A data-drivenmodeling approach for predictive control ofmixed traffic flow

A data-drivenmodeling approach for predictive control ofmixedtraffic flow

关键词:  Mixed traffic flow · Model predictive control · Cellular automata model · Koopman operator

Key words:  Mixed traffic flow · Model predictive control · Cellular automata model · Koopman operator

摘要:

Effective control of mixed traffic flow remains challenging due to vehicle behavior uncertainty and complex interactions. This paper proposes a data-driven control strategy for connected and autonomous vehicles (CAVs) in mixed traffic flow, implemented through variable speed limits and lane-changing guidance. First, a cellular automata model of mixed traffic flow is developed, with adjustable CAVs’ maximum speed limit and lane-changing probability, thereby linking microscopic CAV operating rules to macroscopic traffic flow dynamics. Second, a recurrent neural network (RNN) is employed to capture the temporal dynamics of the traffic system and predict the evolution of traffic flow states. The RNN is then linearized via the Koopman operator, transforming the complex nonlinear model into a linear representation for the design of a computationally efficient model predictive controller. Finally, simulation results demonstrate that the strategy increases the average traffic speed by 14.2% across 12 traffic scenarios. Specifically, under the challenging conditions of high traffic density with low CAVpenetration, it achieves a 5.22% improvement and promotes amore uniform vehicle distribution. These findings highlight the potential of the proposed approach for mixed traffic flow regulation.

Abstract:

Effective control of mixed traffic flow remains challenging due to vehicle behavior uncertainty and complex interactions. This paper proposes a data-driven control strategy for connected and autonomous vehicles (CAVs) in mixed traffic flow, implemented through variable speed limits and lane-changing guidance. First, a cellular automata model of mixed traffic flow is developed, with adjustable CAVs’ maximum speed limit and lane-changing probability, thereby linking microscopic CAV operating rules to macroscopic traffic flow dynamics. Second, a recurrent neural network (RNN) is employed to capture the temporal dynamics of the traffic system and predict the evolution of traffic flow states. The RNN is then linearized via the Koopman operator, transforming the complex nonlinear model into a linear representation for the design of a computationally efficient model predictive controller. Finally, simulation results demonstrate that the strategy increases the average traffic speed by 14.2% across 12 traffic scenarios. Specifically, under the challenging conditions of high traffic density with low CAVpenetration, it achieves a 5.22% improvement and promotes amore uniform vehicle distribution. These findings highlight the potential of the proposed approach for mixed traffic flow regulation.