Thermal Management Strategy of EV Battery Packs Based on Driving Condition Prediction

Electric vehicles have become a major research focus in the automotive industry because of their advantages in energy saving, environmental protection, and low noise. One of the most critical technical barriers is the thermal behavior of the energy storage system. In particular, the performance of lithium-ion batteries is strongly temperature dependent. Under aggressive driving conditions, such as rapid acceleration or hill climbing, the discharge current drawn from an EV battery pack can increase sharply, leading to a rapid rise in heat generation. If this heat is not removed promptly, the battery temperature can exceed its safe operating window, which accelerates capacity fade and may even trigger thermal runaway. Therefore, designing a rational thermal management strategy for an EV battery pack under high-temperature conditions is essential for improving the safety and reliability of electric vehicles.

In my research, I take a commercial lithium-ion battery as the research object and propose a thermal management strategy for an EV battery pack based on driving-condition prediction. The designed heat-dissipation structure is applied in a liquid-cooling system model, and the predicted driving cycle is coupled with the cooling controller so that the battery can be cooled in advance. This approach overcomes the large time lag of conventional battery thermal management systems. In this article, I summarize the major findings of my thesis work, with an emphasis on thermal modelling, cooling structure design, condition recognition, and predictive control strategy.

1. Thermal Characteristics of Lithium-Ion Batteries

To build a reliable thermal model, I first analyzed the structure and operating principle of a lithium-ion cell inside an EV battery pack. The cell used in this study is a prismatic lithium iron phosphate (LiFePO₄) battery with a nominal capacity of 50 Ah, a nominal voltage of 3.2 V, an internal resistance not exceeding 0.7 mΩ, and an operating voltage range from 2.5 V to 3.65 V. The principal parameters of this cell are summarized in Table 1.

Table 1. Key parameters of the selected lithium-ion cell
Parameter Value
Nominal capacity 50 Ah
Nominal voltage 3.2 V
Maximum continuous charge current 3C
Maximum discharge current 5C
Charge cut-off voltage 3.65 V
Discharge cut-off voltage 2.5 V
Charge temperature 0 °C to 55 °C
Discharge temperature −20 °C to 55 °C

The heat generated by a lithium-ion battery consists of four major parts: reaction heat \(Q_r\), Joule heat \(Q_j\), polarization heat \(Q_p\), and side-reaction heat \(Q_s\). In normal operation, side-reaction heat is negligible. Thus the total heat generation \(Q_t\) of the EV battery pack can be expressed as:

$$
Q_t = Q_r + Q_j + Q_p + Q_s
$$

The Bernardi model is widely used for estimating the heat-generation rate of a cell while accounting for its reversible entropy change and irreversible resistive loss. Under the assumption of uniform heat generation and constant internal resistance, the volumetric heat-generation rate \(q\) is:

$$
q = \frac{I}{V_0} \left[ I R – T \frac{dE_{oc}}{dT} \right]
$$

where \(I\) is the current, \(V_0\) is the battery volume, \(R\) is the equivalent internal resistance, \(T\) is the absolute temperature, and \(dE_{oc}/dT\) is the entropy coefficient. Since the open-circuit voltage \(E_{oc}\) and terminal voltage \(U\) satisfy \(E_{oc}-U = IR\), the above formula reduces to a practical expression for the heat-generation rate.

I measured the internal resistances of six cells at ambient temperature using a battery internal-resistance tester based on the four-terminal Kelvin method. The measured values are given in Table 2.

Table 2. Measured internal resistance values of the tested cells
Cell number 1 2 3 4 5 6
Internal resistance (mΩ) 0.41 0.42 0.42 0.40 0.41 0.40

From these measurements, I used a representative internal resistance of 0.41 mΩ for the thermal calculation. The battery dimensions are 135 mm × 180 mm × 29 mm. Based on the material densities and mass fractions, the equivalent density, specific heat capacity, and anisotropic thermal conductivities of the cell were determined. These parameters are presented in Table 3.

Table 3. Effective thermal properties of the LiFePO₄ cell
Property Symbol Value
Density \(\rho\) 1971.83 kg/m³
Specific heat capacity \(C_p\) 950 J/(kg·K)
In-plane thermal conductivity \(\lambda_x\) 1 W/(m·K)
Through-plane thermal conductivity \(\lambda_y=\lambda_z\) 29 W/(m·K)

Using these properties and the measured internal resistance, I calculated the volumetric heat-generation rates at different discharge rates. The results are reported in Table 4.

Table 4. Heat-generation rates of a single cell at several discharge rates
Discharge rate 1C 2C
Cell heat-generation rate (W/m³) 2242.8 7396.1
Positive tab heat-generation rate (W/m³) 722.6 2890.6
Negative tab heat-generation rate (W/m³) 433.7 1774.9

I built a three-dimensional model of the single cell in SolidWorks and performed finite-volume simulations in Ansys Fluent. The initial temperature was set to 25 °C, and natural convection was considered on the external surfaces. The predicted maximum temperatures at the end of 1C and 2C discharges were 33.08 °C and 40.91 °C, respectively.

To validate the cell model, I performed discharge experiments in a constant-temperature chamber. The battery was first charged at 1C to 3.65 V and then charged at constant voltage until the current dropped to 0.15 A. After resting for one hour, the battery was discharged at 1C or 2C until the cut-off voltage of 2.5 V was reached. The surface temperature at the center of the cell was recorded. The comparison is shown in Table 5.

Table 5. Comparison between simulated and experimental temperatures
Discharge rate Experiment (°C) Simulation (°C) Difference (°C)
1C 32.7 33.0 0.3
2C 40.1 40.7 0.6

The maximum difference between the simulation and the experiment was less than 1 °C, which demonstrates that the established thermal model of the battery is accurate and suitable for subsequent thermal management studies of an EV battery pack.

2. Design and Simulation of Liquid-Cooling Structures

After analyzing various cooling methods, I selected indirect liquid cooling as the thermal management approach for the EV battery pack. In this method, coolant flows through channels inside a metal cold plate, and the battery cells are attached to the cold plate surface. Heat generated inside the cells is conducted through the casing to the cold plate and is finally removed by the coolant. Since the coolant does not directly contact the cells, the method achieves a good balance among cooling performance, safety, and cost.

I designed two families of cold plates: serpentine-channel cold plates and straight-channel cold plates. For each family, two arrangements were considered: the cold plate placed vertically between two layers of cells, and the cold plate placed horizontally under the cells. Thus, four heat-dissipation structures were constructed. The dimensions of all cold plates are identical: total thickness 12 mm, wall thickness 3 mm, and channel thickness 6 mm. Six cells form one module, with a gap of 5 mm between adjacent cells. The resulting structures are identified as:

  • Structure I: serpentine-channel cold plate vertically placed between cells
  • Structure II: serpentine-channel cold plate placed horizontally under the cells
  • Structure III: straight-channel cold plate vertically placed between cells
  • Structure IV: straight-channel cold plate placed horizontally under the cells

Before conducting CFD analyses, I determined the flow regime of the coolant. The Reynolds number is defined as:

$$
Re = \frac{\rho v d_h}{\mu}
$$

where \(\rho\) is the coolant density, \(v\) is the mean velocity, \(\mu\) is the dynamic viscosity, and \(d_h\) is the hydraulic diameter given by:

$$
d_h = \frac{4 A_s}{P_s}
$$

Here, \(A_s\) is the cross-sectional area of the flow passage and \(P_s\) is the wetted perimeter. For water at 25 °C flowing at 0.2 m/s in the proposed channels, the calculated Reynolds number is about 1731.7, which is below 2320. Therefore, the laminar model is appropriate for the simulations.

2.1 Meshing and Grid Independence

I generated mesh models for the complete battery module with the cold plate in Fluent Meshing. Polyhedral cells were used because they offer a balanced combination of accuracy and computational efficiency. The number of mesh elements was gradually increased from 143,592 to 316,815. By comparing the predicted maximum battery temperature and maximum temperature difference under the same boundary condition, I found that when the grid number exceeded 232,110, further refinement changed the results by less than 0.1 °C. Thus, a mesh size of 232,110 elements was selected for the final simulations. The grid-independence data are shown in Table 6.

Table 6. Grid independence check for Structure I
Grid number Maximum temperature (°C) Maximum temperature difference (°C)
143,592 31.71 6.12
180,967 31.44 5.92
232,110 31.27 5.74
284,774 31.24 5.70
316,815 31.23 5.69

2.2 CFD Setup and Comparison of Cooling Structures

In the CFD model, the coolant was pure water with an inlet velocity of 0.2 m/s and an inlet temperature of 25 °C. The pressure outlet boundary condition was set at the outlet. A conjugate heat-transfer solver was enabled, assuming good thermal contact between the batteries and the cold plate. The battery module was discharged at 2C for 1800 seconds. The final temperature and temperature-difference results for the four structures are summarized in Table 7.

Table 7. Cooling performance comparison of four structures
Structure Maximum temperature (°C) Maximum temperature difference (°C)
I (serpentine, vertical) 31.27 5.74
II (serpentine, horizontal) 33.29 7.81
III (straight, vertical) 31.13 5.59
IV (straight, horizontal) 33.12 7.63

The results show that structures with a vertical cold plate between two layers of cells outperform structures with the cold plate placed at the bottom. This is mainly because the direct heat path from the cell core to the coolant is shorter in the vertical arrangement, and the active cooling area is used more effectively. Between the two vertical designs, the straight-channel cold plate (Structure III) produced both a lower maximum temperature and a smaller temperature spread. This structure therefore was selected as the optimum heat-dissipation configuration for subsequent studies.

2.3 Experimental Validation of the Optimum Structure

I constructed an experimental bench consisting of six series-connected cells arranged exactly as Structure III. The cold plate was an aluminum plate with the same channel dimensions as in the simulation. A peristaltic pump was used to drive water at a flow rate of 1.15 L/min, which corresponds to a flow velocity of 0.2 m/s in the channels. The battery module was discharged at 2C, and the surface temperature at the hottest measurement point was recorded. The measured final temperature was 31.0 °C, while the predicted value was 31.13 °C. The maximum deviation over the whole discharge process was only 0.13 °C, confirming the feasibility of the numerical model.

2.4 Influence of Various Factors on Cooling Performance

I then used the optimum Structure III to investigate parametric influences on the cooling performance of the EV battery pack. First, three coolants were compared: pure water, a 50% ethylene-glycol/water mixture, and pure ethylene glycol. Table 8 lists the thermal properties of these fluids.

Table 8. Thermophysical properties of candidate coolants
Coolant Density (kg/m³) Specific heat (J/(kg·K)) Thermal conductivity (W/(m·K)) Dynamic viscosity (Pa·s)
Water 998.2 4182 0.6 0.001003
50% ethylene-glycol/water 1071.13 3301 0.384 0.00373
Ethylene glycol 1113.36 2350 0.29 0.00210

The simulation results under the same boundary conditions are summarized in Table 9. Water showed the best cooling capability, whereas pure ethylene glycol led to the highest temperature rise. However, for real vehicle applications, pure water may freeze at low ambient temperatures. The 50% glycol mixture provides a sufficiently low freezing point while maintaining reasonable thermal conductivity and heat capacity. Based on this trade-off, I selected the 50% ethylene-glycol/water mixture as the preferred coolant for the final thermal management strategy.

Table 9. Simulation results for different coolants
Coolant Maximum temperature (°C) Maximum temperature difference (°C)
Water 31.26 5.63
50% ethylene-glycol/water 31.29 5.66
Ethylene glycol 32.37 7.37

Next, the inlet coolant velocity was varied from 0.05 m/s to 0.4 m/s. The flow regime changes from laminar to turbulent when the velocity exceeds about 0.25 m/s. Table 10 shows the resulting maximum temperatures and temperature differences.

Table 10. Influence of inlet velocity on battery temperature
Inlet velocity (m/s) Reynolds number Maximum temperature (°C) Maximum temperature difference (°C)
0.05 432.9 33.06 7.25
0.1 865.8 31.72 6.31
0.2 1731.7 31.26 5.63
0.3 2597.5 30.90 5.05
0.4 3463.3 30.49 4.94

When the velocity was increased above 0.2 m/s, the temperature reduction became small while the pressure drop in the flow channel rose steeply. Thus, 0.2 m/s was regarded as the optimal inlet velocity for the cooling design.

Finally, I investigated the effect of coolant inlet temperature. The results are summarized in Table 11. A lower coolant temperature decreases the maximum cell temperature but increases the temperature difference inside the module. Considering the need to avoid condensation and maintain temperature uniformity, 25 °C was selected as the baseline coolant inlet temperature.

Table 11. Influence of coolant inlet temperature on battery temperature
Coolant inlet temperature (°C) Maximum temperature (°C) Maximum temperature difference (°C)
20 26.33 6.08
25 31.26 5.63
30 36.08 5.29
35 40.71 4.62

3. Real-World Driving Cycle Collection and Condition Prediction

Traditional thermal management systems react only after the measured temperature has exceeded a threshold. Owing to the thermal mass and delayed heat transfer inside an EV battery pack, this reactive approach can lead to temperature overshoots. To enable predictive cooling, I collected actual driving data from an electric van. The vehicle has an unladen mass of 2995 kg, a fully loaded mass of 4465 kg, a top speed of 100 km/h, a peak motor power of 100 kW, and a peak motor torque of 1000 N·m. The vehicle was driven on public roads using the free-driving method. The speed, state of charge, voltage, current, and other signals were recorded at a sampling interval of one second. After removing invalid data caused by long stops or parking events, the complete time-series speed profile was obtained.

3.1 Segmentation and Feature Extraction

A fixed-step segmentation method was used to divide the overall driving data into small fragments. Each fragment lasts 60 seconds, giving a total of 358 fragments. For every fragment, four characteristic parameters were calculated: average vehicle speed, stopping proportion, deceleration proportion, and constant-speed proportion. These parameters are defined as follows:

$$
\bar{v} = \frac{1}{T} \sum_{t=1}^{T} v(t)
$$

$$
P_{\text{stop}} = \frac{T_{\text{stop}}}{T}, \quad P_{\text{dec}} = \frac{T_{\text{dec}}}{T}, \quad P_{\text{const}} = \frac{T_{\text{const}}}{T}
$$

where \(T\) is the fragment duration, \(T_{\text{stop}}\), \(T_{\text{dec}}\), and \(T_{\text{const}}\) represent the time spent in stopping, decelerating, and nearly constant speed, respectively. An example of the extracted fragments is given in Table 12.

Table 12. Example of characteristic parameters for several fragments
Fragment Average speed (km/h) Stop proportion Deceleration proportion Constant-speed proportion
1 10.92 0.350 0.322 0.317
2 23.10 0.150 0.153 0.167
3 41.95 0 0.119 0.300
4 34.95 0 0.475 0.083
5 6.62 0.767 0.034 0
… … … … …
358 17.40 0.005 0.371 0.267

3.2 K-Means Clustering of Driving Conditions

The K-means algorithm was applied to classify the 358 fragments into representative driving-condition categories. The silhouette coefficient was used to select the number of clusters. By comparing \(k=3,4,5,6\), I found that \(k=4\) produced the highest average silhouette coefficient with the fewest misclassified samples. Therefore, four types of driving conditions were determined. The resulting cluster centers are listed in Table 13.

Table 13. Cluster centers of the four driving conditions
Class Average speed (km/h) Stop proportion Deceleration proportion Constant-speed proportion Interpretation
\(c_1\) 4.93 0.676 0.135 0.071 Congested traffic
\(c_2\) 66.61 0.005 0.359 0.237 Highway condition
\(c_3\) 23.62 0.144 0.381 0.154 Urban condition
\(c_4\) 43.77 0.026 0.426 0.155 Suburban condition

Class \(c_1\) represents severe congestion with low speed and high stop frequency. Class \(c_2\) corresponds to smooth highway driving with a high average speed and very few stops. Class \(c_3\) is a typical city driving condition with moderate speed and frequent stops, while class \(c_4\) represents a suburban driving scenario with moderate-to-high speed and low stop frequency. These four categories cover the diversified operating conditions encountered by an EV battery pack in real use.

3.3 BP Neural Network for Driving-Condition Prediction

To predict the next 60-second driving condition from historical data, I established a back-propagation (BP) neural network. The input layer contains 20 neurons: 16 historical speed values sampled every second from the previous 16 seconds, plus the four characteristic parameters (average speed, stop proportion, deceleration proportion, and constant-speed proportion) of the preceding fragment. The output layer contains 60 neurons, representing the predicted speed values for the next 60 seconds. The network architecture has one hidden layer with 10 neurons. The tansig activation function was used in the hidden layer, while the output layer was kept linear. Table 14 summarizes the training parameters.

Table 14. Configuration of the BP neural network used for condition prediction
Parameter Value or method
Number of input layer neurons 20
Number of hidden layer neurons 10
Number of output layer neurons 60
Hidden layer activation function \(\tanh\)
Training algorithm Levenberg-Marquardt
Maximum iterations 1000
Maximum validation iterations 150
Learning rate 0.1
Training target error \(10^{-6}\)

The first 300 fragments were used to train the network, and the remaining 58 fragments were used for validation. I evaluated the model using the root-mean-square error (RMSE), defined as:

$$
RMSE = \sqrt{ \frac{1}{n} \sum_{i=1}^{n} \left( y_i – \hat{y}_i \right)^2 }
$$

The training RMSE was 0.764, and the maximum relative error between the predicted and actual values was less than 6.3%. For the test set, the RMSE was 0.722, with a maximum relative error of 6.2%. These results show that the BP neural network is able to predict short-term driving conditions with satisfactory accuracy for the subsequent thermal management strategy.

4. Thermal Management Strategy Based on Condition Prediction

4.1 Current Profiles of Different Driving Conditions

Using the four classified driving conditions, I analyzed the current distribution of the EV battery pack under each condition. The highest current values observed in the four categories are listed in Table 15. The congested condition (\(c_1\)) has a very low maximum current because the vehicle speed is low. The highway condition (\(c_2\)) has the highest maximum current because the motor demands high power at high speed. The urban condition (\(c_3\)) has a maximum current of 50 A, and the suburban condition (\(c_4\)) can reach 100 A during rapid acceleration.

Table 15. Maximum discharge current and corresponding coolant flow settings
Condition class Description Maximum discharge current (A) Maximum discharge rate (C) Required flow velocity (m/s) Mass flow rate (kg/s)
\(c_1\) Congested 20 0.4 0.05 0.005
\(c_2\) Highway 120 2.4 0.30 0.031
\(c_3\) Urban 50 1.0 0.10 0.010
\(c_4\) Suburban 100 2.0 0.20 0.021

The relation between the coolant mass flow rate \(\dot{m}\) and the flow velocity \(v\) is:

$$
\dot{m} = \rho v A_c
$$

where \(A_c\) is the total cross-sectional area of the coolant channels. I calculated the required mass flow rate for each condition category to maintain the maximum cell temperature within the safe range. These values are used as the control target of the thermal management strategy.

4.2 Structure of the Proposed Thermal Management Strategy

The proposed thermal management strategy combines the BP-neural-network condition predictor with a PID controller. The control architecture is shown conceptually in the following nested description. At the beginning of each 60-second prediction horizon, the historical speed data are fed into the trained BP network to generate a predicted speed profile of the next 60 seconds. The characteristics of this predicted profile are compared with the four cluster centers, and the nearest condition class is determined by the Euclidean distance. Based on the identified class, the PID controller adjusts the pump speed to realize the required coolant mass flow rate. In this way, the cooling action is taken before the EV battery pack actually enters the predicted demanding condition.

The block diagram of the control strategy can be represented by the following functional relationships:

$$
\dot{m}_{ref} = f\left( \text{mode}_{k}(t+1) \right)
$$

$$
u(t) = k_p e(t) + k_i \int_{0}^{t} e(\tau)d\tau + k_d \frac{de(t)}{dt}
$$

where \(\dot{m}_{ref}\) is the reference coolant mass flow rate determined by the predicted condition mode, \(e(t)\) is the error between the desired battery temperature and the measured battery temperature, and \(k_p, k_i, k_d\) are the PID gains. In the first set of simulations, the PID controller was tuned using the Ziegler-Nichols critical-oscillation method followed by manual fine-tuning. The resulting parameters for the actual (non-predictive) mode and predictive mode are given in Table 16.

Table 16. PID parameters used in simulations
Mode \(k_p\) \(k_i\) \(k_d\)
Actual driving condition 12 0.001 0.0001
Predicted driving condition 10 0.001 0.0001

4.3 AMESim Simulation Models

I built a liquid-cooling system model in AMESim using the optimum straight-channel cold plate structure from the CFD studies. The model includes the battery module, coolant channels, radiator, water pump, and temperature sensor. The cells are connected as 120 series cells and 2 parallel branches, resulting in a battery pack capacity of 100 Ah and a nominal voltage of 384 V. The relevant parameters of the EV battery pack model are summarized in Table 17.

Table 17. Main parameters of the EV battery pack model in AMESim
Parameter Value
Connection configuration 120S2P
Capacity 100 Ah
Nominal voltage 384 V
Coolant type 50% ethylene-glycol/water
Coolant inlet temperature 25 °C
Ambient temperature 25 °C

In addition to the liquid-cooling circuit, I established a complete electric-vehicle driving model in AMESim, which contains longitudinal vehicle dynamics, driver model, motor controller, electric motor, and battery pack. This model allows the EV battery pack to follow a given speed profile while computing the instantaneous discharge current and state of charge.

4.4 Simulation Results: Predictive versus Non-Predictive Strategy

I extracted a test cycle consisting of 20 consecutive 60-second fragments from the recorded data, corresponding to a total driving duration of 1200 s. The predicted speed profile was obtained from the BP model for this test cycle. The predicted condition category for every fragment exactly matched the category obtained by direct classification of the real measured fragment, confirming the reliability of the prediction method. The thermal management strategy was simulated under two scenarios:

  • Actual driving condition scenario: the coolant mass flow rate for the current fragment is determined by the condition class of the previous fragment, which represents a conventional non-predictive delay.
  • Predicted driving condition scenario: the condition class of the next fragment is known from the prediction, and the coolant flow rate is adjusted one fragment in advance.

Figure 7 of the original thesis shows the actual and predicted speed profiles. The simulation results are summarized in Table 18. In the predicted-condition case, the maximum battery temperature was 30.7 °C, whereas in the actual-condition case it was 34.5 °C. Thus, predictive cooling reduced the maximum temperature by 3.8 °C. The maximum module temperature difference was reduced from 1.0 °C to 0.7 °C, indicating better temperature uniformity. Moreover, the rate of temperature rise in the predictive scenario was more gentle and stable.

Table 18. Comparison between actual and predictive thermal management strategies
Metric Actual condition Predicted condition
Maximum battery temperature (°C) 34.5 30.7
Maximum temperature difference (°C) 1.0 0.7
Temperature reduction (°C) — 3.8

4.5 Experimental Validation

To validate the simulation results, I built an experimental platform with a single battery module of six cells. The coolant used was the 50% ethylene-glycol/water mixture. Since the capacity of the tested module is half of the full EV battery pack capacity, the real measured current profile was scaled by 0.5, and the mass flow rate was also halved accordingly. The experiments were performed in a constant-temperature chamber set to 25 °C. Three temperature sensors were placed on the center of the top surface of three representative cells.

In the actual-driving-condition experiment, the coolant flow rate was adjusted according to the previous fragment class, causing a time delay. The maximum measured temperature was 34.5 °C at the cell nearest to the coolant outlet. In the predicted-condition experiment, the flow rate was adjusted one fragment ahead based on the predicted class, and the maximum temperature was 31.1 °C. The maximum temperature difference among the three measurement points in the predictive case was approximately 0.7 °C. These experimental results agree well with the AMESim simulations, with a small discrepancy of about 0.4 °C, and confirm that the proposed predictive thermal management strategy can reduce the maximum temperature of an EV battery pack by more than 3 °C compared with the delayed non-predictive baseline.

4.6 Robustness Analysis

I performed a robustness analysis by varying the ambient temperature from 25 °C to 40 °C while keeping all other settings unchanged. The resulting maximum battery temperatures at the end of the predictive test are shown in Table 19.

Table 19. Robustness of the proposed strategy under different ambient temperatures
Ambient temperature (°C) Maximum battery temperature (°C)
25 30.7
30 30.3
35 30.8
40 31.0

The maximum variation over the considered ambient-temperature range was only about 0.7 °C. The mean absolute deviation was 0.2 °C, indicating that the proposed thermal management strategy is robust enough to maintain the EV battery pack temperature near the design point under various ambient conditions.

4.7 Optimization of the Predictive Thermal Management Strategy

Although the predictive strategy works well, unnecessary coolant flow may still occur when the battery temperature is low or when the predicted condition is not sufficiently severe. To reduce the energy consumption of the cooling pump, I added a temperature-feedback loop to the liquid-cooling system model. The logic is as follows: if the current battery temperature is below 31 °C, the pump speed is reduced so that the coolant mass flow rate is lower than the reference value; if the battery temperature reaches or exceeds 31 °C, the pump speed is increased to the required value corresponding to the predicted condition class. This feedback adjustment can be expressed as:

$$
\dot{m}_{opt} =
\begin{cases}
\alpha \dot{m}_{ref}, & T_{bat} < 31^\circ C \\
\dot{m}_{ref}, & T_{bat} \geq 31^\circ C
\end{cases}
$$

where \(\alpha\) is a reduction coefficient determined by the temperature error through the PID controller. After including this feedback path, I repeated the predictive simulation. The evolution of the coolant mass flow rate before and after optimization is compared in Table 20, where the maximum and minimum flow rates are listed.

Table 20. Comparison of coolant mass flow rates before and after optimization
Metric Before optimization After optimization
Maximum mass flow rate (kg/s) 0.031 0.029
Minimum mass flow rate (kg/s) 0.005 0.0043
Maximum reduction in mass flow rate — 17.6%

The optimized predictive strategy maintained essentially the same cooling performance: the maximum battery temperature remained at 30.7 °C and the maximum temperature difference remained at 0.7 °C. However, the total coolant consumption in the most demanding period was reduced by 17.6%. Therefore, the feedback-optimized predictive thermal management strategy can effectively reduce pump energy consumption without degrading the thermal performance of the EV battery pack.

5. Conclusions

In this work, I investigated the thermal management of an EV battery pack from multiple perspectives: thermal characterization, liquid-cooling structure design, driving-condition prediction, and predictive control strategy. The following conclusions can be drawn from my research:

  1. A finite-volume thermal model of the LiFePO₄ cell was established. The predicted temperatures at 1C and 2C discharges agree with experimental data within 1 °C, which verifies the accuracy of the thermal model.
  2. Among the four proposed liquid-cooling structures, the straight-channel cold plate vertically placed between two layers of cells achieved the best thermal performance. At a 2C discharge rate, the maximum battery temperature was 31.13 °C and the temperature difference was 5.59 °C in the simulation, while the experiment gave a maximum temperature of 31.0 °C.
  3. Water exhibits the best heat-transfer performance among the tested coolants, but a 50% ethylene-glycol/water mixture is preferred for real EV applications because it provides adequate cooling with a lower freezing point.
  4. Both the inlet velocity and the inlet temperature of the coolant significantly affect the maximum temperature and temperature uniformity. An inlet velocity of 0.2 m/s and an inlet temperature of 25 °C were selected as a balanced design.
  5. A K-means clustering method successfully divided the real-world driving data into four classes: congested, highway, urban, and suburban. A BP neural network with 20 inputs, 10 hidden neurons, and 60 outputs predicted the next 60-second speed profile with an RMSE below 0.8 and a maximum relative error below 6.3%.
  6. By using the predicted driving condition to adjust the coolant mass flow rate one fragment in advance, the maximum temperature of the EV battery pack was reduced by 3.8 °C compared with the non-predictive baseline, while the maximum temperature difference was reduced to 0.7 °C. Experimental validation confirmed the simulation trends, with a measured maximum temperature of 31.1 °C under the predictive strategy.
  7. Adding a temperature-feedback loop to the predictive thermal management strategy preserves the cooling performance while reducing the maximum coolant mass flow rate by 17.6%, thereby improving energy efficiency without sacrificing safety.

My proposed predictive thermal management strategy provides a feasible solution to overcome the inherent delay of conventional battery cooling systems. In future work, the strategy can be extended to the entire vehicle thermal management architecture, including the cabin heating and cooling system, and can be integrated with more advanced reinforcement-learning-based controllers to further reduce energy consumption.

Scroll to Top