Optimizing Fuel Consumption of Trucks Using a Machine Learning-Based Gear Selection Strategy

Sammanfattning

Efficient gear selection is a crucial factor for fuel efficiency and drivability in trucks. Traditional rule-based or model-based gear shift strategies may struggle to adapt to varying driving conditions and powertrain configurations, resulting in suboptimal fuel efficiency. This thesis develops a reinforcement learning-based gear selection strategy using a Deep Q-Network to optimize fuel efficiency while generalizing across different driving routes and vehicle configurations. The methodology involves training and validating the reinforcement learning-based model in an in-house Software-in-the-Loop simulation environment. The reward function balances fuel savings and gear shift quality to preserve drivability. The trained model is evaluated in simulation against a commercially leading rule-based gear selection strategy using a paired t-test to assess its effectiveness in fuel savings. Results indicate that the reinforcement learning-based strategy achieves a statistically significant and an average reduction of 2.04% in fuel consumption compared to the baseline on both seen and unseen routes. However, a key limitation of the study is that these fuel consumption improvements were measured under differing average speeds, despite the same target cruise speed being set. This indicates that while the agent achieves better fuel economy, it may do so at the cost of drivability, specifically by settling for slower or less consistent velocity profiles. Nevertheless, this work demonstrates the potential of reinforcement learning for gear selection strategies in trucks. It highlights the need for further exploration of generalization capabilities across different driving conditions and vehicle configurations. Future work could focus on enhancing the model’s robustness and adaptability to real-world implementations, potentially integrating additional factors such as driver behaviour and environmental conditions. The findings contribute to the ongoing research in intelligent transportation systems, particularly in optimizing vehicle performance through advanced machine learning techniques.