Abstract
This paper presents a novel approach to eco-driving for autonomous vehicles (AVs) navigating multiple signalized intersections. By leveraging Vehicle-to-Infrastructure (V2I) communication, the authors develop an optimal acceleration and speed profile that minimizes both travel time and energy consumption. The key innovation is the theoretical proof that the optimal acceleration profile is piecewise linear. This structural identification transforms a complex, infinite-dimensional control problem into a finite-dimensional parametric optimization, enabling real-time analytical solutions. Simulation results demonstrate that joint optimization across multiple intersections significantly outperforms individual intersection management, yielding a 25.26% improvement in energy efficiency.
Introduction
Urban driving is characterized by inefficient "stop-and-go" patterns, which are major contributors to fuel waste and carbon emissions. While Connected and Automated Vehicles (CAVs) can receive traffic light timings in advance, calculating the most efficient trajectory through a sequence of intersections remains a significant computational challenge. Existing numerical methods, such as Dynamic Programming (DP) or Model Predictive Control (MPC), are often too computationally expensive for real-time deployment, especially as the number of intersections increases. Furthermore, many current models treat intersections as isolated events, missing the efficiency gains possible through holistic corridor optimization.
Methodology
Structural Identification
The authors model vehicle dynamics using a double-integrator system (position and velocity) and define a cost function that balances travel time against the integral of acceleration squared (a proxy for energy consumption). Using Pontryagin’s Minimum Principle and the calculus of variations, they prove that the optimal acceleration profile ($u^*(t)$) must take a piecewise linear form ($at + b$).
Parametric Transformation
Because the "shape" of the optimal solution is known to be a series of line segments, the problem shifts from searching for an unknown function to finding a finite set of parameters: slopes, intercepts, and switching times. For $N$ intersections, the profile is characterized by $4(N-1)+3$ triplets of parameters. This reduction in complexity allows the system to solve for the optimal trajectory analytically in real-time.
Safety and Interference
The model is extended to include safety constraints, such as maintaining minimum following distances. This ensures that the eco-driving logic remains robust even when the AV must interact with slower-moving or non-automated interfering traffic.
Key Results
The proposed algorithm was evaluated in a MATLAB simulation environment featuring a vehicle crossing two intersections with 40-second light cycles. The "joint" optimization approach was compared against a baseline that optimized for each intersection individually.
- Overall Efficiency: The multi-intersection parametric approach reduced the total cost (time + energy) by 10.29% compared to the baseline.
- Energy Savings: The energy component of the cost function saw a substantial 25.26% improvement.
- Smooth Jerk: As shown in the acceleration profiles, the proposed method (blue line) maintains continuity, leading to a "smooth jerk" profile that enhances passenger comfort and energy efficiency compared to the abrupt changes seen in individual optimization (red line).
- Holistic Gains: The simulation confirms that "greedy" optimization is sub-optimal. By considering multiple intersections simultaneously, the vehicle may adopt a slightly less efficient trajectory for the first light to ensure a significantly more efficient passage through subsequent signals.
Conclusion and Future Work
This research demonstrates that identifying the underlying structure of optimal control can dramatically simplify real-time AV navigation. By transforming the problem into a parametric optimization, the authors achieve significant energy savings without the computational overhead of traditional numerical methods.
Future work will focus on incorporating more complex, non-linear powertrain dynamics and adapting the model for stochastic environments where traffic light timings may change dynamically based on real-time sensor data. Additionally, scaling the framework to handle saturated traffic flows remains a key area for further investigation.
Summary updated