Neural Network-Enhanced Model Predictive Cruise Control for a Heavy Truck with LLM-Assisted Data Collection for Training

Predictive Cruise Control (PCC) is a promising approach for improving fuel efficiency and reducing operational costs in heavy trucks. However, its implementation using conventional Nonlinear Model Predictive Control (NMPC) methods is hindered by computational limitations, often restricting the use of long-horizon slope information. This paper addresses these challenges by proposing a neural network-enhanced slope-adaptive NMPC framework. A Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) architecture is employed to integrate long-horizon slope information and dynamically update control parameters, effectively overcoming computational constraints of traditional NMPC. To further enhance efficiency, an automated simulation scheduling system is developed, leveraging Large Language Models (LLMs) and expert knowledge to optimize parameter tuning and streamline data collection, significantly reducing training overhead. Validation on a high-fidelity simulation platform demonstrates that the proposed method achieves fuel savings of 0.53% on long downhill slopes and 0.88% during incline-to-flat transitions at a constant speed of 72 km/h, outperforming fixed-parameter PCC approaches. The automated simulation scheduling system reduces human involvement in data preparation by 60%, highlighting the potential of integrating LLMs into control systems. These results confirm the feasibility and advantages of the proposed method for real-world applications in fuel-efficient heavy truck operations.