Abstract
Featuring fast response and flexibility in control allocation, electric vehicles with in-wheel motors (IMDVs) are ideal platforms for advanced vehicle dynamics control. Lateral stability, realized through torque vectoring, is a critical safety function for these vehicles. This paper proposes a novel data-driven Nonlinear Model Predictive Control (NMPC) based on Recurrent High-Order Neural Network (RHONN) modelling. Unlike conventional physics-based methods, the RHONN model uses high-order polynomials to represent nonlinear dynamics directly from data. This model facilitates iterative optimization of steady-state responses (yaw rate and sideslip angle) and improves prediction accuracy during receding horizon control. Validated in CarSim/Simulink, the RHONN-based NMPC demonstrates superior performance and an increased stability margin compared to conventional methods under various driving scenarios.
The Challenge of Lateral Stability
Controlling the lateral stability of a vehicle—preventing dangerous skidding or spinning—is a significant engineering challenge. This is particularly true for in-wheel motor vehicles where each wheel's torque can be independently modulated. Traditional controllers typically rely on simplified physics-based "bicycle models."
While these models are effective for standard driving, they often fail during "limit handling" scenarios, such as high-speed cornering or maneuvers on slippery roads where tire behavior becomes highly nonlinear. Existing empirical solutions, like the "Magic Formula," require extensive pre-calculated datasets and lack the ability to adapt to real-time environmental changes, such as sudden drops in road friction. This research seeks to bridge this gap by employing a neural network that learns and adapts to vehicle physics dynamically.
Methodology: RHONN-based Hierarchical Control
The proposed system utilizes a hierarchical control architecture designed to handle complex nonlinearities without the need for explicit, rigid physical equations.
1. RHONN Modeling
Standard Recurrent Neural Networks (RNNs) often use linear combinations of states. In contrast, the Recurrent High-Order Neural Network (RHONN) employs high-order polynomials, allowing it to capture the intricate, nonlinear relationships inherent in tire dynamics. The model is continuously updated using an Extended Kalman Filter (EKF), enabling it to adapt to varying road friction levels ($\mu$) without manual recalibration.
2. Predictive Control and Target Setting
The NMPC system leverages the RHONN in two critical capacities:
- Target Setting: A "neighbors-searching" algorithm uses the RHONN to identify the optimal steady-state yaw rate and sideslip angle for current conditions, ensuring the controller does not pursue physically impossible states.
- Receding Horizon Control: The NMPC uses the RHONN as an internal predictive model to forecast vehicle reactions to torque commands, selecting the distribution that maximizes stability.
Key Results and Performance Validation
The system was evaluated using high-fidelity CarSim/Simulink co-simulations, focusing on demanding maneuvers like the Double-Lane Change (DLC) on both high-friction ($\mu=0.7$) and low-friction ($\mu=0.35$) surfaces.
- Model Accuracy: The RHONN model significantly outperformed traditional 7-Degree-of-Freedom (7-DoF) models. In longitudinal velocity estimation, the RHONN reduced the Root Mean Square Error (RMSE) by over 90% compared to linear models.
- Expanded Stability Envelope: During slippery road tests, the RHONN-based NMPC maintained stability at speeds up to 74.4 km/h, whereas standard Linear MPC (LMPC) failed above 62.8 km/h—representing an 18.4% improvement in the vehicle's safe operating range.
- Real-time Feasibility: Despite the neural network's complexity, the average computation time was approximately 0.017 ms, well within the requirements for real-time automotive control loops.
Conclusion and Future Directions
The RHONN-based NMPC approach demonstrates that data-driven models can effectively mimic and even surpass traditional physics-based models in complex, nonlinear driving regimes. By providing a more accurate "crystal ball" for predictive control, the system enhances the safety and handling of IMDVs.
Future work will focus on validating these results on physical vehicle prototypes to account for real-world sensor noise and latency. Additionally, researchers aim to explore adaptive weighting strategies for the NMPC cost function to further optimize performance across an even broader range of driving maneuvers.
Summary updated