TL;DR
This paper introduces a novel hierarchical control architecture, DP-ECMS, for optimizing the fuel efficiency of connected and automated hybrid electric vehicles. By integrating Approximate Dynamic Programming (ADP) with an Equivalent Consumption Minimization Strategy (ECMS), the approach achieves near-optimal eco-driving performance with significantly reduced computational demands, paving the way for real-time implementation.
What Problem Does This Paper Address?
The increasing prevalence of Connected and Automated Vehicles (CAVs) presents a significant opportunity to improve energy efficiency, particularly for hybrid electric powertrains. Eco-driving, the practice of optimizing vehicle speed and powertrain operation to minimize energy consumption, is a key area for these improvements. However, optimizing both vehicle motion planning (velocity trajectory) and powertrain energy management simultaneously for hybrid vehicles is a complex, computationally intensive problem. Traditional Dynamic Programming (DP) approaches, while optimal, suffer from the "curse of dimensionality," making them impractical for real-time applications. This paper addresses the need for an efficient, yet near-optimal, control strategy that can handle the complexities of hybrid powertrains and adapt to dynamic driving conditions in real-time.
Methods
The core of this work is the DP-ECMS (Dynamic Programming - Equivalent Consumption Minimization Strategy) approach. It's a multi-layer hierarchical control architecture:
- Full-Route Optimization (DP-ECMS): This layer uses DP to determine the optimal velocity trajectory and powertrain torque split over the entire planned route. To manage computational complexity, the state space is reduced by using vehicle speed squared and battery State-of-Charge (SoC) as states, and the powertrain torque as the control input. The ECMS is embedded within the DP framework to handle the torque split, effectively optimizing the balance between engine and electric motor usage.
- Look-Ahead Control (Receding Horizon DP-ECMS): To adapt to real-time variabilities and uncertainties (like traffic signal timing or unexpected speed limit changes), a receding horizon approach is employed. This truncates the full-route horizon to a shorter, manageable look-ahead window. It uses a rollout algorithm, leveraging the value function from the full-route optimization as a terminal cost, to improve policy and adapt to short-term changes without requiring explicit terminal constraints on battery SoC.
- A-Tuning Strategy: A crucial component is the dynamic tuning of the equivalence factor ($\lambda$) for the ECMS. This factor balances the cost of actual fuel consumption against the equivalent cost of battery energy usage. For full-route optimization, a single optimal $\lambda$ is found using a shooting method. For look-ahead control, a grid of $\lambda$ values is explored dynamically for each receding horizon, allowing for more adaptive energy management.
Datasets
The paper uses a forward-looking dynamic powertrain model, which is calibrated and verified using experimental data from the FTP (Federal Test Procedure) regulatory drive cycle, obtained from chassis dynamometer testing. The test route for evaluation is a 7 km mixed urban/highway route in Columbus, Ohio, featuring 6 stop signs and varying speed limits.
Key Insights and Ideas
- Synergy of DP and ECMS: Embedding ECMS within a DP framework allows for leveraging the global optimality of DP while benefiting from the real-time solvability of ECMS for the torque split. This significantly reduces the dimensionality of the control problem for DP.
- Spatial Domain Formulation: Casting the problem in the spatial domain simplifies the incorporation of position-based route information (speed limits, topography).
- Adaptive Equivalence Factor ($\lambda$): Dynamically tuning $\lambda$ in the look-ahead control is crucial for adapting to changing route conditions and achieving better battery usage and "proximity to optimality."
- Computational Efficiency: The DP-ECMS approach offers a substantial reduction in computational effort (over 10x claimed) compared to traditional DP, making real-time implementation feasible.
- Trade-off Management: The $\gamma$ parameter allows for explicit control over the trade-off between travel time and fuel consumption, catering to different driver aggressiveness levels.
Key Results
- Full-Route Optimization: The DP-ECMS approach achieved Pareto fronts that were very close to the benchmark DP, consuming only 0.5%-1% more fuel while taking 1.5%-2% longer to travel the route. Quantitative evaluation showed cost increments always less than 2% compared to the benchmark.
- Look-Ahead Control: This approach further reduced cost increments, consistently staying less than 1% sub-optimal over different $\lambda$ grid sizes. In some cases, cost increments were as low as 0% compared to the benchmark. The look-ahead scheme also showed improved battery usage and a "proximity to optimality."
- Model Validation: The forward vehicle model accurately predicted fuel consumption and battery SoC profiles, with final fuel consumption error less than 4% relative to experimental data.
Limitations and Future Work
- Horizon Length: The choice of horizon length ($N_H$) for the look-ahead control is noted as a compromise and may need generalization based on powertrain configuration and route features.
- Traffic Signal Timing: The current full-route optimization deterministically incorporates signal phase information but cannot handle varying timing.
- Future Work: The paper suggests that the proposed DP-ECMS framework provides a solid foundation for real-time HEV eco-driving applications. Further research could explore more sophisticated adaptive $\lambda$ tuning strategies and investigate the impact of more complex traffic scenarios.
Summary updated