To address the traffic accident risks caused by livestock crossing grassland highways, this study develops an Unmanned Aerial Vehicle (UAV) adaptive cruise model based on the Ensemble Q-networks and Delayed policy update-based Soft Actor-Critic (ED-SAC) algorithm, aiming to ensure vehicle driving safety on grassland highways. Firstly, a set of mutually independent Q-value networks is constructed. When calculating target values, the minimum value is selected from randomly sampled subsets to break the correlation of estimation errors, thereby effectively suppressing value overestimation bias. Secondly, the update frequencies of the policy and value networks are decoupled, and a delayed policy update mechanism is adopted to ensure that the policy is optimized under the guidance of more convergent and stable value functions. Finally, simulation experiments are conducted in a built environment. The results show that compared with the Proximal Policy Optimization (PPO) and standard SAC algorithms, the ED-SAC algorithm achieves superior performance: in terms of performance upper bound, the average tracking error of the learned final policy is as low as 0.27 m, representing a 40% reduction compared to the 0.45 m of the standard SAC. In terms of task reliability, the algorithm achieves a maximum task success rate of 98.7% in disturbance-free environments, and maintains a high success rate of 96.2% even under severe conditions with continuous random disturbances. These findings fully validate that the ED-SAC algorithm exhibits higher control precision, task success rate, and stronger anti-disturbance capability in various complex trajectory tracking tasks, regardless of the presence of external disturbances. This ensures the UAV's adaptive cruise capability on grassland highways and contributes to enhancing traffic safety levels on such highways.