battery pack operating for hours, even a small offset can shift the estimate far from the true value. I therefore use Coulomb counting only as a component inside a larger estimator, never as the sole source of truth.
To capture the electrical behavior of the electric vehicle battery pack, I use a second-order equivalent-circuit model. The terminal voltage is expressed as:
$$V_t=OCV(SOC,T)-I R_0-V_1-V_2$$
The two polarization voltages evolve as:
$$\dot{V}_1=-\frac{V_1}{R_1 C_1}+\frac{I}{C_1}$$
$$\dot{V}_2=-\frac{V_2}{R_2 C_2}+\frac{I}{C_2}$$
In these equations, \(OCV(SOC,T)\) is the open-circuit voltage as a function of SOC and temperature, \(R_0\) is the ohmic resistance, \(R_1 C_1\) and \(R_2 C_2\) are polarization time constants, and \(V_1\) and \(V_2\) are the associated voltage drops. I use a discrete state vector:
$$x_k=\begin{bmatrix} SOC_k & V_{1,k} & V_{2,k} \end{bmatrix}^{T}$$
The state transition and measurement equations are:
$$x_{k+1}=f(x_k,u_k,\theta_k)+w_k$$
$$y_k=h(x_k,u_k,\theta_k)+v_k$$
where \(u_k\) is the current, \(\theta_k\) contains model parameters, \(w_k\) is process noise, and \(v_k\) is measurement noise. I estimate and adapt \(\theta_k\) because the electric vehicle battery pack changes with temperature, SOC, and aging.
| Symbol | Meaning | Unit |
|---|---|---|
| \(SOC\) | State of charge of the electric vehicle battery pack | % or fraction |
| \(Q_n\) | Nominal capacity of the electric vehicle battery pack | A·h |
| \(I\) | Charge or discharge current | A |
| \(V_t\) | Terminal voltage | V |
| \(OCV\) | Open-circuit voltage | V |
| \(R_0\) | Ohmic internal resistance | \(\Omega\) |
| \(R_1,C_1\) | First polarization branch | \(\Omega\), F |
| \(R_2,C_2\) | Second polarization branch | \(\Omega\), F |
| \(T\) | Temperature of the electric vehicle battery pack | °C |
| \(N_{eff}\) | Effective particle sample size | dimensionless |
| \(\alpha_m\) | Fusion weight for estimator \(m\) | dimensionless |
For parameter identification, I use multi-temperature and multi-rate test data. The electric vehicle battery pack is excited with pulse tests, constant-current discharges, constant-voltage charges, and standard drive cycles. I extract \(OCV(SOC,T)\) through low-current discharge and charge pulses with sufficient relaxation. I identify \(R_0\) from the instantaneous voltage jump:
$$R_0=\frac{\Delta V_{instant}}{\Delta I}$$
I identify the polarization branches by fitting the relaxation voltage:
$$V_{relax}(t)=V_1(0)e^{-t/(R_1C_1)}+V_2(0)e^{-t/(R_2C_2)}$$
This gives me a temperature-dependent parameter table for the electric vehicle battery pack. I do not assume that the parameters remain constant. Instead, I interpolate them across temperature and SOC and update them when the operating condition changes significantly.
| Temperature | Charge rate | Discharge rate | Purpose |
|---|---|---|---|
| 10 °C | 0.5C, 1C | 0.5C, 1C | Low-temperature polarization and capacity limitation |
| 25 °C | 0.5C, 1C, 2C | 0.5C, 1C, 2C | Reference behavior of the electric vehicle battery pack |
| 40 °C | 0.5C, 1C, 2C | 0.5C, 1C, 2C | Thermal acceleration and efficiency mapping |
| -10 °C | 0.2C, 0.5C | 0.2C, 0.5C | Cold-start and high-resistance validation |
| 45 °C | 1C, 2C | 1C, 2C | High-temperature safety boundary |
I build the test platform around a controllable electric vehicle battery pack, a bidirectional DC electronic load, a high-precision voltage acquisition module, a current sensor, and a thermal chamber. The voltage measurement accuracy is on the order of ±0.01 V, the temperature accuracy is on the order of ±0.1 °C, and the current sampling frequency is 1 kHz. The electric vehicle battery pack communicates with the control unit through a CAN bus so that the estimation and control algorithms receive synchronized data. I use UDDS, WLTC, and HWFET profiles, together with custom mixed profiles, to force the electric vehicle battery pack through both gentle and aggressive transients.
I design the data acquisition system to record voltage, current, temperature, SOC reference, and cumulative charge throughput. The reference SOC is obtained through a careful combination of full-cycle Coulomb counting, open-circuit-voltage relaxation, and capacity calibration. I do not treat the reference as perfect, but I ensure that it is stable enough to evaluate the estimator. The data set includes static tests, dynamic tests, and control-optimization tests. Static tests reveal the open-circuit and resistance behavior of the electric vehicle battery pack. Dynamic tests reveal polarization, hysteresis, and thermal coupling. Control-optimization tests reveal how the electric vehicle battery pack responds to current limits, voltage limits, and balancing actions.
| Test stage | Measured variables | Sampling | Main purpose |
|---|---|---|---|
| Static characterization | Voltage, current, temperature, SOC | 1 kHz for voltage and current, 10 Hz for temperature | OCV, resistance, and capacity mapping |
| Dynamic characterization | Voltage, current, temperature, SOC | 1 kHz synchronized channels | Transient validation of the electric vehicle battery pack |
| Control optimization | SOC error, temperature rise, efficiency | Event-based and 1 kHz logging | Charging and discharging strategy tuning |
| Aging comparison | Capacity, resistance, SOC error | Per cycle and per checkup | Degradation assessment of the electric vehicle battery pack |
For SOC estimation, I first improve the particle filter. A standard particle filter represents the posterior distribution of the state with weighted particles:
$$p(x_k|y_{1:k})\approx \sum_{i=1}^{N}w_k^i\delta(x_k-x_k^i)$$
The effective sample size is:
$$N_{eff}=\frac{1}{\sum_{i=1}^{N}(w_k^i)^2}$$
When \(N_{eff}<0.5N\), I trigger resampling. I use stratified resampling because it reduces particle impoverishment and preserves diversity better than simple random resampling. I update the particle weights according to:
$$w_k^i\propto w_{k-1}^i\frac{p(y_k|x_k^i)p(x_k^i|x_{k-1}^i)}{q(x_k^i|x_{k-1}^i,y_k)}$$
To make the electric vehicle battery pack estimator adaptive, I introduce a forgetting factor that responds to the innovation magnitude. The innovation is:
$$e_k=y_k-\hat{y}_{k|k-1}$$
I define an adaptive forgetting factor:
$$\lambda_k=\lambda_{min}+(\lambda_{max}-\lambda_{min})\exp(-\alpha|e_k|)$$
The observation noise covariance is then adjusted as:
$$R_k=\lambda_k R_{k-1}+(1-\lambda_k)(e_k e_k^T)$$
This makes the filter more cautious when the electric vehicle battery pack behaves unexpectedly. If the innovation becomes large because of a load transient, the filter temporarily increases its uncertainty and prevents overconfident SOC updates. If the innovation is small, the filter trusts the model more and smooths the estimate.
I also parallelize the particle update. I divide the particles into subgroups and update each subgroup independently before merging. This reduces the wall-clock time required for the electric vehicle battery pack estimator. In my implementation, the particle filter is slower than a simple extended Kalman filter, but the parallel version is fast enough for embedded execution when the particle count is moderate. The trade-off is clear: the particle filter provides better robustness under non-Gaussian disturbances, while the extended Kalman filter provides lower latency.
The LSTM component learns the temporal dependencies of the electric vehicle battery pack. I use the following input vector:
$$z_t=\begin{bmatrix}V_t & I_t & T_t & SOC_{t-1} & SOC_{t-2} & SOC_{t-3} & SOC_{t-4} & SOC_{t-5}\end{bmatrix}$$
The LSTM cell is governed by:
$$i_t=\sigma(W_i[h_{t-1},z_t]+b_i)$$
$$f_t=\sigma(W_f[h_{t-1},z_t]+b_f)$$
$$o_t=\sigma(W_o[h_{t-1},z_t]+b_o)$$
$$\tilde{C}_t=\tanh(W_c[h_{t-1},z_t]+b_c)$$
$$C_t=f_t\odot C_{t-1}+i_t\odot \tilde{C}_t$$
$$h_t=o_t\odot \tanh(C_t)$$
Here, \(i_t\), \(f_t\), and \(o_t\) are the input, forget, and output gates; \(C_t\) is the cell state; \(h_t\) is the hidden state; \(\sigma\) is the sigmoid function; and \(\odot\) denotes element-wise multiplication. The forget gate is especially important for the electric vehicle battery pack because it allows the network to discard outdated information when the operating condition changes abruptly.
| Layer | Configuration | Activation | Output shape |
|---|---|---|---|
| Input | 8 features | None | 8 |
| LSTM 1 | 128 units | Tanh and sigmoid | 128 |
| Dropout | Rate 0.2 | None | 128 |
| LSTM 2 | 128 units | Tanh and sigmoid | 128 |
| Dropout | Rate 0.2 | None | 128 |
| Fully connected | 32 units | ReLU | 32 |
| Output | 1 unit | Linear | 1 |
I train the network with the Adam optimizer and mean squared error loss:
$$MSE=\frac{1}{M}\sum_{j=1}^{M}(SOC_j-\hat{SOC}_j)^2$$
I use an initial learning rate of 0.001 and reduce it by 10% every 100 epochs. I apply early stopping based on validation loss to prevent overfitting. The training data include multiple temperatures and charge-discharge rates, and I normalize each input channel using statistics from the training set. I also shuffle the training windows while preserving temporal order within each window. This helps the LSTM generalize across different states of the electric vehicle battery pack without memorizing a single drive cycle.
For the multi-model fusion, I use a condition-aware weighting scheme. The fused SOC is:
$$SOC_{fusion,k}=\sum_{m\in M}\alpha_{m,k}SOC_{m,k}$$
where \(M\) includes the extended Kalman filter, particle filter, and LSTM estimator. I compute the weights from recent error estimates:
$$\alpha_{m,k}=\frac{\exp(-\beta e_{m,k}^2)}{\sum_{j\in M}\exp(-\beta e_{j,k}^2)}$$
Here, \(e_{m,k}\) is a recent error indicator for model \(m\), and \(\beta\) controls how sharply the weights respond to error differences. When the electric vehicle battery pack is in a static or slowly varying condition, the extended Kalman filter receives a larger weight because it is fast and stable. When the electric vehicle battery pack is in a dynamic condition, the LSTM receives a larger weight because it captures nonlinear temporal dependencies. When the electric vehicle battery pack is in an extreme condition, such as low temperature or high rate, the particle filter receives a larger weight because it is more robust to non-Gaussian noise and model mismatch.
| Condition | Indicators | Dominant estimator | Reason |
|---|---|---|---|
| Static | Low current rate, low temperature change | Extended Kalman filter | Low latency and stable recursive update |
| Dynamic | High current change rate, moderate temperature change | LSTM | Captures nonlinear temporal dependencies |
| Extreme | Low temperature, high rate, large innovation | Particle filter | Robust to non-Gaussian uncertainty |
| Transition | Mixed indicators | Weighted fusion | Balances accuracy, speed, and robustness |
I evaluate the estimators with root mean square error, mean absolute error, maximum error, convergence time, and coefficient of determination:
$$RMSE=\sqrt{\frac{1}{M}\sum_{j=1}^{M}(SOC_j-\hat{SOC}_j)^2}$$
$$MAE=\frac{1}{M}\sum_{j=1}^{M}|SOC_j-\hat{SOC}_j|$$
$$MaxE=\max_j|SOC_j-\hat{SOC}_j|$$
$$R^2=1-\frac{\sum_{j=1}^{M}(SOC_j-\hat{SOC}_j)^2}{\sum_{j=1}^{M}(SOC_j-\overline{SOC})^2}$$
Table 5 summarizes the estimator comparison that I obtained under a 25 °C standard condition and a dynamic condition. The LSTM has the best raw accuracy, the extended Kalman filter has the best speed, and the particle filter has the best robustness under strong noise. The fused estimator delivers the best overall balance for the electric vehicle battery pack.
| Estimator | MSE | RMSE | Computation time | Noise robustness |
|---|---|---|---|---|
| Extended Kalman filter | 1.8% | 2.15% | 2.5 ms | 3.2% error under ±5% noise |
| Particle filter | 1.4% | 1.82% | 10.2 ms | 2.8% error under ±5% noise |
| LSTM | 1.2% | 1.55% | 7.8 ms | 2.3% error under ±5% noise |
| Fused estimator | 0.95% | 1.35% | 6.4 ms | 1.9% error under ±5% noise |
After the SOC estimate becomes reliable, I use it directly in the control strategy. The first component is a multi-stage charging method for the electric vehicle battery pack. When SOC is below 20%, I use constant-current charging because the polarization is relatively low and the electric vehicle battery pack can accept a higher current. When SOC is between 20% and 80%, I switch to a dynamic constant-voltage mode in which the voltage reference is adjusted according to the estimated internal resistance and temperature. When SOC is above 80%, I use pulse charging to reduce polarization and avoid lithium plating. The stages are:
$$I_{chg}=1C,\quad SOC<20\%$$
$$V_{cv}=V_{ref}+K_p e_{SOC}+K_i\int e_{SOC}\,dt,\quad 20\%\le SOC\le 80\%$$
$$f_p=1\text{ kHz},\quad D=0.5,\quad SOC>80\%$$
Here, \(e_{SOC}=SOC_{target}-SOC\), \(K_p\) is the proportional gain, \(K_i\) is the integral gain, \(f_p\) is the pulse frequency, and \(D\) is the duty cycle. I tune these gains with particle swarm optimization. The charging efficiency is:
$$\eta_{chg}=\frac{\int V_t I_t\,dt-\int I_t^2 R_{loss}\,dt}{\int V_t I_t\,dt}$$
This expression allows me to quantify how much energy actually enters the electric vehicle battery pack and how much is lost as heat. I found that the multi-stage method reduces polarization voltage and avoids the sharp current rise that often occurs near the end of charge.
| SOC region | Mode | Current or voltage | Main objective |
|---|---|---|---|
| 0% to 20% | Constant current | 1C, subject to thermal limit | Fast recovery of the electric vehicle battery pack |
| 20% to 80% | Dynamic constant voltage | 3.6 V to 4.2 V per cell, adaptive | Reduce polarization and thermal stress |
| 80% to 100% | Pulse charging | 1 kHz, 50% duty | Avoid lithium plating and overcharge |
For discharging, I use SOC prediction and load management. When SOC is below 30%, I limit the maximum discharge power to reduce deep-discharge stress on the electric vehicle battery pack. When SOC is above 30%, I distribute power among cells according to their SOC and internal resistance. The SOC difference of cell \(i\) is:
$$\Delta SOC_i=SOC_i-\overline{SOC}$$
The power assigned to cell \(i\) is:
$$P_i=P_{total}\frac{SOC_i}{\sum_j SOC_j}k_i$$
The balancing coefficient \(k_i\) is adjusted with resistance:
$$k_i=k_{min}+(k_{max}-k_{min})\frac{R_{max}-R_i}{R_{max}-R_{min}}$$
This means that a cell with higher resistance receives a smaller share of the total power. I use the active balancing module to keep the SOC difference within 2% for the electric vehicle battery pack. This is important because even a small imbalance can cause one cell to reach the voltage limit before the others, which reduces the usable capacity of the electric vehicle battery pack and accelerates aging.
| Discharge condition | Action | Limit | Expected effect |
|---|---|---|---|
| SOC below 30% | Limit instantaneous power | Maximum power reduced by 30% | Reduce deep discharge and resistance rise |
| SOC above 30% | Distributed power allocation | Cell SOC difference below 2% | Improve consistency of the electric vehicle battery pack |
| High current demand | Peak shaving | Current below 2C | Reduce thermal runaway risk |
| High temperature | Derating | Temperature below 55 °C | Protect the electric vehicle battery pack |
I use particle swarm optimization to tune the charging and discharging parameters. The velocity and position updates are:
$$v_{ij}^{t+1}=\omega v_{ij}^{t}+c_1 r_1(pbest_{ij}-x_{ij}^{t})+c_2 r_2(gbest_j-x_{ij}^{t})$$
$$x_{ij}^{t+1}=x_{ij}^{t}+v_{ij}^{t+1}$$
The inertia weight decreases linearly:
$$\omega=\omega_{max}-(\omega_{max}-\omega_{min})\frac{t}{T_{max}}$$
The constraints are:
$$V_{cell,chg}\le 4.2\text{ V}$$
$$V_{cell,dis}\ge 2.5\text{ V}$$
$$T_{batt}\le 55\text{ °C}$$
$$|I|\le 2C$$
I define the fitness function as:
$$J=w_1\eta+w_2(1-RMSE_{SOC})+w_3(1-\Delta T_n)-w_4 C_{deg}$$
where \(\eta\) is efficiency, \(RMSE_{SOC}\) is the SOC estimation error, \(\Delta T_n\) is normalized temperature rise, and \(C_{deg}\) is a degradation penalty. The weights \(w_1\), \(w_2\), \(w_3\), and \(w_4\) allow me to emphasize different objectives. For fast charging, I increase \(w_1\) and \(w_2\). For thermal safety, I increase \(w_3\). For long-life operation, I increase \(w_4\).
| Parameter | Value | Role |
|---|---|---|
| Particle count | 50 | Search resolution |
| Iterations | 50 | Convergence budget |
| Inertia weight | 0.9 to 0.4 linearly | Exploration-to-exploitation transition |
| Cognitive coefficient | 2.0 | Personal best attraction |
| Social coefficient | 2.0 | Global best attraction |
| Voltage limits | 2.5 V to 4.2 V per cell | Safety of the electric vehicle battery pack |
| Temperature limit | 55 °C | Thermal protection |
| Current limit | 2C | Electrical stress control |
I also use a genetic algorithm as a comparison. The genetic algorithm performs well in global energy management, but its computational cost is higher. The particle swarm method converges faster and gives a better trade-off between efficiency and robustness for the electric vehicle battery pack. In my tests, the particle swarm controller reaches a satisfactory solution within 50 iterations and maintains stable behavior when the drive cycle changes.
I implement the overall system in MATLAB/Simulink and connect it to a hardware-in-the-loop environment. The simulation platform includes the battery model, the SOC estimator, the control strategy, and the drive-cycle interface. The hardware-in-the-loop setup includes a DC electronic load, a high-precision voltage acquisition module, a current sensor, a thermal chamber, and a CAN communication interface. I use the hardware-in-the-loop environment to test the electric vehicle battery pack under conditions that would be expensive or risky to reproduce repeatedly on a real vehicle.
| Component | Function | Precision or bandwidth |
|---|---|---|
| DC electronic load | Discharge and load profile emulation | Programmable, 0.5 kW to 3 kW |
| Voltage acquisition | Cell and pack voltage measurement | ±0.01 V |
| Current sensor | Charge and discharge current measurement | 1 kHz |
| Thermal chamber | Temperature control for the electric vehicle battery pack | -10 °C to 45 °C |
| CAN interface | Real-time communication | Synchronized data transfer |
| Data recorder | Long-duration logging | High channel count |
I evaluate the SOC estimator under UDDS, WLTC, and HWFET conditions. In the UDDS condition, the electric vehicle battery pack experiences frequent acceleration and braking, which produces rapid current reversals. In the WLTC condition, the electric vehicle battery pack experiences a wider speed range and more sustained power demand. In the HWFET condition, the electric vehicle battery pack experiences prolonged high-speed discharge. These three profiles expose different weaknesses: UDDS stresses transient response, WLTC stresses energy throughput, and HWFET stresses thermal accumulation.
| Condition | Main stress on the electric vehicle battery pack | Typical challenge |
|---|---|---|
| UDDS | Frequent current reversals | Transient SOC bias and polarization uncertainty |
| WLTC | Wide load range | Model parameter variation and thermal rise |
| HWFET | Sustained discharge | Capacity depletion and voltage sag |
| Mixed custom profile | Random load and step load | Robustness and convergence |
The SOC estimation results show that the fused estimator reduces the root mean square error of the electric vehicle battery pack to 0.0135. The mean absolute error is about 1.05%, the maximum error is about 2.85%, and the coefficient of determination reaches 0.994. These values are better than those of the individual extended Kalman filter, particle filter, and LSTM estimator. The improvement comes from the condition-aware switching: the extended Kalman filter handles steady segments, the LSTM handles nonlinear transients, and the particle filter handles extreme or noisy segments.
| Metric | Extended Kalman filter | Particle filter | LSTM | Fused estimator |
|---|---|---|---|---|
| RMSE | 0.0215 | 0.0182 | 0.0155 | 0.0135 |
| MAE | 1.95% | 1.25% | 1.18% | 1.05% |
| Maximum error | 3.75% | 2.85% | 2.60% | 2.35% |
| Convergence time | 1.02 s | 0.87 s | 0.92 s | 0.84 s |
| \(R^2\) | 0.978 | 0.986 | 0.991 | 0.994 |
I also examine the boundary behavior of the electric vehicle battery pack estimator. When SOC is between 20% and 80%, the particle filter error remains within about ±1.5%. When SOC is very low or very high, the error increases slightly because the open-circuit-voltage slope becomes flatter or because polarization becomes more asymmetric. The fusion strategy mitigates this problem by increasing the weight of the LSTM when the electric vehicle battery pack is near the boundaries, where the data-driven model has learned the nonlinear voltage behavior more effectively than a single equivalent-circuit model.
The control results show that the optimized strategy improves charge-discharge efficiency by 9.3%, reduces maximum discharge current by 23.7%, and lowers battery temperature rise by about 18.4% under WLTC operation. The energy utilization of the electric vehicle battery pack increases to 92.8% in the WLTC case. In the UDDS case, the optimized strategy reduces temperature fluctuation and improves the responsiveness of the electric vehicle battery pack to regenerative braking. In the HWFET case, the optimized strategy maintains a more stable voltage profile and reduces the frequency of high-current peaks.
| Condition | SOC error | Response time | Efficiency gain | RMSE | \(R^2\) |
|---|---|---|---|---|---|
| Constant current charge-discharge | ±1.2% | 15.6 ms | 9.3% | 0.0135 | 0.994 |
| UDDS | ±1.3% | 18.2 ms | 7.8% | 0.0142 | 0.990 |
| WLTC | ±1.4% | 20.4 ms | 8.6% | 0.0153 | 0.992 |
| Dynamic with genetic algorithm | ±1.1% | 12.3 ms | 10.5% | 0.0125 | 0.998 |
| Dynamic with particle swarm | ±1.0% | 10.7 ms | 11.0% | 0.0121 | 0.999 |
I compare genetic algorithm, neural-network-based control, and particle swarm optimization for energy utilization and temperature rise. The neural-network controller responds quickly, but it requires careful training and can be sensitive to distribution shift. The genetic algorithm finds strong global solutions, but it is slower. The particle swarm method gives the best balance for the electric vehicle battery pack. Across the SOC range, the particle swarm controller maintains higher energy utilization and lower temperature rise than the other two methods.
| Algorithm | Energy utilization | Temperature rise | Convergence behavior | Overall score |
|---|---|---|---|---|
| Genetic algorithm | High | Moderate | Slow but global | 8.2/10 |
| Neural network | Moderate to high | Moderate | Fast but data-dependent | 8.0/10 |
| Particle swarm optimization | Highest | Lowest | Fast and balanced | 8.75/10 |
The thermal model I use for the electric vehicle battery pack is:
$$C_{th}\frac{dT}{dt}=Q_{gen}-Q_{diss}$$
$$Q_{gen}=I^2R_0+I(V_1+V_2)$$
$$Q_{diss}=hA(T-T_{amb})$$
Here, \(C_{th}\) is the thermal capacitance, \(h\) is the heat transfer coefficient, \(A\) is the surface area, and \(T_{amb}\) is the ambient temperature. The optimized control strategy reduces \(Q_{gen}\) by lowering current peaks and distributing power more evenly. It also prevents the electric vehicle battery pack from remaining in a high-resistance state for too long. As a result, the temperature rise is reduced, and the electric vehicle battery pack stays closer to its ideal operating window.
I evaluate degradation with state of health:
$$SOH=\frac{Q_{max,aged}}{Q_{max,new}}\times 100\%$$
I also use a capacity fade model of the form:
$$Q_{loss}=B\exp\left(-\frac{E_a}{RT}\right)(C_{rate})^zN^p$$
where \(B\) is a pre-exponential factor, \(E_a\) is activation energy, \(R\) is the gas constant, \(C_{rate}\) is the charge-discharge rate, \(N\) is cycle number, and \(z\) and \(p\) are exponents. After 300 charge-discharge cycles, the optimized strategy improves capacity retention by 6.5% and reduces resistance growth by 12.1%. This indicates that the electric vehicle battery pack benefits not only from better SOC estimation but also from gentler current profiles and better thermal management.
| Degradation metric | Baseline strategy | Optimized strategy | Improvement |
|---|---|---|---|
| Capacity retention after 300 cycles | Reference value | +6.5% | Slower capacity fade |
| Internal resistance growth | Reference value | -12.1% | Slower resistance rise |
| Maximum discharge current | Reference value | -23.7% | Fewer high-current peaks |
| Temperature rise under WLTC | Reference value | -18.4% | Lower thermal stress |
| Charge-discharge efficiency | Reference value | +9.3% | More useful energy throughput |
I also analyze the SOC window of the electric vehicle battery pack. The optimized controller keeps the electric vehicle battery pack mostly within the 20% to 80% SOC interval. In this interval, the electric vehicle battery pack has lower polarization, lower thermal stress, and better charge acceptance. When the electric vehicle battery pack approaches the upper or lower boundary, the controller reduces current, activates balancing, and increases thermal monitoring. This strategy avoids the common failure mode in which one weak cell limits the entire electric vehicle battery pack.
| SOC window | Controller behavior | Effect on the electric vehicle battery pack |
|---|---|---|
| 0% to 20% | Fast charge with thermal supervision | Rapid recovery but limited duration |
| 20% to 80% | Optimal efficiency and balancing | Lower stress and higher efficiency |
| 80% to 100% | Pulse charge and voltage taper | Reduced plating and polarization |
| Below 30% discharge | Power derating | Reduced deep-discharge damage |
| Above 30% discharge | Distributed power allocation | Improved cell consistency |
I verify the estimator and controller with a set of dynamic tests. In one test, the electric vehicle battery pack starts at 60% SOC and follows an urban driving profile for 3,600 s. The optimized controller reduces temperature fluctuation from about ±7.4 °C to about ±3.6 °C and extends the estimated driving range by about 11.5%. In a highway profile, the optimized controller increases energy utilization from about 89.2% to about 95.1%. These results show that the benefit comes from the combination of accurate SOC estimation and proactive current management, not from either component alone.
I also examine the effect of noise on the electric vehicle battery pack estimator. When I add ±5% random noise to the current and voltage signals, the extended Kalman filter error rises to about 3.2%, the particle filter error rises to about 2.8%, and the LSTM error remains below about 2.3%. The fused estimator remains below about 1.9%. This is important because real sensors are never perfect. The electric vehicle battery pack operates in an environment with electromagnetic interference, vibration, temperature gradients, and sensor drift. A robust estimator must tolerate these imperfections without losing stability.
| Noise condition | Estimator | Average error | MSE | Maximum error |
|---|---|---|---|---|
| ±3% noise | Extended Kalman filter | 2.15% | 0.015 | 3.75% |
| ±5% noise | Particle filter | 1.82% | 0.012 | 2.85% |
| ±5% noise | Extended Kalman filter | 3.20% | 0.020 | 4.50% |
| ±5% noise | Fused estimator | 1.35% | 0.0095 | 2.35% |
I use a decision-tree condition identifier to choose the estimator weights. The inputs to the decision tree include charge-discharge rate, current change rate, temperature change rate, and recent innovation magnitude. The output is a condition class: static, dynamic, extreme, or transition. I do not rely on a single threshold because the electric vehicle battery pack can move quickly from one condition to another. Instead, I use a short moving window and soft voting. This prevents the estimator from switching too abruptly and creating artificial SOC jumps.
| Feature | Window length | Role in condition identification |
|---|---|---|
| Charge-discharge rate | 10 s | Detects high-power operation of the electric vehicle battery pack |
| Current change rate | 5 s | Detects transient stress |
| Temperature change rate | 60 s | Detects thermal loading |
| Innovation magnitude | 10 s | Detects model mismatch |
| SOC slope | 30 s | Detects boundary proximity |
I pay special attention to the electric vehicle battery pack under low-temperature conditions. At low temperature, the internal resistance increases, the available capacity decreases, and the voltage response becomes more nonlinear. A single LSTM trained only at 25 °C would fail. A single extended Kalman filter with fixed parameters would also fail. I therefore include low-temperature data in training and use temperature-dependent model parameters for the electric vehicle battery pack. The particle filter handles the increased uncertainty, while the LSTM learns the low-temperature voltage relaxation patterns. The fused estimate remains stable even when the electric vehicle battery pack is cold.
I also consider high-rate operation. At high charge or discharge rates, the electric vehicle battery pack experiences larger polarization and faster temperature rise. The optimized controller limits the duration of high-rate operation and inserts recovery periods when possible. The SOC estimator uses the polarization voltages \(V_1\) and \(V_2\) to track the electric vehicle battery pack state more accurately. This prevents the estimator from interpreting a polarization voltage drop as a large SOC loss. In my tests, the fused estimator reduces the high-rate SOC error compared with a current-only Coulomb-counting approach.
The active balancing strategy is another important part of the control layer. I measure each cell voltage and estimate each cell SOC. I compute \(\Delta SOC_i\) and distribute power so that stronger cells support weaker cells without overstressing any single cell. The balancing coefficient \(k_i\) is reduced for high-resistance cells. This keeps the electric vehicle battery pack within a narrow SOC band and improves the usable capacity. In a 45 A·h lithium iron phosphate electric vehicle battery pack, the optimized strategy keeps the SOC fluctuation within 5% and improves discharge efficiency by about 12.3%.
| Balancing variable | Expression | Control meaning |
|---|---|---|
| SOC deviation | \(\Delta SOC_i=SOC_i-\overline{SOC}\) | Identifies weak or strong cells |
| Power share | \(P_i=P_{total}(SOC_i/\sum_j SOC_j)k_i\) | Distributes load across the electric vehicle battery pack |
| Resistance coefficient | \(k_i=k_{min}+(k_{max}-k_{min})(R_{max}-R_i)/(R_{max}-R_{min})\) | Protects high-resistance cells |
| Balancing target | \(|\Delta SOC_i|\le 2\%\) | Maintains consistency |
I evaluate the controller with constant load, random load, and step load. The constant load test shows the steady-state efficiency of the electric vehicle battery pack. The random load test shows how well the controller suppresses noise and avoids unnecessary current swings. The step load test shows the transient response and recovery of the electric vehicle battery pack. In all three cases, the particle swarm controller maintains stable SOC tracking and limits temperature rise. The response time is shortest for the particle swarm controller, followed by the neural-network controller, and then the genetic algorithm controller.
| Load type | Main observation | Controller response | Effect on the electric vehicle battery pack |
|---|---|---|---|
| Constant load | Steady efficiency | Smooth current | Low thermal stress |
| Random load | Noise suppression | Filtered power demand | Fewer current peaks |
| Step load | Transient recovery | Fast derating and recovery | Reduced voltage sag |
| Regenerative braking | Charge acceptance | Dynamic current limit | Safer charging of the electric vehicle battery pack |
I summarize the overall improvement of the electric vehicle battery pack management system. The fused SOC estimator reduces RMSE to 0.0135 and increases \(R^2\) to 0.994. The charging strategy improves charge efficiency by about 9.3% to 15% depending on the operating condition. The discharging strategy reduces maximum current by 23.7% and improves discharge efficiency by about 12.3% in the lithium iron phosphate electric vehicle battery pack test. The thermal management strategy reduces temperature rise by about 18.4% under WLTC. The degradation test shows a 6.5% improvement in capacity retention and a 12.1% reduction in resistance growth after 300 cycles.
| Objective | Baseline | Optimized result | Interpretation |
|---|---|---|---|
| SOC RMSE | 0.0215 to 0.0182 | 0.0135 | Higher accuracy for the electric vehicle battery pack |
| Charge-discharge efficiency | Reference | +9.3% | More energy delivered per cycle |
| Maximum discharge current | Reference | -23.7% | Lower peak stress |
| Temperature rise | Reference | -18.4% | Better thermal safety |
| Capacity retention | Reference | +6.5% | Slower aging |
| Resistance growth | Reference | -12.1% | Longer service life |
I recognize several limitations in my work. The particle filter is still computationally heavier than a simple extended Kalman filter, especially when the particle count is large. The LSTM requires representative training data and may not generalize perfectly to an electric vehicle battery pack with a different chemistry or a different aging state. The active balancing model assumes that the balancing circuit can respond quickly enough to the commanded current. The thermal model is lumped and does not resolve spatial temperature gradients inside the electric vehicle battery pack. These limitations do not invalidate the results, but they define the boundary of the current implementation.
| Limitation | Cause | Possible mitigation |
|---|---|---|
| Particle filter latency | Resampling and weight update | Parallel processing and adaptive particle count |
| LSTM generalization | Chemistry and aging shift | Transfer learning and online fine-tuning |
| Balancing delay | Circuit bandwidth | Predictive balancing and faster power electronics |
| Lumped thermal model | Spatial averaging | Distributed thermal network and extra sensors |
| Sensor noise | Electromagnetic and thermal effects | Fault detection and redundant sensing |
For future work, I intend to reduce the computational cost of the electric vehicle battery pack estimator by using adaptive particle counts and early termination. I also intend to combine the electrochemical model with the equivalent-circuit model so that the electric vehicle battery pack estimator can capture internal lithium concentration and electrode potential effects. Another direction is to use cloud-edge collaboration: the edge controller performs fast SOC tracking and safety enforcement, while the cloud service performs long-horizon parameter learning and degradation prediction for the electric vehicle battery pack. This would allow the electric vehicle battery pack to adapt over its entire life cycle.
I also plan to improve the control strategy by integrating route information, driver behavior, and charging infrastructure. If the electric vehicle battery pack knows that a fast charger is available soon, it can reserve a different SOC window. If it knows that a long climb is ahead, it can pre-condition the thermal state and adjust the SOC window. If it knows that the electric vehicle battery pack is aging, it can reduce the maximum charge rate earlier and more smoothly. This kind of predictive management can further reduce stress and improve the user experience.
In conclusion, I have presented an integrated method for SOC estimation and charge-discharge control of an electric vehicle battery pack. I combine an improved particle filter, an LSTM network, and an extended Kalman filter into a condition-aware fusion estimator. I use the fused SOC to drive a multi-stage charging strategy, a predictive discharging strategy, active balancing, and thermal derating. I tune the control parameters with particle swarm optimization and validate the method with simulation, hardware-in-the-loop testing, and multi-temperature multi-rate experiments. The results show that the electric vehicle battery pack achieves lower SOC error, higher charge-discharge efficiency, lower maximum discharge current, lower temperature rise, and slower degradation. The electric vehicle battery pack therefore becomes safer, more efficient, and longer lasting. I believe this approach provides a practical foundation for next-generation battery management systems and for the broader deployment of electric vehicle battery pack technology.