Driven by the global low-carbon transition and the increasing penetration of electric vehicles, the charging management of electric vehicles has become a critical research topic that involves not only convenience for individuals but also grid stability and battery lifetime. In this work, I focus on the household slow-charging scenario and develop a multi-objective intelligent charging strategy for electric vehicles. The framework couples a battery aging model, a data-driven travel energy prediction model, and a deep reinforcement learning power dispatcher so that the charging cutoff state of charge, charging cost, and battery health can be optimized in a coordinated manner.
My research addresses three questions that are not fully resolved in existing electric vehicle charging studies. First, battery models are often used offline while charging policies are trained from highly simplified battery descriptions, which makes the resulting strategy unable to adapt to the actual state of charge, temperature, and aging state. Second, most charging schedulers fix the charging target, for example at 80% or 90% state of charge, even though each trip has a substantially different energy requirement and each driver has a different charging habit. Third, the objectives of charging economy, charging reliability, and battery lifetime are usually studied separately. A policy that aggressively minimizes electricity cost may let the electric vehicle battery charge in unfavorable state-of-charge ranges or with excessive power, while an over-conservative policy may not satisfy the user’s driving demand. My goal is therefore to provide a decision pipeline in which the charging target is dynamically determined and the charging power is continuously adjusted.
The complete strategy consists of three main components. The first component is an electro-thermal-aging coupled model of the battery pack, which provides an accurate virtual environment for policy evaluation and health assessment. The second component is a personalized charging cutoff SOC prediction method based on travel energy consumption prediction. The third component is a deep deterministic policy gradient based scheduling agent that determines the instantaneous charging power according to the time-of-use tariff, charging urgency, battery temperature, and remaining time. I integrate these three components and verify them through battery experiments, ablation tests, and 2000 long-term charging cycles.
1. Coupled Battery Modeling for Electric Vehicles
Battery modeling is the foundation of the proposed electric vehicle charging strategy. In actual operation, the electrical, thermal, and aging phenomena in a lithium-ion cell are coupled. For example, high charging power raises the temperature, high temperature accelerates internal resistance growth, and increased internal resistance further changes voltage response and heat dissipation. I establish a coupled single-cell model and then extend it to a pack configuration of 109 series and 2 parallel cells. I use a second-order RC equivalent circuit to describe electrical dynamics, a lumped-parameter thermal balance equation to describe temperature change, and a semi-empirical aging model to describe calendar and cycle capacity loss.
1.1 Electrical Model
The electrical model treats the open-circuit voltage as a function of state of charge, which is obtained from the measured SOC-OCV curve and represented by cubic spline interpolation. The ohmic resistance \(R_0\) depends on SOC, temperature, and aging. The polarization overpotential is expressed by two RC branches whose voltages \(U_1\) and \(U_2\) evolve according to the following equations:
\begin{aligned}
\frac{dU_1}{dt} &= -\frac{U_1}{R_1 C_1} + \frac{I}{C_1},\\
\frac{dU_2}{dt} &= -\frac{U_2}{R_2 C_2} + \frac{I}{C_2}.
\end{aligned}
$$
The terminal voltage is then given by:
V_t = OCV(SOC) – I R_0(SOC,T,\text{aging}) – U_1 – U_2.
$$
Since the purpose of the model is to support charging optimization rather than high-frequency estimation, the SOC is updated by ampere-hour integration with the instantaneous current \(I\) and the state-of-charge is corrected by the accumulated capacity loss from the aging model.
1.2 Thermal Model
I use a lumped thermal model based on the Bernardi heat generation equation. The total heat generation rate consists of irreversible Joule heat and reversible entropic heat:
Q_{\text{gen}} = I^2 R_0 + I T \frac{dU}{dT}.
$$
Heat dissipation includes natural convection and radiation. The energy balance equation is:
mC_p \frac{dT}{dt} = Q_{\text{gen}} – hA_s(T-T_{\text{env}}) – \varepsilon \sigma A_s\left(T^4-T_{\text{env}}^4\right).
$$
For numerical implementation, I use the factorized form of the radiation law to improve stability near room temperature. The thermal capacity \(mC_p\) and the convection coefficient \(h\) are obtained from charging temperature-rise experiments.
1.3 Aging Model
Battery aging strongly affects the long-term value of an electric vehicle. My aging model separates calendar aging and cycle aging. Calendar aging mainly follows the growth of the solid-electrolyte interface layer and is aggravated by high storage SOC. In order to handle time-varying temperature and SOC, I define an equivalent stress accumulation approach:
t_{\text{eff}} = \int_0^t \sigma_{\text{cal}}\left(T(\tau),SOC(\tau)\right)\, d\tau,
$$
where \(\sigma_{\text{cal}}\) is the calendar stress factor relative to a reference condition. The calendar capacity loss \(L_{\text{cal}}\) is expressed by a three-stage piecewise function with two knee points, representing decelerating growth, near-linear growth, and final accelerated growth.
Cycle aging is driven by repeated charge and discharge processes. The cycle stress factor includes temperature stress, C-rate stress, DOD stress, and SOC-location stress. The equivalent throughput is defined as:
Q_{\text{eff}} = \int_0^t \sigma_{\text{cyc}}\left(T(SOC,C,DOD)\right)\, dQ_{\text{tp}},
$$
where \(Q_{\text{tp}}\) is the actual throughput. The cycle capacity loss \(L_{\text{cyc}}\) is again described by a three-stage piecewise function whose initial exponent is \(n_{\text{cyc}}\). The total capacity state-of-health is written as:
SOH(t) = 1 – L_{\text{cal}}(t_{\text{eff}}) – L_{\text{cyc}}(Q_{\text{eff}}).
$$
In addition to capacity loss, the internal resistance also increases with aging. I model the increase as the superposition of a calendar component and a cycle component, both following a square-root law. These aging variables are fed back into the electrical model, so the model can represent the interaction between aging and voltage response.
Some aging parameters are obtained from published NCM cell datasets, while other parameters are calibrated from my cell experiments in Section 5. I deliberately focus on the relative ranking of different charging strategies rather than absolute lifetime prediction. Since all strategies are evaluated with the same aging model, systematic parameter uncertainty has a limited effect on the ranking conclusions. A sensitivity analysis with \(±20\%\) parameter perturbation confirms that the order of capacity degradation under different thermal/rate conditions remains unchanged.
1.4 Parameter Identification and Model Validation
I use a commercial NCM pouch cell with a nominal capacity of 93 Ah and a nominal voltage of 3.7 V as the experimental object. The cell is tested in an Arbin battery test system and a temperature-controlled chamber. I first perform capacity calibration, then SOC-OCV characterization with 0.1 SOC intervals, and then HPPC tests. The HPPC responses provide the ohmic resistance and the two RC parameters at different SOC values. The temperature-rise tests at 0.2C are used to fit the thermal parameters, while the 0.3C temperature curve is used for validation. Table 1 summarizes the basic cell parameters used in the model.
| Parameter | Value |
|---|---|
| Cell format | NCM pouch cell |
| Nominal voltage | 3.7 V |
| Working voltage | 2.85–4.2 V |
| Nominal capacity | 93 Ah |
| Nominal energy | 344 Wh |
| Testing ambient temperature | 25 °C |
| Calibrated capacity used for SOC-OCV test | 94.70 Ah |
The electrical model is validated by comparing the simulated discharge voltage with the measured voltage under 0.3C constant-current discharge. The maximum voltage error is less than 15 mV. The thermal model is validated by comparing the simulated and measured temperature curve during charging; the temperature error is within \(0.2\,^\circ C\). The aging model is validated under three simulated operating conditions with different temperatures and C-rates. The results show that the model correctly reproduces the degradation trend, including the accelerating degradation at high temperature and high charge C-rate.
1.5 109S2P Battery Pack Model
I extend the single-cell model to a battery pack with 109 series-connected modules and two parallel cells per module, i.e. 109S2P. The pack nominal voltage is 403.3 V, the nominal energy is 75 kWh, and the total mass is approximately 370.6 kg. The parallel cells in each module are treated as an equivalent module whose capacity is doubled and resistance is halved. For thermal modeling, the pack is divided into three thermal regions, so a spatial temperature gradient inside the battery pack can be represented. The overall pack temperature is the weighted average of the regional temperatures. The pack-level SOH follows the weakest-cell principle, which mimics the fact that the discharge capability of an electric vehicle battery pack is limited by its worst module.
For the charging strategy, commands are expressed as pack-level power rather than cell current. Therefore I define the conversion relationship between pack power and single-cell current:
P_{\text{pack}} = 218\, V_{\text{cell}}\, I_{\text{cell}},
$$
or equivalently:
I_{\text{cell}} = \frac{P_{\text{pack}}}{218\,V_{\text{cell}}}.
$$
Since \(V_{\text{cell}}\) gradually increases during charging, the equivalent cell current under constant pack power slowly decreases. For example, at 21 kW pack power the equivalent cell current decreases from approximately 28.6 A at 10% SOC to about 23.6 A at 90% SOC. This conversion is used in the battery experiment to reproduce the pack-level DDPG charging power on a single real cell.
2. Personalized Charging Cutoff SOC Prediction
The charging cutoff SOC is not necessarily 80% or 100% for every electric vehicle user. The required energy differs with trip distance, traffic conditions, weather, driving style, and auxiliary load. I therefore propose a data-driven framework that predicts the energy consumption of the upcoming trip chain, combines it with a safety reserve, and then adjusts the cutoff SOC using personalized charging habits and battery-health considerations.
2.1 Feature Engineering for Energy Consumption
I train and validate the energy prediction model using real-world electric vehicle data. The raw data are cleaned by removing invalid records and reconstructing trip-level fragments. For each independent trip, I construct multidimensional features from four categories: environment, driving behavior, vehicle state, and trip character. Table 2 lists the key features and their operational ranges.
| Feature dimension | Feature definition | Typical range |
|---|---|---|
| Environment | Calendar time, weather category, ambient temperature effect \(|T-25|\) | \(-20 \sim 45\,^\circ C\) |
| Driving behavior | Energy-style index, average speed, speed-mileage ratio; | \(0 \sim 120\) km/h |
| Vehicle state | Current battery SOC, initial trip SOC | \(0\% \sim 100\%\) |
| Trip character | Estimated trip distance, historical charging intervals | \(0 \sim 200\) km |
A number of derived features are introduced. The environmental temperature effect \(T_{\text{effect}}\) is computed as \(|T_{\text{amb}}-25|\) because both cold and hot conditions increase thermal load and thus trip energy. The driving style index combines the number of hard acceleration events, hard deceleration events, and average speed. These derived variables reduce the nonlinearity that a regression model must learn.
2.2 General Model and Personalized Calibration
I compare six machine-learning algorithms: ridge regression, random forest, gradient boosting, extremely randomized trees, LightGBM, and XGBoost. The comparison is based on 5-fold cross-validation and repeated on the test dataset. Table 3 reports the RMSE, MAE, training R², and test R² for each algorithm.
| Algorithm | RMSE (kWh) | MAE (kWh) | Train R² | Test R² |
|---|---|---|---|---|
| Ridge regression | 0.696 | 0.557 | 0.902 | 0.901 |
| Random forest | 0.766 | 0.611 | 0.974 | 0.878 |
| Gradient boosting | 0.767 | 0.614 | 0.984 | 0.878 |
| Extremely randomized trees | 0.762 | 0.609 | 0.990 | 0.880 |
| LightGBM | 0.745 | 0.601 | 0.975 | 0.885 |
| XGBoost | 0.750 | 0.603 | 0.992 | 0.899 |
Ridge regression achieves a test R² of 0.90, comparable with XGBoost but with less overfitting and better interpretability. More importantly, the linear parameter space of ridge regression is convenient for small-sample transfer learning. I therefore choose ridge regression as the general model.
Although the general model captures population-level trends, an individual electric vehicle driver may show systematic differences in energy consumption. I personalize the model using transfer learning. The initial weights are the general model weights \(w_0\), and the personalized objective function is:
\min_{w} \; \lVert y_u – X_u w \rVert_2^2 + \lambda_p \lVert w – w_0 \rVert_2^2,
$$
where \(X_u\) and \(y_u\) are the historical trips of the target user, and \(\lambda_p=0.5\) constrains the personal model not to deviate too far from the general model. In the test on one individual user with 121 trips, the general model achieves an MAE of 0.682 kWh and RMSE of 0.893 kWh, while the personalized model achieves an MAE of 0.426 kWh and RMSE of 0.541 kWh. The MAPE decreases by about 37.5%. This improvement demonstrates the benefit of personalized calibration for electric vehicle charging-demand prediction.
2.3 Three-Module Charging Cutoff SOC Framework
I design a cascaded framework for charging cutoff SOC prediction. Module 1 calculates a baseline target SOC from the expected energy consumption and safety reserve:
SOC_{\text{target}} = SOC_{\text{current}} + \frac{E_{\text{total}} + E_{\text{reserve}}}{C_{\text{battery}}}\times 100\%,
$$
with \(E_{\text{reserve}}=0.15\,E_{\text{total}}\). The safety reserve covers prediction uncertainty and unexpected route changes.
Module 2 incorporates the user’s historical charging behavior. Let \(SOC_{\text{avg}}\) be the average cutoff SOC selected by the user over past events. The habit-corrected target is:
SOC_{\text{habit}} = SOC_{\text{target}} + \lambda_h \left(SOC_{\text{avg}}-SOC_{\text{target}}\right),
$$
where \(\lambda_h=0.5\). This half-correction prevents the model from copying historical habits blindly when the predicted travel energy is reliable. The resulting target is limited to a maximum safe value of 95%.
Module 3 performs an aging-score optimization. I define an aging score consisting of voltage stress, DOD stress, and low-SOC stress:
J_{\text{aging}} = a\left(\frac{SOC_{\text{high}}}{100}\right)^2 + b\left(\frac{x}{100}\right) + c\exp\left(-\frac{SOC_{\text{low}}}{\tau}\right),
$$
where \(x=SOC_{\text{high}}-SOC_{\text{low}}\) is the depth of discharge, \(a=3\), \(b=0.5\), \(c=2\), and \(\tau=20\). Assuming a fixed value of \(x\) required by the trip, the optimization decides how to place the SOC window within the battery operating range. The first-order optimality condition is:
\frac{2a\left(SOC_{\text{low}}+x\right)}{10000} = \frac{c}{\tau} \exp\left(-\frac{SOC_{\text{low}}}{\tau}\right).
$$
Because this is a transcendental equation, I use the bisection method to solve for the optimal \(SOC_{\text{low}}\). The advantage of this approach is that it changes only the position of the SOC window without changing the charge quantity, so travel security is preserved while aging stress is reduced.
2.4 Validation on Typical User Profiles
I validate the personalized cutoff SOC prediction on two user profiles. The first profile is a low-frequency deep-charging user who charges fewer than three times per week and often fills the battery to a high SOC. The second profile is a high-frequency shallow-charging user who charges three or more times per week and keeps the SOC in a moderate range. For each profile, I evaluate 200 charging cycles. Table 4 shows the summary results.
| User profile | Range coverage | MAE reduction | Aging score improvement |
|---|---|---|---|
| Low-frequency deep charging | 99.5% | 16.28% | 21.1% |
| High-frequency shallow charging | 100% | 3.52% | 4.4% |
The low-frequency user has larger per-cycle energy variation, and the habit-correction module is able to compensate for systematic underprediction, increasing the range coverage from 97% to 99.5%. The high-frequency user has smaller charging depth, and the aging optimization still achieves positive improvement over all validation cycles. For a fixed discharge depth of 30%, the aging score has a minimum at an optimal starting SOC of approximately 22.9% in one representative test, which demonstrates that the optimal SOC window is a compromise between high-SOC voltage stress and low-SOC lithium-plating risk.
3. Residential Charging Power Scheduling with DDPG
After the cutoff SOC has been determined, the next task is to determine when and at what power the electric vehicle should be charged. In a household scenario, the user plugs the electric vehicle into an AC charging pile with a certain departure time. The price of electricity varies by time of use. I model the charging scheduling as a Markov decision process and use the deep deterministic policy gradient algorithm to optimize the instantaneous charging power.
3.1 Scenario Modeling
The household charging scenario is modeled with the following elements: the charging pile, the on-board charger, the battery pack, the time-of-use tariff, and the user trip requirement. I assume that the charging pile can continuously adjust the allowable charging current, and the charging power is bounded between zero and \(P_{\max}=21\) kW. The time-of-use electricity price is shown in Table 5.
| Period type | Time interval | Price (CNY/kWh) |
|---|---|---|
| Valley | 00:00–08:00, 22:00–24:00 | 0.34 |
| Peak | 08:00–11:00, 17:00–20:00 | 0.62 |
| Flat | 11:00–17:00, 20:00–22:00 | 0.52 |
User behavior is characterized by the arrival time, departure time, required target SOC, and the requested charging time window. The measured charging records show that most electric vehicle drivers arrive home between 18:00 and 21:00 and leave between 7:00 and 9:00. Since some periods have sparse samples, I augment the data by translating complete original charging sessions into underrepresented time windows. The augmentation preserves the charging duration and SOC change while shifting the session start time.
3.2 MDP Formulation
At each decision step \(t\), the environmental state is composed of the charging urgency \(U_{\text{urgency}}\), the current SOC, the battery pack temperature, the recent electricity prices, and the previous power command:
s_t = \left[U_{\text{urgency},t},\, SOC_t,\, T_{\text{pack},t},\, p_{t-N}, \ldots, p_t,\, P_{t-1}\right].
$$
The action \(a_t\) is the charging power in the continuous interval \([0, P_{\max}]\). The charging urgency is defined as:
U_{\text{urgency},t} = \frac{SOC_{\text{tar}} – SOC_t}{t_{\text{dep}}-t},
$$
which is the rate at which SOC must be increased to meet the departure target. The reward is a weighted combination of several terms. I define instantaneous charging incentive, charging cost, charging progress, battery health, power smoothness, and terminal achievement. Table 6 gives the weighting factors used in the reward function.
| Reward component | Weighting parameter | Value |
|---|---|---|
| Active charging incentive | \(\lambda_c,\lambda_d\) | 0.8, 30 |
| Charging cost | \(\lambda_{\text{price}}\) | 42 |
| Charging progress | \(\lambda_e,\lambda_f,\beta\) | 12, 16, 80 |
| High temperature penalty | \(\lambda_T\) | 5.5 |
| SOC location penalty | \(\alpha_{\text{SOC}}\) | 8.5 |
| Direct SOH penalty | \(\alpha_{\text{SOH}}\) | 23 |
| Success reward | \(R_{\text{success}}\) | 900 |
| Failure penalty coefficient | \(G\) | 420 |
Although all reward terms are summed, the terminal condition is deliberately designed as the highest priority. Failing to reach the target SOC by the departure time causes a large negative penalty. The control constraints are implemented in different layers: power saturation is enforced by the actor output layer; temperature and SOC constraints are reflected as penalties in the reward; and the charging time window is enforced by resetting the environment after departure.
3.3 DDPG Training Implementation
The deep deterministic policy gradient architecture consists of an actor network \(\mu(s|\theta^\mu)\) and a critic network \(Q(s,a|\theta^Q)\). Each network has two hidden layers with 256 neurons. The actor outputs a continuous power command, while the critic estimates the expected cumulative reward. The critic loss is the temporal-difference error:
L_Q = \mathbb{E}\left[\left(Q(s_t,a_t \mid \theta^Q) – y_t\right)^2\right],
$$
where
y_t = r_t + \gamma Q’\left(s_{t+1}, \mu'(s_{t+1}\mid \theta^{\mu’}) \mid \theta^{Q’}\right).
$$
The actor is updated by the deterministic policy gradient:
\nabla_{\theta^\mu} J = \mathbb{E}\left[\nabla_a Q(s,a \mid \theta^Q)\big|_{a=\mu(s)} \nabla_{\theta^\mu} \mu(s\mid\theta^\mu)\right].
$$
Ornstein-Uhlenbeck noise is used for exploration during training. Target networks are updated by soft interpolation. Table 7 summarizes the main training hyperparameters. The total training length is 1.5 million steps. The cumulative reward increases from an initially negative regime to a stable plateau, indicating that the agent has learned a sensible charging policy.
| Hyperparameter | Value |
|---|---|
| Actor learning rate | 0.0001 |
| Critic learning rate | 0.001 |
| Discount factor | 0.99 |
| Soft update coefficient | 0.001 |
| Batch size | 64 |
| Hidden layer size | 256 |
| OU noise \(\sigma\) | 0.2 |
| OU noise \(\theta\) | 0.15 |
3.4 Typical-Charging-Scenario Behavior
I evaluate the learned DDPG policy on four representative scenarios. The first scenario is overnight charging from 20:00 to 06:20. The policy avoids charging during the evening flat period, starts charging at the valley price, and keeps the power at about 3–5 kW until the SOC approaches the target. The power curve is very smooth, which is beneficial to battery lifetime. The second scenario is evening-peak charging from 18:40 to 22:00. The agent does not charge during the peak price; as the flat-price period approaches, the power quickly increases to about 14 kW, and then decreases when the target SOC is nearly reached. The third scenario is morning charging from 06:30 to 11:10. The agent charges at 21 kW during the final valley hours and then gradually lowers the power after the price rises to the peak level. This illustrates that the DDPG policy successfully shifts most energy into the valley-price period. The fourth scenario is a short peak-price charging window from 08:00 to 10:20, where there is no low-price period to exploit. The policy initially charges at about 17 kW to guarantee completion, then lowers the power as the SOC approaches the target, thereby reducing battery stress as much as possible. These four cases show that my policy adapts to time availability, electricity price, and remaining SOC in a coherent way.
4. Experimental Validation and System Performance
A complete multi-objective strategy should be validated through both physical experiments and long-term simulations. I first convert the pack-level charging power into equivalent single-cell current using the pack-to-cell relationship. Then I run a parallel controlled cell-aging experiment for 100 cycles between a CCCV benchmark and a DDPG-optimized current profile. The experimental capacity fade results are then used to calibrate the aging model. Finally, the calibrated model is used in 2000 repeated charging cycles to compare the proposed comprehensive strategy with baseline strategies and to assess the contribution of each module.
4.1 Parallel Cell Aging Experiment
Two cells of the same batch are used. Cell 1 is charged with a standard 28 A CCCV protocol. Cell 2 is charged with a DDPG-optimized current sequence that is produced by converting the training scenario into single-cell commands. Both cells are discharged at 0.3C to the same lower SOC boundary. The ambient temperature is controlled at 25°C. After each cycle, the actual discharge capacity is recorded. Table 8 shows the cumulative capacity after 100 cycles. The DDPG group loses about 2.46 Ah, while the CCCV group loses about 3.16 Ah. Although the difference is not large, it represents a consistent trend that the optimized current waveform is less damaging than the constant-current constant-voltage protocol.
| Charging strategy | Initial measured capacity | Capacity after 100 cycles | Accumulated capacity loss |
|---|---|---|---|
| CCCV benchmark | 94.70 Ah | 91.54 Ah | 3.16 Ah |
| DDPG optimized strategy | 94.56 Ah | 92.10 Ah | 2.46 Ah |
I use the 100-cycle experimental data to calibrate three aging parameters in the semi-empirical model: the cycle aging exponent \(n_{\text{cyc}}\), the initial cycle aging coefficient \(k_{\text{cyc}}\), and the C-rate stress exponent \(\beta_c\). The calibrated values are \(n_{\text{cyc}}=0.513\), \(k_{\text{cyc}}=1.85\times10^{-4}\), and \(\beta_c=0.63\). With this calibration, the simulation RMSE is only 0.17% SOH for the CCCV protocol and 0.19% SOH for the DDPG protocol over 100 experimental cycles. The calibrated model retains the same relative ranking between the two aging curves while improving the absolute agreement with measured data.
4.2 Ablation Study
I design four strategies to separate the contribution of the SOC prediction module and the DDPG scheduling module. Strategy A uses both modules. Strategy B replaces the predicted cutoff SOC by a fixed target of 80% but keeps the DDPG scheduler. Strategy C keeps the predicted cutoff SOC but charges with fixed 21 kW power. Strategy D uses both fixed 80% target and fixed 21 kW power. Table 9 reports the final SOH, charging satisfaction rate, average charging cost, and valley-period energy proportion after 2000 charging cycles.
| Strategy | Final SOH | Satisfaction | Average cost | Valley share |
|---|---|---|---|---|
| A: SOC prediction + DDPG | 93.01% | 99.8% | 14.1 CNY | 68.1% |
| B: fixed 80% + DDPG | 92.77% | 89.9% | 14.1 CNY | 68.7% |
| C: SOC prediction + 21 kW | 90.62% | 99.8% | 17.8 CNY | 53.1% |
| D: fixed 80% + 21 kW | 90.63% | 89.9% | 17.8 CNY | 52.5% |
Three conclusions can be drawn from this ablation study. First, the DDPG scheduling module is extremely important for battery aging and cost. Compared with Strategy D, Strategy B reduces the relative SOH loss by about 22.8% and cuts charging cost by about 20.8%. Second, the SOC prediction module is critical for user demand satisfaction. Compared with Strategy B, Strategy A increases the satisfaction rate from 89.9% to 99.8% while also improving battery aging by about 3.3% relative loss reduction. Third, the two modules are complementary. Removing either one significantly degrades at least one objective.
4.3 Multi-Strategy Comparison
I compare the proposed comprehensive strategy with three practical baseline strategies for electric vehicle charging: a constant-power strategy, a SOC-section strategy that reduces power at high and low SOC, and a price-table strategy that only follows the time-of-use tariff. Table 10 summarizes the final SOH, satisfaction rate, average charging cost, and time-of-use energy distribution.
| Strategy | Final SOH | Satisfaction | Average cost | Valley share | Peak share |
|---|---|---|---|---|---|
| Constant power, target 90% | 90.01% | 96.4% | 18.2 CNY | 52.6% | 18.2% |
| SOC-section, target 80% | 91.34% | 90.0% | 15.9 CNY | 58.9% | 14.6% |
| TOU schedule, target 90% | 90.72% | 96.2% | 16.2 CNY | 65.5% | 0% |
| Proposed comprehensive strategy | 93.01% | 99.8% | 14.1 CNY | 68.1% | 8.3% |
Compared with the constant-power benchmark, the proposed strategy reduces the SOH loss by 30.03%, saves 22.5% of charging cost, and improves the charging satisfaction rate from 96.4% to 99.8%. The proposed strategy also has the highest valley-period charging proportion among all four strategies, while still retaining a small peak-period charging fraction when high urgency makes it necessary.
Among the baseline strategies, the SOC-section strategy improves battery health by avoiding high power at extreme SOC, but its fixed target SOC of 80% leads to a lower satisfaction rate. The price following strategy maximizes the ability to avoid peak periods, but its battery-health benefit is smaller because it does not coordinate the SOC target with the state of the battery. These observations confirm that a single-objective optimization method is insufficient for multi-objective electric vehicle charging.
5. Discussion and Outlook
The experimental and simulation results show that the proposed integrated strategy provides a favorable balance among charging reliability, economic cost, and battery lifetime. The combination of travel-energy-based cutoff SOC estimation and DDPG-based charging power control is central to the overall performance. In the proposed framework, the SOC prediction module makes the system aware of what the electric vehicle truly needs, while the DDPG scheduler makes the system aware of when and at what price energy should be delivered.
Several limitations remain. First, the aging model is calibrated with only 100 cycles of experimental data, so the nonlinear knee-point regime is not directly verified by physical experiments. Second, the energy consumption predictor relies on a specific open dataset; rare weather conditions or unusual driving styles can degrade its accuracy. Third, the DDPG algorithm is known to be sensitive to hyperparameter selection. Future research should investigate TD3 or model-based reinforcement learning algorithms and include vehicle-to-grid capability. Another promising extension is to integrate real-time charging pile scheduling with renewable energy sharing among multiple electric vehicles, where the multi-objective formulation becomes significantly more complex but also more realistic.
6. Conclusion
In this work, I present a multi-objective intelligent charging strategy for electric vehicles based on deep reinforcement learning. The strategy is built around three mutually reinforcing components. A high-fidelity electro-thermal-aging coupled model provides the simulation environment and health indicator. A personalized charging cutoff SOC prediction framework, based on travel energy consumption and aging-score optimization, provides the charging target. A DDPG power dispatcher provides the real-time power command under time-of-use tariffs and user departure constraints.
My validation results show that the comprehensive strategy achieves a range coverage above 99.5% for typical users, a charging satisfaction rate of 99.8%, a relative SOH loss reduction of 30.03%, and a charging cost reduction of 22.5% compared with a constant-power benchmark. Ablation experiments demonstrate that the SOC prediction module and the DDPG scheduling module exhibit clear synergy. The proposed method offers a practical and extensible framework for smart charging of electric vehicles in residential scenarios.