A Driving Cycle Recognition-Based Hierarchical Control Strategy for Regenerative Braking in Distributed Drive Battery EV Cars

In the pursuit of sustainable mobility, enhancing the energy efficiency of battery EV cars is paramount. During deceleration and braking, a significant portion of the vehicle’s kinetic energy is traditionally dissipated as heat through friction brakes. For a typical battery EV car, this wasted energy can account for up to 50% of the total driving energy in urban environments. Effectively recovering this energy and storing it back into the battery pack is a crucial method for substantially extending the driving range of battery EV cars. Distributed drive architectures, where each wheel is independently powered by an electric motor, offer superior control flexibility for torque vectoring and regenerative braking compared to centralized single-motor configurations. This paper presents a novel hierarchical control strategy for regenerative braking in distributed drive battery EV cars, based on real-time driving cycle recognition, to optimally coordinate mechanical and regenerative braking forces, thereby maximizing energy recuperation while ensuring braking safety.

The fundamental architecture of a distributed drive battery EV car is centered around its powertrain and control modules. This system intelligently manages longitudinal force distribution to coordinate torque output from all four wheels. The core vehicle control module integrates key components such as the Vehicle Control Unit (VCU), the high-voltage battery pack, a DC/DC converter, Space Vector Pulse Width Modulation (SVPWM) inverters, and the in-wheel or near-wheel motors. In this configuration for a battery EV car, the VCU collects vehicle state information and executes control algorithms to manage longitudinal dynamics. The battery provides energy for all electrical systems and simultaneously receives recovered energy during regenerative braking. The power electronics convert DC power to the appropriate AC form for the motors. Crucially, the independent motors not only provide propulsion but also act as generators during braking. A separate hydraulic system, linked to the brake pedal, provides the foundational mechanical braking force. The interaction of these systems in a battery EV car is summarized below:

Component Primary Function in Battery EV Car
Vehicle Control Unit (VCU) Central processor for state monitoring and control algorithm execution.
Battery Pack Energy storage for propulsion and receptacle for regenerated energy.
DC/DC Converter & Inverter Power conversion and conditioning for motor/generator operation.
Distributed Electric Motors Provide independent wheel torque for driving and regenerative braking.
Hydraulic Brake System Provides reliable mechanical braking force, especially for high-demand scenarios.

The foundation for any braking control strategy in a battery EV car is a precise longitudinal dynamics model. The forces acting on the vehicle include the total driving force $F_t$, and the opposing forces: rolling resistance $F_f$, aerodynamic drag $F_w$, grade resistance $F_i$, and acceleration resistance $F_j$. The total driving force from the motors is given by:
$$F_t = \frac{T_t \beta \eta_t}{r}$$
where $T_t$ is the motor’s nominal torque, $\beta$ is the pedal position (0-1), $\eta_t$ is the driveline efficiency, and $r$ is the wheel radius. The grade and acceleration resistances are:
$$F_i = mg\sin\alpha$$
$$F_j = \xi m \frac{du}{dt}$$
Here, $m$ is the mass of the battery EV car, $g$ is gravity, $\alpha$ is the road slope, $u$ is the vehicle speed, $\xi$ is the mass factor accounting for rotational inertia, and $du/dt$ is the acceleration.

During braking, the normal load on each wheel shifts due to longitudinal weight transfer. Ignoring smaller resistances for the ideal case, the normal forces on the front ($F_{z1}$) and rear ($F_{z2}$) axles are:
$$F_{z1} = \frac{mgb \cos\alpha + m \frac{du}{dt} h_g}{L}$$
$$F_{z2} = \frac{mga \cos\alpha – m \frac{du}{dt} h_g}{L}$$
where $a$ and $b$ are distances from the center of gravity to the front and rear axles, $h_g$ is the center of gravity height, and $L$ is the wheelbase ($L=a+b$). The force for an individual wheel $x$ (where x=1: front-left, 2: front-right, 3: rear-left, 4: rear-right) is simply $F_{zx} = F_{z1}/2$ for front wheels or $F_{z2}/2$ for rear wheels.

The wheel slip ratio $\lambda_x$, a critical parameter for braking stability, is defined for each wheel as:
$$\lambda_x = \frac{u – \omega_x r}{u}$$
with $\omega_x$ being the rotational speed of wheel $x$. The tire-road interaction is modeled using the Pacejka Magic Formula, which calculates the longitudinal force $F_x$ as a function of slip:
$$F_x = D \sin[C \arctan(B \phi – E(B \phi – \arctan(B\phi)))]$$
$$\phi = (1-E)(\lambda + S_h) + \frac{E}{B} \arctan(B(\lambda + S_h))$$
where $B$, $C$, $D$, $E$ are stiffness, shape, peak, and curvature factors, and $S_h$ is a horizontal shift. Consequently, the corrected braking torque $T_{nx}$ available at each wheel of the battery EV car, considering the friction limit, is:
$$T_{nx} = r \cdot D \sin[C \arctan(B_s – E(B_s – \arctan(B_s)))]$$
This $T_{nx}$ represents the maximum braking torque that can be applied before wheel lock-up, forming the safety boundary for total brake force allocation.

The core innovation of this work is a hierarchical regenerative braking strategy for battery EV cars that dynamically adapts based on recognized driving conditions. The overarching goal is to maximize the proportion of regenerative braking ($\delta = T_{regen} / T_{total}$) within the constraints of safety and actuator limits. The strategy operates in two layers: an upper-layer driving cycle recognizer and a lower-layer brake force distributor. The upper classifier categorizes real-time driving into “Simple” or “Complex” conditions. Based on this classification, the lower layer employs one of two specialized control algorithms: a Neural Network Inverse Model for complex urban-like conditions prioritizing safety and adaptability, and a Global Optimal Control within safety boundaries for simple, predictable conditions prioritizing pure energy recovery efficiency. This conditional blending allows the battery EV car to always operate with the most appropriate strategy.

Accurate driving cycle recognition is essential for this adaptive battery EV car control system. Simple conditions (e.g., highway, steady suburban driving) are characterized by higher average speed, lower acceleration variance, and minimal stopping. Complex conditions (e.g., dense urban traffic) feature low average speed, frequent high-intensity acceleration/deceleration, and significant idling time. Using a dataset of 55 kinematic segments from standard cycles like UDDS and NEDC, key feature parameters are extracted. The average values for these parameters distinguish the two classes:

Feature Parameter Simple Condition Complex Condition
Average Speed (m/s) 14.92 4.90
Maximum Speed (m/s) 18.61 10.73
Average Acceleration (m/s²) 0.21 0.35
Average Deceleration (m/s²) -0.30 -0.42
Stop Time Percentage (%) 3.15 23.71

Recognition is performed using a Euclidean distance classifier. The feature vector of a current driving snippet is normalized and its distance to the pre-defined simple and complex condition centroids is calculated. The condition class with the smaller Euclidean distance is assigned. This method provides a robust and computationally efficient way for the battery EV car’s controller to understand its operational context in real-time.

For Complex Conditions, safety and adaptability are paramount. A Neural Network (NN) Inverse Model is employed to determine the optimal regenerative braking torque ratio $\delta$. This model learns the complex, non-linear relationship between vehicle state, driver demand, and the optimal safe regeneration level. The network is designed with vehicle speed $u$, braking intensity $z$, and road slope $\alpha$ as inputs, and the regenerative torque ratio $\delta$ as the single output. A feedforward structure with two hidden layers is used. The Levenberg-Marquardt backpropagation algorithm trains the network using a dataset of 378 sample points, with 70% for training, 15% for validation, and 15% for testing. The model employs sigmoid activation functions in the hidden layers and a linear function in the output layer. The performance is excellent, with a Mean Squared Error (MSE) of 0.0010145 and a regression correlation value of 0.98758 for the overall dataset, confirming its high accuracy and generalization capability for the battery EV car’s use in unpredictable environments.

Neural Network Parameter Setting/Choice
Network Type Feedforward, Inverse Model
Inputs Vehicle Speed ($u$), Braking Intensity ($z$), Road Slope ($\alpha$)
Output Regenerative Braking Ratio ($\delta$)
Hidden Layers 2
Training Algorithm Levenberg-Marquardt Backpropagation
Performance (MSE) 0.0010145

For Simple Conditions, where the driving demand is smooth and predictable, the objective shifts to maximizing energy recovery within a clearly defined safety boundary. The safety boundary is established based on braking intensity $z$ (deceleration divided by gravity, $g$). The total required braking torque $T_{req}$ is first calculated from the driver’s pedal input. The allowable range for the regenerative torque ratio $\delta$ is then determined by the braking intensity zone:

  • Low Intensity ($-0.3 < z \leq 0$): Full regeneration is allowed and desired. $\delta = 1$, so $T_{regen} = T_{req}$ (capped at motor capability).
  • Medium Intensity ($-0.5 < z \leq -0.3$): Regeneration must be blended. The safe boundary is $(0.7 – |z|)/0.4 \leq \delta \leq 1$.
  • High Intensity ($-0.7 < z \leq -0.5$): Similar boundary: $(0.7 – |z|)/0.4 \leq \delta \leq 1$, but the lower limit is higher, reducing the regeneration window for stability.
  • Emergency Braking ($z \leq -0.7$): Regeneration is disabled for maximum, predictable mechanical braking. $\delta = 0$.

Within this dynamic safety boundary $[\delta_{min}, \delta_{max}]$, the controller seeks the global optimum $\delta^*$ that maximizes the recovered energy. This requires a model of the motor-generator efficiency map. The load torque $T_L$ of a typical BLDC motor used in battery EV cars can be expressed as:
$$T_L = \frac{e_A i_A + e_B i_B + e_C i_C}{\omega} – J\frac{d\omega}{dt} – B_v$$
where $e$ and $i$ are phase back-EMF and current, $\omega$ is angular speed, $J$ is inertia, and $B_v$ is viscous friction. A simplified efficiency characteristic $\eta_{gen}(T_{regen}, \omega)$ is obtained from this model. The optimization problem for the battery EV car at each time step is then:
$$\delta^* = \arg \max_{\delta \in [\delta_{min}, \delta_{max}]} \left( \int P_{regen} \, dt \right) = \arg \max_{\delta \in [\delta_{min}, \delta_{max}]} \left( \int \omega \cdot T_{regen} \cdot \eta_{gen}(T_{regen}, \omega) \, dt \right)$$
where $T_{regen} = \delta \cdot T_{req}$. A global search algorithm (e.g., GlobalSearch in conjunction with fmincon) is used to reliably find $\delta^*$ within the bounded interval, ensuring the battery EV car captures the maximum possible energy during gentle to moderate braking events on open roads.

The proposed hierarchical control strategy for the battery EV car was rigorously evaluated through co-simulation using MATLAB/Simulink and CarSim, followed by validation on a chassis dynamometer with a real vehicle. The high-fidelity vehicle model in CarSim was coupled with the control algorithm in Simulink. The distributed drive battery EV car model incorporated the longitudinal dynamics, the tire model, and the detailed models of the four independent motor-generators and the hydraulic brake system. The key parameters for the subject battery EV car are listed below:

Vehicle Parameter Value
Vehicle Mass ($m$) 1110 kg
Wheel Radius ($r$) 0.310 m
Wheelbase ($L$) 2.600 m
Front Axle to CG Distance ($a$) 1.040 m
Center of Gravity Height ($h_g$) 0.540 m
Drag Area 1.6 m²

The Worldwide Harmonized Light Vehicles Test Cycle (WLTC) was used as the test cycle, representing a mix of urban, suburban, and highway driving with a total distance of 23.25 km, making it an ideal benchmark for a modern battery EV car. The performance was measured using the Energy Saving Contribution Degree $\varphi_s$, defined as:
$$\varphi_s = \frac{E_{off} – E_{test}}{E_{off}} \times 100\%$$
where $E_{test}$ is the net energy consumed from the battery using a regenerative braking strategy, and $E_{off}$ is the net energy consumed with regenerative braking disabled (only mechanical braking).

The results clearly demonstrate the superiority of the proposed Driving Cycle Recognition-based Hybrid Control (CIC) strategy for the battery EV car. In the WLTC simulation, the proposed (CIC) strategy achieved a $\varphi_s$ of 23.55%. In contrast, a traditional Hierarchical Control (HC) strategy, which does not employ situational adaptation, achieved only 20.32%. This represents a significant improvement of 3.23 percentage points in energy recovery contribution for the battery EV car. The dynamometer tests on the real vehicle confirmed the simulation trends with remarkable consistency. The measured $\varphi_s$ was 23.87% for the CIC strategy and 20.58% for the HC strategy, yielding an improvement of 3.19 percentage points. The close match between simulation and experimental validation underscores the accuracy of the models and the practical effectiveness of the proposed strategy in a real-world battery EV car.

Control Strategy Description Simulation $\varphi_s$ (WLTC) Experimental $\varphi_s$ (WLTC)
No Recovery (NC) Baseline – Only friction braking. 0.00% 0.00%
Hierarchical Control (HC) Traditional fixed-rule strategy. 20.32% 20.58%
Cycle-Recognition Hybrid Control (CIC) Proposed adaptive strategy. 23.55% 23.87%

In conclusion, this paper has developed and validated a sophisticated hierarchical control strategy for regenerative braking in distributed drive battery EV cars. The strategy’s innovation lies in its situational awareness: an upper-layer classifier identifies whether the battery EV car is in a Simple or Complex driving condition. This recognition triggers the most appropriate lower-layer controller—a safety-adaptive Neural Network for complex urban settings or a recuperation-maximizing Global Optimal Controller for predictable highway settings. The co-simulation and real-vehicle dynamometer testing under the comprehensive WLTC cycle conclusively prove the strategy’s efficacy. Compared to a conventional non-adaptive hierarchical control approach, the proposed method increased the Energy Saving Contribution Degree by approximately 3.2%. This substantial gain in energy recovery efficiency directly translates to an extended driving range, highlighting a significant step forward in optimizing the energy management of next-generation battery EV cars. The fusion of driving cycle recognition with specialized control laws presents a powerful framework for enhancing the sustainability and performance of electric vehicles.

Scroll to Top