Abstract
As a disruptive technology in the power and transport sectors, electric mobility (e-mobility) is receiving increasing attention. E-mobility encompasses the electrification of transportation by means of diversified electric vehicles (EVs) instead of fuel-based vehicles (FVs), promoting low-carbon and sustainable energy development. However, the electrification of driving systems, stochasticity of environments, mobility of infrastructures, and interaction of power and transportation systems significantly increase the analytic difficulty of e-mobility. This paper addresses fundamental issues in e-mobility by proposing a self-learning Digital Twin (DT) framework designed specifically for the unique challenges of EV autonomous driving. By integrating power system dynamics with transportation data, the system enables safer, more efficient driving and route planning without requiring explicit knowledge of complex physical environments.
Introduction: The E-mobility Challenge
Autonomous driving for electric vehicles is significantly more complex than for traditional fuel-based vehicles. EVs are not merely transport; they function as "mobile batteries" that interact deeply with the power grid. This introduces critical variables such as range anxiety, fluctuating electricity prices, and the potential for vehicle-to-grid (V2G) energy trading.
Traditional "mechanism models"—which rely on rigid mathematical or physics-based equations—often struggle to capture the stochastic nature of these interactions. Furthermore, building high-fidelity simulations for large-scale e-mobility is time-consuming and often inaccurate because the inherent dynamics of the system are frequently unknown. This research bridges the gap by using Digital Twins to create a virtual "sandbox" where an EV can learn and plan in a digital space before acting in the physical one.
Methodology: The DT-Empowered Framework
The proposed reference framework consists of four distinct layers:
- Physical E-mobility Layer: Real-world EVs and infrastructure.
- Digital Twin Layer: The core replica containing visualization and modeling components.
- Cyber Layer: Provides computational support via Cloud, 5G/6G, and IoT.
- Service Layer: The interface for user preferences, such as comfort and speed.
The Self-Learning DT Model
The technical core is a self-learning model utilizing three interconnected Neural Networks (NNs):
- NN-Transform: Converts raw physical observations into a "hidden digital state."
- NN-Dynamic: Predicts environmental changes (next state and reward) based on chosen actions.
- NN-Prediction: Generates driving policies and calculates the "value" of specific states.
To handle the continuous variables typical of driving (e.g., steering angle or acceleration), the researchers adapted the MuZero learning framework—originally designed for discrete games like Chess—by implementing discretization and a "progressive widening" strategy within the Monte Carlo tree search (MCTS).
Key Insights: The "Mobile Battery" Paradigm
This framework shifts the focus of autonomous driving from simple navigation to the simultaneous optimization of energy and transport. A key concept is Value Equivalence: instead of modeling every environmental detail, the dynamic model ensures that cumulative rewards in virtual space match real-world outcomes, making the system robust against unknown dynamics.
Furthermore, the system enhances safety through multi-step "lookahead" planning in the digital space, allowing the vehicle to anticipate and avoid accidents that reactive systems might miss. Because the DT model is trained offline and used for fast inference online, it scales effectively to complex multi-agent scenarios, such as EV fleet management.
Experimental Results
The framework was evaluated using simulation environments including Highway-v0 (OpenAI Gym), EV-sim (a specialized Python-based EV driving simulator), and the Sioux Falls Transportation Network benchmark.
- Case I: Highway Driving: The DT-based method achieved a collision rate of only 0.1%, significantly outperforming the Deep Q-Network (DQN) at 1.5% and the Discrete Soft Actor-Critic (DSAC-SF) at 1.1%. It also reached stable, high episode rewards faster than competing methods.
- Case II: Route Planning: When balancing travel time and charging costs under stochastic traffic conditions, the DT method consistently identified superior routes compared to the standard Shortest Path-based Actor-Critic (SP+A2C) method.
Conclusion and Future Work
While the results demonstrate the potential of Digital Twins in e-mobility, the authors identify several areas for future research. These include more accurately estimating the computational and communication overhead of maintaining high-fidelity DTs in real-time and conducting validation on physical EVs in real-world environments. The creation of an open-access digital platform is also suggested to allow the broader research community to refine these models.
Summary updated