Data-driven energy management for electric vehicles using offline reinforcement learning.

Area of Science:

• Sustainable Energy
• Artificial Intelligence
• Automotive Engineering

Background:

• Electric vehicle (EV) energy management is crucial for sustainability, but real-world application lags due to simulation limitations.
• Current methods often rely on manual rules or high-fidelity simulations, failing to bridge the gap between theory and practice.
• Optimizing EV energy consumption and system longevity requires practical, data-driven solutions.

Purpose of the Study:

• To introduce a real-world data-driven energy management framework for electric vehicles.
• To leverage offline reinforcement learning (RL) with operational data, eliminating the need for manual rules or simulations.
• To enhance EV performance and reduce system degradation through adaptive learning.

Main Methods:

• Developed an offline reinforcement learning framework utilizing extensive electric vehicle operation data.
• Integrated the framework into existing systems for post-deployment performance enhancement.
• Validated the approach on fuel cell electric vehicles (FCEVs) using real-world data from China.

Main Results:

• The data-driven framework demonstrated superior energy optimization and reduced system degradation in FCEVs.
• Performance improved from 88% to 98.6% of the theoretical optimum with data updates.
• Training on over 60 million kilometers of data allowed the RL agent to generalize across diverse and unseen scenarios.

Conclusions:

• Real-world data-driven methods, particularly offline RL, offer a viable path to optimize EV energy management.
• Large-scale vehicle data utilization significantly enhances energy efficiency and extends vehicle longevity.
• This approach bridges the gap between simulation and practice, improving EV performance and sustainability.